1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/kernel.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/bpf.h>
17 #include <linux/bpf_verifier.h>
18 #include <linux/filter.h>
19 #include <net/netlink.h>
20 #include <linux/file.h>
21 #include <linux/vmalloc.h>
22 #include <linux/stringify.h>
23 #include <linux/bsearch.h>
24 #include <linux/sort.h>
25 #include <linux/perf_event.h>
29 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
30 #define BPF_PROG_TYPE(_id, _name) \
31 [_id] = & _name ## _verifier_ops,
32 #define BPF_MAP_TYPE(_id, _ops)
33 #include <linux/bpf_types.h>
38 /* bpf_check() is a static code analyzer that walks eBPF program
39 * instruction by instruction and updates register/stack state.
40 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
42 * The first pass is depth-first-search to check that the program is a DAG.
43 * It rejects the following programs:
44 * - larger than BPF_MAXINSNS insns
45 * - if loop is present (detected via back-edge)
46 * - unreachable insns exist (shouldn't be a forest. program = one function)
47 * - out of bounds or malformed jumps
48 * The second pass is all possible path descent from the 1st insn.
49 * Since it's analyzing all pathes through the program, the length of the
50 * analysis is limited to 64k insn, which may be hit even if total number of
51 * insn is less then 4K, but there are too many branches that change stack/regs.
52 * Number of 'branches to be analyzed' is limited to 1k
54 * On entry to each instruction, each register has a type, and the instruction
55 * changes the types of the registers depending on instruction semantics.
56 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
59 * All registers are 64-bit.
60 * R0 - return register
61 * R1-R5 argument passing registers
62 * R6-R9 callee saved registers
63 * R10 - frame pointer read-only
65 * At the start of BPF program the register R1 contains a pointer to bpf_context
66 * and has type PTR_TO_CTX.
68 * Verifier tracks arithmetic operations on pointers in case:
69 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
70 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
71 * 1st insn copies R10 (which has FRAME_PTR) type into R1
72 * and 2nd arithmetic instruction is pattern matched to recognize
73 * that it wants to construct a pointer to some element within stack.
74 * So after 2nd insn, the register R1 has type PTR_TO_STACK
75 * (and -20 constant is saved for further stack bounds checking).
76 * Meaning that this reg is a pointer to stack plus known immediate constant.
78 * Most of the time the registers have SCALAR_VALUE type, which
79 * means the register has some value, but it's not a valid pointer.
80 * (like pointer plus pointer becomes SCALAR_VALUE type)
82 * When verifier sees load or store instructions the type of base register
83 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
84 * types recognized by check_mem_access() function.
86 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
87 * and the range of [ptr, ptr + map's value_size) is accessible.
89 * registers used to pass values to function calls are checked against
90 * function argument constraints.
92 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
93 * It means that the register type passed to this function must be
94 * PTR_TO_STACK and it will be used inside the function as
95 * 'pointer to map element key'
97 * For example the argument constraints for bpf_map_lookup_elem():
98 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
99 * .arg1_type = ARG_CONST_MAP_PTR,
100 * .arg2_type = ARG_PTR_TO_MAP_KEY,
102 * ret_type says that this function returns 'pointer to map elem value or null'
103 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
104 * 2nd argument should be a pointer to stack, which will be used inside
105 * the helper function as a pointer to map element key.
107 * On the kernel side the helper function looks like:
108 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
110 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
111 * void *key = (void *) (unsigned long) r2;
114 * here kernel can access 'key' and 'map' pointers safely, knowing that
115 * [key, key + map->key_size) bytes are valid and were initialized on
116 * the stack of eBPF program.
119 * Corresponding eBPF program may look like:
120 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
121 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
122 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
123 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
124 * here verifier looks at prototype of map_lookup_elem() and sees:
125 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
126 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
128 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
129 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
130 * and were initialized prior to this call.
131 * If it's ok, then verifier allows this BPF_CALL insn and looks at
132 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
133 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
134 * returns ether pointer to map value or NULL.
136 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
137 * insn, the register holding that pointer in the true branch changes state to
138 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
139 * branch. See check_cond_jmp_op().
141 * After the call R0 is set to return type of the function and registers R1-R5
142 * are set to NOT_INIT to indicate that they are no longer readable.
145 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
146 struct bpf_verifier_stack_elem {
147 /* verifer state is 'st'
148 * before processing instruction 'insn_idx'
149 * and after processing instruction 'prev_insn_idx'
151 struct bpf_verifier_state st;
154 struct bpf_verifier_stack_elem *next;
157 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
158 #define BPF_COMPLEXITY_LIMIT_STACK 1024
160 #define BPF_MAP_PTR_UNPRIV 1UL
161 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
162 POISON_POINTER_DELTA))
163 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
165 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
167 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
170 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
172 return aux->map_state & BPF_MAP_PTR_UNPRIV;
175 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
176 const struct bpf_map *map, bool unpriv)
178 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
179 unpriv |= bpf_map_ptr_unpriv(aux);
180 aux->map_state = (unsigned long)map |
181 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
184 struct bpf_call_arg_meta {
185 struct bpf_map *map_ptr;
190 s64 msize_smax_value;
191 u64 msize_umax_value;
194 static DEFINE_MUTEX(bpf_verifier_lock);
196 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
201 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
203 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
204 "verifier log line truncated - local buffer too short\n");
206 n = min(log->len_total - log->len_used - 1, n);
209 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
215 /* log_level controls verbosity level of eBPF verifier.
216 * bpf_verifier_log_write() is used to dump the verification trace to the log,
217 * so the user can figure out what's wrong with the program
219 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
220 const char *fmt, ...)
224 if (!bpf_verifier_log_needed(&env->log))
228 bpf_verifier_vlog(&env->log, fmt, args);
231 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
233 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
235 struct bpf_verifier_env *env = private_data;
238 if (!bpf_verifier_log_needed(&env->log))
242 bpf_verifier_vlog(&env->log, fmt, args);
246 static bool type_is_pkt_pointer(enum bpf_reg_type type)
248 return type == PTR_TO_PACKET ||
249 type == PTR_TO_PACKET_META;
252 /* string representation of 'enum bpf_reg_type' */
253 static const char * const reg_type_str[] = {
255 [SCALAR_VALUE] = "inv",
256 [PTR_TO_CTX] = "ctx",
257 [CONST_PTR_TO_MAP] = "map_ptr",
258 [PTR_TO_MAP_VALUE] = "map_value",
259 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
260 [PTR_TO_STACK] = "fp",
261 [PTR_TO_PACKET] = "pkt",
262 [PTR_TO_PACKET_META] = "pkt_meta",
263 [PTR_TO_PACKET_END] = "pkt_end",
264 [PTR_TO_FLOW_KEYS] = "flow_keys",
267 static char slot_type_char[] = {
268 [STACK_INVALID] = '?',
274 static void print_liveness(struct bpf_verifier_env *env,
275 enum bpf_reg_liveness live)
277 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
279 if (live & REG_LIVE_READ)
281 if (live & REG_LIVE_WRITTEN)
285 static struct bpf_func_state *func(struct bpf_verifier_env *env,
286 const struct bpf_reg_state *reg)
288 struct bpf_verifier_state *cur = env->cur_state;
290 return cur->frame[reg->frameno];
293 static void print_verifier_state(struct bpf_verifier_env *env,
294 const struct bpf_func_state *state)
296 const struct bpf_reg_state *reg;
301 verbose(env, " frame%d:", state->frameno);
302 for (i = 0; i < MAX_BPF_REG; i++) {
303 reg = &state->regs[i];
307 verbose(env, " R%d", i);
308 print_liveness(env, reg->live);
309 verbose(env, "=%s", reg_type_str[t]);
310 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
311 tnum_is_const(reg->var_off)) {
312 /* reg->off should be 0 for SCALAR_VALUE */
313 verbose(env, "%lld", reg->var_off.value + reg->off);
314 if (t == PTR_TO_STACK)
315 verbose(env, ",call_%d", func(env, reg)->callsite);
317 verbose(env, "(id=%d", reg->id);
318 if (t != SCALAR_VALUE)
319 verbose(env, ",off=%d", reg->off);
320 if (type_is_pkt_pointer(t))
321 verbose(env, ",r=%d", reg->range);
322 else if (t == CONST_PTR_TO_MAP ||
323 t == PTR_TO_MAP_VALUE ||
324 t == PTR_TO_MAP_VALUE_OR_NULL)
325 verbose(env, ",ks=%d,vs=%d",
326 reg->map_ptr->key_size,
327 reg->map_ptr->value_size);
328 if (tnum_is_const(reg->var_off)) {
329 /* Typically an immediate SCALAR_VALUE, but
330 * could be a pointer whose offset is too big
333 verbose(env, ",imm=%llx", reg->var_off.value);
335 if (reg->smin_value != reg->umin_value &&
336 reg->smin_value != S64_MIN)
337 verbose(env, ",smin_value=%lld",
338 (long long)reg->smin_value);
339 if (reg->smax_value != reg->umax_value &&
340 reg->smax_value != S64_MAX)
341 verbose(env, ",smax_value=%lld",
342 (long long)reg->smax_value);
343 if (reg->umin_value != 0)
344 verbose(env, ",umin_value=%llu",
345 (unsigned long long)reg->umin_value);
346 if (reg->umax_value != U64_MAX)
347 verbose(env, ",umax_value=%llu",
348 (unsigned long long)reg->umax_value);
349 if (!tnum_is_unknown(reg->var_off)) {
352 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
353 verbose(env, ",var_off=%s", tn_buf);
359 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
360 char types_buf[BPF_REG_SIZE + 1];
364 for (j = 0; j < BPF_REG_SIZE; j++) {
365 if (state->stack[i].slot_type[j] != STACK_INVALID)
367 types_buf[j] = slot_type_char[
368 state->stack[i].slot_type[j]];
370 types_buf[BPF_REG_SIZE] = 0;
373 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
374 print_liveness(env, state->stack[i].spilled_ptr.live);
375 if (state->stack[i].slot_type[0] == STACK_SPILL)
377 reg_type_str[state->stack[i].spilled_ptr.type]);
379 verbose(env, "=%s", types_buf);
384 static int copy_stack_state(struct bpf_func_state *dst,
385 const struct bpf_func_state *src)
389 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
390 /* internal bug, make state invalid to reject the program */
391 memset(dst, 0, sizeof(*dst));
394 memcpy(dst->stack, src->stack,
395 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
399 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
400 * make it consume minimal amount of memory. check_stack_write() access from
401 * the program calls into realloc_func_state() to grow the stack size.
402 * Note there is a non-zero parent pointer inside each reg of bpf_verifier_state
403 * which this function copies over. It points to corresponding reg in previous
404 * bpf_verifier_state which is never reallocated
406 static int realloc_func_state(struct bpf_func_state *state, int size,
409 u32 old_size = state->allocated_stack;
410 struct bpf_stack_state *new_stack;
411 int slot = size / BPF_REG_SIZE;
413 if (size <= old_size || !size) {
416 state->allocated_stack = slot * BPF_REG_SIZE;
417 if (!size && old_size) {
423 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
429 memcpy(new_stack, state->stack,
430 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
431 memset(new_stack + old_size / BPF_REG_SIZE, 0,
432 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
434 state->allocated_stack = slot * BPF_REG_SIZE;
436 state->stack = new_stack;
440 static void free_func_state(struct bpf_func_state *state)
448 static void free_verifier_state(struct bpf_verifier_state *state,
453 for (i = 0; i <= state->curframe; i++) {
454 free_func_state(state->frame[i]);
455 state->frame[i] = NULL;
461 /* copy verifier state from src to dst growing dst stack space
462 * when necessary to accommodate larger src stack
464 static int copy_func_state(struct bpf_func_state *dst,
465 const struct bpf_func_state *src)
469 err = realloc_func_state(dst, src->allocated_stack, false);
472 memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack));
473 return copy_stack_state(dst, src);
476 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
477 const struct bpf_verifier_state *src)
479 struct bpf_func_state *dst;
482 /* if dst has more stack frames then src frame, free them */
483 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
484 free_func_state(dst_state->frame[i]);
485 dst_state->frame[i] = NULL;
487 dst_state->curframe = src->curframe;
488 for (i = 0; i <= src->curframe; i++) {
489 dst = dst_state->frame[i];
491 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
494 dst_state->frame[i] = dst;
496 err = copy_func_state(dst, src->frame[i]);
503 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
506 struct bpf_verifier_state *cur = env->cur_state;
507 struct bpf_verifier_stack_elem *elem, *head = env->head;
510 if (env->head == NULL)
514 err = copy_verifier_state(cur, &head->st);
519 *insn_idx = head->insn_idx;
521 *prev_insn_idx = head->prev_insn_idx;
523 free_verifier_state(&head->st, false);
530 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
531 int insn_idx, int prev_insn_idx)
533 struct bpf_verifier_state *cur = env->cur_state;
534 struct bpf_verifier_stack_elem *elem;
537 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
541 elem->insn_idx = insn_idx;
542 elem->prev_insn_idx = prev_insn_idx;
543 elem->next = env->head;
546 err = copy_verifier_state(&elem->st, cur);
549 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
550 verbose(env, "BPF program is too complex\n");
555 free_verifier_state(env->cur_state, true);
556 env->cur_state = NULL;
557 /* pop all elements and return */
558 while (!pop_stack(env, NULL, NULL));
562 #define CALLER_SAVED_REGS 6
563 static const int caller_saved[CALLER_SAVED_REGS] = {
564 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
567 static void __mark_reg_not_init(struct bpf_reg_state *reg);
569 /* Mark the unknown part of a register (variable offset or scalar value) as
570 * known to have the value @imm.
572 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
574 /* Clear id, off, and union(map_ptr, range) */
575 memset(((u8 *)reg) + sizeof(reg->type), 0,
576 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
577 reg->var_off = tnum_const(imm);
578 reg->smin_value = (s64)imm;
579 reg->smax_value = (s64)imm;
580 reg->umin_value = imm;
581 reg->umax_value = imm;
584 /* Mark the 'variable offset' part of a register as zero. This should be
585 * used only on registers holding a pointer type.
587 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
589 __mark_reg_known(reg, 0);
592 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
594 __mark_reg_known(reg, 0);
595 reg->type = SCALAR_VALUE;
598 static void mark_reg_known_zero(struct bpf_verifier_env *env,
599 struct bpf_reg_state *regs, u32 regno)
601 if (WARN_ON(regno >= MAX_BPF_REG)) {
602 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
603 /* Something bad happened, let's kill all regs */
604 for (regno = 0; regno < MAX_BPF_REG; regno++)
605 __mark_reg_not_init(regs + regno);
608 __mark_reg_known_zero(regs + regno);
611 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
613 return type_is_pkt_pointer(reg->type);
616 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
618 return reg_is_pkt_pointer(reg) ||
619 reg->type == PTR_TO_PACKET_END;
622 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
623 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
624 enum bpf_reg_type which)
626 /* The register can already have a range from prior markings.
627 * This is fine as long as it hasn't been advanced from its
630 return reg->type == which &&
633 tnum_equals_const(reg->var_off, 0);
636 /* Attempts to improve min/max values based on var_off information */
637 static void __update_reg_bounds(struct bpf_reg_state *reg)
639 /* min signed is max(sign bit) | min(other bits) */
640 reg->smin_value = max_t(s64, reg->smin_value,
641 reg->var_off.value | (reg->var_off.mask & S64_MIN));
642 /* max signed is min(sign bit) | max(other bits) */
643 reg->smax_value = min_t(s64, reg->smax_value,
644 reg->var_off.value | (reg->var_off.mask & S64_MAX));
645 reg->umin_value = max(reg->umin_value, reg->var_off.value);
646 reg->umax_value = min(reg->umax_value,
647 reg->var_off.value | reg->var_off.mask);
650 /* Uses signed min/max values to inform unsigned, and vice-versa */
651 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
653 /* Learn sign from signed bounds.
654 * If we cannot cross the sign boundary, then signed and unsigned bounds
655 * are the same, so combine. This works even in the negative case, e.g.
656 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
658 if (reg->smin_value >= 0 || reg->smax_value < 0) {
659 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
661 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
665 /* Learn sign from unsigned bounds. Signed bounds cross the sign
666 * boundary, so we must be careful.
668 if ((s64)reg->umax_value >= 0) {
669 /* Positive. We can't learn anything from the smin, but smax
670 * is positive, hence safe.
672 reg->smin_value = reg->umin_value;
673 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
675 } else if ((s64)reg->umin_value < 0) {
676 /* Negative. We can't learn anything from the smax, but smin
677 * is negative, hence safe.
679 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
681 reg->smax_value = reg->umax_value;
685 /* Attempts to improve var_off based on unsigned min/max information */
686 static void __reg_bound_offset(struct bpf_reg_state *reg)
688 reg->var_off = tnum_intersect(reg->var_off,
689 tnum_range(reg->umin_value,
693 /* Reset the min/max bounds of a register */
694 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
696 reg->smin_value = S64_MIN;
697 reg->smax_value = S64_MAX;
699 reg->umax_value = U64_MAX;
702 /* Mark a register as having a completely unknown (scalar) value. */
703 static void __mark_reg_unknown(struct bpf_reg_state *reg)
706 * Clear type, id, off, and union(map_ptr, range) and
707 * padding between 'type' and union
709 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
710 reg->type = SCALAR_VALUE;
711 reg->var_off = tnum_unknown;
713 __mark_reg_unbounded(reg);
716 static void mark_reg_unknown(struct bpf_verifier_env *env,
717 struct bpf_reg_state *regs, u32 regno)
719 if (WARN_ON(regno >= MAX_BPF_REG)) {
720 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
721 /* Something bad happened, let's kill all regs except FP */
722 for (regno = 0; regno < BPF_REG_FP; regno++)
723 __mark_reg_not_init(regs + regno);
726 __mark_reg_unknown(regs + regno);
729 static void __mark_reg_not_init(struct bpf_reg_state *reg)
731 __mark_reg_unknown(reg);
732 reg->type = NOT_INIT;
735 static void mark_reg_not_init(struct bpf_verifier_env *env,
736 struct bpf_reg_state *regs, u32 regno)
738 if (WARN_ON(regno >= MAX_BPF_REG)) {
739 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
740 /* Something bad happened, let's kill all regs except FP */
741 for (regno = 0; regno < BPF_REG_FP; regno++)
742 __mark_reg_not_init(regs + regno);
745 __mark_reg_not_init(regs + regno);
748 static void init_reg_state(struct bpf_verifier_env *env,
749 struct bpf_func_state *state)
751 struct bpf_reg_state *regs = state->regs;
754 for (i = 0; i < MAX_BPF_REG; i++) {
755 mark_reg_not_init(env, regs, i);
756 regs[i].live = REG_LIVE_NONE;
757 regs[i].parent = NULL;
761 regs[BPF_REG_FP].type = PTR_TO_STACK;
762 mark_reg_known_zero(env, regs, BPF_REG_FP);
763 regs[BPF_REG_FP].frameno = state->frameno;
765 /* 1st arg to a function */
766 regs[BPF_REG_1].type = PTR_TO_CTX;
767 mark_reg_known_zero(env, regs, BPF_REG_1);
770 #define BPF_MAIN_FUNC (-1)
771 static void init_func_state(struct bpf_verifier_env *env,
772 struct bpf_func_state *state,
773 int callsite, int frameno, int subprogno)
775 state->callsite = callsite;
776 state->frameno = frameno;
777 state->subprogno = subprogno;
778 init_reg_state(env, state);
782 SRC_OP, /* register is used as source operand */
783 DST_OP, /* register is used as destination operand */
784 DST_OP_NO_MARK /* same as above, check only, don't mark */
787 static int cmp_subprogs(const void *a, const void *b)
789 return ((struct bpf_subprog_info *)a)->start -
790 ((struct bpf_subprog_info *)b)->start;
793 static int find_subprog(struct bpf_verifier_env *env, int off)
795 struct bpf_subprog_info *p;
797 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
798 sizeof(env->subprog_info[0]), cmp_subprogs);
801 return p - env->subprog_info;
805 static int add_subprog(struct bpf_verifier_env *env, int off)
807 int insn_cnt = env->prog->len;
810 if (off >= insn_cnt || off < 0) {
811 verbose(env, "call to invalid destination\n");
814 ret = find_subprog(env, off);
817 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
818 verbose(env, "too many subprograms\n");
821 env->subprog_info[env->subprog_cnt++].start = off;
822 sort(env->subprog_info, env->subprog_cnt,
823 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
827 static int check_subprogs(struct bpf_verifier_env *env)
829 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
830 struct bpf_subprog_info *subprog = env->subprog_info;
831 struct bpf_insn *insn = env->prog->insnsi;
832 int insn_cnt = env->prog->len;
834 /* Add entry function. */
835 ret = add_subprog(env, 0);
839 /* determine subprog starts. The end is one before the next starts */
840 for (i = 0; i < insn_cnt; i++) {
841 if (insn[i].code != (BPF_JMP | BPF_CALL))
843 if (insn[i].src_reg != BPF_PSEUDO_CALL)
845 if (!env->allow_ptr_leaks) {
846 verbose(env, "function calls to other bpf functions are allowed for root only\n");
849 if (bpf_prog_is_dev_bound(env->prog->aux)) {
850 verbose(env, "function calls in offloaded programs are not supported yet\n");
853 ret = add_subprog(env, i + insn[i].imm + 1);
858 /* Add a fake 'exit' subprog which could simplify subprog iteration
859 * logic. 'subprog_cnt' should not be increased.
861 subprog[env->subprog_cnt].start = insn_cnt;
863 if (env->log.level > 1)
864 for (i = 0; i < env->subprog_cnt; i++)
865 verbose(env, "func#%d @%d\n", i, subprog[i].start);
867 /* now check that all jumps are within the same subprog */
868 subprog_start = subprog[cur_subprog].start;
869 subprog_end = subprog[cur_subprog + 1].start;
870 for (i = 0; i < insn_cnt; i++) {
871 u8 code = insn[i].code;
873 if (BPF_CLASS(code) != BPF_JMP)
875 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
877 off = i + insn[i].off + 1;
878 if (off < subprog_start || off >= subprog_end) {
879 verbose(env, "jump out of range from insn %d to %d\n", i, off);
883 if (i == subprog_end - 1) {
884 /* to avoid fall-through from one subprog into another
885 * the last insn of the subprog should be either exit
886 * or unconditional jump back
888 if (code != (BPF_JMP | BPF_EXIT) &&
889 code != (BPF_JMP | BPF_JA)) {
890 verbose(env, "last insn is not an exit or jmp\n");
893 subprog_start = subprog_end;
895 if (cur_subprog < env->subprog_cnt)
896 subprog_end = subprog[cur_subprog + 1].start;
902 /* Parentage chain of this register (or stack slot) should take care of all
903 * issues like callee-saved registers, stack slot allocation time, etc.
905 static int mark_reg_read(struct bpf_verifier_env *env,
906 const struct bpf_reg_state *state,
907 struct bpf_reg_state *parent)
909 bool writes = parent == state->parent; /* Observe write marks */
912 /* if read wasn't screened by an earlier write ... */
913 if (writes && state->live & REG_LIVE_WRITTEN)
915 /* ... then we depend on parent's value */
916 parent->live |= REG_LIVE_READ;
918 parent = state->parent;
924 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
927 struct bpf_verifier_state *vstate = env->cur_state;
928 struct bpf_func_state *state = vstate->frame[vstate->curframe];
929 struct bpf_reg_state *regs = state->regs;
931 if (regno >= MAX_BPF_REG) {
932 verbose(env, "R%d is invalid\n", regno);
937 /* check whether register used as source operand can be read */
938 if (regs[regno].type == NOT_INIT) {
939 verbose(env, "R%d !read_ok\n", regno);
942 /* We don't need to worry about FP liveness because it's read-only */
943 if (regno != BPF_REG_FP)
944 return mark_reg_read(env, ®s[regno],
947 /* check whether register used as dest operand can be written to */
948 if (regno == BPF_REG_FP) {
949 verbose(env, "frame pointer is read only\n");
952 regs[regno].live |= REG_LIVE_WRITTEN;
954 mark_reg_unknown(env, regs, regno);
959 static bool is_spillable_regtype(enum bpf_reg_type type)
962 case PTR_TO_MAP_VALUE:
963 case PTR_TO_MAP_VALUE_OR_NULL:
967 case PTR_TO_PACKET_META:
968 case PTR_TO_PACKET_END:
969 case PTR_TO_FLOW_KEYS:
970 case CONST_PTR_TO_MAP:
977 /* Does this register contain a constant zero? */
978 static bool register_is_null(struct bpf_reg_state *reg)
980 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
983 /* check_stack_read/write functions track spill/fill of registers,
984 * stack boundary and alignment are checked in check_mem_access()
986 static int check_stack_write(struct bpf_verifier_env *env,
987 struct bpf_func_state *state, /* func where register points to */
988 int off, int size, int value_regno, int insn_idx)
990 struct bpf_func_state *cur; /* state of the current function */
991 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
992 enum bpf_reg_type type;
994 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
998 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
999 * so it's aligned access and [off, off + size) are within stack limits
1001 if (!env->allow_ptr_leaks &&
1002 state->stack[spi].slot_type[0] == STACK_SPILL &&
1003 size != BPF_REG_SIZE) {
1004 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1008 cur = env->cur_state->frame[env->cur_state->curframe];
1009 if (value_regno >= 0 &&
1010 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1012 /* register containing pointer is being spilled into stack */
1013 if (size != BPF_REG_SIZE) {
1014 verbose(env, "invalid size of register spill\n");
1018 if (state != cur && type == PTR_TO_STACK) {
1019 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1023 /* save register state */
1024 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1025 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1027 for (i = 0; i < BPF_REG_SIZE; i++) {
1028 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1029 !env->allow_ptr_leaks) {
1030 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1031 int soff = (-spi - 1) * BPF_REG_SIZE;
1033 /* detected reuse of integer stack slot with a pointer
1034 * which means either llvm is reusing stack slot or
1035 * an attacker is trying to exploit CVE-2018-3639
1036 * (speculative store bypass)
1037 * Have to sanitize that slot with preemptive
1040 if (*poff && *poff != soff) {
1041 /* disallow programs where single insn stores
1042 * into two different stack slots, since verifier
1043 * cannot sanitize them
1046 "insn %d cannot access two stack slots fp%d and fp%d",
1047 insn_idx, *poff, soff);
1052 state->stack[spi].slot_type[i] = STACK_SPILL;
1055 u8 type = STACK_MISC;
1057 /* regular write of data into stack destroys any spilled ptr */
1058 state->stack[spi].spilled_ptr.type = NOT_INIT;
1060 /* only mark the slot as written if all 8 bytes were written
1061 * otherwise read propagation may incorrectly stop too soon
1062 * when stack slots are partially written.
1063 * This heuristic means that read propagation will be
1064 * conservative, since it will add reg_live_read marks
1065 * to stack slots all the way to first state when programs
1066 * writes+reads less than 8 bytes
1068 if (size == BPF_REG_SIZE)
1069 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1071 /* when we zero initialize stack slots mark them as such */
1072 if (value_regno >= 0 &&
1073 register_is_null(&cur->regs[value_regno]))
1076 for (i = 0; i < size; i++)
1077 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1083 static int check_stack_read(struct bpf_verifier_env *env,
1084 struct bpf_func_state *reg_state /* func where register points to */,
1085 int off, int size, int value_regno)
1087 struct bpf_verifier_state *vstate = env->cur_state;
1088 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1089 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1092 if (reg_state->allocated_stack <= slot) {
1093 verbose(env, "invalid read from stack off %d+0 size %d\n",
1097 stype = reg_state->stack[spi].slot_type;
1099 if (stype[0] == STACK_SPILL) {
1100 if (size != BPF_REG_SIZE) {
1101 verbose(env, "invalid size of register spill\n");
1104 for (i = 1; i < BPF_REG_SIZE; i++) {
1105 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1106 verbose(env, "corrupted spill memory\n");
1111 if (value_regno >= 0) {
1112 /* restore register state from stack */
1113 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1114 /* mark reg as written since spilled pointer state likely
1115 * has its liveness marks cleared by is_state_visited()
1116 * which resets stack/reg liveness for state transitions
1118 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1120 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1121 reg_state->stack[spi].spilled_ptr.parent);
1126 for (i = 0; i < size; i++) {
1127 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1129 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1133 verbose(env, "invalid read from stack off %d+%d size %d\n",
1137 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1138 reg_state->stack[spi].spilled_ptr.parent);
1139 if (value_regno >= 0) {
1140 if (zeros == size) {
1141 /* any size read into register is zero extended,
1142 * so the whole register == const_zero
1144 __mark_reg_const_zero(&state->regs[value_regno]);
1146 /* have read misc data from the stack */
1147 mark_reg_unknown(env, state->regs, value_regno);
1149 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1155 /* check read/write into map element returned by bpf_map_lookup_elem() */
1156 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1157 int size, bool zero_size_allowed)
1159 struct bpf_reg_state *regs = cur_regs(env);
1160 struct bpf_map *map = regs[regno].map_ptr;
1162 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1163 off + size > map->value_size) {
1164 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1165 map->value_size, off, size);
1171 /* check read/write into a map element with possible variable offset */
1172 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1173 int off, int size, bool zero_size_allowed)
1175 struct bpf_verifier_state *vstate = env->cur_state;
1176 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1177 struct bpf_reg_state *reg = &state->regs[regno];
1180 /* We may have adjusted the register to this map value, so we
1181 * need to try adding each of min_value and max_value to off
1182 * to make sure our theoretical access will be safe.
1185 print_verifier_state(env, state);
1186 /* The minimum value is only important with signed
1187 * comparisons where we can't assume the floor of a
1188 * value is 0. If we are using signed variables for our
1189 * index'es we need to make sure that whatever we use
1190 * will have a set floor within our range.
1192 if (reg->smin_value < 0) {
1193 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1197 err = __check_map_access(env, regno, reg->smin_value + off, size,
1200 verbose(env, "R%d min value is outside of the array range\n",
1205 /* If we haven't set a max value then we need to bail since we can't be
1206 * sure we won't do bad things.
1207 * If reg->umax_value + off could overflow, treat that as unbounded too.
1209 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1210 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1214 err = __check_map_access(env, regno, reg->umax_value + off, size,
1217 verbose(env, "R%d max value is outside of the array range\n",
1222 #define MAX_PACKET_OFF 0xffff
1224 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1225 const struct bpf_call_arg_meta *meta,
1226 enum bpf_access_type t)
1228 switch (env->prog->type) {
1229 case BPF_PROG_TYPE_LWT_IN:
1230 case BPF_PROG_TYPE_LWT_OUT:
1231 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1232 case BPF_PROG_TYPE_SK_REUSEPORT:
1233 /* dst_input() and dst_output() can't write for now */
1237 case BPF_PROG_TYPE_SCHED_CLS:
1238 case BPF_PROG_TYPE_SCHED_ACT:
1239 case BPF_PROG_TYPE_XDP:
1240 case BPF_PROG_TYPE_LWT_XMIT:
1241 case BPF_PROG_TYPE_SK_SKB:
1242 case BPF_PROG_TYPE_SK_MSG:
1243 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1245 return meta->pkt_access;
1247 env->seen_direct_write = true;
1254 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1255 int off, int size, bool zero_size_allowed)
1257 struct bpf_reg_state *regs = cur_regs(env);
1258 struct bpf_reg_state *reg = ®s[regno];
1260 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1261 (u64)off + size > reg->range) {
1262 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1263 off, size, regno, reg->id, reg->off, reg->range);
1269 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1270 int size, bool zero_size_allowed)
1272 struct bpf_reg_state *regs = cur_regs(env);
1273 struct bpf_reg_state *reg = ®s[regno];
1276 /* We may have added a variable offset to the packet pointer; but any
1277 * reg->range we have comes after that. We are only checking the fixed
1281 /* We don't allow negative numbers, because we aren't tracking enough
1282 * detail to prove they're safe.
1284 if (reg->smin_value < 0) {
1285 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1289 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1291 verbose(env, "R%d offset is outside of the packet\n", regno);
1297 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1298 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1299 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1301 struct bpf_insn_access_aux info = {
1302 .reg_type = *reg_type,
1305 if (env->ops->is_valid_access &&
1306 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1307 /* A non zero info.ctx_field_size indicates that this field is a
1308 * candidate for later verifier transformation to load the whole
1309 * field and then apply a mask when accessed with a narrower
1310 * access than actual ctx access size. A zero info.ctx_field_size
1311 * will only allow for whole field access and rejects any other
1312 * type of narrower access.
1314 *reg_type = info.reg_type;
1316 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1317 /* remember the offset of last byte accessed in ctx */
1318 if (env->prog->aux->max_ctx_offset < off + size)
1319 env->prog->aux->max_ctx_offset = off + size;
1323 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1327 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1330 if (size < 0 || off < 0 ||
1331 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1332 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1339 static bool __is_pointer_value(bool allow_ptr_leaks,
1340 const struct bpf_reg_state *reg)
1342 if (allow_ptr_leaks)
1345 return reg->type != SCALAR_VALUE;
1348 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1350 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1353 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1355 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1357 return reg->type == PTR_TO_CTX;
1360 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1362 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1364 return type_is_pkt_pointer(reg->type);
1367 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1368 const struct bpf_reg_state *reg,
1369 int off, int size, bool strict)
1371 struct tnum reg_off;
1374 /* Byte size accesses are always allowed. */
1375 if (!strict || size == 1)
1378 /* For platforms that do not have a Kconfig enabling
1379 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1380 * NET_IP_ALIGN is universally set to '2'. And on platforms
1381 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1382 * to this code only in strict mode where we want to emulate
1383 * the NET_IP_ALIGN==2 checking. Therefore use an
1384 * unconditional IP align value of '2'.
1388 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1389 if (!tnum_is_aligned(reg_off, size)) {
1392 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1394 "misaligned packet access off %d+%s+%d+%d size %d\n",
1395 ip_align, tn_buf, reg->off, off, size);
1402 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1403 const struct bpf_reg_state *reg,
1404 const char *pointer_desc,
1405 int off, int size, bool strict)
1407 struct tnum reg_off;
1409 /* Byte size accesses are always allowed. */
1410 if (!strict || size == 1)
1413 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1414 if (!tnum_is_aligned(reg_off, size)) {
1417 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1418 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1419 pointer_desc, tn_buf, reg->off, off, size);
1426 static int check_ptr_alignment(struct bpf_verifier_env *env,
1427 const struct bpf_reg_state *reg, int off,
1428 int size, bool strict_alignment_once)
1430 bool strict = env->strict_alignment || strict_alignment_once;
1431 const char *pointer_desc = "";
1433 switch (reg->type) {
1435 case PTR_TO_PACKET_META:
1436 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1437 * right in front, treat it the very same way.
1439 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1440 case PTR_TO_FLOW_KEYS:
1441 pointer_desc = "flow keys ";
1443 case PTR_TO_MAP_VALUE:
1444 pointer_desc = "value ";
1447 pointer_desc = "context ";
1450 pointer_desc = "stack ";
1451 /* The stack spill tracking logic in check_stack_write()
1452 * and check_stack_read() relies on stack accesses being
1460 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1464 static int update_stack_depth(struct bpf_verifier_env *env,
1465 const struct bpf_func_state *func,
1468 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1473 /* update known max for given subprogram */
1474 env->subprog_info[func->subprogno].stack_depth = -off;
1478 /* starting from main bpf function walk all instructions of the function
1479 * and recursively walk all callees that given function can call.
1480 * Ignore jump and exit insns.
1481 * Since recursion is prevented by check_cfg() this algorithm
1482 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1484 static int check_max_stack_depth(struct bpf_verifier_env *env)
1486 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1487 struct bpf_subprog_info *subprog = env->subprog_info;
1488 struct bpf_insn *insn = env->prog->insnsi;
1489 int ret_insn[MAX_CALL_FRAMES];
1490 int ret_prog[MAX_CALL_FRAMES];
1493 /* round up to 32-bytes, since this is granularity
1494 * of interpreter stack size
1496 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1497 if (depth > MAX_BPF_STACK) {
1498 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1503 subprog_end = subprog[idx + 1].start;
1504 for (; i < subprog_end; i++) {
1505 if (insn[i].code != (BPF_JMP | BPF_CALL))
1507 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1509 /* remember insn and function to return to */
1510 ret_insn[frame] = i + 1;
1511 ret_prog[frame] = idx;
1513 /* find the callee */
1514 i = i + insn[i].imm + 1;
1515 idx = find_subprog(env, i);
1517 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1522 if (frame >= MAX_CALL_FRAMES) {
1523 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1528 /* end of for() loop means the last insn of the 'subprog'
1529 * was reached. Doesn't matter whether it was JA or EXIT
1533 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1535 i = ret_insn[frame];
1536 idx = ret_prog[frame];
1540 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1541 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1542 const struct bpf_insn *insn, int idx)
1544 int start = idx + insn->imm + 1, subprog;
1546 subprog = find_subprog(env, start);
1548 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1552 return env->subprog_info[subprog].stack_depth;
1556 static int check_ctx_reg(struct bpf_verifier_env *env,
1557 const struct bpf_reg_state *reg, int regno)
1559 /* Access to ctx or passing it to a helper is only allowed in
1560 * its original, unmodified form.
1564 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1569 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1572 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1573 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1580 /* truncate register to smaller size (in bytes)
1581 * must be called with size < BPF_REG_SIZE
1583 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1587 /* clear high bits in bit representation */
1588 reg->var_off = tnum_cast(reg->var_off, size);
1590 /* fix arithmetic bounds */
1591 mask = ((u64)1 << (size * 8)) - 1;
1592 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1593 reg->umin_value &= mask;
1594 reg->umax_value &= mask;
1596 reg->umin_value = 0;
1597 reg->umax_value = mask;
1599 reg->smin_value = reg->umin_value;
1600 reg->smax_value = reg->umax_value;
1603 /* check whether memory at (regno + off) is accessible for t = (read | write)
1604 * if t==write, value_regno is a register which value is stored into memory
1605 * if t==read, value_regno is a register which will receive the value from memory
1606 * if t==write && value_regno==-1, some unknown value is stored into memory
1607 * if t==read && value_regno==-1, don't care what we read from memory
1609 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1610 int off, int bpf_size, enum bpf_access_type t,
1611 int value_regno, bool strict_alignment_once)
1613 struct bpf_reg_state *regs = cur_regs(env);
1614 struct bpf_reg_state *reg = regs + regno;
1615 struct bpf_func_state *state;
1618 size = bpf_size_to_bytes(bpf_size);
1622 /* alignment checks will add in reg->off themselves */
1623 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1627 /* for access checks, reg->off is just part of off */
1630 if (reg->type == PTR_TO_MAP_VALUE) {
1631 if (t == BPF_WRITE && value_regno >= 0 &&
1632 is_pointer_value(env, value_regno)) {
1633 verbose(env, "R%d leaks addr into map\n", value_regno);
1637 err = check_map_access(env, regno, off, size, false);
1638 if (!err && t == BPF_READ && value_regno >= 0)
1639 mark_reg_unknown(env, regs, value_regno);
1641 } else if (reg->type == PTR_TO_CTX) {
1642 enum bpf_reg_type reg_type = SCALAR_VALUE;
1644 if (t == BPF_WRITE && value_regno >= 0 &&
1645 is_pointer_value(env, value_regno)) {
1646 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1650 err = check_ctx_reg(env, reg, regno);
1654 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1655 if (!err && t == BPF_READ && value_regno >= 0) {
1656 /* ctx access returns either a scalar, or a
1657 * PTR_TO_PACKET[_META,_END]. In the latter
1658 * case, we know the offset is zero.
1660 if (reg_type == SCALAR_VALUE)
1661 mark_reg_unknown(env, regs, value_regno);
1663 mark_reg_known_zero(env, regs,
1665 regs[value_regno].type = reg_type;
1668 } else if (reg->type == PTR_TO_STACK) {
1669 /* stack accesses must be at a fixed offset, so that we can
1670 * determine what type of data were returned.
1671 * See check_stack_read().
1673 if (!tnum_is_const(reg->var_off)) {
1676 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1677 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1681 off += reg->var_off.value;
1682 if (off >= 0 || off < -MAX_BPF_STACK) {
1683 verbose(env, "invalid stack off=%d size=%d\n", off,
1688 state = func(env, reg);
1689 err = update_stack_depth(env, state, off);
1694 err = check_stack_write(env, state, off, size,
1695 value_regno, insn_idx);
1697 err = check_stack_read(env, state, off, size,
1699 } else if (reg_is_pkt_pointer(reg)) {
1700 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1701 verbose(env, "cannot write into packet\n");
1704 if (t == BPF_WRITE && value_regno >= 0 &&
1705 is_pointer_value(env, value_regno)) {
1706 verbose(env, "R%d leaks addr into packet\n",
1710 err = check_packet_access(env, regno, off, size, false);
1711 if (!err && t == BPF_READ && value_regno >= 0)
1712 mark_reg_unknown(env, regs, value_regno);
1713 } else if (reg->type == PTR_TO_FLOW_KEYS) {
1714 if (t == BPF_WRITE && value_regno >= 0 &&
1715 is_pointer_value(env, value_regno)) {
1716 verbose(env, "R%d leaks addr into flow keys\n",
1721 err = check_flow_keys_access(env, off, size);
1722 if (!err && t == BPF_READ && value_regno >= 0)
1723 mark_reg_unknown(env, regs, value_regno);
1725 verbose(env, "R%d invalid mem access '%s'\n", regno,
1726 reg_type_str[reg->type]);
1730 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1731 regs[value_regno].type == SCALAR_VALUE) {
1732 /* b/h/w load zero-extends, mark upper bits as known 0 */
1733 coerce_reg_to_size(®s[value_regno], size);
1738 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1742 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1744 verbose(env, "BPF_XADD uses reserved fields\n");
1748 /* check src1 operand */
1749 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1753 /* check src2 operand */
1754 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1758 if (is_pointer_value(env, insn->src_reg)) {
1759 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1763 if (is_ctx_reg(env, insn->dst_reg) ||
1764 is_pkt_reg(env, insn->dst_reg)) {
1765 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1766 insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ?
1767 "context" : "packet");
1771 /* check whether atomic_add can read the memory */
1772 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1773 BPF_SIZE(insn->code), BPF_READ, -1, true);
1777 /* check whether atomic_add can write into the same memory */
1778 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1779 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1782 /* when register 'regno' is passed into function that will read 'access_size'
1783 * bytes from that pointer, make sure that it's within stack boundary
1784 * and all elements of stack are initialized.
1785 * Unlike most pointer bounds-checking functions, this one doesn't take an
1786 * 'off' argument, so it has to add in reg->off itself.
1788 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1789 int access_size, bool zero_size_allowed,
1790 struct bpf_call_arg_meta *meta)
1792 struct bpf_reg_state *reg = cur_regs(env) + regno;
1793 struct bpf_func_state *state = func(env, reg);
1794 int off, i, slot, spi;
1796 if (reg->type != PTR_TO_STACK) {
1797 /* Allow zero-byte read from NULL, regardless of pointer type */
1798 if (zero_size_allowed && access_size == 0 &&
1799 register_is_null(reg))
1802 verbose(env, "R%d type=%s expected=%s\n", regno,
1803 reg_type_str[reg->type],
1804 reg_type_str[PTR_TO_STACK]);
1808 /* Only allow fixed-offset stack reads */
1809 if (!tnum_is_const(reg->var_off)) {
1812 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1813 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1817 off = reg->off + reg->var_off.value;
1818 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1819 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1820 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1821 regno, off, access_size);
1825 if (meta && meta->raw_mode) {
1826 meta->access_size = access_size;
1827 meta->regno = regno;
1831 for (i = 0; i < access_size; i++) {
1834 slot = -(off + i) - 1;
1835 spi = slot / BPF_REG_SIZE;
1836 if (state->allocated_stack <= slot)
1838 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
1839 if (*stype == STACK_MISC)
1841 if (*stype == STACK_ZERO) {
1842 /* helper can write anything into the stack */
1843 *stype = STACK_MISC;
1847 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1848 off, i, access_size);
1851 /* reading any byte out of 8-byte 'spill_slot' will cause
1852 * the whole slot to be marked as 'read'
1854 mark_reg_read(env, &state->stack[spi].spilled_ptr,
1855 state->stack[spi].spilled_ptr.parent);
1857 return update_stack_depth(env, state, off);
1860 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1861 int access_size, bool zero_size_allowed,
1862 struct bpf_call_arg_meta *meta)
1864 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1866 switch (reg->type) {
1868 case PTR_TO_PACKET_META:
1869 return check_packet_access(env, regno, reg->off, access_size,
1871 case PTR_TO_FLOW_KEYS:
1872 return check_flow_keys_access(env, reg->off, access_size);
1873 case PTR_TO_MAP_VALUE:
1874 return check_map_access(env, regno, reg->off, access_size,
1876 default: /* scalar_value|ptr_to_stack or invalid ptr */
1877 return check_stack_boundary(env, regno, access_size,
1878 zero_size_allowed, meta);
1882 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
1884 return type == ARG_PTR_TO_MEM ||
1885 type == ARG_PTR_TO_MEM_OR_NULL ||
1886 type == ARG_PTR_TO_UNINIT_MEM;
1889 static bool arg_type_is_mem_size(enum bpf_arg_type type)
1891 return type == ARG_CONST_SIZE ||
1892 type == ARG_CONST_SIZE_OR_ZERO;
1895 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1896 enum bpf_arg_type arg_type,
1897 struct bpf_call_arg_meta *meta)
1899 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1900 enum bpf_reg_type expected_type, type = reg->type;
1903 if (arg_type == ARG_DONTCARE)
1906 err = check_reg_arg(env, regno, SRC_OP);
1910 if (arg_type == ARG_ANYTHING) {
1911 if (is_pointer_value(env, regno)) {
1912 verbose(env, "R%d leaks addr into helper function\n",
1919 if (type_is_pkt_pointer(type) &&
1920 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1921 verbose(env, "helper access to the packet is not allowed\n");
1925 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1926 arg_type == ARG_PTR_TO_MAP_VALUE) {
1927 expected_type = PTR_TO_STACK;
1928 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
1929 type != expected_type)
1931 } else if (arg_type == ARG_CONST_SIZE ||
1932 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1933 expected_type = SCALAR_VALUE;
1934 if (type != expected_type)
1936 } else if (arg_type == ARG_CONST_MAP_PTR) {
1937 expected_type = CONST_PTR_TO_MAP;
1938 if (type != expected_type)
1940 } else if (arg_type == ARG_PTR_TO_CTX) {
1941 expected_type = PTR_TO_CTX;
1942 if (type != expected_type)
1944 err = check_ctx_reg(env, reg, regno);
1947 } else if (arg_type_is_mem_ptr(arg_type)) {
1948 expected_type = PTR_TO_STACK;
1949 /* One exception here. In case function allows for NULL to be
1950 * passed in as argument, it's a SCALAR_VALUE type. Final test
1951 * happens during stack boundary checking.
1953 if (register_is_null(reg) &&
1954 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1955 /* final test in check_stack_boundary() */;
1956 else if (!type_is_pkt_pointer(type) &&
1957 type != PTR_TO_MAP_VALUE &&
1958 type != expected_type)
1960 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1962 verbose(env, "unsupported arg_type %d\n", arg_type);
1966 if (arg_type == ARG_CONST_MAP_PTR) {
1967 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1968 meta->map_ptr = reg->map_ptr;
1969 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1970 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1971 * check that [key, key + map->key_size) are within
1972 * stack limits and initialized
1974 if (!meta->map_ptr) {
1975 /* in function declaration map_ptr must come before
1976 * map_key, so that it's verified and known before
1977 * we have to check map_key here. Otherwise it means
1978 * that kernel subsystem misconfigured verifier
1980 verbose(env, "invalid map_ptr to access map->key\n");
1983 err = check_helper_mem_access(env, regno,
1984 meta->map_ptr->key_size, false,
1986 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1987 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1988 * check [value, value + map->value_size) validity
1990 if (!meta->map_ptr) {
1991 /* kernel subsystem misconfigured verifier */
1992 verbose(env, "invalid map_ptr to access map->value\n");
1995 err = check_helper_mem_access(env, regno,
1996 meta->map_ptr->value_size, false,
1998 } else if (arg_type_is_mem_size(arg_type)) {
1999 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2001 /* remember the mem_size which may be used later
2002 * to refine return values.
2004 meta->msize_smax_value = reg->smax_value;
2005 meta->msize_umax_value = reg->umax_value;
2007 /* The register is SCALAR_VALUE; the access check
2008 * happens using its boundaries.
2010 if (!tnum_is_const(reg->var_off))
2011 /* For unprivileged variable accesses, disable raw
2012 * mode so that the program is required to
2013 * initialize all the memory that the helper could
2014 * just partially fill up.
2018 if (reg->smin_value < 0) {
2019 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2024 if (reg->umin_value == 0) {
2025 err = check_helper_mem_access(env, regno - 1, 0,
2032 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2033 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2037 err = check_helper_mem_access(env, regno - 1,
2039 zero_size_allowed, meta);
2044 verbose(env, "R%d type=%s expected=%s\n", regno,
2045 reg_type_str[type], reg_type_str[expected_type]);
2049 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2050 struct bpf_map *map, int func_id)
2055 /* We need a two way check, first is from map perspective ... */
2056 switch (map->map_type) {
2057 case BPF_MAP_TYPE_PROG_ARRAY:
2058 if (func_id != BPF_FUNC_tail_call)
2061 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2062 if (func_id != BPF_FUNC_perf_event_read &&
2063 func_id != BPF_FUNC_perf_event_output &&
2064 func_id != BPF_FUNC_perf_event_read_value)
2067 case BPF_MAP_TYPE_STACK_TRACE:
2068 if (func_id != BPF_FUNC_get_stackid)
2071 case BPF_MAP_TYPE_CGROUP_ARRAY:
2072 if (func_id != BPF_FUNC_skb_under_cgroup &&
2073 func_id != BPF_FUNC_current_task_under_cgroup)
2076 case BPF_MAP_TYPE_CGROUP_STORAGE:
2077 if (func_id != BPF_FUNC_get_local_storage)
2080 /* devmap returns a pointer to a live net_device ifindex that we cannot
2081 * allow to be modified from bpf side. So do not allow lookup elements
2084 case BPF_MAP_TYPE_DEVMAP:
2085 if (func_id != BPF_FUNC_redirect_map)
2088 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2091 case BPF_MAP_TYPE_CPUMAP:
2092 case BPF_MAP_TYPE_XSKMAP:
2093 if (func_id != BPF_FUNC_redirect_map)
2096 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2097 case BPF_MAP_TYPE_HASH_OF_MAPS:
2098 if (func_id != BPF_FUNC_map_lookup_elem)
2101 case BPF_MAP_TYPE_SOCKMAP:
2102 if (func_id != BPF_FUNC_sk_redirect_map &&
2103 func_id != BPF_FUNC_sock_map_update &&
2104 func_id != BPF_FUNC_map_delete_elem &&
2105 func_id != BPF_FUNC_msg_redirect_map)
2108 case BPF_MAP_TYPE_SOCKHASH:
2109 if (func_id != BPF_FUNC_sk_redirect_hash &&
2110 func_id != BPF_FUNC_sock_hash_update &&
2111 func_id != BPF_FUNC_map_delete_elem &&
2112 func_id != BPF_FUNC_msg_redirect_hash)
2115 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2116 if (func_id != BPF_FUNC_sk_select_reuseport)
2123 /* ... and second from the function itself. */
2125 case BPF_FUNC_tail_call:
2126 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2128 if (env->subprog_cnt > 1) {
2129 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2133 case BPF_FUNC_perf_event_read:
2134 case BPF_FUNC_perf_event_output:
2135 case BPF_FUNC_perf_event_read_value:
2136 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2139 case BPF_FUNC_get_stackid:
2140 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2143 case BPF_FUNC_current_task_under_cgroup:
2144 case BPF_FUNC_skb_under_cgroup:
2145 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2148 case BPF_FUNC_redirect_map:
2149 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2150 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2151 map->map_type != BPF_MAP_TYPE_XSKMAP)
2154 case BPF_FUNC_sk_redirect_map:
2155 case BPF_FUNC_msg_redirect_map:
2156 case BPF_FUNC_sock_map_update:
2157 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2160 case BPF_FUNC_sk_redirect_hash:
2161 case BPF_FUNC_msg_redirect_hash:
2162 case BPF_FUNC_sock_hash_update:
2163 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2166 case BPF_FUNC_get_local_storage:
2167 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE)
2170 case BPF_FUNC_sk_select_reuseport:
2171 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2180 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2181 map->map_type, func_id_name(func_id), func_id);
2185 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2189 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2191 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2193 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2195 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2197 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2200 /* We only support one arg being in raw mode at the moment,
2201 * which is sufficient for the helper functions we have
2207 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2208 enum bpf_arg_type arg_next)
2210 return (arg_type_is_mem_ptr(arg_curr) &&
2211 !arg_type_is_mem_size(arg_next)) ||
2212 (!arg_type_is_mem_ptr(arg_curr) &&
2213 arg_type_is_mem_size(arg_next));
2216 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2218 /* bpf_xxx(..., buf, len) call will access 'len'
2219 * bytes from memory 'buf'. Both arg types need
2220 * to be paired, so make sure there's no buggy
2221 * helper function specification.
2223 if (arg_type_is_mem_size(fn->arg1_type) ||
2224 arg_type_is_mem_ptr(fn->arg5_type) ||
2225 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2226 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2227 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2228 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2234 static int check_func_proto(const struct bpf_func_proto *fn)
2236 return check_raw_mode_ok(fn) &&
2237 check_arg_pair_ok(fn) ? 0 : -EINVAL;
2240 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2241 * are now invalid, so turn them into unknown SCALAR_VALUE.
2243 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2244 struct bpf_func_state *state)
2246 struct bpf_reg_state *regs = state->regs, *reg;
2249 for (i = 0; i < MAX_BPF_REG; i++)
2250 if (reg_is_pkt_pointer_any(®s[i]))
2251 mark_reg_unknown(env, regs, i);
2253 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2254 if (state->stack[i].slot_type[0] != STACK_SPILL)
2256 reg = &state->stack[i].spilled_ptr;
2257 if (reg_is_pkt_pointer_any(reg))
2258 __mark_reg_unknown(reg);
2262 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2264 struct bpf_verifier_state *vstate = env->cur_state;
2267 for (i = 0; i <= vstate->curframe; i++)
2268 __clear_all_pkt_pointers(env, vstate->frame[i]);
2271 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2274 struct bpf_verifier_state *state = env->cur_state;
2275 struct bpf_func_state *caller, *callee;
2276 int i, subprog, target_insn;
2278 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2279 verbose(env, "the call stack of %d frames is too deep\n",
2280 state->curframe + 2);
2284 target_insn = *insn_idx + insn->imm;
2285 subprog = find_subprog(env, target_insn + 1);
2287 verbose(env, "verifier bug. No program starts at insn %d\n",
2292 caller = state->frame[state->curframe];
2293 if (state->frame[state->curframe + 1]) {
2294 verbose(env, "verifier bug. Frame %d already allocated\n",
2295 state->curframe + 1);
2299 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2302 state->frame[state->curframe + 1] = callee;
2304 /* callee cannot access r0, r6 - r9 for reading and has to write
2305 * into its own stack before reading from it.
2306 * callee can read/write into caller's stack
2308 init_func_state(env, callee,
2309 /* remember the callsite, it will be used by bpf_exit */
2310 *insn_idx /* callsite */,
2311 state->curframe + 1 /* frameno within this callchain */,
2312 subprog /* subprog number within this prog */);
2314 /* copy r1 - r5 args that callee can access. The copy includes parent
2315 * pointers, which connects us up to the liveness chain
2317 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2318 callee->regs[i] = caller->regs[i];
2320 /* after the call registers r0 - r5 were scratched */
2321 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2322 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2323 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2326 /* only increment it after check_reg_arg() finished */
2329 /* and go analyze first insn of the callee */
2330 *insn_idx = target_insn;
2332 if (env->log.level) {
2333 verbose(env, "caller:\n");
2334 print_verifier_state(env, caller);
2335 verbose(env, "callee:\n");
2336 print_verifier_state(env, callee);
2341 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2343 struct bpf_verifier_state *state = env->cur_state;
2344 struct bpf_func_state *caller, *callee;
2345 struct bpf_reg_state *r0;
2347 callee = state->frame[state->curframe];
2348 r0 = &callee->regs[BPF_REG_0];
2349 if (r0->type == PTR_TO_STACK) {
2350 /* technically it's ok to return caller's stack pointer
2351 * (or caller's caller's pointer) back to the caller,
2352 * since these pointers are valid. Only current stack
2353 * pointer will be invalid as soon as function exits,
2354 * but let's be conservative
2356 verbose(env, "cannot return stack pointer to the caller\n");
2361 caller = state->frame[state->curframe];
2362 /* return to the caller whatever r0 had in the callee */
2363 caller->regs[BPF_REG_0] = *r0;
2365 *insn_idx = callee->callsite + 1;
2366 if (env->log.level) {
2367 verbose(env, "returning from callee:\n");
2368 print_verifier_state(env, callee);
2369 verbose(env, "to caller at %d:\n", *insn_idx);
2370 print_verifier_state(env, caller);
2372 /* clear everything in the callee */
2373 free_func_state(callee);
2374 state->frame[state->curframe + 1] = NULL;
2378 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2380 struct bpf_call_arg_meta *meta)
2382 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2384 if (ret_type != RET_INTEGER ||
2385 (func_id != BPF_FUNC_get_stack &&
2386 func_id != BPF_FUNC_probe_read_str))
2389 ret_reg->smax_value = meta->msize_smax_value;
2390 ret_reg->umax_value = meta->msize_umax_value;
2391 __reg_deduce_bounds(ret_reg);
2392 __reg_bound_offset(ret_reg);
2396 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2397 int func_id, int insn_idx)
2399 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2401 if (func_id != BPF_FUNC_tail_call &&
2402 func_id != BPF_FUNC_map_lookup_elem &&
2403 func_id != BPF_FUNC_map_update_elem &&
2404 func_id != BPF_FUNC_map_delete_elem)
2407 if (meta->map_ptr == NULL) {
2408 verbose(env, "kernel subsystem misconfigured verifier\n");
2412 if (!BPF_MAP_PTR(aux->map_state))
2413 bpf_map_ptr_store(aux, meta->map_ptr,
2414 meta->map_ptr->unpriv_array);
2415 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2416 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2417 meta->map_ptr->unpriv_array);
2421 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2423 const struct bpf_func_proto *fn = NULL;
2424 struct bpf_reg_state *regs;
2425 struct bpf_call_arg_meta meta;
2429 /* find function prototype */
2430 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2431 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2436 if (env->ops->get_func_proto)
2437 fn = env->ops->get_func_proto(func_id, env->prog);
2439 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2444 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2445 if (!env->prog->gpl_compatible && fn->gpl_only) {
2446 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2450 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2451 changes_data = bpf_helper_changes_pkt_data(fn->func);
2452 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2453 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2454 func_id_name(func_id), func_id);
2458 memset(&meta, 0, sizeof(meta));
2459 meta.pkt_access = fn->pkt_access;
2461 err = check_func_proto(fn);
2463 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2464 func_id_name(func_id), func_id);
2469 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2472 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2475 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2478 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2481 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2485 err = record_func_map(env, &meta, func_id, insn_idx);
2489 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2490 * is inferred from register state.
2492 for (i = 0; i < meta.access_size; i++) {
2493 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2494 BPF_WRITE, -1, false);
2499 regs = cur_regs(env);
2501 /* check that flags argument in get_local_storage(map, flags) is 0,
2502 * this is required because get_local_storage() can't return an error.
2504 if (func_id == BPF_FUNC_get_local_storage &&
2505 !register_is_null(®s[BPF_REG_2])) {
2506 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2510 /* reset caller saved regs */
2511 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2512 mark_reg_not_init(env, regs, caller_saved[i]);
2513 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2516 /* update return register (already marked as written above) */
2517 if (fn->ret_type == RET_INTEGER) {
2518 /* sets type to SCALAR_VALUE */
2519 mark_reg_unknown(env, regs, BPF_REG_0);
2520 } else if (fn->ret_type == RET_VOID) {
2521 regs[BPF_REG_0].type = NOT_INIT;
2522 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2523 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2524 if (fn->ret_type == RET_PTR_TO_MAP_VALUE)
2525 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2527 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2528 /* There is no offset yet applied, variable or fixed */
2529 mark_reg_known_zero(env, regs, BPF_REG_0);
2530 /* remember map_ptr, so that check_map_access()
2531 * can check 'value_size' boundary of memory access
2532 * to map element returned from bpf_map_lookup_elem()
2534 if (meta.map_ptr == NULL) {
2536 "kernel subsystem misconfigured verifier\n");
2539 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2540 regs[BPF_REG_0].id = ++env->id_gen;
2542 verbose(env, "unknown return type %d of func %s#%d\n",
2543 fn->ret_type, func_id_name(func_id), func_id);
2547 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2549 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2553 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2554 const char *err_str;
2556 #ifdef CONFIG_PERF_EVENTS
2557 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2558 err_str = "cannot get callchain buffer for func %s#%d\n";
2561 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2564 verbose(env, err_str, func_id_name(func_id), func_id);
2568 env->prog->has_callchain_buf = true;
2572 clear_all_pkt_pointers(env);
2576 static bool signed_add_overflows(s64 a, s64 b)
2578 /* Do the add in u64, where overflow is well-defined */
2579 s64 res = (s64)((u64)a + (u64)b);
2586 static bool signed_sub_overflows(s64 a, s64 b)
2588 /* Do the sub in u64, where overflow is well-defined */
2589 s64 res = (s64)((u64)a - (u64)b);
2596 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2597 const struct bpf_reg_state *reg,
2598 enum bpf_reg_type type)
2600 bool known = tnum_is_const(reg->var_off);
2601 s64 val = reg->var_off.value;
2602 s64 smin = reg->smin_value;
2604 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2605 verbose(env, "math between %s pointer and %lld is not allowed\n",
2606 reg_type_str[type], val);
2610 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2611 verbose(env, "%s pointer offset %d is not allowed\n",
2612 reg_type_str[type], reg->off);
2616 if (smin == S64_MIN) {
2617 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2618 reg_type_str[type]);
2622 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2623 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2624 smin, reg_type_str[type]);
2631 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2632 * Caller should also handle BPF_MOV case separately.
2633 * If we return -EACCES, caller may want to try again treating pointer as a
2634 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2636 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2637 struct bpf_insn *insn,
2638 const struct bpf_reg_state *ptr_reg,
2639 const struct bpf_reg_state *off_reg)
2641 struct bpf_verifier_state *vstate = env->cur_state;
2642 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2643 struct bpf_reg_state *regs = state->regs, *dst_reg;
2644 bool known = tnum_is_const(off_reg->var_off);
2645 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2646 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2647 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2648 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2649 u8 opcode = BPF_OP(insn->code);
2650 u32 dst = insn->dst_reg;
2652 dst_reg = ®s[dst];
2654 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2655 smin_val > smax_val || umin_val > umax_val) {
2656 /* Taint dst register if offset had invalid bounds derived from
2657 * e.g. dead branches.
2659 __mark_reg_unknown(dst_reg);
2663 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2664 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2666 "R%d 32-bit pointer arithmetic prohibited\n",
2671 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2672 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2676 if (ptr_reg->type == CONST_PTR_TO_MAP) {
2677 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2681 if (ptr_reg->type == PTR_TO_PACKET_END) {
2682 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2687 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2688 * The id may be overwritten later if we create a new variable offset.
2690 dst_reg->type = ptr_reg->type;
2691 dst_reg->id = ptr_reg->id;
2693 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
2694 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
2699 /* We can take a fixed offset as long as it doesn't overflow
2700 * the s32 'off' field
2702 if (known && (ptr_reg->off + smin_val ==
2703 (s64)(s32)(ptr_reg->off + smin_val))) {
2704 /* pointer += K. Accumulate it into fixed offset */
2705 dst_reg->smin_value = smin_ptr;
2706 dst_reg->smax_value = smax_ptr;
2707 dst_reg->umin_value = umin_ptr;
2708 dst_reg->umax_value = umax_ptr;
2709 dst_reg->var_off = ptr_reg->var_off;
2710 dst_reg->off = ptr_reg->off + smin_val;
2711 dst_reg->range = ptr_reg->range;
2714 /* A new variable offset is created. Note that off_reg->off
2715 * == 0, since it's a scalar.
2716 * dst_reg gets the pointer type and since some positive
2717 * integer value was added to the pointer, give it a new 'id'
2718 * if it's a PTR_TO_PACKET.
2719 * this creates a new 'base' pointer, off_reg (variable) gets
2720 * added into the variable offset, and we copy the fixed offset
2723 if (signed_add_overflows(smin_ptr, smin_val) ||
2724 signed_add_overflows(smax_ptr, smax_val)) {
2725 dst_reg->smin_value = S64_MIN;
2726 dst_reg->smax_value = S64_MAX;
2728 dst_reg->smin_value = smin_ptr + smin_val;
2729 dst_reg->smax_value = smax_ptr + smax_val;
2731 if (umin_ptr + umin_val < umin_ptr ||
2732 umax_ptr + umax_val < umax_ptr) {
2733 dst_reg->umin_value = 0;
2734 dst_reg->umax_value = U64_MAX;
2736 dst_reg->umin_value = umin_ptr + umin_val;
2737 dst_reg->umax_value = umax_ptr + umax_val;
2739 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
2740 dst_reg->off = ptr_reg->off;
2741 if (reg_is_pkt_pointer(ptr_reg)) {
2742 dst_reg->id = ++env->id_gen;
2743 /* something was added to pkt_ptr, set range to zero */
2748 if (dst_reg == off_reg) {
2749 /* scalar -= pointer. Creates an unknown scalar */
2750 verbose(env, "R%d tried to subtract pointer from scalar\n",
2754 /* We don't allow subtraction from FP, because (according to
2755 * test_verifier.c test "invalid fp arithmetic", JITs might not
2756 * be able to deal with it.
2758 if (ptr_reg->type == PTR_TO_STACK) {
2759 verbose(env, "R%d subtraction from stack pointer prohibited\n",
2763 if (known && (ptr_reg->off - smin_val ==
2764 (s64)(s32)(ptr_reg->off - smin_val))) {
2765 /* pointer -= K. Subtract it from fixed offset */
2766 dst_reg->smin_value = smin_ptr;
2767 dst_reg->smax_value = smax_ptr;
2768 dst_reg->umin_value = umin_ptr;
2769 dst_reg->umax_value = umax_ptr;
2770 dst_reg->var_off = ptr_reg->var_off;
2771 dst_reg->id = ptr_reg->id;
2772 dst_reg->off = ptr_reg->off - smin_val;
2773 dst_reg->range = ptr_reg->range;
2776 /* A new variable offset is created. If the subtrahend is known
2777 * nonnegative, then any reg->range we had before is still good.
2779 if (signed_sub_overflows(smin_ptr, smax_val) ||
2780 signed_sub_overflows(smax_ptr, smin_val)) {
2781 /* Overflow possible, we know nothing */
2782 dst_reg->smin_value = S64_MIN;
2783 dst_reg->smax_value = S64_MAX;
2785 dst_reg->smin_value = smin_ptr - smax_val;
2786 dst_reg->smax_value = smax_ptr - smin_val;
2788 if (umin_ptr < umax_val) {
2789 /* Overflow possible, we know nothing */
2790 dst_reg->umin_value = 0;
2791 dst_reg->umax_value = U64_MAX;
2793 /* Cannot overflow (as long as bounds are consistent) */
2794 dst_reg->umin_value = umin_ptr - umax_val;
2795 dst_reg->umax_value = umax_ptr - umin_val;
2797 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2798 dst_reg->off = ptr_reg->off;
2799 if (reg_is_pkt_pointer(ptr_reg)) {
2800 dst_reg->id = ++env->id_gen;
2801 /* something was added to pkt_ptr, set range to zero */
2809 /* bitwise ops on pointers are troublesome, prohibit. */
2810 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2811 dst, bpf_alu_string[opcode >> 4]);
2814 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2815 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2816 dst, bpf_alu_string[opcode >> 4]);
2820 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2823 __update_reg_bounds(dst_reg);
2824 __reg_deduce_bounds(dst_reg);
2825 __reg_bound_offset(dst_reg);
2829 /* WARNING: This function does calculations on 64-bit values, but the actual
2830 * execution may occur on 32-bit values. Therefore, things like bitshifts
2831 * need extra checks in the 32-bit case.
2833 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2834 struct bpf_insn *insn,
2835 struct bpf_reg_state *dst_reg,
2836 struct bpf_reg_state src_reg)
2838 struct bpf_reg_state *regs = cur_regs(env);
2839 u8 opcode = BPF_OP(insn->code);
2840 bool src_known, dst_known;
2841 s64 smin_val, smax_val;
2842 u64 umin_val, umax_val;
2843 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2845 if (insn_bitness == 32) {
2846 /* Relevant for 32-bit RSH: Information can propagate towards
2847 * LSB, so it isn't sufficient to only truncate the output to
2850 coerce_reg_to_size(dst_reg, 4);
2851 coerce_reg_to_size(&src_reg, 4);
2854 smin_val = src_reg.smin_value;
2855 smax_val = src_reg.smax_value;
2856 umin_val = src_reg.umin_value;
2857 umax_val = src_reg.umax_value;
2858 src_known = tnum_is_const(src_reg.var_off);
2859 dst_known = tnum_is_const(dst_reg->var_off);
2861 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2862 smin_val > smax_val || umin_val > umax_val) {
2863 /* Taint dst register if offset had invalid bounds derived from
2864 * e.g. dead branches.
2866 __mark_reg_unknown(dst_reg);
2871 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2872 __mark_reg_unknown(dst_reg);
2878 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2879 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2880 dst_reg->smin_value = S64_MIN;
2881 dst_reg->smax_value = S64_MAX;
2883 dst_reg->smin_value += smin_val;
2884 dst_reg->smax_value += smax_val;
2886 if (dst_reg->umin_value + umin_val < umin_val ||
2887 dst_reg->umax_value + umax_val < umax_val) {
2888 dst_reg->umin_value = 0;
2889 dst_reg->umax_value = U64_MAX;
2891 dst_reg->umin_value += umin_val;
2892 dst_reg->umax_value += umax_val;
2894 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2897 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2898 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2899 /* Overflow possible, we know nothing */
2900 dst_reg->smin_value = S64_MIN;
2901 dst_reg->smax_value = S64_MAX;
2903 dst_reg->smin_value -= smax_val;
2904 dst_reg->smax_value -= smin_val;
2906 if (dst_reg->umin_value < umax_val) {
2907 /* Overflow possible, we know nothing */
2908 dst_reg->umin_value = 0;
2909 dst_reg->umax_value = U64_MAX;
2911 /* Cannot overflow (as long as bounds are consistent) */
2912 dst_reg->umin_value -= umax_val;
2913 dst_reg->umax_value -= umin_val;
2915 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2918 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2919 if (smin_val < 0 || dst_reg->smin_value < 0) {
2920 /* Ain't nobody got time to multiply that sign */
2921 __mark_reg_unbounded(dst_reg);
2922 __update_reg_bounds(dst_reg);
2925 /* Both values are positive, so we can work with unsigned and
2926 * copy the result to signed (unless it exceeds S64_MAX).
2928 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2929 /* Potential overflow, we know nothing */
2930 __mark_reg_unbounded(dst_reg);
2931 /* (except what we can learn from the var_off) */
2932 __update_reg_bounds(dst_reg);
2935 dst_reg->umin_value *= umin_val;
2936 dst_reg->umax_value *= umax_val;
2937 if (dst_reg->umax_value > S64_MAX) {
2938 /* Overflow possible, we know nothing */
2939 dst_reg->smin_value = S64_MIN;
2940 dst_reg->smax_value = S64_MAX;
2942 dst_reg->smin_value = dst_reg->umin_value;
2943 dst_reg->smax_value = dst_reg->umax_value;
2947 if (src_known && dst_known) {
2948 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2949 src_reg.var_off.value);
2952 /* We get our minimum from the var_off, since that's inherently
2953 * bitwise. Our maximum is the minimum of the operands' maxima.
2955 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2956 dst_reg->umin_value = dst_reg->var_off.value;
2957 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2958 if (dst_reg->smin_value < 0 || smin_val < 0) {
2959 /* Lose signed bounds when ANDing negative numbers,
2960 * ain't nobody got time for that.
2962 dst_reg->smin_value = S64_MIN;
2963 dst_reg->smax_value = S64_MAX;
2965 /* ANDing two positives gives a positive, so safe to
2966 * cast result into s64.
2968 dst_reg->smin_value = dst_reg->umin_value;
2969 dst_reg->smax_value = dst_reg->umax_value;
2971 /* We may learn something more from the var_off */
2972 __update_reg_bounds(dst_reg);
2975 if (src_known && dst_known) {
2976 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2977 src_reg.var_off.value);
2980 /* We get our maximum from the var_off, and our minimum is the
2981 * maximum of the operands' minima
2983 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2984 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2985 dst_reg->umax_value = dst_reg->var_off.value |
2986 dst_reg->var_off.mask;
2987 if (dst_reg->smin_value < 0 || smin_val < 0) {
2988 /* Lose signed bounds when ORing negative numbers,
2989 * ain't nobody got time for that.
2991 dst_reg->smin_value = S64_MIN;
2992 dst_reg->smax_value = S64_MAX;
2994 /* ORing two positives gives a positive, so safe to
2995 * cast result into s64.
2997 dst_reg->smin_value = dst_reg->umin_value;
2998 dst_reg->smax_value = dst_reg->umax_value;
3000 /* We may learn something more from the var_off */
3001 __update_reg_bounds(dst_reg);
3004 if (umax_val >= insn_bitness) {
3005 /* Shifts greater than 31 or 63 are undefined.
3006 * This includes shifts by a negative number.
3008 mark_reg_unknown(env, regs, insn->dst_reg);
3011 /* We lose all sign bit information (except what we can pick
3014 dst_reg->smin_value = S64_MIN;
3015 dst_reg->smax_value = S64_MAX;
3016 /* If we might shift our top bit out, then we know nothing */
3017 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3018 dst_reg->umin_value = 0;
3019 dst_reg->umax_value = U64_MAX;
3021 dst_reg->umin_value <<= umin_val;
3022 dst_reg->umax_value <<= umax_val;
3024 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3025 /* We may learn something more from the var_off */
3026 __update_reg_bounds(dst_reg);
3029 if (umax_val >= insn_bitness) {
3030 /* Shifts greater than 31 or 63 are undefined.
3031 * This includes shifts by a negative number.
3033 mark_reg_unknown(env, regs, insn->dst_reg);
3036 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3037 * be negative, then either:
3038 * 1) src_reg might be zero, so the sign bit of the result is
3039 * unknown, so we lose our signed bounds
3040 * 2) it's known negative, thus the unsigned bounds capture the
3042 * 3) the signed bounds cross zero, so they tell us nothing
3044 * If the value in dst_reg is known nonnegative, then again the
3045 * unsigned bounts capture the signed bounds.
3046 * Thus, in all cases it suffices to blow away our signed bounds
3047 * and rely on inferring new ones from the unsigned bounds and
3048 * var_off of the result.
3050 dst_reg->smin_value = S64_MIN;
3051 dst_reg->smax_value = S64_MAX;
3052 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3053 dst_reg->umin_value >>= umax_val;
3054 dst_reg->umax_value >>= umin_val;
3055 /* We may learn something more from the var_off */
3056 __update_reg_bounds(dst_reg);
3059 if (umax_val >= insn_bitness) {
3060 /* Shifts greater than 31 or 63 are undefined.
3061 * This includes shifts by a negative number.
3063 mark_reg_unknown(env, regs, insn->dst_reg);
3067 /* Upon reaching here, src_known is true and
3068 * umax_val is equal to umin_val.
3070 dst_reg->smin_value >>= umin_val;
3071 dst_reg->smax_value >>= umin_val;
3072 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3074 /* blow away the dst_reg umin_value/umax_value and rely on
3075 * dst_reg var_off to refine the result.
3077 dst_reg->umin_value = 0;
3078 dst_reg->umax_value = U64_MAX;
3079 __update_reg_bounds(dst_reg);
3082 mark_reg_unknown(env, regs, insn->dst_reg);
3086 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3087 /* 32-bit ALU ops are (32,32)->32 */
3088 coerce_reg_to_size(dst_reg, 4);
3091 __reg_deduce_bounds(dst_reg);
3092 __reg_bound_offset(dst_reg);
3096 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3099 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3100 struct bpf_insn *insn)
3102 struct bpf_verifier_state *vstate = env->cur_state;
3103 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3104 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3105 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3106 u8 opcode = BPF_OP(insn->code);
3108 dst_reg = ®s[insn->dst_reg];
3110 if (dst_reg->type != SCALAR_VALUE)
3112 if (BPF_SRC(insn->code) == BPF_X) {
3113 src_reg = ®s[insn->src_reg];
3114 if (src_reg->type != SCALAR_VALUE) {
3115 if (dst_reg->type != SCALAR_VALUE) {
3116 /* Combining two pointers by any ALU op yields
3117 * an arbitrary scalar. Disallow all math except
3118 * pointer subtraction
3120 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3121 mark_reg_unknown(env, regs, insn->dst_reg);
3124 verbose(env, "R%d pointer %s pointer prohibited\n",
3126 bpf_alu_string[opcode >> 4]);
3129 /* scalar += pointer
3130 * This is legal, but we have to reverse our
3131 * src/dest handling in computing the range
3133 return adjust_ptr_min_max_vals(env, insn,
3136 } else if (ptr_reg) {
3137 /* pointer += scalar */
3138 return adjust_ptr_min_max_vals(env, insn,
3142 /* Pretend the src is a reg with a known value, since we only
3143 * need to be able to read from this state.
3145 off_reg.type = SCALAR_VALUE;
3146 __mark_reg_known(&off_reg, insn->imm);
3148 if (ptr_reg) /* pointer += K */
3149 return adjust_ptr_min_max_vals(env, insn,
3153 /* Got here implies adding two SCALAR_VALUEs */
3154 if (WARN_ON_ONCE(ptr_reg)) {
3155 print_verifier_state(env, state);
3156 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3159 if (WARN_ON(!src_reg)) {
3160 print_verifier_state(env, state);
3161 verbose(env, "verifier internal error: no src_reg\n");
3164 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3167 /* check validity of 32-bit and 64-bit arithmetic operations */
3168 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3170 struct bpf_reg_state *regs = cur_regs(env);
3171 u8 opcode = BPF_OP(insn->code);
3174 if (opcode == BPF_END || opcode == BPF_NEG) {
3175 if (opcode == BPF_NEG) {
3176 if (BPF_SRC(insn->code) != 0 ||
3177 insn->src_reg != BPF_REG_0 ||
3178 insn->off != 0 || insn->imm != 0) {
3179 verbose(env, "BPF_NEG uses reserved fields\n");
3183 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3184 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3185 BPF_CLASS(insn->code) == BPF_ALU64) {
3186 verbose(env, "BPF_END uses reserved fields\n");
3191 /* check src operand */
3192 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3196 if (is_pointer_value(env, insn->dst_reg)) {
3197 verbose(env, "R%d pointer arithmetic prohibited\n",
3202 /* check dest operand */
3203 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3207 } else if (opcode == BPF_MOV) {
3209 if (BPF_SRC(insn->code) == BPF_X) {
3210 if (insn->imm != 0 || insn->off != 0) {
3211 verbose(env, "BPF_MOV uses reserved fields\n");
3215 /* check src operand */
3216 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3220 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3221 verbose(env, "BPF_MOV uses reserved fields\n");
3226 /* check dest operand, mark as required later */
3227 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3231 if (BPF_SRC(insn->code) == BPF_X) {
3232 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3234 * copy register state to dest reg
3236 regs[insn->dst_reg] = regs[insn->src_reg];
3237 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3240 if (is_pointer_value(env, insn->src_reg)) {
3242 "R%d partial copy of pointer\n",
3246 mark_reg_unknown(env, regs, insn->dst_reg);
3247 coerce_reg_to_size(®s[insn->dst_reg], 4);
3251 * remember the value we stored into this reg
3253 /* clear any state __mark_reg_known doesn't set */
3254 mark_reg_unknown(env, regs, insn->dst_reg);
3255 regs[insn->dst_reg].type = SCALAR_VALUE;
3256 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3257 __mark_reg_known(regs + insn->dst_reg,
3260 __mark_reg_known(regs + insn->dst_reg,
3265 } else if (opcode > BPF_END) {
3266 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3269 } else { /* all other ALU ops: and, sub, xor, add, ... */
3271 if (BPF_SRC(insn->code) == BPF_X) {
3272 if (insn->imm != 0 || insn->off != 0) {
3273 verbose(env, "BPF_ALU uses reserved fields\n");
3276 /* check src1 operand */
3277 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3281 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3282 verbose(env, "BPF_ALU uses reserved fields\n");
3287 /* check src2 operand */
3288 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3292 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3293 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3294 verbose(env, "div by zero\n");
3298 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3299 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3303 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3304 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3305 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3307 if (insn->imm < 0 || insn->imm >= size) {
3308 verbose(env, "invalid shift %d\n", insn->imm);
3313 /* check dest operand */
3314 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3318 return adjust_reg_min_max_vals(env, insn);
3324 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3325 struct bpf_reg_state *dst_reg,
3326 enum bpf_reg_type type,
3327 bool range_right_open)
3329 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3330 struct bpf_reg_state *regs = state->regs, *reg;
3334 if (dst_reg->off < 0 ||
3335 (dst_reg->off == 0 && range_right_open))
3336 /* This doesn't give us any range */
3339 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3340 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3341 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3342 * than pkt_end, but that's because it's also less than pkt.
3346 new_range = dst_reg->off;
3347 if (range_right_open)
3350 /* Examples for register markings:
3352 * pkt_data in dst register:
3356 * if (r2 > pkt_end) goto <handle exception>
3361 * if (r2 < pkt_end) goto <access okay>
3362 * <handle exception>
3365 * r2 == dst_reg, pkt_end == src_reg
3366 * r2=pkt(id=n,off=8,r=0)
3367 * r3=pkt(id=n,off=0,r=0)
3369 * pkt_data in src register:
3373 * if (pkt_end >= r2) goto <access okay>
3374 * <handle exception>
3378 * if (pkt_end <= r2) goto <handle exception>
3382 * pkt_end == dst_reg, r2 == src_reg
3383 * r2=pkt(id=n,off=8,r=0)
3384 * r3=pkt(id=n,off=0,r=0)
3386 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3387 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3388 * and [r3, r3 + 8-1) respectively is safe to access depending on
3392 /* If our ids match, then we must have the same max_value. And we
3393 * don't care about the other reg's fixed offset, since if it's too big
3394 * the range won't allow anything.
3395 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3397 for (i = 0; i < MAX_BPF_REG; i++)
3398 if (regs[i].type == type && regs[i].id == dst_reg->id)
3399 /* keep the maximum range already checked */
3400 regs[i].range = max(regs[i].range, new_range);
3402 for (j = 0; j <= vstate->curframe; j++) {
3403 state = vstate->frame[j];
3404 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3405 if (state->stack[i].slot_type[0] != STACK_SPILL)
3407 reg = &state->stack[i].spilled_ptr;
3408 if (reg->type == type && reg->id == dst_reg->id)
3409 reg->range = max(reg->range, new_range);
3414 /* Adjusts the register min/max values in the case that the dst_reg is the
3415 * variable register that we are working on, and src_reg is a constant or we're
3416 * simply doing a BPF_K check.
3417 * In JEQ/JNE cases we also adjust the var_off values.
3419 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3420 struct bpf_reg_state *false_reg, u64 val,
3423 /* If the dst_reg is a pointer, we can't learn anything about its
3424 * variable offset from the compare (unless src_reg were a pointer into
3425 * the same object, but we don't bother with that.
3426 * Since false_reg and true_reg have the same type by construction, we
3427 * only need to check one of them for pointerness.
3429 if (__is_pointer_value(false, false_reg))
3434 /* If this is false then we know nothing Jon Snow, but if it is
3435 * true then we know for sure.
3437 __mark_reg_known(true_reg, val);
3440 /* If this is true we know nothing Jon Snow, but if it is false
3441 * we know the value for sure;
3443 __mark_reg_known(false_reg, val);
3446 false_reg->umax_value = min(false_reg->umax_value, val);
3447 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3450 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3451 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3454 false_reg->umin_value = max(false_reg->umin_value, val);
3455 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3458 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3459 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3462 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3463 true_reg->umin_value = max(true_reg->umin_value, val);
3466 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3467 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3470 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3471 true_reg->umax_value = min(true_reg->umax_value, val);
3474 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3475 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3481 __reg_deduce_bounds(false_reg);
3482 __reg_deduce_bounds(true_reg);
3483 /* We might have learned some bits from the bounds. */
3484 __reg_bound_offset(false_reg);
3485 __reg_bound_offset(true_reg);
3486 /* Intersecting with the old var_off might have improved our bounds
3487 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3488 * then new var_off is (0; 0x7f...fc) which improves our umax.
3490 __update_reg_bounds(false_reg);
3491 __update_reg_bounds(true_reg);
3494 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3497 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3498 struct bpf_reg_state *false_reg, u64 val,
3501 if (__is_pointer_value(false, false_reg))
3506 /* If this is false then we know nothing Jon Snow, but if it is
3507 * true then we know for sure.
3509 __mark_reg_known(true_reg, val);
3512 /* If this is true we know nothing Jon Snow, but if it is false
3513 * we know the value for sure;
3515 __mark_reg_known(false_reg, val);
3518 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3519 false_reg->umin_value = max(false_reg->umin_value, val);
3522 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3523 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3526 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3527 false_reg->umax_value = min(false_reg->umax_value, val);
3530 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3531 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3534 true_reg->umax_value = min(true_reg->umax_value, val);
3535 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3538 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3539 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3542 true_reg->umin_value = max(true_reg->umin_value, val);
3543 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3546 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3547 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3553 __reg_deduce_bounds(false_reg);
3554 __reg_deduce_bounds(true_reg);
3555 /* We might have learned some bits from the bounds. */
3556 __reg_bound_offset(false_reg);
3557 __reg_bound_offset(true_reg);
3558 /* Intersecting with the old var_off might have improved our bounds
3559 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3560 * then new var_off is (0; 0x7f...fc) which improves our umax.
3562 __update_reg_bounds(false_reg);
3563 __update_reg_bounds(true_reg);
3566 /* Regs are known to be equal, so intersect their min/max/var_off */
3567 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3568 struct bpf_reg_state *dst_reg)
3570 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3571 dst_reg->umin_value);
3572 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3573 dst_reg->umax_value);
3574 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3575 dst_reg->smin_value);
3576 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3577 dst_reg->smax_value);
3578 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3580 /* We might have learned new bounds from the var_off. */
3581 __update_reg_bounds(src_reg);
3582 __update_reg_bounds(dst_reg);
3583 /* We might have learned something about the sign bit. */
3584 __reg_deduce_bounds(src_reg);
3585 __reg_deduce_bounds(dst_reg);
3586 /* We might have learned some bits from the bounds. */
3587 __reg_bound_offset(src_reg);
3588 __reg_bound_offset(dst_reg);
3589 /* Intersecting with the old var_off might have improved our bounds
3590 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3591 * then new var_off is (0; 0x7f...fc) which improves our umax.
3593 __update_reg_bounds(src_reg);
3594 __update_reg_bounds(dst_reg);
3597 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3598 struct bpf_reg_state *true_dst,
3599 struct bpf_reg_state *false_src,
3600 struct bpf_reg_state *false_dst,
3605 __reg_combine_min_max(true_src, true_dst);
3608 __reg_combine_min_max(false_src, false_dst);
3613 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
3616 struct bpf_reg_state *reg = ®s[regno];
3618 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
3619 /* Old offset (both fixed and variable parts) should
3620 * have been known-zero, because we don't allow pointer
3621 * arithmetic on pointers that might be NULL.
3623 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3624 !tnum_equals_const(reg->var_off, 0) ||
3626 __mark_reg_known_zero(reg);
3630 reg->type = SCALAR_VALUE;
3631 } else if (reg->map_ptr->inner_map_meta) {
3632 reg->type = CONST_PTR_TO_MAP;
3633 reg->map_ptr = reg->map_ptr->inner_map_meta;
3635 reg->type = PTR_TO_MAP_VALUE;
3637 /* We don't need id from this point onwards anymore, thus we
3638 * should better reset it, so that state pruning has chances
3645 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3646 * be folded together at some point.
3648 static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno,
3651 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3652 struct bpf_reg_state *regs = state->regs;
3653 u32 id = regs[regno].id;
3656 for (i = 0; i < MAX_BPF_REG; i++)
3657 mark_map_reg(regs, i, id, is_null);
3659 for (j = 0; j <= vstate->curframe; j++) {
3660 state = vstate->frame[j];
3661 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3662 if (state->stack[i].slot_type[0] != STACK_SPILL)
3664 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
3669 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3670 struct bpf_reg_state *dst_reg,
3671 struct bpf_reg_state *src_reg,
3672 struct bpf_verifier_state *this_branch,
3673 struct bpf_verifier_state *other_branch)
3675 if (BPF_SRC(insn->code) != BPF_X)
3678 switch (BPF_OP(insn->code)) {
3680 if ((dst_reg->type == PTR_TO_PACKET &&
3681 src_reg->type == PTR_TO_PACKET_END) ||
3682 (dst_reg->type == PTR_TO_PACKET_META &&
3683 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3684 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3685 find_good_pkt_pointers(this_branch, dst_reg,
3686 dst_reg->type, false);
3687 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3688 src_reg->type == PTR_TO_PACKET) ||
3689 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3690 src_reg->type == PTR_TO_PACKET_META)) {
3691 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3692 find_good_pkt_pointers(other_branch, src_reg,
3693 src_reg->type, true);
3699 if ((dst_reg->type == PTR_TO_PACKET &&
3700 src_reg->type == PTR_TO_PACKET_END) ||
3701 (dst_reg->type == PTR_TO_PACKET_META &&
3702 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3703 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3704 find_good_pkt_pointers(other_branch, dst_reg,
3705 dst_reg->type, true);
3706 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3707 src_reg->type == PTR_TO_PACKET) ||
3708 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3709 src_reg->type == PTR_TO_PACKET_META)) {
3710 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3711 find_good_pkt_pointers(this_branch, src_reg,
3712 src_reg->type, false);
3718 if ((dst_reg->type == PTR_TO_PACKET &&
3719 src_reg->type == PTR_TO_PACKET_END) ||
3720 (dst_reg->type == PTR_TO_PACKET_META &&
3721 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3722 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3723 find_good_pkt_pointers(this_branch, dst_reg,
3724 dst_reg->type, true);
3725 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3726 src_reg->type == PTR_TO_PACKET) ||
3727 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3728 src_reg->type == PTR_TO_PACKET_META)) {
3729 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3730 find_good_pkt_pointers(other_branch, src_reg,
3731 src_reg->type, false);
3737 if ((dst_reg->type == PTR_TO_PACKET &&
3738 src_reg->type == PTR_TO_PACKET_END) ||
3739 (dst_reg->type == PTR_TO_PACKET_META &&
3740 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3741 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3742 find_good_pkt_pointers(other_branch, dst_reg,
3743 dst_reg->type, false);
3744 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3745 src_reg->type == PTR_TO_PACKET) ||
3746 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3747 src_reg->type == PTR_TO_PACKET_META)) {
3748 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3749 find_good_pkt_pointers(this_branch, src_reg,
3750 src_reg->type, true);
3762 static int check_cond_jmp_op(struct bpf_verifier_env *env,
3763 struct bpf_insn *insn, int *insn_idx)
3765 struct bpf_verifier_state *this_branch = env->cur_state;
3766 struct bpf_verifier_state *other_branch;
3767 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
3768 struct bpf_reg_state *dst_reg, *other_branch_regs;
3769 u8 opcode = BPF_OP(insn->code);
3772 if (opcode > BPF_JSLE) {
3773 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
3777 if (BPF_SRC(insn->code) == BPF_X) {
3778 if (insn->imm != 0) {
3779 verbose(env, "BPF_JMP uses reserved fields\n");
3783 /* check src1 operand */
3784 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3788 if (is_pointer_value(env, insn->src_reg)) {
3789 verbose(env, "R%d pointer comparison prohibited\n",
3794 if (insn->src_reg != BPF_REG_0) {
3795 verbose(env, "BPF_JMP uses reserved fields\n");
3800 /* check src2 operand */
3801 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3805 dst_reg = ®s[insn->dst_reg];
3807 /* detect if R == 0 where R was initialized to zero earlier */
3808 if (BPF_SRC(insn->code) == BPF_K &&
3809 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3810 dst_reg->type == SCALAR_VALUE &&
3811 tnum_is_const(dst_reg->var_off)) {
3812 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
3813 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
3814 /* if (imm == imm) goto pc+off;
3815 * only follow the goto, ignore fall-through
3817 *insn_idx += insn->off;
3820 /* if (imm != imm) goto pc+off;
3821 * only follow fall-through branch, since
3822 * that's where the program will go
3828 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3831 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
3833 /* detect if we are comparing against a constant value so we can adjust
3834 * our min/max values for our dst register.
3835 * this is only legit if both are scalars (or pointers to the same
3836 * object, I suppose, but we don't support that right now), because
3837 * otherwise the different base pointers mean the offsets aren't
3840 if (BPF_SRC(insn->code) == BPF_X) {
3841 if (dst_reg->type == SCALAR_VALUE &&
3842 regs[insn->src_reg].type == SCALAR_VALUE) {
3843 if (tnum_is_const(regs[insn->src_reg].var_off))
3844 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3845 dst_reg, regs[insn->src_reg].var_off.value,
3847 else if (tnum_is_const(dst_reg->var_off))
3848 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
3849 ®s[insn->src_reg],
3850 dst_reg->var_off.value, opcode);
3851 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3852 /* Comparing for equality, we can combine knowledge */
3853 reg_combine_min_max(&other_branch_regs[insn->src_reg],
3854 &other_branch_regs[insn->dst_reg],
3855 ®s[insn->src_reg],
3856 ®s[insn->dst_reg], opcode);
3858 } else if (dst_reg->type == SCALAR_VALUE) {
3859 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3860 dst_reg, insn->imm, opcode);
3863 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3864 if (BPF_SRC(insn->code) == BPF_K &&
3865 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3866 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3867 /* Mark all identical map registers in each branch as either
3868 * safe or unknown depending R == 0 or R != 0 conditional.
3870 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3871 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3872 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3873 this_branch, other_branch) &&
3874 is_pointer_value(env, insn->dst_reg)) {
3875 verbose(env, "R%d pointer comparison prohibited\n",
3880 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
3884 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3885 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3887 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3889 return (struct bpf_map *) (unsigned long) imm64;
3892 /* verify BPF_LD_IMM64 instruction */
3893 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3895 struct bpf_reg_state *regs = cur_regs(env);
3898 if (BPF_SIZE(insn->code) != BPF_DW) {
3899 verbose(env, "invalid BPF_LD_IMM insn\n");
3902 if (insn->off != 0) {
3903 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3907 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3911 if (insn->src_reg == 0) {
3912 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3914 regs[insn->dst_reg].type = SCALAR_VALUE;
3915 __mark_reg_known(®s[insn->dst_reg], imm);
3919 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3920 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3922 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3923 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3927 static bool may_access_skb(enum bpf_prog_type type)
3930 case BPF_PROG_TYPE_SOCKET_FILTER:
3931 case BPF_PROG_TYPE_SCHED_CLS:
3932 case BPF_PROG_TYPE_SCHED_ACT:
3939 /* verify safety of LD_ABS|LD_IND instructions:
3940 * - they can only appear in the programs where ctx == skb
3941 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3942 * preserve R6-R9, and store return value into R0
3945 * ctx == skb == R6 == CTX
3948 * SRC == any register
3949 * IMM == 32-bit immediate
3952 * R0 - 8/16/32-bit skb data converted to cpu endianness
3954 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3956 struct bpf_reg_state *regs = cur_regs(env);
3957 u8 mode = BPF_MODE(insn->code);
3960 if (!may_access_skb(env->prog->type)) {
3961 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3965 if (!env->ops->gen_ld_abs) {
3966 verbose(env, "bpf verifier is misconfigured\n");
3970 if (env->subprog_cnt > 1) {
3971 /* when program has LD_ABS insn JITs and interpreter assume
3972 * that r1 == ctx == skb which is not the case for callees
3973 * that can have arbitrary arguments. It's problematic
3974 * for main prog as well since JITs would need to analyze
3975 * all functions in order to make proper register save/restore
3976 * decisions in the main prog. Hence disallow LD_ABS with calls
3978 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
3982 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3983 BPF_SIZE(insn->code) == BPF_DW ||
3984 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3985 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3989 /* check whether implicit source operand (register R6) is readable */
3990 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3994 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3996 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4000 if (mode == BPF_IND) {
4001 /* check explicit source operand */
4002 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4007 /* reset caller saved regs to unreadable */
4008 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4009 mark_reg_not_init(env, regs, caller_saved[i]);
4010 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4013 /* mark destination R0 register as readable, since it contains
4014 * the value fetched from the packet.
4015 * Already marked as written above.
4017 mark_reg_unknown(env, regs, BPF_REG_0);
4021 static int check_return_code(struct bpf_verifier_env *env)
4023 struct bpf_reg_state *reg;
4024 struct tnum range = tnum_range(0, 1);
4026 switch (env->prog->type) {
4027 case BPF_PROG_TYPE_CGROUP_SKB:
4028 case BPF_PROG_TYPE_CGROUP_SOCK:
4029 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4030 case BPF_PROG_TYPE_SOCK_OPS:
4031 case BPF_PROG_TYPE_CGROUP_DEVICE:
4037 reg = cur_regs(env) + BPF_REG_0;
4038 if (reg->type != SCALAR_VALUE) {
4039 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4040 reg_type_str[reg->type]);
4044 if (!tnum_in(range, reg->var_off)) {
4045 verbose(env, "At program exit the register R0 ");
4046 if (!tnum_is_unknown(reg->var_off)) {
4049 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4050 verbose(env, "has value %s", tn_buf);
4052 verbose(env, "has unknown scalar value");
4054 verbose(env, " should have been 0 or 1\n");
4060 /* non-recursive DFS pseudo code
4061 * 1 procedure DFS-iterative(G,v):
4062 * 2 label v as discovered
4063 * 3 let S be a stack
4065 * 5 while S is not empty
4067 * 7 if t is what we're looking for:
4069 * 9 for all edges e in G.adjacentEdges(t) do
4070 * 10 if edge e is already labelled
4071 * 11 continue with the next edge
4072 * 12 w <- G.adjacentVertex(t,e)
4073 * 13 if vertex w is not discovered and not explored
4074 * 14 label e as tree-edge
4075 * 15 label w as discovered
4078 * 18 else if vertex w is discovered
4079 * 19 label e as back-edge
4081 * 21 // vertex w is explored
4082 * 22 label e as forward- or cross-edge
4083 * 23 label t as explored
4088 * 0x11 - discovered and fall-through edge labelled
4089 * 0x12 - discovered and fall-through and branch edges labelled
4100 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4102 static int *insn_stack; /* stack of insns to process */
4103 static int cur_stack; /* current stack index */
4104 static int *insn_state;
4106 /* t, w, e - match pseudo-code above:
4107 * t - index of current instruction
4108 * w - next instruction
4111 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4113 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4116 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4119 if (w < 0 || w >= env->prog->len) {
4120 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4125 /* mark branch target for state pruning */
4126 env->explored_states[w] = STATE_LIST_MARK;
4128 if (insn_state[w] == 0) {
4130 insn_state[t] = DISCOVERED | e;
4131 insn_state[w] = DISCOVERED;
4132 if (cur_stack >= env->prog->len)
4134 insn_stack[cur_stack++] = w;
4136 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4137 verbose(env, "back-edge from insn %d to %d\n", t, w);
4139 } else if (insn_state[w] == EXPLORED) {
4140 /* forward- or cross-edge */
4141 insn_state[t] = DISCOVERED | e;
4143 verbose(env, "insn state internal bug\n");
4149 /* non-recursive depth-first-search to detect loops in BPF program
4150 * loop == back-edge in directed graph
4152 static int check_cfg(struct bpf_verifier_env *env)
4154 struct bpf_insn *insns = env->prog->insnsi;
4155 int insn_cnt = env->prog->len;
4159 ret = check_subprogs(env);
4163 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4167 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4173 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4174 insn_stack[0] = 0; /* 0 is the first instruction */
4180 t = insn_stack[cur_stack - 1];
4182 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4183 u8 opcode = BPF_OP(insns[t].code);
4185 if (opcode == BPF_EXIT) {
4187 } else if (opcode == BPF_CALL) {
4188 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4193 if (t + 1 < insn_cnt)
4194 env->explored_states[t + 1] = STATE_LIST_MARK;
4195 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4196 env->explored_states[t] = STATE_LIST_MARK;
4197 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4203 } else if (opcode == BPF_JA) {
4204 if (BPF_SRC(insns[t].code) != BPF_K) {
4208 /* unconditional jump with single edge */
4209 ret = push_insn(t, t + insns[t].off + 1,
4215 /* tell verifier to check for equivalent states
4216 * after every call and jump
4218 if (t + 1 < insn_cnt)
4219 env->explored_states[t + 1] = STATE_LIST_MARK;
4221 /* conditional jump with two edges */
4222 env->explored_states[t] = STATE_LIST_MARK;
4223 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4229 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4236 /* all other non-branch instructions with single
4239 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4247 insn_state[t] = EXPLORED;
4248 if (cur_stack-- <= 0) {
4249 verbose(env, "pop stack internal bug\n");
4256 for (i = 0; i < insn_cnt; i++) {
4257 if (insn_state[i] != EXPLORED) {
4258 verbose(env, "unreachable insn %d\n", i);
4263 ret = 0; /* cfg looks good */
4271 /* check %cur's range satisfies %old's */
4272 static bool range_within(struct bpf_reg_state *old,
4273 struct bpf_reg_state *cur)
4275 return old->umin_value <= cur->umin_value &&
4276 old->umax_value >= cur->umax_value &&
4277 old->smin_value <= cur->smin_value &&
4278 old->smax_value >= cur->smax_value;
4281 /* Maximum number of register states that can exist at once */
4282 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4288 /* If in the old state two registers had the same id, then they need to have
4289 * the same id in the new state as well. But that id could be different from
4290 * the old state, so we need to track the mapping from old to new ids.
4291 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4292 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4293 * regs with a different old id could still have new id 9, we don't care about
4295 * So we look through our idmap to see if this old id has been seen before. If
4296 * so, we require the new id to match; otherwise, we add the id pair to the map.
4298 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4302 for (i = 0; i < ID_MAP_SIZE; i++) {
4303 if (!idmap[i].old) {
4304 /* Reached an empty slot; haven't seen this id before */
4305 idmap[i].old = old_id;
4306 idmap[i].cur = cur_id;
4309 if (idmap[i].old == old_id)
4310 return idmap[i].cur == cur_id;
4312 /* We ran out of idmap slots, which should be impossible */
4317 /* Returns true if (rold safe implies rcur safe) */
4318 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4319 struct idpair *idmap)
4323 if (!(rold->live & REG_LIVE_READ))
4324 /* explored state didn't use this */
4327 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
4329 if (rold->type == PTR_TO_STACK)
4330 /* two stack pointers are equal only if they're pointing to
4331 * the same stack frame, since fp-8 in foo != fp-8 in bar
4333 return equal && rold->frameno == rcur->frameno;
4338 if (rold->type == NOT_INIT)
4339 /* explored state can't have used this */
4341 if (rcur->type == NOT_INIT)
4343 switch (rold->type) {
4345 if (rcur->type == SCALAR_VALUE) {
4346 /* new val must satisfy old val knowledge */
4347 return range_within(rold, rcur) &&
4348 tnum_in(rold->var_off, rcur->var_off);
4350 /* We're trying to use a pointer in place of a scalar.
4351 * Even if the scalar was unbounded, this could lead to
4352 * pointer leaks because scalars are allowed to leak
4353 * while pointers are not. We could make this safe in
4354 * special cases if root is calling us, but it's
4355 * probably not worth the hassle.
4359 case PTR_TO_MAP_VALUE:
4360 /* If the new min/max/var_off satisfy the old ones and
4361 * everything else matches, we are OK.
4362 * We don't care about the 'id' value, because nothing
4363 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4365 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4366 range_within(rold, rcur) &&
4367 tnum_in(rold->var_off, rcur->var_off);
4368 case PTR_TO_MAP_VALUE_OR_NULL:
4369 /* a PTR_TO_MAP_VALUE could be safe to use as a
4370 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4371 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4372 * checked, doing so could have affected others with the same
4373 * id, and we can't check for that because we lost the id when
4374 * we converted to a PTR_TO_MAP_VALUE.
4376 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4378 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4380 /* Check our ids match any regs they're supposed to */
4381 return check_ids(rold->id, rcur->id, idmap);
4382 case PTR_TO_PACKET_META:
4384 if (rcur->type != rold->type)
4386 /* We must have at least as much range as the old ptr
4387 * did, so that any accesses which were safe before are
4388 * still safe. This is true even if old range < old off,
4389 * since someone could have accessed through (ptr - k), or
4390 * even done ptr -= k in a register, to get a safe access.
4392 if (rold->range > rcur->range)
4394 /* If the offsets don't match, we can't trust our alignment;
4395 * nor can we be sure that we won't fall out of range.
4397 if (rold->off != rcur->off)
4399 /* id relations must be preserved */
4400 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4402 /* new val must satisfy old val knowledge */
4403 return range_within(rold, rcur) &&
4404 tnum_in(rold->var_off, rcur->var_off);
4406 case CONST_PTR_TO_MAP:
4407 case PTR_TO_PACKET_END:
4408 case PTR_TO_FLOW_KEYS:
4409 /* Only valid matches are exact, which memcmp() above
4410 * would have accepted
4413 /* Don't know what's going on, just say it's not safe */
4417 /* Shouldn't get here; if we do, say it's not safe */
4422 static bool stacksafe(struct bpf_func_state *old,
4423 struct bpf_func_state *cur,
4424 struct idpair *idmap)
4428 /* if explored stack has more populated slots than current stack
4429 * such stacks are not equivalent
4431 if (old->allocated_stack > cur->allocated_stack)
4434 /* walk slots of the explored stack and ignore any additional
4435 * slots in the current stack, since explored(safe) state
4438 for (i = 0; i < old->allocated_stack; i++) {
4439 spi = i / BPF_REG_SIZE;
4441 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4442 /* explored state didn't use this */
4445 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4447 /* if old state was safe with misc data in the stack
4448 * it will be safe with zero-initialized stack.
4449 * The opposite is not true
4451 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4452 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4454 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4455 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4456 /* Ex: old explored (safe) state has STACK_SPILL in
4457 * this stack slot, but current has has STACK_MISC ->
4458 * this verifier states are not equivalent,
4459 * return false to continue verification of this path
4462 if (i % BPF_REG_SIZE)
4464 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4466 if (!regsafe(&old->stack[spi].spilled_ptr,
4467 &cur->stack[spi].spilled_ptr,
4469 /* when explored and current stack slot are both storing
4470 * spilled registers, check that stored pointers types
4471 * are the same as well.
4472 * Ex: explored safe path could have stored
4473 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4474 * but current path has stored:
4475 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4476 * such verifier states are not equivalent.
4477 * return false to continue verification of this path
4484 /* compare two verifier states
4486 * all states stored in state_list are known to be valid, since
4487 * verifier reached 'bpf_exit' instruction through them
4489 * this function is called when verifier exploring different branches of
4490 * execution popped from the state stack. If it sees an old state that has
4491 * more strict register state and more strict stack state then this execution
4492 * branch doesn't need to be explored further, since verifier already
4493 * concluded that more strict state leads to valid finish.
4495 * Therefore two states are equivalent if register state is more conservative
4496 * and explored stack state is more conservative than the current one.
4499 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4500 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4502 * In other words if current stack state (one being explored) has more
4503 * valid slots than old one that already passed validation, it means
4504 * the verifier can stop exploring and conclude that current state is valid too
4506 * Similarly with registers. If explored state has register type as invalid
4507 * whereas register type in current state is meaningful, it means that
4508 * the current state will reach 'bpf_exit' instruction safely
4510 static bool func_states_equal(struct bpf_func_state *old,
4511 struct bpf_func_state *cur)
4513 struct idpair *idmap;
4517 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4518 /* If we failed to allocate the idmap, just say it's not safe */
4522 for (i = 0; i < MAX_BPF_REG; i++) {
4523 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4527 if (!stacksafe(old, cur, idmap))
4535 static bool states_equal(struct bpf_verifier_env *env,
4536 struct bpf_verifier_state *old,
4537 struct bpf_verifier_state *cur)
4541 if (old->curframe != cur->curframe)
4544 /* for states to be equal callsites have to be the same
4545 * and all frame states need to be equivalent
4547 for (i = 0; i <= old->curframe; i++) {
4548 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4550 if (!func_states_equal(old->frame[i], cur->frame[i]))
4556 /* A write screens off any subsequent reads; but write marks come from the
4557 * straight-line code between a state and its parent. When we arrive at an
4558 * equivalent state (jump target or such) we didn't arrive by the straight-line
4559 * code, so read marks in the state must propagate to the parent regardless
4560 * of the state's write marks. That's what 'parent == state->parent' comparison
4561 * in mark_reg_read() is for.
4563 static int propagate_liveness(struct bpf_verifier_env *env,
4564 const struct bpf_verifier_state *vstate,
4565 struct bpf_verifier_state *vparent)
4567 int i, frame, err = 0;
4568 struct bpf_func_state *state, *parent;
4570 if (vparent->curframe != vstate->curframe) {
4571 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4572 vparent->curframe, vstate->curframe);
4575 /* Propagate read liveness of registers... */
4576 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4577 /* We don't need to worry about FP liveness because it's read-only */
4578 for (i = 0; i < BPF_REG_FP; i++) {
4579 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4581 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4582 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
4583 &vparent->frame[vstate->curframe]->regs[i]);
4589 /* ... and stack slots */
4590 for (frame = 0; frame <= vstate->curframe; frame++) {
4591 state = vstate->frame[frame];
4592 parent = vparent->frame[frame];
4593 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4594 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4595 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4597 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4598 mark_reg_read(env, &state->stack[i].spilled_ptr,
4599 &parent->stack[i].spilled_ptr);
4605 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4607 struct bpf_verifier_state_list *new_sl;
4608 struct bpf_verifier_state_list *sl;
4609 struct bpf_verifier_state *cur = env->cur_state, *new;
4612 sl = env->explored_states[insn_idx];
4614 /* this 'insn_idx' instruction wasn't marked, so we will not
4615 * be doing state search here
4619 while (sl != STATE_LIST_MARK) {
4620 if (states_equal(env, &sl->state, cur)) {
4621 /* reached equivalent register/stack state,
4623 * Registers read by the continuation are read by us.
4624 * If we have any write marks in env->cur_state, they
4625 * will prevent corresponding reads in the continuation
4626 * from reaching our parent (an explored_state). Our
4627 * own state will get the read marks recorded, but
4628 * they'll be immediately forgotten as we're pruning
4629 * this state and will pop a new one.
4631 err = propagate_liveness(env, &sl->state, cur);
4639 /* there were no equivalent states, remember current one.
4640 * technically the current state is not proven to be safe yet,
4641 * but it will either reach outer most bpf_exit (which means it's safe)
4642 * or it will be rejected. Since there are no loops, we won't be
4643 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4644 * again on the way to bpf_exit
4646 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
4650 /* add new state to the head of linked list */
4651 new = &new_sl->state;
4652 err = copy_verifier_state(new, cur);
4654 free_verifier_state(new, false);
4658 new_sl->next = env->explored_states[insn_idx];
4659 env->explored_states[insn_idx] = new_sl;
4660 /* connect new state to parentage chain */
4661 for (i = 0; i < BPF_REG_FP; i++)
4662 cur_regs(env)[i].parent = &new->frame[new->curframe]->regs[i];
4663 /* clear write marks in current state: the writes we did are not writes
4664 * our child did, so they don't screen off its reads from us.
4665 * (There are no read marks in current state, because reads always mark
4666 * their parent and current state never has children yet. Only
4667 * explored_states can get read marks.)
4669 for (i = 0; i < BPF_REG_FP; i++)
4670 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
4672 /* all stack frames are accessible from callee, clear them all */
4673 for (j = 0; j <= cur->curframe; j++) {
4674 struct bpf_func_state *frame = cur->frame[j];
4675 struct bpf_func_state *newframe = new->frame[j];
4677 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
4678 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
4679 frame->stack[i].spilled_ptr.parent =
4680 &newframe->stack[i].spilled_ptr;
4686 static int do_check(struct bpf_verifier_env *env)
4688 struct bpf_verifier_state *state;
4689 struct bpf_insn *insns = env->prog->insnsi;
4690 struct bpf_reg_state *regs;
4691 int insn_cnt = env->prog->len, i;
4692 int insn_idx, prev_insn_idx = 0;
4693 int insn_processed = 0;
4694 bool do_print_state = false;
4696 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
4699 state->curframe = 0;
4700 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
4701 if (!state->frame[0]) {
4705 env->cur_state = state;
4706 init_func_state(env, state->frame[0],
4707 BPF_MAIN_FUNC /* callsite */,
4709 0 /* subprogno, zero == main subprog */);
4712 struct bpf_insn *insn;
4716 if (insn_idx >= insn_cnt) {
4717 verbose(env, "invalid insn idx %d insn_cnt %d\n",
4718 insn_idx, insn_cnt);
4722 insn = &insns[insn_idx];
4723 class = BPF_CLASS(insn->code);
4725 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
4727 "BPF program is too large. Processed %d insn\n",
4732 err = is_state_visited(env, insn_idx);
4736 /* found equivalent state, can prune the search */
4737 if (env->log.level) {
4739 verbose(env, "\nfrom %d to %d: safe\n",
4740 prev_insn_idx, insn_idx);
4742 verbose(env, "%d: safe\n", insn_idx);
4744 goto process_bpf_exit;
4750 if (env->log.level > 1 || (env->log.level && do_print_state)) {
4751 if (env->log.level > 1)
4752 verbose(env, "%d:", insn_idx);
4754 verbose(env, "\nfrom %d to %d:",
4755 prev_insn_idx, insn_idx);
4756 print_verifier_state(env, state->frame[state->curframe]);
4757 do_print_state = false;
4760 if (env->log.level) {
4761 const struct bpf_insn_cbs cbs = {
4762 .cb_print = verbose,
4763 .private_data = env,
4766 verbose(env, "%d: ", insn_idx);
4767 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4770 if (bpf_prog_is_dev_bound(env->prog->aux)) {
4771 err = bpf_prog_offload_verify_insn(env, insn_idx,
4777 regs = cur_regs(env);
4778 env->insn_aux_data[insn_idx].seen = true;
4779 if (class == BPF_ALU || class == BPF_ALU64) {
4780 err = check_alu_op(env, insn);
4784 } else if (class == BPF_LDX) {
4785 enum bpf_reg_type *prev_src_type, src_reg_type;
4787 /* check for reserved fields is already done */
4789 /* check src operand */
4790 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4794 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4798 src_reg_type = regs[insn->src_reg].type;
4800 /* check that memory (src_reg + off) is readable,
4801 * the state of dst_reg will be updated by this func
4803 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
4804 BPF_SIZE(insn->code), BPF_READ,
4805 insn->dst_reg, false);
4809 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
4811 if (*prev_src_type == NOT_INIT) {
4813 * dst_reg = *(u32 *)(src_reg + off)
4814 * save type to validate intersecting paths
4816 *prev_src_type = src_reg_type;
4818 } else if (src_reg_type != *prev_src_type &&
4819 (src_reg_type == PTR_TO_CTX ||
4820 *prev_src_type == PTR_TO_CTX)) {
4821 /* ABuser program is trying to use the same insn
4822 * dst_reg = *(u32*) (src_reg + off)
4823 * with different pointer types:
4824 * src_reg == ctx in one branch and
4825 * src_reg == stack|map in some other branch.
4828 verbose(env, "same insn cannot be used with different pointers\n");
4832 } else if (class == BPF_STX) {
4833 enum bpf_reg_type *prev_dst_type, dst_reg_type;
4835 if (BPF_MODE(insn->code) == BPF_XADD) {
4836 err = check_xadd(env, insn_idx, insn);
4843 /* check src1 operand */
4844 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4847 /* check src2 operand */
4848 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4852 dst_reg_type = regs[insn->dst_reg].type;
4854 /* check that memory (dst_reg + off) is writeable */
4855 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4856 BPF_SIZE(insn->code), BPF_WRITE,
4857 insn->src_reg, false);
4861 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
4863 if (*prev_dst_type == NOT_INIT) {
4864 *prev_dst_type = dst_reg_type;
4865 } else if (dst_reg_type != *prev_dst_type &&
4866 (dst_reg_type == PTR_TO_CTX ||
4867 *prev_dst_type == PTR_TO_CTX)) {
4868 verbose(env, "same insn cannot be used with different pointers\n");
4872 } else if (class == BPF_ST) {
4873 if (BPF_MODE(insn->code) != BPF_MEM ||
4874 insn->src_reg != BPF_REG_0) {
4875 verbose(env, "BPF_ST uses reserved fields\n");
4878 /* check src operand */
4879 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4883 if (is_ctx_reg(env, insn->dst_reg)) {
4884 verbose(env, "BPF_ST stores into R%d context is not allowed\n",
4889 /* check that memory (dst_reg + off) is writeable */
4890 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4891 BPF_SIZE(insn->code), BPF_WRITE,
4896 } else if (class == BPF_JMP) {
4897 u8 opcode = BPF_OP(insn->code);
4899 if (opcode == BPF_CALL) {
4900 if (BPF_SRC(insn->code) != BPF_K ||
4902 (insn->src_reg != BPF_REG_0 &&
4903 insn->src_reg != BPF_PSEUDO_CALL) ||
4904 insn->dst_reg != BPF_REG_0) {
4905 verbose(env, "BPF_CALL uses reserved fields\n");
4909 if (insn->src_reg == BPF_PSEUDO_CALL)
4910 err = check_func_call(env, insn, &insn_idx);
4912 err = check_helper_call(env, insn->imm, insn_idx);
4916 } else if (opcode == BPF_JA) {
4917 if (BPF_SRC(insn->code) != BPF_K ||
4919 insn->src_reg != BPF_REG_0 ||
4920 insn->dst_reg != BPF_REG_0) {
4921 verbose(env, "BPF_JA uses reserved fields\n");
4925 insn_idx += insn->off + 1;
4928 } else if (opcode == BPF_EXIT) {
4929 if (BPF_SRC(insn->code) != BPF_K ||
4931 insn->src_reg != BPF_REG_0 ||
4932 insn->dst_reg != BPF_REG_0) {
4933 verbose(env, "BPF_EXIT uses reserved fields\n");
4937 if (state->curframe) {
4938 /* exit from nested function */
4939 prev_insn_idx = insn_idx;
4940 err = prepare_func_exit(env, &insn_idx);
4943 do_print_state = true;
4947 /* eBPF calling convetion is such that R0 is used
4948 * to return the value from eBPF program.
4949 * Make sure that it's readable at this time
4950 * of bpf_exit, which means that program wrote
4951 * something into it earlier
4953 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4957 if (is_pointer_value(env, BPF_REG_0)) {
4958 verbose(env, "R0 leaks addr as return value\n");
4962 err = check_return_code(env);
4966 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4972 do_print_state = true;
4976 err = check_cond_jmp_op(env, insn, &insn_idx);
4980 } else if (class == BPF_LD) {
4981 u8 mode = BPF_MODE(insn->code);
4983 if (mode == BPF_ABS || mode == BPF_IND) {
4984 err = check_ld_abs(env, insn);
4988 } else if (mode == BPF_IMM) {
4989 err = check_ld_imm(env, insn);
4994 env->insn_aux_data[insn_idx].seen = true;
4996 verbose(env, "invalid BPF_LD mode\n");
5000 verbose(env, "unknown insn class %d\n", class);
5007 verbose(env, "processed %d insns (limit %d), stack depth ",
5008 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5009 for (i = 0; i < env->subprog_cnt; i++) {
5010 u32 depth = env->subprog_info[i].stack_depth;
5012 verbose(env, "%d", depth);
5013 if (i + 1 < env->subprog_cnt)
5017 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
5021 static int check_map_prealloc(struct bpf_map *map)
5023 return (map->map_type != BPF_MAP_TYPE_HASH &&
5024 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
5025 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
5026 !(map->map_flags & BPF_F_NO_PREALLOC);
5029 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
5030 struct bpf_map *map,
5031 struct bpf_prog *prog)
5034 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5035 * preallocated hash maps, since doing memory allocation
5036 * in overflow_handler can crash depending on where nmi got
5039 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
5040 if (!check_map_prealloc(map)) {
5041 verbose(env, "perf_event programs can only use preallocated hash map\n");
5044 if (map->inner_map_meta &&
5045 !check_map_prealloc(map->inner_map_meta)) {
5046 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
5051 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
5052 !bpf_offload_prog_map_match(prog, map)) {
5053 verbose(env, "offload device mismatch between prog and map\n");
5060 /* look for pseudo eBPF instructions that access map FDs and
5061 * replace them with actual map pointers
5063 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
5065 struct bpf_insn *insn = env->prog->insnsi;
5066 int insn_cnt = env->prog->len;
5069 err = bpf_prog_calc_tag(env->prog);
5073 for (i = 0; i < insn_cnt; i++, insn++) {
5074 if (BPF_CLASS(insn->code) == BPF_LDX &&
5075 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
5076 verbose(env, "BPF_LDX uses reserved fields\n");
5080 if (BPF_CLASS(insn->code) == BPF_STX &&
5081 ((BPF_MODE(insn->code) != BPF_MEM &&
5082 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
5083 verbose(env, "BPF_STX uses reserved fields\n");
5087 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5088 struct bpf_map *map;
5091 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5092 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5094 verbose(env, "invalid bpf_ld_imm64 insn\n");
5098 if (insn->src_reg == 0)
5099 /* valid generic load 64-bit imm */
5102 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5104 "unrecognized bpf_ld_imm64 insn\n");
5108 f = fdget(insn->imm);
5109 map = __bpf_map_get(f);
5111 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5113 return PTR_ERR(map);
5116 err = check_map_prog_compatibility(env, map, env->prog);
5122 /* store map pointer inside BPF_LD_IMM64 instruction */
5123 insn[0].imm = (u32) (unsigned long) map;
5124 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5126 /* check whether we recorded this map already */
5127 for (j = 0; j < env->used_map_cnt; j++)
5128 if (env->used_maps[j] == map) {
5133 if (env->used_map_cnt >= MAX_USED_MAPS) {
5138 /* hold the map. If the program is rejected by verifier,
5139 * the map will be released by release_maps() or it
5140 * will be used by the valid program until it's unloaded
5141 * and all maps are released in free_used_maps()
5143 map = bpf_map_inc(map, false);
5146 return PTR_ERR(map);
5148 env->used_maps[env->used_map_cnt++] = map;
5150 if (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE &&
5151 bpf_cgroup_storage_assign(env->prog, map)) {
5153 "only one cgroup storage is allowed\n");
5165 /* Basic sanity check before we invest more work here. */
5166 if (!bpf_opcode_in_insntable(insn->code)) {
5167 verbose(env, "unknown opcode %02x\n", insn->code);
5172 /* now all pseudo BPF_LD_IMM64 instructions load valid
5173 * 'struct bpf_map *' into a register instead of user map_fd.
5174 * These pointers will be used later by verifier to validate map access.
5179 /* drop refcnt of maps used by the rejected program */
5180 static void release_maps(struct bpf_verifier_env *env)
5184 if (env->prog->aux->cgroup_storage)
5185 bpf_cgroup_storage_release(env->prog,
5186 env->prog->aux->cgroup_storage);
5188 for (i = 0; i < env->used_map_cnt; i++)
5189 bpf_map_put(env->used_maps[i]);
5192 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5193 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5195 struct bpf_insn *insn = env->prog->insnsi;
5196 int insn_cnt = env->prog->len;
5199 for (i = 0; i < insn_cnt; i++, insn++)
5200 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5204 /* single env->prog->insni[off] instruction was replaced with the range
5205 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5206 * [0, off) and [off, end) to new locations, so the patched range stays zero
5208 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5211 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5216 new_data = vzalloc(array_size(prog_len,
5217 sizeof(struct bpf_insn_aux_data)));
5220 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5221 memcpy(new_data + off + cnt - 1, old_data + off,
5222 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5223 for (i = off; i < off + cnt - 1; i++)
5224 new_data[i].seen = true;
5225 env->insn_aux_data = new_data;
5230 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5236 /* NOTE: fake 'exit' subprog should be updated as well. */
5237 for (i = 0; i <= env->subprog_cnt; i++) {
5238 if (env->subprog_info[i].start < off)
5240 env->subprog_info[i].start += len - 1;
5244 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5245 const struct bpf_insn *patch, u32 len)
5247 struct bpf_prog *new_prog;
5249 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5252 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5254 adjust_subprog_starts(env, off, len);
5258 /* The verifier does more data flow analysis than llvm and will not
5259 * explore branches that are dead at run time. Malicious programs can
5260 * have dead code too. Therefore replace all dead at-run-time code
5263 * Just nops are not optimal, e.g. if they would sit at the end of the
5264 * program and through another bug we would manage to jump there, then
5265 * we'd execute beyond program memory otherwise. Returning exception
5266 * code also wouldn't work since we can have subprogs where the dead
5267 * code could be located.
5269 static void sanitize_dead_code(struct bpf_verifier_env *env)
5271 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5272 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5273 struct bpf_insn *insn = env->prog->insnsi;
5274 const int insn_cnt = env->prog->len;
5277 for (i = 0; i < insn_cnt; i++) {
5278 if (aux_data[i].seen)
5280 memcpy(insn + i, &trap, sizeof(trap));
5284 /* convert load instructions that access fields of 'struct __sk_buff'
5285 * into sequence of instructions that access fields of 'struct sk_buff'
5287 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5289 const struct bpf_verifier_ops *ops = env->ops;
5290 int i, cnt, size, ctx_field_size, delta = 0;
5291 const int insn_cnt = env->prog->len;
5292 struct bpf_insn insn_buf[16], *insn;
5293 struct bpf_prog *new_prog;
5294 enum bpf_access_type type;
5295 bool is_narrower_load;
5298 if (ops->gen_prologue) {
5299 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5301 if (cnt >= ARRAY_SIZE(insn_buf)) {
5302 verbose(env, "bpf verifier is misconfigured\n");
5305 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5309 env->prog = new_prog;
5314 if (!ops->convert_ctx_access || bpf_prog_is_dev_bound(env->prog->aux))
5317 insn = env->prog->insnsi + delta;
5319 for (i = 0; i < insn_cnt; i++, insn++) {
5320 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5321 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5322 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5323 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5325 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5326 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5327 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5328 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5333 if (type == BPF_WRITE &&
5334 env->insn_aux_data[i + delta].sanitize_stack_off) {
5335 struct bpf_insn patch[] = {
5336 /* Sanitize suspicious stack slot with zero.
5337 * There are no memory dependencies for this store,
5338 * since it's only using frame pointer and immediate
5341 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5342 env->insn_aux_data[i + delta].sanitize_stack_off,
5344 /* the original STX instruction will immediately
5345 * overwrite the same stack slot with appropriate value
5350 cnt = ARRAY_SIZE(patch);
5351 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5356 env->prog = new_prog;
5357 insn = new_prog->insnsi + i + delta;
5361 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
5364 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5365 size = BPF_LDST_BYTES(insn);
5367 /* If the read access is a narrower load of the field,
5368 * convert to a 4/8-byte load, to minimum program type specific
5369 * convert_ctx_access changes. If conversion is successful,
5370 * we will apply proper mask to the result.
5372 is_narrower_load = size < ctx_field_size;
5373 if (is_narrower_load) {
5374 u32 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
5375 u32 off = insn->off;
5378 if (type == BPF_WRITE) {
5379 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5384 if (ctx_field_size == 4)
5386 else if (ctx_field_size == 8)
5389 insn->off = off & ~(size_default - 1);
5390 insn->code = BPF_LDX | BPF_MEM | size_code;
5394 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
5396 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5397 (ctx_field_size && !target_size)) {
5398 verbose(env, "bpf verifier is misconfigured\n");
5402 if (is_narrower_load && size < target_size) {
5403 if (ctx_field_size <= 4)
5404 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5405 (1 << size * 8) - 1);
5407 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5408 (1 << size * 8) - 1);
5411 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5417 /* keep walking new program and skip insns we just inserted */
5418 env->prog = new_prog;
5419 insn = new_prog->insnsi + i + delta;
5425 static int jit_subprogs(struct bpf_verifier_env *env)
5427 struct bpf_prog *prog = env->prog, **func, *tmp;
5428 int i, j, subprog_start, subprog_end = 0, len, subprog;
5429 struct bpf_insn *insn;
5433 if (env->subprog_cnt <= 1)
5436 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5437 if (insn->code != (BPF_JMP | BPF_CALL) ||
5438 insn->src_reg != BPF_PSEUDO_CALL)
5440 /* Upon error here we cannot fall back to interpreter but
5441 * need a hard reject of the program. Thus -EFAULT is
5442 * propagated in any case.
5444 subprog = find_subprog(env, i + insn->imm + 1);
5446 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5450 /* temporarily remember subprog id inside insn instead of
5451 * aux_data, since next loop will split up all insns into funcs
5453 insn->off = subprog;
5454 /* remember original imm in case JIT fails and fallback
5455 * to interpreter will be needed
5457 env->insn_aux_data[i].call_imm = insn->imm;
5458 /* point imm to __bpf_call_base+1 from JITs point of view */
5462 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
5466 for (i = 0; i < env->subprog_cnt; i++) {
5467 subprog_start = subprog_end;
5468 subprog_end = env->subprog_info[i + 1].start;
5470 len = subprog_end - subprog_start;
5471 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5474 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5475 len * sizeof(struct bpf_insn));
5476 func[i]->type = prog->type;
5478 if (bpf_prog_calc_tag(func[i]))
5480 func[i]->is_func = 1;
5481 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5482 * Long term would need debug info to populate names
5484 func[i]->aux->name[0] = 'F';
5485 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
5486 func[i]->jit_requested = 1;
5487 func[i] = bpf_int_jit_compile(func[i]);
5488 if (!func[i]->jited) {
5494 /* at this point all bpf functions were successfully JITed
5495 * now populate all bpf_calls with correct addresses and
5496 * run last pass of JIT
5498 for (i = 0; i < env->subprog_cnt; i++) {
5499 insn = func[i]->insnsi;
5500 for (j = 0; j < func[i]->len; j++, insn++) {
5501 if (insn->code != (BPF_JMP | BPF_CALL) ||
5502 insn->src_reg != BPF_PSEUDO_CALL)
5504 subprog = insn->off;
5505 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5506 func[subprog]->bpf_func -
5510 /* we use the aux data to keep a list of the start addresses
5511 * of the JITed images for each function in the program
5513 * for some architectures, such as powerpc64, the imm field
5514 * might not be large enough to hold the offset of the start
5515 * address of the callee's JITed image from __bpf_call_base
5517 * in such cases, we can lookup the start address of a callee
5518 * by using its subprog id, available from the off field of
5519 * the call instruction, as an index for this list
5521 func[i]->aux->func = func;
5522 func[i]->aux->func_cnt = env->subprog_cnt;
5524 for (i = 0; i < env->subprog_cnt; i++) {
5525 old_bpf_func = func[i]->bpf_func;
5526 tmp = bpf_int_jit_compile(func[i]);
5527 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5528 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5535 /* finally lock prog and jit images for all functions and
5538 for (i = 0; i < env->subprog_cnt; i++) {
5539 bpf_prog_lock_ro(func[i]);
5540 bpf_prog_kallsyms_add(func[i]);
5543 /* Last step: make now unused interpreter insns from main
5544 * prog consistent for later dump requests, so they can
5545 * later look the same as if they were interpreted only.
5547 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5548 if (insn->code != (BPF_JMP | BPF_CALL) ||
5549 insn->src_reg != BPF_PSEUDO_CALL)
5551 insn->off = env->insn_aux_data[i].call_imm;
5552 subprog = find_subprog(env, i + insn->off + 1);
5553 insn->imm = subprog;
5557 prog->bpf_func = func[0]->bpf_func;
5558 prog->aux->func = func;
5559 prog->aux->func_cnt = env->subprog_cnt;
5562 for (i = 0; i < env->subprog_cnt; i++)
5564 bpf_jit_free(func[i]);
5567 /* cleanup main prog to be interpreted */
5568 prog->jit_requested = 0;
5569 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5570 if (insn->code != (BPF_JMP | BPF_CALL) ||
5571 insn->src_reg != BPF_PSEUDO_CALL)
5574 insn->imm = env->insn_aux_data[i].call_imm;
5579 static int fixup_call_args(struct bpf_verifier_env *env)
5581 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5582 struct bpf_prog *prog = env->prog;
5583 struct bpf_insn *insn = prog->insnsi;
5589 if (env->prog->jit_requested) {
5590 err = jit_subprogs(env);
5596 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5597 for (i = 0; i < prog->len; i++, insn++) {
5598 if (insn->code != (BPF_JMP | BPF_CALL) ||
5599 insn->src_reg != BPF_PSEUDO_CALL)
5601 depth = get_callee_stack_depth(env, insn, i);
5604 bpf_patch_call_args(insn, depth);
5611 /* fixup insn->imm field of bpf_call instructions
5612 * and inline eligible helpers as explicit sequence of BPF instructions
5614 * this function is called after eBPF program passed verification
5616 static int fixup_bpf_calls(struct bpf_verifier_env *env)
5618 struct bpf_prog *prog = env->prog;
5619 struct bpf_insn *insn = prog->insnsi;
5620 const struct bpf_func_proto *fn;
5621 const int insn_cnt = prog->len;
5622 const struct bpf_map_ops *ops;
5623 struct bpf_insn_aux_data *aux;
5624 struct bpf_insn insn_buf[16];
5625 struct bpf_prog *new_prog;
5626 struct bpf_map *map_ptr;
5627 int i, cnt, delta = 0;
5629 for (i = 0; i < insn_cnt; i++, insn++) {
5630 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
5631 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5632 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
5633 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5634 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
5635 struct bpf_insn mask_and_div[] = {
5636 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5638 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
5639 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
5640 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
5643 struct bpf_insn mask_and_mod[] = {
5644 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5645 /* Rx mod 0 -> Rx */
5646 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
5649 struct bpf_insn *patchlet;
5651 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5652 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5653 patchlet = mask_and_div + (is64 ? 1 : 0);
5654 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
5656 patchlet = mask_and_mod + (is64 ? 1 : 0);
5657 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
5660 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
5665 env->prog = prog = new_prog;
5666 insn = new_prog->insnsi + i + delta;
5670 if (BPF_CLASS(insn->code) == BPF_LD &&
5671 (BPF_MODE(insn->code) == BPF_ABS ||
5672 BPF_MODE(insn->code) == BPF_IND)) {
5673 cnt = env->ops->gen_ld_abs(insn, insn_buf);
5674 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5675 verbose(env, "bpf verifier is misconfigured\n");
5679 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5684 env->prog = prog = new_prog;
5685 insn = new_prog->insnsi + i + delta;
5689 if (insn->code != (BPF_JMP | BPF_CALL))
5691 if (insn->src_reg == BPF_PSEUDO_CALL)
5694 if (insn->imm == BPF_FUNC_get_route_realm)
5695 prog->dst_needed = 1;
5696 if (insn->imm == BPF_FUNC_get_prandom_u32)
5697 bpf_user_rnd_init_once();
5698 if (insn->imm == BPF_FUNC_override_return)
5699 prog->kprobe_override = 1;
5700 if (insn->imm == BPF_FUNC_tail_call) {
5701 /* If we tail call into other programs, we
5702 * cannot make any assumptions since they can
5703 * be replaced dynamically during runtime in
5704 * the program array.
5706 prog->cb_access = 1;
5707 env->prog->aux->stack_depth = MAX_BPF_STACK;
5709 /* mark bpf_tail_call as different opcode to avoid
5710 * conditional branch in the interpeter for every normal
5711 * call and to prevent accidental JITing by JIT compiler
5712 * that doesn't support bpf_tail_call yet
5715 insn->code = BPF_JMP | BPF_TAIL_CALL;
5717 aux = &env->insn_aux_data[i + delta];
5718 if (!bpf_map_ptr_unpriv(aux))
5721 /* instead of changing every JIT dealing with tail_call
5722 * emit two extra insns:
5723 * if (index >= max_entries) goto out;
5724 * index &= array->index_mask;
5725 * to avoid out-of-bounds cpu speculation
5727 if (bpf_map_ptr_poisoned(aux)) {
5728 verbose(env, "tail_call abusing map_ptr\n");
5732 map_ptr = BPF_MAP_PTR(aux->map_state);
5733 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
5734 map_ptr->max_entries, 2);
5735 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
5736 container_of(map_ptr,
5739 insn_buf[2] = *insn;
5741 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5746 env->prog = prog = new_prog;
5747 insn = new_prog->insnsi + i + delta;
5751 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5752 * and other inlining handlers are currently limited to 64 bit
5755 if (prog->jit_requested && BITS_PER_LONG == 64 &&
5756 (insn->imm == BPF_FUNC_map_lookup_elem ||
5757 insn->imm == BPF_FUNC_map_update_elem ||
5758 insn->imm == BPF_FUNC_map_delete_elem)) {
5759 aux = &env->insn_aux_data[i + delta];
5760 if (bpf_map_ptr_poisoned(aux))
5761 goto patch_call_imm;
5763 map_ptr = BPF_MAP_PTR(aux->map_state);
5765 if (insn->imm == BPF_FUNC_map_lookup_elem &&
5766 ops->map_gen_lookup) {
5767 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
5768 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5769 verbose(env, "bpf verifier is misconfigured\n");
5773 new_prog = bpf_patch_insn_data(env, i + delta,
5779 env->prog = prog = new_prog;
5780 insn = new_prog->insnsi + i + delta;
5784 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
5785 (void *(*)(struct bpf_map *map, void *key))NULL));
5786 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
5787 (int (*)(struct bpf_map *map, void *key))NULL));
5788 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
5789 (int (*)(struct bpf_map *map, void *key, void *value,
5791 switch (insn->imm) {
5792 case BPF_FUNC_map_lookup_elem:
5793 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
5796 case BPF_FUNC_map_update_elem:
5797 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
5800 case BPF_FUNC_map_delete_elem:
5801 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
5806 goto patch_call_imm;
5810 fn = env->ops->get_func_proto(insn->imm, env->prog);
5811 /* all functions that have prototype and verifier allowed
5812 * programs to call them, must be real in-kernel functions
5816 "kernel subsystem misconfigured func %s#%d\n",
5817 func_id_name(insn->imm), insn->imm);
5820 insn->imm = fn->func - __bpf_call_base;
5826 static void free_states(struct bpf_verifier_env *env)
5828 struct bpf_verifier_state_list *sl, *sln;
5831 if (!env->explored_states)
5834 for (i = 0; i < env->prog->len; i++) {
5835 sl = env->explored_states[i];
5838 while (sl != STATE_LIST_MARK) {
5840 free_verifier_state(&sl->state, false);
5846 kfree(env->explored_states);
5849 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
5851 struct bpf_verifier_env *env;
5852 struct bpf_verifier_log *log;
5855 /* no program is valid */
5856 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
5859 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5860 * allocate/free it every time bpf_check() is called
5862 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
5867 env->insn_aux_data =
5868 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
5871 if (!env->insn_aux_data)
5874 env->ops = bpf_verifier_ops[env->prog->type];
5876 /* grab the mutex to protect few globals used by verifier */
5877 mutex_lock(&bpf_verifier_lock);
5879 if (attr->log_level || attr->log_buf || attr->log_size) {
5880 /* user requested verbose verifier output
5881 * and supplied buffer to store the verification trace
5883 log->level = attr->log_level;
5884 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
5885 log->len_total = attr->log_size;
5888 /* log attributes have to be sane */
5889 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
5890 !log->level || !log->ubuf)
5894 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
5895 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
5896 env->strict_alignment = true;
5898 ret = replace_map_fd_with_map_ptr(env);
5900 goto skip_full_check;
5902 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5903 ret = bpf_prog_offload_verifier_prep(env);
5905 goto skip_full_check;
5908 env->explored_states = kcalloc(env->prog->len,
5909 sizeof(struct bpf_verifier_state_list *),
5912 if (!env->explored_states)
5913 goto skip_full_check;
5915 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
5917 ret = check_cfg(env);
5919 goto skip_full_check;
5921 ret = do_check(env);
5922 if (env->cur_state) {
5923 free_verifier_state(env->cur_state, true);
5924 env->cur_state = NULL;
5928 while (!pop_stack(env, NULL, NULL));
5932 sanitize_dead_code(env);
5935 ret = check_max_stack_depth(env);
5938 /* program is valid, convert *(u32*)(ctx + off) accesses */
5939 ret = convert_ctx_accesses(env);
5942 ret = fixup_bpf_calls(env);
5945 ret = fixup_call_args(env);
5947 if (log->level && bpf_verifier_log_full(log))
5949 if (log->level && !log->ubuf) {
5951 goto err_release_maps;
5954 if (ret == 0 && env->used_map_cnt) {
5955 /* if program passed verifier, update used_maps in bpf_prog_info */
5956 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
5957 sizeof(env->used_maps[0]),
5960 if (!env->prog->aux->used_maps) {
5962 goto err_release_maps;
5965 memcpy(env->prog->aux->used_maps, env->used_maps,
5966 sizeof(env->used_maps[0]) * env->used_map_cnt);
5967 env->prog->aux->used_map_cnt = env->used_map_cnt;
5969 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
5970 * bpf_ld_imm64 instructions
5972 convert_pseudo_ld_imm64(env);
5976 if (!env->prog->aux->used_maps)
5977 /* if we didn't copy map pointers into bpf_prog_info, release
5978 * them now. Otherwise free_used_maps() will release them.
5983 mutex_unlock(&bpf_verifier_lock);
5984 vfree(env->insn_aux_data);