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>
23 #include <linux/bsearch.h>
24 #include <linux/sort.h>
25 #include <linux/perf_event.h>
29 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
30 #define BPF_PROG_TYPE(_id, _name) \
31 [_id] = & _name ## _verifier_ops,
32 #define BPF_MAP_TYPE(_id, _ops)
33 #include <linux/bpf_types.h>
38 /* bpf_check() is a static code analyzer that walks eBPF program
39 * instruction by instruction and updates register/stack state.
40 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
42 * The first pass is depth-first-search to check that the program is a DAG.
43 * It rejects the following programs:
44 * - larger than BPF_MAXINSNS insns
45 * - if loop is present (detected via back-edge)
46 * - unreachable insns exist (shouldn't be a forest. program = one function)
47 * - out of bounds or malformed jumps
48 * The second pass is all possible path descent from the 1st insn.
49 * Since it's analyzing all pathes through the program, the length of the
50 * analysis is limited to 64k insn, which may be hit even if total number of
51 * insn is less then 4K, but there are too many branches that change stack/regs.
52 * Number of 'branches to be analyzed' is limited to 1k
54 * On entry to each instruction, each register has a type, and the instruction
55 * changes the types of the registers depending on instruction semantics.
56 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
59 * All registers are 64-bit.
60 * R0 - return register
61 * R1-R5 argument passing registers
62 * R6-R9 callee saved registers
63 * R10 - frame pointer read-only
65 * At the start of BPF program the register R1 contains a pointer to bpf_context
66 * and has type PTR_TO_CTX.
68 * Verifier tracks arithmetic operations on pointers in case:
69 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
70 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
71 * 1st insn copies R10 (which has FRAME_PTR) type into R1
72 * and 2nd arithmetic instruction is pattern matched to recognize
73 * that it wants to construct a pointer to some element within stack.
74 * So after 2nd insn, the register R1 has type PTR_TO_STACK
75 * (and -20 constant is saved for further stack bounds checking).
76 * Meaning that this reg is a pointer to stack plus known immediate constant.
78 * Most of the time the registers have SCALAR_VALUE type, which
79 * means the register has some value, but it's not a valid pointer.
80 * (like pointer plus pointer becomes SCALAR_VALUE type)
82 * When verifier sees load or store instructions the type of base register
83 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
84 * types recognized by check_mem_access() function.
86 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
87 * and the range of [ptr, ptr + map's value_size) is accessible.
89 * registers used to pass values to function calls are checked against
90 * function argument constraints.
92 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
93 * It means that the register type passed to this function must be
94 * PTR_TO_STACK and it will be used inside the function as
95 * 'pointer to map element key'
97 * For example the argument constraints for bpf_map_lookup_elem():
98 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
99 * .arg1_type = ARG_CONST_MAP_PTR,
100 * .arg2_type = ARG_PTR_TO_MAP_KEY,
102 * ret_type says that this function returns 'pointer to map elem value or null'
103 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
104 * 2nd argument should be a pointer to stack, which will be used inside
105 * the helper function as a pointer to map element key.
107 * On the kernel side the helper function looks like:
108 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
110 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
111 * void *key = (void *) (unsigned long) r2;
114 * here kernel can access 'key' and 'map' pointers safely, knowing that
115 * [key, key + map->key_size) bytes are valid and were initialized on
116 * the stack of eBPF program.
119 * Corresponding eBPF program may look like:
120 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
121 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
122 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
123 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
124 * here verifier looks at prototype of map_lookup_elem() and sees:
125 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
126 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
128 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
129 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
130 * and were initialized prior to this call.
131 * If it's ok, then verifier allows this BPF_CALL insn and looks at
132 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
133 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
134 * returns ether pointer to map value or NULL.
136 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
137 * insn, the register holding that pointer in the true branch changes state to
138 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
139 * branch. See check_cond_jmp_op().
141 * After the call R0 is set to return type of the function and registers R1-R5
142 * are set to NOT_INIT to indicate that they are no longer readable.
145 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
146 struct bpf_verifier_stack_elem {
147 /* verifer state is 'st'
148 * before processing instruction 'insn_idx'
149 * and after processing instruction 'prev_insn_idx'
151 struct bpf_verifier_state st;
154 struct bpf_verifier_stack_elem *next;
157 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
158 #define BPF_COMPLEXITY_LIMIT_STACK 1024
160 #define BPF_MAP_PTR_UNPRIV 1UL
161 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
162 POISON_POINTER_DELTA))
163 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
165 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
167 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
170 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
172 return aux->map_state & BPF_MAP_PTR_UNPRIV;
175 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
176 const struct bpf_map *map, bool unpriv)
178 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
179 unpriv |= bpf_map_ptr_unpriv(aux);
180 aux->map_state = (unsigned long)map |
181 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
184 struct bpf_call_arg_meta {
185 struct bpf_map *map_ptr;
190 s64 msize_smax_value;
191 u64 msize_umax_value;
194 static DEFINE_MUTEX(bpf_verifier_lock);
196 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
201 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
203 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
204 "verifier log line truncated - local buffer too short\n");
206 n = min(log->len_total - log->len_used - 1, n);
209 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
215 /* log_level controls verbosity level of eBPF verifier.
216 * bpf_verifier_log_write() is used to dump the verification trace to the log,
217 * so the user can figure out what's wrong with the program
219 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
220 const char *fmt, ...)
224 if (!bpf_verifier_log_needed(&env->log))
228 bpf_verifier_vlog(&env->log, fmt, args);
231 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
233 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
235 struct bpf_verifier_env *env = private_data;
238 if (!bpf_verifier_log_needed(&env->log))
242 bpf_verifier_vlog(&env->log, fmt, args);
246 static bool type_is_pkt_pointer(enum bpf_reg_type type)
248 return type == PTR_TO_PACKET ||
249 type == PTR_TO_PACKET_META;
252 /* string representation of 'enum bpf_reg_type' */
253 static const char * const reg_type_str[] = {
255 [SCALAR_VALUE] = "inv",
256 [PTR_TO_CTX] = "ctx",
257 [CONST_PTR_TO_MAP] = "map_ptr",
258 [PTR_TO_MAP_VALUE] = "map_value",
259 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
260 [PTR_TO_STACK] = "fp",
261 [PTR_TO_PACKET] = "pkt",
262 [PTR_TO_PACKET_META] = "pkt_meta",
263 [PTR_TO_PACKET_END] = "pkt_end",
266 static void print_liveness(struct bpf_verifier_env *env,
267 enum bpf_reg_liveness live)
269 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
271 if (live & REG_LIVE_READ)
273 if (live & REG_LIVE_WRITTEN)
277 static struct bpf_func_state *func(struct bpf_verifier_env *env,
278 const struct bpf_reg_state *reg)
280 struct bpf_verifier_state *cur = env->cur_state;
282 return cur->frame[reg->frameno];
285 static void print_verifier_state(struct bpf_verifier_env *env,
286 const struct bpf_func_state *state)
288 const struct bpf_reg_state *reg;
293 verbose(env, " frame%d:", state->frameno);
294 for (i = 0; i < MAX_BPF_REG; i++) {
295 reg = &state->regs[i];
299 verbose(env, " R%d", i);
300 print_liveness(env, reg->live);
301 verbose(env, "=%s", reg_type_str[t]);
302 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
303 tnum_is_const(reg->var_off)) {
304 /* reg->off should be 0 for SCALAR_VALUE */
305 verbose(env, "%lld", reg->var_off.value + reg->off);
306 if (t == PTR_TO_STACK)
307 verbose(env, ",call_%d", func(env, reg)->callsite);
309 verbose(env, "(id=%d", reg->id);
310 if (t != SCALAR_VALUE)
311 verbose(env, ",off=%d", reg->off);
312 if (type_is_pkt_pointer(t))
313 verbose(env, ",r=%d", reg->range);
314 else if (t == CONST_PTR_TO_MAP ||
315 t == PTR_TO_MAP_VALUE ||
316 t == PTR_TO_MAP_VALUE_OR_NULL)
317 verbose(env, ",ks=%d,vs=%d",
318 reg->map_ptr->key_size,
319 reg->map_ptr->value_size);
320 if (tnum_is_const(reg->var_off)) {
321 /* Typically an immediate SCALAR_VALUE, but
322 * could be a pointer whose offset is too big
325 verbose(env, ",imm=%llx", reg->var_off.value);
327 if (reg->smin_value != reg->umin_value &&
328 reg->smin_value != S64_MIN)
329 verbose(env, ",smin_value=%lld",
330 (long long)reg->smin_value);
331 if (reg->smax_value != reg->umax_value &&
332 reg->smax_value != S64_MAX)
333 verbose(env, ",smax_value=%lld",
334 (long long)reg->smax_value);
335 if (reg->umin_value != 0)
336 verbose(env, ",umin_value=%llu",
337 (unsigned long long)reg->umin_value);
338 if (reg->umax_value != U64_MAX)
339 verbose(env, ",umax_value=%llu",
340 (unsigned long long)reg->umax_value);
341 if (!tnum_is_unknown(reg->var_off)) {
344 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
345 verbose(env, ",var_off=%s", tn_buf);
351 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
352 if (state->stack[i].slot_type[0] == STACK_SPILL) {
353 verbose(env, " fp%d",
354 (-i - 1) * BPF_REG_SIZE);
355 print_liveness(env, state->stack[i].spilled_ptr.live);
357 reg_type_str[state->stack[i].spilled_ptr.type]);
359 if (state->stack[i].slot_type[0] == STACK_ZERO)
360 verbose(env, " fp%d=0", (-i - 1) * BPF_REG_SIZE);
365 static int copy_stack_state(struct bpf_func_state *dst,
366 const struct bpf_func_state *src)
370 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
371 /* internal bug, make state invalid to reject the program */
372 memset(dst, 0, sizeof(*dst));
375 memcpy(dst->stack, src->stack,
376 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
380 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
381 * make it consume minimal amount of memory. check_stack_write() access from
382 * the program calls into realloc_func_state() to grow the stack size.
383 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
384 * which this function copies over. It points to previous bpf_verifier_state
385 * which is never reallocated
387 static int realloc_func_state(struct bpf_func_state *state, int size,
390 u32 old_size = state->allocated_stack;
391 struct bpf_stack_state *new_stack;
392 int slot = size / BPF_REG_SIZE;
394 if (size <= old_size || !size) {
397 state->allocated_stack = slot * BPF_REG_SIZE;
398 if (!size && old_size) {
404 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
410 memcpy(new_stack, state->stack,
411 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
412 memset(new_stack + old_size / BPF_REG_SIZE, 0,
413 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
415 state->allocated_stack = slot * BPF_REG_SIZE;
417 state->stack = new_stack;
421 static void free_func_state(struct bpf_func_state *state)
429 static void free_verifier_state(struct bpf_verifier_state *state,
434 for (i = 0; i <= state->curframe; i++) {
435 free_func_state(state->frame[i]);
436 state->frame[i] = NULL;
442 /* copy verifier state from src to dst growing dst stack space
443 * when necessary to accommodate larger src stack
445 static int copy_func_state(struct bpf_func_state *dst,
446 const struct bpf_func_state *src)
450 err = realloc_func_state(dst, src->allocated_stack, false);
453 memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack));
454 return copy_stack_state(dst, src);
457 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
458 const struct bpf_verifier_state *src)
460 struct bpf_func_state *dst;
463 /* if dst has more stack frames then src frame, free them */
464 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
465 free_func_state(dst_state->frame[i]);
466 dst_state->frame[i] = NULL;
468 dst_state->curframe = src->curframe;
469 dst_state->parent = src->parent;
470 for (i = 0; i <= src->curframe; i++) {
471 dst = dst_state->frame[i];
473 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
476 dst_state->frame[i] = dst;
478 err = copy_func_state(dst, src->frame[i]);
485 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
488 struct bpf_verifier_state *cur = env->cur_state;
489 struct bpf_verifier_stack_elem *elem, *head = env->head;
492 if (env->head == NULL)
496 err = copy_verifier_state(cur, &head->st);
501 *insn_idx = head->insn_idx;
503 *prev_insn_idx = head->prev_insn_idx;
505 free_verifier_state(&head->st, false);
512 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
513 int insn_idx, int prev_insn_idx)
515 struct bpf_verifier_state *cur = env->cur_state;
516 struct bpf_verifier_stack_elem *elem;
519 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
523 elem->insn_idx = insn_idx;
524 elem->prev_insn_idx = prev_insn_idx;
525 elem->next = env->head;
528 err = copy_verifier_state(&elem->st, cur);
531 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
532 verbose(env, "BPF program is too complex\n");
537 free_verifier_state(env->cur_state, true);
538 env->cur_state = NULL;
539 /* pop all elements and return */
540 while (!pop_stack(env, NULL, NULL));
544 #define CALLER_SAVED_REGS 6
545 static const int caller_saved[CALLER_SAVED_REGS] = {
546 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
549 static void __mark_reg_not_init(struct bpf_reg_state *reg);
551 /* Mark the unknown part of a register (variable offset or scalar value) as
552 * known to have the value @imm.
554 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
557 reg->var_off = tnum_const(imm);
558 reg->smin_value = (s64)imm;
559 reg->smax_value = (s64)imm;
560 reg->umin_value = imm;
561 reg->umax_value = imm;
564 /* Mark the 'variable offset' part of a register as zero. This should be
565 * used only on registers holding a pointer type.
567 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
569 __mark_reg_known(reg, 0);
572 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
574 __mark_reg_known(reg, 0);
576 reg->type = SCALAR_VALUE;
579 static void mark_reg_known_zero(struct bpf_verifier_env *env,
580 struct bpf_reg_state *regs, u32 regno)
582 if (WARN_ON(regno >= MAX_BPF_REG)) {
583 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
584 /* Something bad happened, let's kill all regs */
585 for (regno = 0; regno < MAX_BPF_REG; regno++)
586 __mark_reg_not_init(regs + regno);
589 __mark_reg_known_zero(regs + regno);
592 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
594 return type_is_pkt_pointer(reg->type);
597 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
599 return reg_is_pkt_pointer(reg) ||
600 reg->type == PTR_TO_PACKET_END;
603 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
604 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
605 enum bpf_reg_type which)
607 /* The register can already have a range from prior markings.
608 * This is fine as long as it hasn't been advanced from its
611 return reg->type == which &&
614 tnum_equals_const(reg->var_off, 0);
617 /* Attempts to improve min/max values based on var_off information */
618 static void __update_reg_bounds(struct bpf_reg_state *reg)
620 /* min signed is max(sign bit) | min(other bits) */
621 reg->smin_value = max_t(s64, reg->smin_value,
622 reg->var_off.value | (reg->var_off.mask & S64_MIN));
623 /* max signed is min(sign bit) | max(other bits) */
624 reg->smax_value = min_t(s64, reg->smax_value,
625 reg->var_off.value | (reg->var_off.mask & S64_MAX));
626 reg->umin_value = max(reg->umin_value, reg->var_off.value);
627 reg->umax_value = min(reg->umax_value,
628 reg->var_off.value | reg->var_off.mask);
631 /* Uses signed min/max values to inform unsigned, and vice-versa */
632 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
634 /* Learn sign from signed bounds.
635 * If we cannot cross the sign boundary, then signed and unsigned bounds
636 * are the same, so combine. This works even in the negative case, e.g.
637 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
639 if (reg->smin_value >= 0 || reg->smax_value < 0) {
640 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
642 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
646 /* Learn sign from unsigned bounds. Signed bounds cross the sign
647 * boundary, so we must be careful.
649 if ((s64)reg->umax_value >= 0) {
650 /* Positive. We can't learn anything from the smin, but smax
651 * is positive, hence safe.
653 reg->smin_value = reg->umin_value;
654 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
656 } else if ((s64)reg->umin_value < 0) {
657 /* Negative. We can't learn anything from the smax, but smin
658 * is negative, hence safe.
660 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
662 reg->smax_value = reg->umax_value;
666 /* Attempts to improve var_off based on unsigned min/max information */
667 static void __reg_bound_offset(struct bpf_reg_state *reg)
669 reg->var_off = tnum_intersect(reg->var_off,
670 tnum_range(reg->umin_value,
674 /* Reset the min/max bounds of a register */
675 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
677 reg->smin_value = S64_MIN;
678 reg->smax_value = S64_MAX;
680 reg->umax_value = U64_MAX;
683 /* Mark a register as having a completely unknown (scalar) value. */
684 static void __mark_reg_unknown(struct bpf_reg_state *reg)
686 reg->type = SCALAR_VALUE;
689 reg->var_off = tnum_unknown;
691 __mark_reg_unbounded(reg);
694 static void mark_reg_unknown(struct bpf_verifier_env *env,
695 struct bpf_reg_state *regs, u32 regno)
697 if (WARN_ON(regno >= MAX_BPF_REG)) {
698 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
699 /* Something bad happened, let's kill all regs except FP */
700 for (regno = 0; regno < BPF_REG_FP; regno++)
701 __mark_reg_not_init(regs + regno);
704 __mark_reg_unknown(regs + regno);
707 static void __mark_reg_not_init(struct bpf_reg_state *reg)
709 __mark_reg_unknown(reg);
710 reg->type = NOT_INIT;
713 static void mark_reg_not_init(struct bpf_verifier_env *env,
714 struct bpf_reg_state *regs, u32 regno)
716 if (WARN_ON(regno >= MAX_BPF_REG)) {
717 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
718 /* Something bad happened, let's kill all regs except FP */
719 for (regno = 0; regno < BPF_REG_FP; regno++)
720 __mark_reg_not_init(regs + regno);
723 __mark_reg_not_init(regs + regno);
726 static void init_reg_state(struct bpf_verifier_env *env,
727 struct bpf_func_state *state)
729 struct bpf_reg_state *regs = state->regs;
732 for (i = 0; i < MAX_BPF_REG; i++) {
733 mark_reg_not_init(env, regs, i);
734 regs[i].live = REG_LIVE_NONE;
738 regs[BPF_REG_FP].type = PTR_TO_STACK;
739 mark_reg_known_zero(env, regs, BPF_REG_FP);
740 regs[BPF_REG_FP].frameno = state->frameno;
742 /* 1st arg to a function */
743 regs[BPF_REG_1].type = PTR_TO_CTX;
744 mark_reg_known_zero(env, regs, BPF_REG_1);
747 #define BPF_MAIN_FUNC (-1)
748 static void init_func_state(struct bpf_verifier_env *env,
749 struct bpf_func_state *state,
750 int callsite, int frameno, int subprogno)
752 state->callsite = callsite;
753 state->frameno = frameno;
754 state->subprogno = subprogno;
755 init_reg_state(env, state);
759 SRC_OP, /* register is used as source operand */
760 DST_OP, /* register is used as destination operand */
761 DST_OP_NO_MARK /* same as above, check only, don't mark */
764 static int cmp_subprogs(const void *a, const void *b)
766 return ((struct bpf_subprog_info *)a)->start -
767 ((struct bpf_subprog_info *)b)->start;
770 static int find_subprog(struct bpf_verifier_env *env, int off)
772 struct bpf_subprog_info *p;
774 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
775 sizeof(env->subprog_info[0]), cmp_subprogs);
778 return p - env->subprog_info;
782 static int add_subprog(struct bpf_verifier_env *env, int off)
784 int insn_cnt = env->prog->len;
787 if (off >= insn_cnt || off < 0) {
788 verbose(env, "call to invalid destination\n");
791 ret = find_subprog(env, off);
794 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
795 verbose(env, "too many subprograms\n");
798 env->subprog_info[env->subprog_cnt++].start = off;
799 sort(env->subprog_info, env->subprog_cnt,
800 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
804 static int check_subprogs(struct bpf_verifier_env *env)
806 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
807 struct bpf_subprog_info *subprog = env->subprog_info;
808 struct bpf_insn *insn = env->prog->insnsi;
809 int insn_cnt = env->prog->len;
811 /* Add entry function. */
812 ret = add_subprog(env, 0);
816 /* determine subprog starts. The end is one before the next starts */
817 for (i = 0; i < insn_cnt; i++) {
818 if (insn[i].code != (BPF_JMP | BPF_CALL))
820 if (insn[i].src_reg != BPF_PSEUDO_CALL)
822 if (!env->allow_ptr_leaks) {
823 verbose(env, "function calls to other bpf functions are allowed for root only\n");
826 if (bpf_prog_is_dev_bound(env->prog->aux)) {
827 verbose(env, "function calls in offloaded programs are not supported yet\n");
830 ret = add_subprog(env, i + insn[i].imm + 1);
835 /* Add a fake 'exit' subprog which could simplify subprog iteration
836 * logic. 'subprog_cnt' should not be increased.
838 subprog[env->subprog_cnt].start = insn_cnt;
840 if (env->log.level > 1)
841 for (i = 0; i < env->subprog_cnt; i++)
842 verbose(env, "func#%d @%d\n", i, subprog[i].start);
844 /* now check that all jumps are within the same subprog */
845 subprog_start = subprog[cur_subprog].start;
846 subprog_end = subprog[cur_subprog + 1].start;
847 for (i = 0; i < insn_cnt; i++) {
848 u8 code = insn[i].code;
850 if (BPF_CLASS(code) != BPF_JMP)
852 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
854 off = i + insn[i].off + 1;
855 if (off < subprog_start || off >= subprog_end) {
856 verbose(env, "jump out of range from insn %d to %d\n", i, off);
860 if (i == subprog_end - 1) {
861 /* to avoid fall-through from one subprog into another
862 * the last insn of the subprog should be either exit
863 * or unconditional jump back
865 if (code != (BPF_JMP | BPF_EXIT) &&
866 code != (BPF_JMP | BPF_JA)) {
867 verbose(env, "last insn is not an exit or jmp\n");
870 subprog_start = subprog_end;
872 if (cur_subprog < env->subprog_cnt)
873 subprog_end = subprog[cur_subprog + 1].start;
880 struct bpf_verifier_state *skip_callee(struct bpf_verifier_env *env,
881 const struct bpf_verifier_state *state,
882 struct bpf_verifier_state *parent,
885 struct bpf_verifier_state *tmp = NULL;
887 /* 'parent' could be a state of caller and
888 * 'state' could be a state of callee. In such case
889 * parent->curframe < state->curframe
890 * and it's ok for r1 - r5 registers
892 * 'parent' could be a callee's state after it bpf_exit-ed.
893 * In such case parent->curframe > state->curframe
894 * and it's ok for r0 only
896 if (parent->curframe == state->curframe ||
897 (parent->curframe < state->curframe &&
898 regno >= BPF_REG_1 && regno <= BPF_REG_5) ||
899 (parent->curframe > state->curframe &&
903 if (parent->curframe > state->curframe &&
904 regno >= BPF_REG_6) {
905 /* for callee saved regs we have to skip the whole chain
906 * of states that belong to callee and mark as LIVE_READ
907 * the registers before the call
910 while (tmp && tmp->curframe != state->curframe) {
921 verbose(env, "verifier bug regno %d tmp %p\n", regno, tmp);
922 verbose(env, "regno %d parent frame %d current frame %d\n",
923 regno, parent->curframe, state->curframe);
927 static int mark_reg_read(struct bpf_verifier_env *env,
928 const struct bpf_verifier_state *state,
929 struct bpf_verifier_state *parent,
932 bool writes = parent == state->parent; /* Observe write marks */
934 if (regno == BPF_REG_FP)
935 /* We don't need to worry about FP liveness because it's read-only */
939 /* if read wasn't screened by an earlier write ... */
940 if (writes && state->frame[state->curframe]->regs[regno].live & REG_LIVE_WRITTEN)
942 parent = skip_callee(env, state, parent, regno);
945 /* ... then we depend on parent's value */
946 parent->frame[parent->curframe]->regs[regno].live |= REG_LIVE_READ;
948 parent = state->parent;
954 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
957 struct bpf_verifier_state *vstate = env->cur_state;
958 struct bpf_func_state *state = vstate->frame[vstate->curframe];
959 struct bpf_reg_state *regs = state->regs;
961 if (regno >= MAX_BPF_REG) {
962 verbose(env, "R%d is invalid\n", regno);
967 /* check whether register used as source operand can be read */
968 if (regs[regno].type == NOT_INIT) {
969 verbose(env, "R%d !read_ok\n", regno);
972 return mark_reg_read(env, vstate, vstate->parent, regno);
974 /* check whether register used as dest operand can be written to */
975 if (regno == BPF_REG_FP) {
976 verbose(env, "frame pointer is read only\n");
979 regs[regno].live |= REG_LIVE_WRITTEN;
981 mark_reg_unknown(env, regs, regno);
986 static bool is_spillable_regtype(enum bpf_reg_type type)
989 case PTR_TO_MAP_VALUE:
990 case PTR_TO_MAP_VALUE_OR_NULL:
994 case PTR_TO_PACKET_META:
995 case PTR_TO_PACKET_END:
996 case CONST_PTR_TO_MAP:
1003 /* Does this register contain a constant zero? */
1004 static bool register_is_null(struct bpf_reg_state *reg)
1006 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1009 /* check_stack_read/write functions track spill/fill of registers,
1010 * stack boundary and alignment are checked in check_mem_access()
1012 static int check_stack_write(struct bpf_verifier_env *env,
1013 struct bpf_func_state *state, /* func where register points to */
1014 int off, int size, int value_regno, int insn_idx)
1016 struct bpf_func_state *cur; /* state of the current function */
1017 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1018 enum bpf_reg_type type;
1020 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1024 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1025 * so it's aligned access and [off, off + size) are within stack limits
1027 if (!env->allow_ptr_leaks &&
1028 state->stack[spi].slot_type[0] == STACK_SPILL &&
1029 size != BPF_REG_SIZE) {
1030 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1034 cur = env->cur_state->frame[env->cur_state->curframe];
1035 if (value_regno >= 0 &&
1036 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1038 /* register containing pointer is being spilled into stack */
1039 if (size != BPF_REG_SIZE) {
1040 verbose(env, "invalid size of register spill\n");
1044 if (state != cur && type == PTR_TO_STACK) {
1045 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1049 /* save register state */
1050 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1051 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1053 for (i = 0; i < BPF_REG_SIZE; i++) {
1054 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1055 !env->allow_ptr_leaks) {
1056 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1057 int soff = (-spi - 1) * BPF_REG_SIZE;
1059 /* detected reuse of integer stack slot with a pointer
1060 * which means either llvm is reusing stack slot or
1061 * an attacker is trying to exploit CVE-2018-3639
1062 * (speculative store bypass)
1063 * Have to sanitize that slot with preemptive
1066 if (*poff && *poff != soff) {
1067 /* disallow programs where single insn stores
1068 * into two different stack slots, since verifier
1069 * cannot sanitize them
1072 "insn %d cannot access two stack slots fp%d and fp%d",
1073 insn_idx, *poff, soff);
1078 state->stack[spi].slot_type[i] = STACK_SPILL;
1081 u8 type = STACK_MISC;
1083 /* regular write of data into stack */
1084 state->stack[spi].spilled_ptr = (struct bpf_reg_state) {};
1086 /* only mark the slot as written if all 8 bytes were written
1087 * otherwise read propagation may incorrectly stop too soon
1088 * when stack slots are partially written.
1089 * This heuristic means that read propagation will be
1090 * conservative, since it will add reg_live_read marks
1091 * to stack slots all the way to first state when programs
1092 * writes+reads less than 8 bytes
1094 if (size == BPF_REG_SIZE)
1095 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1097 /* when we zero initialize stack slots mark them as such */
1098 if (value_regno >= 0 &&
1099 register_is_null(&cur->regs[value_regno]))
1102 for (i = 0; i < size; i++)
1103 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1109 /* registers of every function are unique and mark_reg_read() propagates
1110 * the liveness in the following cases:
1111 * - from callee into caller for R1 - R5 that were used as arguments
1112 * - from caller into callee for R0 that used as result of the call
1113 * - from caller to the same caller skipping states of the callee for R6 - R9,
1114 * since R6 - R9 are callee saved by implicit function prologue and
1115 * caller's R6 != callee's R6, so when we propagate liveness up to
1116 * parent states we need to skip callee states for R6 - R9.
1118 * stack slot marking is different, since stacks of caller and callee are
1119 * accessible in both (since caller can pass a pointer to caller's stack to
1120 * callee which can pass it to another function), hence mark_stack_slot_read()
1121 * has to propagate the stack liveness to all parent states at given frame number.
1131 * First *ptr is reading from f1's stack and mark_stack_slot_read() has
1132 * to mark liveness at the f1's frame and not f2's frame.
1133 * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has
1134 * to propagate liveness to f2 states at f1's frame level and further into
1135 * f1 states at f1's frame level until write into that stack slot
1137 static void mark_stack_slot_read(struct bpf_verifier_env *env,
1138 const struct bpf_verifier_state *state,
1139 struct bpf_verifier_state *parent,
1140 int slot, int frameno)
1142 bool writes = parent == state->parent; /* Observe write marks */
1145 if (parent->frame[frameno]->allocated_stack <= slot * BPF_REG_SIZE)
1146 /* since LIVE_WRITTEN mark is only done for full 8-byte
1147 * write the read marks are conservative and parent
1148 * state may not even have the stack allocated. In such case
1149 * end the propagation, since the loop reached beginning
1153 /* if read wasn't screened by an earlier write ... */
1154 if (writes && state->frame[frameno]->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN)
1156 /* ... then we depend on parent's value */
1157 parent->frame[frameno]->stack[slot].spilled_ptr.live |= REG_LIVE_READ;
1159 parent = state->parent;
1164 static int check_stack_read(struct bpf_verifier_env *env,
1165 struct bpf_func_state *reg_state /* func where register points to */,
1166 int off, int size, int value_regno)
1168 struct bpf_verifier_state *vstate = env->cur_state;
1169 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1170 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1173 if (reg_state->allocated_stack <= slot) {
1174 verbose(env, "invalid read from stack off %d+0 size %d\n",
1178 stype = reg_state->stack[spi].slot_type;
1180 if (stype[0] == STACK_SPILL) {
1181 if (size != BPF_REG_SIZE) {
1182 verbose(env, "invalid size of register spill\n");
1185 for (i = 1; i < BPF_REG_SIZE; i++) {
1186 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1187 verbose(env, "corrupted spill memory\n");
1192 if (value_regno >= 0) {
1193 /* restore register state from stack */
1194 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1195 /* mark reg as written since spilled pointer state likely
1196 * has its liveness marks cleared by is_state_visited()
1197 * which resets stack/reg liveness for state transitions
1199 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1201 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1202 reg_state->frameno);
1207 for (i = 0; i < size; i++) {
1208 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1210 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1214 verbose(env, "invalid read from stack off %d+%d size %d\n",
1218 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1219 reg_state->frameno);
1220 if (value_regno >= 0) {
1221 if (zeros == size) {
1222 /* any size read into register is zero extended,
1223 * so the whole register == const_zero
1225 __mark_reg_const_zero(&state->regs[value_regno]);
1227 /* have read misc data from the stack */
1228 mark_reg_unknown(env, state->regs, value_regno);
1230 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1236 /* check read/write into map element returned by bpf_map_lookup_elem() */
1237 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1238 int size, bool zero_size_allowed)
1240 struct bpf_reg_state *regs = cur_regs(env);
1241 struct bpf_map *map = regs[regno].map_ptr;
1243 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1244 off + size > map->value_size) {
1245 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1246 map->value_size, off, size);
1252 /* check read/write into a map element with possible variable offset */
1253 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1254 int off, int size, bool zero_size_allowed)
1256 struct bpf_verifier_state *vstate = env->cur_state;
1257 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1258 struct bpf_reg_state *reg = &state->regs[regno];
1261 /* We may have adjusted the register to this map value, so we
1262 * need to try adding each of min_value and max_value to off
1263 * to make sure our theoretical access will be safe.
1266 print_verifier_state(env, state);
1267 /* The minimum value is only important with signed
1268 * comparisons where we can't assume the floor of a
1269 * value is 0. If we are using signed variables for our
1270 * index'es we need to make sure that whatever we use
1271 * will have a set floor within our range.
1273 if (reg->smin_value < 0) {
1274 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1278 err = __check_map_access(env, regno, reg->smin_value + off, size,
1281 verbose(env, "R%d min value is outside of the array range\n",
1286 /* If we haven't set a max value then we need to bail since we can't be
1287 * sure we won't do bad things.
1288 * If reg->umax_value + off could overflow, treat that as unbounded too.
1290 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1291 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1295 err = __check_map_access(env, regno, reg->umax_value + off, size,
1298 verbose(env, "R%d max value is outside of the array range\n",
1303 #define MAX_PACKET_OFF 0xffff
1305 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1306 const struct bpf_call_arg_meta *meta,
1307 enum bpf_access_type t)
1309 switch (env->prog->type) {
1310 case BPF_PROG_TYPE_LWT_IN:
1311 case BPF_PROG_TYPE_LWT_OUT:
1312 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1313 /* dst_input() and dst_output() can't write for now */
1317 case BPF_PROG_TYPE_SCHED_CLS:
1318 case BPF_PROG_TYPE_SCHED_ACT:
1319 case BPF_PROG_TYPE_XDP:
1320 case BPF_PROG_TYPE_LWT_XMIT:
1321 case BPF_PROG_TYPE_SK_SKB:
1322 case BPF_PROG_TYPE_SK_MSG:
1324 return meta->pkt_access;
1326 env->seen_direct_write = true;
1333 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1334 int off, int size, bool zero_size_allowed)
1336 struct bpf_reg_state *regs = cur_regs(env);
1337 struct bpf_reg_state *reg = ®s[regno];
1339 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1340 (u64)off + size > reg->range) {
1341 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1342 off, size, regno, reg->id, reg->off, reg->range);
1348 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1349 int size, bool zero_size_allowed)
1351 struct bpf_reg_state *regs = cur_regs(env);
1352 struct bpf_reg_state *reg = ®s[regno];
1355 /* We may have added a variable offset to the packet pointer; but any
1356 * reg->range we have comes after that. We are only checking the fixed
1360 /* We don't allow negative numbers, because we aren't tracking enough
1361 * detail to prove they're safe.
1363 if (reg->smin_value < 0) {
1364 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1368 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1370 verbose(env, "R%d offset is outside of the packet\n", regno);
1376 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1377 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1378 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1380 struct bpf_insn_access_aux info = {
1381 .reg_type = *reg_type,
1384 if (env->ops->is_valid_access &&
1385 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1386 /* A non zero info.ctx_field_size indicates that this field is a
1387 * candidate for later verifier transformation to load the whole
1388 * field and then apply a mask when accessed with a narrower
1389 * access than actual ctx access size. A zero info.ctx_field_size
1390 * will only allow for whole field access and rejects any other
1391 * type of narrower access.
1393 *reg_type = info.reg_type;
1395 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1396 /* remember the offset of last byte accessed in ctx */
1397 if (env->prog->aux->max_ctx_offset < off + size)
1398 env->prog->aux->max_ctx_offset = off + size;
1402 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1406 static bool __is_pointer_value(bool allow_ptr_leaks,
1407 const struct bpf_reg_state *reg)
1409 if (allow_ptr_leaks)
1412 return reg->type != SCALAR_VALUE;
1415 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1417 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1420 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1422 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1424 return reg->type == PTR_TO_CTX;
1427 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1429 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1431 return type_is_pkt_pointer(reg->type);
1434 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1435 const struct bpf_reg_state *reg,
1436 int off, int size, bool strict)
1438 struct tnum reg_off;
1441 /* Byte size accesses are always allowed. */
1442 if (!strict || size == 1)
1445 /* For platforms that do not have a Kconfig enabling
1446 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1447 * NET_IP_ALIGN is universally set to '2'. And on platforms
1448 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1449 * to this code only in strict mode where we want to emulate
1450 * the NET_IP_ALIGN==2 checking. Therefore use an
1451 * unconditional IP align value of '2'.
1455 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1456 if (!tnum_is_aligned(reg_off, size)) {
1459 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1461 "misaligned packet access off %d+%s+%d+%d size %d\n",
1462 ip_align, tn_buf, reg->off, off, size);
1469 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1470 const struct bpf_reg_state *reg,
1471 const char *pointer_desc,
1472 int off, int size, bool strict)
1474 struct tnum reg_off;
1476 /* Byte size accesses are always allowed. */
1477 if (!strict || size == 1)
1480 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1481 if (!tnum_is_aligned(reg_off, size)) {
1484 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1485 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1486 pointer_desc, tn_buf, reg->off, off, size);
1493 static int check_ptr_alignment(struct bpf_verifier_env *env,
1494 const struct bpf_reg_state *reg, int off,
1495 int size, bool strict_alignment_once)
1497 bool strict = env->strict_alignment || strict_alignment_once;
1498 const char *pointer_desc = "";
1500 switch (reg->type) {
1502 case PTR_TO_PACKET_META:
1503 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1504 * right in front, treat it the very same way.
1506 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1507 case PTR_TO_MAP_VALUE:
1508 pointer_desc = "value ";
1511 pointer_desc = "context ";
1514 pointer_desc = "stack ";
1515 /* The stack spill tracking logic in check_stack_write()
1516 * and check_stack_read() relies on stack accesses being
1524 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1528 static int update_stack_depth(struct bpf_verifier_env *env,
1529 const struct bpf_func_state *func,
1532 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1537 /* update known max for given subprogram */
1538 env->subprog_info[func->subprogno].stack_depth = -off;
1542 /* starting from main bpf function walk all instructions of the function
1543 * and recursively walk all callees that given function can call.
1544 * Ignore jump and exit insns.
1545 * Since recursion is prevented by check_cfg() this algorithm
1546 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1548 static int check_max_stack_depth(struct bpf_verifier_env *env)
1550 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1551 struct bpf_subprog_info *subprog = env->subprog_info;
1552 struct bpf_insn *insn = env->prog->insnsi;
1553 int ret_insn[MAX_CALL_FRAMES];
1554 int ret_prog[MAX_CALL_FRAMES];
1557 /* round up to 32-bytes, since this is granularity
1558 * of interpreter stack size
1560 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1561 if (depth > MAX_BPF_STACK) {
1562 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1567 subprog_end = subprog[idx + 1].start;
1568 for (; i < subprog_end; i++) {
1569 if (insn[i].code != (BPF_JMP | BPF_CALL))
1571 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1573 /* remember insn and function to return to */
1574 ret_insn[frame] = i + 1;
1575 ret_prog[frame] = idx;
1577 /* find the callee */
1578 i = i + insn[i].imm + 1;
1579 idx = find_subprog(env, i);
1581 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1586 if (frame >= MAX_CALL_FRAMES) {
1587 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1592 /* end of for() loop means the last insn of the 'subprog'
1593 * was reached. Doesn't matter whether it was JA or EXIT
1597 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1599 i = ret_insn[frame];
1600 idx = ret_prog[frame];
1604 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1605 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1606 const struct bpf_insn *insn, int idx)
1608 int start = idx + insn->imm + 1, subprog;
1610 subprog = find_subprog(env, start);
1612 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1616 return env->subprog_info[subprog].stack_depth;
1620 static int check_ctx_reg(struct bpf_verifier_env *env,
1621 const struct bpf_reg_state *reg, int regno)
1623 /* Access to ctx or passing it to a helper is only allowed in
1624 * its original, unmodified form.
1628 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1633 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1636 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1637 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1644 /* truncate register to smaller size (in bytes)
1645 * must be called with size < BPF_REG_SIZE
1647 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1651 /* clear high bits in bit representation */
1652 reg->var_off = tnum_cast(reg->var_off, size);
1654 /* fix arithmetic bounds */
1655 mask = ((u64)1 << (size * 8)) - 1;
1656 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1657 reg->umin_value &= mask;
1658 reg->umax_value &= mask;
1660 reg->umin_value = 0;
1661 reg->umax_value = mask;
1663 reg->smin_value = reg->umin_value;
1664 reg->smax_value = reg->umax_value;
1667 /* check whether memory at (regno + off) is accessible for t = (read | write)
1668 * if t==write, value_regno is a register which value is stored into memory
1669 * if t==read, value_regno is a register which will receive the value from memory
1670 * if t==write && value_regno==-1, some unknown value is stored into memory
1671 * if t==read && value_regno==-1, don't care what we read from memory
1673 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1674 int off, int bpf_size, enum bpf_access_type t,
1675 int value_regno, bool strict_alignment_once)
1677 struct bpf_reg_state *regs = cur_regs(env);
1678 struct bpf_reg_state *reg = regs + regno;
1679 struct bpf_func_state *state;
1682 size = bpf_size_to_bytes(bpf_size);
1686 /* alignment checks will add in reg->off themselves */
1687 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1691 /* for access checks, reg->off is just part of off */
1694 if (reg->type == PTR_TO_MAP_VALUE) {
1695 if (t == BPF_WRITE && value_regno >= 0 &&
1696 is_pointer_value(env, value_regno)) {
1697 verbose(env, "R%d leaks addr into map\n", value_regno);
1701 err = check_map_access(env, regno, off, size, false);
1702 if (!err && t == BPF_READ && value_regno >= 0)
1703 mark_reg_unknown(env, regs, value_regno);
1705 } else if (reg->type == PTR_TO_CTX) {
1706 enum bpf_reg_type reg_type = SCALAR_VALUE;
1708 if (t == BPF_WRITE && value_regno >= 0 &&
1709 is_pointer_value(env, value_regno)) {
1710 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1714 err = check_ctx_reg(env, reg, regno);
1718 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1719 if (!err && t == BPF_READ && value_regno >= 0) {
1720 /* ctx access returns either a scalar, or a
1721 * PTR_TO_PACKET[_META,_END]. In the latter
1722 * case, we know the offset is zero.
1724 if (reg_type == SCALAR_VALUE)
1725 mark_reg_unknown(env, regs, value_regno);
1727 mark_reg_known_zero(env, regs,
1729 regs[value_regno].id = 0;
1730 regs[value_regno].off = 0;
1731 regs[value_regno].range = 0;
1732 regs[value_regno].type = reg_type;
1735 } else if (reg->type == PTR_TO_STACK) {
1736 /* stack accesses must be at a fixed offset, so that we can
1737 * determine what type of data were returned.
1738 * See check_stack_read().
1740 if (!tnum_is_const(reg->var_off)) {
1743 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1744 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1748 off += reg->var_off.value;
1749 if (off >= 0 || off < -MAX_BPF_STACK) {
1750 verbose(env, "invalid stack off=%d size=%d\n", off,
1755 state = func(env, reg);
1756 err = update_stack_depth(env, state, off);
1761 err = check_stack_write(env, state, off, size,
1762 value_regno, insn_idx);
1764 err = check_stack_read(env, state, off, size,
1766 } else if (reg_is_pkt_pointer(reg)) {
1767 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1768 verbose(env, "cannot write into packet\n");
1771 if (t == BPF_WRITE && value_regno >= 0 &&
1772 is_pointer_value(env, value_regno)) {
1773 verbose(env, "R%d leaks addr into packet\n",
1777 err = check_packet_access(env, regno, off, size, false);
1778 if (!err && t == BPF_READ && value_regno >= 0)
1779 mark_reg_unknown(env, regs, value_regno);
1781 verbose(env, "R%d invalid mem access '%s'\n", regno,
1782 reg_type_str[reg->type]);
1786 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1787 regs[value_regno].type == SCALAR_VALUE) {
1788 /* b/h/w load zero-extends, mark upper bits as known 0 */
1789 coerce_reg_to_size(®s[value_regno], size);
1794 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1798 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1800 verbose(env, "BPF_XADD uses reserved fields\n");
1804 /* check src1 operand */
1805 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1809 /* check src2 operand */
1810 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1814 if (is_pointer_value(env, insn->src_reg)) {
1815 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1819 if (is_ctx_reg(env, insn->dst_reg) ||
1820 is_pkt_reg(env, insn->dst_reg)) {
1821 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1822 insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ?
1823 "context" : "packet");
1827 /* check whether atomic_add can read the memory */
1828 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1829 BPF_SIZE(insn->code), BPF_READ, -1, true);
1833 /* check whether atomic_add can write into the same memory */
1834 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1835 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1838 /* when register 'regno' is passed into function that will read 'access_size'
1839 * bytes from that pointer, make sure that it's within stack boundary
1840 * and all elements of stack are initialized.
1841 * Unlike most pointer bounds-checking functions, this one doesn't take an
1842 * 'off' argument, so it has to add in reg->off itself.
1844 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1845 int access_size, bool zero_size_allowed,
1846 struct bpf_call_arg_meta *meta)
1848 struct bpf_reg_state *reg = cur_regs(env) + regno;
1849 struct bpf_func_state *state = func(env, reg);
1850 int off, i, slot, spi;
1852 if (reg->type != PTR_TO_STACK) {
1853 /* Allow zero-byte read from NULL, regardless of pointer type */
1854 if (zero_size_allowed && access_size == 0 &&
1855 register_is_null(reg))
1858 verbose(env, "R%d type=%s expected=%s\n", regno,
1859 reg_type_str[reg->type],
1860 reg_type_str[PTR_TO_STACK]);
1864 /* Only allow fixed-offset stack reads */
1865 if (!tnum_is_const(reg->var_off)) {
1868 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1869 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1873 off = reg->off + reg->var_off.value;
1874 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1875 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1876 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1877 regno, off, access_size);
1881 if (meta && meta->raw_mode) {
1882 meta->access_size = access_size;
1883 meta->regno = regno;
1887 for (i = 0; i < access_size; i++) {
1890 slot = -(off + i) - 1;
1891 spi = slot / BPF_REG_SIZE;
1892 if (state->allocated_stack <= slot)
1894 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
1895 if (*stype == STACK_MISC)
1897 if (*stype == STACK_ZERO) {
1898 /* helper can write anything into the stack */
1899 *stype = STACK_MISC;
1903 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1904 off, i, access_size);
1907 /* reading any byte out of 8-byte 'spill_slot' will cause
1908 * the whole slot to be marked as 'read'
1910 mark_stack_slot_read(env, env->cur_state, env->cur_state->parent,
1911 spi, state->frameno);
1913 return update_stack_depth(env, state, off);
1916 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1917 int access_size, bool zero_size_allowed,
1918 struct bpf_call_arg_meta *meta)
1920 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1922 switch (reg->type) {
1924 case PTR_TO_PACKET_META:
1925 return check_packet_access(env, regno, reg->off, access_size,
1927 case PTR_TO_MAP_VALUE:
1928 return check_map_access(env, regno, reg->off, access_size,
1930 default: /* scalar_value|ptr_to_stack or invalid ptr */
1931 return check_stack_boundary(env, regno, access_size,
1932 zero_size_allowed, meta);
1936 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
1938 return type == ARG_PTR_TO_MEM ||
1939 type == ARG_PTR_TO_MEM_OR_NULL ||
1940 type == ARG_PTR_TO_UNINIT_MEM;
1943 static bool arg_type_is_mem_size(enum bpf_arg_type type)
1945 return type == ARG_CONST_SIZE ||
1946 type == ARG_CONST_SIZE_OR_ZERO;
1949 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1950 enum bpf_arg_type arg_type,
1951 struct bpf_call_arg_meta *meta)
1953 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1954 enum bpf_reg_type expected_type, type = reg->type;
1957 if (arg_type == ARG_DONTCARE)
1960 err = check_reg_arg(env, regno, SRC_OP);
1964 if (arg_type == ARG_ANYTHING) {
1965 if (is_pointer_value(env, regno)) {
1966 verbose(env, "R%d leaks addr into helper function\n",
1973 if (type_is_pkt_pointer(type) &&
1974 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1975 verbose(env, "helper access to the packet is not allowed\n");
1979 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1980 arg_type == ARG_PTR_TO_MAP_VALUE) {
1981 expected_type = PTR_TO_STACK;
1982 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
1983 type != expected_type)
1985 } else if (arg_type == ARG_CONST_SIZE ||
1986 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1987 expected_type = SCALAR_VALUE;
1988 if (type != expected_type)
1990 } else if (arg_type == ARG_CONST_MAP_PTR) {
1991 expected_type = CONST_PTR_TO_MAP;
1992 if (type != expected_type)
1994 } else if (arg_type == ARG_PTR_TO_CTX) {
1995 expected_type = PTR_TO_CTX;
1996 if (type != expected_type)
1998 err = check_ctx_reg(env, reg, regno);
2001 } else if (arg_type_is_mem_ptr(arg_type)) {
2002 expected_type = PTR_TO_STACK;
2003 /* One exception here. In case function allows for NULL to be
2004 * passed in as argument, it's a SCALAR_VALUE type. Final test
2005 * happens during stack boundary checking.
2007 if (register_is_null(reg) &&
2008 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2009 /* final test in check_stack_boundary() */;
2010 else if (!type_is_pkt_pointer(type) &&
2011 type != PTR_TO_MAP_VALUE &&
2012 type != expected_type)
2014 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2016 verbose(env, "unsupported arg_type %d\n", arg_type);
2020 if (arg_type == ARG_CONST_MAP_PTR) {
2021 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2022 meta->map_ptr = reg->map_ptr;
2023 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2024 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2025 * check that [key, key + map->key_size) are within
2026 * stack limits and initialized
2028 if (!meta->map_ptr) {
2029 /* in function declaration map_ptr must come before
2030 * map_key, so that it's verified and known before
2031 * we have to check map_key here. Otherwise it means
2032 * that kernel subsystem misconfigured verifier
2034 verbose(env, "invalid map_ptr to access map->key\n");
2037 err = check_helper_mem_access(env, regno,
2038 meta->map_ptr->key_size, false,
2040 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
2041 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2042 * check [value, value + map->value_size) validity
2044 if (!meta->map_ptr) {
2045 /* kernel subsystem misconfigured verifier */
2046 verbose(env, "invalid map_ptr to access map->value\n");
2049 err = check_helper_mem_access(env, regno,
2050 meta->map_ptr->value_size, false,
2052 } else if (arg_type_is_mem_size(arg_type)) {
2053 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2055 /* remember the mem_size which may be used later
2056 * to refine return values.
2058 meta->msize_smax_value = reg->smax_value;
2059 meta->msize_umax_value = reg->umax_value;
2061 /* The register is SCALAR_VALUE; the access check
2062 * happens using its boundaries.
2064 if (!tnum_is_const(reg->var_off))
2065 /* For unprivileged variable accesses, disable raw
2066 * mode so that the program is required to
2067 * initialize all the memory that the helper could
2068 * just partially fill up.
2072 if (reg->smin_value < 0) {
2073 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2078 if (reg->umin_value == 0) {
2079 err = check_helper_mem_access(env, regno - 1, 0,
2086 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2087 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2091 err = check_helper_mem_access(env, regno - 1,
2093 zero_size_allowed, meta);
2098 verbose(env, "R%d type=%s expected=%s\n", regno,
2099 reg_type_str[type], reg_type_str[expected_type]);
2103 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2104 struct bpf_map *map, int func_id)
2109 /* We need a two way check, first is from map perspective ... */
2110 switch (map->map_type) {
2111 case BPF_MAP_TYPE_PROG_ARRAY:
2112 if (func_id != BPF_FUNC_tail_call)
2115 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2116 if (func_id != BPF_FUNC_perf_event_read &&
2117 func_id != BPF_FUNC_perf_event_output &&
2118 func_id != BPF_FUNC_perf_event_read_value)
2121 case BPF_MAP_TYPE_STACK_TRACE:
2122 if (func_id != BPF_FUNC_get_stackid)
2125 case BPF_MAP_TYPE_CGROUP_ARRAY:
2126 if (func_id != BPF_FUNC_skb_under_cgroup &&
2127 func_id != BPF_FUNC_current_task_under_cgroup)
2130 /* devmap returns a pointer to a live net_device ifindex that we cannot
2131 * allow to be modified from bpf side. So do not allow lookup elements
2134 case BPF_MAP_TYPE_DEVMAP:
2135 if (func_id != BPF_FUNC_redirect_map)
2138 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2141 case BPF_MAP_TYPE_CPUMAP:
2142 case BPF_MAP_TYPE_XSKMAP:
2143 if (func_id != BPF_FUNC_redirect_map)
2146 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2147 case BPF_MAP_TYPE_HASH_OF_MAPS:
2148 if (func_id != BPF_FUNC_map_lookup_elem)
2151 case BPF_MAP_TYPE_SOCKMAP:
2152 if (func_id != BPF_FUNC_sk_redirect_map &&
2153 func_id != BPF_FUNC_sock_map_update &&
2154 func_id != BPF_FUNC_map_delete_elem &&
2155 func_id != BPF_FUNC_msg_redirect_map)
2158 case BPF_MAP_TYPE_SOCKHASH:
2159 if (func_id != BPF_FUNC_sk_redirect_hash &&
2160 func_id != BPF_FUNC_sock_hash_update &&
2161 func_id != BPF_FUNC_map_delete_elem &&
2162 func_id != BPF_FUNC_msg_redirect_hash)
2169 /* ... and second from the function itself. */
2171 case BPF_FUNC_tail_call:
2172 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2174 if (env->subprog_cnt > 1) {
2175 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2179 case BPF_FUNC_perf_event_read:
2180 case BPF_FUNC_perf_event_output:
2181 case BPF_FUNC_perf_event_read_value:
2182 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2185 case BPF_FUNC_get_stackid:
2186 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2189 case BPF_FUNC_current_task_under_cgroup:
2190 case BPF_FUNC_skb_under_cgroup:
2191 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2194 case BPF_FUNC_redirect_map:
2195 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2196 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2197 map->map_type != BPF_MAP_TYPE_XSKMAP)
2200 case BPF_FUNC_sk_redirect_map:
2201 case BPF_FUNC_msg_redirect_map:
2202 case BPF_FUNC_sock_map_update:
2203 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2206 case BPF_FUNC_sk_redirect_hash:
2207 case BPF_FUNC_msg_redirect_hash:
2208 case BPF_FUNC_sock_hash_update:
2209 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2218 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2219 map->map_type, func_id_name(func_id), func_id);
2223 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2227 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2229 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2231 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2233 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2235 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2238 /* We only support one arg being in raw mode at the moment,
2239 * which is sufficient for the helper functions we have
2245 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2246 enum bpf_arg_type arg_next)
2248 return (arg_type_is_mem_ptr(arg_curr) &&
2249 !arg_type_is_mem_size(arg_next)) ||
2250 (!arg_type_is_mem_ptr(arg_curr) &&
2251 arg_type_is_mem_size(arg_next));
2254 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2256 /* bpf_xxx(..., buf, len) call will access 'len'
2257 * bytes from memory 'buf'. Both arg types need
2258 * to be paired, so make sure there's no buggy
2259 * helper function specification.
2261 if (arg_type_is_mem_size(fn->arg1_type) ||
2262 arg_type_is_mem_ptr(fn->arg5_type) ||
2263 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2264 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2265 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2266 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2272 static int check_func_proto(const struct bpf_func_proto *fn)
2274 return check_raw_mode_ok(fn) &&
2275 check_arg_pair_ok(fn) ? 0 : -EINVAL;
2278 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2279 * are now invalid, so turn them into unknown SCALAR_VALUE.
2281 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2282 struct bpf_func_state *state)
2284 struct bpf_reg_state *regs = state->regs, *reg;
2287 for (i = 0; i < MAX_BPF_REG; i++)
2288 if (reg_is_pkt_pointer_any(®s[i]))
2289 mark_reg_unknown(env, regs, i);
2291 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2292 if (state->stack[i].slot_type[0] != STACK_SPILL)
2294 reg = &state->stack[i].spilled_ptr;
2295 if (reg_is_pkt_pointer_any(reg))
2296 __mark_reg_unknown(reg);
2300 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2302 struct bpf_verifier_state *vstate = env->cur_state;
2305 for (i = 0; i <= vstate->curframe; i++)
2306 __clear_all_pkt_pointers(env, vstate->frame[i]);
2309 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2312 struct bpf_verifier_state *state = env->cur_state;
2313 struct bpf_func_state *caller, *callee;
2314 int i, subprog, target_insn;
2316 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2317 verbose(env, "the call stack of %d frames is too deep\n",
2318 state->curframe + 2);
2322 target_insn = *insn_idx + insn->imm;
2323 subprog = find_subprog(env, target_insn + 1);
2325 verbose(env, "verifier bug. No program starts at insn %d\n",
2330 caller = state->frame[state->curframe];
2331 if (state->frame[state->curframe + 1]) {
2332 verbose(env, "verifier bug. Frame %d already allocated\n",
2333 state->curframe + 1);
2337 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2340 state->frame[state->curframe + 1] = callee;
2342 /* callee cannot access r0, r6 - r9 for reading and has to write
2343 * into its own stack before reading from it.
2344 * callee can read/write into caller's stack
2346 init_func_state(env, callee,
2347 /* remember the callsite, it will be used by bpf_exit */
2348 *insn_idx /* callsite */,
2349 state->curframe + 1 /* frameno within this callchain */,
2350 subprog /* subprog number within this prog */);
2352 /* copy r1 - r5 args that callee can access */
2353 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2354 callee->regs[i] = caller->regs[i];
2356 /* after the call regsiters r0 - r5 were scratched */
2357 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2358 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2359 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2362 /* only increment it after check_reg_arg() finished */
2365 /* and go analyze first insn of the callee */
2366 *insn_idx = target_insn;
2368 if (env->log.level) {
2369 verbose(env, "caller:\n");
2370 print_verifier_state(env, caller);
2371 verbose(env, "callee:\n");
2372 print_verifier_state(env, callee);
2377 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2379 struct bpf_verifier_state *state = env->cur_state;
2380 struct bpf_func_state *caller, *callee;
2381 struct bpf_reg_state *r0;
2383 callee = state->frame[state->curframe];
2384 r0 = &callee->regs[BPF_REG_0];
2385 if (r0->type == PTR_TO_STACK) {
2386 /* technically it's ok to return caller's stack pointer
2387 * (or caller's caller's pointer) back to the caller,
2388 * since these pointers are valid. Only current stack
2389 * pointer will be invalid as soon as function exits,
2390 * but let's be conservative
2392 verbose(env, "cannot return stack pointer to the caller\n");
2397 caller = state->frame[state->curframe];
2398 /* return to the caller whatever r0 had in the callee */
2399 caller->regs[BPF_REG_0] = *r0;
2401 *insn_idx = callee->callsite + 1;
2402 if (env->log.level) {
2403 verbose(env, "returning from callee:\n");
2404 print_verifier_state(env, callee);
2405 verbose(env, "to caller at %d:\n", *insn_idx);
2406 print_verifier_state(env, caller);
2408 /* clear everything in the callee */
2409 free_func_state(callee);
2410 state->frame[state->curframe + 1] = NULL;
2414 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2416 struct bpf_call_arg_meta *meta)
2418 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2420 if (ret_type != RET_INTEGER ||
2421 (func_id != BPF_FUNC_get_stack &&
2422 func_id != BPF_FUNC_probe_read_str))
2425 ret_reg->smax_value = meta->msize_smax_value;
2426 ret_reg->umax_value = meta->msize_umax_value;
2427 __reg_deduce_bounds(ret_reg);
2428 __reg_bound_offset(ret_reg);
2432 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2433 int func_id, int insn_idx)
2435 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2437 if (func_id != BPF_FUNC_tail_call &&
2438 func_id != BPF_FUNC_map_lookup_elem &&
2439 func_id != BPF_FUNC_map_update_elem &&
2440 func_id != BPF_FUNC_map_delete_elem)
2443 if (meta->map_ptr == NULL) {
2444 verbose(env, "kernel subsystem misconfigured verifier\n");
2448 if (!BPF_MAP_PTR(aux->map_state))
2449 bpf_map_ptr_store(aux, meta->map_ptr,
2450 meta->map_ptr->unpriv_array);
2451 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2452 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2453 meta->map_ptr->unpriv_array);
2457 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2459 const struct bpf_func_proto *fn = NULL;
2460 struct bpf_reg_state *regs;
2461 struct bpf_call_arg_meta meta;
2465 /* find function prototype */
2466 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2467 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2472 if (env->ops->get_func_proto)
2473 fn = env->ops->get_func_proto(func_id, env->prog);
2475 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2480 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2481 if (!env->prog->gpl_compatible && fn->gpl_only) {
2482 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2486 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2487 changes_data = bpf_helper_changes_pkt_data(fn->func);
2488 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2489 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2490 func_id_name(func_id), func_id);
2494 memset(&meta, 0, sizeof(meta));
2495 meta.pkt_access = fn->pkt_access;
2497 err = check_func_proto(fn);
2499 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2500 func_id_name(func_id), func_id);
2505 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2508 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2511 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2514 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2517 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2521 err = record_func_map(env, &meta, func_id, insn_idx);
2525 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2526 * is inferred from register state.
2528 for (i = 0; i < meta.access_size; i++) {
2529 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2530 BPF_WRITE, -1, false);
2535 regs = cur_regs(env);
2536 /* reset caller saved regs */
2537 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2538 mark_reg_not_init(env, regs, caller_saved[i]);
2539 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2542 /* update return register (already marked as written above) */
2543 if (fn->ret_type == RET_INTEGER) {
2544 /* sets type to SCALAR_VALUE */
2545 mark_reg_unknown(env, regs, BPF_REG_0);
2546 } else if (fn->ret_type == RET_VOID) {
2547 regs[BPF_REG_0].type = NOT_INIT;
2548 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2549 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2550 if (fn->ret_type == RET_PTR_TO_MAP_VALUE)
2551 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2553 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2554 /* There is no offset yet applied, variable or fixed */
2555 mark_reg_known_zero(env, regs, BPF_REG_0);
2556 regs[BPF_REG_0].off = 0;
2557 /* remember map_ptr, so that check_map_access()
2558 * can check 'value_size' boundary of memory access
2559 * to map element returned from bpf_map_lookup_elem()
2561 if (meta.map_ptr == NULL) {
2563 "kernel subsystem misconfigured verifier\n");
2566 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2567 regs[BPF_REG_0].id = ++env->id_gen;
2569 verbose(env, "unknown return type %d of func %s#%d\n",
2570 fn->ret_type, func_id_name(func_id), func_id);
2574 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2576 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2580 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2581 const char *err_str;
2583 #ifdef CONFIG_PERF_EVENTS
2584 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2585 err_str = "cannot get callchain buffer for func %s#%d\n";
2588 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2591 verbose(env, err_str, func_id_name(func_id), func_id);
2595 env->prog->has_callchain_buf = true;
2599 clear_all_pkt_pointers(env);
2603 static bool signed_add_overflows(s64 a, s64 b)
2605 /* Do the add in u64, where overflow is well-defined */
2606 s64 res = (s64)((u64)a + (u64)b);
2613 static bool signed_sub_overflows(s64 a, s64 b)
2615 /* Do the sub in u64, where overflow is well-defined */
2616 s64 res = (s64)((u64)a - (u64)b);
2623 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2624 const struct bpf_reg_state *reg,
2625 enum bpf_reg_type type)
2627 bool known = tnum_is_const(reg->var_off);
2628 s64 val = reg->var_off.value;
2629 s64 smin = reg->smin_value;
2631 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2632 verbose(env, "math between %s pointer and %lld is not allowed\n",
2633 reg_type_str[type], val);
2637 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2638 verbose(env, "%s pointer offset %d is not allowed\n",
2639 reg_type_str[type], reg->off);
2643 if (smin == S64_MIN) {
2644 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2645 reg_type_str[type]);
2649 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2650 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2651 smin, reg_type_str[type]);
2658 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2659 * Caller should also handle BPF_MOV case separately.
2660 * If we return -EACCES, caller may want to try again treating pointer as a
2661 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2663 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2664 struct bpf_insn *insn,
2665 const struct bpf_reg_state *ptr_reg,
2666 const struct bpf_reg_state *off_reg)
2668 struct bpf_verifier_state *vstate = env->cur_state;
2669 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2670 struct bpf_reg_state *regs = state->regs, *dst_reg;
2671 bool known = tnum_is_const(off_reg->var_off);
2672 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2673 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2674 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2675 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2676 u8 opcode = BPF_OP(insn->code);
2677 u32 dst = insn->dst_reg;
2679 dst_reg = ®s[dst];
2681 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2682 smin_val > smax_val || umin_val > umax_val) {
2683 /* Taint dst register if offset had invalid bounds derived from
2684 * e.g. dead branches.
2686 __mark_reg_unknown(dst_reg);
2690 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2691 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2693 "R%d 32-bit pointer arithmetic prohibited\n",
2698 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2699 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2703 if (ptr_reg->type == CONST_PTR_TO_MAP) {
2704 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2708 if (ptr_reg->type == PTR_TO_PACKET_END) {
2709 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2714 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2715 * The id may be overwritten later if we create a new variable offset.
2717 dst_reg->type = ptr_reg->type;
2718 dst_reg->id = ptr_reg->id;
2720 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
2721 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
2726 /* We can take a fixed offset as long as it doesn't overflow
2727 * the s32 'off' field
2729 if (known && (ptr_reg->off + smin_val ==
2730 (s64)(s32)(ptr_reg->off + smin_val))) {
2731 /* pointer += K. Accumulate it into fixed offset */
2732 dst_reg->smin_value = smin_ptr;
2733 dst_reg->smax_value = smax_ptr;
2734 dst_reg->umin_value = umin_ptr;
2735 dst_reg->umax_value = umax_ptr;
2736 dst_reg->var_off = ptr_reg->var_off;
2737 dst_reg->off = ptr_reg->off + smin_val;
2738 dst_reg->range = ptr_reg->range;
2741 /* A new variable offset is created. Note that off_reg->off
2742 * == 0, since it's a scalar.
2743 * dst_reg gets the pointer type and since some positive
2744 * integer value was added to the pointer, give it a new 'id'
2745 * if it's a PTR_TO_PACKET.
2746 * this creates a new 'base' pointer, off_reg (variable) gets
2747 * added into the variable offset, and we copy the fixed offset
2750 if (signed_add_overflows(smin_ptr, smin_val) ||
2751 signed_add_overflows(smax_ptr, smax_val)) {
2752 dst_reg->smin_value = S64_MIN;
2753 dst_reg->smax_value = S64_MAX;
2755 dst_reg->smin_value = smin_ptr + smin_val;
2756 dst_reg->smax_value = smax_ptr + smax_val;
2758 if (umin_ptr + umin_val < umin_ptr ||
2759 umax_ptr + umax_val < umax_ptr) {
2760 dst_reg->umin_value = 0;
2761 dst_reg->umax_value = U64_MAX;
2763 dst_reg->umin_value = umin_ptr + umin_val;
2764 dst_reg->umax_value = umax_ptr + umax_val;
2766 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
2767 dst_reg->off = ptr_reg->off;
2768 if (reg_is_pkt_pointer(ptr_reg)) {
2769 dst_reg->id = ++env->id_gen;
2770 /* something was added to pkt_ptr, set range to zero */
2775 if (dst_reg == off_reg) {
2776 /* scalar -= pointer. Creates an unknown scalar */
2777 verbose(env, "R%d tried to subtract pointer from scalar\n",
2781 /* We don't allow subtraction from FP, because (according to
2782 * test_verifier.c test "invalid fp arithmetic", JITs might not
2783 * be able to deal with it.
2785 if (ptr_reg->type == PTR_TO_STACK) {
2786 verbose(env, "R%d subtraction from stack pointer prohibited\n",
2790 if (known && (ptr_reg->off - smin_val ==
2791 (s64)(s32)(ptr_reg->off - smin_val))) {
2792 /* pointer -= K. Subtract it from fixed offset */
2793 dst_reg->smin_value = smin_ptr;
2794 dst_reg->smax_value = smax_ptr;
2795 dst_reg->umin_value = umin_ptr;
2796 dst_reg->umax_value = umax_ptr;
2797 dst_reg->var_off = ptr_reg->var_off;
2798 dst_reg->id = ptr_reg->id;
2799 dst_reg->off = ptr_reg->off - smin_val;
2800 dst_reg->range = ptr_reg->range;
2803 /* A new variable offset is created. If the subtrahend is known
2804 * nonnegative, then any reg->range we had before is still good.
2806 if (signed_sub_overflows(smin_ptr, smax_val) ||
2807 signed_sub_overflows(smax_ptr, smin_val)) {
2808 /* Overflow possible, we know nothing */
2809 dst_reg->smin_value = S64_MIN;
2810 dst_reg->smax_value = S64_MAX;
2812 dst_reg->smin_value = smin_ptr - smax_val;
2813 dst_reg->smax_value = smax_ptr - smin_val;
2815 if (umin_ptr < umax_val) {
2816 /* Overflow possible, we know nothing */
2817 dst_reg->umin_value = 0;
2818 dst_reg->umax_value = U64_MAX;
2820 /* Cannot overflow (as long as bounds are consistent) */
2821 dst_reg->umin_value = umin_ptr - umax_val;
2822 dst_reg->umax_value = umax_ptr - umin_val;
2824 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2825 dst_reg->off = ptr_reg->off;
2826 if (reg_is_pkt_pointer(ptr_reg)) {
2827 dst_reg->id = ++env->id_gen;
2828 /* something was added to pkt_ptr, set range to zero */
2836 /* bitwise ops on pointers are troublesome, prohibit. */
2837 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2838 dst, bpf_alu_string[opcode >> 4]);
2841 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2842 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2843 dst, bpf_alu_string[opcode >> 4]);
2847 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2850 __update_reg_bounds(dst_reg);
2851 __reg_deduce_bounds(dst_reg);
2852 __reg_bound_offset(dst_reg);
2856 /* WARNING: This function does calculations on 64-bit values, but the actual
2857 * execution may occur on 32-bit values. Therefore, things like bitshifts
2858 * need extra checks in the 32-bit case.
2860 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2861 struct bpf_insn *insn,
2862 struct bpf_reg_state *dst_reg,
2863 struct bpf_reg_state src_reg)
2865 struct bpf_reg_state *regs = cur_regs(env);
2866 u8 opcode = BPF_OP(insn->code);
2867 bool src_known, dst_known;
2868 s64 smin_val, smax_val;
2869 u64 umin_val, umax_val;
2870 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2872 smin_val = src_reg.smin_value;
2873 smax_val = src_reg.smax_value;
2874 umin_val = src_reg.umin_value;
2875 umax_val = src_reg.umax_value;
2876 src_known = tnum_is_const(src_reg.var_off);
2877 dst_known = tnum_is_const(dst_reg->var_off);
2879 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2880 smin_val > smax_val || umin_val > umax_val) {
2881 /* Taint dst register if offset had invalid bounds derived from
2882 * e.g. dead branches.
2884 __mark_reg_unknown(dst_reg);
2889 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2890 __mark_reg_unknown(dst_reg);
2896 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2897 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2898 dst_reg->smin_value = S64_MIN;
2899 dst_reg->smax_value = S64_MAX;
2901 dst_reg->smin_value += smin_val;
2902 dst_reg->smax_value += smax_val;
2904 if (dst_reg->umin_value + umin_val < umin_val ||
2905 dst_reg->umax_value + umax_val < umax_val) {
2906 dst_reg->umin_value = 0;
2907 dst_reg->umax_value = U64_MAX;
2909 dst_reg->umin_value += umin_val;
2910 dst_reg->umax_value += umax_val;
2912 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2915 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2916 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2917 /* Overflow possible, we know nothing */
2918 dst_reg->smin_value = S64_MIN;
2919 dst_reg->smax_value = S64_MAX;
2921 dst_reg->smin_value -= smax_val;
2922 dst_reg->smax_value -= smin_val;
2924 if (dst_reg->umin_value < umax_val) {
2925 /* Overflow possible, we know nothing */
2926 dst_reg->umin_value = 0;
2927 dst_reg->umax_value = U64_MAX;
2929 /* Cannot overflow (as long as bounds are consistent) */
2930 dst_reg->umin_value -= umax_val;
2931 dst_reg->umax_value -= umin_val;
2933 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2936 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2937 if (smin_val < 0 || dst_reg->smin_value < 0) {
2938 /* Ain't nobody got time to multiply that sign */
2939 __mark_reg_unbounded(dst_reg);
2940 __update_reg_bounds(dst_reg);
2943 /* Both values are positive, so we can work with unsigned and
2944 * copy the result to signed (unless it exceeds S64_MAX).
2946 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2947 /* Potential overflow, we know nothing */
2948 __mark_reg_unbounded(dst_reg);
2949 /* (except what we can learn from the var_off) */
2950 __update_reg_bounds(dst_reg);
2953 dst_reg->umin_value *= umin_val;
2954 dst_reg->umax_value *= umax_val;
2955 if (dst_reg->umax_value > S64_MAX) {
2956 /* Overflow possible, we know nothing */
2957 dst_reg->smin_value = S64_MIN;
2958 dst_reg->smax_value = S64_MAX;
2960 dst_reg->smin_value = dst_reg->umin_value;
2961 dst_reg->smax_value = dst_reg->umax_value;
2965 if (src_known && dst_known) {
2966 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2967 src_reg.var_off.value);
2970 /* We get our minimum from the var_off, since that's inherently
2971 * bitwise. Our maximum is the minimum of the operands' maxima.
2973 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2974 dst_reg->umin_value = dst_reg->var_off.value;
2975 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2976 if (dst_reg->smin_value < 0 || smin_val < 0) {
2977 /* Lose signed bounds when ANDing negative numbers,
2978 * ain't nobody got time for that.
2980 dst_reg->smin_value = S64_MIN;
2981 dst_reg->smax_value = S64_MAX;
2983 /* ANDing two positives gives a positive, so safe to
2984 * cast result into s64.
2986 dst_reg->smin_value = dst_reg->umin_value;
2987 dst_reg->smax_value = dst_reg->umax_value;
2989 /* We may learn something more from the var_off */
2990 __update_reg_bounds(dst_reg);
2993 if (src_known && dst_known) {
2994 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2995 src_reg.var_off.value);
2998 /* We get our maximum from the var_off, and our minimum is the
2999 * maximum of the operands' minima
3001 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3002 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3003 dst_reg->umax_value = dst_reg->var_off.value |
3004 dst_reg->var_off.mask;
3005 if (dst_reg->smin_value < 0 || smin_val < 0) {
3006 /* Lose signed bounds when ORing negative numbers,
3007 * ain't nobody got time for that.
3009 dst_reg->smin_value = S64_MIN;
3010 dst_reg->smax_value = S64_MAX;
3012 /* ORing two positives gives a positive, so safe to
3013 * cast result into s64.
3015 dst_reg->smin_value = dst_reg->umin_value;
3016 dst_reg->smax_value = dst_reg->umax_value;
3018 /* We may learn something more from the var_off */
3019 __update_reg_bounds(dst_reg);
3022 if (umax_val >= insn_bitness) {
3023 /* Shifts greater than 31 or 63 are undefined.
3024 * This includes shifts by a negative number.
3026 mark_reg_unknown(env, regs, insn->dst_reg);
3029 /* We lose all sign bit information (except what we can pick
3032 dst_reg->smin_value = S64_MIN;
3033 dst_reg->smax_value = S64_MAX;
3034 /* If we might shift our top bit out, then we know nothing */
3035 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3036 dst_reg->umin_value = 0;
3037 dst_reg->umax_value = U64_MAX;
3039 dst_reg->umin_value <<= umin_val;
3040 dst_reg->umax_value <<= umax_val;
3042 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3043 /* We may learn something more from the var_off */
3044 __update_reg_bounds(dst_reg);
3047 if (umax_val >= insn_bitness) {
3048 /* Shifts greater than 31 or 63 are undefined.
3049 * This includes shifts by a negative number.
3051 mark_reg_unknown(env, regs, insn->dst_reg);
3054 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3055 * be negative, then either:
3056 * 1) src_reg might be zero, so the sign bit of the result is
3057 * unknown, so we lose our signed bounds
3058 * 2) it's known negative, thus the unsigned bounds capture the
3060 * 3) the signed bounds cross zero, so they tell us nothing
3062 * If the value in dst_reg is known nonnegative, then again the
3063 * unsigned bounts capture the signed bounds.
3064 * Thus, in all cases it suffices to blow away our signed bounds
3065 * and rely on inferring new ones from the unsigned bounds and
3066 * var_off of the result.
3068 dst_reg->smin_value = S64_MIN;
3069 dst_reg->smax_value = S64_MAX;
3070 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3071 dst_reg->umin_value >>= umax_val;
3072 dst_reg->umax_value >>= umin_val;
3073 /* We may learn something more from the var_off */
3074 __update_reg_bounds(dst_reg);
3077 if (umax_val >= insn_bitness) {
3078 /* Shifts greater than 31 or 63 are undefined.
3079 * This includes shifts by a negative number.
3081 mark_reg_unknown(env, regs, insn->dst_reg);
3085 /* Upon reaching here, src_known is true and
3086 * umax_val is equal to umin_val.
3088 dst_reg->smin_value >>= umin_val;
3089 dst_reg->smax_value >>= umin_val;
3090 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3092 /* blow away the dst_reg umin_value/umax_value and rely on
3093 * dst_reg var_off to refine the result.
3095 dst_reg->umin_value = 0;
3096 dst_reg->umax_value = U64_MAX;
3097 __update_reg_bounds(dst_reg);
3100 mark_reg_unknown(env, regs, insn->dst_reg);
3104 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3105 /* 32-bit ALU ops are (32,32)->32 */
3106 coerce_reg_to_size(dst_reg, 4);
3107 coerce_reg_to_size(&src_reg, 4);
3110 __reg_deduce_bounds(dst_reg);
3111 __reg_bound_offset(dst_reg);
3115 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3118 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3119 struct bpf_insn *insn)
3121 struct bpf_verifier_state *vstate = env->cur_state;
3122 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3123 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3124 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3125 u8 opcode = BPF_OP(insn->code);
3127 dst_reg = ®s[insn->dst_reg];
3129 if (dst_reg->type != SCALAR_VALUE)
3131 if (BPF_SRC(insn->code) == BPF_X) {
3132 src_reg = ®s[insn->src_reg];
3133 if (src_reg->type != SCALAR_VALUE) {
3134 if (dst_reg->type != SCALAR_VALUE) {
3135 /* Combining two pointers by any ALU op yields
3136 * an arbitrary scalar. Disallow all math except
3137 * pointer subtraction
3139 if (opcode == BPF_SUB){
3140 mark_reg_unknown(env, regs, insn->dst_reg);
3143 verbose(env, "R%d pointer %s pointer prohibited\n",
3145 bpf_alu_string[opcode >> 4]);
3148 /* scalar += pointer
3149 * This is legal, but we have to reverse our
3150 * src/dest handling in computing the range
3152 return adjust_ptr_min_max_vals(env, insn,
3155 } else if (ptr_reg) {
3156 /* pointer += scalar */
3157 return adjust_ptr_min_max_vals(env, insn,
3161 /* Pretend the src is a reg with a known value, since we only
3162 * need to be able to read from this state.
3164 off_reg.type = SCALAR_VALUE;
3165 __mark_reg_known(&off_reg, insn->imm);
3167 if (ptr_reg) /* pointer += K */
3168 return adjust_ptr_min_max_vals(env, insn,
3172 /* Got here implies adding two SCALAR_VALUEs */
3173 if (WARN_ON_ONCE(ptr_reg)) {
3174 print_verifier_state(env, state);
3175 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3178 if (WARN_ON(!src_reg)) {
3179 print_verifier_state(env, state);
3180 verbose(env, "verifier internal error: no src_reg\n");
3183 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3186 /* check validity of 32-bit and 64-bit arithmetic operations */
3187 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3189 struct bpf_reg_state *regs = cur_regs(env);
3190 u8 opcode = BPF_OP(insn->code);
3193 if (opcode == BPF_END || opcode == BPF_NEG) {
3194 if (opcode == BPF_NEG) {
3195 if (BPF_SRC(insn->code) != 0 ||
3196 insn->src_reg != BPF_REG_0 ||
3197 insn->off != 0 || insn->imm != 0) {
3198 verbose(env, "BPF_NEG uses reserved fields\n");
3202 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3203 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3204 BPF_CLASS(insn->code) == BPF_ALU64) {
3205 verbose(env, "BPF_END uses reserved fields\n");
3210 /* check src operand */
3211 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3215 if (is_pointer_value(env, insn->dst_reg)) {
3216 verbose(env, "R%d pointer arithmetic prohibited\n",
3221 /* check dest operand */
3222 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3226 } else if (opcode == BPF_MOV) {
3228 if (BPF_SRC(insn->code) == BPF_X) {
3229 if (insn->imm != 0 || insn->off != 0) {
3230 verbose(env, "BPF_MOV uses reserved fields\n");
3234 /* check src operand */
3235 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3239 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3240 verbose(env, "BPF_MOV uses reserved fields\n");
3245 /* check dest operand, mark as required later */
3246 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3250 if (BPF_SRC(insn->code) == BPF_X) {
3251 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3253 * copy register state to dest reg
3255 regs[insn->dst_reg] = regs[insn->src_reg];
3256 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3259 if (is_pointer_value(env, insn->src_reg)) {
3261 "R%d partial copy of pointer\n",
3265 mark_reg_unknown(env, regs, insn->dst_reg);
3266 coerce_reg_to_size(®s[insn->dst_reg], 4);
3270 * remember the value we stored into this reg
3272 /* clear any state __mark_reg_known doesn't set */
3273 mark_reg_unknown(env, regs, insn->dst_reg);
3274 regs[insn->dst_reg].type = SCALAR_VALUE;
3275 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3276 __mark_reg_known(regs + insn->dst_reg,
3279 __mark_reg_known(regs + insn->dst_reg,
3284 } else if (opcode > BPF_END) {
3285 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3288 } else { /* all other ALU ops: and, sub, xor, add, ... */
3290 if (BPF_SRC(insn->code) == BPF_X) {
3291 if (insn->imm != 0 || insn->off != 0) {
3292 verbose(env, "BPF_ALU uses reserved fields\n");
3295 /* check src1 operand */
3296 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3300 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3301 verbose(env, "BPF_ALU uses reserved fields\n");
3306 /* check src2 operand */
3307 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3311 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3312 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3313 verbose(env, "div by zero\n");
3317 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3318 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3322 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3323 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3324 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3326 if (insn->imm < 0 || insn->imm >= size) {
3327 verbose(env, "invalid shift %d\n", insn->imm);
3332 /* check dest operand */
3333 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3337 return adjust_reg_min_max_vals(env, insn);
3343 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3344 struct bpf_reg_state *dst_reg,
3345 enum bpf_reg_type type,
3346 bool range_right_open)
3348 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3349 struct bpf_reg_state *regs = state->regs, *reg;
3353 if (dst_reg->off < 0 ||
3354 (dst_reg->off == 0 && range_right_open))
3355 /* This doesn't give us any range */
3358 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3359 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3360 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3361 * than pkt_end, but that's because it's also less than pkt.
3365 new_range = dst_reg->off;
3366 if (range_right_open)
3369 /* Examples for register markings:
3371 * pkt_data in dst register:
3375 * if (r2 > pkt_end) goto <handle exception>
3380 * if (r2 < pkt_end) goto <access okay>
3381 * <handle exception>
3384 * r2 == dst_reg, pkt_end == src_reg
3385 * r2=pkt(id=n,off=8,r=0)
3386 * r3=pkt(id=n,off=0,r=0)
3388 * pkt_data in src register:
3392 * if (pkt_end >= r2) goto <access okay>
3393 * <handle exception>
3397 * if (pkt_end <= r2) goto <handle exception>
3401 * pkt_end == dst_reg, r2 == src_reg
3402 * r2=pkt(id=n,off=8,r=0)
3403 * r3=pkt(id=n,off=0,r=0)
3405 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3406 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3407 * and [r3, r3 + 8-1) respectively is safe to access depending on
3411 /* If our ids match, then we must have the same max_value. And we
3412 * don't care about the other reg's fixed offset, since if it's too big
3413 * the range won't allow anything.
3414 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3416 for (i = 0; i < MAX_BPF_REG; i++)
3417 if (regs[i].type == type && regs[i].id == dst_reg->id)
3418 /* keep the maximum range already checked */
3419 regs[i].range = max(regs[i].range, new_range);
3421 for (j = 0; j <= vstate->curframe; j++) {
3422 state = vstate->frame[j];
3423 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3424 if (state->stack[i].slot_type[0] != STACK_SPILL)
3426 reg = &state->stack[i].spilled_ptr;
3427 if (reg->type == type && reg->id == dst_reg->id)
3428 reg->range = max(reg->range, new_range);
3433 /* Adjusts the register min/max values in the case that the dst_reg is the
3434 * variable register that we are working on, and src_reg is a constant or we're
3435 * simply doing a BPF_K check.
3436 * In JEQ/JNE cases we also adjust the var_off values.
3438 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3439 struct bpf_reg_state *false_reg, u64 val,
3442 /* If the dst_reg is a pointer, we can't learn anything about its
3443 * variable offset from the compare (unless src_reg were a pointer into
3444 * the same object, but we don't bother with that.
3445 * Since false_reg and true_reg have the same type by construction, we
3446 * only need to check one of them for pointerness.
3448 if (__is_pointer_value(false, false_reg))
3453 /* If this is false then we know nothing Jon Snow, but if it is
3454 * true then we know for sure.
3456 __mark_reg_known(true_reg, val);
3459 /* If this is true we know nothing Jon Snow, but if it is false
3460 * we know the value for sure;
3462 __mark_reg_known(false_reg, val);
3465 false_reg->umax_value = min(false_reg->umax_value, val);
3466 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3469 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3470 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3473 false_reg->umin_value = max(false_reg->umin_value, val);
3474 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3477 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3478 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3481 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3482 true_reg->umin_value = max(true_reg->umin_value, val);
3485 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3486 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3489 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3490 true_reg->umax_value = min(true_reg->umax_value, val);
3493 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3494 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3500 __reg_deduce_bounds(false_reg);
3501 __reg_deduce_bounds(true_reg);
3502 /* We might have learned some bits from the bounds. */
3503 __reg_bound_offset(false_reg);
3504 __reg_bound_offset(true_reg);
3505 /* Intersecting with the old var_off might have improved our bounds
3506 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3507 * then new var_off is (0; 0x7f...fc) which improves our umax.
3509 __update_reg_bounds(false_reg);
3510 __update_reg_bounds(true_reg);
3513 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3516 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3517 struct bpf_reg_state *false_reg, u64 val,
3520 if (__is_pointer_value(false, false_reg))
3525 /* If this is false then we know nothing Jon Snow, but if it is
3526 * true then we know for sure.
3528 __mark_reg_known(true_reg, val);
3531 /* If this is true we know nothing Jon Snow, but if it is false
3532 * we know the value for sure;
3534 __mark_reg_known(false_reg, val);
3537 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3538 false_reg->umin_value = max(false_reg->umin_value, val);
3541 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3542 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3545 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3546 false_reg->umax_value = min(false_reg->umax_value, val);
3549 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3550 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3553 true_reg->umax_value = min(true_reg->umax_value, val);
3554 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3557 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3558 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3561 true_reg->umin_value = max(true_reg->umin_value, val);
3562 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3565 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3566 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3572 __reg_deduce_bounds(false_reg);
3573 __reg_deduce_bounds(true_reg);
3574 /* We might have learned some bits from the bounds. */
3575 __reg_bound_offset(false_reg);
3576 __reg_bound_offset(true_reg);
3577 /* Intersecting with the old var_off might have improved our bounds
3578 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3579 * then new var_off is (0; 0x7f...fc) which improves our umax.
3581 __update_reg_bounds(false_reg);
3582 __update_reg_bounds(true_reg);
3585 /* Regs are known to be equal, so intersect their min/max/var_off */
3586 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3587 struct bpf_reg_state *dst_reg)
3589 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3590 dst_reg->umin_value);
3591 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3592 dst_reg->umax_value);
3593 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3594 dst_reg->smin_value);
3595 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3596 dst_reg->smax_value);
3597 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3599 /* We might have learned new bounds from the var_off. */
3600 __update_reg_bounds(src_reg);
3601 __update_reg_bounds(dst_reg);
3602 /* We might have learned something about the sign bit. */
3603 __reg_deduce_bounds(src_reg);
3604 __reg_deduce_bounds(dst_reg);
3605 /* We might have learned some bits from the bounds. */
3606 __reg_bound_offset(src_reg);
3607 __reg_bound_offset(dst_reg);
3608 /* Intersecting with the old var_off might have improved our bounds
3609 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3610 * then new var_off is (0; 0x7f...fc) which improves our umax.
3612 __update_reg_bounds(src_reg);
3613 __update_reg_bounds(dst_reg);
3616 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3617 struct bpf_reg_state *true_dst,
3618 struct bpf_reg_state *false_src,
3619 struct bpf_reg_state *false_dst,
3624 __reg_combine_min_max(true_src, true_dst);
3627 __reg_combine_min_max(false_src, false_dst);
3632 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
3635 struct bpf_reg_state *reg = ®s[regno];
3637 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
3638 /* Old offset (both fixed and variable parts) should
3639 * have been known-zero, because we don't allow pointer
3640 * arithmetic on pointers that might be NULL.
3642 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3643 !tnum_equals_const(reg->var_off, 0) ||
3645 __mark_reg_known_zero(reg);
3649 reg->type = SCALAR_VALUE;
3650 } else if (reg->map_ptr->inner_map_meta) {
3651 reg->type = CONST_PTR_TO_MAP;
3652 reg->map_ptr = reg->map_ptr->inner_map_meta;
3654 reg->type = PTR_TO_MAP_VALUE;
3656 /* We don't need id from this point onwards anymore, thus we
3657 * should better reset it, so that state pruning has chances
3664 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3665 * be folded together at some point.
3667 static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno,
3670 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3671 struct bpf_reg_state *regs = state->regs;
3672 u32 id = regs[regno].id;
3675 for (i = 0; i < MAX_BPF_REG; i++)
3676 mark_map_reg(regs, i, id, is_null);
3678 for (j = 0; j <= vstate->curframe; j++) {
3679 state = vstate->frame[j];
3680 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3681 if (state->stack[i].slot_type[0] != STACK_SPILL)
3683 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
3688 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3689 struct bpf_reg_state *dst_reg,
3690 struct bpf_reg_state *src_reg,
3691 struct bpf_verifier_state *this_branch,
3692 struct bpf_verifier_state *other_branch)
3694 if (BPF_SRC(insn->code) != BPF_X)
3697 switch (BPF_OP(insn->code)) {
3699 if ((dst_reg->type == PTR_TO_PACKET &&
3700 src_reg->type == PTR_TO_PACKET_END) ||
3701 (dst_reg->type == PTR_TO_PACKET_META &&
3702 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3703 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3704 find_good_pkt_pointers(this_branch, dst_reg,
3705 dst_reg->type, false);
3706 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3707 src_reg->type == PTR_TO_PACKET) ||
3708 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3709 src_reg->type == PTR_TO_PACKET_META)) {
3710 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3711 find_good_pkt_pointers(other_branch, src_reg,
3712 src_reg->type, true);
3718 if ((dst_reg->type == PTR_TO_PACKET &&
3719 src_reg->type == PTR_TO_PACKET_END) ||
3720 (dst_reg->type == PTR_TO_PACKET_META &&
3721 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3722 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3723 find_good_pkt_pointers(other_branch, dst_reg,
3724 dst_reg->type, true);
3725 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3726 src_reg->type == PTR_TO_PACKET) ||
3727 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3728 src_reg->type == PTR_TO_PACKET_META)) {
3729 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3730 find_good_pkt_pointers(this_branch, src_reg,
3731 src_reg->type, false);
3737 if ((dst_reg->type == PTR_TO_PACKET &&
3738 src_reg->type == PTR_TO_PACKET_END) ||
3739 (dst_reg->type == PTR_TO_PACKET_META &&
3740 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3741 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3742 find_good_pkt_pointers(this_branch, dst_reg,
3743 dst_reg->type, true);
3744 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3745 src_reg->type == PTR_TO_PACKET) ||
3746 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3747 src_reg->type == PTR_TO_PACKET_META)) {
3748 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3749 find_good_pkt_pointers(other_branch, src_reg,
3750 src_reg->type, false);
3756 if ((dst_reg->type == PTR_TO_PACKET &&
3757 src_reg->type == PTR_TO_PACKET_END) ||
3758 (dst_reg->type == PTR_TO_PACKET_META &&
3759 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3760 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3761 find_good_pkt_pointers(other_branch, dst_reg,
3762 dst_reg->type, false);
3763 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3764 src_reg->type == PTR_TO_PACKET) ||
3765 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3766 src_reg->type == PTR_TO_PACKET_META)) {
3767 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3768 find_good_pkt_pointers(this_branch, src_reg,
3769 src_reg->type, true);
3781 static int check_cond_jmp_op(struct bpf_verifier_env *env,
3782 struct bpf_insn *insn, int *insn_idx)
3784 struct bpf_verifier_state *this_branch = env->cur_state;
3785 struct bpf_verifier_state *other_branch;
3786 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
3787 struct bpf_reg_state *dst_reg, *other_branch_regs;
3788 u8 opcode = BPF_OP(insn->code);
3791 if (opcode > BPF_JSLE) {
3792 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
3796 if (BPF_SRC(insn->code) == BPF_X) {
3797 if (insn->imm != 0) {
3798 verbose(env, "BPF_JMP uses reserved fields\n");
3802 /* check src1 operand */
3803 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3807 if (is_pointer_value(env, insn->src_reg)) {
3808 verbose(env, "R%d pointer comparison prohibited\n",
3813 if (insn->src_reg != BPF_REG_0) {
3814 verbose(env, "BPF_JMP uses reserved fields\n");
3819 /* check src2 operand */
3820 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3824 dst_reg = ®s[insn->dst_reg];
3826 /* detect if R == 0 where R was initialized to zero earlier */
3827 if (BPF_SRC(insn->code) == BPF_K &&
3828 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3829 dst_reg->type == SCALAR_VALUE &&
3830 tnum_is_const(dst_reg->var_off)) {
3831 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
3832 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
3833 /* if (imm == imm) goto pc+off;
3834 * only follow the goto, ignore fall-through
3836 *insn_idx += insn->off;
3839 /* if (imm != imm) goto pc+off;
3840 * only follow fall-through branch, since
3841 * that's where the program will go
3847 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3850 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
3852 /* detect if we are comparing against a constant value so we can adjust
3853 * our min/max values for our dst register.
3854 * this is only legit if both are scalars (or pointers to the same
3855 * object, I suppose, but we don't support that right now), because
3856 * otherwise the different base pointers mean the offsets aren't
3859 if (BPF_SRC(insn->code) == BPF_X) {
3860 if (dst_reg->type == SCALAR_VALUE &&
3861 regs[insn->src_reg].type == SCALAR_VALUE) {
3862 if (tnum_is_const(regs[insn->src_reg].var_off))
3863 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3864 dst_reg, regs[insn->src_reg].var_off.value,
3866 else if (tnum_is_const(dst_reg->var_off))
3867 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
3868 ®s[insn->src_reg],
3869 dst_reg->var_off.value, opcode);
3870 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3871 /* Comparing for equality, we can combine knowledge */
3872 reg_combine_min_max(&other_branch_regs[insn->src_reg],
3873 &other_branch_regs[insn->dst_reg],
3874 ®s[insn->src_reg],
3875 ®s[insn->dst_reg], opcode);
3877 } else if (dst_reg->type == SCALAR_VALUE) {
3878 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3879 dst_reg, insn->imm, opcode);
3882 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3883 if (BPF_SRC(insn->code) == BPF_K &&
3884 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3885 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3886 /* Mark all identical map registers in each branch as either
3887 * safe or unknown depending R == 0 or R != 0 conditional.
3889 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3890 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3891 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3892 this_branch, other_branch) &&
3893 is_pointer_value(env, insn->dst_reg)) {
3894 verbose(env, "R%d pointer comparison prohibited\n",
3899 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
3903 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3904 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3906 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3908 return (struct bpf_map *) (unsigned long) imm64;
3911 /* verify BPF_LD_IMM64 instruction */
3912 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3914 struct bpf_reg_state *regs = cur_regs(env);
3917 if (BPF_SIZE(insn->code) != BPF_DW) {
3918 verbose(env, "invalid BPF_LD_IMM insn\n");
3921 if (insn->off != 0) {
3922 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3926 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3930 if (insn->src_reg == 0) {
3931 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3933 regs[insn->dst_reg].type = SCALAR_VALUE;
3934 __mark_reg_known(®s[insn->dst_reg], imm);
3938 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3939 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3941 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3942 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3946 static bool may_access_skb(enum bpf_prog_type type)
3949 case BPF_PROG_TYPE_SOCKET_FILTER:
3950 case BPF_PROG_TYPE_SCHED_CLS:
3951 case BPF_PROG_TYPE_SCHED_ACT:
3958 /* verify safety of LD_ABS|LD_IND instructions:
3959 * - they can only appear in the programs where ctx == skb
3960 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3961 * preserve R6-R9, and store return value into R0
3964 * ctx == skb == R6 == CTX
3967 * SRC == any register
3968 * IMM == 32-bit immediate
3971 * R0 - 8/16/32-bit skb data converted to cpu endianness
3973 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3975 struct bpf_reg_state *regs = cur_regs(env);
3976 u8 mode = BPF_MODE(insn->code);
3979 if (!may_access_skb(env->prog->type)) {
3980 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3984 if (!env->ops->gen_ld_abs) {
3985 verbose(env, "bpf verifier is misconfigured\n");
3989 if (env->subprog_cnt > 1) {
3990 /* when program has LD_ABS insn JITs and interpreter assume
3991 * that r1 == ctx == skb which is not the case for callees
3992 * that can have arbitrary arguments. It's problematic
3993 * for main prog as well since JITs would need to analyze
3994 * all functions in order to make proper register save/restore
3995 * decisions in the main prog. Hence disallow LD_ABS with calls
3997 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4001 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4002 BPF_SIZE(insn->code) == BPF_DW ||
4003 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4004 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4008 /* check whether implicit source operand (register R6) is readable */
4009 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4013 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4015 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4019 if (mode == BPF_IND) {
4020 /* check explicit source operand */
4021 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4026 /* reset caller saved regs to unreadable */
4027 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4028 mark_reg_not_init(env, regs, caller_saved[i]);
4029 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4032 /* mark destination R0 register as readable, since it contains
4033 * the value fetched from the packet.
4034 * Already marked as written above.
4036 mark_reg_unknown(env, regs, BPF_REG_0);
4040 static int check_return_code(struct bpf_verifier_env *env)
4042 struct bpf_reg_state *reg;
4043 struct tnum range = tnum_range(0, 1);
4045 switch (env->prog->type) {
4046 case BPF_PROG_TYPE_CGROUP_SKB:
4047 case BPF_PROG_TYPE_CGROUP_SOCK:
4048 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4049 case BPF_PROG_TYPE_SOCK_OPS:
4050 case BPF_PROG_TYPE_CGROUP_DEVICE:
4056 reg = cur_regs(env) + BPF_REG_0;
4057 if (reg->type != SCALAR_VALUE) {
4058 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4059 reg_type_str[reg->type]);
4063 if (!tnum_in(range, reg->var_off)) {
4064 verbose(env, "At program exit the register R0 ");
4065 if (!tnum_is_unknown(reg->var_off)) {
4068 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4069 verbose(env, "has value %s", tn_buf);
4071 verbose(env, "has unknown scalar value");
4073 verbose(env, " should have been 0 or 1\n");
4079 /* non-recursive DFS pseudo code
4080 * 1 procedure DFS-iterative(G,v):
4081 * 2 label v as discovered
4082 * 3 let S be a stack
4084 * 5 while S is not empty
4086 * 7 if t is what we're looking for:
4088 * 9 for all edges e in G.adjacentEdges(t) do
4089 * 10 if edge e is already labelled
4090 * 11 continue with the next edge
4091 * 12 w <- G.adjacentVertex(t,e)
4092 * 13 if vertex w is not discovered and not explored
4093 * 14 label e as tree-edge
4094 * 15 label w as discovered
4097 * 18 else if vertex w is discovered
4098 * 19 label e as back-edge
4100 * 21 // vertex w is explored
4101 * 22 label e as forward- or cross-edge
4102 * 23 label t as explored
4107 * 0x11 - discovered and fall-through edge labelled
4108 * 0x12 - discovered and fall-through and branch edges labelled
4119 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4121 static int *insn_stack; /* stack of insns to process */
4122 static int cur_stack; /* current stack index */
4123 static int *insn_state;
4125 /* t, w, e - match pseudo-code above:
4126 * t - index of current instruction
4127 * w - next instruction
4130 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4132 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4135 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4138 if (w < 0 || w >= env->prog->len) {
4139 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4144 /* mark branch target for state pruning */
4145 env->explored_states[w] = STATE_LIST_MARK;
4147 if (insn_state[w] == 0) {
4149 insn_state[t] = DISCOVERED | e;
4150 insn_state[w] = DISCOVERED;
4151 if (cur_stack >= env->prog->len)
4153 insn_stack[cur_stack++] = w;
4155 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4156 verbose(env, "back-edge from insn %d to %d\n", t, w);
4158 } else if (insn_state[w] == EXPLORED) {
4159 /* forward- or cross-edge */
4160 insn_state[t] = DISCOVERED | e;
4162 verbose(env, "insn state internal bug\n");
4168 /* non-recursive depth-first-search to detect loops in BPF program
4169 * loop == back-edge in directed graph
4171 static int check_cfg(struct bpf_verifier_env *env)
4173 struct bpf_insn *insns = env->prog->insnsi;
4174 int insn_cnt = env->prog->len;
4178 ret = check_subprogs(env);
4182 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4186 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4192 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4193 insn_stack[0] = 0; /* 0 is the first instruction */
4199 t = insn_stack[cur_stack - 1];
4201 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4202 u8 opcode = BPF_OP(insns[t].code);
4204 if (opcode == BPF_EXIT) {
4206 } else if (opcode == BPF_CALL) {
4207 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4212 if (t + 1 < insn_cnt)
4213 env->explored_states[t + 1] = STATE_LIST_MARK;
4214 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4215 env->explored_states[t] = STATE_LIST_MARK;
4216 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4222 } else if (opcode == BPF_JA) {
4223 if (BPF_SRC(insns[t].code) != BPF_K) {
4227 /* unconditional jump with single edge */
4228 ret = push_insn(t, t + insns[t].off + 1,
4234 /* tell verifier to check for equivalent states
4235 * after every call and jump
4237 if (t + 1 < insn_cnt)
4238 env->explored_states[t + 1] = STATE_LIST_MARK;
4240 /* conditional jump with two edges */
4241 env->explored_states[t] = STATE_LIST_MARK;
4242 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4248 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4255 /* all other non-branch instructions with single
4258 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4266 insn_state[t] = EXPLORED;
4267 if (cur_stack-- <= 0) {
4268 verbose(env, "pop stack internal bug\n");
4275 for (i = 0; i < insn_cnt; i++) {
4276 if (insn_state[i] != EXPLORED) {
4277 verbose(env, "unreachable insn %d\n", i);
4282 ret = 0; /* cfg looks good */
4290 /* check %cur's range satisfies %old's */
4291 static bool range_within(struct bpf_reg_state *old,
4292 struct bpf_reg_state *cur)
4294 return old->umin_value <= cur->umin_value &&
4295 old->umax_value >= cur->umax_value &&
4296 old->smin_value <= cur->smin_value &&
4297 old->smax_value >= cur->smax_value;
4300 /* Maximum number of register states that can exist at once */
4301 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4307 /* If in the old state two registers had the same id, then they need to have
4308 * the same id in the new state as well. But that id could be different from
4309 * the old state, so we need to track the mapping from old to new ids.
4310 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4311 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4312 * regs with a different old id could still have new id 9, we don't care about
4314 * So we look through our idmap to see if this old id has been seen before. If
4315 * so, we require the new id to match; otherwise, we add the id pair to the map.
4317 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4321 for (i = 0; i < ID_MAP_SIZE; i++) {
4322 if (!idmap[i].old) {
4323 /* Reached an empty slot; haven't seen this id before */
4324 idmap[i].old = old_id;
4325 idmap[i].cur = cur_id;
4328 if (idmap[i].old == old_id)
4329 return idmap[i].cur == cur_id;
4331 /* We ran out of idmap slots, which should be impossible */
4336 /* Returns true if (rold safe implies rcur safe) */
4337 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4338 struct idpair *idmap)
4342 if (!(rold->live & REG_LIVE_READ))
4343 /* explored state didn't use this */
4346 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, frameno)) == 0;
4348 if (rold->type == PTR_TO_STACK)
4349 /* two stack pointers are equal only if they're pointing to
4350 * the same stack frame, since fp-8 in foo != fp-8 in bar
4352 return equal && rold->frameno == rcur->frameno;
4357 if (rold->type == NOT_INIT)
4358 /* explored state can't have used this */
4360 if (rcur->type == NOT_INIT)
4362 switch (rold->type) {
4364 if (rcur->type == SCALAR_VALUE) {
4365 /* new val must satisfy old val knowledge */
4366 return range_within(rold, rcur) &&
4367 tnum_in(rold->var_off, rcur->var_off);
4369 /* We're trying to use a pointer in place of a scalar.
4370 * Even if the scalar was unbounded, this could lead to
4371 * pointer leaks because scalars are allowed to leak
4372 * while pointers are not. We could make this safe in
4373 * special cases if root is calling us, but it's
4374 * probably not worth the hassle.
4378 case PTR_TO_MAP_VALUE:
4379 /* If the new min/max/var_off satisfy the old ones and
4380 * everything else matches, we are OK.
4381 * We don't care about the 'id' value, because nothing
4382 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4384 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4385 range_within(rold, rcur) &&
4386 tnum_in(rold->var_off, rcur->var_off);
4387 case PTR_TO_MAP_VALUE_OR_NULL:
4388 /* a PTR_TO_MAP_VALUE could be safe to use as a
4389 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4390 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4391 * checked, doing so could have affected others with the same
4392 * id, and we can't check for that because we lost the id when
4393 * we converted to a PTR_TO_MAP_VALUE.
4395 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4397 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4399 /* Check our ids match any regs they're supposed to */
4400 return check_ids(rold->id, rcur->id, idmap);
4401 case PTR_TO_PACKET_META:
4403 if (rcur->type != rold->type)
4405 /* We must have at least as much range as the old ptr
4406 * did, so that any accesses which were safe before are
4407 * still safe. This is true even if old range < old off,
4408 * since someone could have accessed through (ptr - k), or
4409 * even done ptr -= k in a register, to get a safe access.
4411 if (rold->range > rcur->range)
4413 /* If the offsets don't match, we can't trust our alignment;
4414 * nor can we be sure that we won't fall out of range.
4416 if (rold->off != rcur->off)
4418 /* id relations must be preserved */
4419 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4421 /* new val must satisfy old val knowledge */
4422 return range_within(rold, rcur) &&
4423 tnum_in(rold->var_off, rcur->var_off);
4425 case CONST_PTR_TO_MAP:
4426 case PTR_TO_PACKET_END:
4427 /* Only valid matches are exact, which memcmp() above
4428 * would have accepted
4431 /* Don't know what's going on, just say it's not safe */
4435 /* Shouldn't get here; if we do, say it's not safe */
4440 static bool stacksafe(struct bpf_func_state *old,
4441 struct bpf_func_state *cur,
4442 struct idpair *idmap)
4446 /* if explored stack has more populated slots than current stack
4447 * such stacks are not equivalent
4449 if (old->allocated_stack > cur->allocated_stack)
4452 /* walk slots of the explored stack and ignore any additional
4453 * slots in the current stack, since explored(safe) state
4456 for (i = 0; i < old->allocated_stack; i++) {
4457 spi = i / BPF_REG_SIZE;
4459 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4460 /* explored state didn't use this */
4463 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4465 /* if old state was safe with misc data in the stack
4466 * it will be safe with zero-initialized stack.
4467 * The opposite is not true
4469 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4470 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4472 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4473 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4474 /* Ex: old explored (safe) state has STACK_SPILL in
4475 * this stack slot, but current has has STACK_MISC ->
4476 * this verifier states are not equivalent,
4477 * return false to continue verification of this path
4480 if (i % BPF_REG_SIZE)
4482 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4484 if (!regsafe(&old->stack[spi].spilled_ptr,
4485 &cur->stack[spi].spilled_ptr,
4487 /* when explored and current stack slot are both storing
4488 * spilled registers, check that stored pointers types
4489 * are the same as well.
4490 * Ex: explored safe path could have stored
4491 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4492 * but current path has stored:
4493 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4494 * such verifier states are not equivalent.
4495 * return false to continue verification of this path
4502 /* compare two verifier states
4504 * all states stored in state_list are known to be valid, since
4505 * verifier reached 'bpf_exit' instruction through them
4507 * this function is called when verifier exploring different branches of
4508 * execution popped from the state stack. If it sees an old state that has
4509 * more strict register state and more strict stack state then this execution
4510 * branch doesn't need to be explored further, since verifier already
4511 * concluded that more strict state leads to valid finish.
4513 * Therefore two states are equivalent if register state is more conservative
4514 * and explored stack state is more conservative than the current one.
4517 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4518 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4520 * In other words if current stack state (one being explored) has more
4521 * valid slots than old one that already passed validation, it means
4522 * the verifier can stop exploring and conclude that current state is valid too
4524 * Similarly with registers. If explored state has register type as invalid
4525 * whereas register type in current state is meaningful, it means that
4526 * the current state will reach 'bpf_exit' instruction safely
4528 static bool func_states_equal(struct bpf_func_state *old,
4529 struct bpf_func_state *cur)
4531 struct idpair *idmap;
4535 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4536 /* If we failed to allocate the idmap, just say it's not safe */
4540 for (i = 0; i < MAX_BPF_REG; i++) {
4541 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4545 if (!stacksafe(old, cur, idmap))
4553 static bool states_equal(struct bpf_verifier_env *env,
4554 struct bpf_verifier_state *old,
4555 struct bpf_verifier_state *cur)
4559 if (old->curframe != cur->curframe)
4562 /* for states to be equal callsites have to be the same
4563 * and all frame states need to be equivalent
4565 for (i = 0; i <= old->curframe; i++) {
4566 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4568 if (!func_states_equal(old->frame[i], cur->frame[i]))
4574 /* A write screens off any subsequent reads; but write marks come from the
4575 * straight-line code between a state and its parent. When we arrive at an
4576 * equivalent state (jump target or such) we didn't arrive by the straight-line
4577 * code, so read marks in the state must propagate to the parent regardless
4578 * of the state's write marks. That's what 'parent == state->parent' comparison
4579 * in mark_reg_read() and mark_stack_slot_read() is for.
4581 static int propagate_liveness(struct bpf_verifier_env *env,
4582 const struct bpf_verifier_state *vstate,
4583 struct bpf_verifier_state *vparent)
4585 int i, frame, err = 0;
4586 struct bpf_func_state *state, *parent;
4588 if (vparent->curframe != vstate->curframe) {
4589 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4590 vparent->curframe, vstate->curframe);
4593 /* Propagate read liveness of registers... */
4594 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4595 /* We don't need to worry about FP liveness because it's read-only */
4596 for (i = 0; i < BPF_REG_FP; i++) {
4597 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4599 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4600 err = mark_reg_read(env, vstate, vparent, i);
4606 /* ... and stack slots */
4607 for (frame = 0; frame <= vstate->curframe; frame++) {
4608 state = vstate->frame[frame];
4609 parent = vparent->frame[frame];
4610 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4611 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4612 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4614 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4615 mark_stack_slot_read(env, vstate, vparent, i, frame);
4621 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4623 struct bpf_verifier_state_list *new_sl;
4624 struct bpf_verifier_state_list *sl;
4625 struct bpf_verifier_state *cur = env->cur_state;
4628 sl = env->explored_states[insn_idx];
4630 /* this 'insn_idx' instruction wasn't marked, so we will not
4631 * be doing state search here
4635 while (sl != STATE_LIST_MARK) {
4636 if (states_equal(env, &sl->state, cur)) {
4637 /* reached equivalent register/stack state,
4639 * Registers read by the continuation are read by us.
4640 * If we have any write marks in env->cur_state, they
4641 * will prevent corresponding reads in the continuation
4642 * from reaching our parent (an explored_state). Our
4643 * own state will get the read marks recorded, but
4644 * they'll be immediately forgotten as we're pruning
4645 * this state and will pop a new one.
4647 err = propagate_liveness(env, &sl->state, cur);
4655 /* there were no equivalent states, remember current one.
4656 * technically the current state is not proven to be safe yet,
4657 * but it will either reach outer most bpf_exit (which means it's safe)
4658 * or it will be rejected. Since there are no loops, we won't be
4659 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4660 * again on the way to bpf_exit
4662 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
4666 /* add new state to the head of linked list */
4667 err = copy_verifier_state(&new_sl->state, cur);
4669 free_verifier_state(&new_sl->state, false);
4673 new_sl->next = env->explored_states[insn_idx];
4674 env->explored_states[insn_idx] = new_sl;
4675 /* connect new state to parentage chain */
4676 cur->parent = &new_sl->state;
4677 /* clear write marks in current state: the writes we did are not writes
4678 * our child did, so they don't screen off its reads from us.
4679 * (There are no read marks in current state, because reads always mark
4680 * their parent and current state never has children yet. Only
4681 * explored_states can get read marks.)
4683 for (i = 0; i < BPF_REG_FP; i++)
4684 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
4686 /* all stack frames are accessible from callee, clear them all */
4687 for (j = 0; j <= cur->curframe; j++) {
4688 struct bpf_func_state *frame = cur->frame[j];
4690 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++)
4691 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
4696 static int do_check(struct bpf_verifier_env *env)
4698 struct bpf_verifier_state *state;
4699 struct bpf_insn *insns = env->prog->insnsi;
4700 struct bpf_reg_state *regs;
4701 int insn_cnt = env->prog->len, i;
4702 int insn_idx, prev_insn_idx = 0;
4703 int insn_processed = 0;
4704 bool do_print_state = false;
4706 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
4709 state->curframe = 0;
4710 state->parent = NULL;
4711 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
4712 if (!state->frame[0]) {
4716 env->cur_state = state;
4717 init_func_state(env, state->frame[0],
4718 BPF_MAIN_FUNC /* callsite */,
4720 0 /* subprogno, zero == main subprog */);
4723 struct bpf_insn *insn;
4727 if (insn_idx >= insn_cnt) {
4728 verbose(env, "invalid insn idx %d insn_cnt %d\n",
4729 insn_idx, insn_cnt);
4733 insn = &insns[insn_idx];
4734 class = BPF_CLASS(insn->code);
4736 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
4738 "BPF program is too large. Processed %d insn\n",
4743 err = is_state_visited(env, insn_idx);
4747 /* found equivalent state, can prune the search */
4748 if (env->log.level) {
4750 verbose(env, "\nfrom %d to %d: safe\n",
4751 prev_insn_idx, insn_idx);
4753 verbose(env, "%d: safe\n", insn_idx);
4755 goto process_bpf_exit;
4761 if (env->log.level > 1 || (env->log.level && do_print_state)) {
4762 if (env->log.level > 1)
4763 verbose(env, "%d:", insn_idx);
4765 verbose(env, "\nfrom %d to %d:",
4766 prev_insn_idx, insn_idx);
4767 print_verifier_state(env, state->frame[state->curframe]);
4768 do_print_state = false;
4771 if (env->log.level) {
4772 const struct bpf_insn_cbs cbs = {
4773 .cb_print = verbose,
4774 .private_data = env,
4777 verbose(env, "%d: ", insn_idx);
4778 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4781 if (bpf_prog_is_dev_bound(env->prog->aux)) {
4782 err = bpf_prog_offload_verify_insn(env, insn_idx,
4788 regs = cur_regs(env);
4789 env->insn_aux_data[insn_idx].seen = true;
4790 if (class == BPF_ALU || class == BPF_ALU64) {
4791 err = check_alu_op(env, insn);
4795 } else if (class == BPF_LDX) {
4796 enum bpf_reg_type *prev_src_type, src_reg_type;
4798 /* check for reserved fields is already done */
4800 /* check src operand */
4801 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4805 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4809 src_reg_type = regs[insn->src_reg].type;
4811 /* check that memory (src_reg + off) is readable,
4812 * the state of dst_reg will be updated by this func
4814 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
4815 BPF_SIZE(insn->code), BPF_READ,
4816 insn->dst_reg, false);
4820 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
4822 if (*prev_src_type == NOT_INIT) {
4824 * dst_reg = *(u32 *)(src_reg + off)
4825 * save type to validate intersecting paths
4827 *prev_src_type = src_reg_type;
4829 } else if (src_reg_type != *prev_src_type &&
4830 (src_reg_type == PTR_TO_CTX ||
4831 *prev_src_type == PTR_TO_CTX)) {
4832 /* ABuser program is trying to use the same insn
4833 * dst_reg = *(u32*) (src_reg + off)
4834 * with different pointer types:
4835 * src_reg == ctx in one branch and
4836 * src_reg == stack|map in some other branch.
4839 verbose(env, "same insn cannot be used with different pointers\n");
4843 } else if (class == BPF_STX) {
4844 enum bpf_reg_type *prev_dst_type, dst_reg_type;
4846 if (BPF_MODE(insn->code) == BPF_XADD) {
4847 err = check_xadd(env, insn_idx, insn);
4854 /* check src1 operand */
4855 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4858 /* check src2 operand */
4859 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4863 dst_reg_type = regs[insn->dst_reg].type;
4865 /* check that memory (dst_reg + off) is writeable */
4866 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4867 BPF_SIZE(insn->code), BPF_WRITE,
4868 insn->src_reg, false);
4872 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
4874 if (*prev_dst_type == NOT_INIT) {
4875 *prev_dst_type = dst_reg_type;
4876 } else if (dst_reg_type != *prev_dst_type &&
4877 (dst_reg_type == PTR_TO_CTX ||
4878 *prev_dst_type == PTR_TO_CTX)) {
4879 verbose(env, "same insn cannot be used with different pointers\n");
4883 } else if (class == BPF_ST) {
4884 if (BPF_MODE(insn->code) != BPF_MEM ||
4885 insn->src_reg != BPF_REG_0) {
4886 verbose(env, "BPF_ST uses reserved fields\n");
4889 /* check src operand */
4890 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4894 if (is_ctx_reg(env, insn->dst_reg)) {
4895 verbose(env, "BPF_ST stores into R%d context is not allowed\n",
4900 /* check that memory (dst_reg + off) is writeable */
4901 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4902 BPF_SIZE(insn->code), BPF_WRITE,
4907 } else if (class == BPF_JMP) {
4908 u8 opcode = BPF_OP(insn->code);
4910 if (opcode == BPF_CALL) {
4911 if (BPF_SRC(insn->code) != BPF_K ||
4913 (insn->src_reg != BPF_REG_0 &&
4914 insn->src_reg != BPF_PSEUDO_CALL) ||
4915 insn->dst_reg != BPF_REG_0) {
4916 verbose(env, "BPF_CALL uses reserved fields\n");
4920 if (insn->src_reg == BPF_PSEUDO_CALL)
4921 err = check_func_call(env, insn, &insn_idx);
4923 err = check_helper_call(env, insn->imm, insn_idx);
4927 } else if (opcode == BPF_JA) {
4928 if (BPF_SRC(insn->code) != BPF_K ||
4930 insn->src_reg != BPF_REG_0 ||
4931 insn->dst_reg != BPF_REG_0) {
4932 verbose(env, "BPF_JA uses reserved fields\n");
4936 insn_idx += insn->off + 1;
4939 } else if (opcode == BPF_EXIT) {
4940 if (BPF_SRC(insn->code) != BPF_K ||
4942 insn->src_reg != BPF_REG_0 ||
4943 insn->dst_reg != BPF_REG_0) {
4944 verbose(env, "BPF_EXIT uses reserved fields\n");
4948 if (state->curframe) {
4949 /* exit from nested function */
4950 prev_insn_idx = insn_idx;
4951 err = prepare_func_exit(env, &insn_idx);
4954 do_print_state = true;
4958 /* eBPF calling convetion is such that R0 is used
4959 * to return the value from eBPF program.
4960 * Make sure that it's readable at this time
4961 * of bpf_exit, which means that program wrote
4962 * something into it earlier
4964 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4968 if (is_pointer_value(env, BPF_REG_0)) {
4969 verbose(env, "R0 leaks addr as return value\n");
4973 err = check_return_code(env);
4977 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4983 do_print_state = true;
4987 err = check_cond_jmp_op(env, insn, &insn_idx);
4991 } else if (class == BPF_LD) {
4992 u8 mode = BPF_MODE(insn->code);
4994 if (mode == BPF_ABS || mode == BPF_IND) {
4995 err = check_ld_abs(env, insn);
4999 } else if (mode == BPF_IMM) {
5000 err = check_ld_imm(env, insn);
5005 env->insn_aux_data[insn_idx].seen = true;
5007 verbose(env, "invalid BPF_LD mode\n");
5011 verbose(env, "unknown insn class %d\n", class);
5018 verbose(env, "processed %d insns (limit %d), stack depth ",
5019 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5020 for (i = 0; i < env->subprog_cnt; i++) {
5021 u32 depth = env->subprog_info[i].stack_depth;
5023 verbose(env, "%d", depth);
5024 if (i + 1 < env->subprog_cnt)
5028 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
5032 static int check_map_prealloc(struct bpf_map *map)
5034 return (map->map_type != BPF_MAP_TYPE_HASH &&
5035 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
5036 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
5037 !(map->map_flags & BPF_F_NO_PREALLOC);
5040 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
5041 struct bpf_map *map,
5042 struct bpf_prog *prog)
5045 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5046 * preallocated hash maps, since doing memory allocation
5047 * in overflow_handler can crash depending on where nmi got
5050 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
5051 if (!check_map_prealloc(map)) {
5052 verbose(env, "perf_event programs can only use preallocated hash map\n");
5055 if (map->inner_map_meta &&
5056 !check_map_prealloc(map->inner_map_meta)) {
5057 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
5062 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
5063 !bpf_offload_prog_map_match(prog, map)) {
5064 verbose(env, "offload device mismatch between prog and map\n");
5071 /* look for pseudo eBPF instructions that access map FDs and
5072 * replace them with actual map pointers
5074 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
5076 struct bpf_insn *insn = env->prog->insnsi;
5077 int insn_cnt = env->prog->len;
5080 err = bpf_prog_calc_tag(env->prog);
5084 for (i = 0; i < insn_cnt; i++, insn++) {
5085 if (BPF_CLASS(insn->code) == BPF_LDX &&
5086 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
5087 verbose(env, "BPF_LDX uses reserved fields\n");
5091 if (BPF_CLASS(insn->code) == BPF_STX &&
5092 ((BPF_MODE(insn->code) != BPF_MEM &&
5093 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
5094 verbose(env, "BPF_STX uses reserved fields\n");
5098 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5099 struct bpf_map *map;
5102 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5103 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5105 verbose(env, "invalid bpf_ld_imm64 insn\n");
5109 if (insn->src_reg == 0)
5110 /* valid generic load 64-bit imm */
5113 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5115 "unrecognized bpf_ld_imm64 insn\n");
5119 f = fdget(insn->imm);
5120 map = __bpf_map_get(f);
5122 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5124 return PTR_ERR(map);
5127 err = check_map_prog_compatibility(env, map, env->prog);
5133 /* store map pointer inside BPF_LD_IMM64 instruction */
5134 insn[0].imm = (u32) (unsigned long) map;
5135 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5137 /* check whether we recorded this map already */
5138 for (j = 0; j < env->used_map_cnt; j++)
5139 if (env->used_maps[j] == map) {
5144 if (env->used_map_cnt >= MAX_USED_MAPS) {
5149 /* hold the map. If the program is rejected by verifier,
5150 * the map will be released by release_maps() or it
5151 * will be used by the valid program until it's unloaded
5152 * and all maps are released in free_used_maps()
5154 map = bpf_map_inc(map, false);
5157 return PTR_ERR(map);
5159 env->used_maps[env->used_map_cnt++] = map;
5161 if (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE &&
5162 bpf_cgroup_storage_assign(env->prog, map)) {
5164 "only one cgroup storage is allowed\n");
5176 /* Basic sanity check before we invest more work here. */
5177 if (!bpf_opcode_in_insntable(insn->code)) {
5178 verbose(env, "unknown opcode %02x\n", insn->code);
5183 /* now all pseudo BPF_LD_IMM64 instructions load valid
5184 * 'struct bpf_map *' into a register instead of user map_fd.
5185 * These pointers will be used later by verifier to validate map access.
5190 /* drop refcnt of maps used by the rejected program */
5191 static void release_maps(struct bpf_verifier_env *env)
5195 if (env->prog->aux->cgroup_storage)
5196 bpf_cgroup_storage_release(env->prog,
5197 env->prog->aux->cgroup_storage);
5199 for (i = 0; i < env->used_map_cnt; i++)
5200 bpf_map_put(env->used_maps[i]);
5203 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5204 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5206 struct bpf_insn *insn = env->prog->insnsi;
5207 int insn_cnt = env->prog->len;
5210 for (i = 0; i < insn_cnt; i++, insn++)
5211 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5215 /* single env->prog->insni[off] instruction was replaced with the range
5216 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5217 * [0, off) and [off, end) to new locations, so the patched range stays zero
5219 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5222 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5227 new_data = vzalloc(array_size(prog_len,
5228 sizeof(struct bpf_insn_aux_data)));
5231 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5232 memcpy(new_data + off + cnt - 1, old_data + off,
5233 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5234 for (i = off; i < off + cnt - 1; i++)
5235 new_data[i].seen = true;
5236 env->insn_aux_data = new_data;
5241 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5247 /* NOTE: fake 'exit' subprog should be updated as well. */
5248 for (i = 0; i <= env->subprog_cnt; i++) {
5249 if (env->subprog_info[i].start < off)
5251 env->subprog_info[i].start += len - 1;
5255 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5256 const struct bpf_insn *patch, u32 len)
5258 struct bpf_prog *new_prog;
5260 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5263 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5265 adjust_subprog_starts(env, off, len);
5269 /* The verifier does more data flow analysis than llvm and will not
5270 * explore branches that are dead at run time. Malicious programs can
5271 * have dead code too. Therefore replace all dead at-run-time code
5274 * Just nops are not optimal, e.g. if they would sit at the end of the
5275 * program and through another bug we would manage to jump there, then
5276 * we'd execute beyond program memory otherwise. Returning exception
5277 * code also wouldn't work since we can have subprogs where the dead
5278 * code could be located.
5280 static void sanitize_dead_code(struct bpf_verifier_env *env)
5282 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5283 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5284 struct bpf_insn *insn = env->prog->insnsi;
5285 const int insn_cnt = env->prog->len;
5288 for (i = 0; i < insn_cnt; i++) {
5289 if (aux_data[i].seen)
5291 memcpy(insn + i, &trap, sizeof(trap));
5295 /* convert load instructions that access fields of 'struct __sk_buff'
5296 * into sequence of instructions that access fields of 'struct sk_buff'
5298 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5300 const struct bpf_verifier_ops *ops = env->ops;
5301 int i, cnt, size, ctx_field_size, delta = 0;
5302 const int insn_cnt = env->prog->len;
5303 struct bpf_insn insn_buf[16], *insn;
5304 struct bpf_prog *new_prog;
5305 enum bpf_access_type type;
5306 bool is_narrower_load;
5309 if (ops->gen_prologue) {
5310 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5312 if (cnt >= ARRAY_SIZE(insn_buf)) {
5313 verbose(env, "bpf verifier is misconfigured\n");
5316 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5320 env->prog = new_prog;
5325 if (!ops->convert_ctx_access || bpf_prog_is_dev_bound(env->prog->aux))
5328 insn = env->prog->insnsi + delta;
5330 for (i = 0; i < insn_cnt; i++, insn++) {
5331 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5332 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5333 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5334 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5336 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5337 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5338 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5339 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5344 if (type == BPF_WRITE &&
5345 env->insn_aux_data[i + delta].sanitize_stack_off) {
5346 struct bpf_insn patch[] = {
5347 /* Sanitize suspicious stack slot with zero.
5348 * There are no memory dependencies for this store,
5349 * since it's only using frame pointer and immediate
5352 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5353 env->insn_aux_data[i + delta].sanitize_stack_off,
5355 /* the original STX instruction will immediately
5356 * overwrite the same stack slot with appropriate value
5361 cnt = ARRAY_SIZE(patch);
5362 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5367 env->prog = new_prog;
5368 insn = new_prog->insnsi + i + delta;
5372 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
5375 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5376 size = BPF_LDST_BYTES(insn);
5378 /* If the read access is a narrower load of the field,
5379 * convert to a 4/8-byte load, to minimum program type specific
5380 * convert_ctx_access changes. If conversion is successful,
5381 * we will apply proper mask to the result.
5383 is_narrower_load = size < ctx_field_size;
5384 if (is_narrower_load) {
5385 u32 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
5386 u32 off = insn->off;
5389 if (type == BPF_WRITE) {
5390 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5395 if (ctx_field_size == 4)
5397 else if (ctx_field_size == 8)
5400 insn->off = off & ~(size_default - 1);
5401 insn->code = BPF_LDX | BPF_MEM | size_code;
5405 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
5407 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5408 (ctx_field_size && !target_size)) {
5409 verbose(env, "bpf verifier is misconfigured\n");
5413 if (is_narrower_load && size < target_size) {
5414 if (ctx_field_size <= 4)
5415 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5416 (1 << size * 8) - 1);
5418 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5419 (1 << size * 8) - 1);
5422 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5428 /* keep walking new program and skip insns we just inserted */
5429 env->prog = new_prog;
5430 insn = new_prog->insnsi + i + delta;
5436 static int jit_subprogs(struct bpf_verifier_env *env)
5438 struct bpf_prog *prog = env->prog, **func, *tmp;
5439 int i, j, subprog_start, subprog_end = 0, len, subprog;
5440 struct bpf_insn *insn;
5444 if (env->subprog_cnt <= 1)
5447 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5448 if (insn->code != (BPF_JMP | BPF_CALL) ||
5449 insn->src_reg != BPF_PSEUDO_CALL)
5451 /* Upon error here we cannot fall back to interpreter but
5452 * need a hard reject of the program. Thus -EFAULT is
5453 * propagated in any case.
5455 subprog = find_subprog(env, i + insn->imm + 1);
5457 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5461 /* temporarily remember subprog id inside insn instead of
5462 * aux_data, since next loop will split up all insns into funcs
5464 insn->off = subprog;
5465 /* remember original imm in case JIT fails and fallback
5466 * to interpreter will be needed
5468 env->insn_aux_data[i].call_imm = insn->imm;
5469 /* point imm to __bpf_call_base+1 from JITs point of view */
5473 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
5477 for (i = 0; i < env->subprog_cnt; i++) {
5478 subprog_start = subprog_end;
5479 subprog_end = env->subprog_info[i + 1].start;
5481 len = subprog_end - subprog_start;
5482 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5485 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5486 len * sizeof(struct bpf_insn));
5487 func[i]->type = prog->type;
5489 if (bpf_prog_calc_tag(func[i]))
5491 func[i]->is_func = 1;
5492 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5493 * Long term would need debug info to populate names
5495 func[i]->aux->name[0] = 'F';
5496 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
5497 func[i]->jit_requested = 1;
5498 func[i] = bpf_int_jit_compile(func[i]);
5499 if (!func[i]->jited) {
5505 /* at this point all bpf functions were successfully JITed
5506 * now populate all bpf_calls with correct addresses and
5507 * run last pass of JIT
5509 for (i = 0; i < env->subprog_cnt; i++) {
5510 insn = func[i]->insnsi;
5511 for (j = 0; j < func[i]->len; j++, insn++) {
5512 if (insn->code != (BPF_JMP | BPF_CALL) ||
5513 insn->src_reg != BPF_PSEUDO_CALL)
5515 subprog = insn->off;
5516 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5517 func[subprog]->bpf_func -
5521 /* we use the aux data to keep a list of the start addresses
5522 * of the JITed images for each function in the program
5524 * for some architectures, such as powerpc64, the imm field
5525 * might not be large enough to hold the offset of the start
5526 * address of the callee's JITed image from __bpf_call_base
5528 * in such cases, we can lookup the start address of a callee
5529 * by using its subprog id, available from the off field of
5530 * the call instruction, as an index for this list
5532 func[i]->aux->func = func;
5533 func[i]->aux->func_cnt = env->subprog_cnt;
5535 for (i = 0; i < env->subprog_cnt; i++) {
5536 old_bpf_func = func[i]->bpf_func;
5537 tmp = bpf_int_jit_compile(func[i]);
5538 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5539 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5546 /* finally lock prog and jit images for all functions and
5549 for (i = 0; i < env->subprog_cnt; i++) {
5550 bpf_prog_lock_ro(func[i]);
5551 bpf_prog_kallsyms_add(func[i]);
5554 /* Last step: make now unused interpreter insns from main
5555 * prog consistent for later dump requests, so they can
5556 * later look the same as if they were interpreted only.
5558 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5559 if (insn->code != (BPF_JMP | BPF_CALL) ||
5560 insn->src_reg != BPF_PSEUDO_CALL)
5562 insn->off = env->insn_aux_data[i].call_imm;
5563 subprog = find_subprog(env, i + insn->off + 1);
5564 insn->imm = subprog;
5568 prog->bpf_func = func[0]->bpf_func;
5569 prog->aux->func = func;
5570 prog->aux->func_cnt = env->subprog_cnt;
5573 for (i = 0; i < env->subprog_cnt; i++)
5575 bpf_jit_free(func[i]);
5578 /* cleanup main prog to be interpreted */
5579 prog->jit_requested = 0;
5580 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5581 if (insn->code != (BPF_JMP | BPF_CALL) ||
5582 insn->src_reg != BPF_PSEUDO_CALL)
5585 insn->imm = env->insn_aux_data[i].call_imm;
5590 static int fixup_call_args(struct bpf_verifier_env *env)
5592 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5593 struct bpf_prog *prog = env->prog;
5594 struct bpf_insn *insn = prog->insnsi;
5600 if (env->prog->jit_requested) {
5601 err = jit_subprogs(env);
5607 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5608 for (i = 0; i < prog->len; i++, insn++) {
5609 if (insn->code != (BPF_JMP | BPF_CALL) ||
5610 insn->src_reg != BPF_PSEUDO_CALL)
5612 depth = get_callee_stack_depth(env, insn, i);
5615 bpf_patch_call_args(insn, depth);
5622 /* fixup insn->imm field of bpf_call instructions
5623 * and inline eligible helpers as explicit sequence of BPF instructions
5625 * this function is called after eBPF program passed verification
5627 static int fixup_bpf_calls(struct bpf_verifier_env *env)
5629 struct bpf_prog *prog = env->prog;
5630 struct bpf_insn *insn = prog->insnsi;
5631 const struct bpf_func_proto *fn;
5632 const int insn_cnt = prog->len;
5633 const struct bpf_map_ops *ops;
5634 struct bpf_insn_aux_data *aux;
5635 struct bpf_insn insn_buf[16];
5636 struct bpf_prog *new_prog;
5637 struct bpf_map *map_ptr;
5638 int i, cnt, delta = 0;
5640 for (i = 0; i < insn_cnt; i++, insn++) {
5641 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
5642 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5643 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
5644 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5645 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
5646 struct bpf_insn mask_and_div[] = {
5647 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5649 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
5650 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
5651 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
5654 struct bpf_insn mask_and_mod[] = {
5655 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5656 /* Rx mod 0 -> Rx */
5657 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
5660 struct bpf_insn *patchlet;
5662 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5663 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5664 patchlet = mask_and_div + (is64 ? 1 : 0);
5665 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
5667 patchlet = mask_and_mod + (is64 ? 1 : 0);
5668 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
5671 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
5676 env->prog = prog = new_prog;
5677 insn = new_prog->insnsi + i + delta;
5681 if (BPF_CLASS(insn->code) == BPF_LD &&
5682 (BPF_MODE(insn->code) == BPF_ABS ||
5683 BPF_MODE(insn->code) == BPF_IND)) {
5684 cnt = env->ops->gen_ld_abs(insn, insn_buf);
5685 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5686 verbose(env, "bpf verifier is misconfigured\n");
5690 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5695 env->prog = prog = new_prog;
5696 insn = new_prog->insnsi + i + delta;
5700 if (insn->code != (BPF_JMP | BPF_CALL))
5702 if (insn->src_reg == BPF_PSEUDO_CALL)
5705 if (insn->imm == BPF_FUNC_get_route_realm)
5706 prog->dst_needed = 1;
5707 if (insn->imm == BPF_FUNC_get_prandom_u32)
5708 bpf_user_rnd_init_once();
5709 if (insn->imm == BPF_FUNC_override_return)
5710 prog->kprobe_override = 1;
5711 if (insn->imm == BPF_FUNC_tail_call) {
5712 /* If we tail call into other programs, we
5713 * cannot make any assumptions since they can
5714 * be replaced dynamically during runtime in
5715 * the program array.
5717 prog->cb_access = 1;
5718 env->prog->aux->stack_depth = MAX_BPF_STACK;
5720 /* mark bpf_tail_call as different opcode to avoid
5721 * conditional branch in the interpeter for every normal
5722 * call and to prevent accidental JITing by JIT compiler
5723 * that doesn't support bpf_tail_call yet
5726 insn->code = BPF_JMP | BPF_TAIL_CALL;
5728 aux = &env->insn_aux_data[i + delta];
5729 if (!bpf_map_ptr_unpriv(aux))
5732 /* instead of changing every JIT dealing with tail_call
5733 * emit two extra insns:
5734 * if (index >= max_entries) goto out;
5735 * index &= array->index_mask;
5736 * to avoid out-of-bounds cpu speculation
5738 if (bpf_map_ptr_poisoned(aux)) {
5739 verbose(env, "tail_call abusing map_ptr\n");
5743 map_ptr = BPF_MAP_PTR(aux->map_state);
5744 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
5745 map_ptr->max_entries, 2);
5746 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
5747 container_of(map_ptr,
5750 insn_buf[2] = *insn;
5752 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5757 env->prog = prog = new_prog;
5758 insn = new_prog->insnsi + i + delta;
5762 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5763 * and other inlining handlers are currently limited to 64 bit
5766 if (prog->jit_requested && BITS_PER_LONG == 64 &&
5767 (insn->imm == BPF_FUNC_map_lookup_elem ||
5768 insn->imm == BPF_FUNC_map_update_elem ||
5769 insn->imm == BPF_FUNC_map_delete_elem)) {
5770 aux = &env->insn_aux_data[i + delta];
5771 if (bpf_map_ptr_poisoned(aux))
5772 goto patch_call_imm;
5774 map_ptr = BPF_MAP_PTR(aux->map_state);
5776 if (insn->imm == BPF_FUNC_map_lookup_elem &&
5777 ops->map_gen_lookup) {
5778 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
5779 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5780 verbose(env, "bpf verifier is misconfigured\n");
5784 new_prog = bpf_patch_insn_data(env, i + delta,
5790 env->prog = prog = new_prog;
5791 insn = new_prog->insnsi + i + delta;
5795 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
5796 (void *(*)(struct bpf_map *map, void *key))NULL));
5797 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
5798 (int (*)(struct bpf_map *map, void *key))NULL));
5799 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
5800 (int (*)(struct bpf_map *map, void *key, void *value,
5802 switch (insn->imm) {
5803 case BPF_FUNC_map_lookup_elem:
5804 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
5807 case BPF_FUNC_map_update_elem:
5808 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
5811 case BPF_FUNC_map_delete_elem:
5812 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
5817 goto patch_call_imm;
5820 if (insn->imm == BPF_FUNC_redirect_map) {
5821 /* Note, we cannot use prog directly as imm as subsequent
5822 * rewrites would still change the prog pointer. The only
5823 * stable address we can use is aux, which also works with
5824 * prog clones during blinding.
5826 u64 addr = (unsigned long)prog->aux;
5827 struct bpf_insn r4_ld[] = {
5828 BPF_LD_IMM64(BPF_REG_4, addr),
5831 cnt = ARRAY_SIZE(r4_ld);
5833 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
5838 env->prog = prog = new_prog;
5839 insn = new_prog->insnsi + i + delta;
5842 fn = env->ops->get_func_proto(insn->imm, env->prog);
5843 /* all functions that have prototype and verifier allowed
5844 * programs to call them, must be real in-kernel functions
5848 "kernel subsystem misconfigured func %s#%d\n",
5849 func_id_name(insn->imm), insn->imm);
5852 insn->imm = fn->func - __bpf_call_base;
5858 static void free_states(struct bpf_verifier_env *env)
5860 struct bpf_verifier_state_list *sl, *sln;
5863 if (!env->explored_states)
5866 for (i = 0; i < env->prog->len; i++) {
5867 sl = env->explored_states[i];
5870 while (sl != STATE_LIST_MARK) {
5872 free_verifier_state(&sl->state, false);
5878 kfree(env->explored_states);
5881 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
5883 struct bpf_verifier_env *env;
5884 struct bpf_verifier_log *log;
5887 /* no program is valid */
5888 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
5891 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5892 * allocate/free it every time bpf_check() is called
5894 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
5899 env->insn_aux_data =
5900 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
5903 if (!env->insn_aux_data)
5906 env->ops = bpf_verifier_ops[env->prog->type];
5908 /* grab the mutex to protect few globals used by verifier */
5909 mutex_lock(&bpf_verifier_lock);
5911 if (attr->log_level || attr->log_buf || attr->log_size) {
5912 /* user requested verbose verifier output
5913 * and supplied buffer to store the verification trace
5915 log->level = attr->log_level;
5916 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
5917 log->len_total = attr->log_size;
5920 /* log attributes have to be sane */
5921 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
5922 !log->level || !log->ubuf)
5926 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
5927 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
5928 env->strict_alignment = true;
5930 ret = replace_map_fd_with_map_ptr(env);
5932 goto skip_full_check;
5934 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5935 ret = bpf_prog_offload_verifier_prep(env);
5937 goto skip_full_check;
5940 env->explored_states = kcalloc(env->prog->len,
5941 sizeof(struct bpf_verifier_state_list *),
5944 if (!env->explored_states)
5945 goto skip_full_check;
5947 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
5949 ret = check_cfg(env);
5951 goto skip_full_check;
5953 ret = do_check(env);
5954 if (env->cur_state) {
5955 free_verifier_state(env->cur_state, true);
5956 env->cur_state = NULL;
5960 while (!pop_stack(env, NULL, NULL));
5964 sanitize_dead_code(env);
5967 ret = check_max_stack_depth(env);
5970 /* program is valid, convert *(u32*)(ctx + off) accesses */
5971 ret = convert_ctx_accesses(env);
5974 ret = fixup_bpf_calls(env);
5977 ret = fixup_call_args(env);
5979 if (log->level && bpf_verifier_log_full(log))
5981 if (log->level && !log->ubuf) {
5983 goto err_release_maps;
5986 if (ret == 0 && env->used_map_cnt) {
5987 /* if program passed verifier, update used_maps in bpf_prog_info */
5988 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
5989 sizeof(env->used_maps[0]),
5992 if (!env->prog->aux->used_maps) {
5994 goto err_release_maps;
5997 memcpy(env->prog->aux->used_maps, env->used_maps,
5998 sizeof(env->used_maps[0]) * env->used_map_cnt);
5999 env->prog->aux->used_map_cnt = env->used_map_cnt;
6001 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
6002 * bpf_ld_imm64 instructions
6004 convert_pseudo_ld_imm64(env);
6008 if (!env->prog->aux->used_maps)
6009 /* if we didn't copy map pointers into bpf_prog_info, release
6010 * them now. Otherwise free_used_maps() will release them.
6015 mutex_unlock(&bpf_verifier_lock);
6016 vfree(env->insn_aux_data);