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 ";
1062 /* The stack spill tracking logic in check_stack_write()
1063 * and check_stack_read() relies on stack accesses being
1071 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1075 /* truncate register to smaller size (in bytes)
1076 * must be called with size < BPF_REG_SIZE
1078 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1082 /* clear high bits in bit representation */
1083 reg->var_off = tnum_cast(reg->var_off, size);
1085 /* fix arithmetic bounds */
1086 mask = ((u64)1 << (size * 8)) - 1;
1087 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1088 reg->umin_value &= mask;
1089 reg->umax_value &= mask;
1091 reg->umin_value = 0;
1092 reg->umax_value = mask;
1094 reg->smin_value = reg->umin_value;
1095 reg->smax_value = reg->umax_value;
1098 /* check whether memory at (regno + off) is accessible for t = (read | write)
1099 * if t==write, value_regno is a register which value is stored into memory
1100 * if t==read, value_regno is a register which will receive the value from memory
1101 * if t==write && value_regno==-1, some unknown value is stored into memory
1102 * if t==read && value_regno==-1, don't care what we read from memory
1104 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, int off,
1105 int bpf_size, enum bpf_access_type t,
1108 struct bpf_verifier_state *state = env->cur_state;
1109 struct bpf_reg_state *regs = cur_regs(env);
1110 struct bpf_reg_state *reg = regs + regno;
1113 size = bpf_size_to_bytes(bpf_size);
1117 /* alignment checks will add in reg->off themselves */
1118 err = check_ptr_alignment(env, reg, off, size);
1122 /* for access checks, reg->off is just part of off */
1125 if (reg->type == PTR_TO_MAP_VALUE) {
1126 if (t == BPF_WRITE && value_regno >= 0 &&
1127 is_pointer_value(env, value_regno)) {
1128 verbose(env, "R%d leaks addr into map\n", value_regno);
1132 err = check_map_access(env, regno, off, size, false);
1133 if (!err && t == BPF_READ && value_regno >= 0)
1134 mark_reg_unknown(env, regs, value_regno);
1136 } else if (reg->type == PTR_TO_CTX) {
1137 enum bpf_reg_type reg_type = SCALAR_VALUE;
1139 if (t == BPF_WRITE && value_regno >= 0 &&
1140 is_pointer_value(env, value_regno)) {
1141 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1144 /* ctx accesses must be at a fixed offset, so that we can
1145 * determine what type of data were returned.
1149 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1150 regno, reg->off, off - reg->off);
1153 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1156 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1158 "variable ctx access var_off=%s off=%d size=%d",
1162 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1163 if (!err && t == BPF_READ && value_regno >= 0) {
1164 /* ctx access returns either a scalar, or a
1165 * PTR_TO_PACKET[_META,_END]. In the latter
1166 * case, we know the offset is zero.
1168 if (reg_type == SCALAR_VALUE)
1169 mark_reg_unknown(env, regs, value_regno);
1171 mark_reg_known_zero(env, regs,
1173 regs[value_regno].id = 0;
1174 regs[value_regno].off = 0;
1175 regs[value_regno].range = 0;
1176 regs[value_regno].type = reg_type;
1179 } else if (reg->type == PTR_TO_STACK) {
1180 /* stack accesses must be at a fixed offset, so that we can
1181 * determine what type of data were returned.
1182 * See check_stack_read().
1184 if (!tnum_is_const(reg->var_off)) {
1187 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1188 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1192 off += reg->var_off.value;
1193 if (off >= 0 || off < -MAX_BPF_STACK) {
1194 verbose(env, "invalid stack off=%d size=%d\n", off,
1199 if (env->prog->aux->stack_depth < -off)
1200 env->prog->aux->stack_depth = -off;
1203 err = check_stack_write(env, state, off, size,
1206 err = check_stack_read(env, state, off, size,
1208 } else if (reg_is_pkt_pointer(reg)) {
1209 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1210 verbose(env, "cannot write into packet\n");
1213 if (t == BPF_WRITE && value_regno >= 0 &&
1214 is_pointer_value(env, value_regno)) {
1215 verbose(env, "R%d leaks addr into packet\n",
1219 err = check_packet_access(env, regno, off, size, false);
1220 if (!err && t == BPF_READ && value_regno >= 0)
1221 mark_reg_unknown(env, regs, value_regno);
1223 verbose(env, "R%d invalid mem access '%s'\n", regno,
1224 reg_type_str[reg->type]);
1228 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1229 regs[value_regno].type == SCALAR_VALUE) {
1230 /* b/h/w load zero-extends, mark upper bits as known 0 */
1231 coerce_reg_to_size(®s[value_regno], size);
1236 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1240 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1242 verbose(env, "BPF_XADD uses reserved fields\n");
1246 /* check src1 operand */
1247 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1251 /* check src2 operand */
1252 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1256 if (is_pointer_value(env, insn->src_reg)) {
1257 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1261 /* check whether atomic_add can read the memory */
1262 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1263 BPF_SIZE(insn->code), BPF_READ, -1);
1267 /* check whether atomic_add can write into the same memory */
1268 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1269 BPF_SIZE(insn->code), BPF_WRITE, -1);
1272 /* Does this register contain a constant zero? */
1273 static bool register_is_null(struct bpf_reg_state reg)
1275 return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
1278 /* when register 'regno' is passed into function that will read 'access_size'
1279 * bytes from that pointer, make sure that it's within stack boundary
1280 * and all elements of stack are initialized.
1281 * Unlike most pointer bounds-checking functions, this one doesn't take an
1282 * 'off' argument, so it has to add in reg->off itself.
1284 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1285 int access_size, bool zero_size_allowed,
1286 struct bpf_call_arg_meta *meta)
1288 struct bpf_verifier_state *state = env->cur_state;
1289 struct bpf_reg_state *regs = state->regs;
1290 int off, i, slot, spi;
1292 if (regs[regno].type != PTR_TO_STACK) {
1293 /* Allow zero-byte read from NULL, regardless of pointer type */
1294 if (zero_size_allowed && access_size == 0 &&
1295 register_is_null(regs[regno]))
1298 verbose(env, "R%d type=%s expected=%s\n", regno,
1299 reg_type_str[regs[regno].type],
1300 reg_type_str[PTR_TO_STACK]);
1304 /* Only allow fixed-offset stack reads */
1305 if (!tnum_is_const(regs[regno].var_off)) {
1308 tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
1309 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1313 off = regs[regno].off + regs[regno].var_off.value;
1314 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1315 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1316 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1317 regno, off, access_size);
1321 if (env->prog->aux->stack_depth < -off)
1322 env->prog->aux->stack_depth = -off;
1324 if (meta && meta->raw_mode) {
1325 meta->access_size = access_size;
1326 meta->regno = regno;
1330 for (i = 0; i < access_size; i++) {
1331 slot = -(off + i) - 1;
1332 spi = slot / BPF_REG_SIZE;
1333 if (state->allocated_stack <= slot ||
1334 state->stack[spi].slot_type[slot % BPF_REG_SIZE] !=
1336 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1337 off, i, access_size);
1344 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1345 int access_size, bool zero_size_allowed,
1346 struct bpf_call_arg_meta *meta)
1348 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1350 switch (reg->type) {
1352 case PTR_TO_PACKET_META:
1353 return check_packet_access(env, regno, reg->off, access_size,
1355 case PTR_TO_MAP_VALUE:
1356 return check_map_access(env, regno, reg->off, access_size,
1358 default: /* scalar_value|ptr_to_stack or invalid ptr */
1359 return check_stack_boundary(env, regno, access_size,
1360 zero_size_allowed, meta);
1364 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1365 enum bpf_arg_type arg_type,
1366 struct bpf_call_arg_meta *meta)
1368 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1369 enum bpf_reg_type expected_type, type = reg->type;
1372 if (arg_type == ARG_DONTCARE)
1375 err = check_reg_arg(env, regno, SRC_OP);
1379 if (arg_type == ARG_ANYTHING) {
1380 if (is_pointer_value(env, regno)) {
1381 verbose(env, "R%d leaks addr into helper function\n",
1388 if (type_is_pkt_pointer(type) &&
1389 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1390 verbose(env, "helper access to the packet is not allowed\n");
1394 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1395 arg_type == ARG_PTR_TO_MAP_VALUE) {
1396 expected_type = PTR_TO_STACK;
1397 if (!type_is_pkt_pointer(type) &&
1398 type != expected_type)
1400 } else if (arg_type == ARG_CONST_SIZE ||
1401 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1402 expected_type = SCALAR_VALUE;
1403 if (type != expected_type)
1405 } else if (arg_type == ARG_CONST_MAP_PTR) {
1406 expected_type = CONST_PTR_TO_MAP;
1407 if (type != expected_type)
1409 } else if (arg_type == ARG_PTR_TO_CTX) {
1410 expected_type = PTR_TO_CTX;
1411 if (type != expected_type)
1413 } else if (arg_type == ARG_PTR_TO_MEM ||
1414 arg_type == ARG_PTR_TO_MEM_OR_NULL ||
1415 arg_type == ARG_PTR_TO_UNINIT_MEM) {
1416 expected_type = PTR_TO_STACK;
1417 /* One exception here. In case function allows for NULL to be
1418 * passed in as argument, it's a SCALAR_VALUE type. Final test
1419 * happens during stack boundary checking.
1421 if (register_is_null(*reg) &&
1422 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1423 /* final test in check_stack_boundary() */;
1424 else if (!type_is_pkt_pointer(type) &&
1425 type != PTR_TO_MAP_VALUE &&
1426 type != expected_type)
1428 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1430 verbose(env, "unsupported arg_type %d\n", arg_type);
1434 if (arg_type == ARG_CONST_MAP_PTR) {
1435 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1436 meta->map_ptr = reg->map_ptr;
1437 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1438 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1439 * check that [key, key + map->key_size) are within
1440 * stack limits and initialized
1442 if (!meta->map_ptr) {
1443 /* in function declaration map_ptr must come before
1444 * map_key, so that it's verified and known before
1445 * we have to check map_key here. Otherwise it means
1446 * that kernel subsystem misconfigured verifier
1448 verbose(env, "invalid map_ptr to access map->key\n");
1451 if (type_is_pkt_pointer(type))
1452 err = check_packet_access(env, regno, reg->off,
1453 meta->map_ptr->key_size,
1456 err = check_stack_boundary(env, regno,
1457 meta->map_ptr->key_size,
1459 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1460 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1461 * check [value, value + map->value_size) validity
1463 if (!meta->map_ptr) {
1464 /* kernel subsystem misconfigured verifier */
1465 verbose(env, "invalid map_ptr to access map->value\n");
1468 if (type_is_pkt_pointer(type))
1469 err = check_packet_access(env, regno, reg->off,
1470 meta->map_ptr->value_size,
1473 err = check_stack_boundary(env, regno,
1474 meta->map_ptr->value_size,
1476 } else if (arg_type == ARG_CONST_SIZE ||
1477 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1478 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1480 /* bpf_xxx(..., buf, len) call will access 'len' bytes
1481 * from stack pointer 'buf'. Check it
1482 * note: regno == len, regno - 1 == buf
1485 /* kernel subsystem misconfigured verifier */
1487 "ARG_CONST_SIZE cannot be first argument\n");
1491 /* The register is SCALAR_VALUE; the access check
1492 * happens using its boundaries.
1495 if (!tnum_is_const(reg->var_off))
1496 /* For unprivileged variable accesses, disable raw
1497 * mode so that the program is required to
1498 * initialize all the memory that the helper could
1499 * just partially fill up.
1503 if (reg->smin_value < 0) {
1504 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
1509 if (reg->umin_value == 0) {
1510 err = check_helper_mem_access(env, regno - 1, 0,
1517 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
1518 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
1522 err = check_helper_mem_access(env, regno - 1,
1524 zero_size_allowed, meta);
1529 verbose(env, "R%d type=%s expected=%s\n", regno,
1530 reg_type_str[type], reg_type_str[expected_type]);
1534 static int check_map_func_compatibility(struct bpf_verifier_env *env,
1535 struct bpf_map *map, int func_id)
1540 /* We need a two way check, first is from map perspective ... */
1541 switch (map->map_type) {
1542 case BPF_MAP_TYPE_PROG_ARRAY:
1543 if (func_id != BPF_FUNC_tail_call)
1546 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
1547 if (func_id != BPF_FUNC_perf_event_read &&
1548 func_id != BPF_FUNC_perf_event_output &&
1549 func_id != BPF_FUNC_perf_event_read_value)
1552 case BPF_MAP_TYPE_STACK_TRACE:
1553 if (func_id != BPF_FUNC_get_stackid)
1556 case BPF_MAP_TYPE_CGROUP_ARRAY:
1557 if (func_id != BPF_FUNC_skb_under_cgroup &&
1558 func_id != BPF_FUNC_current_task_under_cgroup)
1561 /* devmap returns a pointer to a live net_device ifindex that we cannot
1562 * allow to be modified from bpf side. So do not allow lookup elements
1565 case BPF_MAP_TYPE_DEVMAP:
1566 if (func_id != BPF_FUNC_redirect_map)
1569 /* Restrict bpf side of cpumap, open when use-cases appear */
1570 case BPF_MAP_TYPE_CPUMAP:
1571 if (func_id != BPF_FUNC_redirect_map)
1574 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
1575 case BPF_MAP_TYPE_HASH_OF_MAPS:
1576 if (func_id != BPF_FUNC_map_lookup_elem)
1579 case BPF_MAP_TYPE_SOCKMAP:
1580 if (func_id != BPF_FUNC_sk_redirect_map &&
1581 func_id != BPF_FUNC_sock_map_update &&
1582 func_id != BPF_FUNC_map_delete_elem)
1589 /* ... and second from the function itself. */
1591 case BPF_FUNC_tail_call:
1592 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1595 case BPF_FUNC_perf_event_read:
1596 case BPF_FUNC_perf_event_output:
1597 case BPF_FUNC_perf_event_read_value:
1598 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
1601 case BPF_FUNC_get_stackid:
1602 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
1605 case BPF_FUNC_current_task_under_cgroup:
1606 case BPF_FUNC_skb_under_cgroup:
1607 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
1610 case BPF_FUNC_redirect_map:
1611 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
1612 map->map_type != BPF_MAP_TYPE_CPUMAP)
1615 case BPF_FUNC_sk_redirect_map:
1616 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1619 case BPF_FUNC_sock_map_update:
1620 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
1629 verbose(env, "cannot pass map_type %d into func %s#%d\n",
1630 map->map_type, func_id_name(func_id), func_id);
1634 static int check_raw_mode(const struct bpf_func_proto *fn)
1638 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
1640 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
1642 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
1644 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
1646 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
1649 return count > 1 ? -EINVAL : 0;
1652 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
1653 * are now invalid, so turn them into unknown SCALAR_VALUE.
1655 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
1657 struct bpf_verifier_state *state = env->cur_state;
1658 struct bpf_reg_state *regs = state->regs, *reg;
1661 for (i = 0; i < MAX_BPF_REG; i++)
1662 if (reg_is_pkt_pointer_any(®s[i]))
1663 mark_reg_unknown(env, regs, i);
1665 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
1666 if (state->stack[i].slot_type[0] != STACK_SPILL)
1668 reg = &state->stack[i].spilled_ptr;
1669 if (reg_is_pkt_pointer_any(reg))
1670 __mark_reg_unknown(reg);
1674 static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
1676 const struct bpf_func_proto *fn = NULL;
1677 struct bpf_reg_state *regs;
1678 struct bpf_call_arg_meta meta;
1682 /* find function prototype */
1683 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
1684 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
1689 if (env->ops->get_func_proto)
1690 fn = env->ops->get_func_proto(func_id);
1693 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
1698 /* eBPF programs must be GPL compatible to use GPL-ed functions */
1699 if (!env->prog->gpl_compatible && fn->gpl_only) {
1700 verbose(env, "cannot call GPL only function from proprietary program\n");
1704 /* With LD_ABS/IND some JITs save/restore skb from r1. */
1705 changes_data = bpf_helper_changes_pkt_data(fn->func);
1706 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
1707 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
1708 func_id_name(func_id), func_id);
1712 memset(&meta, 0, sizeof(meta));
1713 meta.pkt_access = fn->pkt_access;
1715 /* We only support one arg being in raw mode at the moment, which
1716 * is sufficient for the helper functions we have right now.
1718 err = check_raw_mode(fn);
1720 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
1721 func_id_name(func_id), func_id);
1726 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
1729 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
1732 if (func_id == BPF_FUNC_tail_call) {
1733 if (meta.map_ptr == NULL) {
1734 verbose(env, "verifier bug\n");
1737 env->insn_aux_data[insn_idx].map_ptr = meta.map_ptr;
1739 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
1742 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
1745 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
1749 /* Mark slots with STACK_MISC in case of raw mode, stack offset
1750 * is inferred from register state.
1752 for (i = 0; i < meta.access_size; i++) {
1753 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, BPF_WRITE, -1);
1758 regs = cur_regs(env);
1759 /* reset caller saved regs */
1760 for (i = 0; i < CALLER_SAVED_REGS; i++) {
1761 mark_reg_not_init(env, regs, caller_saved[i]);
1762 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
1765 /* update return register (already marked as written above) */
1766 if (fn->ret_type == RET_INTEGER) {
1767 /* sets type to SCALAR_VALUE */
1768 mark_reg_unknown(env, regs, BPF_REG_0);
1769 } else if (fn->ret_type == RET_VOID) {
1770 regs[BPF_REG_0].type = NOT_INIT;
1771 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
1772 struct bpf_insn_aux_data *insn_aux;
1774 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
1775 /* There is no offset yet applied, variable or fixed */
1776 mark_reg_known_zero(env, regs, BPF_REG_0);
1777 regs[BPF_REG_0].off = 0;
1778 /* remember map_ptr, so that check_map_access()
1779 * can check 'value_size' boundary of memory access
1780 * to map element returned from bpf_map_lookup_elem()
1782 if (meta.map_ptr == NULL) {
1784 "kernel subsystem misconfigured verifier\n");
1787 regs[BPF_REG_0].map_ptr = meta.map_ptr;
1788 regs[BPF_REG_0].id = ++env->id_gen;
1789 insn_aux = &env->insn_aux_data[insn_idx];
1790 if (!insn_aux->map_ptr)
1791 insn_aux->map_ptr = meta.map_ptr;
1792 else if (insn_aux->map_ptr != meta.map_ptr)
1793 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
1795 verbose(env, "unknown return type %d of func %s#%d\n",
1796 fn->ret_type, func_id_name(func_id), func_id);
1800 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
1805 clear_all_pkt_pointers(env);
1809 static bool signed_add_overflows(s64 a, s64 b)
1811 /* Do the add in u64, where overflow is well-defined */
1812 s64 res = (s64)((u64)a + (u64)b);
1819 static bool signed_sub_overflows(s64 a, s64 b)
1821 /* Do the sub in u64, where overflow is well-defined */
1822 s64 res = (s64)((u64)a - (u64)b);
1829 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
1830 const struct bpf_reg_state *reg,
1831 enum bpf_reg_type type)
1833 bool known = tnum_is_const(reg->var_off);
1834 s64 val = reg->var_off.value;
1835 s64 smin = reg->smin_value;
1837 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
1838 verbose(env, "math between %s pointer and %lld is not allowed\n",
1839 reg_type_str[type], val);
1843 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
1844 verbose(env, "%s pointer offset %d is not allowed\n",
1845 reg_type_str[type], reg->off);
1849 if (smin == S64_MIN) {
1850 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
1851 reg_type_str[type]);
1855 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
1856 verbose(env, "value %lld makes %s pointer be out of bounds\n",
1857 smin, reg_type_str[type]);
1864 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
1865 * Caller should also handle BPF_MOV case separately.
1866 * If we return -EACCES, caller may want to try again treating pointer as a
1867 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
1869 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
1870 struct bpf_insn *insn,
1871 const struct bpf_reg_state *ptr_reg,
1872 const struct bpf_reg_state *off_reg)
1874 struct bpf_reg_state *regs = cur_regs(env), *dst_reg;
1875 bool known = tnum_is_const(off_reg->var_off);
1876 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
1877 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
1878 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
1879 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
1880 u8 opcode = BPF_OP(insn->code);
1881 u32 dst = insn->dst_reg;
1883 dst_reg = ®s[dst];
1885 if (WARN_ON_ONCE(known && (smin_val != smax_val))) {
1886 print_verifier_state(env, env->cur_state);
1888 "verifier internal error: known but bad sbounds\n");
1891 if (WARN_ON_ONCE(known && (umin_val != umax_val))) {
1892 print_verifier_state(env, env->cur_state);
1894 "verifier internal error: known but bad ubounds\n");
1898 if (BPF_CLASS(insn->code) != BPF_ALU64) {
1899 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
1901 "R%d 32-bit pointer arithmetic prohibited\n",
1906 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
1907 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
1911 if (ptr_reg->type == CONST_PTR_TO_MAP) {
1912 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
1916 if (ptr_reg->type == PTR_TO_PACKET_END) {
1917 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
1922 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
1923 * The id may be overwritten later if we create a new variable offset.
1925 dst_reg->type = ptr_reg->type;
1926 dst_reg->id = ptr_reg->id;
1928 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
1929 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
1934 /* We can take a fixed offset as long as it doesn't overflow
1935 * the s32 'off' field
1937 if (known && (ptr_reg->off + smin_val ==
1938 (s64)(s32)(ptr_reg->off + smin_val))) {
1939 /* pointer += K. Accumulate it into fixed offset */
1940 dst_reg->smin_value = smin_ptr;
1941 dst_reg->smax_value = smax_ptr;
1942 dst_reg->umin_value = umin_ptr;
1943 dst_reg->umax_value = umax_ptr;
1944 dst_reg->var_off = ptr_reg->var_off;
1945 dst_reg->off = ptr_reg->off + smin_val;
1946 dst_reg->range = ptr_reg->range;
1949 /* A new variable offset is created. Note that off_reg->off
1950 * == 0, since it's a scalar.
1951 * dst_reg gets the pointer type and since some positive
1952 * integer value was added to the pointer, give it a new 'id'
1953 * if it's a PTR_TO_PACKET.
1954 * this creates a new 'base' pointer, off_reg (variable) gets
1955 * added into the variable offset, and we copy the fixed offset
1958 if (signed_add_overflows(smin_ptr, smin_val) ||
1959 signed_add_overflows(smax_ptr, smax_val)) {
1960 dst_reg->smin_value = S64_MIN;
1961 dst_reg->smax_value = S64_MAX;
1963 dst_reg->smin_value = smin_ptr + smin_val;
1964 dst_reg->smax_value = smax_ptr + smax_val;
1966 if (umin_ptr + umin_val < umin_ptr ||
1967 umax_ptr + umax_val < umax_ptr) {
1968 dst_reg->umin_value = 0;
1969 dst_reg->umax_value = U64_MAX;
1971 dst_reg->umin_value = umin_ptr + umin_val;
1972 dst_reg->umax_value = umax_ptr + umax_val;
1974 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
1975 dst_reg->off = ptr_reg->off;
1976 if (reg_is_pkt_pointer(ptr_reg)) {
1977 dst_reg->id = ++env->id_gen;
1978 /* something was added to pkt_ptr, set range to zero */
1983 if (dst_reg == off_reg) {
1984 /* scalar -= pointer. Creates an unknown scalar */
1985 verbose(env, "R%d tried to subtract pointer from scalar\n",
1989 /* We don't allow subtraction from FP, because (according to
1990 * test_verifier.c test "invalid fp arithmetic", JITs might not
1991 * be able to deal with it.
1993 if (ptr_reg->type == PTR_TO_STACK) {
1994 verbose(env, "R%d subtraction from stack pointer prohibited\n",
1998 if (known && (ptr_reg->off - smin_val ==
1999 (s64)(s32)(ptr_reg->off - smin_val))) {
2000 /* pointer -= K. Subtract it from fixed offset */
2001 dst_reg->smin_value = smin_ptr;
2002 dst_reg->smax_value = smax_ptr;
2003 dst_reg->umin_value = umin_ptr;
2004 dst_reg->umax_value = umax_ptr;
2005 dst_reg->var_off = ptr_reg->var_off;
2006 dst_reg->id = ptr_reg->id;
2007 dst_reg->off = ptr_reg->off - smin_val;
2008 dst_reg->range = ptr_reg->range;
2011 /* A new variable offset is created. If the subtrahend is known
2012 * nonnegative, then any reg->range we had before is still good.
2014 if (signed_sub_overflows(smin_ptr, smax_val) ||
2015 signed_sub_overflows(smax_ptr, smin_val)) {
2016 /* Overflow possible, we know nothing */
2017 dst_reg->smin_value = S64_MIN;
2018 dst_reg->smax_value = S64_MAX;
2020 dst_reg->smin_value = smin_ptr - smax_val;
2021 dst_reg->smax_value = smax_ptr - smin_val;
2023 if (umin_ptr < umax_val) {
2024 /* Overflow possible, we know nothing */
2025 dst_reg->umin_value = 0;
2026 dst_reg->umax_value = U64_MAX;
2028 /* Cannot overflow (as long as bounds are consistent) */
2029 dst_reg->umin_value = umin_ptr - umax_val;
2030 dst_reg->umax_value = umax_ptr - umin_val;
2032 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2033 dst_reg->off = ptr_reg->off;
2034 if (reg_is_pkt_pointer(ptr_reg)) {
2035 dst_reg->id = ++env->id_gen;
2036 /* something was added to pkt_ptr, set range to zero */
2044 /* bitwise ops on pointers are troublesome, prohibit. */
2045 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2046 dst, bpf_alu_string[opcode >> 4]);
2049 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2050 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2051 dst, bpf_alu_string[opcode >> 4]);
2055 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2058 __update_reg_bounds(dst_reg);
2059 __reg_deduce_bounds(dst_reg);
2060 __reg_bound_offset(dst_reg);
2064 /* WARNING: This function does calculations on 64-bit values, but the actual
2065 * execution may occur on 32-bit values. Therefore, things like bitshifts
2066 * need extra checks in the 32-bit case.
2068 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2069 struct bpf_insn *insn,
2070 struct bpf_reg_state *dst_reg,
2071 struct bpf_reg_state src_reg)
2073 struct bpf_reg_state *regs = cur_regs(env);
2074 u8 opcode = BPF_OP(insn->code);
2075 bool src_known, dst_known;
2076 s64 smin_val, smax_val;
2077 u64 umin_val, umax_val;
2078 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2080 smin_val = src_reg.smin_value;
2081 smax_val = src_reg.smax_value;
2082 umin_val = src_reg.umin_value;
2083 umax_val = src_reg.umax_value;
2084 src_known = tnum_is_const(src_reg.var_off);
2085 dst_known = tnum_is_const(dst_reg->var_off);
2088 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2089 __mark_reg_unknown(dst_reg);
2095 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2096 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2097 dst_reg->smin_value = S64_MIN;
2098 dst_reg->smax_value = S64_MAX;
2100 dst_reg->smin_value += smin_val;
2101 dst_reg->smax_value += smax_val;
2103 if (dst_reg->umin_value + umin_val < umin_val ||
2104 dst_reg->umax_value + umax_val < umax_val) {
2105 dst_reg->umin_value = 0;
2106 dst_reg->umax_value = U64_MAX;
2108 dst_reg->umin_value += umin_val;
2109 dst_reg->umax_value += umax_val;
2111 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2114 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2115 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2116 /* Overflow possible, we know nothing */
2117 dst_reg->smin_value = S64_MIN;
2118 dst_reg->smax_value = S64_MAX;
2120 dst_reg->smin_value -= smax_val;
2121 dst_reg->smax_value -= smin_val;
2123 if (dst_reg->umin_value < umax_val) {
2124 /* Overflow possible, we know nothing */
2125 dst_reg->umin_value = 0;
2126 dst_reg->umax_value = U64_MAX;
2128 /* Cannot overflow (as long as bounds are consistent) */
2129 dst_reg->umin_value -= umax_val;
2130 dst_reg->umax_value -= umin_val;
2132 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2135 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2136 if (smin_val < 0 || dst_reg->smin_value < 0) {
2137 /* Ain't nobody got time to multiply that sign */
2138 __mark_reg_unbounded(dst_reg);
2139 __update_reg_bounds(dst_reg);
2142 /* Both values are positive, so we can work with unsigned and
2143 * copy the result to signed (unless it exceeds S64_MAX).
2145 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2146 /* Potential overflow, we know nothing */
2147 __mark_reg_unbounded(dst_reg);
2148 /* (except what we can learn from the var_off) */
2149 __update_reg_bounds(dst_reg);
2152 dst_reg->umin_value *= umin_val;
2153 dst_reg->umax_value *= umax_val;
2154 if (dst_reg->umax_value > S64_MAX) {
2155 /* Overflow possible, we know nothing */
2156 dst_reg->smin_value = S64_MIN;
2157 dst_reg->smax_value = S64_MAX;
2159 dst_reg->smin_value = dst_reg->umin_value;
2160 dst_reg->smax_value = dst_reg->umax_value;
2164 if (src_known && dst_known) {
2165 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2166 src_reg.var_off.value);
2169 /* We get our minimum from the var_off, since that's inherently
2170 * bitwise. Our maximum is the minimum of the operands' maxima.
2172 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2173 dst_reg->umin_value = dst_reg->var_off.value;
2174 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2175 if (dst_reg->smin_value < 0 || smin_val < 0) {
2176 /* Lose signed bounds when ANDing negative numbers,
2177 * ain't nobody got time for that.
2179 dst_reg->smin_value = S64_MIN;
2180 dst_reg->smax_value = S64_MAX;
2182 /* ANDing two positives gives a positive, so safe to
2183 * cast result into s64.
2185 dst_reg->smin_value = dst_reg->umin_value;
2186 dst_reg->smax_value = dst_reg->umax_value;
2188 /* We may learn something more from the var_off */
2189 __update_reg_bounds(dst_reg);
2192 if (src_known && dst_known) {
2193 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2194 src_reg.var_off.value);
2197 /* We get our maximum from the var_off, and our minimum is the
2198 * maximum of the operands' minima
2200 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2201 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2202 dst_reg->umax_value = dst_reg->var_off.value |
2203 dst_reg->var_off.mask;
2204 if (dst_reg->smin_value < 0 || smin_val < 0) {
2205 /* Lose signed bounds when ORing negative numbers,
2206 * ain't nobody got time for that.
2208 dst_reg->smin_value = S64_MIN;
2209 dst_reg->smax_value = S64_MAX;
2211 /* ORing two positives gives a positive, so safe to
2212 * cast result into s64.
2214 dst_reg->smin_value = dst_reg->umin_value;
2215 dst_reg->smax_value = dst_reg->umax_value;
2217 /* We may learn something more from the var_off */
2218 __update_reg_bounds(dst_reg);
2221 if (umax_val >= insn_bitness) {
2222 /* Shifts greater than 31 or 63 are undefined.
2223 * This includes shifts by a negative number.
2225 mark_reg_unknown(env, regs, insn->dst_reg);
2228 /* We lose all sign bit information (except what we can pick
2231 dst_reg->smin_value = S64_MIN;
2232 dst_reg->smax_value = S64_MAX;
2233 /* If we might shift our top bit out, then we know nothing */
2234 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2235 dst_reg->umin_value = 0;
2236 dst_reg->umax_value = U64_MAX;
2238 dst_reg->umin_value <<= umin_val;
2239 dst_reg->umax_value <<= umax_val;
2242 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2244 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2245 /* We may learn something more from the var_off */
2246 __update_reg_bounds(dst_reg);
2249 if (umax_val >= insn_bitness) {
2250 /* Shifts greater than 31 or 63 are undefined.
2251 * This includes shifts by a negative number.
2253 mark_reg_unknown(env, regs, insn->dst_reg);
2256 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2257 * be negative, then either:
2258 * 1) src_reg might be zero, so the sign bit of the result is
2259 * unknown, so we lose our signed bounds
2260 * 2) it's known negative, thus the unsigned bounds capture the
2262 * 3) the signed bounds cross zero, so they tell us nothing
2264 * If the value in dst_reg is known nonnegative, then again the
2265 * unsigned bounts capture the signed bounds.
2266 * Thus, in all cases it suffices to blow away our signed bounds
2267 * and rely on inferring new ones from the unsigned bounds and
2268 * var_off of the result.
2270 dst_reg->smin_value = S64_MIN;
2271 dst_reg->smax_value = S64_MAX;
2273 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2276 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2277 dst_reg->umin_value >>= umax_val;
2278 dst_reg->umax_value >>= umin_val;
2279 /* We may learn something more from the var_off */
2280 __update_reg_bounds(dst_reg);
2283 mark_reg_unknown(env, regs, insn->dst_reg);
2287 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2288 /* 32-bit ALU ops are (32,32)->32 */
2289 coerce_reg_to_size(dst_reg, 4);
2290 coerce_reg_to_size(&src_reg, 4);
2293 __reg_deduce_bounds(dst_reg);
2294 __reg_bound_offset(dst_reg);
2298 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
2301 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
2302 struct bpf_insn *insn)
2304 struct bpf_reg_state *regs = cur_regs(env), *dst_reg, *src_reg;
2305 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
2306 u8 opcode = BPF_OP(insn->code);
2308 dst_reg = ®s[insn->dst_reg];
2310 if (dst_reg->type != SCALAR_VALUE)
2312 if (BPF_SRC(insn->code) == BPF_X) {
2313 src_reg = ®s[insn->src_reg];
2314 if (src_reg->type != SCALAR_VALUE) {
2315 if (dst_reg->type != SCALAR_VALUE) {
2316 /* Combining two pointers by any ALU op yields
2317 * an arbitrary scalar. Disallow all math except
2318 * pointer subtraction
2320 if (opcode == BPF_SUB){
2321 mark_reg_unknown(env, regs, insn->dst_reg);
2324 verbose(env, "R%d pointer %s pointer prohibited\n",
2326 bpf_alu_string[opcode >> 4]);
2329 /* scalar += pointer
2330 * This is legal, but we have to reverse our
2331 * src/dest handling in computing the range
2333 return adjust_ptr_min_max_vals(env, insn,
2336 } else if (ptr_reg) {
2337 /* pointer += scalar */
2338 return adjust_ptr_min_max_vals(env, insn,
2342 /* Pretend the src is a reg with a known value, since we only
2343 * need to be able to read from this state.
2345 off_reg.type = SCALAR_VALUE;
2346 __mark_reg_known(&off_reg, insn->imm);
2348 if (ptr_reg) /* pointer += K */
2349 return adjust_ptr_min_max_vals(env, insn,
2353 /* Got here implies adding two SCALAR_VALUEs */
2354 if (WARN_ON_ONCE(ptr_reg)) {
2355 print_verifier_state(env, env->cur_state);
2356 verbose(env, "verifier internal error: unexpected ptr_reg\n");
2359 if (WARN_ON(!src_reg)) {
2360 print_verifier_state(env, env->cur_state);
2361 verbose(env, "verifier internal error: no src_reg\n");
2364 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
2367 /* check validity of 32-bit and 64-bit arithmetic operations */
2368 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
2370 struct bpf_reg_state *regs = cur_regs(env);
2371 u8 opcode = BPF_OP(insn->code);
2374 if (opcode == BPF_END || opcode == BPF_NEG) {
2375 if (opcode == BPF_NEG) {
2376 if (BPF_SRC(insn->code) != 0 ||
2377 insn->src_reg != BPF_REG_0 ||
2378 insn->off != 0 || insn->imm != 0) {
2379 verbose(env, "BPF_NEG uses reserved fields\n");
2383 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
2384 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
2385 BPF_CLASS(insn->code) == BPF_ALU64) {
2386 verbose(env, "BPF_END uses reserved fields\n");
2391 /* check src operand */
2392 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2396 if (is_pointer_value(env, insn->dst_reg)) {
2397 verbose(env, "R%d pointer arithmetic prohibited\n",
2402 /* check dest operand */
2403 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2407 } else if (opcode == BPF_MOV) {
2409 if (BPF_SRC(insn->code) == BPF_X) {
2410 if (insn->imm != 0 || insn->off != 0) {
2411 verbose(env, "BPF_MOV uses reserved fields\n");
2415 /* check src operand */
2416 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2420 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2421 verbose(env, "BPF_MOV uses reserved fields\n");
2426 /* check dest operand */
2427 err = check_reg_arg(env, insn->dst_reg, DST_OP);
2431 if (BPF_SRC(insn->code) == BPF_X) {
2432 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2434 * copy register state to dest reg
2436 regs[insn->dst_reg] = regs[insn->src_reg];
2437 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
2440 if (is_pointer_value(env, insn->src_reg)) {
2442 "R%d partial copy of pointer\n",
2446 mark_reg_unknown(env, regs, insn->dst_reg);
2447 coerce_reg_to_size(®s[insn->dst_reg], 4);
2451 * remember the value we stored into this reg
2453 regs[insn->dst_reg].type = SCALAR_VALUE;
2454 if (BPF_CLASS(insn->code) == BPF_ALU64) {
2455 __mark_reg_known(regs + insn->dst_reg,
2458 __mark_reg_known(regs + insn->dst_reg,
2463 } else if (opcode > BPF_END) {
2464 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
2467 } else { /* all other ALU ops: and, sub, xor, add, ... */
2469 if (BPF_SRC(insn->code) == BPF_X) {
2470 if (insn->imm != 0 || insn->off != 0) {
2471 verbose(env, "BPF_ALU uses reserved fields\n");
2474 /* check src1 operand */
2475 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2479 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
2480 verbose(env, "BPF_ALU uses reserved fields\n");
2485 /* check src2 operand */
2486 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2490 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
2491 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
2492 verbose(env, "div by zero\n");
2496 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
2497 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
2498 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
2500 if (insn->imm < 0 || insn->imm >= size) {
2501 verbose(env, "invalid shift %d\n", insn->imm);
2506 /* check dest operand */
2507 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
2511 return adjust_reg_min_max_vals(env, insn);
2517 static void find_good_pkt_pointers(struct bpf_verifier_state *state,
2518 struct bpf_reg_state *dst_reg,
2519 enum bpf_reg_type type,
2520 bool range_right_open)
2522 struct bpf_reg_state *regs = state->regs, *reg;
2526 if (dst_reg->off < 0 ||
2527 (dst_reg->off == 0 && range_right_open))
2528 /* This doesn't give us any range */
2531 if (dst_reg->umax_value > MAX_PACKET_OFF ||
2532 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
2533 /* Risk of overflow. For instance, ptr + (1<<63) may be less
2534 * than pkt_end, but that's because it's also less than pkt.
2538 new_range = dst_reg->off;
2539 if (range_right_open)
2542 /* Examples for register markings:
2544 * pkt_data in dst register:
2548 * if (r2 > pkt_end) goto <handle exception>
2553 * if (r2 < pkt_end) goto <access okay>
2554 * <handle exception>
2557 * r2 == dst_reg, pkt_end == src_reg
2558 * r2=pkt(id=n,off=8,r=0)
2559 * r3=pkt(id=n,off=0,r=0)
2561 * pkt_data in src register:
2565 * if (pkt_end >= r2) goto <access okay>
2566 * <handle exception>
2570 * if (pkt_end <= r2) goto <handle exception>
2574 * pkt_end == dst_reg, r2 == src_reg
2575 * r2=pkt(id=n,off=8,r=0)
2576 * r3=pkt(id=n,off=0,r=0)
2578 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
2579 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
2580 * and [r3, r3 + 8-1) respectively is safe to access depending on
2584 /* If our ids match, then we must have the same max_value. And we
2585 * don't care about the other reg's fixed offset, since if it's too big
2586 * the range won't allow anything.
2587 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
2589 for (i = 0; i < MAX_BPF_REG; i++)
2590 if (regs[i].type == type && regs[i].id == dst_reg->id)
2591 /* keep the maximum range already checked */
2592 regs[i].range = max(regs[i].range, new_range);
2594 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2595 if (state->stack[i].slot_type[0] != STACK_SPILL)
2597 reg = &state->stack[i].spilled_ptr;
2598 if (reg->type == type && reg->id == dst_reg->id)
2599 reg->range = max(reg->range, new_range);
2603 /* Adjusts the register min/max values in the case that the dst_reg is the
2604 * variable register that we are working on, and src_reg is a constant or we're
2605 * simply doing a BPF_K check.
2606 * In JEQ/JNE cases we also adjust the var_off values.
2608 static void reg_set_min_max(struct bpf_reg_state *true_reg,
2609 struct bpf_reg_state *false_reg, u64 val,
2612 /* If the dst_reg is a pointer, we can't learn anything about its
2613 * variable offset from the compare (unless src_reg were a pointer into
2614 * the same object, but we don't bother with that.
2615 * Since false_reg and true_reg have the same type by construction, we
2616 * only need to check one of them for pointerness.
2618 if (__is_pointer_value(false, false_reg))
2623 /* If this is false then we know nothing Jon Snow, but if it is
2624 * true then we know for sure.
2626 __mark_reg_known(true_reg, val);
2629 /* If this is true we know nothing Jon Snow, but if it is false
2630 * we know the value for sure;
2632 __mark_reg_known(false_reg, val);
2635 false_reg->umax_value = min(false_reg->umax_value, val);
2636 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2639 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2640 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2643 false_reg->umin_value = max(false_reg->umin_value, val);
2644 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2647 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2648 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2651 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2652 true_reg->umin_value = max(true_reg->umin_value, val);
2655 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2656 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2659 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2660 true_reg->umax_value = min(true_reg->umax_value, val);
2663 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2664 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2670 __reg_deduce_bounds(false_reg);
2671 __reg_deduce_bounds(true_reg);
2672 /* We might have learned some bits from the bounds. */
2673 __reg_bound_offset(false_reg);
2674 __reg_bound_offset(true_reg);
2675 /* Intersecting with the old var_off might have improved our bounds
2676 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2677 * then new var_off is (0; 0x7f...fc) which improves our umax.
2679 __update_reg_bounds(false_reg);
2680 __update_reg_bounds(true_reg);
2683 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
2686 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
2687 struct bpf_reg_state *false_reg, u64 val,
2690 if (__is_pointer_value(false, false_reg))
2695 /* If this is false then we know nothing Jon Snow, but if it is
2696 * true then we know for sure.
2698 __mark_reg_known(true_reg, val);
2701 /* If this is true we know nothing Jon Snow, but if it is false
2702 * we know the value for sure;
2704 __mark_reg_known(false_reg, val);
2707 true_reg->umax_value = min(true_reg->umax_value, val - 1);
2708 false_reg->umin_value = max(false_reg->umin_value, val);
2711 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
2712 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
2715 true_reg->umin_value = max(true_reg->umin_value, val + 1);
2716 false_reg->umax_value = min(false_reg->umax_value, val);
2719 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
2720 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
2723 true_reg->umax_value = min(true_reg->umax_value, val);
2724 false_reg->umin_value = max(false_reg->umin_value, val + 1);
2727 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
2728 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
2731 true_reg->umin_value = max(true_reg->umin_value, val);
2732 false_reg->umax_value = min(false_reg->umax_value, val - 1);
2735 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
2736 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
2742 __reg_deduce_bounds(false_reg);
2743 __reg_deduce_bounds(true_reg);
2744 /* We might have learned some bits from the bounds. */
2745 __reg_bound_offset(false_reg);
2746 __reg_bound_offset(true_reg);
2747 /* Intersecting with the old var_off might have improved our bounds
2748 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2749 * then new var_off is (0; 0x7f...fc) which improves our umax.
2751 __update_reg_bounds(false_reg);
2752 __update_reg_bounds(true_reg);
2755 /* Regs are known to be equal, so intersect their min/max/var_off */
2756 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
2757 struct bpf_reg_state *dst_reg)
2759 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
2760 dst_reg->umin_value);
2761 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
2762 dst_reg->umax_value);
2763 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
2764 dst_reg->smin_value);
2765 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
2766 dst_reg->smax_value);
2767 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
2769 /* We might have learned new bounds from the var_off. */
2770 __update_reg_bounds(src_reg);
2771 __update_reg_bounds(dst_reg);
2772 /* We might have learned something about the sign bit. */
2773 __reg_deduce_bounds(src_reg);
2774 __reg_deduce_bounds(dst_reg);
2775 /* We might have learned some bits from the bounds. */
2776 __reg_bound_offset(src_reg);
2777 __reg_bound_offset(dst_reg);
2778 /* Intersecting with the old var_off might have improved our bounds
2779 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2780 * then new var_off is (0; 0x7f...fc) which improves our umax.
2782 __update_reg_bounds(src_reg);
2783 __update_reg_bounds(dst_reg);
2786 static void reg_combine_min_max(struct bpf_reg_state *true_src,
2787 struct bpf_reg_state *true_dst,
2788 struct bpf_reg_state *false_src,
2789 struct bpf_reg_state *false_dst,
2794 __reg_combine_min_max(true_src, true_dst);
2797 __reg_combine_min_max(false_src, false_dst);
2802 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
2805 struct bpf_reg_state *reg = ®s[regno];
2807 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
2808 /* Old offset (both fixed and variable parts) should
2809 * have been known-zero, because we don't allow pointer
2810 * arithmetic on pointers that might be NULL.
2812 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
2813 !tnum_equals_const(reg->var_off, 0) ||
2815 __mark_reg_known_zero(reg);
2819 reg->type = SCALAR_VALUE;
2820 } else if (reg->map_ptr->inner_map_meta) {
2821 reg->type = CONST_PTR_TO_MAP;
2822 reg->map_ptr = reg->map_ptr->inner_map_meta;
2824 reg->type = PTR_TO_MAP_VALUE;
2826 /* We don't need id from this point onwards anymore, thus we
2827 * should better reset it, so that state pruning has chances
2834 /* The logic is similar to find_good_pkt_pointers(), both could eventually
2835 * be folded together at some point.
2837 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
2840 struct bpf_reg_state *regs = state->regs;
2841 u32 id = regs[regno].id;
2844 for (i = 0; i < MAX_BPF_REG; i++)
2845 mark_map_reg(regs, i, id, is_null);
2847 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2848 if (state->stack[i].slot_type[0] != STACK_SPILL)
2850 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
2854 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
2855 struct bpf_reg_state *dst_reg,
2856 struct bpf_reg_state *src_reg,
2857 struct bpf_verifier_state *this_branch,
2858 struct bpf_verifier_state *other_branch)
2860 if (BPF_SRC(insn->code) != BPF_X)
2863 switch (BPF_OP(insn->code)) {
2865 if ((dst_reg->type == PTR_TO_PACKET &&
2866 src_reg->type == PTR_TO_PACKET_END) ||
2867 (dst_reg->type == PTR_TO_PACKET_META &&
2868 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2869 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
2870 find_good_pkt_pointers(this_branch, dst_reg,
2871 dst_reg->type, false);
2872 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2873 src_reg->type == PTR_TO_PACKET) ||
2874 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2875 src_reg->type == PTR_TO_PACKET_META)) {
2876 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
2877 find_good_pkt_pointers(other_branch, src_reg,
2878 src_reg->type, true);
2884 if ((dst_reg->type == PTR_TO_PACKET &&
2885 src_reg->type == PTR_TO_PACKET_END) ||
2886 (dst_reg->type == PTR_TO_PACKET_META &&
2887 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2888 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
2889 find_good_pkt_pointers(other_branch, dst_reg,
2890 dst_reg->type, true);
2891 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2892 src_reg->type == PTR_TO_PACKET) ||
2893 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2894 src_reg->type == PTR_TO_PACKET_META)) {
2895 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
2896 find_good_pkt_pointers(this_branch, src_reg,
2897 src_reg->type, false);
2903 if ((dst_reg->type == PTR_TO_PACKET &&
2904 src_reg->type == PTR_TO_PACKET_END) ||
2905 (dst_reg->type == PTR_TO_PACKET_META &&
2906 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2907 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
2908 find_good_pkt_pointers(this_branch, dst_reg,
2909 dst_reg->type, true);
2910 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2911 src_reg->type == PTR_TO_PACKET) ||
2912 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2913 src_reg->type == PTR_TO_PACKET_META)) {
2914 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
2915 find_good_pkt_pointers(other_branch, src_reg,
2916 src_reg->type, false);
2922 if ((dst_reg->type == PTR_TO_PACKET &&
2923 src_reg->type == PTR_TO_PACKET_END) ||
2924 (dst_reg->type == PTR_TO_PACKET_META &&
2925 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
2926 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
2927 find_good_pkt_pointers(other_branch, dst_reg,
2928 dst_reg->type, false);
2929 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
2930 src_reg->type == PTR_TO_PACKET) ||
2931 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
2932 src_reg->type == PTR_TO_PACKET_META)) {
2933 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
2934 find_good_pkt_pointers(this_branch, src_reg,
2935 src_reg->type, true);
2947 static int check_cond_jmp_op(struct bpf_verifier_env *env,
2948 struct bpf_insn *insn, int *insn_idx)
2950 struct bpf_verifier_state *other_branch, *this_branch = env->cur_state;
2951 struct bpf_reg_state *regs = this_branch->regs, *dst_reg;
2952 u8 opcode = BPF_OP(insn->code);
2955 if (opcode > BPF_JSLE) {
2956 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
2960 if (BPF_SRC(insn->code) == BPF_X) {
2961 if (insn->imm != 0) {
2962 verbose(env, "BPF_JMP uses reserved fields\n");
2966 /* check src1 operand */
2967 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2971 if (is_pointer_value(env, insn->src_reg)) {
2972 verbose(env, "R%d pointer comparison prohibited\n",
2977 if (insn->src_reg != BPF_REG_0) {
2978 verbose(env, "BPF_JMP uses reserved fields\n");
2983 /* check src2 operand */
2984 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2988 dst_reg = ®s[insn->dst_reg];
2990 /* detect if R == 0 where R was initialized to zero earlier */
2991 if (BPF_SRC(insn->code) == BPF_K &&
2992 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
2993 dst_reg->type == SCALAR_VALUE &&
2994 tnum_equals_const(dst_reg->var_off, insn->imm)) {
2995 if (opcode == BPF_JEQ) {
2996 /* if (imm == imm) goto pc+off;
2997 * only follow the goto, ignore fall-through
2999 *insn_idx += insn->off;
3002 /* if (imm != imm) goto pc+off;
3003 * only follow fall-through branch, since
3004 * that's where the program will go
3010 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3014 /* detect if we are comparing against a constant value so we can adjust
3015 * our min/max values for our dst register.
3016 * this is only legit if both are scalars (or pointers to the same
3017 * object, I suppose, but we don't support that right now), because
3018 * otherwise the different base pointers mean the offsets aren't
3021 if (BPF_SRC(insn->code) == BPF_X) {
3022 if (dst_reg->type == SCALAR_VALUE &&
3023 regs[insn->src_reg].type == SCALAR_VALUE) {
3024 if (tnum_is_const(regs[insn->src_reg].var_off))
3025 reg_set_min_max(&other_branch->regs[insn->dst_reg],
3026 dst_reg, regs[insn->src_reg].var_off.value,
3028 else if (tnum_is_const(dst_reg->var_off))
3029 reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
3030 ®s[insn->src_reg],
3031 dst_reg->var_off.value, opcode);
3032 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3033 /* Comparing for equality, we can combine knowledge */
3034 reg_combine_min_max(&other_branch->regs[insn->src_reg],
3035 &other_branch->regs[insn->dst_reg],
3036 ®s[insn->src_reg],
3037 ®s[insn->dst_reg], opcode);
3039 } else if (dst_reg->type == SCALAR_VALUE) {
3040 reg_set_min_max(&other_branch->regs[insn->dst_reg],
3041 dst_reg, insn->imm, opcode);
3044 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3045 if (BPF_SRC(insn->code) == BPF_K &&
3046 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3047 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3048 /* Mark all identical map registers in each branch as either
3049 * safe or unknown depending R == 0 or R != 0 conditional.
3051 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3052 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3053 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3054 this_branch, other_branch) &&
3055 is_pointer_value(env, insn->dst_reg)) {
3056 verbose(env, "R%d pointer comparison prohibited\n",
3061 print_verifier_state(env, this_branch);
3065 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3066 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3068 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3070 return (struct bpf_map *) (unsigned long) imm64;
3073 /* verify BPF_LD_IMM64 instruction */
3074 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3076 struct bpf_reg_state *regs = cur_regs(env);
3079 if (BPF_SIZE(insn->code) != BPF_DW) {
3080 verbose(env, "invalid BPF_LD_IMM insn\n");
3083 if (insn->off != 0) {
3084 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3088 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3092 if (insn->src_reg == 0) {
3093 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3095 regs[insn->dst_reg].type = SCALAR_VALUE;
3096 __mark_reg_known(®s[insn->dst_reg], imm);
3100 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3101 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3103 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3104 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3108 static bool may_access_skb(enum bpf_prog_type type)
3111 case BPF_PROG_TYPE_SOCKET_FILTER:
3112 case BPF_PROG_TYPE_SCHED_CLS:
3113 case BPF_PROG_TYPE_SCHED_ACT:
3120 /* verify safety of LD_ABS|LD_IND instructions:
3121 * - they can only appear in the programs where ctx == skb
3122 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3123 * preserve R6-R9, and store return value into R0
3126 * ctx == skb == R6 == CTX
3129 * SRC == any register
3130 * IMM == 32-bit immediate
3133 * R0 - 8/16/32-bit skb data converted to cpu endianness
3135 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3137 struct bpf_reg_state *regs = cur_regs(env);
3138 u8 mode = BPF_MODE(insn->code);
3141 if (!may_access_skb(env->prog->type)) {
3142 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3146 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3147 BPF_SIZE(insn->code) == BPF_DW ||
3148 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3149 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3153 /* check whether implicit source operand (register R6) is readable */
3154 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3158 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3160 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3164 if (mode == BPF_IND) {
3165 /* check explicit source operand */
3166 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3171 /* reset caller saved regs to unreadable */
3172 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3173 mark_reg_not_init(env, regs, caller_saved[i]);
3174 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3177 /* mark destination R0 register as readable, since it contains
3178 * the value fetched from the packet.
3179 * Already marked as written above.
3181 mark_reg_unknown(env, regs, BPF_REG_0);
3185 static int check_return_code(struct bpf_verifier_env *env)
3187 struct bpf_reg_state *reg;
3188 struct tnum range = tnum_range(0, 1);
3190 switch (env->prog->type) {
3191 case BPF_PROG_TYPE_CGROUP_SKB:
3192 case BPF_PROG_TYPE_CGROUP_SOCK:
3193 case BPF_PROG_TYPE_SOCK_OPS:
3194 case BPF_PROG_TYPE_CGROUP_DEVICE:
3200 reg = cur_regs(env) + BPF_REG_0;
3201 if (reg->type != SCALAR_VALUE) {
3202 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
3203 reg_type_str[reg->type]);
3207 if (!tnum_in(range, reg->var_off)) {
3208 verbose(env, "At program exit the register R0 ");
3209 if (!tnum_is_unknown(reg->var_off)) {
3212 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3213 verbose(env, "has value %s", tn_buf);
3215 verbose(env, "has unknown scalar value");
3217 verbose(env, " should have been 0 or 1\n");
3223 /* non-recursive DFS pseudo code
3224 * 1 procedure DFS-iterative(G,v):
3225 * 2 label v as discovered
3226 * 3 let S be a stack
3228 * 5 while S is not empty
3230 * 7 if t is what we're looking for:
3232 * 9 for all edges e in G.adjacentEdges(t) do
3233 * 10 if edge e is already labelled
3234 * 11 continue with the next edge
3235 * 12 w <- G.adjacentVertex(t,e)
3236 * 13 if vertex w is not discovered and not explored
3237 * 14 label e as tree-edge
3238 * 15 label w as discovered
3241 * 18 else if vertex w is discovered
3242 * 19 label e as back-edge
3244 * 21 // vertex w is explored
3245 * 22 label e as forward- or cross-edge
3246 * 23 label t as explored
3251 * 0x11 - discovered and fall-through edge labelled
3252 * 0x12 - discovered and fall-through and branch edges labelled
3263 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3265 static int *insn_stack; /* stack of insns to process */
3266 static int cur_stack; /* current stack index */
3267 static int *insn_state;
3269 /* t, w, e - match pseudo-code above:
3270 * t - index of current instruction
3271 * w - next instruction
3274 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
3276 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
3279 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
3282 if (w < 0 || w >= env->prog->len) {
3283 verbose(env, "jump out of range from insn %d to %d\n", t, w);
3288 /* mark branch target for state pruning */
3289 env->explored_states[w] = STATE_LIST_MARK;
3291 if (insn_state[w] == 0) {
3293 insn_state[t] = DISCOVERED | e;
3294 insn_state[w] = DISCOVERED;
3295 if (cur_stack >= env->prog->len)
3297 insn_stack[cur_stack++] = w;
3299 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
3300 verbose(env, "back-edge from insn %d to %d\n", t, w);
3302 } else if (insn_state[w] == EXPLORED) {
3303 /* forward- or cross-edge */
3304 insn_state[t] = DISCOVERED | e;
3306 verbose(env, "insn state internal bug\n");
3312 /* non-recursive depth-first-search to detect loops in BPF program
3313 * loop == back-edge in directed graph
3315 static int check_cfg(struct bpf_verifier_env *env)
3317 struct bpf_insn *insns = env->prog->insnsi;
3318 int insn_cnt = env->prog->len;
3322 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3326 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
3332 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
3333 insn_stack[0] = 0; /* 0 is the first instruction */
3339 t = insn_stack[cur_stack - 1];
3341 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
3342 u8 opcode = BPF_OP(insns[t].code);
3344 if (opcode == BPF_EXIT) {
3346 } else if (opcode == BPF_CALL) {
3347 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3352 if (t + 1 < insn_cnt)
3353 env->explored_states[t + 1] = STATE_LIST_MARK;
3354 } else if (opcode == BPF_JA) {
3355 if (BPF_SRC(insns[t].code) != BPF_K) {
3359 /* unconditional jump with single edge */
3360 ret = push_insn(t, t + insns[t].off + 1,
3366 /* tell verifier to check for equivalent states
3367 * after every call and jump
3369 if (t + 1 < insn_cnt)
3370 env->explored_states[t + 1] = STATE_LIST_MARK;
3372 /* conditional jump with two edges */
3373 env->explored_states[t] = STATE_LIST_MARK;
3374 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3380 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
3387 /* all other non-branch instructions with single
3390 ret = push_insn(t, t + 1, FALLTHROUGH, env);
3398 insn_state[t] = EXPLORED;
3399 if (cur_stack-- <= 0) {
3400 verbose(env, "pop stack internal bug\n");
3407 for (i = 0; i < insn_cnt; i++) {
3408 if (insn_state[i] != EXPLORED) {
3409 verbose(env, "unreachable insn %d\n", i);
3414 ret = 0; /* cfg looks good */
3422 /* check %cur's range satisfies %old's */
3423 static bool range_within(struct bpf_reg_state *old,
3424 struct bpf_reg_state *cur)
3426 return old->umin_value <= cur->umin_value &&
3427 old->umax_value >= cur->umax_value &&
3428 old->smin_value <= cur->smin_value &&
3429 old->smax_value >= cur->smax_value;
3432 /* Maximum number of register states that can exist at once */
3433 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
3439 /* If in the old state two registers had the same id, then they need to have
3440 * the same id in the new state as well. But that id could be different from
3441 * the old state, so we need to track the mapping from old to new ids.
3442 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
3443 * regs with old id 5 must also have new id 9 for the new state to be safe. But
3444 * regs with a different old id could still have new id 9, we don't care about
3446 * So we look through our idmap to see if this old id has been seen before. If
3447 * so, we require the new id to match; otherwise, we add the id pair to the map.
3449 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
3453 for (i = 0; i < ID_MAP_SIZE; i++) {
3454 if (!idmap[i].old) {
3455 /* Reached an empty slot; haven't seen this id before */
3456 idmap[i].old = old_id;
3457 idmap[i].cur = cur_id;
3460 if (idmap[i].old == old_id)
3461 return idmap[i].cur == cur_id;
3463 /* We ran out of idmap slots, which should be impossible */
3468 /* Returns true if (rold safe implies rcur safe) */
3469 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
3470 struct idpair *idmap)
3472 if (!(rold->live & REG_LIVE_READ))
3473 /* explored state didn't use this */
3476 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, live)) == 0)
3479 if (rold->type == NOT_INIT)
3480 /* explored state can't have used this */
3482 if (rcur->type == NOT_INIT)
3484 switch (rold->type) {
3486 if (rcur->type == SCALAR_VALUE) {
3487 /* new val must satisfy old val knowledge */
3488 return range_within(rold, rcur) &&
3489 tnum_in(rold->var_off, rcur->var_off);
3491 /* We're trying to use a pointer in place of a scalar.
3492 * Even if the scalar was unbounded, this could lead to
3493 * pointer leaks because scalars are allowed to leak
3494 * while pointers are not. We could make this safe in
3495 * special cases if root is calling us, but it's
3496 * probably not worth the hassle.
3500 case PTR_TO_MAP_VALUE:
3501 /* If the new min/max/var_off satisfy the old ones and
3502 * everything else matches, we are OK.
3503 * We don't care about the 'id' value, because nothing
3504 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
3506 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
3507 range_within(rold, rcur) &&
3508 tnum_in(rold->var_off, rcur->var_off);
3509 case PTR_TO_MAP_VALUE_OR_NULL:
3510 /* a PTR_TO_MAP_VALUE could be safe to use as a
3511 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
3512 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
3513 * checked, doing so could have affected others with the same
3514 * id, and we can't check for that because we lost the id when
3515 * we converted to a PTR_TO_MAP_VALUE.
3517 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
3519 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
3521 /* Check our ids match any regs they're supposed to */
3522 return check_ids(rold->id, rcur->id, idmap);
3523 case PTR_TO_PACKET_META:
3525 if (rcur->type != rold->type)
3527 /* We must have at least as much range as the old ptr
3528 * did, so that any accesses which were safe before are
3529 * still safe. This is true even if old range < old off,
3530 * since someone could have accessed through (ptr - k), or
3531 * even done ptr -= k in a register, to get a safe access.
3533 if (rold->range > rcur->range)
3535 /* If the offsets don't match, we can't trust our alignment;
3536 * nor can we be sure that we won't fall out of range.
3538 if (rold->off != rcur->off)
3540 /* id relations must be preserved */
3541 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
3543 /* new val must satisfy old val knowledge */
3544 return range_within(rold, rcur) &&
3545 tnum_in(rold->var_off, rcur->var_off);
3547 case CONST_PTR_TO_MAP:
3549 case PTR_TO_PACKET_END:
3550 /* Only valid matches are exact, which memcmp() above
3551 * would have accepted
3554 /* Don't know what's going on, just say it's not safe */
3558 /* Shouldn't get here; if we do, say it's not safe */
3563 static bool stacksafe(struct bpf_verifier_state *old,
3564 struct bpf_verifier_state *cur,
3565 struct idpair *idmap)
3569 /* if explored stack has more populated slots than current stack
3570 * such stacks are not equivalent
3572 if (old->allocated_stack > cur->allocated_stack)
3575 /* walk slots of the explored stack and ignore any additional
3576 * slots in the current stack, since explored(safe) state
3579 for (i = 0; i < old->allocated_stack; i++) {
3580 spi = i / BPF_REG_SIZE;
3582 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
3584 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
3585 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
3586 /* Ex: old explored (safe) state has STACK_SPILL in
3587 * this stack slot, but current has has STACK_MISC ->
3588 * this verifier states are not equivalent,
3589 * return false to continue verification of this path
3592 if (i % BPF_REG_SIZE)
3594 if (old->stack[spi].slot_type[0] != STACK_SPILL)
3596 if (!regsafe(&old->stack[spi].spilled_ptr,
3597 &cur->stack[spi].spilled_ptr,
3599 /* when explored and current stack slot are both storing
3600 * spilled registers, check that stored pointers types
3601 * are the same as well.
3602 * Ex: explored safe path could have stored
3603 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
3604 * but current path has stored:
3605 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
3606 * such verifier states are not equivalent.
3607 * return false to continue verification of this path
3614 /* compare two verifier states
3616 * all states stored in state_list are known to be valid, since
3617 * verifier reached 'bpf_exit' instruction through them
3619 * this function is called when verifier exploring different branches of
3620 * execution popped from the state stack. If it sees an old state that has
3621 * more strict register state and more strict stack state then this execution
3622 * branch doesn't need to be explored further, since verifier already
3623 * concluded that more strict state leads to valid finish.
3625 * Therefore two states are equivalent if register state is more conservative
3626 * and explored stack state is more conservative than the current one.
3629 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
3630 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
3632 * In other words if current stack state (one being explored) has more
3633 * valid slots than old one that already passed validation, it means
3634 * the verifier can stop exploring and conclude that current state is valid too
3636 * Similarly with registers. If explored state has register type as invalid
3637 * whereas register type in current state is meaningful, it means that
3638 * the current state will reach 'bpf_exit' instruction safely
3640 static bool states_equal(struct bpf_verifier_env *env,
3641 struct bpf_verifier_state *old,
3642 struct bpf_verifier_state *cur)
3644 struct idpair *idmap;
3648 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
3649 /* If we failed to allocate the idmap, just say it's not safe */
3653 for (i = 0; i < MAX_BPF_REG; i++) {
3654 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
3658 if (!stacksafe(old, cur, idmap))
3666 /* A write screens off any subsequent reads; but write marks come from the
3667 * straight-line code between a state and its parent. When we arrive at a
3668 * jump target (in the first iteration of the propagate_liveness() loop),
3669 * we didn't arrive by the straight-line code, so read marks in state must
3670 * propagate to parent regardless of state's write marks.
3672 static bool do_propagate_liveness(const struct bpf_verifier_state *state,
3673 struct bpf_verifier_state *parent)
3675 bool writes = parent == state->parent; /* Observe write marks */
3676 bool touched = false; /* any changes made? */
3681 /* Propagate read liveness of registers... */
3682 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
3683 /* We don't need to worry about FP liveness because it's read-only */
3684 for (i = 0; i < BPF_REG_FP; i++) {
3685 if (parent->regs[i].live & REG_LIVE_READ)
3687 if (writes && (state->regs[i].live & REG_LIVE_WRITTEN))
3689 if (state->regs[i].live & REG_LIVE_READ) {
3690 parent->regs[i].live |= REG_LIVE_READ;
3694 /* ... and stack slots */
3695 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
3696 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
3697 if (parent->stack[i].slot_type[0] != STACK_SPILL)
3699 if (state->stack[i].slot_type[0] != STACK_SPILL)
3701 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
3704 (state->stack[i].spilled_ptr.live & REG_LIVE_WRITTEN))
3706 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ) {
3707 parent->stack[i].spilled_ptr.live |= REG_LIVE_READ;
3714 /* "parent" is "a state from which we reach the current state", but initially
3715 * it is not the state->parent (i.e. "the state whose straight-line code leads
3716 * to the current state"), instead it is the state that happened to arrive at
3717 * a (prunable) equivalent of the current state. See comment above
3718 * do_propagate_liveness() for consequences of this.
3719 * This function is just a more efficient way of calling mark_reg_read() or
3720 * mark_stack_slot_read() on each reg in "parent" that is read in "state",
3721 * though it requires that parent != state->parent in the call arguments.
3723 static void propagate_liveness(const struct bpf_verifier_state *state,
3724 struct bpf_verifier_state *parent)
3726 while (do_propagate_liveness(state, parent)) {
3727 /* Something changed, so we need to feed those changes onward */
3729 parent = state->parent;
3733 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
3735 struct bpf_verifier_state_list *new_sl;
3736 struct bpf_verifier_state_list *sl;
3737 struct bpf_verifier_state *cur = env->cur_state;
3740 sl = env->explored_states[insn_idx];
3742 /* this 'insn_idx' instruction wasn't marked, so we will not
3743 * be doing state search here
3747 while (sl != STATE_LIST_MARK) {
3748 if (states_equal(env, &sl->state, cur)) {
3749 /* reached equivalent register/stack state,
3751 * Registers read by the continuation are read by us.
3752 * If we have any write marks in env->cur_state, they
3753 * will prevent corresponding reads in the continuation
3754 * from reaching our parent (an explored_state). Our
3755 * own state will get the read marks recorded, but
3756 * they'll be immediately forgotten as we're pruning
3757 * this state and will pop a new one.
3759 propagate_liveness(&sl->state, cur);
3765 /* there were no equivalent states, remember current one.
3766 * technically the current state is not proven to be safe yet,
3767 * but it will either reach bpf_exit (which means it's safe) or
3768 * it will be rejected. Since there are no loops, we won't be
3769 * seeing this 'insn_idx' instruction again on the way to bpf_exit
3771 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
3775 /* add new state to the head of linked list */
3776 err = copy_verifier_state(&new_sl->state, cur);
3778 free_verifier_state(&new_sl->state, false);
3782 new_sl->next = env->explored_states[insn_idx];
3783 env->explored_states[insn_idx] = new_sl;
3784 /* connect new state to parentage chain */
3785 cur->parent = &new_sl->state;
3786 /* clear write marks in current state: the writes we did are not writes
3787 * our child did, so they don't screen off its reads from us.
3788 * (There are no read marks in current state, because reads always mark
3789 * their parent and current state never has children yet. Only
3790 * explored_states can get read marks.)
3792 for (i = 0; i < BPF_REG_FP; i++)
3793 cur->regs[i].live = REG_LIVE_NONE;
3794 for (i = 0; i < cur->allocated_stack / BPF_REG_SIZE; i++)
3795 if (cur->stack[i].slot_type[0] == STACK_SPILL)
3796 cur->stack[i].spilled_ptr.live = REG_LIVE_NONE;
3800 static int ext_analyzer_insn_hook(struct bpf_verifier_env *env,
3801 int insn_idx, int prev_insn_idx)
3803 if (env->dev_ops && env->dev_ops->insn_hook)
3804 return env->dev_ops->insn_hook(env, insn_idx, prev_insn_idx);
3809 static int do_check(struct bpf_verifier_env *env)
3811 struct bpf_verifier_state *state;
3812 struct bpf_insn *insns = env->prog->insnsi;
3813 struct bpf_reg_state *regs;
3814 int insn_cnt = env->prog->len;
3815 int insn_idx, prev_insn_idx = 0;
3816 int insn_processed = 0;
3817 bool do_print_state = false;
3819 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
3822 env->cur_state = state;
3823 init_reg_state(env, state->regs);
3824 state->parent = NULL;
3827 struct bpf_insn *insn;
3831 if (insn_idx >= insn_cnt) {
3832 verbose(env, "invalid insn idx %d insn_cnt %d\n",
3833 insn_idx, insn_cnt);
3837 insn = &insns[insn_idx];
3838 class = BPF_CLASS(insn->code);
3840 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
3842 "BPF program is too large. Processed %d insn\n",
3847 err = is_state_visited(env, insn_idx);
3851 /* found equivalent state, can prune the search */
3852 if (env->log.level) {
3854 verbose(env, "\nfrom %d to %d: safe\n",
3855 prev_insn_idx, insn_idx);
3857 verbose(env, "%d: safe\n", insn_idx);
3859 goto process_bpf_exit;
3865 if (env->log.level > 1 || (env->log.level && do_print_state)) {
3866 if (env->log.level > 1)
3867 verbose(env, "%d:", insn_idx);
3869 verbose(env, "\nfrom %d to %d:",
3870 prev_insn_idx, insn_idx);
3871 print_verifier_state(env, state);
3872 do_print_state = false;
3875 if (env->log.level) {
3876 verbose(env, "%d: ", insn_idx);
3877 print_bpf_insn(verbose, env, insn,
3878 env->allow_ptr_leaks);
3881 err = ext_analyzer_insn_hook(env, insn_idx, prev_insn_idx);
3885 regs = cur_regs(env);
3886 env->insn_aux_data[insn_idx].seen = true;
3887 if (class == BPF_ALU || class == BPF_ALU64) {
3888 err = check_alu_op(env, insn);
3892 } else if (class == BPF_LDX) {
3893 enum bpf_reg_type *prev_src_type, src_reg_type;
3895 /* check for reserved fields is already done */
3897 /* check src operand */
3898 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3902 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3906 src_reg_type = regs[insn->src_reg].type;
3908 /* check that memory (src_reg + off) is readable,
3909 * the state of dst_reg will be updated by this func
3911 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
3912 BPF_SIZE(insn->code), BPF_READ,
3917 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
3919 if (*prev_src_type == NOT_INIT) {
3921 * dst_reg = *(u32 *)(src_reg + off)
3922 * save type to validate intersecting paths
3924 *prev_src_type = src_reg_type;
3926 } else if (src_reg_type != *prev_src_type &&
3927 (src_reg_type == PTR_TO_CTX ||
3928 *prev_src_type == PTR_TO_CTX)) {
3929 /* ABuser program is trying to use the same insn
3930 * dst_reg = *(u32*) (src_reg + off)
3931 * with different pointer types:
3932 * src_reg == ctx in one branch and
3933 * src_reg == stack|map in some other branch.
3936 verbose(env, "same insn cannot be used with different pointers\n");
3940 } else if (class == BPF_STX) {
3941 enum bpf_reg_type *prev_dst_type, dst_reg_type;
3943 if (BPF_MODE(insn->code) == BPF_XADD) {
3944 err = check_xadd(env, insn_idx, insn);
3951 /* check src1 operand */
3952 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3955 /* check src2 operand */
3956 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3960 dst_reg_type = regs[insn->dst_reg].type;
3962 /* check that memory (dst_reg + off) is writeable */
3963 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3964 BPF_SIZE(insn->code), BPF_WRITE,
3969 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
3971 if (*prev_dst_type == NOT_INIT) {
3972 *prev_dst_type = dst_reg_type;
3973 } else if (dst_reg_type != *prev_dst_type &&
3974 (dst_reg_type == PTR_TO_CTX ||
3975 *prev_dst_type == PTR_TO_CTX)) {
3976 verbose(env, "same insn cannot be used with different pointers\n");
3980 } else if (class == BPF_ST) {
3981 if (BPF_MODE(insn->code) != BPF_MEM ||
3982 insn->src_reg != BPF_REG_0) {
3983 verbose(env, "BPF_ST uses reserved fields\n");
3986 /* check src operand */
3987 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3991 /* check that memory (dst_reg + off) is writeable */
3992 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3993 BPF_SIZE(insn->code), BPF_WRITE,
3998 } else if (class == BPF_JMP) {
3999 u8 opcode = BPF_OP(insn->code);
4001 if (opcode == BPF_CALL) {
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_CALL uses reserved fields\n");
4010 err = check_call(env, insn->imm, insn_idx);
4014 } else if (opcode == BPF_JA) {
4015 if (BPF_SRC(insn->code) != BPF_K ||
4017 insn->src_reg != BPF_REG_0 ||
4018 insn->dst_reg != BPF_REG_0) {
4019 verbose(env, "BPF_JA uses reserved fields\n");
4023 insn_idx += insn->off + 1;
4026 } else if (opcode == BPF_EXIT) {
4027 if (BPF_SRC(insn->code) != BPF_K ||
4029 insn->src_reg != BPF_REG_0 ||
4030 insn->dst_reg != BPF_REG_0) {
4031 verbose(env, "BPF_EXIT uses reserved fields\n");
4035 /* eBPF calling convetion is such that R0 is used
4036 * to return the value from eBPF program.
4037 * Make sure that it's readable at this time
4038 * of bpf_exit, which means that program wrote
4039 * something into it earlier
4041 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4045 if (is_pointer_value(env, BPF_REG_0)) {
4046 verbose(env, "R0 leaks addr as return value\n");
4050 err = check_return_code(env);
4054 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4060 do_print_state = true;
4064 err = check_cond_jmp_op(env, insn, &insn_idx);
4068 } else if (class == BPF_LD) {
4069 u8 mode = BPF_MODE(insn->code);
4071 if (mode == BPF_ABS || mode == BPF_IND) {
4072 err = check_ld_abs(env, insn);
4076 } else if (mode == BPF_IMM) {
4077 err = check_ld_imm(env, insn);
4082 env->insn_aux_data[insn_idx].seen = true;
4084 verbose(env, "invalid BPF_LD mode\n");
4088 verbose(env, "unknown insn class %d\n", class);
4095 verbose(env, "processed %d insns, stack depth %d\n", insn_processed,
4096 env->prog->aux->stack_depth);
4100 static int check_map_prealloc(struct bpf_map *map)
4102 return (map->map_type != BPF_MAP_TYPE_HASH &&
4103 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
4104 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
4105 !(map->map_flags & BPF_F_NO_PREALLOC);
4108 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
4109 struct bpf_map *map,
4110 struct bpf_prog *prog)
4113 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4114 * preallocated hash maps, since doing memory allocation
4115 * in overflow_handler can crash depending on where nmi got
4118 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4119 if (!check_map_prealloc(map)) {
4120 verbose(env, "perf_event programs can only use preallocated hash map\n");
4123 if (map->inner_map_meta &&
4124 !check_map_prealloc(map->inner_map_meta)) {
4125 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
4132 /* look for pseudo eBPF instructions that access map FDs and
4133 * replace them with actual map pointers
4135 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4137 struct bpf_insn *insn = env->prog->insnsi;
4138 int insn_cnt = env->prog->len;
4141 err = bpf_prog_calc_tag(env->prog);
4145 for (i = 0; i < insn_cnt; i++, insn++) {
4146 if (BPF_CLASS(insn->code) == BPF_LDX &&
4147 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4148 verbose(env, "BPF_LDX uses reserved fields\n");
4152 if (BPF_CLASS(insn->code) == BPF_STX &&
4153 ((BPF_MODE(insn->code) != BPF_MEM &&
4154 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4155 verbose(env, "BPF_STX uses reserved fields\n");
4159 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4160 struct bpf_map *map;
4163 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4164 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4166 verbose(env, "invalid bpf_ld_imm64 insn\n");
4170 if (insn->src_reg == 0)
4171 /* valid generic load 64-bit imm */
4174 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4176 "unrecognized bpf_ld_imm64 insn\n");
4180 f = fdget(insn->imm);
4181 map = __bpf_map_get(f);
4183 verbose(env, "fd %d is not pointing to valid bpf_map\n",
4185 return PTR_ERR(map);
4188 err = check_map_prog_compatibility(env, map, env->prog);
4194 /* store map pointer inside BPF_LD_IMM64 instruction */
4195 insn[0].imm = (u32) (unsigned long) map;
4196 insn[1].imm = ((u64) (unsigned long) map) >> 32;
4198 /* check whether we recorded this map already */
4199 for (j = 0; j < env->used_map_cnt; j++)
4200 if (env->used_maps[j] == map) {
4205 if (env->used_map_cnt >= MAX_USED_MAPS) {
4210 /* hold the map. If the program is rejected by verifier,
4211 * the map will be released by release_maps() or it
4212 * will be used by the valid program until it's unloaded
4213 * and all maps are released in free_bpf_prog_info()
4215 map = bpf_map_inc(map, false);
4218 return PTR_ERR(map);
4220 env->used_maps[env->used_map_cnt++] = map;
4229 /* now all pseudo BPF_LD_IMM64 instructions load valid
4230 * 'struct bpf_map *' into a register instead of user map_fd.
4231 * These pointers will be used later by verifier to validate map access.
4236 /* drop refcnt of maps used by the rejected program */
4237 static void release_maps(struct bpf_verifier_env *env)
4241 for (i = 0; i < env->used_map_cnt; i++)
4242 bpf_map_put(env->used_maps[i]);
4245 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
4246 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
4248 struct bpf_insn *insn = env->prog->insnsi;
4249 int insn_cnt = env->prog->len;
4252 for (i = 0; i < insn_cnt; i++, insn++)
4253 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
4257 /* single env->prog->insni[off] instruction was replaced with the range
4258 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
4259 * [0, off) and [off, end) to new locations, so the patched range stays zero
4261 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
4264 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
4269 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
4272 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
4273 memcpy(new_data + off + cnt - 1, old_data + off,
4274 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
4275 for (i = off; i < off + cnt - 1; i++)
4276 new_data[i].seen = true;
4277 env->insn_aux_data = new_data;
4282 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
4283 const struct bpf_insn *patch, u32 len)
4285 struct bpf_prog *new_prog;
4287 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
4290 if (adjust_insn_aux_data(env, new_prog->len, off, len))
4295 /* The verifier does more data flow analysis than llvm and will not explore
4296 * branches that are dead at run time. Malicious programs can have dead code
4297 * too. Therefore replace all dead at-run-time code with nops.
4299 static void sanitize_dead_code(struct bpf_verifier_env *env)
4301 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
4302 struct bpf_insn nop = BPF_MOV64_REG(BPF_REG_0, BPF_REG_0);
4303 struct bpf_insn *insn = env->prog->insnsi;
4304 const int insn_cnt = env->prog->len;
4307 for (i = 0; i < insn_cnt; i++) {
4308 if (aux_data[i].seen)
4310 memcpy(insn + i, &nop, sizeof(nop));
4314 /* convert load instructions that access fields of 'struct __sk_buff'
4315 * into sequence of instructions that access fields of 'struct sk_buff'
4317 static int convert_ctx_accesses(struct bpf_verifier_env *env)
4319 const struct bpf_verifier_ops *ops = env->ops;
4320 int i, cnt, size, ctx_field_size, delta = 0;
4321 const int insn_cnt = env->prog->len;
4322 struct bpf_insn insn_buf[16], *insn;
4323 struct bpf_prog *new_prog;
4324 enum bpf_access_type type;
4325 bool is_narrower_load;
4328 if (ops->gen_prologue) {
4329 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
4331 if (cnt >= ARRAY_SIZE(insn_buf)) {
4332 verbose(env, "bpf verifier is misconfigured\n");
4335 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
4339 env->prog = new_prog;
4344 if (!ops->convert_ctx_access)
4347 insn = env->prog->insnsi + delta;
4349 for (i = 0; i < insn_cnt; i++, insn++) {
4350 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
4351 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
4352 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
4353 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
4355 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
4356 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
4357 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
4358 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
4363 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
4366 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
4367 size = BPF_LDST_BYTES(insn);
4369 /* If the read access is a narrower load of the field,
4370 * convert to a 4/8-byte load, to minimum program type specific
4371 * convert_ctx_access changes. If conversion is successful,
4372 * we will apply proper mask to the result.
4374 is_narrower_load = size < ctx_field_size;
4375 if (is_narrower_load) {
4376 u32 off = insn->off;
4379 if (type == BPF_WRITE) {
4380 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
4385 if (ctx_field_size == 4)
4387 else if (ctx_field_size == 8)
4390 insn->off = off & ~(ctx_field_size - 1);
4391 insn->code = BPF_LDX | BPF_MEM | size_code;
4395 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
4397 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
4398 (ctx_field_size && !target_size)) {
4399 verbose(env, "bpf verifier is misconfigured\n");
4403 if (is_narrower_load && size < target_size) {
4404 if (ctx_field_size <= 4)
4405 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
4406 (1 << size * 8) - 1);
4408 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
4409 (1 << size * 8) - 1);
4412 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4418 /* keep walking new program and skip insns we just inserted */
4419 env->prog = new_prog;
4420 insn = new_prog->insnsi + i + delta;
4426 /* fixup insn->imm field of bpf_call instructions
4427 * and inline eligible helpers as explicit sequence of BPF instructions
4429 * this function is called after eBPF program passed verification
4431 static int fixup_bpf_calls(struct bpf_verifier_env *env)
4433 struct bpf_prog *prog = env->prog;
4434 struct bpf_insn *insn = prog->insnsi;
4435 const struct bpf_func_proto *fn;
4436 const int insn_cnt = prog->len;
4437 struct bpf_insn insn_buf[16];
4438 struct bpf_prog *new_prog;
4439 struct bpf_map *map_ptr;
4440 int i, cnt, delta = 0;
4442 for (i = 0; i < insn_cnt; i++, insn++) {
4443 if (insn->code != (BPF_JMP | BPF_CALL))
4446 if (insn->imm == BPF_FUNC_get_route_realm)
4447 prog->dst_needed = 1;
4448 if (insn->imm == BPF_FUNC_get_prandom_u32)
4449 bpf_user_rnd_init_once();
4450 if (insn->imm == BPF_FUNC_tail_call) {
4451 /* If we tail call into other programs, we
4452 * cannot make any assumptions since they can
4453 * be replaced dynamically during runtime in
4454 * the program array.
4456 prog->cb_access = 1;
4457 env->prog->aux->stack_depth = MAX_BPF_STACK;
4459 /* mark bpf_tail_call as different opcode to avoid
4460 * conditional branch in the interpeter for every normal
4461 * call and to prevent accidental JITing by JIT compiler
4462 * that doesn't support bpf_tail_call yet
4465 insn->code = BPF_JMP | BPF_TAIL_CALL;
4467 /* instead of changing every JIT dealing with tail_call
4468 * emit two extra insns:
4469 * if (index >= max_entries) goto out;
4470 * index &= array->index_mask;
4471 * to avoid out-of-bounds cpu speculation
4473 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4474 if (map_ptr == BPF_MAP_PTR_POISON) {
4475 verbose(env, "tail_call obusing map_ptr\n");
4478 if (!map_ptr->unpriv_array)
4480 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
4481 map_ptr->max_entries, 2);
4482 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
4483 container_of(map_ptr,
4486 insn_buf[2] = *insn;
4488 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
4493 env->prog = prog = new_prog;
4494 insn = new_prog->insnsi + i + delta;
4498 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
4499 * handlers are currently limited to 64 bit only.
4501 if (ebpf_jit_enabled() && BITS_PER_LONG == 64 &&
4502 insn->imm == BPF_FUNC_map_lookup_elem) {
4503 map_ptr = env->insn_aux_data[i + delta].map_ptr;
4504 if (map_ptr == BPF_MAP_PTR_POISON ||
4505 !map_ptr->ops->map_gen_lookup)
4506 goto patch_call_imm;
4508 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
4509 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
4510 verbose(env, "bpf verifier is misconfigured\n");
4514 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
4521 /* keep walking new program and skip insns we just inserted */
4522 env->prog = prog = new_prog;
4523 insn = new_prog->insnsi + i + delta;
4527 if (insn->imm == BPF_FUNC_redirect_map) {
4528 /* Note, we cannot use prog directly as imm as subsequent
4529 * rewrites would still change the prog pointer. The only
4530 * stable address we can use is aux, which also works with
4531 * prog clones during blinding.
4533 u64 addr = (unsigned long)prog->aux;
4534 struct bpf_insn r4_ld[] = {
4535 BPF_LD_IMM64(BPF_REG_4, addr),
4538 cnt = ARRAY_SIZE(r4_ld);
4540 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
4545 env->prog = prog = new_prog;
4546 insn = new_prog->insnsi + i + delta;
4549 fn = env->ops->get_func_proto(insn->imm);
4550 /* all functions that have prototype and verifier allowed
4551 * programs to call them, must be real in-kernel functions
4555 "kernel subsystem misconfigured func %s#%d\n",
4556 func_id_name(insn->imm), insn->imm);
4559 insn->imm = fn->func - __bpf_call_base;
4565 static void free_states(struct bpf_verifier_env *env)
4567 struct bpf_verifier_state_list *sl, *sln;
4570 if (!env->explored_states)
4573 for (i = 0; i < env->prog->len; i++) {
4574 sl = env->explored_states[i];
4577 while (sl != STATE_LIST_MARK) {
4579 free_verifier_state(&sl->state, false);
4585 kfree(env->explored_states);
4588 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
4590 struct bpf_verifier_env *env;
4591 struct bpf_verifer_log *log;
4594 /* no program is valid */
4595 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
4598 /* 'struct bpf_verifier_env' can be global, but since it's not small,
4599 * allocate/free it every time bpf_check() is called
4601 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
4606 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
4609 if (!env->insn_aux_data)
4612 env->ops = bpf_verifier_ops[env->prog->type];
4614 /* grab the mutex to protect few globals used by verifier */
4615 mutex_lock(&bpf_verifier_lock);
4617 if (attr->log_level || attr->log_buf || attr->log_size) {
4618 /* user requested verbose verifier output
4619 * and supplied buffer to store the verification trace
4621 log->level = attr->log_level;
4622 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
4623 log->len_total = attr->log_size;
4626 /* log attributes have to be sane */
4627 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
4628 !log->level || !log->ubuf)
4632 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
4633 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
4634 env->strict_alignment = true;
4636 if (env->prog->aux->offload) {
4637 ret = bpf_prog_offload_verifier_prep(env);
4642 ret = replace_map_fd_with_map_ptr(env);
4644 goto skip_full_check;
4646 env->explored_states = kcalloc(env->prog->len,
4647 sizeof(struct bpf_verifier_state_list *),
4650 if (!env->explored_states)
4651 goto skip_full_check;
4653 ret = check_cfg(env);
4655 goto skip_full_check;
4657 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
4659 ret = do_check(env);
4660 if (env->cur_state) {
4661 free_verifier_state(env->cur_state, true);
4662 env->cur_state = NULL;
4666 while (!pop_stack(env, NULL, NULL));
4670 sanitize_dead_code(env);
4673 /* program is valid, convert *(u32*)(ctx + off) accesses */
4674 ret = convert_ctx_accesses(env);
4677 ret = fixup_bpf_calls(env);
4679 if (log->level && bpf_verifier_log_full(log))
4681 if (log->level && !log->ubuf) {
4683 goto err_release_maps;
4686 if (ret == 0 && env->used_map_cnt) {
4687 /* if program passed verifier, update used_maps in bpf_prog_info */
4688 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
4689 sizeof(env->used_maps[0]),
4692 if (!env->prog->aux->used_maps) {
4694 goto err_release_maps;
4697 memcpy(env->prog->aux->used_maps, env->used_maps,
4698 sizeof(env->used_maps[0]) * env->used_map_cnt);
4699 env->prog->aux->used_map_cnt = env->used_map_cnt;
4701 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
4702 * bpf_ld_imm64 instructions
4704 convert_pseudo_ld_imm64(env);
4708 if (!env->prog->aux->used_maps)
4709 /* if we didn't copy map pointers into bpf_prog_info, release
4710 * them now. Otherwise free_bpf_prog_info() will release them.
4715 mutex_unlock(&bpf_verifier_lock);
4716 vfree(env->insn_aux_data);