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>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #include <linux/bpf_types.h>
37 /* bpf_check() is a static code analyzer that walks eBPF program
38 * instruction by instruction and updates register/stack state.
39 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
41 * The first pass is depth-first-search to check that the program is a DAG.
42 * It rejects the following programs:
43 * - larger than BPF_MAXINSNS insns
44 * - if loop is present (detected via back-edge)
45 * - unreachable insns exist (shouldn't be a forest. program = one function)
46 * - out of bounds or malformed jumps
47 * The second pass is all possible path descent from the 1st insn.
48 * Since it's analyzing all pathes through the program, the length of the
49 * analysis is limited to 64k insn, which may be hit even if total number of
50 * insn is less then 4K, but there are too many branches that change stack/regs.
51 * Number of 'branches to be analyzed' is limited to 1k
53 * On entry to each instruction, each register has a type, and the instruction
54 * changes the types of the registers depending on instruction semantics.
55 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
58 * All registers are 64-bit.
59 * R0 - return register
60 * R1-R5 argument passing registers
61 * R6-R9 callee saved registers
62 * R10 - frame pointer read-only
64 * At the start of BPF program the register R1 contains a pointer to bpf_context
65 * and has type PTR_TO_CTX.
67 * Verifier tracks arithmetic operations on pointers in case:
68 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
69 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
70 * 1st insn copies R10 (which has FRAME_PTR) type into R1
71 * and 2nd arithmetic instruction is pattern matched to recognize
72 * that it wants to construct a pointer to some element within stack.
73 * So after 2nd insn, the register R1 has type PTR_TO_STACK
74 * (and -20 constant is saved for further stack bounds checking).
75 * Meaning that this reg is a pointer to stack plus known immediate constant.
77 * Most of the time the registers have SCALAR_VALUE type, which
78 * means the register has some value, but it's not a valid pointer.
79 * (like pointer plus pointer becomes SCALAR_VALUE type)
81 * When verifier sees load or store instructions the type of base register
82 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
83 * types recognized by check_mem_access() function.
85 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
86 * and the range of [ptr, ptr + map's value_size) is accessible.
88 * registers used to pass values to function calls are checked against
89 * function argument constraints.
91 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
92 * It means that the register type passed to this function must be
93 * PTR_TO_STACK and it will be used inside the function as
94 * 'pointer to map element key'
96 * For example the argument constraints for bpf_map_lookup_elem():
97 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
98 * .arg1_type = ARG_CONST_MAP_PTR,
99 * .arg2_type = ARG_PTR_TO_MAP_KEY,
101 * ret_type says that this function returns 'pointer to map elem value or null'
102 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
103 * 2nd argument should be a pointer to stack, which will be used inside
104 * the helper function as a pointer to map element key.
106 * On the kernel side the helper function looks like:
107 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
109 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
110 * void *key = (void *) (unsigned long) r2;
113 * here kernel can access 'key' and 'map' pointers safely, knowing that
114 * [key, key + map->key_size) bytes are valid and were initialized on
115 * the stack of eBPF program.
118 * Corresponding eBPF program may look like:
119 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
120 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
121 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
122 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
123 * here verifier looks at prototype of map_lookup_elem() and sees:
124 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
125 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
127 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
128 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
129 * and were initialized prior to this call.
130 * If it's ok, then verifier allows this BPF_CALL insn and looks at
131 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
132 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
133 * returns ether pointer to map value or NULL.
135 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
136 * insn, the register holding that pointer in the true branch changes state to
137 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
138 * branch. See check_cond_jmp_op().
140 * After the call R0 is set to return type of the function and registers R1-R5
141 * are set to NOT_INIT to indicate that they are no longer readable.
144 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
145 struct bpf_verifier_stack_elem {
146 /* verifer state is 'st'
147 * before processing instruction 'insn_idx'
148 * and after processing instruction 'prev_insn_idx'
150 struct bpf_verifier_state st;
153 struct bpf_verifier_stack_elem *next;
156 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
157 #define BPF_COMPLEXITY_LIMIT_STACK 1024
159 #define BPF_MAP_PTR_UNPRIV 1UL
160 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
161 POISON_POINTER_DELTA))
162 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
164 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
166 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
169 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
171 return aux->map_state & BPF_MAP_PTR_UNPRIV;
174 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
175 const struct bpf_map *map, bool unpriv)
177 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
178 unpriv |= bpf_map_ptr_unpriv(aux);
179 aux->map_state = (unsigned long)map |
180 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
183 struct bpf_call_arg_meta {
184 struct bpf_map *map_ptr;
191 static DEFINE_MUTEX(bpf_verifier_lock);
193 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
198 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
200 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
201 "verifier log line truncated - local buffer too short\n");
203 n = min(log->len_total - log->len_used - 1, n);
206 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
212 /* log_level controls verbosity level of eBPF verifier.
213 * bpf_verifier_log_write() is used to dump the verification trace to the log,
214 * so the user can figure out what's wrong with the program
216 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
217 const char *fmt, ...)
221 if (!bpf_verifier_log_needed(&env->log))
225 bpf_verifier_vlog(&env->log, fmt, args);
228 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
230 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
232 struct bpf_verifier_env *env = private_data;
235 if (!bpf_verifier_log_needed(&env->log))
239 bpf_verifier_vlog(&env->log, fmt, args);
243 static bool type_is_pkt_pointer(enum bpf_reg_type type)
245 return type == PTR_TO_PACKET ||
246 type == PTR_TO_PACKET_META;
249 /* string representation of 'enum bpf_reg_type' */
250 static const char * const reg_type_str[] = {
252 [SCALAR_VALUE] = "inv",
253 [PTR_TO_CTX] = "ctx",
254 [CONST_PTR_TO_MAP] = "map_ptr",
255 [PTR_TO_MAP_VALUE] = "map_value",
256 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
257 [PTR_TO_STACK] = "fp",
258 [PTR_TO_PACKET] = "pkt",
259 [PTR_TO_PACKET_META] = "pkt_meta",
260 [PTR_TO_PACKET_END] = "pkt_end",
263 static void print_liveness(struct bpf_verifier_env *env,
264 enum bpf_reg_liveness live)
266 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
268 if (live & REG_LIVE_READ)
270 if (live & REG_LIVE_WRITTEN)
274 static struct bpf_func_state *func(struct bpf_verifier_env *env,
275 const struct bpf_reg_state *reg)
277 struct bpf_verifier_state *cur = env->cur_state;
279 return cur->frame[reg->frameno];
282 static void print_verifier_state(struct bpf_verifier_env *env,
283 const struct bpf_func_state *state)
285 const struct bpf_reg_state *reg;
290 verbose(env, " frame%d:", state->frameno);
291 for (i = 0; i < MAX_BPF_REG; i++) {
292 reg = &state->regs[i];
296 verbose(env, " R%d", i);
297 print_liveness(env, reg->live);
298 verbose(env, "=%s", reg_type_str[t]);
299 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
300 tnum_is_const(reg->var_off)) {
301 /* reg->off should be 0 for SCALAR_VALUE */
302 verbose(env, "%lld", reg->var_off.value + reg->off);
303 if (t == PTR_TO_STACK)
304 verbose(env, ",call_%d", func(env, reg)->callsite);
306 verbose(env, "(id=%d", reg->id);
307 if (t != SCALAR_VALUE)
308 verbose(env, ",off=%d", reg->off);
309 if (type_is_pkt_pointer(t))
310 verbose(env, ",r=%d", reg->range);
311 else if (t == CONST_PTR_TO_MAP ||
312 t == PTR_TO_MAP_VALUE ||
313 t == PTR_TO_MAP_VALUE_OR_NULL)
314 verbose(env, ",ks=%d,vs=%d",
315 reg->map_ptr->key_size,
316 reg->map_ptr->value_size);
317 if (tnum_is_const(reg->var_off)) {
318 /* Typically an immediate SCALAR_VALUE, but
319 * could be a pointer whose offset is too big
322 verbose(env, ",imm=%llx", reg->var_off.value);
324 if (reg->smin_value != reg->umin_value &&
325 reg->smin_value != S64_MIN)
326 verbose(env, ",smin_value=%lld",
327 (long long)reg->smin_value);
328 if (reg->smax_value != reg->umax_value &&
329 reg->smax_value != S64_MAX)
330 verbose(env, ",smax_value=%lld",
331 (long long)reg->smax_value);
332 if (reg->umin_value != 0)
333 verbose(env, ",umin_value=%llu",
334 (unsigned long long)reg->umin_value);
335 if (reg->umax_value != U64_MAX)
336 verbose(env, ",umax_value=%llu",
337 (unsigned long long)reg->umax_value);
338 if (!tnum_is_unknown(reg->var_off)) {
341 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
342 verbose(env, ",var_off=%s", tn_buf);
348 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
349 if (state->stack[i].slot_type[0] == STACK_SPILL) {
350 verbose(env, " fp%d",
351 (-i - 1) * BPF_REG_SIZE);
352 print_liveness(env, state->stack[i].spilled_ptr.live);
354 reg_type_str[state->stack[i].spilled_ptr.type]);
356 if (state->stack[i].slot_type[0] == STACK_ZERO)
357 verbose(env, " fp%d=0", (-i - 1) * BPF_REG_SIZE);
362 static int copy_stack_state(struct bpf_func_state *dst,
363 const struct bpf_func_state *src)
367 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
368 /* internal bug, make state invalid to reject the program */
369 memset(dst, 0, sizeof(*dst));
372 memcpy(dst->stack, src->stack,
373 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
377 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
378 * make it consume minimal amount of memory. check_stack_write() access from
379 * the program calls into realloc_func_state() to grow the stack size.
380 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
381 * which this function copies over. It points to previous bpf_verifier_state
382 * which is never reallocated
384 static int realloc_func_state(struct bpf_func_state *state, int size,
387 u32 old_size = state->allocated_stack;
388 struct bpf_stack_state *new_stack;
389 int slot = size / BPF_REG_SIZE;
391 if (size <= old_size || !size) {
394 state->allocated_stack = slot * BPF_REG_SIZE;
395 if (!size && old_size) {
401 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
407 memcpy(new_stack, state->stack,
408 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
409 memset(new_stack + old_size / BPF_REG_SIZE, 0,
410 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
412 state->allocated_stack = slot * BPF_REG_SIZE;
414 state->stack = new_stack;
418 static void free_func_state(struct bpf_func_state *state)
426 static void free_verifier_state(struct bpf_verifier_state *state,
431 for (i = 0; i <= state->curframe; i++) {
432 free_func_state(state->frame[i]);
433 state->frame[i] = NULL;
439 /* copy verifier state from src to dst growing dst stack space
440 * when necessary to accommodate larger src stack
442 static int copy_func_state(struct bpf_func_state *dst,
443 const struct bpf_func_state *src)
447 err = realloc_func_state(dst, src->allocated_stack, false);
450 memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack));
451 return copy_stack_state(dst, src);
454 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
455 const struct bpf_verifier_state *src)
457 struct bpf_func_state *dst;
460 /* if dst has more stack frames then src frame, free them */
461 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
462 free_func_state(dst_state->frame[i]);
463 dst_state->frame[i] = NULL;
465 dst_state->curframe = src->curframe;
466 dst_state->parent = src->parent;
467 for (i = 0; i <= src->curframe; i++) {
468 dst = dst_state->frame[i];
470 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
473 dst_state->frame[i] = dst;
475 err = copy_func_state(dst, src->frame[i]);
482 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
485 struct bpf_verifier_state *cur = env->cur_state;
486 struct bpf_verifier_stack_elem *elem, *head = env->head;
489 if (env->head == NULL)
493 err = copy_verifier_state(cur, &head->st);
498 *insn_idx = head->insn_idx;
500 *prev_insn_idx = head->prev_insn_idx;
502 free_verifier_state(&head->st, false);
509 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
510 int insn_idx, int prev_insn_idx)
512 struct bpf_verifier_state *cur = env->cur_state;
513 struct bpf_verifier_stack_elem *elem;
516 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
520 elem->insn_idx = insn_idx;
521 elem->prev_insn_idx = prev_insn_idx;
522 elem->next = env->head;
525 err = copy_verifier_state(&elem->st, cur);
528 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
529 verbose(env, "BPF program is too complex\n");
534 free_verifier_state(env->cur_state, true);
535 env->cur_state = NULL;
536 /* pop all elements and return */
537 while (!pop_stack(env, NULL, NULL));
541 #define CALLER_SAVED_REGS 6
542 static const int caller_saved[CALLER_SAVED_REGS] = {
543 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
546 static void __mark_reg_not_init(struct bpf_reg_state *reg);
548 /* Mark the unknown part of a register (variable offset or scalar value) as
549 * known to have the value @imm.
551 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
554 reg->var_off = tnum_const(imm);
555 reg->smin_value = (s64)imm;
556 reg->smax_value = (s64)imm;
557 reg->umin_value = imm;
558 reg->umax_value = imm;
561 /* Mark the 'variable offset' part of a register as zero. This should be
562 * used only on registers holding a pointer type.
564 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
566 __mark_reg_known(reg, 0);
569 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
571 __mark_reg_known(reg, 0);
573 reg->type = SCALAR_VALUE;
576 static void mark_reg_known_zero(struct bpf_verifier_env *env,
577 struct bpf_reg_state *regs, u32 regno)
579 if (WARN_ON(regno >= MAX_BPF_REG)) {
580 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
581 /* Something bad happened, let's kill all regs */
582 for (regno = 0; regno < MAX_BPF_REG; regno++)
583 __mark_reg_not_init(regs + regno);
586 __mark_reg_known_zero(regs + regno);
589 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
591 return type_is_pkt_pointer(reg->type);
594 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
596 return reg_is_pkt_pointer(reg) ||
597 reg->type == PTR_TO_PACKET_END;
600 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
601 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
602 enum bpf_reg_type which)
604 /* The register can already have a range from prior markings.
605 * This is fine as long as it hasn't been advanced from its
608 return reg->type == which &&
611 tnum_equals_const(reg->var_off, 0);
614 /* Attempts to improve min/max values based on var_off information */
615 static void __update_reg_bounds(struct bpf_reg_state *reg)
617 /* min signed is max(sign bit) | min(other bits) */
618 reg->smin_value = max_t(s64, reg->smin_value,
619 reg->var_off.value | (reg->var_off.mask & S64_MIN));
620 /* max signed is min(sign bit) | max(other bits) */
621 reg->smax_value = min_t(s64, reg->smax_value,
622 reg->var_off.value | (reg->var_off.mask & S64_MAX));
623 reg->umin_value = max(reg->umin_value, reg->var_off.value);
624 reg->umax_value = min(reg->umax_value,
625 reg->var_off.value | reg->var_off.mask);
628 /* Uses signed min/max values to inform unsigned, and vice-versa */
629 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
631 /* Learn sign from signed bounds.
632 * If we cannot cross the sign boundary, then signed and unsigned bounds
633 * are the same, so combine. This works even in the negative case, e.g.
634 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
636 if (reg->smin_value >= 0 || reg->smax_value < 0) {
637 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
639 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
643 /* Learn sign from unsigned bounds. Signed bounds cross the sign
644 * boundary, so we must be careful.
646 if ((s64)reg->umax_value >= 0) {
647 /* Positive. We can't learn anything from the smin, but smax
648 * is positive, hence safe.
650 reg->smin_value = reg->umin_value;
651 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
653 } else if ((s64)reg->umin_value < 0) {
654 /* Negative. We can't learn anything from the smax, but smin
655 * is negative, hence safe.
657 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
659 reg->smax_value = reg->umax_value;
663 /* Attempts to improve var_off based on unsigned min/max information */
664 static void __reg_bound_offset(struct bpf_reg_state *reg)
666 reg->var_off = tnum_intersect(reg->var_off,
667 tnum_range(reg->umin_value,
671 /* Reset the min/max bounds of a register */
672 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
674 reg->smin_value = S64_MIN;
675 reg->smax_value = S64_MAX;
677 reg->umax_value = U64_MAX;
680 /* Mark a register as having a completely unknown (scalar) value. */
681 static void __mark_reg_unknown(struct bpf_reg_state *reg)
683 reg->type = SCALAR_VALUE;
686 reg->var_off = tnum_unknown;
688 __mark_reg_unbounded(reg);
691 static void mark_reg_unknown(struct bpf_verifier_env *env,
692 struct bpf_reg_state *regs, u32 regno)
694 if (WARN_ON(regno >= MAX_BPF_REG)) {
695 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
696 /* Something bad happened, let's kill all regs except FP */
697 for (regno = 0; regno < BPF_REG_FP; regno++)
698 __mark_reg_not_init(regs + regno);
701 __mark_reg_unknown(regs + regno);
704 static void __mark_reg_not_init(struct bpf_reg_state *reg)
706 __mark_reg_unknown(reg);
707 reg->type = NOT_INIT;
710 static void mark_reg_not_init(struct bpf_verifier_env *env,
711 struct bpf_reg_state *regs, u32 regno)
713 if (WARN_ON(regno >= MAX_BPF_REG)) {
714 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
715 /* Something bad happened, let's kill all regs except FP */
716 for (regno = 0; regno < BPF_REG_FP; regno++)
717 __mark_reg_not_init(regs + regno);
720 __mark_reg_not_init(regs + regno);
723 static void init_reg_state(struct bpf_verifier_env *env,
724 struct bpf_func_state *state)
726 struct bpf_reg_state *regs = state->regs;
729 for (i = 0; i < MAX_BPF_REG; i++) {
730 mark_reg_not_init(env, regs, i);
731 regs[i].live = REG_LIVE_NONE;
735 regs[BPF_REG_FP].type = PTR_TO_STACK;
736 mark_reg_known_zero(env, regs, BPF_REG_FP);
737 regs[BPF_REG_FP].frameno = state->frameno;
739 /* 1st arg to a function */
740 regs[BPF_REG_1].type = PTR_TO_CTX;
741 mark_reg_known_zero(env, regs, BPF_REG_1);
744 #define BPF_MAIN_FUNC (-1)
745 static void init_func_state(struct bpf_verifier_env *env,
746 struct bpf_func_state *state,
747 int callsite, int frameno, int subprogno)
749 state->callsite = callsite;
750 state->frameno = frameno;
751 state->subprogno = subprogno;
752 init_reg_state(env, state);
756 SRC_OP, /* register is used as source operand */
757 DST_OP, /* register is used as destination operand */
758 DST_OP_NO_MARK /* same as above, check only, don't mark */
761 static int cmp_subprogs(const void *a, const void *b)
763 return *(int *)a - *(int *)b;
766 static int find_subprog(struct bpf_verifier_env *env, int off)
770 p = bsearch(&off, env->subprog_starts, env->subprog_cnt,
771 sizeof(env->subprog_starts[0]), cmp_subprogs);
774 return p - env->subprog_starts;
778 static int add_subprog(struct bpf_verifier_env *env, int off)
780 int insn_cnt = env->prog->len;
783 if (off >= insn_cnt || off < 0) {
784 verbose(env, "call to invalid destination\n");
787 ret = find_subprog(env, off);
790 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
791 verbose(env, "too many subprograms\n");
794 env->subprog_starts[env->subprog_cnt++] = off;
795 sort(env->subprog_starts, env->subprog_cnt,
796 sizeof(env->subprog_starts[0]), cmp_subprogs, NULL);
800 static int check_subprogs(struct bpf_verifier_env *env)
802 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
803 struct bpf_insn *insn = env->prog->insnsi;
804 int insn_cnt = env->prog->len;
806 /* determine subprog starts. The end is one before the next starts */
807 for (i = 0; i < insn_cnt; i++) {
808 if (insn[i].code != (BPF_JMP | BPF_CALL))
810 if (insn[i].src_reg != BPF_PSEUDO_CALL)
812 if (!env->allow_ptr_leaks) {
813 verbose(env, "function calls to other bpf functions are allowed for root only\n");
816 if (bpf_prog_is_dev_bound(env->prog->aux)) {
817 verbose(env, "function calls in offloaded programs are not supported yet\n");
820 ret = add_subprog(env, i + insn[i].imm + 1);
825 if (env->log.level > 1)
826 for (i = 0; i < env->subprog_cnt; i++)
827 verbose(env, "func#%d @%d\n", i, env->subprog_starts[i]);
829 /* now check that all jumps are within the same subprog */
831 if (env->subprog_cnt == cur_subprog)
832 subprog_end = insn_cnt;
834 subprog_end = env->subprog_starts[cur_subprog++];
835 for (i = 0; i < insn_cnt; i++) {
836 u8 code = insn[i].code;
838 if (BPF_CLASS(code) != BPF_JMP)
840 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
842 off = i + insn[i].off + 1;
843 if (off < subprog_start || off >= subprog_end) {
844 verbose(env, "jump out of range from insn %d to %d\n", i, off);
848 if (i == subprog_end - 1) {
849 /* to avoid fall-through from one subprog into another
850 * the last insn of the subprog should be either exit
851 * or unconditional jump back
853 if (code != (BPF_JMP | BPF_EXIT) &&
854 code != (BPF_JMP | BPF_JA)) {
855 verbose(env, "last insn is not an exit or jmp\n");
858 subprog_start = subprog_end;
859 if (env->subprog_cnt == cur_subprog)
860 subprog_end = insn_cnt;
862 subprog_end = env->subprog_starts[cur_subprog++];
869 struct bpf_verifier_state *skip_callee(struct bpf_verifier_env *env,
870 const struct bpf_verifier_state *state,
871 struct bpf_verifier_state *parent,
874 struct bpf_verifier_state *tmp = NULL;
876 /* 'parent' could be a state of caller and
877 * 'state' could be a state of callee. In such case
878 * parent->curframe < state->curframe
879 * and it's ok for r1 - r5 registers
881 * 'parent' could be a callee's state after it bpf_exit-ed.
882 * In such case parent->curframe > state->curframe
883 * and it's ok for r0 only
885 if (parent->curframe == state->curframe ||
886 (parent->curframe < state->curframe &&
887 regno >= BPF_REG_1 && regno <= BPF_REG_5) ||
888 (parent->curframe > state->curframe &&
892 if (parent->curframe > state->curframe &&
893 regno >= BPF_REG_6) {
894 /* for callee saved regs we have to skip the whole chain
895 * of states that belong to callee and mark as LIVE_READ
896 * the registers before the call
899 while (tmp && tmp->curframe != state->curframe) {
910 verbose(env, "verifier bug regno %d tmp %p\n", regno, tmp);
911 verbose(env, "regno %d parent frame %d current frame %d\n",
912 regno, parent->curframe, state->curframe);
916 static int mark_reg_read(struct bpf_verifier_env *env,
917 const struct bpf_verifier_state *state,
918 struct bpf_verifier_state *parent,
921 bool writes = parent == state->parent; /* Observe write marks */
923 if (regno == BPF_REG_FP)
924 /* We don't need to worry about FP liveness because it's read-only */
928 /* if read wasn't screened by an earlier write ... */
929 if (writes && state->frame[state->curframe]->regs[regno].live & REG_LIVE_WRITTEN)
931 parent = skip_callee(env, state, parent, regno);
934 /* ... then we depend on parent's value */
935 parent->frame[parent->curframe]->regs[regno].live |= REG_LIVE_READ;
937 parent = state->parent;
943 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
946 struct bpf_verifier_state *vstate = env->cur_state;
947 struct bpf_func_state *state = vstate->frame[vstate->curframe];
948 struct bpf_reg_state *regs = state->regs;
950 if (regno >= MAX_BPF_REG) {
951 verbose(env, "R%d is invalid\n", regno);
956 /* check whether register used as source operand can be read */
957 if (regs[regno].type == NOT_INIT) {
958 verbose(env, "R%d !read_ok\n", regno);
961 return mark_reg_read(env, vstate, vstate->parent, regno);
963 /* check whether register used as dest operand can be written to */
964 if (regno == BPF_REG_FP) {
965 verbose(env, "frame pointer is read only\n");
968 regs[regno].live |= REG_LIVE_WRITTEN;
970 mark_reg_unknown(env, regs, regno);
975 static bool is_spillable_regtype(enum bpf_reg_type type)
978 case PTR_TO_MAP_VALUE:
979 case PTR_TO_MAP_VALUE_OR_NULL:
983 case PTR_TO_PACKET_META:
984 case PTR_TO_PACKET_END:
985 case CONST_PTR_TO_MAP:
992 /* Does this register contain a constant zero? */
993 static bool register_is_null(struct bpf_reg_state *reg)
995 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
998 /* check_stack_read/write functions track spill/fill of registers,
999 * stack boundary and alignment are checked in check_mem_access()
1001 static int check_stack_write(struct bpf_verifier_env *env,
1002 struct bpf_func_state *state, /* func where register points to */
1003 int off, int size, int value_regno)
1005 struct bpf_func_state *cur; /* state of the current function */
1006 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1007 enum bpf_reg_type type;
1009 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1013 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1014 * so it's aligned access and [off, off + size) are within stack limits
1016 if (!env->allow_ptr_leaks &&
1017 state->stack[spi].slot_type[0] == STACK_SPILL &&
1018 size != BPF_REG_SIZE) {
1019 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1023 cur = env->cur_state->frame[env->cur_state->curframe];
1024 if (value_regno >= 0 &&
1025 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1027 /* register containing pointer is being spilled into stack */
1028 if (size != BPF_REG_SIZE) {
1029 verbose(env, "invalid size of register spill\n");
1033 if (state != cur && type == PTR_TO_STACK) {
1034 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1038 /* save register state */
1039 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1040 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1042 for (i = 0; i < BPF_REG_SIZE; i++)
1043 state->stack[spi].slot_type[i] = STACK_SPILL;
1045 u8 type = STACK_MISC;
1047 /* regular write of data into stack */
1048 state->stack[spi].spilled_ptr = (struct bpf_reg_state) {};
1050 /* only mark the slot as written if all 8 bytes were written
1051 * otherwise read propagation may incorrectly stop too soon
1052 * when stack slots are partially written.
1053 * This heuristic means that read propagation will be
1054 * conservative, since it will add reg_live_read marks
1055 * to stack slots all the way to first state when programs
1056 * writes+reads less than 8 bytes
1058 if (size == BPF_REG_SIZE)
1059 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1061 /* when we zero initialize stack slots mark them as such */
1062 if (value_regno >= 0 &&
1063 register_is_null(&cur->regs[value_regno]))
1066 for (i = 0; i < size; i++)
1067 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1073 /* registers of every function are unique and mark_reg_read() propagates
1074 * the liveness in the following cases:
1075 * - from callee into caller for R1 - R5 that were used as arguments
1076 * - from caller into callee for R0 that used as result of the call
1077 * - from caller to the same caller skipping states of the callee for R6 - R9,
1078 * since R6 - R9 are callee saved by implicit function prologue and
1079 * caller's R6 != callee's R6, so when we propagate liveness up to
1080 * parent states we need to skip callee states for R6 - R9.
1082 * stack slot marking is different, since stacks of caller and callee are
1083 * accessible in both (since caller can pass a pointer to caller's stack to
1084 * callee which can pass it to another function), hence mark_stack_slot_read()
1085 * has to propagate the stack liveness to all parent states at given frame number.
1095 * First *ptr is reading from f1's stack and mark_stack_slot_read() has
1096 * to mark liveness at the f1's frame and not f2's frame.
1097 * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has
1098 * to propagate liveness to f2 states at f1's frame level and further into
1099 * f1 states at f1's frame level until write into that stack slot
1101 static void mark_stack_slot_read(struct bpf_verifier_env *env,
1102 const struct bpf_verifier_state *state,
1103 struct bpf_verifier_state *parent,
1104 int slot, int frameno)
1106 bool writes = parent == state->parent; /* Observe write marks */
1109 if (parent->frame[frameno]->allocated_stack <= slot * BPF_REG_SIZE)
1110 /* since LIVE_WRITTEN mark is only done for full 8-byte
1111 * write the read marks are conservative and parent
1112 * state may not even have the stack allocated. In such case
1113 * end the propagation, since the loop reached beginning
1117 /* if read wasn't screened by an earlier write ... */
1118 if (writes && state->frame[frameno]->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN)
1120 /* ... then we depend on parent's value */
1121 parent->frame[frameno]->stack[slot].spilled_ptr.live |= REG_LIVE_READ;
1123 parent = state->parent;
1128 static int check_stack_read(struct bpf_verifier_env *env,
1129 struct bpf_func_state *reg_state /* func where register points to */,
1130 int off, int size, int value_regno)
1132 struct bpf_verifier_state *vstate = env->cur_state;
1133 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1134 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1137 if (reg_state->allocated_stack <= slot) {
1138 verbose(env, "invalid read from stack off %d+0 size %d\n",
1142 stype = reg_state->stack[spi].slot_type;
1144 if (stype[0] == STACK_SPILL) {
1145 if (size != BPF_REG_SIZE) {
1146 verbose(env, "invalid size of register spill\n");
1149 for (i = 1; i < BPF_REG_SIZE; i++) {
1150 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1151 verbose(env, "corrupted spill memory\n");
1156 if (value_regno >= 0) {
1157 /* restore register state from stack */
1158 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1159 /* mark reg as written since spilled pointer state likely
1160 * has its liveness marks cleared by is_state_visited()
1161 * which resets stack/reg liveness for state transitions
1163 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1165 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1166 reg_state->frameno);
1171 for (i = 0; i < size; i++) {
1172 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1174 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1178 verbose(env, "invalid read from stack off %d+%d size %d\n",
1182 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1183 reg_state->frameno);
1184 if (value_regno >= 0) {
1185 if (zeros == size) {
1186 /* any size read into register is zero extended,
1187 * so the whole register == const_zero
1189 __mark_reg_const_zero(&state->regs[value_regno]);
1191 /* have read misc data from the stack */
1192 mark_reg_unknown(env, state->regs, value_regno);
1194 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1200 /* check read/write into map element returned by bpf_map_lookup_elem() */
1201 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1202 int size, bool zero_size_allowed)
1204 struct bpf_reg_state *regs = cur_regs(env);
1205 struct bpf_map *map = regs[regno].map_ptr;
1207 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1208 off + size > map->value_size) {
1209 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1210 map->value_size, off, size);
1216 /* check read/write into a map element with possible variable offset */
1217 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1218 int off, int size, bool zero_size_allowed)
1220 struct bpf_verifier_state *vstate = env->cur_state;
1221 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1222 struct bpf_reg_state *reg = &state->regs[regno];
1225 /* We may have adjusted the register to this map value, so we
1226 * need to try adding each of min_value and max_value to off
1227 * to make sure our theoretical access will be safe.
1230 print_verifier_state(env, state);
1231 /* The minimum value is only important with signed
1232 * comparisons where we can't assume the floor of a
1233 * value is 0. If we are using signed variables for our
1234 * index'es we need to make sure that whatever we use
1235 * will have a set floor within our range.
1237 if (reg->smin_value < 0) {
1238 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1242 err = __check_map_access(env, regno, reg->smin_value + off, size,
1245 verbose(env, "R%d min value is outside of the array range\n",
1250 /* If we haven't set a max value then we need to bail since we can't be
1251 * sure we won't do bad things.
1252 * If reg->umax_value + off could overflow, treat that as unbounded too.
1254 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1255 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1259 err = __check_map_access(env, regno, reg->umax_value + off, size,
1262 verbose(env, "R%d max value is outside of the array range\n",
1267 #define MAX_PACKET_OFF 0xffff
1269 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1270 const struct bpf_call_arg_meta *meta,
1271 enum bpf_access_type t)
1273 switch (env->prog->type) {
1274 case BPF_PROG_TYPE_LWT_IN:
1275 case BPF_PROG_TYPE_LWT_OUT:
1276 /* dst_input() and dst_output() can't write for now */
1280 case BPF_PROG_TYPE_SCHED_CLS:
1281 case BPF_PROG_TYPE_SCHED_ACT:
1282 case BPF_PROG_TYPE_XDP:
1283 case BPF_PROG_TYPE_LWT_XMIT:
1284 case BPF_PROG_TYPE_SK_SKB:
1285 case BPF_PROG_TYPE_SK_MSG:
1287 return meta->pkt_access;
1289 env->seen_direct_write = true;
1296 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1297 int off, int size, bool zero_size_allowed)
1299 struct bpf_reg_state *regs = cur_regs(env);
1300 struct bpf_reg_state *reg = ®s[regno];
1302 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1303 (u64)off + size > reg->range) {
1304 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1305 off, size, regno, reg->id, reg->off, reg->range);
1311 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1312 int size, bool zero_size_allowed)
1314 struct bpf_reg_state *regs = cur_regs(env);
1315 struct bpf_reg_state *reg = ®s[regno];
1318 /* We may have added a variable offset to the packet pointer; but any
1319 * reg->range we have comes after that. We are only checking the fixed
1323 /* We don't allow negative numbers, because we aren't tracking enough
1324 * detail to prove they're safe.
1326 if (reg->smin_value < 0) {
1327 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1331 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1333 verbose(env, "R%d offset is outside of the packet\n", regno);
1339 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1340 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1341 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1343 struct bpf_insn_access_aux info = {
1344 .reg_type = *reg_type,
1347 if (env->ops->is_valid_access &&
1348 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1349 /* A non zero info.ctx_field_size indicates that this field is a
1350 * candidate for later verifier transformation to load the whole
1351 * field and then apply a mask when accessed with a narrower
1352 * access than actual ctx access size. A zero info.ctx_field_size
1353 * will only allow for whole field access and rejects any other
1354 * type of narrower access.
1356 *reg_type = info.reg_type;
1358 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1359 /* remember the offset of last byte accessed in ctx */
1360 if (env->prog->aux->max_ctx_offset < off + size)
1361 env->prog->aux->max_ctx_offset = off + size;
1365 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1369 static bool __is_pointer_value(bool allow_ptr_leaks,
1370 const struct bpf_reg_state *reg)
1372 if (allow_ptr_leaks)
1375 return reg->type != SCALAR_VALUE;
1378 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1380 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1383 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1385 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1387 return reg->type == PTR_TO_CTX;
1390 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1392 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1394 return type_is_pkt_pointer(reg->type);
1397 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1398 const struct bpf_reg_state *reg,
1399 int off, int size, bool strict)
1401 struct tnum reg_off;
1404 /* Byte size accesses are always allowed. */
1405 if (!strict || size == 1)
1408 /* For platforms that do not have a Kconfig enabling
1409 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1410 * NET_IP_ALIGN is universally set to '2'. And on platforms
1411 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1412 * to this code only in strict mode where we want to emulate
1413 * the NET_IP_ALIGN==2 checking. Therefore use an
1414 * unconditional IP align value of '2'.
1418 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1419 if (!tnum_is_aligned(reg_off, size)) {
1422 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1424 "misaligned packet access off %d+%s+%d+%d size %d\n",
1425 ip_align, tn_buf, reg->off, off, size);
1432 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1433 const struct bpf_reg_state *reg,
1434 const char *pointer_desc,
1435 int off, int size, bool strict)
1437 struct tnum reg_off;
1439 /* Byte size accesses are always allowed. */
1440 if (!strict || size == 1)
1443 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1444 if (!tnum_is_aligned(reg_off, size)) {
1447 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1448 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1449 pointer_desc, tn_buf, reg->off, off, size);
1456 static int check_ptr_alignment(struct bpf_verifier_env *env,
1457 const struct bpf_reg_state *reg, int off,
1458 int size, bool strict_alignment_once)
1460 bool strict = env->strict_alignment || strict_alignment_once;
1461 const char *pointer_desc = "";
1463 switch (reg->type) {
1465 case PTR_TO_PACKET_META:
1466 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1467 * right in front, treat it the very same way.
1469 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1470 case PTR_TO_MAP_VALUE:
1471 pointer_desc = "value ";
1474 pointer_desc = "context ";
1477 pointer_desc = "stack ";
1478 /* The stack spill tracking logic in check_stack_write()
1479 * and check_stack_read() relies on stack accesses being
1487 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1491 static int update_stack_depth(struct bpf_verifier_env *env,
1492 const struct bpf_func_state *func,
1495 u16 stack = env->subprog_stack_depth[func->subprogno];
1500 /* update known max for given subprogram */
1501 env->subprog_stack_depth[func->subprogno] = -off;
1505 /* starting from main bpf function walk all instructions of the function
1506 * and recursively walk all callees that given function can call.
1507 * Ignore jump and exit insns.
1508 * Since recursion is prevented by check_cfg() this algorithm
1509 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1511 static int check_max_stack_depth(struct bpf_verifier_env *env)
1513 int depth = 0, frame = 0, subprog = 0, i = 0, subprog_end;
1514 struct bpf_insn *insn = env->prog->insnsi;
1515 int insn_cnt = env->prog->len;
1516 int ret_insn[MAX_CALL_FRAMES];
1517 int ret_prog[MAX_CALL_FRAMES];
1520 /* round up to 32-bytes, since this is granularity
1521 * of interpreter stack size
1523 depth += round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1524 if (depth > MAX_BPF_STACK) {
1525 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1530 if (env->subprog_cnt == subprog)
1531 subprog_end = insn_cnt;
1533 subprog_end = env->subprog_starts[subprog];
1534 for (; i < subprog_end; i++) {
1535 if (insn[i].code != (BPF_JMP | BPF_CALL))
1537 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1539 /* remember insn and function to return to */
1540 ret_insn[frame] = i + 1;
1541 ret_prog[frame] = subprog;
1543 /* find the callee */
1544 i = i + insn[i].imm + 1;
1545 subprog = find_subprog(env, i);
1547 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1553 if (frame >= MAX_CALL_FRAMES) {
1554 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1559 /* end of for() loop means the last insn of the 'subprog'
1560 * was reached. Doesn't matter whether it was JA or EXIT
1564 depth -= round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1566 i = ret_insn[frame];
1567 subprog = ret_prog[frame];
1571 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1572 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1573 const struct bpf_insn *insn, int idx)
1575 int start = idx + insn->imm + 1, subprog;
1577 subprog = find_subprog(env, start);
1579 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1584 return env->subprog_stack_depth[subprog];
1588 /* truncate register to smaller size (in bytes)
1589 * must be called with size < BPF_REG_SIZE
1591 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1595 /* clear high bits in bit representation */
1596 reg->var_off = tnum_cast(reg->var_off, size);
1598 /* fix arithmetic bounds */
1599 mask = ((u64)1 << (size * 8)) - 1;
1600 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1601 reg->umin_value &= mask;
1602 reg->umax_value &= mask;
1604 reg->umin_value = 0;
1605 reg->umax_value = mask;
1607 reg->smin_value = reg->umin_value;
1608 reg->smax_value = reg->umax_value;
1611 /* check whether memory at (regno + off) is accessible for t = (read | write)
1612 * if t==write, value_regno is a register which value is stored into memory
1613 * if t==read, value_regno is a register which will receive the value from memory
1614 * if t==write && value_regno==-1, some unknown value is stored into memory
1615 * if t==read && value_regno==-1, don't care what we read from memory
1617 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1618 int off, int bpf_size, enum bpf_access_type t,
1619 int value_regno, bool strict_alignment_once)
1621 struct bpf_reg_state *regs = cur_regs(env);
1622 struct bpf_reg_state *reg = regs + regno;
1623 struct bpf_func_state *state;
1626 size = bpf_size_to_bytes(bpf_size);
1630 /* alignment checks will add in reg->off themselves */
1631 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1635 /* for access checks, reg->off is just part of off */
1638 if (reg->type == PTR_TO_MAP_VALUE) {
1639 if (t == BPF_WRITE && value_regno >= 0 &&
1640 is_pointer_value(env, value_regno)) {
1641 verbose(env, "R%d leaks addr into map\n", value_regno);
1645 err = check_map_access(env, regno, off, size, false);
1646 if (!err && t == BPF_READ && value_regno >= 0)
1647 mark_reg_unknown(env, regs, value_regno);
1649 } else if (reg->type == PTR_TO_CTX) {
1650 enum bpf_reg_type reg_type = SCALAR_VALUE;
1652 if (t == BPF_WRITE && value_regno >= 0 &&
1653 is_pointer_value(env, value_regno)) {
1654 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1657 /* ctx accesses must be at a fixed offset, so that we can
1658 * determine what type of data were returned.
1662 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1663 regno, reg->off, off - reg->off);
1666 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1669 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1671 "variable ctx access var_off=%s off=%d size=%d",
1675 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1676 if (!err && t == BPF_READ && value_regno >= 0) {
1677 /* ctx access returns either a scalar, or a
1678 * PTR_TO_PACKET[_META,_END]. In the latter
1679 * case, we know the offset is zero.
1681 if (reg_type == SCALAR_VALUE)
1682 mark_reg_unknown(env, regs, value_regno);
1684 mark_reg_known_zero(env, regs,
1686 regs[value_regno].id = 0;
1687 regs[value_regno].off = 0;
1688 regs[value_regno].range = 0;
1689 regs[value_regno].type = reg_type;
1692 } else if (reg->type == PTR_TO_STACK) {
1693 /* stack accesses must be at a fixed offset, so that we can
1694 * determine what type of data were returned.
1695 * See check_stack_read().
1697 if (!tnum_is_const(reg->var_off)) {
1700 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1701 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1705 off += reg->var_off.value;
1706 if (off >= 0 || off < -MAX_BPF_STACK) {
1707 verbose(env, "invalid stack off=%d size=%d\n", off,
1712 state = func(env, reg);
1713 err = update_stack_depth(env, state, off);
1718 err = check_stack_write(env, state, off, size,
1721 err = check_stack_read(env, state, off, size,
1723 } else if (reg_is_pkt_pointer(reg)) {
1724 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1725 verbose(env, "cannot write into packet\n");
1728 if (t == BPF_WRITE && value_regno >= 0 &&
1729 is_pointer_value(env, value_regno)) {
1730 verbose(env, "R%d leaks addr into packet\n",
1734 err = check_packet_access(env, regno, off, size, false);
1735 if (!err && t == BPF_READ && value_regno >= 0)
1736 mark_reg_unknown(env, regs, value_regno);
1738 verbose(env, "R%d invalid mem access '%s'\n", regno,
1739 reg_type_str[reg->type]);
1743 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1744 regs[value_regno].type == SCALAR_VALUE) {
1745 /* b/h/w load zero-extends, mark upper bits as known 0 */
1746 coerce_reg_to_size(®s[value_regno], size);
1751 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1755 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1757 verbose(env, "BPF_XADD uses reserved fields\n");
1761 /* check src1 operand */
1762 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1766 /* check src2 operand */
1767 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1771 if (is_pointer_value(env, insn->src_reg)) {
1772 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1776 if (is_ctx_reg(env, insn->dst_reg) ||
1777 is_pkt_reg(env, insn->dst_reg)) {
1778 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1779 insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ?
1780 "context" : "packet");
1784 /* check whether atomic_add can read the memory */
1785 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1786 BPF_SIZE(insn->code), BPF_READ, -1, true);
1790 /* check whether atomic_add can write into the same memory */
1791 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1792 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1795 /* when register 'regno' is passed into function that will read 'access_size'
1796 * bytes from that pointer, make sure that it's within stack boundary
1797 * and all elements of stack are initialized.
1798 * Unlike most pointer bounds-checking functions, this one doesn't take an
1799 * 'off' argument, so it has to add in reg->off itself.
1801 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1802 int access_size, bool zero_size_allowed,
1803 struct bpf_call_arg_meta *meta)
1805 struct bpf_reg_state *reg = cur_regs(env) + regno;
1806 struct bpf_func_state *state = func(env, reg);
1807 int off, i, slot, spi;
1809 if (reg->type != PTR_TO_STACK) {
1810 /* Allow zero-byte read from NULL, regardless of pointer type */
1811 if (zero_size_allowed && access_size == 0 &&
1812 register_is_null(reg))
1815 verbose(env, "R%d type=%s expected=%s\n", regno,
1816 reg_type_str[reg->type],
1817 reg_type_str[PTR_TO_STACK]);
1821 /* Only allow fixed-offset stack reads */
1822 if (!tnum_is_const(reg->var_off)) {
1825 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1826 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1830 off = reg->off + reg->var_off.value;
1831 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1832 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1833 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1834 regno, off, access_size);
1838 if (meta && meta->raw_mode) {
1839 meta->access_size = access_size;
1840 meta->regno = regno;
1844 for (i = 0; i < access_size; i++) {
1847 slot = -(off + i) - 1;
1848 spi = slot / BPF_REG_SIZE;
1849 if (state->allocated_stack <= slot)
1851 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
1852 if (*stype == STACK_MISC)
1854 if (*stype == STACK_ZERO) {
1855 /* helper can write anything into the stack */
1856 *stype = STACK_MISC;
1860 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1861 off, i, access_size);
1864 /* reading any byte out of 8-byte 'spill_slot' will cause
1865 * the whole slot to be marked as 'read'
1867 mark_stack_slot_read(env, env->cur_state, env->cur_state->parent,
1868 spi, state->frameno);
1870 return update_stack_depth(env, state, off);
1873 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1874 int access_size, bool zero_size_allowed,
1875 struct bpf_call_arg_meta *meta)
1877 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1879 switch (reg->type) {
1881 case PTR_TO_PACKET_META:
1882 return check_packet_access(env, regno, reg->off, access_size,
1884 case PTR_TO_MAP_VALUE:
1885 return check_map_access(env, regno, reg->off, access_size,
1887 default: /* scalar_value|ptr_to_stack or invalid ptr */
1888 return check_stack_boundary(env, regno, access_size,
1889 zero_size_allowed, meta);
1893 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
1895 return type == ARG_PTR_TO_MEM ||
1896 type == ARG_PTR_TO_MEM_OR_NULL ||
1897 type == ARG_PTR_TO_UNINIT_MEM;
1900 static bool arg_type_is_mem_size(enum bpf_arg_type type)
1902 return type == ARG_CONST_SIZE ||
1903 type == ARG_CONST_SIZE_OR_ZERO;
1906 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1907 enum bpf_arg_type arg_type,
1908 struct bpf_call_arg_meta *meta)
1910 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1911 enum bpf_reg_type expected_type, type = reg->type;
1914 if (arg_type == ARG_DONTCARE)
1917 err = check_reg_arg(env, regno, SRC_OP);
1921 if (arg_type == ARG_ANYTHING) {
1922 if (is_pointer_value(env, regno)) {
1923 verbose(env, "R%d leaks addr into helper function\n",
1930 if (type_is_pkt_pointer(type) &&
1931 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1932 verbose(env, "helper access to the packet is not allowed\n");
1936 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1937 arg_type == ARG_PTR_TO_MAP_VALUE) {
1938 expected_type = PTR_TO_STACK;
1939 if (!type_is_pkt_pointer(type) &&
1940 type != expected_type)
1942 } else if (arg_type == ARG_CONST_SIZE ||
1943 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1944 expected_type = SCALAR_VALUE;
1945 if (type != expected_type)
1947 } else if (arg_type == ARG_CONST_MAP_PTR) {
1948 expected_type = CONST_PTR_TO_MAP;
1949 if (type != expected_type)
1951 } else if (arg_type == ARG_PTR_TO_CTX) {
1952 expected_type = PTR_TO_CTX;
1953 if (type != expected_type)
1955 } else if (arg_type_is_mem_ptr(arg_type)) {
1956 expected_type = PTR_TO_STACK;
1957 /* One exception here. In case function allows for NULL to be
1958 * passed in as argument, it's a SCALAR_VALUE type. Final test
1959 * happens during stack boundary checking.
1961 if (register_is_null(reg) &&
1962 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1963 /* final test in check_stack_boundary() */;
1964 else if (!type_is_pkt_pointer(type) &&
1965 type != PTR_TO_MAP_VALUE &&
1966 type != expected_type)
1968 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1970 verbose(env, "unsupported arg_type %d\n", arg_type);
1974 if (arg_type == ARG_CONST_MAP_PTR) {
1975 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1976 meta->map_ptr = reg->map_ptr;
1977 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1978 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1979 * check that [key, key + map->key_size) are within
1980 * stack limits and initialized
1982 if (!meta->map_ptr) {
1983 /* in function declaration map_ptr must come before
1984 * map_key, so that it's verified and known before
1985 * we have to check map_key here. Otherwise it means
1986 * that kernel subsystem misconfigured verifier
1988 verbose(env, "invalid map_ptr to access map->key\n");
1991 if (type_is_pkt_pointer(type))
1992 err = check_packet_access(env, regno, reg->off,
1993 meta->map_ptr->key_size,
1996 err = check_stack_boundary(env, regno,
1997 meta->map_ptr->key_size,
1999 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
2000 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2001 * check [value, value + map->value_size) validity
2003 if (!meta->map_ptr) {
2004 /* kernel subsystem misconfigured verifier */
2005 verbose(env, "invalid map_ptr to access map->value\n");
2008 if (type_is_pkt_pointer(type))
2009 err = check_packet_access(env, regno, reg->off,
2010 meta->map_ptr->value_size,
2013 err = check_stack_boundary(env, regno,
2014 meta->map_ptr->value_size,
2016 } else if (arg_type_is_mem_size(arg_type)) {
2017 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2019 /* The register is SCALAR_VALUE; the access check
2020 * happens using its boundaries.
2022 if (!tnum_is_const(reg->var_off))
2023 /* For unprivileged variable accesses, disable raw
2024 * mode so that the program is required to
2025 * initialize all the memory that the helper could
2026 * just partially fill up.
2030 if (reg->smin_value < 0) {
2031 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2036 if (reg->umin_value == 0) {
2037 err = check_helper_mem_access(env, regno - 1, 0,
2044 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2045 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2049 err = check_helper_mem_access(env, regno - 1,
2051 zero_size_allowed, meta);
2056 verbose(env, "R%d type=%s expected=%s\n", regno,
2057 reg_type_str[type], reg_type_str[expected_type]);
2061 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2062 struct bpf_map *map, int func_id)
2067 /* We need a two way check, first is from map perspective ... */
2068 switch (map->map_type) {
2069 case BPF_MAP_TYPE_PROG_ARRAY:
2070 if (func_id != BPF_FUNC_tail_call)
2073 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2074 if (func_id != BPF_FUNC_perf_event_read &&
2075 func_id != BPF_FUNC_perf_event_output &&
2076 func_id != BPF_FUNC_perf_event_read_value)
2079 case BPF_MAP_TYPE_STACK_TRACE:
2080 if (func_id != BPF_FUNC_get_stackid)
2083 case BPF_MAP_TYPE_CGROUP_ARRAY:
2084 if (func_id != BPF_FUNC_skb_under_cgroup &&
2085 func_id != BPF_FUNC_current_task_under_cgroup)
2088 /* devmap returns a pointer to a live net_device ifindex that we cannot
2089 * allow to be modified from bpf side. So do not allow lookup elements
2092 case BPF_MAP_TYPE_DEVMAP:
2093 if (func_id != BPF_FUNC_redirect_map)
2096 /* Restrict bpf side of cpumap, open when use-cases appear */
2097 case BPF_MAP_TYPE_CPUMAP:
2098 if (func_id != BPF_FUNC_redirect_map)
2101 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2102 case BPF_MAP_TYPE_HASH_OF_MAPS:
2103 if (func_id != BPF_FUNC_map_lookup_elem)
2106 case BPF_MAP_TYPE_SOCKMAP:
2107 if (func_id != BPF_FUNC_sk_redirect_map &&
2108 func_id != BPF_FUNC_sock_map_update &&
2109 func_id != BPF_FUNC_map_delete_elem &&
2110 func_id != BPF_FUNC_msg_redirect_map)
2117 /* ... and second from the function itself. */
2119 case BPF_FUNC_tail_call:
2120 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2122 if (env->subprog_cnt) {
2123 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2127 case BPF_FUNC_perf_event_read:
2128 case BPF_FUNC_perf_event_output:
2129 case BPF_FUNC_perf_event_read_value:
2130 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2133 case BPF_FUNC_get_stackid:
2134 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2137 case BPF_FUNC_current_task_under_cgroup:
2138 case BPF_FUNC_skb_under_cgroup:
2139 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2142 case BPF_FUNC_redirect_map:
2143 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2144 map->map_type != BPF_MAP_TYPE_CPUMAP)
2147 case BPF_FUNC_sk_redirect_map:
2148 case BPF_FUNC_msg_redirect_map:
2149 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2152 case BPF_FUNC_sock_map_update:
2153 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2162 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2163 map->map_type, func_id_name(func_id), func_id);
2167 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2171 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2173 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2175 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2177 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2179 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2182 /* We only support one arg being in raw mode at the moment,
2183 * which is sufficient for the helper functions we have
2189 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2190 enum bpf_arg_type arg_next)
2192 return (arg_type_is_mem_ptr(arg_curr) &&
2193 !arg_type_is_mem_size(arg_next)) ||
2194 (!arg_type_is_mem_ptr(arg_curr) &&
2195 arg_type_is_mem_size(arg_next));
2198 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2200 /* bpf_xxx(..., buf, len) call will access 'len'
2201 * bytes from memory 'buf'. Both arg types need
2202 * to be paired, so make sure there's no buggy
2203 * helper function specification.
2205 if (arg_type_is_mem_size(fn->arg1_type) ||
2206 arg_type_is_mem_ptr(fn->arg5_type) ||
2207 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2208 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2209 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2210 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2216 static int check_func_proto(const struct bpf_func_proto *fn)
2218 return check_raw_mode_ok(fn) &&
2219 check_arg_pair_ok(fn) ? 0 : -EINVAL;
2222 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2223 * are now invalid, so turn them into unknown SCALAR_VALUE.
2225 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2226 struct bpf_func_state *state)
2228 struct bpf_reg_state *regs = state->regs, *reg;
2231 for (i = 0; i < MAX_BPF_REG; i++)
2232 if (reg_is_pkt_pointer_any(®s[i]))
2233 mark_reg_unknown(env, regs, i);
2235 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2236 if (state->stack[i].slot_type[0] != STACK_SPILL)
2238 reg = &state->stack[i].spilled_ptr;
2239 if (reg_is_pkt_pointer_any(reg))
2240 __mark_reg_unknown(reg);
2244 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2246 struct bpf_verifier_state *vstate = env->cur_state;
2249 for (i = 0; i <= vstate->curframe; i++)
2250 __clear_all_pkt_pointers(env, vstate->frame[i]);
2253 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2256 struct bpf_verifier_state *state = env->cur_state;
2257 struct bpf_func_state *caller, *callee;
2258 int i, subprog, target_insn;
2260 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2261 verbose(env, "the call stack of %d frames is too deep\n",
2262 state->curframe + 2);
2266 target_insn = *insn_idx + insn->imm;
2267 subprog = find_subprog(env, target_insn + 1);
2269 verbose(env, "verifier bug. No program starts at insn %d\n",
2274 caller = state->frame[state->curframe];
2275 if (state->frame[state->curframe + 1]) {
2276 verbose(env, "verifier bug. Frame %d already allocated\n",
2277 state->curframe + 1);
2281 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2284 state->frame[state->curframe + 1] = callee;
2286 /* callee cannot access r0, r6 - r9 for reading and has to write
2287 * into its own stack before reading from it.
2288 * callee can read/write into caller's stack
2290 init_func_state(env, callee,
2291 /* remember the callsite, it will be used by bpf_exit */
2292 *insn_idx /* callsite */,
2293 state->curframe + 1 /* frameno within this callchain */,
2294 subprog + 1 /* subprog number within this prog */);
2296 /* copy r1 - r5 args that callee can access */
2297 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2298 callee->regs[i] = caller->regs[i];
2300 /* after the call regsiters r0 - r5 were scratched */
2301 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2302 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2303 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2306 /* only increment it after check_reg_arg() finished */
2309 /* and go analyze first insn of the callee */
2310 *insn_idx = target_insn;
2312 if (env->log.level) {
2313 verbose(env, "caller:\n");
2314 print_verifier_state(env, caller);
2315 verbose(env, "callee:\n");
2316 print_verifier_state(env, callee);
2321 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2323 struct bpf_verifier_state *state = env->cur_state;
2324 struct bpf_func_state *caller, *callee;
2325 struct bpf_reg_state *r0;
2327 callee = state->frame[state->curframe];
2328 r0 = &callee->regs[BPF_REG_0];
2329 if (r0->type == PTR_TO_STACK) {
2330 /* technically it's ok to return caller's stack pointer
2331 * (or caller's caller's pointer) back to the caller,
2332 * since these pointers are valid. Only current stack
2333 * pointer will be invalid as soon as function exits,
2334 * but let's be conservative
2336 verbose(env, "cannot return stack pointer to the caller\n");
2341 caller = state->frame[state->curframe];
2342 /* return to the caller whatever r0 had in the callee */
2343 caller->regs[BPF_REG_0] = *r0;
2345 *insn_idx = callee->callsite + 1;
2346 if (env->log.level) {
2347 verbose(env, "returning from callee:\n");
2348 print_verifier_state(env, callee);
2349 verbose(env, "to caller at %d:\n", *insn_idx);
2350 print_verifier_state(env, caller);
2352 /* clear everything in the callee */
2353 free_func_state(callee);
2354 state->frame[state->curframe + 1] = NULL;
2359 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2360 int func_id, int insn_idx)
2362 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2364 if (func_id != BPF_FUNC_tail_call &&
2365 func_id != BPF_FUNC_map_lookup_elem)
2367 if (meta->map_ptr == NULL) {
2368 verbose(env, "kernel subsystem misconfigured verifier\n");
2372 if (!BPF_MAP_PTR(aux->map_state))
2373 bpf_map_ptr_store(aux, meta->map_ptr,
2374 meta->map_ptr->unpriv_array);
2375 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2376 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2377 meta->map_ptr->unpriv_array);
2381 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2383 const struct bpf_func_proto *fn = NULL;
2384 struct bpf_reg_state *regs;
2385 struct bpf_call_arg_meta meta;
2389 /* find function prototype */
2390 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2391 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2396 if (env->ops->get_func_proto)
2397 fn = env->ops->get_func_proto(func_id, env->prog);
2399 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2404 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2405 if (!env->prog->gpl_compatible && fn->gpl_only) {
2406 verbose(env, "cannot call GPL only function from proprietary program\n");
2410 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2411 changes_data = bpf_helper_changes_pkt_data(fn->func);
2412 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2413 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2414 func_id_name(func_id), func_id);
2418 memset(&meta, 0, sizeof(meta));
2419 meta.pkt_access = fn->pkt_access;
2421 err = check_func_proto(fn);
2423 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2424 func_id_name(func_id), func_id);
2429 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2432 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2435 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2438 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2441 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2445 err = record_func_map(env, &meta, func_id, insn_idx);
2449 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2450 * is inferred from register state.
2452 for (i = 0; i < meta.access_size; i++) {
2453 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2454 BPF_WRITE, -1, false);
2459 regs = cur_regs(env);
2460 /* reset caller saved regs */
2461 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2462 mark_reg_not_init(env, regs, caller_saved[i]);
2463 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2466 /* update return register (already marked as written above) */
2467 if (fn->ret_type == RET_INTEGER) {
2468 /* sets type to SCALAR_VALUE */
2469 mark_reg_unknown(env, regs, BPF_REG_0);
2470 } else if (fn->ret_type == RET_VOID) {
2471 regs[BPF_REG_0].type = NOT_INIT;
2472 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
2473 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2474 /* There is no offset yet applied, variable or fixed */
2475 mark_reg_known_zero(env, regs, BPF_REG_0);
2476 regs[BPF_REG_0].off = 0;
2477 /* remember map_ptr, so that check_map_access()
2478 * can check 'value_size' boundary of memory access
2479 * to map element returned from bpf_map_lookup_elem()
2481 if (meta.map_ptr == NULL) {
2483 "kernel subsystem misconfigured verifier\n");
2486 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2487 regs[BPF_REG_0].id = ++env->id_gen;
2489 verbose(env, "unknown return type %d of func %s#%d\n",
2490 fn->ret_type, func_id_name(func_id), func_id);
2494 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2499 clear_all_pkt_pointers(env);
2503 static bool signed_add_overflows(s64 a, s64 b)
2505 /* Do the add in u64, where overflow is well-defined */
2506 s64 res = (s64)((u64)a + (u64)b);
2513 static bool signed_sub_overflows(s64 a, s64 b)
2515 /* Do the sub in u64, where overflow is well-defined */
2516 s64 res = (s64)((u64)a - (u64)b);
2523 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2524 const struct bpf_reg_state *reg,
2525 enum bpf_reg_type type)
2527 bool known = tnum_is_const(reg->var_off);
2528 s64 val = reg->var_off.value;
2529 s64 smin = reg->smin_value;
2531 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2532 verbose(env, "math between %s pointer and %lld is not allowed\n",
2533 reg_type_str[type], val);
2537 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2538 verbose(env, "%s pointer offset %d is not allowed\n",
2539 reg_type_str[type], reg->off);
2543 if (smin == S64_MIN) {
2544 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2545 reg_type_str[type]);
2549 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2550 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2551 smin, reg_type_str[type]);
2558 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2559 * Caller should also handle BPF_MOV case separately.
2560 * If we return -EACCES, caller may want to try again treating pointer as a
2561 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2563 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2564 struct bpf_insn *insn,
2565 const struct bpf_reg_state *ptr_reg,
2566 const struct bpf_reg_state *off_reg)
2568 struct bpf_verifier_state *vstate = env->cur_state;
2569 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2570 struct bpf_reg_state *regs = state->regs, *dst_reg;
2571 bool known = tnum_is_const(off_reg->var_off);
2572 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2573 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2574 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2575 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2576 u8 opcode = BPF_OP(insn->code);
2577 u32 dst = insn->dst_reg;
2579 dst_reg = ®s[dst];
2581 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2582 smin_val > smax_val || umin_val > umax_val) {
2583 /* Taint dst register if offset had invalid bounds derived from
2584 * e.g. dead branches.
2586 __mark_reg_unknown(dst_reg);
2590 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2591 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2593 "R%d 32-bit pointer arithmetic prohibited\n",
2598 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2599 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2603 if (ptr_reg->type == CONST_PTR_TO_MAP) {
2604 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2608 if (ptr_reg->type == PTR_TO_PACKET_END) {
2609 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2614 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2615 * The id may be overwritten later if we create a new variable offset.
2617 dst_reg->type = ptr_reg->type;
2618 dst_reg->id = ptr_reg->id;
2620 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
2621 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
2626 /* We can take a fixed offset as long as it doesn't overflow
2627 * the s32 'off' field
2629 if (known && (ptr_reg->off + smin_val ==
2630 (s64)(s32)(ptr_reg->off + smin_val))) {
2631 /* pointer += K. Accumulate it into fixed offset */
2632 dst_reg->smin_value = smin_ptr;
2633 dst_reg->smax_value = smax_ptr;
2634 dst_reg->umin_value = umin_ptr;
2635 dst_reg->umax_value = umax_ptr;
2636 dst_reg->var_off = ptr_reg->var_off;
2637 dst_reg->off = ptr_reg->off + smin_val;
2638 dst_reg->range = ptr_reg->range;
2641 /* A new variable offset is created. Note that off_reg->off
2642 * == 0, since it's a scalar.
2643 * dst_reg gets the pointer type and since some positive
2644 * integer value was added to the pointer, give it a new 'id'
2645 * if it's a PTR_TO_PACKET.
2646 * this creates a new 'base' pointer, off_reg (variable) gets
2647 * added into the variable offset, and we copy the fixed offset
2650 if (signed_add_overflows(smin_ptr, smin_val) ||
2651 signed_add_overflows(smax_ptr, smax_val)) {
2652 dst_reg->smin_value = S64_MIN;
2653 dst_reg->smax_value = S64_MAX;
2655 dst_reg->smin_value = smin_ptr + smin_val;
2656 dst_reg->smax_value = smax_ptr + smax_val;
2658 if (umin_ptr + umin_val < umin_ptr ||
2659 umax_ptr + umax_val < umax_ptr) {
2660 dst_reg->umin_value = 0;
2661 dst_reg->umax_value = U64_MAX;
2663 dst_reg->umin_value = umin_ptr + umin_val;
2664 dst_reg->umax_value = umax_ptr + umax_val;
2666 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
2667 dst_reg->off = ptr_reg->off;
2668 if (reg_is_pkt_pointer(ptr_reg)) {
2669 dst_reg->id = ++env->id_gen;
2670 /* something was added to pkt_ptr, set range to zero */
2675 if (dst_reg == off_reg) {
2676 /* scalar -= pointer. Creates an unknown scalar */
2677 verbose(env, "R%d tried to subtract pointer from scalar\n",
2681 /* We don't allow subtraction from FP, because (according to
2682 * test_verifier.c test "invalid fp arithmetic", JITs might not
2683 * be able to deal with it.
2685 if (ptr_reg->type == PTR_TO_STACK) {
2686 verbose(env, "R%d subtraction from stack pointer prohibited\n",
2690 if (known && (ptr_reg->off - smin_val ==
2691 (s64)(s32)(ptr_reg->off - smin_val))) {
2692 /* pointer -= K. Subtract it from fixed offset */
2693 dst_reg->smin_value = smin_ptr;
2694 dst_reg->smax_value = smax_ptr;
2695 dst_reg->umin_value = umin_ptr;
2696 dst_reg->umax_value = umax_ptr;
2697 dst_reg->var_off = ptr_reg->var_off;
2698 dst_reg->id = ptr_reg->id;
2699 dst_reg->off = ptr_reg->off - smin_val;
2700 dst_reg->range = ptr_reg->range;
2703 /* A new variable offset is created. If the subtrahend is known
2704 * nonnegative, then any reg->range we had before is still good.
2706 if (signed_sub_overflows(smin_ptr, smax_val) ||
2707 signed_sub_overflows(smax_ptr, smin_val)) {
2708 /* Overflow possible, we know nothing */
2709 dst_reg->smin_value = S64_MIN;
2710 dst_reg->smax_value = S64_MAX;
2712 dst_reg->smin_value = smin_ptr - smax_val;
2713 dst_reg->smax_value = smax_ptr - smin_val;
2715 if (umin_ptr < umax_val) {
2716 /* Overflow possible, we know nothing */
2717 dst_reg->umin_value = 0;
2718 dst_reg->umax_value = U64_MAX;
2720 /* Cannot overflow (as long as bounds are consistent) */
2721 dst_reg->umin_value = umin_ptr - umax_val;
2722 dst_reg->umax_value = umax_ptr - umin_val;
2724 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2725 dst_reg->off = ptr_reg->off;
2726 if (reg_is_pkt_pointer(ptr_reg)) {
2727 dst_reg->id = ++env->id_gen;
2728 /* something was added to pkt_ptr, set range to zero */
2736 /* bitwise ops on pointers are troublesome, prohibit. */
2737 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2738 dst, bpf_alu_string[opcode >> 4]);
2741 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2742 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2743 dst, bpf_alu_string[opcode >> 4]);
2747 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2750 __update_reg_bounds(dst_reg);
2751 __reg_deduce_bounds(dst_reg);
2752 __reg_bound_offset(dst_reg);
2756 /* WARNING: This function does calculations on 64-bit values, but the actual
2757 * execution may occur on 32-bit values. Therefore, things like bitshifts
2758 * need extra checks in the 32-bit case.
2760 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2761 struct bpf_insn *insn,
2762 struct bpf_reg_state *dst_reg,
2763 struct bpf_reg_state src_reg)
2765 struct bpf_reg_state *regs = cur_regs(env);
2766 u8 opcode = BPF_OP(insn->code);
2767 bool src_known, dst_known;
2768 s64 smin_val, smax_val;
2769 u64 umin_val, umax_val;
2770 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2772 smin_val = src_reg.smin_value;
2773 smax_val = src_reg.smax_value;
2774 umin_val = src_reg.umin_value;
2775 umax_val = src_reg.umax_value;
2776 src_known = tnum_is_const(src_reg.var_off);
2777 dst_known = tnum_is_const(dst_reg->var_off);
2779 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2780 smin_val > smax_val || umin_val > umax_val) {
2781 /* Taint dst register if offset had invalid bounds derived from
2782 * e.g. dead branches.
2784 __mark_reg_unknown(dst_reg);
2789 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2790 __mark_reg_unknown(dst_reg);
2796 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2797 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2798 dst_reg->smin_value = S64_MIN;
2799 dst_reg->smax_value = S64_MAX;
2801 dst_reg->smin_value += smin_val;
2802 dst_reg->smax_value += smax_val;
2804 if (dst_reg->umin_value + umin_val < umin_val ||
2805 dst_reg->umax_value + umax_val < umax_val) {
2806 dst_reg->umin_value = 0;
2807 dst_reg->umax_value = U64_MAX;
2809 dst_reg->umin_value += umin_val;
2810 dst_reg->umax_value += umax_val;
2812 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2815 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2816 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2817 /* Overflow possible, we know nothing */
2818 dst_reg->smin_value = S64_MIN;
2819 dst_reg->smax_value = S64_MAX;
2821 dst_reg->smin_value -= smax_val;
2822 dst_reg->smax_value -= smin_val;
2824 if (dst_reg->umin_value < umax_val) {
2825 /* Overflow possible, we know nothing */
2826 dst_reg->umin_value = 0;
2827 dst_reg->umax_value = U64_MAX;
2829 /* Cannot overflow (as long as bounds are consistent) */
2830 dst_reg->umin_value -= umax_val;
2831 dst_reg->umax_value -= umin_val;
2833 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2836 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2837 if (smin_val < 0 || dst_reg->smin_value < 0) {
2838 /* Ain't nobody got time to multiply that sign */
2839 __mark_reg_unbounded(dst_reg);
2840 __update_reg_bounds(dst_reg);
2843 /* Both values are positive, so we can work with unsigned and
2844 * copy the result to signed (unless it exceeds S64_MAX).
2846 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2847 /* Potential overflow, we know nothing */
2848 __mark_reg_unbounded(dst_reg);
2849 /* (except what we can learn from the var_off) */
2850 __update_reg_bounds(dst_reg);
2853 dst_reg->umin_value *= umin_val;
2854 dst_reg->umax_value *= umax_val;
2855 if (dst_reg->umax_value > S64_MAX) {
2856 /* Overflow possible, we know nothing */
2857 dst_reg->smin_value = S64_MIN;
2858 dst_reg->smax_value = S64_MAX;
2860 dst_reg->smin_value = dst_reg->umin_value;
2861 dst_reg->smax_value = dst_reg->umax_value;
2865 if (src_known && dst_known) {
2866 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2867 src_reg.var_off.value);
2870 /* We get our minimum from the var_off, since that's inherently
2871 * bitwise. Our maximum is the minimum of the operands' maxima.
2873 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2874 dst_reg->umin_value = dst_reg->var_off.value;
2875 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2876 if (dst_reg->smin_value < 0 || smin_val < 0) {
2877 /* Lose signed bounds when ANDing negative numbers,
2878 * ain't nobody got time for that.
2880 dst_reg->smin_value = S64_MIN;
2881 dst_reg->smax_value = S64_MAX;
2883 /* ANDing two positives gives a positive, so safe to
2884 * cast result into s64.
2886 dst_reg->smin_value = dst_reg->umin_value;
2887 dst_reg->smax_value = dst_reg->umax_value;
2889 /* We may learn something more from the var_off */
2890 __update_reg_bounds(dst_reg);
2893 if (src_known && dst_known) {
2894 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2895 src_reg.var_off.value);
2898 /* We get our maximum from the var_off, and our minimum is the
2899 * maximum of the operands' minima
2901 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2902 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2903 dst_reg->umax_value = dst_reg->var_off.value |
2904 dst_reg->var_off.mask;
2905 if (dst_reg->smin_value < 0 || smin_val < 0) {
2906 /* Lose signed bounds when ORing negative numbers,
2907 * ain't nobody got time for that.
2909 dst_reg->smin_value = S64_MIN;
2910 dst_reg->smax_value = S64_MAX;
2912 /* ORing two positives gives a positive, so safe to
2913 * cast result into s64.
2915 dst_reg->smin_value = dst_reg->umin_value;
2916 dst_reg->smax_value = dst_reg->umax_value;
2918 /* We may learn something more from the var_off */
2919 __update_reg_bounds(dst_reg);
2922 if (umax_val >= insn_bitness) {
2923 /* Shifts greater than 31 or 63 are undefined.
2924 * This includes shifts by a negative number.
2926 mark_reg_unknown(env, regs, insn->dst_reg);
2929 /* We lose all sign bit information (except what we can pick
2932 dst_reg->smin_value = S64_MIN;
2933 dst_reg->smax_value = S64_MAX;
2934 /* If we might shift our top bit out, then we know nothing */
2935 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2936 dst_reg->umin_value = 0;
2937 dst_reg->umax_value = U64_MAX;
2939 dst_reg->umin_value <<= umin_val;
2940 dst_reg->umax_value <<= umax_val;
2943 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2945 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2946 /* We may learn something more from the var_off */
2947 __update_reg_bounds(dst_reg);
2950 if (umax_val >= insn_bitness) {
2951 /* Shifts greater than 31 or 63 are undefined.
2952 * This includes shifts by a negative number.
2954 mark_reg_unknown(env, regs, insn->dst_reg);
2957 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2958 * be negative, then either:
2959 * 1) src_reg might be zero, so the sign bit of the result is
2960 * unknown, so we lose our signed bounds
2961 * 2) it's known negative, thus the unsigned bounds capture the
2963 * 3) the signed bounds cross zero, so they tell us nothing
2965 * If the value in dst_reg is known nonnegative, then again the
2966 * unsigned bounts capture the signed bounds.
2967 * Thus, in all cases it suffices to blow away our signed bounds
2968 * and rely on inferring new ones from the unsigned bounds and
2969 * var_off of the result.
2971 dst_reg->smin_value = S64_MIN;
2972 dst_reg->smax_value = S64_MAX;
2974 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2977 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2978 dst_reg->umin_value >>= umax_val;
2979 dst_reg->umax_value >>= umin_val;
2980 /* We may learn something more from the var_off */
2981 __update_reg_bounds(dst_reg);
2984 mark_reg_unknown(env, regs, insn->dst_reg);
2988 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2989 /* 32-bit ALU ops are (32,32)->32 */
2990 coerce_reg_to_size(dst_reg, 4);
2991 coerce_reg_to_size(&src_reg, 4);
2994 __reg_deduce_bounds(dst_reg);
2995 __reg_bound_offset(dst_reg);
2999 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3002 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3003 struct bpf_insn *insn)
3005 struct bpf_verifier_state *vstate = env->cur_state;
3006 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3007 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3008 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3009 u8 opcode = BPF_OP(insn->code);
3011 dst_reg = ®s[insn->dst_reg];
3013 if (dst_reg->type != SCALAR_VALUE)
3015 if (BPF_SRC(insn->code) == BPF_X) {
3016 src_reg = ®s[insn->src_reg];
3017 if (src_reg->type != SCALAR_VALUE) {
3018 if (dst_reg->type != SCALAR_VALUE) {
3019 /* Combining two pointers by any ALU op yields
3020 * an arbitrary scalar. Disallow all math except
3021 * pointer subtraction
3023 if (opcode == BPF_SUB){
3024 mark_reg_unknown(env, regs, insn->dst_reg);
3027 verbose(env, "R%d pointer %s pointer prohibited\n",
3029 bpf_alu_string[opcode >> 4]);
3032 /* scalar += pointer
3033 * This is legal, but we have to reverse our
3034 * src/dest handling in computing the range
3036 return adjust_ptr_min_max_vals(env, insn,
3039 } else if (ptr_reg) {
3040 /* pointer += scalar */
3041 return adjust_ptr_min_max_vals(env, insn,
3045 /* Pretend the src is a reg with a known value, since we only
3046 * need to be able to read from this state.
3048 off_reg.type = SCALAR_VALUE;
3049 __mark_reg_known(&off_reg, insn->imm);
3051 if (ptr_reg) /* pointer += K */
3052 return adjust_ptr_min_max_vals(env, insn,
3056 /* Got here implies adding two SCALAR_VALUEs */
3057 if (WARN_ON_ONCE(ptr_reg)) {
3058 print_verifier_state(env, state);
3059 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3062 if (WARN_ON(!src_reg)) {
3063 print_verifier_state(env, state);
3064 verbose(env, "verifier internal error: no src_reg\n");
3067 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3070 /* check validity of 32-bit and 64-bit arithmetic operations */
3071 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3073 struct bpf_reg_state *regs = cur_regs(env);
3074 u8 opcode = BPF_OP(insn->code);
3077 if (opcode == BPF_END || opcode == BPF_NEG) {
3078 if (opcode == BPF_NEG) {
3079 if (BPF_SRC(insn->code) != 0 ||
3080 insn->src_reg != BPF_REG_0 ||
3081 insn->off != 0 || insn->imm != 0) {
3082 verbose(env, "BPF_NEG uses reserved fields\n");
3086 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3087 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3088 BPF_CLASS(insn->code) == BPF_ALU64) {
3089 verbose(env, "BPF_END uses reserved fields\n");
3094 /* check src operand */
3095 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3099 if (is_pointer_value(env, insn->dst_reg)) {
3100 verbose(env, "R%d pointer arithmetic prohibited\n",
3105 /* check dest operand */
3106 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3110 } else if (opcode == BPF_MOV) {
3112 if (BPF_SRC(insn->code) == BPF_X) {
3113 if (insn->imm != 0 || insn->off != 0) {
3114 verbose(env, "BPF_MOV uses reserved fields\n");
3118 /* check src operand */
3119 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3123 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3124 verbose(env, "BPF_MOV uses reserved fields\n");
3129 /* check dest operand */
3130 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3134 if (BPF_SRC(insn->code) == BPF_X) {
3135 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3137 * copy register state to dest reg
3139 regs[insn->dst_reg] = regs[insn->src_reg];
3140 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3143 if (is_pointer_value(env, insn->src_reg)) {
3145 "R%d partial copy of pointer\n",
3149 mark_reg_unknown(env, regs, insn->dst_reg);
3150 coerce_reg_to_size(®s[insn->dst_reg], 4);
3154 * remember the value we stored into this reg
3156 regs[insn->dst_reg].type = SCALAR_VALUE;
3157 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3158 __mark_reg_known(regs + insn->dst_reg,
3161 __mark_reg_known(regs + insn->dst_reg,
3166 } else if (opcode > BPF_END) {
3167 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3170 } else { /* all other ALU ops: and, sub, xor, add, ... */
3172 if (BPF_SRC(insn->code) == BPF_X) {
3173 if (insn->imm != 0 || insn->off != 0) {
3174 verbose(env, "BPF_ALU uses reserved fields\n");
3177 /* check src1 operand */
3178 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3182 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3183 verbose(env, "BPF_ALU uses reserved fields\n");
3188 /* check src2 operand */
3189 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3193 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3194 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3195 verbose(env, "div by zero\n");
3199 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3200 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3204 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3205 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3206 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3208 if (insn->imm < 0 || insn->imm >= size) {
3209 verbose(env, "invalid shift %d\n", insn->imm);
3214 /* check dest operand */
3215 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3219 return adjust_reg_min_max_vals(env, insn);
3225 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3226 struct bpf_reg_state *dst_reg,
3227 enum bpf_reg_type type,
3228 bool range_right_open)
3230 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3231 struct bpf_reg_state *regs = state->regs, *reg;
3235 if (dst_reg->off < 0 ||
3236 (dst_reg->off == 0 && range_right_open))
3237 /* This doesn't give us any range */
3240 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3241 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3242 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3243 * than pkt_end, but that's because it's also less than pkt.
3247 new_range = dst_reg->off;
3248 if (range_right_open)
3251 /* Examples for register markings:
3253 * pkt_data in dst register:
3257 * if (r2 > pkt_end) goto <handle exception>
3262 * if (r2 < pkt_end) goto <access okay>
3263 * <handle exception>
3266 * r2 == dst_reg, pkt_end == src_reg
3267 * r2=pkt(id=n,off=8,r=0)
3268 * r3=pkt(id=n,off=0,r=0)
3270 * pkt_data in src register:
3274 * if (pkt_end >= r2) goto <access okay>
3275 * <handle exception>
3279 * if (pkt_end <= r2) goto <handle exception>
3283 * pkt_end == dst_reg, r2 == src_reg
3284 * r2=pkt(id=n,off=8,r=0)
3285 * r3=pkt(id=n,off=0,r=0)
3287 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3288 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3289 * and [r3, r3 + 8-1) respectively is safe to access depending on
3293 /* If our ids match, then we must have the same max_value. And we
3294 * don't care about the other reg's fixed offset, since if it's too big
3295 * the range won't allow anything.
3296 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3298 for (i = 0; i < MAX_BPF_REG; i++)
3299 if (regs[i].type == type && regs[i].id == dst_reg->id)
3300 /* keep the maximum range already checked */
3301 regs[i].range = max(regs[i].range, new_range);
3303 for (j = 0; j <= vstate->curframe; j++) {
3304 state = vstate->frame[j];
3305 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3306 if (state->stack[i].slot_type[0] != STACK_SPILL)
3308 reg = &state->stack[i].spilled_ptr;
3309 if (reg->type == type && reg->id == dst_reg->id)
3310 reg->range = max(reg->range, new_range);
3315 /* Adjusts the register min/max values in the case that the dst_reg is the
3316 * variable register that we are working on, and src_reg is a constant or we're
3317 * simply doing a BPF_K check.
3318 * In JEQ/JNE cases we also adjust the var_off values.
3320 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3321 struct bpf_reg_state *false_reg, u64 val,
3324 /* If the dst_reg is a pointer, we can't learn anything about its
3325 * variable offset from the compare (unless src_reg were a pointer into
3326 * the same object, but we don't bother with that.
3327 * Since false_reg and true_reg have the same type by construction, we
3328 * only need to check one of them for pointerness.
3330 if (__is_pointer_value(false, false_reg))
3335 /* If this is false then we know nothing Jon Snow, but if it is
3336 * true then we know for sure.
3338 __mark_reg_known(true_reg, val);
3341 /* If this is true we know nothing Jon Snow, but if it is false
3342 * we know the value for sure;
3344 __mark_reg_known(false_reg, val);
3347 false_reg->umax_value = min(false_reg->umax_value, val);
3348 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3351 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3352 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3355 false_reg->umin_value = max(false_reg->umin_value, val);
3356 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3359 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3360 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3363 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3364 true_reg->umin_value = max(true_reg->umin_value, val);
3367 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3368 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3371 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3372 true_reg->umax_value = min(true_reg->umax_value, val);
3375 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3376 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3382 __reg_deduce_bounds(false_reg);
3383 __reg_deduce_bounds(true_reg);
3384 /* We might have learned some bits from the bounds. */
3385 __reg_bound_offset(false_reg);
3386 __reg_bound_offset(true_reg);
3387 /* Intersecting with the old var_off might have improved our bounds
3388 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3389 * then new var_off is (0; 0x7f...fc) which improves our umax.
3391 __update_reg_bounds(false_reg);
3392 __update_reg_bounds(true_reg);
3395 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3398 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3399 struct bpf_reg_state *false_reg, u64 val,
3402 if (__is_pointer_value(false, false_reg))
3407 /* If this is false then we know nothing Jon Snow, but if it is
3408 * true then we know for sure.
3410 __mark_reg_known(true_reg, val);
3413 /* If this is true we know nothing Jon Snow, but if it is false
3414 * we know the value for sure;
3416 __mark_reg_known(false_reg, val);
3419 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3420 false_reg->umin_value = max(false_reg->umin_value, val);
3423 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3424 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3427 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3428 false_reg->umax_value = min(false_reg->umax_value, val);
3431 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3432 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3435 true_reg->umax_value = min(true_reg->umax_value, val);
3436 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3439 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3440 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3443 true_reg->umin_value = max(true_reg->umin_value, val);
3444 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3447 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3448 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3454 __reg_deduce_bounds(false_reg);
3455 __reg_deduce_bounds(true_reg);
3456 /* We might have learned some bits from the bounds. */
3457 __reg_bound_offset(false_reg);
3458 __reg_bound_offset(true_reg);
3459 /* Intersecting with the old var_off might have improved our bounds
3460 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3461 * then new var_off is (0; 0x7f...fc) which improves our umax.
3463 __update_reg_bounds(false_reg);
3464 __update_reg_bounds(true_reg);
3467 /* Regs are known to be equal, so intersect their min/max/var_off */
3468 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3469 struct bpf_reg_state *dst_reg)
3471 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3472 dst_reg->umin_value);
3473 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3474 dst_reg->umax_value);
3475 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3476 dst_reg->smin_value);
3477 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3478 dst_reg->smax_value);
3479 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3481 /* We might have learned new bounds from the var_off. */
3482 __update_reg_bounds(src_reg);
3483 __update_reg_bounds(dst_reg);
3484 /* We might have learned something about the sign bit. */
3485 __reg_deduce_bounds(src_reg);
3486 __reg_deduce_bounds(dst_reg);
3487 /* We might have learned some bits from the bounds. */
3488 __reg_bound_offset(src_reg);
3489 __reg_bound_offset(dst_reg);
3490 /* Intersecting with the old var_off might have improved our bounds
3491 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3492 * then new var_off is (0; 0x7f...fc) which improves our umax.
3494 __update_reg_bounds(src_reg);
3495 __update_reg_bounds(dst_reg);
3498 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3499 struct bpf_reg_state *true_dst,
3500 struct bpf_reg_state *false_src,
3501 struct bpf_reg_state *false_dst,
3506 __reg_combine_min_max(true_src, true_dst);
3509 __reg_combine_min_max(false_src, false_dst);
3514 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
3517 struct bpf_reg_state *reg = ®s[regno];
3519 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
3520 /* Old offset (both fixed and variable parts) should
3521 * have been known-zero, because we don't allow pointer
3522 * arithmetic on pointers that might be NULL.
3524 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3525 !tnum_equals_const(reg->var_off, 0) ||
3527 __mark_reg_known_zero(reg);
3531 reg->type = SCALAR_VALUE;
3532 } else if (reg->map_ptr->inner_map_meta) {
3533 reg->type = CONST_PTR_TO_MAP;
3534 reg->map_ptr = reg->map_ptr->inner_map_meta;
3536 reg->type = PTR_TO_MAP_VALUE;
3538 /* We don't need id from this point onwards anymore, thus we
3539 * should better reset it, so that state pruning has chances
3546 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3547 * be folded together at some point.
3549 static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno,
3552 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3553 struct bpf_reg_state *regs = state->regs;
3554 u32 id = regs[regno].id;
3557 for (i = 0; i < MAX_BPF_REG; i++)
3558 mark_map_reg(regs, i, id, is_null);
3560 for (j = 0; j <= vstate->curframe; j++) {
3561 state = vstate->frame[j];
3562 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3563 if (state->stack[i].slot_type[0] != STACK_SPILL)
3565 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
3570 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3571 struct bpf_reg_state *dst_reg,
3572 struct bpf_reg_state *src_reg,
3573 struct bpf_verifier_state *this_branch,
3574 struct bpf_verifier_state *other_branch)
3576 if (BPF_SRC(insn->code) != BPF_X)
3579 switch (BPF_OP(insn->code)) {
3581 if ((dst_reg->type == PTR_TO_PACKET &&
3582 src_reg->type == PTR_TO_PACKET_END) ||
3583 (dst_reg->type == PTR_TO_PACKET_META &&
3584 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3585 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3586 find_good_pkt_pointers(this_branch, dst_reg,
3587 dst_reg->type, false);
3588 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3589 src_reg->type == PTR_TO_PACKET) ||
3590 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3591 src_reg->type == PTR_TO_PACKET_META)) {
3592 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3593 find_good_pkt_pointers(other_branch, src_reg,
3594 src_reg->type, true);
3600 if ((dst_reg->type == PTR_TO_PACKET &&
3601 src_reg->type == PTR_TO_PACKET_END) ||
3602 (dst_reg->type == PTR_TO_PACKET_META &&
3603 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3604 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3605 find_good_pkt_pointers(other_branch, dst_reg,
3606 dst_reg->type, true);
3607 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3608 src_reg->type == PTR_TO_PACKET) ||
3609 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3610 src_reg->type == PTR_TO_PACKET_META)) {
3611 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3612 find_good_pkt_pointers(this_branch, src_reg,
3613 src_reg->type, false);
3619 if ((dst_reg->type == PTR_TO_PACKET &&
3620 src_reg->type == PTR_TO_PACKET_END) ||
3621 (dst_reg->type == PTR_TO_PACKET_META &&
3622 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3623 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3624 find_good_pkt_pointers(this_branch, dst_reg,
3625 dst_reg->type, true);
3626 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3627 src_reg->type == PTR_TO_PACKET) ||
3628 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3629 src_reg->type == PTR_TO_PACKET_META)) {
3630 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3631 find_good_pkt_pointers(other_branch, src_reg,
3632 src_reg->type, false);
3638 if ((dst_reg->type == PTR_TO_PACKET &&
3639 src_reg->type == PTR_TO_PACKET_END) ||
3640 (dst_reg->type == PTR_TO_PACKET_META &&
3641 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3642 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3643 find_good_pkt_pointers(other_branch, dst_reg,
3644 dst_reg->type, false);
3645 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3646 src_reg->type == PTR_TO_PACKET) ||
3647 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3648 src_reg->type == PTR_TO_PACKET_META)) {
3649 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3650 find_good_pkt_pointers(this_branch, src_reg,
3651 src_reg->type, true);
3663 static int check_cond_jmp_op(struct bpf_verifier_env *env,
3664 struct bpf_insn *insn, int *insn_idx)
3666 struct bpf_verifier_state *this_branch = env->cur_state;
3667 struct bpf_verifier_state *other_branch;
3668 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
3669 struct bpf_reg_state *dst_reg, *other_branch_regs;
3670 u8 opcode = BPF_OP(insn->code);
3673 if (opcode > BPF_JSLE) {
3674 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
3678 if (BPF_SRC(insn->code) == BPF_X) {
3679 if (insn->imm != 0) {
3680 verbose(env, "BPF_JMP uses reserved fields\n");
3684 /* check src1 operand */
3685 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3689 if (is_pointer_value(env, insn->src_reg)) {
3690 verbose(env, "R%d pointer comparison prohibited\n",
3695 if (insn->src_reg != BPF_REG_0) {
3696 verbose(env, "BPF_JMP uses reserved fields\n");
3701 /* check src2 operand */
3702 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3706 dst_reg = ®s[insn->dst_reg];
3708 /* detect if R == 0 where R was initialized to zero earlier */
3709 if (BPF_SRC(insn->code) == BPF_K &&
3710 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3711 dst_reg->type == SCALAR_VALUE &&
3712 tnum_is_const(dst_reg->var_off)) {
3713 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
3714 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
3715 /* if (imm == imm) goto pc+off;
3716 * only follow the goto, ignore fall-through
3718 *insn_idx += insn->off;
3721 /* if (imm != imm) goto pc+off;
3722 * only follow fall-through branch, since
3723 * that's where the program will go
3729 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3732 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
3734 /* detect if we are comparing against a constant value so we can adjust
3735 * our min/max values for our dst register.
3736 * this is only legit if both are scalars (or pointers to the same
3737 * object, I suppose, but we don't support that right now), because
3738 * otherwise the different base pointers mean the offsets aren't
3741 if (BPF_SRC(insn->code) == BPF_X) {
3742 if (dst_reg->type == SCALAR_VALUE &&
3743 regs[insn->src_reg].type == SCALAR_VALUE) {
3744 if (tnum_is_const(regs[insn->src_reg].var_off))
3745 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3746 dst_reg, regs[insn->src_reg].var_off.value,
3748 else if (tnum_is_const(dst_reg->var_off))
3749 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
3750 ®s[insn->src_reg],
3751 dst_reg->var_off.value, opcode);
3752 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3753 /* Comparing for equality, we can combine knowledge */
3754 reg_combine_min_max(&other_branch_regs[insn->src_reg],
3755 &other_branch_regs[insn->dst_reg],
3756 ®s[insn->src_reg],
3757 ®s[insn->dst_reg], opcode);
3759 } else if (dst_reg->type == SCALAR_VALUE) {
3760 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3761 dst_reg, insn->imm, opcode);
3764 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3765 if (BPF_SRC(insn->code) == BPF_K &&
3766 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3767 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3768 /* Mark all identical map registers in each branch as either
3769 * safe or unknown depending R == 0 or R != 0 conditional.
3771 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3772 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3773 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3774 this_branch, other_branch) &&
3775 is_pointer_value(env, insn->dst_reg)) {
3776 verbose(env, "R%d pointer comparison prohibited\n",
3781 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
3785 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3786 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3788 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3790 return (struct bpf_map *) (unsigned long) imm64;
3793 /* verify BPF_LD_IMM64 instruction */
3794 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3796 struct bpf_reg_state *regs = cur_regs(env);
3799 if (BPF_SIZE(insn->code) != BPF_DW) {
3800 verbose(env, "invalid BPF_LD_IMM insn\n");
3803 if (insn->off != 0) {
3804 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3808 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3812 if (insn->src_reg == 0) {
3813 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3815 regs[insn->dst_reg].type = SCALAR_VALUE;
3816 __mark_reg_known(®s[insn->dst_reg], imm);
3820 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3821 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3823 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3824 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3828 static bool may_access_skb(enum bpf_prog_type type)
3831 case BPF_PROG_TYPE_SOCKET_FILTER:
3832 case BPF_PROG_TYPE_SCHED_CLS:
3833 case BPF_PROG_TYPE_SCHED_ACT:
3840 /* verify safety of LD_ABS|LD_IND instructions:
3841 * - they can only appear in the programs where ctx == skb
3842 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3843 * preserve R6-R9, and store return value into R0
3846 * ctx == skb == R6 == CTX
3849 * SRC == any register
3850 * IMM == 32-bit immediate
3853 * R0 - 8/16/32-bit skb data converted to cpu endianness
3855 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3857 struct bpf_reg_state *regs = cur_regs(env);
3858 u8 mode = BPF_MODE(insn->code);
3861 if (!may_access_skb(env->prog->type)) {
3862 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3866 if (env->subprog_cnt) {
3867 /* when program has LD_ABS insn JITs and interpreter assume
3868 * that r1 == ctx == skb which is not the case for callees
3869 * that can have arbitrary arguments. It's problematic
3870 * for main prog as well since JITs would need to analyze
3871 * all functions in order to make proper register save/restore
3872 * decisions in the main prog. Hence disallow LD_ABS with calls
3874 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
3878 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3879 BPF_SIZE(insn->code) == BPF_DW ||
3880 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3881 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3885 /* check whether implicit source operand (register R6) is readable */
3886 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3890 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3892 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3896 if (mode == BPF_IND) {
3897 /* check explicit source operand */
3898 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3903 /* reset caller saved regs to unreadable */
3904 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3905 mark_reg_not_init(env, regs, caller_saved[i]);
3906 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3909 /* mark destination R0 register as readable, since it contains
3910 * the value fetched from the packet.
3911 * Already marked as written above.
3913 mark_reg_unknown(env, regs, BPF_REG_0);
3917 static int check_return_code(struct bpf_verifier_env *env)
3919 struct bpf_reg_state *reg;
3920 struct tnum range = tnum_range(0, 1);
3922 switch (env->prog->type) {
3923 case BPF_PROG_TYPE_CGROUP_SKB:
3924 case BPF_PROG_TYPE_CGROUP_SOCK:
3925 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
3926 case BPF_PROG_TYPE_SOCK_OPS:
3927 case BPF_PROG_TYPE_CGROUP_DEVICE:
3933 reg = cur_regs(env) + BPF_REG_0;
3934 if (reg->type != SCALAR_VALUE) {
3935 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
3936 reg_type_str[reg->type]);
3940 if (!tnum_in(range, reg->var_off)) {
3941 verbose(env, "At program exit the register R0 ");
3942 if (!tnum_is_unknown(reg->var_off)) {
3945 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3946 verbose(env, "has value %s", tn_buf);
3948 verbose(env, "has unknown scalar value");
3950 verbose(env, " should have been 0 or 1\n");
3956 /* non-recursive DFS pseudo code
3957 * 1 procedure DFS-iterative(G,v):
3958 * 2 label v as discovered
3959 * 3 let S be a stack
3961 * 5 while S is not empty
3963 * 7 if t is what we're looking for:
3965 * 9 for all edges e in G.adjacentEdges(t) do
3966 * 10 if edge e is already labelled
3967 * 11 continue with the next edge
3968 * 12 w <- G.adjacentVertex(t,e)
3969 * 13 if vertex w is not discovered and not explored
3970 * 14 label e as tree-edge
3971 * 15 label w as discovered
3974 * 18 else if vertex w is discovered
3975 * 19 label e as back-edge
3977 * 21 // vertex w is explored
3978 * 22 label e as forward- or cross-edge
3979 * 23 label t as explored
3984 * 0x11 - discovered and fall-through edge labelled
3985 * 0x12 - discovered and fall-through and branch edges labelled
3996 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3998 static int *insn_stack; /* stack of insns to process */
3999 static int cur_stack; /* current stack index */
4000 static int *insn_state;
4002 /* t, w, e - match pseudo-code above:
4003 * t - index of current instruction
4004 * w - next instruction
4007 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4009 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4012 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4015 if (w < 0 || w >= env->prog->len) {
4016 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4021 /* mark branch target for state pruning */
4022 env->explored_states[w] = STATE_LIST_MARK;
4024 if (insn_state[w] == 0) {
4026 insn_state[t] = DISCOVERED | e;
4027 insn_state[w] = DISCOVERED;
4028 if (cur_stack >= env->prog->len)
4030 insn_stack[cur_stack++] = w;
4032 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4033 verbose(env, "back-edge from insn %d to %d\n", t, w);
4035 } else if (insn_state[w] == EXPLORED) {
4036 /* forward- or cross-edge */
4037 insn_state[t] = DISCOVERED | e;
4039 verbose(env, "insn state internal bug\n");
4045 /* non-recursive depth-first-search to detect loops in BPF program
4046 * loop == back-edge in directed graph
4048 static int check_cfg(struct bpf_verifier_env *env)
4050 struct bpf_insn *insns = env->prog->insnsi;
4051 int insn_cnt = env->prog->len;
4055 ret = check_subprogs(env);
4059 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4063 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4069 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4070 insn_stack[0] = 0; /* 0 is the first instruction */
4076 t = insn_stack[cur_stack - 1];
4078 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4079 u8 opcode = BPF_OP(insns[t].code);
4081 if (opcode == BPF_EXIT) {
4083 } else if (opcode == BPF_CALL) {
4084 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4089 if (t + 1 < insn_cnt)
4090 env->explored_states[t + 1] = STATE_LIST_MARK;
4091 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4092 env->explored_states[t] = STATE_LIST_MARK;
4093 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4099 } else if (opcode == BPF_JA) {
4100 if (BPF_SRC(insns[t].code) != BPF_K) {
4104 /* unconditional jump with single edge */
4105 ret = push_insn(t, t + insns[t].off + 1,
4111 /* tell verifier to check for equivalent states
4112 * after every call and jump
4114 if (t + 1 < insn_cnt)
4115 env->explored_states[t + 1] = STATE_LIST_MARK;
4117 /* conditional jump with two edges */
4118 env->explored_states[t] = STATE_LIST_MARK;
4119 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4125 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4132 /* all other non-branch instructions with single
4135 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4143 insn_state[t] = EXPLORED;
4144 if (cur_stack-- <= 0) {
4145 verbose(env, "pop stack internal bug\n");
4152 for (i = 0; i < insn_cnt; i++) {
4153 if (insn_state[i] != EXPLORED) {
4154 verbose(env, "unreachable insn %d\n", i);
4159 ret = 0; /* cfg looks good */
4167 /* check %cur's range satisfies %old's */
4168 static bool range_within(struct bpf_reg_state *old,
4169 struct bpf_reg_state *cur)
4171 return old->umin_value <= cur->umin_value &&
4172 old->umax_value >= cur->umax_value &&
4173 old->smin_value <= cur->smin_value &&
4174 old->smax_value >= cur->smax_value;
4177 /* Maximum number of register states that can exist at once */
4178 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4184 /* If in the old state two registers had the same id, then they need to have
4185 * the same id in the new state as well. But that id could be different from
4186 * the old state, so we need to track the mapping from old to new ids.
4187 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4188 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4189 * regs with a different old id could still have new id 9, we don't care about
4191 * So we look through our idmap to see if this old id has been seen before. If
4192 * so, we require the new id to match; otherwise, we add the id pair to the map.
4194 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4198 for (i = 0; i < ID_MAP_SIZE; i++) {
4199 if (!idmap[i].old) {
4200 /* Reached an empty slot; haven't seen this id before */
4201 idmap[i].old = old_id;
4202 idmap[i].cur = cur_id;
4205 if (idmap[i].old == old_id)
4206 return idmap[i].cur == cur_id;
4208 /* We ran out of idmap slots, which should be impossible */
4213 /* Returns true if (rold safe implies rcur safe) */
4214 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4215 struct idpair *idmap)
4219 if (!(rold->live & REG_LIVE_READ))
4220 /* explored state didn't use this */
4223 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, frameno)) == 0;
4225 if (rold->type == PTR_TO_STACK)
4226 /* two stack pointers are equal only if they're pointing to
4227 * the same stack frame, since fp-8 in foo != fp-8 in bar
4229 return equal && rold->frameno == rcur->frameno;
4234 if (rold->type == NOT_INIT)
4235 /* explored state can't have used this */
4237 if (rcur->type == NOT_INIT)
4239 switch (rold->type) {
4241 if (rcur->type == SCALAR_VALUE) {
4242 /* new val must satisfy old val knowledge */
4243 return range_within(rold, rcur) &&
4244 tnum_in(rold->var_off, rcur->var_off);
4246 /* We're trying to use a pointer in place of a scalar.
4247 * Even if the scalar was unbounded, this could lead to
4248 * pointer leaks because scalars are allowed to leak
4249 * while pointers are not. We could make this safe in
4250 * special cases if root is calling us, but it's
4251 * probably not worth the hassle.
4255 case PTR_TO_MAP_VALUE:
4256 /* If the new min/max/var_off satisfy the old ones and
4257 * everything else matches, we are OK.
4258 * We don't care about the 'id' value, because nothing
4259 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4261 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4262 range_within(rold, rcur) &&
4263 tnum_in(rold->var_off, rcur->var_off);
4264 case PTR_TO_MAP_VALUE_OR_NULL:
4265 /* a PTR_TO_MAP_VALUE could be safe to use as a
4266 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4267 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4268 * checked, doing so could have affected others with the same
4269 * id, and we can't check for that because we lost the id when
4270 * we converted to a PTR_TO_MAP_VALUE.
4272 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4274 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4276 /* Check our ids match any regs they're supposed to */
4277 return check_ids(rold->id, rcur->id, idmap);
4278 case PTR_TO_PACKET_META:
4280 if (rcur->type != rold->type)
4282 /* We must have at least as much range as the old ptr
4283 * did, so that any accesses which were safe before are
4284 * still safe. This is true even if old range < old off,
4285 * since someone could have accessed through (ptr - k), or
4286 * even done ptr -= k in a register, to get a safe access.
4288 if (rold->range > rcur->range)
4290 /* If the offsets don't match, we can't trust our alignment;
4291 * nor can we be sure that we won't fall out of range.
4293 if (rold->off != rcur->off)
4295 /* id relations must be preserved */
4296 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4298 /* new val must satisfy old val knowledge */
4299 return range_within(rold, rcur) &&
4300 tnum_in(rold->var_off, rcur->var_off);
4302 case CONST_PTR_TO_MAP:
4303 case PTR_TO_PACKET_END:
4304 /* Only valid matches are exact, which memcmp() above
4305 * would have accepted
4308 /* Don't know what's going on, just say it's not safe */
4312 /* Shouldn't get here; if we do, say it's not safe */
4317 static bool stacksafe(struct bpf_func_state *old,
4318 struct bpf_func_state *cur,
4319 struct idpair *idmap)
4323 /* if explored stack has more populated slots than current stack
4324 * such stacks are not equivalent
4326 if (old->allocated_stack > cur->allocated_stack)
4329 /* walk slots of the explored stack and ignore any additional
4330 * slots in the current stack, since explored(safe) state
4333 for (i = 0; i < old->allocated_stack; i++) {
4334 spi = i / BPF_REG_SIZE;
4336 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4337 /* explored state didn't use this */
4340 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4342 /* if old state was safe with misc data in the stack
4343 * it will be safe with zero-initialized stack.
4344 * The opposite is not true
4346 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4347 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4349 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4350 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4351 /* Ex: old explored (safe) state has STACK_SPILL in
4352 * this stack slot, but current has has STACK_MISC ->
4353 * this verifier states are not equivalent,
4354 * return false to continue verification of this path
4357 if (i % BPF_REG_SIZE)
4359 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4361 if (!regsafe(&old->stack[spi].spilled_ptr,
4362 &cur->stack[spi].spilled_ptr,
4364 /* when explored and current stack slot are both storing
4365 * spilled registers, check that stored pointers types
4366 * are the same as well.
4367 * Ex: explored safe path could have stored
4368 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4369 * but current path has stored:
4370 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4371 * such verifier states are not equivalent.
4372 * return false to continue verification of this path
4379 /* compare two verifier states
4381 * all states stored in state_list are known to be valid, since
4382 * verifier reached 'bpf_exit' instruction through them
4384 * this function is called when verifier exploring different branches of
4385 * execution popped from the state stack. If it sees an old state that has
4386 * more strict register state and more strict stack state then this execution
4387 * branch doesn't need to be explored further, since verifier already
4388 * concluded that more strict state leads to valid finish.
4390 * Therefore two states are equivalent if register state is more conservative
4391 * and explored stack state is more conservative than the current one.
4394 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4395 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4397 * In other words if current stack state (one being explored) has more
4398 * valid slots than old one that already passed validation, it means
4399 * the verifier can stop exploring and conclude that current state is valid too
4401 * Similarly with registers. If explored state has register type as invalid
4402 * whereas register type in current state is meaningful, it means that
4403 * the current state will reach 'bpf_exit' instruction safely
4405 static bool func_states_equal(struct bpf_func_state *old,
4406 struct bpf_func_state *cur)
4408 struct idpair *idmap;
4412 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4413 /* If we failed to allocate the idmap, just say it's not safe */
4417 for (i = 0; i < MAX_BPF_REG; i++) {
4418 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4422 if (!stacksafe(old, cur, idmap))
4430 static bool states_equal(struct bpf_verifier_env *env,
4431 struct bpf_verifier_state *old,
4432 struct bpf_verifier_state *cur)
4436 if (old->curframe != cur->curframe)
4439 /* for states to be equal callsites have to be the same
4440 * and all frame states need to be equivalent
4442 for (i = 0; i <= old->curframe; i++) {
4443 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4445 if (!func_states_equal(old->frame[i], cur->frame[i]))
4451 /* A write screens off any subsequent reads; but write marks come from the
4452 * straight-line code between a state and its parent. When we arrive at an
4453 * equivalent state (jump target or such) we didn't arrive by the straight-line
4454 * code, so read marks in the state must propagate to the parent regardless
4455 * of the state's write marks. That's what 'parent == state->parent' comparison
4456 * in mark_reg_read() and mark_stack_slot_read() is for.
4458 static int propagate_liveness(struct bpf_verifier_env *env,
4459 const struct bpf_verifier_state *vstate,
4460 struct bpf_verifier_state *vparent)
4462 int i, frame, err = 0;
4463 struct bpf_func_state *state, *parent;
4465 if (vparent->curframe != vstate->curframe) {
4466 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4467 vparent->curframe, vstate->curframe);
4470 /* Propagate read liveness of registers... */
4471 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4472 /* We don't need to worry about FP liveness because it's read-only */
4473 for (i = 0; i < BPF_REG_FP; i++) {
4474 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4476 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4477 err = mark_reg_read(env, vstate, vparent, i);
4483 /* ... and stack slots */
4484 for (frame = 0; frame <= vstate->curframe; frame++) {
4485 state = vstate->frame[frame];
4486 parent = vparent->frame[frame];
4487 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4488 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4489 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4491 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4492 mark_stack_slot_read(env, vstate, vparent, i, frame);
4498 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4500 struct bpf_verifier_state_list *new_sl;
4501 struct bpf_verifier_state_list *sl;
4502 struct bpf_verifier_state *cur = env->cur_state;
4505 sl = env->explored_states[insn_idx];
4507 /* this 'insn_idx' instruction wasn't marked, so we will not
4508 * be doing state search here
4512 while (sl != STATE_LIST_MARK) {
4513 if (states_equal(env, &sl->state, cur)) {
4514 /* reached equivalent register/stack state,
4516 * Registers read by the continuation are read by us.
4517 * If we have any write marks in env->cur_state, they
4518 * will prevent corresponding reads in the continuation
4519 * from reaching our parent (an explored_state). Our
4520 * own state will get the read marks recorded, but
4521 * they'll be immediately forgotten as we're pruning
4522 * this state and will pop a new one.
4524 err = propagate_liveness(env, &sl->state, cur);
4532 /* there were no equivalent states, remember current one.
4533 * technically the current state is not proven to be safe yet,
4534 * but it will either reach outer most bpf_exit (which means it's safe)
4535 * or it will be rejected. Since there are no loops, we won't be
4536 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4537 * again on the way to bpf_exit
4539 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
4543 /* add new state to the head of linked list */
4544 err = copy_verifier_state(&new_sl->state, cur);
4546 free_verifier_state(&new_sl->state, false);
4550 new_sl->next = env->explored_states[insn_idx];
4551 env->explored_states[insn_idx] = new_sl;
4552 /* connect new state to parentage chain */
4553 cur->parent = &new_sl->state;
4554 /* clear write marks in current state: the writes we did are not writes
4555 * our child did, so they don't screen off its reads from us.
4556 * (There are no read marks in current state, because reads always mark
4557 * their parent and current state never has children yet. Only
4558 * explored_states can get read marks.)
4560 for (i = 0; i < BPF_REG_FP; i++)
4561 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
4563 /* all stack frames are accessible from callee, clear them all */
4564 for (j = 0; j <= cur->curframe; j++) {
4565 struct bpf_func_state *frame = cur->frame[j];
4567 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++)
4568 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
4573 static int do_check(struct bpf_verifier_env *env)
4575 struct bpf_verifier_state *state;
4576 struct bpf_insn *insns = env->prog->insnsi;
4577 struct bpf_reg_state *regs;
4578 int insn_cnt = env->prog->len, i;
4579 int insn_idx, prev_insn_idx = 0;
4580 int insn_processed = 0;
4581 bool do_print_state = false;
4583 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
4586 state->curframe = 0;
4587 state->parent = NULL;
4588 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
4589 if (!state->frame[0]) {
4593 env->cur_state = state;
4594 init_func_state(env, state->frame[0],
4595 BPF_MAIN_FUNC /* callsite */,
4597 0 /* subprogno, zero == main subprog */);
4600 struct bpf_insn *insn;
4604 if (insn_idx >= insn_cnt) {
4605 verbose(env, "invalid insn idx %d insn_cnt %d\n",
4606 insn_idx, insn_cnt);
4610 insn = &insns[insn_idx];
4611 class = BPF_CLASS(insn->code);
4613 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
4615 "BPF program is too large. Processed %d insn\n",
4620 err = is_state_visited(env, insn_idx);
4624 /* found equivalent state, can prune the search */
4625 if (env->log.level) {
4627 verbose(env, "\nfrom %d to %d: safe\n",
4628 prev_insn_idx, insn_idx);
4630 verbose(env, "%d: safe\n", insn_idx);
4632 goto process_bpf_exit;
4638 if (env->log.level > 1 || (env->log.level && do_print_state)) {
4639 if (env->log.level > 1)
4640 verbose(env, "%d:", insn_idx);
4642 verbose(env, "\nfrom %d to %d:",
4643 prev_insn_idx, insn_idx);
4644 print_verifier_state(env, state->frame[state->curframe]);
4645 do_print_state = false;
4648 if (env->log.level) {
4649 const struct bpf_insn_cbs cbs = {
4650 .cb_print = verbose,
4651 .private_data = env,
4654 verbose(env, "%d: ", insn_idx);
4655 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4658 if (bpf_prog_is_dev_bound(env->prog->aux)) {
4659 err = bpf_prog_offload_verify_insn(env, insn_idx,
4665 regs = cur_regs(env);
4666 env->insn_aux_data[insn_idx].seen = true;
4667 if (class == BPF_ALU || class == BPF_ALU64) {
4668 err = check_alu_op(env, insn);
4672 } else if (class == BPF_LDX) {
4673 enum bpf_reg_type *prev_src_type, src_reg_type;
4675 /* check for reserved fields is already done */
4677 /* check src operand */
4678 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4682 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4686 src_reg_type = regs[insn->src_reg].type;
4688 /* check that memory (src_reg + off) is readable,
4689 * the state of dst_reg will be updated by this func
4691 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
4692 BPF_SIZE(insn->code), BPF_READ,
4693 insn->dst_reg, false);
4697 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
4699 if (*prev_src_type == NOT_INIT) {
4701 * dst_reg = *(u32 *)(src_reg + off)
4702 * save type to validate intersecting paths
4704 *prev_src_type = src_reg_type;
4706 } else if (src_reg_type != *prev_src_type &&
4707 (src_reg_type == PTR_TO_CTX ||
4708 *prev_src_type == PTR_TO_CTX)) {
4709 /* ABuser program is trying to use the same insn
4710 * dst_reg = *(u32*) (src_reg + off)
4711 * with different pointer types:
4712 * src_reg == ctx in one branch and
4713 * src_reg == stack|map in some other branch.
4716 verbose(env, "same insn cannot be used with different pointers\n");
4720 } else if (class == BPF_STX) {
4721 enum bpf_reg_type *prev_dst_type, dst_reg_type;
4723 if (BPF_MODE(insn->code) == BPF_XADD) {
4724 err = check_xadd(env, insn_idx, insn);
4731 /* check src1 operand */
4732 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4735 /* check src2 operand */
4736 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4740 dst_reg_type = regs[insn->dst_reg].type;
4742 /* check that memory (dst_reg + off) is writeable */
4743 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4744 BPF_SIZE(insn->code), BPF_WRITE,
4745 insn->src_reg, false);
4749 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
4751 if (*prev_dst_type == NOT_INIT) {
4752 *prev_dst_type = dst_reg_type;
4753 } else if (dst_reg_type != *prev_dst_type &&
4754 (dst_reg_type == PTR_TO_CTX ||
4755 *prev_dst_type == PTR_TO_CTX)) {
4756 verbose(env, "same insn cannot be used with different pointers\n");
4760 } else if (class == BPF_ST) {
4761 if (BPF_MODE(insn->code) != BPF_MEM ||
4762 insn->src_reg != BPF_REG_0) {
4763 verbose(env, "BPF_ST uses reserved fields\n");
4766 /* check src operand */
4767 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4771 if (is_ctx_reg(env, insn->dst_reg)) {
4772 verbose(env, "BPF_ST stores into R%d context is not allowed\n",
4777 /* check that memory (dst_reg + off) is writeable */
4778 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4779 BPF_SIZE(insn->code), BPF_WRITE,
4784 } else if (class == BPF_JMP) {
4785 u8 opcode = BPF_OP(insn->code);
4787 if (opcode == BPF_CALL) {
4788 if (BPF_SRC(insn->code) != BPF_K ||
4790 (insn->src_reg != BPF_REG_0 &&
4791 insn->src_reg != BPF_PSEUDO_CALL) ||
4792 insn->dst_reg != BPF_REG_0) {
4793 verbose(env, "BPF_CALL uses reserved fields\n");
4797 if (insn->src_reg == BPF_PSEUDO_CALL)
4798 err = check_func_call(env, insn, &insn_idx);
4800 err = check_helper_call(env, insn->imm, insn_idx);
4804 } else if (opcode == BPF_JA) {
4805 if (BPF_SRC(insn->code) != BPF_K ||
4807 insn->src_reg != BPF_REG_0 ||
4808 insn->dst_reg != BPF_REG_0) {
4809 verbose(env, "BPF_JA uses reserved fields\n");
4813 insn_idx += insn->off + 1;
4816 } else if (opcode == BPF_EXIT) {
4817 if (BPF_SRC(insn->code) != BPF_K ||
4819 insn->src_reg != BPF_REG_0 ||
4820 insn->dst_reg != BPF_REG_0) {
4821 verbose(env, "BPF_EXIT uses reserved fields\n");
4825 if (state->curframe) {
4826 /* exit from nested function */
4827 prev_insn_idx = insn_idx;
4828 err = prepare_func_exit(env, &insn_idx);
4831 do_print_state = true;
4835 /* eBPF calling convetion is such that R0 is used
4836 * to return the value from eBPF program.
4837 * Make sure that it's readable at this time
4838 * of bpf_exit, which means that program wrote
4839 * something into it earlier
4841 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4845 if (is_pointer_value(env, BPF_REG_0)) {
4846 verbose(env, "R0 leaks addr as return value\n");
4850 err = check_return_code(env);
4854 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4860 do_print_state = true;
4864 err = check_cond_jmp_op(env, insn, &insn_idx);
4868 } else if (class == BPF_LD) {
4869 u8 mode = BPF_MODE(insn->code);
4871 if (mode == BPF_ABS || mode == BPF_IND) {
4872 err = check_ld_abs(env, insn);
4876 } else if (mode == BPF_IMM) {
4877 err = check_ld_imm(env, insn);
4882 env->insn_aux_data[insn_idx].seen = true;
4884 verbose(env, "invalid BPF_LD mode\n");
4888 verbose(env, "unknown insn class %d\n", class);
4895 verbose(env, "processed %d insns (limit %d), stack depth ",
4896 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
4897 for (i = 0; i < env->subprog_cnt + 1; i++) {
4898 u32 depth = env->subprog_stack_depth[i];
4900 verbose(env, "%d", depth);
4901 if (i + 1 < env->subprog_cnt + 1)
4905 env->prog->aux->stack_depth = env->subprog_stack_depth[0];
4909 static int check_map_prealloc(struct bpf_map *map)
4911 return (map->map_type != BPF_MAP_TYPE_HASH &&
4912 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
4913 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
4914 !(map->map_flags & BPF_F_NO_PREALLOC);
4917 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
4918 struct bpf_map *map,
4919 struct bpf_prog *prog)
4922 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4923 * preallocated hash maps, since doing memory allocation
4924 * in overflow_handler can crash depending on where nmi got
4927 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4928 if (!check_map_prealloc(map)) {
4929 verbose(env, "perf_event programs can only use preallocated hash map\n");
4932 if (map->inner_map_meta &&
4933 !check_map_prealloc(map->inner_map_meta)) {
4934 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
4939 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
4940 !bpf_offload_dev_match(prog, map)) {
4941 verbose(env, "offload device mismatch between prog and map\n");
4948 /* look for pseudo eBPF instructions that access map FDs and
4949 * replace them with actual map pointers
4951 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4953 struct bpf_insn *insn = env->prog->insnsi;
4954 int insn_cnt = env->prog->len;
4957 err = bpf_prog_calc_tag(env->prog);
4961 for (i = 0; i < insn_cnt; i++, insn++) {
4962 if (BPF_CLASS(insn->code) == BPF_LDX &&
4963 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4964 verbose(env, "BPF_LDX uses reserved fields\n");
4968 if (BPF_CLASS(insn->code) == BPF_STX &&
4969 ((BPF_MODE(insn->code) != BPF_MEM &&
4970 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4971 verbose(env, "BPF_STX uses reserved fields\n");
4975 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4976 struct bpf_map *map;
4979 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4980 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4982 verbose(env, "invalid bpf_ld_imm64 insn\n");
4986 if (insn->src_reg == 0)
4987 /* valid generic load 64-bit imm */
4990 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4992 "unrecognized bpf_ld_imm64 insn\n");
4996 f = fdget(insn->imm);
4997 map = __bpf_map_get(f);
4999 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5001 return PTR_ERR(map);
5004 err = check_map_prog_compatibility(env, map, env->prog);
5010 /* store map pointer inside BPF_LD_IMM64 instruction */
5011 insn[0].imm = (u32) (unsigned long) map;
5012 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5014 /* check whether we recorded this map already */
5015 for (j = 0; j < env->used_map_cnt; j++)
5016 if (env->used_maps[j] == map) {
5021 if (env->used_map_cnt >= MAX_USED_MAPS) {
5026 /* hold the map. If the program is rejected by verifier,
5027 * the map will be released by release_maps() or it
5028 * will be used by the valid program until it's unloaded
5029 * and all maps are released in free_bpf_prog_info()
5031 map = bpf_map_inc(map, false);
5034 return PTR_ERR(map);
5036 env->used_maps[env->used_map_cnt++] = map;
5045 /* Basic sanity check before we invest more work here. */
5046 if (!bpf_opcode_in_insntable(insn->code)) {
5047 verbose(env, "unknown opcode %02x\n", insn->code);
5052 /* now all pseudo BPF_LD_IMM64 instructions load valid
5053 * 'struct bpf_map *' into a register instead of user map_fd.
5054 * These pointers will be used later by verifier to validate map access.
5059 /* drop refcnt of maps used by the rejected program */
5060 static void release_maps(struct bpf_verifier_env *env)
5064 for (i = 0; i < env->used_map_cnt; i++)
5065 bpf_map_put(env->used_maps[i]);
5068 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5069 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5071 struct bpf_insn *insn = env->prog->insnsi;
5072 int insn_cnt = env->prog->len;
5075 for (i = 0; i < insn_cnt; i++, insn++)
5076 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5080 /* single env->prog->insni[off] instruction was replaced with the range
5081 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5082 * [0, off) and [off, end) to new locations, so the patched range stays zero
5084 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5087 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5092 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
5095 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5096 memcpy(new_data + off + cnt - 1, old_data + off,
5097 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5098 for (i = off; i < off + cnt - 1; i++)
5099 new_data[i].seen = true;
5100 env->insn_aux_data = new_data;
5105 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5111 for (i = 0; i < env->subprog_cnt; i++) {
5112 if (env->subprog_starts[i] < off)
5114 env->subprog_starts[i] += len - 1;
5118 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5119 const struct bpf_insn *patch, u32 len)
5121 struct bpf_prog *new_prog;
5123 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5126 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5128 adjust_subprog_starts(env, off, len);
5132 /* The verifier does more data flow analysis than llvm and will not
5133 * explore branches that are dead at run time. Malicious programs can
5134 * have dead code too. Therefore replace all dead at-run-time code
5137 * Just nops are not optimal, e.g. if they would sit at the end of the
5138 * program and through another bug we would manage to jump there, then
5139 * we'd execute beyond program memory otherwise. Returning exception
5140 * code also wouldn't work since we can have subprogs where the dead
5141 * code could be located.
5143 static void sanitize_dead_code(struct bpf_verifier_env *env)
5145 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5146 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5147 struct bpf_insn *insn = env->prog->insnsi;
5148 const int insn_cnt = env->prog->len;
5151 for (i = 0; i < insn_cnt; i++) {
5152 if (aux_data[i].seen)
5154 memcpy(insn + i, &trap, sizeof(trap));
5158 /* convert load instructions that access fields of 'struct __sk_buff'
5159 * into sequence of instructions that access fields of 'struct sk_buff'
5161 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5163 const struct bpf_verifier_ops *ops = env->ops;
5164 int i, cnt, size, ctx_field_size, delta = 0;
5165 const int insn_cnt = env->prog->len;
5166 struct bpf_insn insn_buf[16], *insn;
5167 struct bpf_prog *new_prog;
5168 enum bpf_access_type type;
5169 bool is_narrower_load;
5172 if (ops->gen_prologue) {
5173 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5175 if (cnt >= ARRAY_SIZE(insn_buf)) {
5176 verbose(env, "bpf verifier is misconfigured\n");
5179 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5183 env->prog = new_prog;
5188 if (!ops->convert_ctx_access)
5191 insn = env->prog->insnsi + delta;
5193 for (i = 0; i < insn_cnt; i++, insn++) {
5194 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5195 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5196 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5197 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5199 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5200 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5201 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5202 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5207 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
5210 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5211 size = BPF_LDST_BYTES(insn);
5213 /* If the read access is a narrower load of the field,
5214 * convert to a 4/8-byte load, to minimum program type specific
5215 * convert_ctx_access changes. If conversion is successful,
5216 * we will apply proper mask to the result.
5218 is_narrower_load = size < ctx_field_size;
5219 if (is_narrower_load) {
5220 u32 off = insn->off;
5223 if (type == BPF_WRITE) {
5224 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5229 if (ctx_field_size == 4)
5231 else if (ctx_field_size == 8)
5234 insn->off = off & ~(ctx_field_size - 1);
5235 insn->code = BPF_LDX | BPF_MEM | size_code;
5239 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
5241 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5242 (ctx_field_size && !target_size)) {
5243 verbose(env, "bpf verifier is misconfigured\n");
5247 if (is_narrower_load && size < target_size) {
5248 if (ctx_field_size <= 4)
5249 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5250 (1 << size * 8) - 1);
5252 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5253 (1 << size * 8) - 1);
5256 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5262 /* keep walking new program and skip insns we just inserted */
5263 env->prog = new_prog;
5264 insn = new_prog->insnsi + i + delta;
5270 static int jit_subprogs(struct bpf_verifier_env *env)
5272 struct bpf_prog *prog = env->prog, **func, *tmp;
5273 int i, j, subprog_start, subprog_end = 0, len, subprog;
5274 struct bpf_insn *insn;
5278 if (env->subprog_cnt == 0)
5281 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5282 if (insn->code != (BPF_JMP | BPF_CALL) ||
5283 insn->src_reg != BPF_PSEUDO_CALL)
5285 subprog = find_subprog(env, i + insn->imm + 1);
5287 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5291 /* temporarily remember subprog id inside insn instead of
5292 * aux_data, since next loop will split up all insns into funcs
5294 insn->off = subprog + 1;
5295 /* remember original imm in case JIT fails and fallback
5296 * to interpreter will be needed
5298 env->insn_aux_data[i].call_imm = insn->imm;
5299 /* point imm to __bpf_call_base+1 from JITs point of view */
5303 func = kzalloc(sizeof(prog) * (env->subprog_cnt + 1), GFP_KERNEL);
5307 for (i = 0; i <= env->subprog_cnt; i++) {
5308 subprog_start = subprog_end;
5309 if (env->subprog_cnt == i)
5310 subprog_end = prog->len;
5312 subprog_end = env->subprog_starts[i];
5314 len = subprog_end - subprog_start;
5315 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5318 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5319 len * sizeof(struct bpf_insn));
5320 func[i]->type = prog->type;
5322 if (bpf_prog_calc_tag(func[i]))
5324 func[i]->is_func = 1;
5325 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5326 * Long term would need debug info to populate names
5328 func[i]->aux->name[0] = 'F';
5329 func[i]->aux->stack_depth = env->subprog_stack_depth[i];
5330 func[i]->jit_requested = 1;
5331 func[i] = bpf_int_jit_compile(func[i]);
5332 if (!func[i]->jited) {
5338 /* at this point all bpf functions were successfully JITed
5339 * now populate all bpf_calls with correct addresses and
5340 * run last pass of JIT
5342 for (i = 0; i <= env->subprog_cnt; i++) {
5343 insn = func[i]->insnsi;
5344 for (j = 0; j < func[i]->len; j++, insn++) {
5345 if (insn->code != (BPF_JMP | BPF_CALL) ||
5346 insn->src_reg != BPF_PSEUDO_CALL)
5348 subprog = insn->off;
5350 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5351 func[subprog]->bpf_func -
5355 for (i = 0; i <= env->subprog_cnt; i++) {
5356 old_bpf_func = func[i]->bpf_func;
5357 tmp = bpf_int_jit_compile(func[i]);
5358 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5359 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5366 /* finally lock prog and jit images for all functions and
5369 for (i = 0; i <= env->subprog_cnt; i++) {
5370 bpf_prog_lock_ro(func[i]);
5371 bpf_prog_kallsyms_add(func[i]);
5374 /* Last step: make now unused interpreter insns from main
5375 * prog consistent for later dump requests, so they can
5376 * later look the same as if they were interpreted only.
5378 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5381 if (insn->code != (BPF_JMP | BPF_CALL) ||
5382 insn->src_reg != BPF_PSEUDO_CALL)
5384 insn->off = env->insn_aux_data[i].call_imm;
5385 subprog = find_subprog(env, i + insn->off + 1);
5386 addr = (unsigned long)func[subprog + 1]->bpf_func;
5388 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5389 addr - __bpf_call_base;
5393 prog->bpf_func = func[0]->bpf_func;
5394 prog->aux->func = func;
5395 prog->aux->func_cnt = env->subprog_cnt + 1;
5398 for (i = 0; i <= env->subprog_cnt; i++)
5400 bpf_jit_free(func[i]);
5402 /* cleanup main prog to be interpreted */
5403 prog->jit_requested = 0;
5404 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5405 if (insn->code != (BPF_JMP | BPF_CALL) ||
5406 insn->src_reg != BPF_PSEUDO_CALL)
5409 insn->imm = env->insn_aux_data[i].call_imm;
5414 static int fixup_call_args(struct bpf_verifier_env *env)
5416 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5417 struct bpf_prog *prog = env->prog;
5418 struct bpf_insn *insn = prog->insnsi;
5424 if (env->prog->jit_requested) {
5425 err = jit_subprogs(env);
5429 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5430 for (i = 0; i < prog->len; i++, insn++) {
5431 if (insn->code != (BPF_JMP | BPF_CALL) ||
5432 insn->src_reg != BPF_PSEUDO_CALL)
5434 depth = get_callee_stack_depth(env, insn, i);
5437 bpf_patch_call_args(insn, depth);
5444 /* fixup insn->imm field of bpf_call instructions
5445 * and inline eligible helpers as explicit sequence of BPF instructions
5447 * this function is called after eBPF program passed verification
5449 static int fixup_bpf_calls(struct bpf_verifier_env *env)
5451 struct bpf_prog *prog = env->prog;
5452 struct bpf_insn *insn = prog->insnsi;
5453 const struct bpf_func_proto *fn;
5454 const int insn_cnt = prog->len;
5455 struct bpf_insn_aux_data *aux;
5456 struct bpf_insn insn_buf[16];
5457 struct bpf_prog *new_prog;
5458 struct bpf_map *map_ptr;
5459 int i, cnt, delta = 0;
5461 for (i = 0; i < insn_cnt; i++, insn++) {
5462 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
5463 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5464 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
5465 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5466 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
5467 struct bpf_insn mask_and_div[] = {
5468 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5470 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
5471 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
5472 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
5475 struct bpf_insn mask_and_mod[] = {
5476 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5477 /* Rx mod 0 -> Rx */
5478 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
5481 struct bpf_insn *patchlet;
5483 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5484 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5485 patchlet = mask_and_div + (is64 ? 1 : 0);
5486 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
5488 patchlet = mask_and_mod + (is64 ? 1 : 0);
5489 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
5492 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
5497 env->prog = prog = new_prog;
5498 insn = new_prog->insnsi + i + delta;
5502 if (insn->code != (BPF_JMP | BPF_CALL))
5504 if (insn->src_reg == BPF_PSEUDO_CALL)
5507 if (insn->imm == BPF_FUNC_get_route_realm)
5508 prog->dst_needed = 1;
5509 if (insn->imm == BPF_FUNC_get_prandom_u32)
5510 bpf_user_rnd_init_once();
5511 if (insn->imm == BPF_FUNC_override_return)
5512 prog->kprobe_override = 1;
5513 if (insn->imm == BPF_FUNC_tail_call) {
5514 /* If we tail call into other programs, we
5515 * cannot make any assumptions since they can
5516 * be replaced dynamically during runtime in
5517 * the program array.
5519 prog->cb_access = 1;
5520 env->prog->aux->stack_depth = MAX_BPF_STACK;
5522 /* mark bpf_tail_call as different opcode to avoid
5523 * conditional branch in the interpeter for every normal
5524 * call and to prevent accidental JITing by JIT compiler
5525 * that doesn't support bpf_tail_call yet
5528 insn->code = BPF_JMP | BPF_TAIL_CALL;
5530 aux = &env->insn_aux_data[i + delta];
5531 if (!bpf_map_ptr_unpriv(aux))
5534 /* instead of changing every JIT dealing with tail_call
5535 * emit two extra insns:
5536 * if (index >= max_entries) goto out;
5537 * index &= array->index_mask;
5538 * to avoid out-of-bounds cpu speculation
5540 if (bpf_map_ptr_poisoned(aux)) {
5541 verbose(env, "tail_call abusing map_ptr\n");
5545 map_ptr = BPF_MAP_PTR(aux->map_state);
5546 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
5547 map_ptr->max_entries, 2);
5548 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
5549 container_of(map_ptr,
5552 insn_buf[2] = *insn;
5554 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5559 env->prog = prog = new_prog;
5560 insn = new_prog->insnsi + i + delta;
5564 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5565 * handlers are currently limited to 64 bit only.
5567 if (prog->jit_requested && BITS_PER_LONG == 64 &&
5568 insn->imm == BPF_FUNC_map_lookup_elem) {
5569 aux = &env->insn_aux_data[i + delta];
5570 if (bpf_map_ptr_poisoned(aux))
5571 goto patch_call_imm;
5573 map_ptr = BPF_MAP_PTR(aux->map_state);
5574 if (!map_ptr->ops->map_gen_lookup)
5575 goto patch_call_imm;
5577 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
5578 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5579 verbose(env, "bpf verifier is misconfigured\n");
5583 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
5590 /* keep walking new program and skip insns we just inserted */
5591 env->prog = prog = new_prog;
5592 insn = new_prog->insnsi + i + delta;
5596 if (insn->imm == BPF_FUNC_redirect_map) {
5597 /* Note, we cannot use prog directly as imm as subsequent
5598 * rewrites would still change the prog pointer. The only
5599 * stable address we can use is aux, which also works with
5600 * prog clones during blinding.
5602 u64 addr = (unsigned long)prog->aux;
5603 struct bpf_insn r4_ld[] = {
5604 BPF_LD_IMM64(BPF_REG_4, addr),
5607 cnt = ARRAY_SIZE(r4_ld);
5609 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
5614 env->prog = prog = new_prog;
5615 insn = new_prog->insnsi + i + delta;
5618 fn = env->ops->get_func_proto(insn->imm, env->prog);
5619 /* all functions that have prototype and verifier allowed
5620 * programs to call them, must be real in-kernel functions
5624 "kernel subsystem misconfigured func %s#%d\n",
5625 func_id_name(insn->imm), insn->imm);
5628 insn->imm = fn->func - __bpf_call_base;
5634 static void free_states(struct bpf_verifier_env *env)
5636 struct bpf_verifier_state_list *sl, *sln;
5639 if (!env->explored_states)
5642 for (i = 0; i < env->prog->len; i++) {
5643 sl = env->explored_states[i];
5646 while (sl != STATE_LIST_MARK) {
5648 free_verifier_state(&sl->state, false);
5654 kfree(env->explored_states);
5657 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
5659 struct bpf_verifier_env *env;
5660 struct bpf_verifier_log *log;
5663 /* no program is valid */
5664 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
5667 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5668 * allocate/free it every time bpf_check() is called
5670 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
5675 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
5678 if (!env->insn_aux_data)
5681 env->ops = bpf_verifier_ops[env->prog->type];
5683 /* grab the mutex to protect few globals used by verifier */
5684 mutex_lock(&bpf_verifier_lock);
5686 if (attr->log_level || attr->log_buf || attr->log_size) {
5687 /* user requested verbose verifier output
5688 * and supplied buffer to store the verification trace
5690 log->level = attr->log_level;
5691 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
5692 log->len_total = attr->log_size;
5695 /* log attributes have to be sane */
5696 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
5697 !log->level || !log->ubuf)
5701 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
5702 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
5703 env->strict_alignment = true;
5705 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5706 ret = bpf_prog_offload_verifier_prep(env);
5711 ret = replace_map_fd_with_map_ptr(env);
5713 goto skip_full_check;
5715 env->explored_states = kcalloc(env->prog->len,
5716 sizeof(struct bpf_verifier_state_list *),
5719 if (!env->explored_states)
5720 goto skip_full_check;
5722 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
5724 ret = check_cfg(env);
5726 goto skip_full_check;
5728 ret = do_check(env);
5729 if (env->cur_state) {
5730 free_verifier_state(env->cur_state, true);
5731 env->cur_state = NULL;
5735 while (!pop_stack(env, NULL, NULL));
5739 sanitize_dead_code(env);
5742 ret = check_max_stack_depth(env);
5745 /* program is valid, convert *(u32*)(ctx + off) accesses */
5746 ret = convert_ctx_accesses(env);
5749 ret = fixup_bpf_calls(env);
5752 ret = fixup_call_args(env);
5754 if (log->level && bpf_verifier_log_full(log))
5756 if (log->level && !log->ubuf) {
5758 goto err_release_maps;
5761 if (ret == 0 && env->used_map_cnt) {
5762 /* if program passed verifier, update used_maps in bpf_prog_info */
5763 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
5764 sizeof(env->used_maps[0]),
5767 if (!env->prog->aux->used_maps) {
5769 goto err_release_maps;
5772 memcpy(env->prog->aux->used_maps, env->used_maps,
5773 sizeof(env->used_maps[0]) * env->used_map_cnt);
5774 env->prog->aux->used_map_cnt = env->used_map_cnt;
5776 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
5777 * bpf_ld_imm64 instructions
5779 convert_pseudo_ld_imm64(env);
5783 if (!env->prog->aux->used_maps)
5784 /* if we didn't copy map pointers into bpf_prog_info, release
5785 * them now. Otherwise free_bpf_prog_info() will release them.
5790 mutex_unlock(&bpf_verifier_lock);
5791 vfree(env->insn_aux_data);