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_POISON ((void *)0xeB9F + POISON_POINTER_DELTA)
162 struct bpf_call_arg_meta {
163 struct bpf_map *map_ptr;
168 s64 msize_smax_value;
169 u64 msize_umax_value;
172 static DEFINE_MUTEX(bpf_verifier_lock);
174 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
179 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
181 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
182 "verifier log line truncated - local buffer too short\n");
184 n = min(log->len_total - log->len_used - 1, n);
187 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
193 /* log_level controls verbosity level of eBPF verifier.
194 * bpf_verifier_log_write() is used to dump the verification trace to the log,
195 * so the user can figure out what's wrong with the program
197 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
198 const char *fmt, ...)
202 if (!bpf_verifier_log_needed(&env->log))
206 bpf_verifier_vlog(&env->log, fmt, args);
209 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
211 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
213 struct bpf_verifier_env *env = private_data;
216 if (!bpf_verifier_log_needed(&env->log))
220 bpf_verifier_vlog(&env->log, fmt, args);
224 static bool type_is_pkt_pointer(enum bpf_reg_type type)
226 return type == PTR_TO_PACKET ||
227 type == PTR_TO_PACKET_META;
230 /* string representation of 'enum bpf_reg_type' */
231 static const char * const reg_type_str[] = {
233 [SCALAR_VALUE] = "inv",
234 [PTR_TO_CTX] = "ctx",
235 [CONST_PTR_TO_MAP] = "map_ptr",
236 [PTR_TO_MAP_VALUE] = "map_value",
237 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
238 [PTR_TO_STACK] = "fp",
239 [PTR_TO_PACKET] = "pkt",
240 [PTR_TO_PACKET_META] = "pkt_meta",
241 [PTR_TO_PACKET_END] = "pkt_end",
244 static void print_liveness(struct bpf_verifier_env *env,
245 enum bpf_reg_liveness live)
247 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
249 if (live & REG_LIVE_READ)
251 if (live & REG_LIVE_WRITTEN)
255 static struct bpf_func_state *func(struct bpf_verifier_env *env,
256 const struct bpf_reg_state *reg)
258 struct bpf_verifier_state *cur = env->cur_state;
260 return cur->frame[reg->frameno];
263 static void print_verifier_state(struct bpf_verifier_env *env,
264 const struct bpf_func_state *state)
266 const struct bpf_reg_state *reg;
271 verbose(env, " frame%d:", state->frameno);
272 for (i = 0; i < MAX_BPF_REG; i++) {
273 reg = &state->regs[i];
277 verbose(env, " R%d", i);
278 print_liveness(env, reg->live);
279 verbose(env, "=%s", reg_type_str[t]);
280 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
281 tnum_is_const(reg->var_off)) {
282 /* reg->off should be 0 for SCALAR_VALUE */
283 verbose(env, "%lld", reg->var_off.value + reg->off);
284 if (t == PTR_TO_STACK)
285 verbose(env, ",call_%d", func(env, reg)->callsite);
287 verbose(env, "(id=%d", reg->id);
288 if (t != SCALAR_VALUE)
289 verbose(env, ",off=%d", reg->off);
290 if (type_is_pkt_pointer(t))
291 verbose(env, ",r=%d", reg->range);
292 else if (t == CONST_PTR_TO_MAP ||
293 t == PTR_TO_MAP_VALUE ||
294 t == PTR_TO_MAP_VALUE_OR_NULL)
295 verbose(env, ",ks=%d,vs=%d",
296 reg->map_ptr->key_size,
297 reg->map_ptr->value_size);
298 if (tnum_is_const(reg->var_off)) {
299 /* Typically an immediate SCALAR_VALUE, but
300 * could be a pointer whose offset is too big
303 verbose(env, ",imm=%llx", reg->var_off.value);
305 if (reg->smin_value != reg->umin_value &&
306 reg->smin_value != S64_MIN)
307 verbose(env, ",smin_value=%lld",
308 (long long)reg->smin_value);
309 if (reg->smax_value != reg->umax_value &&
310 reg->smax_value != S64_MAX)
311 verbose(env, ",smax_value=%lld",
312 (long long)reg->smax_value);
313 if (reg->umin_value != 0)
314 verbose(env, ",umin_value=%llu",
315 (unsigned long long)reg->umin_value);
316 if (reg->umax_value != U64_MAX)
317 verbose(env, ",umax_value=%llu",
318 (unsigned long long)reg->umax_value);
319 if (!tnum_is_unknown(reg->var_off)) {
322 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
323 verbose(env, ",var_off=%s", tn_buf);
329 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
330 if (state->stack[i].slot_type[0] == STACK_SPILL) {
331 verbose(env, " fp%d",
332 (-i - 1) * BPF_REG_SIZE);
333 print_liveness(env, state->stack[i].spilled_ptr.live);
335 reg_type_str[state->stack[i].spilled_ptr.type]);
337 if (state->stack[i].slot_type[0] == STACK_ZERO)
338 verbose(env, " fp%d=0", (-i - 1) * BPF_REG_SIZE);
343 static int copy_stack_state(struct bpf_func_state *dst,
344 const struct bpf_func_state *src)
348 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
349 /* internal bug, make state invalid to reject the program */
350 memset(dst, 0, sizeof(*dst));
353 memcpy(dst->stack, src->stack,
354 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
358 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
359 * make it consume minimal amount of memory. check_stack_write() access from
360 * the program calls into realloc_func_state() to grow the stack size.
361 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
362 * which this function copies over. It points to previous bpf_verifier_state
363 * which is never reallocated
365 static int realloc_func_state(struct bpf_func_state *state, int size,
368 u32 old_size = state->allocated_stack;
369 struct bpf_stack_state *new_stack;
370 int slot = size / BPF_REG_SIZE;
372 if (size <= old_size || !size) {
375 state->allocated_stack = slot * BPF_REG_SIZE;
376 if (!size && old_size) {
382 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
388 memcpy(new_stack, state->stack,
389 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
390 memset(new_stack + old_size / BPF_REG_SIZE, 0,
391 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
393 state->allocated_stack = slot * BPF_REG_SIZE;
395 state->stack = new_stack;
399 static void free_func_state(struct bpf_func_state *state)
407 static void free_verifier_state(struct bpf_verifier_state *state,
412 for (i = 0; i <= state->curframe; i++) {
413 free_func_state(state->frame[i]);
414 state->frame[i] = NULL;
420 /* copy verifier state from src to dst growing dst stack space
421 * when necessary to accommodate larger src stack
423 static int copy_func_state(struct bpf_func_state *dst,
424 const struct bpf_func_state *src)
428 err = realloc_func_state(dst, src->allocated_stack, false);
431 memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack));
432 return copy_stack_state(dst, src);
435 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
436 const struct bpf_verifier_state *src)
438 struct bpf_func_state *dst;
441 /* if dst has more stack frames then src frame, free them */
442 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
443 free_func_state(dst_state->frame[i]);
444 dst_state->frame[i] = NULL;
446 dst_state->curframe = src->curframe;
447 dst_state->parent = src->parent;
448 for (i = 0; i <= src->curframe; i++) {
449 dst = dst_state->frame[i];
451 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
454 dst_state->frame[i] = dst;
456 err = copy_func_state(dst, src->frame[i]);
463 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
466 struct bpf_verifier_state *cur = env->cur_state;
467 struct bpf_verifier_stack_elem *elem, *head = env->head;
470 if (env->head == NULL)
474 err = copy_verifier_state(cur, &head->st);
479 *insn_idx = head->insn_idx;
481 *prev_insn_idx = head->prev_insn_idx;
483 free_verifier_state(&head->st, false);
490 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
491 int insn_idx, int prev_insn_idx)
493 struct bpf_verifier_state *cur = env->cur_state;
494 struct bpf_verifier_stack_elem *elem;
497 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
501 elem->insn_idx = insn_idx;
502 elem->prev_insn_idx = prev_insn_idx;
503 elem->next = env->head;
506 err = copy_verifier_state(&elem->st, cur);
509 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
510 verbose(env, "BPF program is too complex\n");
515 free_verifier_state(env->cur_state, true);
516 env->cur_state = NULL;
517 /* pop all elements and return */
518 while (!pop_stack(env, NULL, NULL));
522 #define CALLER_SAVED_REGS 6
523 static const int caller_saved[CALLER_SAVED_REGS] = {
524 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
527 static void __mark_reg_not_init(struct bpf_reg_state *reg);
529 /* Mark the unknown part of a register (variable offset or scalar value) as
530 * known to have the value @imm.
532 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
535 reg->var_off = tnum_const(imm);
536 reg->smin_value = (s64)imm;
537 reg->smax_value = (s64)imm;
538 reg->umin_value = imm;
539 reg->umax_value = imm;
542 /* Mark the 'variable offset' part of a register as zero. This should be
543 * used only on registers holding a pointer type.
545 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
547 __mark_reg_known(reg, 0);
550 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
552 __mark_reg_known(reg, 0);
554 reg->type = SCALAR_VALUE;
557 static void mark_reg_known_zero(struct bpf_verifier_env *env,
558 struct bpf_reg_state *regs, u32 regno)
560 if (WARN_ON(regno >= MAX_BPF_REG)) {
561 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
562 /* Something bad happened, let's kill all regs */
563 for (regno = 0; regno < MAX_BPF_REG; regno++)
564 __mark_reg_not_init(regs + regno);
567 __mark_reg_known_zero(regs + regno);
570 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
572 return type_is_pkt_pointer(reg->type);
575 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
577 return reg_is_pkt_pointer(reg) ||
578 reg->type == PTR_TO_PACKET_END;
581 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
582 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
583 enum bpf_reg_type which)
585 /* The register can already have a range from prior markings.
586 * This is fine as long as it hasn't been advanced from its
589 return reg->type == which &&
592 tnum_equals_const(reg->var_off, 0);
595 /* Attempts to improve min/max values based on var_off information */
596 static void __update_reg_bounds(struct bpf_reg_state *reg)
598 /* min signed is max(sign bit) | min(other bits) */
599 reg->smin_value = max_t(s64, reg->smin_value,
600 reg->var_off.value | (reg->var_off.mask & S64_MIN));
601 /* max signed is min(sign bit) | max(other bits) */
602 reg->smax_value = min_t(s64, reg->smax_value,
603 reg->var_off.value | (reg->var_off.mask & S64_MAX));
604 reg->umin_value = max(reg->umin_value, reg->var_off.value);
605 reg->umax_value = min(reg->umax_value,
606 reg->var_off.value | reg->var_off.mask);
609 /* Uses signed min/max values to inform unsigned, and vice-versa */
610 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
612 /* Learn sign from signed bounds.
613 * If we cannot cross the sign boundary, then signed and unsigned bounds
614 * are the same, so combine. This works even in the negative case, e.g.
615 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
617 if (reg->smin_value >= 0 || reg->smax_value < 0) {
618 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
620 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
624 /* Learn sign from unsigned bounds. Signed bounds cross the sign
625 * boundary, so we must be careful.
627 if ((s64)reg->umax_value >= 0) {
628 /* Positive. We can't learn anything from the smin, but smax
629 * is positive, hence safe.
631 reg->smin_value = reg->umin_value;
632 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
634 } else if ((s64)reg->umin_value < 0) {
635 /* Negative. We can't learn anything from the smax, but smin
636 * is negative, hence safe.
638 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
640 reg->smax_value = reg->umax_value;
644 /* Attempts to improve var_off based on unsigned min/max information */
645 static void __reg_bound_offset(struct bpf_reg_state *reg)
647 reg->var_off = tnum_intersect(reg->var_off,
648 tnum_range(reg->umin_value,
652 /* Reset the min/max bounds of a register */
653 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
655 reg->smin_value = S64_MIN;
656 reg->smax_value = S64_MAX;
658 reg->umax_value = U64_MAX;
661 /* Mark a register as having a completely unknown (scalar) value. */
662 static void __mark_reg_unknown(struct bpf_reg_state *reg)
664 reg->type = SCALAR_VALUE;
667 reg->var_off = tnum_unknown;
669 __mark_reg_unbounded(reg);
672 static void mark_reg_unknown(struct bpf_verifier_env *env,
673 struct bpf_reg_state *regs, u32 regno)
675 if (WARN_ON(regno >= MAX_BPF_REG)) {
676 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
677 /* Something bad happened, let's kill all regs except FP */
678 for (regno = 0; regno < BPF_REG_FP; regno++)
679 __mark_reg_not_init(regs + regno);
682 __mark_reg_unknown(regs + regno);
685 static void __mark_reg_not_init(struct bpf_reg_state *reg)
687 __mark_reg_unknown(reg);
688 reg->type = NOT_INIT;
691 static void mark_reg_not_init(struct bpf_verifier_env *env,
692 struct bpf_reg_state *regs, u32 regno)
694 if (WARN_ON(regno >= MAX_BPF_REG)) {
695 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
696 /* Something bad happened, let's kill all regs except FP */
697 for (regno = 0; regno < BPF_REG_FP; regno++)
698 __mark_reg_not_init(regs + regno);
701 __mark_reg_not_init(regs + regno);
704 static void init_reg_state(struct bpf_verifier_env *env,
705 struct bpf_func_state *state)
707 struct bpf_reg_state *regs = state->regs;
710 for (i = 0; i < MAX_BPF_REG; i++) {
711 mark_reg_not_init(env, regs, i);
712 regs[i].live = REG_LIVE_NONE;
716 regs[BPF_REG_FP].type = PTR_TO_STACK;
717 mark_reg_known_zero(env, regs, BPF_REG_FP);
718 regs[BPF_REG_FP].frameno = state->frameno;
720 /* 1st arg to a function */
721 regs[BPF_REG_1].type = PTR_TO_CTX;
722 mark_reg_known_zero(env, regs, BPF_REG_1);
725 #define BPF_MAIN_FUNC (-1)
726 static void init_func_state(struct bpf_verifier_env *env,
727 struct bpf_func_state *state,
728 int callsite, int frameno, int subprogno)
730 state->callsite = callsite;
731 state->frameno = frameno;
732 state->subprogno = subprogno;
733 init_reg_state(env, state);
737 SRC_OP, /* register is used as source operand */
738 DST_OP, /* register is used as destination operand */
739 DST_OP_NO_MARK /* same as above, check only, don't mark */
742 static int cmp_subprogs(const void *a, const void *b)
744 return *(int *)a - *(int *)b;
747 static int find_subprog(struct bpf_verifier_env *env, int off)
751 p = bsearch(&off, env->subprog_starts, env->subprog_cnt,
752 sizeof(env->subprog_starts[0]), cmp_subprogs);
755 return p - env->subprog_starts;
759 static int add_subprog(struct bpf_verifier_env *env, int off)
761 int insn_cnt = env->prog->len;
764 if (off >= insn_cnt || off < 0) {
765 verbose(env, "call to invalid destination\n");
768 ret = find_subprog(env, off);
771 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
772 verbose(env, "too many subprograms\n");
775 env->subprog_starts[env->subprog_cnt++] = off;
776 sort(env->subprog_starts, env->subprog_cnt,
777 sizeof(env->subprog_starts[0]), cmp_subprogs, NULL);
781 static int check_subprogs(struct bpf_verifier_env *env)
783 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
784 struct bpf_insn *insn = env->prog->insnsi;
785 int insn_cnt = env->prog->len;
787 /* determine subprog starts. The end is one before the next starts */
788 for (i = 0; i < insn_cnt; i++) {
789 if (insn[i].code != (BPF_JMP | BPF_CALL))
791 if (insn[i].src_reg != BPF_PSEUDO_CALL)
793 if (!env->allow_ptr_leaks) {
794 verbose(env, "function calls to other bpf functions are allowed for root only\n");
797 if (bpf_prog_is_dev_bound(env->prog->aux)) {
798 verbose(env, "function calls in offloaded programs are not supported yet\n");
801 ret = add_subprog(env, i + insn[i].imm + 1);
806 if (env->log.level > 1)
807 for (i = 0; i < env->subprog_cnt; i++)
808 verbose(env, "func#%d @%d\n", i, env->subprog_starts[i]);
810 /* now check that all jumps are within the same subprog */
812 if (env->subprog_cnt == cur_subprog)
813 subprog_end = insn_cnt;
815 subprog_end = env->subprog_starts[cur_subprog++];
816 for (i = 0; i < insn_cnt; i++) {
817 u8 code = insn[i].code;
819 if (BPF_CLASS(code) != BPF_JMP)
821 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
823 off = i + insn[i].off + 1;
824 if (off < subprog_start || off >= subprog_end) {
825 verbose(env, "jump out of range from insn %d to %d\n", i, off);
829 if (i == subprog_end - 1) {
830 /* to avoid fall-through from one subprog into another
831 * the last insn of the subprog should be either exit
832 * or unconditional jump back
834 if (code != (BPF_JMP | BPF_EXIT) &&
835 code != (BPF_JMP | BPF_JA)) {
836 verbose(env, "last insn is not an exit or jmp\n");
839 subprog_start = subprog_end;
840 if (env->subprog_cnt == cur_subprog)
841 subprog_end = insn_cnt;
843 subprog_end = env->subprog_starts[cur_subprog++];
850 struct bpf_verifier_state *skip_callee(struct bpf_verifier_env *env,
851 const struct bpf_verifier_state *state,
852 struct bpf_verifier_state *parent,
855 struct bpf_verifier_state *tmp = NULL;
857 /* 'parent' could be a state of caller and
858 * 'state' could be a state of callee. In such case
859 * parent->curframe < state->curframe
860 * and it's ok for r1 - r5 registers
862 * 'parent' could be a callee's state after it bpf_exit-ed.
863 * In such case parent->curframe > state->curframe
864 * and it's ok for r0 only
866 if (parent->curframe == state->curframe ||
867 (parent->curframe < state->curframe &&
868 regno >= BPF_REG_1 && regno <= BPF_REG_5) ||
869 (parent->curframe > state->curframe &&
873 if (parent->curframe > state->curframe &&
874 regno >= BPF_REG_6) {
875 /* for callee saved regs we have to skip the whole chain
876 * of states that belong to callee and mark as LIVE_READ
877 * the registers before the call
880 while (tmp && tmp->curframe != state->curframe) {
891 verbose(env, "verifier bug regno %d tmp %p\n", regno, tmp);
892 verbose(env, "regno %d parent frame %d current frame %d\n",
893 regno, parent->curframe, state->curframe);
897 static int mark_reg_read(struct bpf_verifier_env *env,
898 const struct bpf_verifier_state *state,
899 struct bpf_verifier_state *parent,
902 bool writes = parent == state->parent; /* Observe write marks */
904 if (regno == BPF_REG_FP)
905 /* We don't need to worry about FP liveness because it's read-only */
909 /* if read wasn't screened by an earlier write ... */
910 if (writes && state->frame[state->curframe]->regs[regno].live & REG_LIVE_WRITTEN)
912 parent = skip_callee(env, state, parent, regno);
915 /* ... then we depend on parent's value */
916 parent->frame[parent->curframe]->regs[regno].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 return mark_reg_read(env, vstate, vstate->parent, regno);
944 /* check whether register used as dest operand can be written to */
945 if (regno == BPF_REG_FP) {
946 verbose(env, "frame pointer is read only\n");
949 regs[regno].live |= REG_LIVE_WRITTEN;
951 mark_reg_unknown(env, regs, regno);
956 static bool is_spillable_regtype(enum bpf_reg_type type)
959 case PTR_TO_MAP_VALUE:
960 case PTR_TO_MAP_VALUE_OR_NULL:
964 case PTR_TO_PACKET_META:
965 case PTR_TO_PACKET_END:
966 case CONST_PTR_TO_MAP:
973 /* Does this register contain a constant zero? */
974 static bool register_is_null(struct bpf_reg_state *reg)
976 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
979 /* check_stack_read/write functions track spill/fill of registers,
980 * stack boundary and alignment are checked in check_mem_access()
982 static int check_stack_write(struct bpf_verifier_env *env,
983 struct bpf_func_state *state, /* func where register points to */
984 int off, int size, int value_regno)
986 struct bpf_func_state *cur; /* state of the current function */
987 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
988 enum bpf_reg_type type;
990 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
994 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
995 * so it's aligned access and [off, off + size) are within stack limits
997 if (!env->allow_ptr_leaks &&
998 state->stack[spi].slot_type[0] == STACK_SPILL &&
999 size != BPF_REG_SIZE) {
1000 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1004 cur = env->cur_state->frame[env->cur_state->curframe];
1005 if (value_regno >= 0 &&
1006 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1008 /* register containing pointer is being spilled into stack */
1009 if (size != BPF_REG_SIZE) {
1010 verbose(env, "invalid size of register spill\n");
1014 if (state != cur && type == PTR_TO_STACK) {
1015 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1019 /* save register state */
1020 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1021 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1023 for (i = 0; i < BPF_REG_SIZE; i++)
1024 state->stack[spi].slot_type[i] = STACK_SPILL;
1026 u8 type = STACK_MISC;
1028 /* regular write of data into stack */
1029 state->stack[spi].spilled_ptr = (struct bpf_reg_state) {};
1031 /* only mark the slot as written if all 8 bytes were written
1032 * otherwise read propagation may incorrectly stop too soon
1033 * when stack slots are partially written.
1034 * This heuristic means that read propagation will be
1035 * conservative, since it will add reg_live_read marks
1036 * to stack slots all the way to first state when programs
1037 * writes+reads less than 8 bytes
1039 if (size == BPF_REG_SIZE)
1040 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1042 /* when we zero initialize stack slots mark them as such */
1043 if (value_regno >= 0 &&
1044 register_is_null(&cur->regs[value_regno]))
1047 for (i = 0; i < size; i++)
1048 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1054 /* registers of every function are unique and mark_reg_read() propagates
1055 * the liveness in the following cases:
1056 * - from callee into caller for R1 - R5 that were used as arguments
1057 * - from caller into callee for R0 that used as result of the call
1058 * - from caller to the same caller skipping states of the callee for R6 - R9,
1059 * since R6 - R9 are callee saved by implicit function prologue and
1060 * caller's R6 != callee's R6, so when we propagate liveness up to
1061 * parent states we need to skip callee states for R6 - R9.
1063 * stack slot marking is different, since stacks of caller and callee are
1064 * accessible in both (since caller can pass a pointer to caller's stack to
1065 * callee which can pass it to another function), hence mark_stack_slot_read()
1066 * has to propagate the stack liveness to all parent states at given frame number.
1076 * First *ptr is reading from f1's stack and mark_stack_slot_read() has
1077 * to mark liveness at the f1's frame and not f2's frame.
1078 * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has
1079 * to propagate liveness to f2 states at f1's frame level and further into
1080 * f1 states at f1's frame level until write into that stack slot
1082 static void mark_stack_slot_read(struct bpf_verifier_env *env,
1083 const struct bpf_verifier_state *state,
1084 struct bpf_verifier_state *parent,
1085 int slot, int frameno)
1087 bool writes = parent == state->parent; /* Observe write marks */
1090 if (parent->frame[frameno]->allocated_stack <= slot * BPF_REG_SIZE)
1091 /* since LIVE_WRITTEN mark is only done for full 8-byte
1092 * write the read marks are conservative and parent
1093 * state may not even have the stack allocated. In such case
1094 * end the propagation, since the loop reached beginning
1098 /* if read wasn't screened by an earlier write ... */
1099 if (writes && state->frame[frameno]->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN)
1101 /* ... then we depend on parent's value */
1102 parent->frame[frameno]->stack[slot].spilled_ptr.live |= REG_LIVE_READ;
1104 parent = state->parent;
1109 static int check_stack_read(struct bpf_verifier_env *env,
1110 struct bpf_func_state *reg_state /* func where register points to */,
1111 int off, int size, int value_regno)
1113 struct bpf_verifier_state *vstate = env->cur_state;
1114 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1115 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1118 if (reg_state->allocated_stack <= slot) {
1119 verbose(env, "invalid read from stack off %d+0 size %d\n",
1123 stype = reg_state->stack[spi].slot_type;
1125 if (stype[0] == STACK_SPILL) {
1126 if (size != BPF_REG_SIZE) {
1127 verbose(env, "invalid size of register spill\n");
1130 for (i = 1; i < BPF_REG_SIZE; i++) {
1131 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1132 verbose(env, "corrupted spill memory\n");
1137 if (value_regno >= 0) {
1138 /* restore register state from stack */
1139 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1140 /* mark reg as written since spilled pointer state likely
1141 * has its liveness marks cleared by is_state_visited()
1142 * which resets stack/reg liveness for state transitions
1144 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1146 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1147 reg_state->frameno);
1152 for (i = 0; i < size; i++) {
1153 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1155 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1159 verbose(env, "invalid read from stack off %d+%d size %d\n",
1163 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1164 reg_state->frameno);
1165 if (value_regno >= 0) {
1166 if (zeros == size) {
1167 /* any size read into register is zero extended,
1168 * so the whole register == const_zero
1170 __mark_reg_const_zero(&state->regs[value_regno]);
1172 /* have read misc data from the stack */
1173 mark_reg_unknown(env, state->regs, value_regno);
1175 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1181 /* check read/write into map element returned by bpf_map_lookup_elem() */
1182 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1183 int size, bool zero_size_allowed)
1185 struct bpf_reg_state *regs = cur_regs(env);
1186 struct bpf_map *map = regs[regno].map_ptr;
1188 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1189 off + size > map->value_size) {
1190 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1191 map->value_size, off, size);
1197 /* check read/write into a map element with possible variable offset */
1198 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1199 int off, int size, bool zero_size_allowed)
1201 struct bpf_verifier_state *vstate = env->cur_state;
1202 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1203 struct bpf_reg_state *reg = &state->regs[regno];
1206 /* We may have adjusted the register to this map value, so we
1207 * need to try adding each of min_value and max_value to off
1208 * to make sure our theoretical access will be safe.
1211 print_verifier_state(env, state);
1212 /* The minimum value is only important with signed
1213 * comparisons where we can't assume the floor of a
1214 * value is 0. If we are using signed variables for our
1215 * index'es we need to make sure that whatever we use
1216 * will have a set floor within our range.
1218 if (reg->smin_value < 0) {
1219 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1223 err = __check_map_access(env, regno, reg->smin_value + off, size,
1226 verbose(env, "R%d min value is outside of the array range\n",
1231 /* If we haven't set a max value then we need to bail since we can't be
1232 * sure we won't do bad things.
1233 * If reg->umax_value + off could overflow, treat that as unbounded too.
1235 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1236 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1240 err = __check_map_access(env, regno, reg->umax_value + off, size,
1243 verbose(env, "R%d max value is outside of the array range\n",
1248 #define MAX_PACKET_OFF 0xffff
1250 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1251 const struct bpf_call_arg_meta *meta,
1252 enum bpf_access_type t)
1254 switch (env->prog->type) {
1255 case BPF_PROG_TYPE_LWT_IN:
1256 case BPF_PROG_TYPE_LWT_OUT:
1257 /* dst_input() and dst_output() can't write for now */
1261 case BPF_PROG_TYPE_SCHED_CLS:
1262 case BPF_PROG_TYPE_SCHED_ACT:
1263 case BPF_PROG_TYPE_XDP:
1264 case BPF_PROG_TYPE_LWT_XMIT:
1265 case BPF_PROG_TYPE_SK_SKB:
1266 case BPF_PROG_TYPE_SK_MSG:
1268 return meta->pkt_access;
1270 env->seen_direct_write = true;
1277 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1278 int off, int size, bool zero_size_allowed)
1280 struct bpf_reg_state *regs = cur_regs(env);
1281 struct bpf_reg_state *reg = ®s[regno];
1283 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1284 (u64)off + size > reg->range) {
1285 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1286 off, size, regno, reg->id, reg->off, reg->range);
1292 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1293 int size, bool zero_size_allowed)
1295 struct bpf_reg_state *regs = cur_regs(env);
1296 struct bpf_reg_state *reg = ®s[regno];
1299 /* We may have added a variable offset to the packet pointer; but any
1300 * reg->range we have comes after that. We are only checking the fixed
1304 /* We don't allow negative numbers, because we aren't tracking enough
1305 * detail to prove they're safe.
1307 if (reg->smin_value < 0) {
1308 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1312 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1314 verbose(env, "R%d offset is outside of the packet\n", regno);
1320 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1321 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1322 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1324 struct bpf_insn_access_aux info = {
1325 .reg_type = *reg_type,
1328 if (env->ops->is_valid_access &&
1329 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1330 /* A non zero info.ctx_field_size indicates that this field is a
1331 * candidate for later verifier transformation to load the whole
1332 * field and then apply a mask when accessed with a narrower
1333 * access than actual ctx access size. A zero info.ctx_field_size
1334 * will only allow for whole field access and rejects any other
1335 * type of narrower access.
1337 *reg_type = info.reg_type;
1339 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1340 /* remember the offset of last byte accessed in ctx */
1341 if (env->prog->aux->max_ctx_offset < off + size)
1342 env->prog->aux->max_ctx_offset = off + size;
1346 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1350 static bool __is_pointer_value(bool allow_ptr_leaks,
1351 const struct bpf_reg_state *reg)
1353 if (allow_ptr_leaks)
1356 return reg->type != SCALAR_VALUE;
1359 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1361 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1364 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1366 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1368 return reg->type == PTR_TO_CTX;
1371 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1373 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1375 return type_is_pkt_pointer(reg->type);
1378 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1379 const struct bpf_reg_state *reg,
1380 int off, int size, bool strict)
1382 struct tnum reg_off;
1385 /* Byte size accesses are always allowed. */
1386 if (!strict || size == 1)
1389 /* For platforms that do not have a Kconfig enabling
1390 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1391 * NET_IP_ALIGN is universally set to '2'. And on platforms
1392 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1393 * to this code only in strict mode where we want to emulate
1394 * the NET_IP_ALIGN==2 checking. Therefore use an
1395 * unconditional IP align value of '2'.
1399 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1400 if (!tnum_is_aligned(reg_off, size)) {
1403 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1405 "misaligned packet access off %d+%s+%d+%d size %d\n",
1406 ip_align, tn_buf, reg->off, off, size);
1413 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1414 const struct bpf_reg_state *reg,
1415 const char *pointer_desc,
1416 int off, int size, bool strict)
1418 struct tnum reg_off;
1420 /* Byte size accesses are always allowed. */
1421 if (!strict || size == 1)
1424 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1425 if (!tnum_is_aligned(reg_off, size)) {
1428 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1429 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1430 pointer_desc, tn_buf, reg->off, off, size);
1437 static int check_ptr_alignment(struct bpf_verifier_env *env,
1438 const struct bpf_reg_state *reg, int off,
1439 int size, bool strict_alignment_once)
1441 bool strict = env->strict_alignment || strict_alignment_once;
1442 const char *pointer_desc = "";
1444 switch (reg->type) {
1446 case PTR_TO_PACKET_META:
1447 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1448 * right in front, treat it the very same way.
1450 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1451 case PTR_TO_MAP_VALUE:
1452 pointer_desc = "value ";
1455 pointer_desc = "context ";
1458 pointer_desc = "stack ";
1459 /* The stack spill tracking logic in check_stack_write()
1460 * and check_stack_read() relies on stack accesses being
1468 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1472 static int update_stack_depth(struct bpf_verifier_env *env,
1473 const struct bpf_func_state *func,
1476 u16 stack = env->subprog_stack_depth[func->subprogno];
1481 /* update known max for given subprogram */
1482 env->subprog_stack_depth[func->subprogno] = -off;
1486 /* starting from main bpf function walk all instructions of the function
1487 * and recursively walk all callees that given function can call.
1488 * Ignore jump and exit insns.
1489 * Since recursion is prevented by check_cfg() this algorithm
1490 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1492 static int check_max_stack_depth(struct bpf_verifier_env *env)
1494 int depth = 0, frame = 0, subprog = 0, i = 0, subprog_end;
1495 struct bpf_insn *insn = env->prog->insnsi;
1496 int insn_cnt = env->prog->len;
1497 int ret_insn[MAX_CALL_FRAMES];
1498 int ret_prog[MAX_CALL_FRAMES];
1501 /* round up to 32-bytes, since this is granularity
1502 * of interpreter stack size
1504 depth += round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1505 if (depth > MAX_BPF_STACK) {
1506 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1511 if (env->subprog_cnt == subprog)
1512 subprog_end = insn_cnt;
1514 subprog_end = env->subprog_starts[subprog];
1515 for (; i < subprog_end; i++) {
1516 if (insn[i].code != (BPF_JMP | BPF_CALL))
1518 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1520 /* remember insn and function to return to */
1521 ret_insn[frame] = i + 1;
1522 ret_prog[frame] = subprog;
1524 /* find the callee */
1525 i = i + insn[i].imm + 1;
1526 subprog = find_subprog(env, i);
1528 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1534 if (frame >= MAX_CALL_FRAMES) {
1535 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1540 /* end of for() loop means the last insn of the 'subprog'
1541 * was reached. Doesn't matter whether it was JA or EXIT
1545 depth -= round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1547 i = ret_insn[frame];
1548 subprog = ret_prog[frame];
1552 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1553 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1554 const struct bpf_insn *insn, int idx)
1556 int start = idx + insn->imm + 1, subprog;
1558 subprog = find_subprog(env, start);
1560 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1565 return env->subprog_stack_depth[subprog];
1569 /* truncate register to smaller size (in bytes)
1570 * must be called with size < BPF_REG_SIZE
1572 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1576 /* clear high bits in bit representation */
1577 reg->var_off = tnum_cast(reg->var_off, size);
1579 /* fix arithmetic bounds */
1580 mask = ((u64)1 << (size * 8)) - 1;
1581 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1582 reg->umin_value &= mask;
1583 reg->umax_value &= mask;
1585 reg->umin_value = 0;
1586 reg->umax_value = mask;
1588 reg->smin_value = reg->umin_value;
1589 reg->smax_value = reg->umax_value;
1592 /* check whether memory at (regno + off) is accessible for t = (read | write)
1593 * if t==write, value_regno is a register which value is stored into memory
1594 * if t==read, value_regno is a register which will receive the value from memory
1595 * if t==write && value_regno==-1, some unknown value is stored into memory
1596 * if t==read && value_regno==-1, don't care what we read from memory
1598 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1599 int off, int bpf_size, enum bpf_access_type t,
1600 int value_regno, bool strict_alignment_once)
1602 struct bpf_reg_state *regs = cur_regs(env);
1603 struct bpf_reg_state *reg = regs + regno;
1604 struct bpf_func_state *state;
1607 size = bpf_size_to_bytes(bpf_size);
1611 /* alignment checks will add in reg->off themselves */
1612 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1616 /* for access checks, reg->off is just part of off */
1619 if (reg->type == PTR_TO_MAP_VALUE) {
1620 if (t == BPF_WRITE && value_regno >= 0 &&
1621 is_pointer_value(env, value_regno)) {
1622 verbose(env, "R%d leaks addr into map\n", value_regno);
1626 err = check_map_access(env, regno, off, size, false);
1627 if (!err && t == BPF_READ && value_regno >= 0)
1628 mark_reg_unknown(env, regs, value_regno);
1630 } else if (reg->type == PTR_TO_CTX) {
1631 enum bpf_reg_type reg_type = SCALAR_VALUE;
1633 if (t == BPF_WRITE && value_regno >= 0 &&
1634 is_pointer_value(env, value_regno)) {
1635 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1638 /* ctx accesses must be at a fixed offset, so that we can
1639 * determine what type of data were returned.
1643 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1644 regno, reg->off, off - reg->off);
1647 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1650 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1652 "variable ctx access var_off=%s off=%d size=%d",
1656 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1657 if (!err && t == BPF_READ && value_regno >= 0) {
1658 /* ctx access returns either a scalar, or a
1659 * PTR_TO_PACKET[_META,_END]. In the latter
1660 * case, we know the offset is zero.
1662 if (reg_type == SCALAR_VALUE)
1663 mark_reg_unknown(env, regs, value_regno);
1665 mark_reg_known_zero(env, regs,
1667 regs[value_regno].id = 0;
1668 regs[value_regno].off = 0;
1669 regs[value_regno].range = 0;
1670 regs[value_regno].type = reg_type;
1673 } else if (reg->type == PTR_TO_STACK) {
1674 /* stack accesses must be at a fixed offset, so that we can
1675 * determine what type of data were returned.
1676 * See check_stack_read().
1678 if (!tnum_is_const(reg->var_off)) {
1681 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1682 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1686 off += reg->var_off.value;
1687 if (off >= 0 || off < -MAX_BPF_STACK) {
1688 verbose(env, "invalid stack off=%d size=%d\n", off,
1693 state = func(env, reg);
1694 err = update_stack_depth(env, state, off);
1699 err = check_stack_write(env, state, off, size,
1702 err = check_stack_read(env, state, off, size,
1704 } else if (reg_is_pkt_pointer(reg)) {
1705 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1706 verbose(env, "cannot write into packet\n");
1709 if (t == BPF_WRITE && value_regno >= 0 &&
1710 is_pointer_value(env, value_regno)) {
1711 verbose(env, "R%d leaks addr into packet\n",
1715 err = check_packet_access(env, regno, off, size, false);
1716 if (!err && t == BPF_READ && value_regno >= 0)
1717 mark_reg_unknown(env, regs, value_regno);
1719 verbose(env, "R%d invalid mem access '%s'\n", regno,
1720 reg_type_str[reg->type]);
1724 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1725 regs[value_regno].type == SCALAR_VALUE) {
1726 /* b/h/w load zero-extends, mark upper bits as known 0 */
1727 coerce_reg_to_size(®s[value_regno], size);
1732 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1736 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1738 verbose(env, "BPF_XADD uses reserved fields\n");
1742 /* check src1 operand */
1743 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1747 /* check src2 operand */
1748 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1752 if (is_pointer_value(env, insn->src_reg)) {
1753 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1757 if (is_ctx_reg(env, insn->dst_reg) ||
1758 is_pkt_reg(env, insn->dst_reg)) {
1759 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1760 insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ?
1761 "context" : "packet");
1765 /* check whether atomic_add can read the memory */
1766 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1767 BPF_SIZE(insn->code), BPF_READ, -1, true);
1771 /* check whether atomic_add can write into the same memory */
1772 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1773 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1776 /* when register 'regno' is passed into function that will read 'access_size'
1777 * bytes from that pointer, make sure that it's within stack boundary
1778 * and all elements of stack are initialized.
1779 * Unlike most pointer bounds-checking functions, this one doesn't take an
1780 * 'off' argument, so it has to add in reg->off itself.
1782 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1783 int access_size, bool zero_size_allowed,
1784 struct bpf_call_arg_meta *meta)
1786 struct bpf_reg_state *reg = cur_regs(env) + regno;
1787 struct bpf_func_state *state = func(env, reg);
1788 int off, i, slot, spi;
1790 if (reg->type != PTR_TO_STACK) {
1791 /* Allow zero-byte read from NULL, regardless of pointer type */
1792 if (zero_size_allowed && access_size == 0 &&
1793 register_is_null(reg))
1796 verbose(env, "R%d type=%s expected=%s\n", regno,
1797 reg_type_str[reg->type],
1798 reg_type_str[PTR_TO_STACK]);
1802 /* Only allow fixed-offset stack reads */
1803 if (!tnum_is_const(reg->var_off)) {
1806 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1807 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1811 off = reg->off + reg->var_off.value;
1812 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1813 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1814 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1815 regno, off, access_size);
1819 if (meta && meta->raw_mode) {
1820 meta->access_size = access_size;
1821 meta->regno = regno;
1825 for (i = 0; i < access_size; i++) {
1828 slot = -(off + i) - 1;
1829 spi = slot / BPF_REG_SIZE;
1830 if (state->allocated_stack <= slot)
1832 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
1833 if (*stype == STACK_MISC)
1835 if (*stype == STACK_ZERO) {
1836 /* helper can write anything into the stack */
1837 *stype = STACK_MISC;
1841 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1842 off, i, access_size);
1845 /* reading any byte out of 8-byte 'spill_slot' will cause
1846 * the whole slot to be marked as 'read'
1848 mark_stack_slot_read(env, env->cur_state, env->cur_state->parent,
1849 spi, state->frameno);
1851 return update_stack_depth(env, state, off);
1854 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1855 int access_size, bool zero_size_allowed,
1856 struct bpf_call_arg_meta *meta)
1858 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1860 switch (reg->type) {
1862 case PTR_TO_PACKET_META:
1863 return check_packet_access(env, regno, reg->off, access_size,
1865 case PTR_TO_MAP_VALUE:
1866 return check_map_access(env, regno, reg->off, access_size,
1868 default: /* scalar_value|ptr_to_stack or invalid ptr */
1869 return check_stack_boundary(env, regno, access_size,
1870 zero_size_allowed, meta);
1874 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
1876 return type == ARG_PTR_TO_MEM ||
1877 type == ARG_PTR_TO_MEM_OR_NULL ||
1878 type == ARG_PTR_TO_UNINIT_MEM;
1881 static bool arg_type_is_mem_size(enum bpf_arg_type type)
1883 return type == ARG_CONST_SIZE ||
1884 type == ARG_CONST_SIZE_OR_ZERO;
1887 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1888 enum bpf_arg_type arg_type,
1889 struct bpf_call_arg_meta *meta)
1891 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1892 enum bpf_reg_type expected_type, type = reg->type;
1895 if (arg_type == ARG_DONTCARE)
1898 err = check_reg_arg(env, regno, SRC_OP);
1902 if (arg_type == ARG_ANYTHING) {
1903 if (is_pointer_value(env, regno)) {
1904 verbose(env, "R%d leaks addr into helper function\n",
1911 if (type_is_pkt_pointer(type) &&
1912 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1913 verbose(env, "helper access to the packet is not allowed\n");
1917 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1918 arg_type == ARG_PTR_TO_MAP_VALUE) {
1919 expected_type = PTR_TO_STACK;
1920 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
1921 type != expected_type)
1923 } else if (arg_type == ARG_CONST_SIZE ||
1924 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1925 expected_type = SCALAR_VALUE;
1926 if (type != expected_type)
1928 } else if (arg_type == ARG_CONST_MAP_PTR) {
1929 expected_type = CONST_PTR_TO_MAP;
1930 if (type != expected_type)
1932 } else if (arg_type == ARG_PTR_TO_CTX) {
1933 expected_type = PTR_TO_CTX;
1934 if (type != expected_type)
1936 } else if (arg_type_is_mem_ptr(arg_type)) {
1937 expected_type = PTR_TO_STACK;
1938 /* One exception here. In case function allows for NULL to be
1939 * passed in as argument, it's a SCALAR_VALUE type. Final test
1940 * happens during stack boundary checking.
1942 if (register_is_null(reg) &&
1943 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1944 /* final test in check_stack_boundary() */;
1945 else if (!type_is_pkt_pointer(type) &&
1946 type != PTR_TO_MAP_VALUE &&
1947 type != expected_type)
1949 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1951 verbose(env, "unsupported arg_type %d\n", arg_type);
1955 if (arg_type == ARG_CONST_MAP_PTR) {
1956 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
1957 meta->map_ptr = reg->map_ptr;
1958 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
1959 /* bpf_map_xxx(..., map_ptr, ..., key) call:
1960 * check that [key, key + map->key_size) are within
1961 * stack limits and initialized
1963 if (!meta->map_ptr) {
1964 /* in function declaration map_ptr must come before
1965 * map_key, so that it's verified and known before
1966 * we have to check map_key here. Otherwise it means
1967 * that kernel subsystem misconfigured verifier
1969 verbose(env, "invalid map_ptr to access map->key\n");
1972 err = check_helper_mem_access(env, regno,
1973 meta->map_ptr->key_size, false,
1975 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
1976 /* bpf_map_xxx(..., map_ptr, ..., value) call:
1977 * check [value, value + map->value_size) validity
1979 if (!meta->map_ptr) {
1980 /* kernel subsystem misconfigured verifier */
1981 verbose(env, "invalid map_ptr to access map->value\n");
1984 err = check_helper_mem_access(env, regno,
1985 meta->map_ptr->value_size, false,
1987 } else if (arg_type_is_mem_size(arg_type)) {
1988 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
1990 /* remember the mem_size which may be used later
1991 * to refine return values.
1993 meta->msize_smax_value = reg->smax_value;
1994 meta->msize_umax_value = reg->umax_value;
1996 /* The register is SCALAR_VALUE; the access check
1997 * happens using its boundaries.
1999 if (!tnum_is_const(reg->var_off))
2000 /* For unprivileged variable accesses, disable raw
2001 * mode so that the program is required to
2002 * initialize all the memory that the helper could
2003 * just partially fill up.
2007 if (reg->smin_value < 0) {
2008 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2013 if (reg->umin_value == 0) {
2014 err = check_helper_mem_access(env, regno - 1, 0,
2021 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2022 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2026 err = check_helper_mem_access(env, regno - 1,
2028 zero_size_allowed, meta);
2033 verbose(env, "R%d type=%s expected=%s\n", regno,
2034 reg_type_str[type], reg_type_str[expected_type]);
2038 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2039 struct bpf_map *map, int func_id)
2044 /* We need a two way check, first is from map perspective ... */
2045 switch (map->map_type) {
2046 case BPF_MAP_TYPE_PROG_ARRAY:
2047 if (func_id != BPF_FUNC_tail_call)
2050 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2051 if (func_id != BPF_FUNC_perf_event_read &&
2052 func_id != BPF_FUNC_perf_event_output &&
2053 func_id != BPF_FUNC_perf_event_read_value)
2056 case BPF_MAP_TYPE_STACK_TRACE:
2057 if (func_id != BPF_FUNC_get_stackid)
2060 case BPF_MAP_TYPE_CGROUP_ARRAY:
2061 if (func_id != BPF_FUNC_skb_under_cgroup &&
2062 func_id != BPF_FUNC_current_task_under_cgroup)
2065 /* devmap returns a pointer to a live net_device ifindex that we cannot
2066 * allow to be modified from bpf side. So do not allow lookup elements
2069 case BPF_MAP_TYPE_DEVMAP:
2070 if (func_id != BPF_FUNC_redirect_map)
2073 /* Restrict bpf side of cpumap, open when use-cases appear */
2074 case BPF_MAP_TYPE_CPUMAP:
2075 if (func_id != BPF_FUNC_redirect_map)
2078 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2079 case BPF_MAP_TYPE_HASH_OF_MAPS:
2080 if (func_id != BPF_FUNC_map_lookup_elem)
2083 case BPF_MAP_TYPE_SOCKMAP:
2084 if (func_id != BPF_FUNC_sk_redirect_map &&
2085 func_id != BPF_FUNC_sock_map_update &&
2086 func_id != BPF_FUNC_map_delete_elem &&
2087 func_id != BPF_FUNC_msg_redirect_map)
2094 /* ... and second from the function itself. */
2096 case BPF_FUNC_tail_call:
2097 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2099 if (env->subprog_cnt) {
2100 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2104 case BPF_FUNC_perf_event_read:
2105 case BPF_FUNC_perf_event_output:
2106 case BPF_FUNC_perf_event_read_value:
2107 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2110 case BPF_FUNC_get_stackid:
2111 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2114 case BPF_FUNC_current_task_under_cgroup:
2115 case BPF_FUNC_skb_under_cgroup:
2116 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2119 case BPF_FUNC_redirect_map:
2120 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2121 map->map_type != BPF_MAP_TYPE_CPUMAP)
2124 case BPF_FUNC_sk_redirect_map:
2125 case BPF_FUNC_msg_redirect_map:
2126 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2129 case BPF_FUNC_sock_map_update:
2130 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2139 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2140 map->map_type, func_id_name(func_id), func_id);
2144 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2148 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2150 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2152 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2154 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2156 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2159 /* We only support one arg being in raw mode at the moment,
2160 * which is sufficient for the helper functions we have
2166 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2167 enum bpf_arg_type arg_next)
2169 return (arg_type_is_mem_ptr(arg_curr) &&
2170 !arg_type_is_mem_size(arg_next)) ||
2171 (!arg_type_is_mem_ptr(arg_curr) &&
2172 arg_type_is_mem_size(arg_next));
2175 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2177 /* bpf_xxx(..., buf, len) call will access 'len'
2178 * bytes from memory 'buf'. Both arg types need
2179 * to be paired, so make sure there's no buggy
2180 * helper function specification.
2182 if (arg_type_is_mem_size(fn->arg1_type) ||
2183 arg_type_is_mem_ptr(fn->arg5_type) ||
2184 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2185 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2186 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2187 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2193 static int check_func_proto(const struct bpf_func_proto *fn)
2195 return check_raw_mode_ok(fn) &&
2196 check_arg_pair_ok(fn) ? 0 : -EINVAL;
2199 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2200 * are now invalid, so turn them into unknown SCALAR_VALUE.
2202 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2203 struct bpf_func_state *state)
2205 struct bpf_reg_state *regs = state->regs, *reg;
2208 for (i = 0; i < MAX_BPF_REG; i++)
2209 if (reg_is_pkt_pointer_any(®s[i]))
2210 mark_reg_unknown(env, regs, i);
2212 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2213 if (state->stack[i].slot_type[0] != STACK_SPILL)
2215 reg = &state->stack[i].spilled_ptr;
2216 if (reg_is_pkt_pointer_any(reg))
2217 __mark_reg_unknown(reg);
2221 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2223 struct bpf_verifier_state *vstate = env->cur_state;
2226 for (i = 0; i <= vstate->curframe; i++)
2227 __clear_all_pkt_pointers(env, vstate->frame[i]);
2230 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2233 struct bpf_verifier_state *state = env->cur_state;
2234 struct bpf_func_state *caller, *callee;
2235 int i, subprog, target_insn;
2237 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2238 verbose(env, "the call stack of %d frames is too deep\n",
2239 state->curframe + 2);
2243 target_insn = *insn_idx + insn->imm;
2244 subprog = find_subprog(env, target_insn + 1);
2246 verbose(env, "verifier bug. No program starts at insn %d\n",
2251 caller = state->frame[state->curframe];
2252 if (state->frame[state->curframe + 1]) {
2253 verbose(env, "verifier bug. Frame %d already allocated\n",
2254 state->curframe + 1);
2258 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2261 state->frame[state->curframe + 1] = callee;
2263 /* callee cannot access r0, r6 - r9 for reading and has to write
2264 * into its own stack before reading from it.
2265 * callee can read/write into caller's stack
2267 init_func_state(env, callee,
2268 /* remember the callsite, it will be used by bpf_exit */
2269 *insn_idx /* callsite */,
2270 state->curframe + 1 /* frameno within this callchain */,
2271 subprog + 1 /* subprog number within this prog */);
2273 /* copy r1 - r5 args that callee can access */
2274 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2275 callee->regs[i] = caller->regs[i];
2277 /* after the call regsiters r0 - r5 were scratched */
2278 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2279 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2280 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2283 /* only increment it after check_reg_arg() finished */
2286 /* and go analyze first insn of the callee */
2287 *insn_idx = target_insn;
2289 if (env->log.level) {
2290 verbose(env, "caller:\n");
2291 print_verifier_state(env, caller);
2292 verbose(env, "callee:\n");
2293 print_verifier_state(env, callee);
2298 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2300 struct bpf_verifier_state *state = env->cur_state;
2301 struct bpf_func_state *caller, *callee;
2302 struct bpf_reg_state *r0;
2304 callee = state->frame[state->curframe];
2305 r0 = &callee->regs[BPF_REG_0];
2306 if (r0->type == PTR_TO_STACK) {
2307 /* technically it's ok to return caller's stack pointer
2308 * (or caller's caller's pointer) back to the caller,
2309 * since these pointers are valid. Only current stack
2310 * pointer will be invalid as soon as function exits,
2311 * but let's be conservative
2313 verbose(env, "cannot return stack pointer to the caller\n");
2318 caller = state->frame[state->curframe];
2319 /* return to the caller whatever r0 had in the callee */
2320 caller->regs[BPF_REG_0] = *r0;
2322 *insn_idx = callee->callsite + 1;
2323 if (env->log.level) {
2324 verbose(env, "returning from callee:\n");
2325 print_verifier_state(env, callee);
2326 verbose(env, "to caller at %d:\n", *insn_idx);
2327 print_verifier_state(env, caller);
2329 /* clear everything in the callee */
2330 free_func_state(callee);
2331 state->frame[state->curframe + 1] = NULL;
2335 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2337 struct bpf_call_arg_meta *meta)
2339 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2341 if (ret_type != RET_INTEGER ||
2342 (func_id != BPF_FUNC_get_stack &&
2343 func_id != BPF_FUNC_probe_read_str))
2346 ret_reg->smax_value = meta->msize_smax_value;
2347 ret_reg->umax_value = meta->msize_umax_value;
2348 __reg_deduce_bounds(ret_reg);
2349 __reg_bound_offset(ret_reg);
2352 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2354 const struct bpf_func_proto *fn = NULL;
2355 struct bpf_reg_state *regs;
2356 struct bpf_call_arg_meta meta;
2360 /* find function prototype */
2361 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2362 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2367 if (env->ops->get_func_proto)
2368 fn = env->ops->get_func_proto(func_id, env->prog);
2370 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2375 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2376 if (!env->prog->gpl_compatible && fn->gpl_only) {
2377 verbose(env, "cannot call GPL only function from proprietary program\n");
2381 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2382 changes_data = bpf_helper_changes_pkt_data(fn->func);
2383 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2384 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2385 func_id_name(func_id), func_id);
2389 memset(&meta, 0, sizeof(meta));
2390 meta.pkt_access = fn->pkt_access;
2392 err = check_func_proto(fn);
2394 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2395 func_id_name(func_id), func_id);
2400 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2403 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2406 if (func_id == BPF_FUNC_tail_call) {
2407 if (meta.map_ptr == NULL) {
2408 verbose(env, "verifier bug\n");
2411 env->insn_aux_data[insn_idx].map_ptr = meta.map_ptr;
2413 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2416 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2419 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2423 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2424 * is inferred from register state.
2426 for (i = 0; i < meta.access_size; i++) {
2427 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2428 BPF_WRITE, -1, false);
2433 regs = cur_regs(env);
2434 /* reset caller saved regs */
2435 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2436 mark_reg_not_init(env, regs, caller_saved[i]);
2437 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2440 /* update return register (already marked as written above) */
2441 if (fn->ret_type == RET_INTEGER) {
2442 /* sets type to SCALAR_VALUE */
2443 mark_reg_unknown(env, regs, BPF_REG_0);
2444 } else if (fn->ret_type == RET_VOID) {
2445 regs[BPF_REG_0].type = NOT_INIT;
2446 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
2447 struct bpf_insn_aux_data *insn_aux;
2449 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2450 /* There is no offset yet applied, variable or fixed */
2451 mark_reg_known_zero(env, regs, BPF_REG_0);
2452 regs[BPF_REG_0].off = 0;
2453 /* remember map_ptr, so that check_map_access()
2454 * can check 'value_size' boundary of memory access
2455 * to map element returned from bpf_map_lookup_elem()
2457 if (meta.map_ptr == NULL) {
2459 "kernel subsystem misconfigured verifier\n");
2462 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2463 regs[BPF_REG_0].id = ++env->id_gen;
2464 insn_aux = &env->insn_aux_data[insn_idx];
2465 if (!insn_aux->map_ptr)
2466 insn_aux->map_ptr = meta.map_ptr;
2467 else if (insn_aux->map_ptr != meta.map_ptr)
2468 insn_aux->map_ptr = BPF_MAP_PTR_POISON;
2470 verbose(env, "unknown return type %d of func %s#%d\n",
2471 fn->ret_type, func_id_name(func_id), func_id);
2475 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2477 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2481 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2482 const char *err_str;
2484 #ifdef CONFIG_PERF_EVENTS
2485 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2486 err_str = "cannot get callchain buffer for func %s#%d\n";
2489 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2492 verbose(env, err_str, func_id_name(func_id), func_id);
2496 env->prog->has_callchain_buf = true;
2500 clear_all_pkt_pointers(env);
2504 static bool signed_add_overflows(s64 a, s64 b)
2506 /* Do the add in u64, where overflow is well-defined */
2507 s64 res = (s64)((u64)a + (u64)b);
2514 static bool signed_sub_overflows(s64 a, s64 b)
2516 /* Do the sub in u64, where overflow is well-defined */
2517 s64 res = (s64)((u64)a - (u64)b);
2524 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2525 const struct bpf_reg_state *reg,
2526 enum bpf_reg_type type)
2528 bool known = tnum_is_const(reg->var_off);
2529 s64 val = reg->var_off.value;
2530 s64 smin = reg->smin_value;
2532 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2533 verbose(env, "math between %s pointer and %lld is not allowed\n",
2534 reg_type_str[type], val);
2538 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2539 verbose(env, "%s pointer offset %d is not allowed\n",
2540 reg_type_str[type], reg->off);
2544 if (smin == S64_MIN) {
2545 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2546 reg_type_str[type]);
2550 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2551 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2552 smin, reg_type_str[type]);
2559 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2560 * Caller should also handle BPF_MOV case separately.
2561 * If we return -EACCES, caller may want to try again treating pointer as a
2562 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2564 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2565 struct bpf_insn *insn,
2566 const struct bpf_reg_state *ptr_reg,
2567 const struct bpf_reg_state *off_reg)
2569 struct bpf_verifier_state *vstate = env->cur_state;
2570 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2571 struct bpf_reg_state *regs = state->regs, *dst_reg;
2572 bool known = tnum_is_const(off_reg->var_off);
2573 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2574 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2575 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2576 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2577 u8 opcode = BPF_OP(insn->code);
2578 u32 dst = insn->dst_reg;
2580 dst_reg = ®s[dst];
2582 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2583 smin_val > smax_val || umin_val > umax_val) {
2584 /* Taint dst register if offset had invalid bounds derived from
2585 * e.g. dead branches.
2587 __mark_reg_unknown(dst_reg);
2591 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2592 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2594 "R%d 32-bit pointer arithmetic prohibited\n",
2599 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2600 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2604 if (ptr_reg->type == CONST_PTR_TO_MAP) {
2605 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2609 if (ptr_reg->type == PTR_TO_PACKET_END) {
2610 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2615 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2616 * The id may be overwritten later if we create a new variable offset.
2618 dst_reg->type = ptr_reg->type;
2619 dst_reg->id = ptr_reg->id;
2621 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
2622 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
2627 /* We can take a fixed offset as long as it doesn't overflow
2628 * the s32 'off' field
2630 if (known && (ptr_reg->off + smin_val ==
2631 (s64)(s32)(ptr_reg->off + smin_val))) {
2632 /* pointer += K. Accumulate it into fixed offset */
2633 dst_reg->smin_value = smin_ptr;
2634 dst_reg->smax_value = smax_ptr;
2635 dst_reg->umin_value = umin_ptr;
2636 dst_reg->umax_value = umax_ptr;
2637 dst_reg->var_off = ptr_reg->var_off;
2638 dst_reg->off = ptr_reg->off + smin_val;
2639 dst_reg->range = ptr_reg->range;
2642 /* A new variable offset is created. Note that off_reg->off
2643 * == 0, since it's a scalar.
2644 * dst_reg gets the pointer type and since some positive
2645 * integer value was added to the pointer, give it a new 'id'
2646 * if it's a PTR_TO_PACKET.
2647 * this creates a new 'base' pointer, off_reg (variable) gets
2648 * added into the variable offset, and we copy the fixed offset
2651 if (signed_add_overflows(smin_ptr, smin_val) ||
2652 signed_add_overflows(smax_ptr, smax_val)) {
2653 dst_reg->smin_value = S64_MIN;
2654 dst_reg->smax_value = S64_MAX;
2656 dst_reg->smin_value = smin_ptr + smin_val;
2657 dst_reg->smax_value = smax_ptr + smax_val;
2659 if (umin_ptr + umin_val < umin_ptr ||
2660 umax_ptr + umax_val < umax_ptr) {
2661 dst_reg->umin_value = 0;
2662 dst_reg->umax_value = U64_MAX;
2664 dst_reg->umin_value = umin_ptr + umin_val;
2665 dst_reg->umax_value = umax_ptr + umax_val;
2667 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
2668 dst_reg->off = ptr_reg->off;
2669 if (reg_is_pkt_pointer(ptr_reg)) {
2670 dst_reg->id = ++env->id_gen;
2671 /* something was added to pkt_ptr, set range to zero */
2676 if (dst_reg == off_reg) {
2677 /* scalar -= pointer. Creates an unknown scalar */
2678 verbose(env, "R%d tried to subtract pointer from scalar\n",
2682 /* We don't allow subtraction from FP, because (according to
2683 * test_verifier.c test "invalid fp arithmetic", JITs might not
2684 * be able to deal with it.
2686 if (ptr_reg->type == PTR_TO_STACK) {
2687 verbose(env, "R%d subtraction from stack pointer prohibited\n",
2691 if (known && (ptr_reg->off - smin_val ==
2692 (s64)(s32)(ptr_reg->off - smin_val))) {
2693 /* pointer -= K. Subtract it from fixed offset */
2694 dst_reg->smin_value = smin_ptr;
2695 dst_reg->smax_value = smax_ptr;
2696 dst_reg->umin_value = umin_ptr;
2697 dst_reg->umax_value = umax_ptr;
2698 dst_reg->var_off = ptr_reg->var_off;
2699 dst_reg->id = ptr_reg->id;
2700 dst_reg->off = ptr_reg->off - smin_val;
2701 dst_reg->range = ptr_reg->range;
2704 /* A new variable offset is created. If the subtrahend is known
2705 * nonnegative, then any reg->range we had before is still good.
2707 if (signed_sub_overflows(smin_ptr, smax_val) ||
2708 signed_sub_overflows(smax_ptr, smin_val)) {
2709 /* Overflow possible, we know nothing */
2710 dst_reg->smin_value = S64_MIN;
2711 dst_reg->smax_value = S64_MAX;
2713 dst_reg->smin_value = smin_ptr - smax_val;
2714 dst_reg->smax_value = smax_ptr - smin_val;
2716 if (umin_ptr < umax_val) {
2717 /* Overflow possible, we know nothing */
2718 dst_reg->umin_value = 0;
2719 dst_reg->umax_value = U64_MAX;
2721 /* Cannot overflow (as long as bounds are consistent) */
2722 dst_reg->umin_value = umin_ptr - umax_val;
2723 dst_reg->umax_value = umax_ptr - umin_val;
2725 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2726 dst_reg->off = ptr_reg->off;
2727 if (reg_is_pkt_pointer(ptr_reg)) {
2728 dst_reg->id = ++env->id_gen;
2729 /* something was added to pkt_ptr, set range to zero */
2737 /* bitwise ops on pointers are troublesome, prohibit. */
2738 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2739 dst, bpf_alu_string[opcode >> 4]);
2742 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2743 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2744 dst, bpf_alu_string[opcode >> 4]);
2748 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2751 __update_reg_bounds(dst_reg);
2752 __reg_deduce_bounds(dst_reg);
2753 __reg_bound_offset(dst_reg);
2757 /* WARNING: This function does calculations on 64-bit values, but the actual
2758 * execution may occur on 32-bit values. Therefore, things like bitshifts
2759 * need extra checks in the 32-bit case.
2761 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2762 struct bpf_insn *insn,
2763 struct bpf_reg_state *dst_reg,
2764 struct bpf_reg_state src_reg)
2766 struct bpf_reg_state *regs = cur_regs(env);
2767 u8 opcode = BPF_OP(insn->code);
2768 bool src_known, dst_known;
2769 s64 smin_val, smax_val;
2770 u64 umin_val, umax_val;
2771 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2773 smin_val = src_reg.smin_value;
2774 smax_val = src_reg.smax_value;
2775 umin_val = src_reg.umin_value;
2776 umax_val = src_reg.umax_value;
2777 src_known = tnum_is_const(src_reg.var_off);
2778 dst_known = tnum_is_const(dst_reg->var_off);
2780 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2781 smin_val > smax_val || umin_val > umax_val) {
2782 /* Taint dst register if offset had invalid bounds derived from
2783 * e.g. dead branches.
2785 __mark_reg_unknown(dst_reg);
2790 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2791 __mark_reg_unknown(dst_reg);
2797 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2798 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2799 dst_reg->smin_value = S64_MIN;
2800 dst_reg->smax_value = S64_MAX;
2802 dst_reg->smin_value += smin_val;
2803 dst_reg->smax_value += smax_val;
2805 if (dst_reg->umin_value + umin_val < umin_val ||
2806 dst_reg->umax_value + umax_val < umax_val) {
2807 dst_reg->umin_value = 0;
2808 dst_reg->umax_value = U64_MAX;
2810 dst_reg->umin_value += umin_val;
2811 dst_reg->umax_value += umax_val;
2813 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2816 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2817 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2818 /* Overflow possible, we know nothing */
2819 dst_reg->smin_value = S64_MIN;
2820 dst_reg->smax_value = S64_MAX;
2822 dst_reg->smin_value -= smax_val;
2823 dst_reg->smax_value -= smin_val;
2825 if (dst_reg->umin_value < umax_val) {
2826 /* Overflow possible, we know nothing */
2827 dst_reg->umin_value = 0;
2828 dst_reg->umax_value = U64_MAX;
2830 /* Cannot overflow (as long as bounds are consistent) */
2831 dst_reg->umin_value -= umax_val;
2832 dst_reg->umax_value -= umin_val;
2834 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2837 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2838 if (smin_val < 0 || dst_reg->smin_value < 0) {
2839 /* Ain't nobody got time to multiply that sign */
2840 __mark_reg_unbounded(dst_reg);
2841 __update_reg_bounds(dst_reg);
2844 /* Both values are positive, so we can work with unsigned and
2845 * copy the result to signed (unless it exceeds S64_MAX).
2847 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2848 /* Potential overflow, we know nothing */
2849 __mark_reg_unbounded(dst_reg);
2850 /* (except what we can learn from the var_off) */
2851 __update_reg_bounds(dst_reg);
2854 dst_reg->umin_value *= umin_val;
2855 dst_reg->umax_value *= umax_val;
2856 if (dst_reg->umax_value > S64_MAX) {
2857 /* Overflow possible, we know nothing */
2858 dst_reg->smin_value = S64_MIN;
2859 dst_reg->smax_value = S64_MAX;
2861 dst_reg->smin_value = dst_reg->umin_value;
2862 dst_reg->smax_value = dst_reg->umax_value;
2866 if (src_known && dst_known) {
2867 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2868 src_reg.var_off.value);
2871 /* We get our minimum from the var_off, since that's inherently
2872 * bitwise. Our maximum is the minimum of the operands' maxima.
2874 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2875 dst_reg->umin_value = dst_reg->var_off.value;
2876 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2877 if (dst_reg->smin_value < 0 || smin_val < 0) {
2878 /* Lose signed bounds when ANDing negative numbers,
2879 * ain't nobody got time for that.
2881 dst_reg->smin_value = S64_MIN;
2882 dst_reg->smax_value = S64_MAX;
2884 /* ANDing two positives gives a positive, so safe to
2885 * cast result into s64.
2887 dst_reg->smin_value = dst_reg->umin_value;
2888 dst_reg->smax_value = dst_reg->umax_value;
2890 /* We may learn something more from the var_off */
2891 __update_reg_bounds(dst_reg);
2894 if (src_known && dst_known) {
2895 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2896 src_reg.var_off.value);
2899 /* We get our maximum from the var_off, and our minimum is the
2900 * maximum of the operands' minima
2902 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2903 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2904 dst_reg->umax_value = dst_reg->var_off.value |
2905 dst_reg->var_off.mask;
2906 if (dst_reg->smin_value < 0 || smin_val < 0) {
2907 /* Lose signed bounds when ORing negative numbers,
2908 * ain't nobody got time for that.
2910 dst_reg->smin_value = S64_MIN;
2911 dst_reg->smax_value = S64_MAX;
2913 /* ORing two positives gives a positive, so safe to
2914 * cast result into s64.
2916 dst_reg->smin_value = dst_reg->umin_value;
2917 dst_reg->smax_value = dst_reg->umax_value;
2919 /* We may learn something more from the var_off */
2920 __update_reg_bounds(dst_reg);
2923 if (umax_val >= insn_bitness) {
2924 /* Shifts greater than 31 or 63 are undefined.
2925 * This includes shifts by a negative number.
2927 mark_reg_unknown(env, regs, insn->dst_reg);
2930 /* We lose all sign bit information (except what we can pick
2933 dst_reg->smin_value = S64_MIN;
2934 dst_reg->smax_value = S64_MAX;
2935 /* If we might shift our top bit out, then we know nothing */
2936 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2937 dst_reg->umin_value = 0;
2938 dst_reg->umax_value = U64_MAX;
2940 dst_reg->umin_value <<= umin_val;
2941 dst_reg->umax_value <<= umax_val;
2944 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2946 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2947 /* We may learn something more from the var_off */
2948 __update_reg_bounds(dst_reg);
2951 if (umax_val >= insn_bitness) {
2952 /* Shifts greater than 31 or 63 are undefined.
2953 * This includes shifts by a negative number.
2955 mark_reg_unknown(env, regs, insn->dst_reg);
2958 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2959 * be negative, then either:
2960 * 1) src_reg might be zero, so the sign bit of the result is
2961 * unknown, so we lose our signed bounds
2962 * 2) it's known negative, thus the unsigned bounds capture the
2964 * 3) the signed bounds cross zero, so they tell us nothing
2966 * If the value in dst_reg is known nonnegative, then again the
2967 * unsigned bounts capture the signed bounds.
2968 * Thus, in all cases it suffices to blow away our signed bounds
2969 * and rely on inferring new ones from the unsigned bounds and
2970 * var_off of the result.
2972 dst_reg->smin_value = S64_MIN;
2973 dst_reg->smax_value = S64_MAX;
2975 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
2978 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
2979 dst_reg->umin_value >>= umax_val;
2980 dst_reg->umax_value >>= umin_val;
2981 /* We may learn something more from the var_off */
2982 __update_reg_bounds(dst_reg);
2985 mark_reg_unknown(env, regs, insn->dst_reg);
2989 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2990 /* 32-bit ALU ops are (32,32)->32 */
2991 coerce_reg_to_size(dst_reg, 4);
2992 coerce_reg_to_size(&src_reg, 4);
2995 __reg_deduce_bounds(dst_reg);
2996 __reg_bound_offset(dst_reg);
3000 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3003 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3004 struct bpf_insn *insn)
3006 struct bpf_verifier_state *vstate = env->cur_state;
3007 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3008 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3009 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3010 u8 opcode = BPF_OP(insn->code);
3012 dst_reg = ®s[insn->dst_reg];
3014 if (dst_reg->type != SCALAR_VALUE)
3016 if (BPF_SRC(insn->code) == BPF_X) {
3017 src_reg = ®s[insn->src_reg];
3018 if (src_reg->type != SCALAR_VALUE) {
3019 if (dst_reg->type != SCALAR_VALUE) {
3020 /* Combining two pointers by any ALU op yields
3021 * an arbitrary scalar. Disallow all math except
3022 * pointer subtraction
3024 if (opcode == BPF_SUB){
3025 mark_reg_unknown(env, regs, insn->dst_reg);
3028 verbose(env, "R%d pointer %s pointer prohibited\n",
3030 bpf_alu_string[opcode >> 4]);
3033 /* scalar += pointer
3034 * This is legal, but we have to reverse our
3035 * src/dest handling in computing the range
3037 return adjust_ptr_min_max_vals(env, insn,
3040 } else if (ptr_reg) {
3041 /* pointer += scalar */
3042 return adjust_ptr_min_max_vals(env, insn,
3046 /* Pretend the src is a reg with a known value, since we only
3047 * need to be able to read from this state.
3049 off_reg.type = SCALAR_VALUE;
3050 __mark_reg_known(&off_reg, insn->imm);
3052 if (ptr_reg) /* pointer += K */
3053 return adjust_ptr_min_max_vals(env, insn,
3057 /* Got here implies adding two SCALAR_VALUEs */
3058 if (WARN_ON_ONCE(ptr_reg)) {
3059 print_verifier_state(env, state);
3060 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3063 if (WARN_ON(!src_reg)) {
3064 print_verifier_state(env, state);
3065 verbose(env, "verifier internal error: no src_reg\n");
3068 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3071 /* check validity of 32-bit and 64-bit arithmetic operations */
3072 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3074 struct bpf_reg_state *regs = cur_regs(env);
3075 u8 opcode = BPF_OP(insn->code);
3078 if (opcode == BPF_END || opcode == BPF_NEG) {
3079 if (opcode == BPF_NEG) {
3080 if (BPF_SRC(insn->code) != 0 ||
3081 insn->src_reg != BPF_REG_0 ||
3082 insn->off != 0 || insn->imm != 0) {
3083 verbose(env, "BPF_NEG uses reserved fields\n");
3087 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3088 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3089 BPF_CLASS(insn->code) == BPF_ALU64) {
3090 verbose(env, "BPF_END uses reserved fields\n");
3095 /* check src operand */
3096 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3100 if (is_pointer_value(env, insn->dst_reg)) {
3101 verbose(env, "R%d pointer arithmetic prohibited\n",
3106 /* check dest operand */
3107 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3111 } else if (opcode == BPF_MOV) {
3113 if (BPF_SRC(insn->code) == BPF_X) {
3114 if (insn->imm != 0 || insn->off != 0) {
3115 verbose(env, "BPF_MOV uses reserved fields\n");
3119 /* check src operand */
3120 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3124 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3125 verbose(env, "BPF_MOV uses reserved fields\n");
3130 /* check dest operand */
3131 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3135 if (BPF_SRC(insn->code) == BPF_X) {
3136 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3138 * copy register state to dest reg
3140 regs[insn->dst_reg] = regs[insn->src_reg];
3141 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3144 if (is_pointer_value(env, insn->src_reg)) {
3146 "R%d partial copy of pointer\n",
3150 mark_reg_unknown(env, regs, insn->dst_reg);
3151 coerce_reg_to_size(®s[insn->dst_reg], 4);
3155 * remember the value we stored into this reg
3157 regs[insn->dst_reg].type = SCALAR_VALUE;
3158 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3159 __mark_reg_known(regs + insn->dst_reg,
3162 __mark_reg_known(regs + insn->dst_reg,
3167 } else if (opcode > BPF_END) {
3168 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3171 } else { /* all other ALU ops: and, sub, xor, add, ... */
3173 if (BPF_SRC(insn->code) == BPF_X) {
3174 if (insn->imm != 0 || insn->off != 0) {
3175 verbose(env, "BPF_ALU uses reserved fields\n");
3178 /* check src1 operand */
3179 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3183 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3184 verbose(env, "BPF_ALU uses reserved fields\n");
3189 /* check src2 operand */
3190 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3194 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3195 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3196 verbose(env, "div by zero\n");
3200 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3201 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3205 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3206 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3207 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3209 if (insn->imm < 0 || insn->imm >= size) {
3210 verbose(env, "invalid shift %d\n", insn->imm);
3215 /* check dest operand */
3216 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3220 return adjust_reg_min_max_vals(env, insn);
3226 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3227 struct bpf_reg_state *dst_reg,
3228 enum bpf_reg_type type,
3229 bool range_right_open)
3231 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3232 struct bpf_reg_state *regs = state->regs, *reg;
3236 if (dst_reg->off < 0 ||
3237 (dst_reg->off == 0 && range_right_open))
3238 /* This doesn't give us any range */
3241 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3242 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3243 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3244 * than pkt_end, but that's because it's also less than pkt.
3248 new_range = dst_reg->off;
3249 if (range_right_open)
3252 /* Examples for register markings:
3254 * pkt_data in dst register:
3258 * if (r2 > pkt_end) goto <handle exception>
3263 * if (r2 < pkt_end) goto <access okay>
3264 * <handle exception>
3267 * r2 == dst_reg, pkt_end == src_reg
3268 * r2=pkt(id=n,off=8,r=0)
3269 * r3=pkt(id=n,off=0,r=0)
3271 * pkt_data in src register:
3275 * if (pkt_end >= r2) goto <access okay>
3276 * <handle exception>
3280 * if (pkt_end <= r2) goto <handle exception>
3284 * pkt_end == dst_reg, r2 == src_reg
3285 * r2=pkt(id=n,off=8,r=0)
3286 * r3=pkt(id=n,off=0,r=0)
3288 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3289 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3290 * and [r3, r3 + 8-1) respectively is safe to access depending on
3294 /* If our ids match, then we must have the same max_value. And we
3295 * don't care about the other reg's fixed offset, since if it's too big
3296 * the range won't allow anything.
3297 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3299 for (i = 0; i < MAX_BPF_REG; i++)
3300 if (regs[i].type == type && regs[i].id == dst_reg->id)
3301 /* keep the maximum range already checked */
3302 regs[i].range = max(regs[i].range, new_range);
3304 for (j = 0; j <= vstate->curframe; j++) {
3305 state = vstate->frame[j];
3306 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3307 if (state->stack[i].slot_type[0] != STACK_SPILL)
3309 reg = &state->stack[i].spilled_ptr;
3310 if (reg->type == type && reg->id == dst_reg->id)
3311 reg->range = max(reg->range, new_range);
3316 /* Adjusts the register min/max values in the case that the dst_reg is the
3317 * variable register that we are working on, and src_reg is a constant or we're
3318 * simply doing a BPF_K check.
3319 * In JEQ/JNE cases we also adjust the var_off values.
3321 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3322 struct bpf_reg_state *false_reg, u64 val,
3325 /* If the dst_reg is a pointer, we can't learn anything about its
3326 * variable offset from the compare (unless src_reg were a pointer into
3327 * the same object, but we don't bother with that.
3328 * Since false_reg and true_reg have the same type by construction, we
3329 * only need to check one of them for pointerness.
3331 if (__is_pointer_value(false, false_reg))
3336 /* If this is false then we know nothing Jon Snow, but if it is
3337 * true then we know for sure.
3339 __mark_reg_known(true_reg, val);
3342 /* If this is true we know nothing Jon Snow, but if it is false
3343 * we know the value for sure;
3345 __mark_reg_known(false_reg, val);
3348 false_reg->umax_value = min(false_reg->umax_value, val);
3349 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3352 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3353 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3356 false_reg->umin_value = max(false_reg->umin_value, val);
3357 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3360 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3361 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3364 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3365 true_reg->umin_value = max(true_reg->umin_value, val);
3368 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3369 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3372 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3373 true_reg->umax_value = min(true_reg->umax_value, val);
3376 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3377 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3383 __reg_deduce_bounds(false_reg);
3384 __reg_deduce_bounds(true_reg);
3385 /* We might have learned some bits from the bounds. */
3386 __reg_bound_offset(false_reg);
3387 __reg_bound_offset(true_reg);
3388 /* Intersecting with the old var_off might have improved our bounds
3389 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3390 * then new var_off is (0; 0x7f...fc) which improves our umax.
3392 __update_reg_bounds(false_reg);
3393 __update_reg_bounds(true_reg);
3396 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3399 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3400 struct bpf_reg_state *false_reg, u64 val,
3403 if (__is_pointer_value(false, false_reg))
3408 /* If this is false then we know nothing Jon Snow, but if it is
3409 * true then we know for sure.
3411 __mark_reg_known(true_reg, val);
3414 /* If this is true we know nothing Jon Snow, but if it is false
3415 * we know the value for sure;
3417 __mark_reg_known(false_reg, val);
3420 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3421 false_reg->umin_value = max(false_reg->umin_value, val);
3424 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3425 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3428 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3429 false_reg->umax_value = min(false_reg->umax_value, val);
3432 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3433 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3436 true_reg->umax_value = min(true_reg->umax_value, val);
3437 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3440 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3441 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3444 true_reg->umin_value = max(true_reg->umin_value, val);
3445 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3448 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3449 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3455 __reg_deduce_bounds(false_reg);
3456 __reg_deduce_bounds(true_reg);
3457 /* We might have learned some bits from the bounds. */
3458 __reg_bound_offset(false_reg);
3459 __reg_bound_offset(true_reg);
3460 /* Intersecting with the old var_off might have improved our bounds
3461 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3462 * then new var_off is (0; 0x7f...fc) which improves our umax.
3464 __update_reg_bounds(false_reg);
3465 __update_reg_bounds(true_reg);
3468 /* Regs are known to be equal, so intersect their min/max/var_off */
3469 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3470 struct bpf_reg_state *dst_reg)
3472 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3473 dst_reg->umin_value);
3474 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3475 dst_reg->umax_value);
3476 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3477 dst_reg->smin_value);
3478 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3479 dst_reg->smax_value);
3480 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3482 /* We might have learned new bounds from the var_off. */
3483 __update_reg_bounds(src_reg);
3484 __update_reg_bounds(dst_reg);
3485 /* We might have learned something about the sign bit. */
3486 __reg_deduce_bounds(src_reg);
3487 __reg_deduce_bounds(dst_reg);
3488 /* We might have learned some bits from the bounds. */
3489 __reg_bound_offset(src_reg);
3490 __reg_bound_offset(dst_reg);
3491 /* Intersecting with the old var_off might have improved our bounds
3492 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3493 * then new var_off is (0; 0x7f...fc) which improves our umax.
3495 __update_reg_bounds(src_reg);
3496 __update_reg_bounds(dst_reg);
3499 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3500 struct bpf_reg_state *true_dst,
3501 struct bpf_reg_state *false_src,
3502 struct bpf_reg_state *false_dst,
3507 __reg_combine_min_max(true_src, true_dst);
3510 __reg_combine_min_max(false_src, false_dst);
3515 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
3518 struct bpf_reg_state *reg = ®s[regno];
3520 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
3521 /* Old offset (both fixed and variable parts) should
3522 * have been known-zero, because we don't allow pointer
3523 * arithmetic on pointers that might be NULL.
3525 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3526 !tnum_equals_const(reg->var_off, 0) ||
3528 __mark_reg_known_zero(reg);
3532 reg->type = SCALAR_VALUE;
3533 } else if (reg->map_ptr->inner_map_meta) {
3534 reg->type = CONST_PTR_TO_MAP;
3535 reg->map_ptr = reg->map_ptr->inner_map_meta;
3537 reg->type = PTR_TO_MAP_VALUE;
3539 /* We don't need id from this point onwards anymore, thus we
3540 * should better reset it, so that state pruning has chances
3547 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3548 * be folded together at some point.
3550 static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno,
3553 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3554 struct bpf_reg_state *regs = state->regs;
3555 u32 id = regs[regno].id;
3558 for (i = 0; i < MAX_BPF_REG; i++)
3559 mark_map_reg(regs, i, id, is_null);
3561 for (j = 0; j <= vstate->curframe; j++) {
3562 state = vstate->frame[j];
3563 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3564 if (state->stack[i].slot_type[0] != STACK_SPILL)
3566 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
3571 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3572 struct bpf_reg_state *dst_reg,
3573 struct bpf_reg_state *src_reg,
3574 struct bpf_verifier_state *this_branch,
3575 struct bpf_verifier_state *other_branch)
3577 if (BPF_SRC(insn->code) != BPF_X)
3580 switch (BPF_OP(insn->code)) {
3582 if ((dst_reg->type == PTR_TO_PACKET &&
3583 src_reg->type == PTR_TO_PACKET_END) ||
3584 (dst_reg->type == PTR_TO_PACKET_META &&
3585 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3586 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3587 find_good_pkt_pointers(this_branch, dst_reg,
3588 dst_reg->type, false);
3589 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3590 src_reg->type == PTR_TO_PACKET) ||
3591 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3592 src_reg->type == PTR_TO_PACKET_META)) {
3593 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3594 find_good_pkt_pointers(other_branch, src_reg,
3595 src_reg->type, true);
3601 if ((dst_reg->type == PTR_TO_PACKET &&
3602 src_reg->type == PTR_TO_PACKET_END) ||
3603 (dst_reg->type == PTR_TO_PACKET_META &&
3604 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3605 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3606 find_good_pkt_pointers(other_branch, dst_reg,
3607 dst_reg->type, true);
3608 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3609 src_reg->type == PTR_TO_PACKET) ||
3610 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3611 src_reg->type == PTR_TO_PACKET_META)) {
3612 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3613 find_good_pkt_pointers(this_branch, src_reg,
3614 src_reg->type, false);
3620 if ((dst_reg->type == PTR_TO_PACKET &&
3621 src_reg->type == PTR_TO_PACKET_END) ||
3622 (dst_reg->type == PTR_TO_PACKET_META &&
3623 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3624 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3625 find_good_pkt_pointers(this_branch, dst_reg,
3626 dst_reg->type, true);
3627 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3628 src_reg->type == PTR_TO_PACKET) ||
3629 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3630 src_reg->type == PTR_TO_PACKET_META)) {
3631 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3632 find_good_pkt_pointers(other_branch, src_reg,
3633 src_reg->type, false);
3639 if ((dst_reg->type == PTR_TO_PACKET &&
3640 src_reg->type == PTR_TO_PACKET_END) ||
3641 (dst_reg->type == PTR_TO_PACKET_META &&
3642 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3643 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3644 find_good_pkt_pointers(other_branch, dst_reg,
3645 dst_reg->type, false);
3646 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3647 src_reg->type == PTR_TO_PACKET) ||
3648 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3649 src_reg->type == PTR_TO_PACKET_META)) {
3650 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3651 find_good_pkt_pointers(this_branch, src_reg,
3652 src_reg->type, true);
3664 static int check_cond_jmp_op(struct bpf_verifier_env *env,
3665 struct bpf_insn *insn, int *insn_idx)
3667 struct bpf_verifier_state *this_branch = env->cur_state;
3668 struct bpf_verifier_state *other_branch;
3669 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
3670 struct bpf_reg_state *dst_reg, *other_branch_regs;
3671 u8 opcode = BPF_OP(insn->code);
3674 if (opcode > BPF_JSLE) {
3675 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
3679 if (BPF_SRC(insn->code) == BPF_X) {
3680 if (insn->imm != 0) {
3681 verbose(env, "BPF_JMP uses reserved fields\n");
3685 /* check src1 operand */
3686 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3690 if (is_pointer_value(env, insn->src_reg)) {
3691 verbose(env, "R%d pointer comparison prohibited\n",
3696 if (insn->src_reg != BPF_REG_0) {
3697 verbose(env, "BPF_JMP uses reserved fields\n");
3702 /* check src2 operand */
3703 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3707 dst_reg = ®s[insn->dst_reg];
3709 /* detect if R == 0 where R was initialized to zero earlier */
3710 if (BPF_SRC(insn->code) == BPF_K &&
3711 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3712 dst_reg->type == SCALAR_VALUE &&
3713 tnum_is_const(dst_reg->var_off)) {
3714 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
3715 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
3716 /* if (imm == imm) goto pc+off;
3717 * only follow the goto, ignore fall-through
3719 *insn_idx += insn->off;
3722 /* if (imm != imm) goto pc+off;
3723 * only follow fall-through branch, since
3724 * that's where the program will go
3730 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3733 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
3735 /* detect if we are comparing against a constant value so we can adjust
3736 * our min/max values for our dst register.
3737 * this is only legit if both are scalars (or pointers to the same
3738 * object, I suppose, but we don't support that right now), because
3739 * otherwise the different base pointers mean the offsets aren't
3742 if (BPF_SRC(insn->code) == BPF_X) {
3743 if (dst_reg->type == SCALAR_VALUE &&
3744 regs[insn->src_reg].type == SCALAR_VALUE) {
3745 if (tnum_is_const(regs[insn->src_reg].var_off))
3746 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3747 dst_reg, regs[insn->src_reg].var_off.value,
3749 else if (tnum_is_const(dst_reg->var_off))
3750 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
3751 ®s[insn->src_reg],
3752 dst_reg->var_off.value, opcode);
3753 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3754 /* Comparing for equality, we can combine knowledge */
3755 reg_combine_min_max(&other_branch_regs[insn->src_reg],
3756 &other_branch_regs[insn->dst_reg],
3757 ®s[insn->src_reg],
3758 ®s[insn->dst_reg], opcode);
3760 } else if (dst_reg->type == SCALAR_VALUE) {
3761 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3762 dst_reg, insn->imm, opcode);
3765 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3766 if (BPF_SRC(insn->code) == BPF_K &&
3767 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3768 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3769 /* Mark all identical map registers in each branch as either
3770 * safe or unknown depending R == 0 or R != 0 conditional.
3772 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3773 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3774 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3775 this_branch, other_branch) &&
3776 is_pointer_value(env, insn->dst_reg)) {
3777 verbose(env, "R%d pointer comparison prohibited\n",
3782 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
3786 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
3787 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3789 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3791 return (struct bpf_map *) (unsigned long) imm64;
3794 /* verify BPF_LD_IMM64 instruction */
3795 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3797 struct bpf_reg_state *regs = cur_regs(env);
3800 if (BPF_SIZE(insn->code) != BPF_DW) {
3801 verbose(env, "invalid BPF_LD_IMM insn\n");
3804 if (insn->off != 0) {
3805 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3809 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3813 if (insn->src_reg == 0) {
3814 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3816 regs[insn->dst_reg].type = SCALAR_VALUE;
3817 __mark_reg_known(®s[insn->dst_reg], imm);
3821 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3822 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3824 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3825 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3829 static bool may_access_skb(enum bpf_prog_type type)
3832 case BPF_PROG_TYPE_SOCKET_FILTER:
3833 case BPF_PROG_TYPE_SCHED_CLS:
3834 case BPF_PROG_TYPE_SCHED_ACT:
3841 /* verify safety of LD_ABS|LD_IND instructions:
3842 * - they can only appear in the programs where ctx == skb
3843 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3844 * preserve R6-R9, and store return value into R0
3847 * ctx == skb == R6 == CTX
3850 * SRC == any register
3851 * IMM == 32-bit immediate
3854 * R0 - 8/16/32-bit skb data converted to cpu endianness
3856 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3858 struct bpf_reg_state *regs = cur_regs(env);
3859 u8 mode = BPF_MODE(insn->code);
3862 if (!may_access_skb(env->prog->type)) {
3863 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3867 if (env->subprog_cnt) {
3868 /* when program has LD_ABS insn JITs and interpreter assume
3869 * that r1 == ctx == skb which is not the case for callees
3870 * that can have arbitrary arguments. It's problematic
3871 * for main prog as well since JITs would need to analyze
3872 * all functions in order to make proper register save/restore
3873 * decisions in the main prog. Hence disallow LD_ABS with calls
3875 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
3879 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3880 BPF_SIZE(insn->code) == BPF_DW ||
3881 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3882 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3886 /* check whether implicit source operand (register R6) is readable */
3887 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3891 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3893 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3897 if (mode == BPF_IND) {
3898 /* check explicit source operand */
3899 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3904 /* reset caller saved regs to unreadable */
3905 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3906 mark_reg_not_init(env, regs, caller_saved[i]);
3907 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3910 /* mark destination R0 register as readable, since it contains
3911 * the value fetched from the packet.
3912 * Already marked as written above.
3914 mark_reg_unknown(env, regs, BPF_REG_0);
3918 static int check_return_code(struct bpf_verifier_env *env)
3920 struct bpf_reg_state *reg;
3921 struct tnum range = tnum_range(0, 1);
3923 switch (env->prog->type) {
3924 case BPF_PROG_TYPE_CGROUP_SKB:
3925 case BPF_PROG_TYPE_CGROUP_SOCK:
3926 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
3927 case BPF_PROG_TYPE_SOCK_OPS:
3928 case BPF_PROG_TYPE_CGROUP_DEVICE:
3934 reg = cur_regs(env) + BPF_REG_0;
3935 if (reg->type != SCALAR_VALUE) {
3936 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
3937 reg_type_str[reg->type]);
3941 if (!tnum_in(range, reg->var_off)) {
3942 verbose(env, "At program exit the register R0 ");
3943 if (!tnum_is_unknown(reg->var_off)) {
3946 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3947 verbose(env, "has value %s", tn_buf);
3949 verbose(env, "has unknown scalar value");
3951 verbose(env, " should have been 0 or 1\n");
3957 /* non-recursive DFS pseudo code
3958 * 1 procedure DFS-iterative(G,v):
3959 * 2 label v as discovered
3960 * 3 let S be a stack
3962 * 5 while S is not empty
3964 * 7 if t is what we're looking for:
3966 * 9 for all edges e in G.adjacentEdges(t) do
3967 * 10 if edge e is already labelled
3968 * 11 continue with the next edge
3969 * 12 w <- G.adjacentVertex(t,e)
3970 * 13 if vertex w is not discovered and not explored
3971 * 14 label e as tree-edge
3972 * 15 label w as discovered
3975 * 18 else if vertex w is discovered
3976 * 19 label e as back-edge
3978 * 21 // vertex w is explored
3979 * 22 label e as forward- or cross-edge
3980 * 23 label t as explored
3985 * 0x11 - discovered and fall-through edge labelled
3986 * 0x12 - discovered and fall-through and branch edges labelled
3997 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
3999 static int *insn_stack; /* stack of insns to process */
4000 static int cur_stack; /* current stack index */
4001 static int *insn_state;
4003 /* t, w, e - match pseudo-code above:
4004 * t - index of current instruction
4005 * w - next instruction
4008 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4010 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4013 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4016 if (w < 0 || w >= env->prog->len) {
4017 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4022 /* mark branch target for state pruning */
4023 env->explored_states[w] = STATE_LIST_MARK;
4025 if (insn_state[w] == 0) {
4027 insn_state[t] = DISCOVERED | e;
4028 insn_state[w] = DISCOVERED;
4029 if (cur_stack >= env->prog->len)
4031 insn_stack[cur_stack++] = w;
4033 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4034 verbose(env, "back-edge from insn %d to %d\n", t, w);
4036 } else if (insn_state[w] == EXPLORED) {
4037 /* forward- or cross-edge */
4038 insn_state[t] = DISCOVERED | e;
4040 verbose(env, "insn state internal bug\n");
4046 /* non-recursive depth-first-search to detect loops in BPF program
4047 * loop == back-edge in directed graph
4049 static int check_cfg(struct bpf_verifier_env *env)
4051 struct bpf_insn *insns = env->prog->insnsi;
4052 int insn_cnt = env->prog->len;
4056 ret = check_subprogs(env);
4060 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4064 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4070 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4071 insn_stack[0] = 0; /* 0 is the first instruction */
4077 t = insn_stack[cur_stack - 1];
4079 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4080 u8 opcode = BPF_OP(insns[t].code);
4082 if (opcode == BPF_EXIT) {
4084 } else if (opcode == BPF_CALL) {
4085 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4090 if (t + 1 < insn_cnt)
4091 env->explored_states[t + 1] = STATE_LIST_MARK;
4092 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4093 env->explored_states[t] = STATE_LIST_MARK;
4094 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4100 } else if (opcode == BPF_JA) {
4101 if (BPF_SRC(insns[t].code) != BPF_K) {
4105 /* unconditional jump with single edge */
4106 ret = push_insn(t, t + insns[t].off + 1,
4112 /* tell verifier to check for equivalent states
4113 * after every call and jump
4115 if (t + 1 < insn_cnt)
4116 env->explored_states[t + 1] = STATE_LIST_MARK;
4118 /* conditional jump with two edges */
4119 env->explored_states[t] = STATE_LIST_MARK;
4120 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4126 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4133 /* all other non-branch instructions with single
4136 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4144 insn_state[t] = EXPLORED;
4145 if (cur_stack-- <= 0) {
4146 verbose(env, "pop stack internal bug\n");
4153 for (i = 0; i < insn_cnt; i++) {
4154 if (insn_state[i] != EXPLORED) {
4155 verbose(env, "unreachable insn %d\n", i);
4160 ret = 0; /* cfg looks good */
4168 /* check %cur's range satisfies %old's */
4169 static bool range_within(struct bpf_reg_state *old,
4170 struct bpf_reg_state *cur)
4172 return old->umin_value <= cur->umin_value &&
4173 old->umax_value >= cur->umax_value &&
4174 old->smin_value <= cur->smin_value &&
4175 old->smax_value >= cur->smax_value;
4178 /* Maximum number of register states that can exist at once */
4179 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4185 /* If in the old state two registers had the same id, then they need to have
4186 * the same id in the new state as well. But that id could be different from
4187 * the old state, so we need to track the mapping from old to new ids.
4188 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4189 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4190 * regs with a different old id could still have new id 9, we don't care about
4192 * So we look through our idmap to see if this old id has been seen before. If
4193 * so, we require the new id to match; otherwise, we add the id pair to the map.
4195 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4199 for (i = 0; i < ID_MAP_SIZE; i++) {
4200 if (!idmap[i].old) {
4201 /* Reached an empty slot; haven't seen this id before */
4202 idmap[i].old = old_id;
4203 idmap[i].cur = cur_id;
4206 if (idmap[i].old == old_id)
4207 return idmap[i].cur == cur_id;
4209 /* We ran out of idmap slots, which should be impossible */
4214 /* Returns true if (rold safe implies rcur safe) */
4215 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4216 struct idpair *idmap)
4220 if (!(rold->live & REG_LIVE_READ))
4221 /* explored state didn't use this */
4224 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, frameno)) == 0;
4226 if (rold->type == PTR_TO_STACK)
4227 /* two stack pointers are equal only if they're pointing to
4228 * the same stack frame, since fp-8 in foo != fp-8 in bar
4230 return equal && rold->frameno == rcur->frameno;
4235 if (rold->type == NOT_INIT)
4236 /* explored state can't have used this */
4238 if (rcur->type == NOT_INIT)
4240 switch (rold->type) {
4242 if (rcur->type == SCALAR_VALUE) {
4243 /* new val must satisfy old val knowledge */
4244 return range_within(rold, rcur) &&
4245 tnum_in(rold->var_off, rcur->var_off);
4247 /* We're trying to use a pointer in place of a scalar.
4248 * Even if the scalar was unbounded, this could lead to
4249 * pointer leaks because scalars are allowed to leak
4250 * while pointers are not. We could make this safe in
4251 * special cases if root is calling us, but it's
4252 * probably not worth the hassle.
4256 case PTR_TO_MAP_VALUE:
4257 /* If the new min/max/var_off satisfy the old ones and
4258 * everything else matches, we are OK.
4259 * We don't care about the 'id' value, because nothing
4260 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4262 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4263 range_within(rold, rcur) &&
4264 tnum_in(rold->var_off, rcur->var_off);
4265 case PTR_TO_MAP_VALUE_OR_NULL:
4266 /* a PTR_TO_MAP_VALUE could be safe to use as a
4267 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4268 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4269 * checked, doing so could have affected others with the same
4270 * id, and we can't check for that because we lost the id when
4271 * we converted to a PTR_TO_MAP_VALUE.
4273 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4275 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4277 /* Check our ids match any regs they're supposed to */
4278 return check_ids(rold->id, rcur->id, idmap);
4279 case PTR_TO_PACKET_META:
4281 if (rcur->type != rold->type)
4283 /* We must have at least as much range as the old ptr
4284 * did, so that any accesses which were safe before are
4285 * still safe. This is true even if old range < old off,
4286 * since someone could have accessed through (ptr - k), or
4287 * even done ptr -= k in a register, to get a safe access.
4289 if (rold->range > rcur->range)
4291 /* If the offsets don't match, we can't trust our alignment;
4292 * nor can we be sure that we won't fall out of range.
4294 if (rold->off != rcur->off)
4296 /* id relations must be preserved */
4297 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4299 /* new val must satisfy old val knowledge */
4300 return range_within(rold, rcur) &&
4301 tnum_in(rold->var_off, rcur->var_off);
4303 case CONST_PTR_TO_MAP:
4304 case PTR_TO_PACKET_END:
4305 /* Only valid matches are exact, which memcmp() above
4306 * would have accepted
4309 /* Don't know what's going on, just say it's not safe */
4313 /* Shouldn't get here; if we do, say it's not safe */
4318 static bool stacksafe(struct bpf_func_state *old,
4319 struct bpf_func_state *cur,
4320 struct idpair *idmap)
4324 /* if explored stack has more populated slots than current stack
4325 * such stacks are not equivalent
4327 if (old->allocated_stack > cur->allocated_stack)
4330 /* walk slots of the explored stack and ignore any additional
4331 * slots in the current stack, since explored(safe) state
4334 for (i = 0; i < old->allocated_stack; i++) {
4335 spi = i / BPF_REG_SIZE;
4337 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4338 /* explored state didn't use this */
4341 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4343 /* if old state was safe with misc data in the stack
4344 * it will be safe with zero-initialized stack.
4345 * The opposite is not true
4347 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4348 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4350 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4351 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4352 /* Ex: old explored (safe) state has STACK_SPILL in
4353 * this stack slot, but current has has STACK_MISC ->
4354 * this verifier states are not equivalent,
4355 * return false to continue verification of this path
4358 if (i % BPF_REG_SIZE)
4360 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4362 if (!regsafe(&old->stack[spi].spilled_ptr,
4363 &cur->stack[spi].spilled_ptr,
4365 /* when explored and current stack slot are both storing
4366 * spilled registers, check that stored pointers types
4367 * are the same as well.
4368 * Ex: explored safe path could have stored
4369 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4370 * but current path has stored:
4371 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4372 * such verifier states are not equivalent.
4373 * return false to continue verification of this path
4380 /* compare two verifier states
4382 * all states stored in state_list are known to be valid, since
4383 * verifier reached 'bpf_exit' instruction through them
4385 * this function is called when verifier exploring different branches of
4386 * execution popped from the state stack. If it sees an old state that has
4387 * more strict register state and more strict stack state then this execution
4388 * branch doesn't need to be explored further, since verifier already
4389 * concluded that more strict state leads to valid finish.
4391 * Therefore two states are equivalent if register state is more conservative
4392 * and explored stack state is more conservative than the current one.
4395 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4396 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4398 * In other words if current stack state (one being explored) has more
4399 * valid slots than old one that already passed validation, it means
4400 * the verifier can stop exploring and conclude that current state is valid too
4402 * Similarly with registers. If explored state has register type as invalid
4403 * whereas register type in current state is meaningful, it means that
4404 * the current state will reach 'bpf_exit' instruction safely
4406 static bool func_states_equal(struct bpf_func_state *old,
4407 struct bpf_func_state *cur)
4409 struct idpair *idmap;
4413 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4414 /* If we failed to allocate the idmap, just say it's not safe */
4418 for (i = 0; i < MAX_BPF_REG; i++) {
4419 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4423 if (!stacksafe(old, cur, idmap))
4431 static bool states_equal(struct bpf_verifier_env *env,
4432 struct bpf_verifier_state *old,
4433 struct bpf_verifier_state *cur)
4437 if (old->curframe != cur->curframe)
4440 /* for states to be equal callsites have to be the same
4441 * and all frame states need to be equivalent
4443 for (i = 0; i <= old->curframe; i++) {
4444 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4446 if (!func_states_equal(old->frame[i], cur->frame[i]))
4452 /* A write screens off any subsequent reads; but write marks come from the
4453 * straight-line code between a state and its parent. When we arrive at an
4454 * equivalent state (jump target or such) we didn't arrive by the straight-line
4455 * code, so read marks in the state must propagate to the parent regardless
4456 * of the state's write marks. That's what 'parent == state->parent' comparison
4457 * in mark_reg_read() and mark_stack_slot_read() is for.
4459 static int propagate_liveness(struct bpf_verifier_env *env,
4460 const struct bpf_verifier_state *vstate,
4461 struct bpf_verifier_state *vparent)
4463 int i, frame, err = 0;
4464 struct bpf_func_state *state, *parent;
4466 if (vparent->curframe != vstate->curframe) {
4467 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4468 vparent->curframe, vstate->curframe);
4471 /* Propagate read liveness of registers... */
4472 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4473 /* We don't need to worry about FP liveness because it's read-only */
4474 for (i = 0; i < BPF_REG_FP; i++) {
4475 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4477 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4478 err = mark_reg_read(env, vstate, vparent, i);
4484 /* ... and stack slots */
4485 for (frame = 0; frame <= vstate->curframe; frame++) {
4486 state = vstate->frame[frame];
4487 parent = vparent->frame[frame];
4488 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4489 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4490 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4492 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4493 mark_stack_slot_read(env, vstate, vparent, i, frame);
4499 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4501 struct bpf_verifier_state_list *new_sl;
4502 struct bpf_verifier_state_list *sl;
4503 struct bpf_verifier_state *cur = env->cur_state;
4506 sl = env->explored_states[insn_idx];
4508 /* this 'insn_idx' instruction wasn't marked, so we will not
4509 * be doing state search here
4513 while (sl != STATE_LIST_MARK) {
4514 if (states_equal(env, &sl->state, cur)) {
4515 /* reached equivalent register/stack state,
4517 * Registers read by the continuation are read by us.
4518 * If we have any write marks in env->cur_state, they
4519 * will prevent corresponding reads in the continuation
4520 * from reaching our parent (an explored_state). Our
4521 * own state will get the read marks recorded, but
4522 * they'll be immediately forgotten as we're pruning
4523 * this state and will pop a new one.
4525 err = propagate_liveness(env, &sl->state, cur);
4533 /* there were no equivalent states, remember current one.
4534 * technically the current state is not proven to be safe yet,
4535 * but it will either reach outer most bpf_exit (which means it's safe)
4536 * or it will be rejected. Since there are no loops, we won't be
4537 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4538 * again on the way to bpf_exit
4540 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
4544 /* add new state to the head of linked list */
4545 err = copy_verifier_state(&new_sl->state, cur);
4547 free_verifier_state(&new_sl->state, false);
4551 new_sl->next = env->explored_states[insn_idx];
4552 env->explored_states[insn_idx] = new_sl;
4553 /* connect new state to parentage chain */
4554 cur->parent = &new_sl->state;
4555 /* clear write marks in current state: the writes we did are not writes
4556 * our child did, so they don't screen off its reads from us.
4557 * (There are no read marks in current state, because reads always mark
4558 * their parent and current state never has children yet. Only
4559 * explored_states can get read marks.)
4561 for (i = 0; i < BPF_REG_FP; i++)
4562 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
4564 /* all stack frames are accessible from callee, clear them all */
4565 for (j = 0; j <= cur->curframe; j++) {
4566 struct bpf_func_state *frame = cur->frame[j];
4568 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++)
4569 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
4574 static int do_check(struct bpf_verifier_env *env)
4576 struct bpf_verifier_state *state;
4577 struct bpf_insn *insns = env->prog->insnsi;
4578 struct bpf_reg_state *regs;
4579 int insn_cnt = env->prog->len, i;
4580 int insn_idx, prev_insn_idx = 0;
4581 int insn_processed = 0;
4582 bool do_print_state = false;
4584 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
4587 state->curframe = 0;
4588 state->parent = NULL;
4589 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
4590 if (!state->frame[0]) {
4594 env->cur_state = state;
4595 init_func_state(env, state->frame[0],
4596 BPF_MAIN_FUNC /* callsite */,
4598 0 /* subprogno, zero == main subprog */);
4601 struct bpf_insn *insn;
4605 if (insn_idx >= insn_cnt) {
4606 verbose(env, "invalid insn idx %d insn_cnt %d\n",
4607 insn_idx, insn_cnt);
4611 insn = &insns[insn_idx];
4612 class = BPF_CLASS(insn->code);
4614 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
4616 "BPF program is too large. Processed %d insn\n",
4621 err = is_state_visited(env, insn_idx);
4625 /* found equivalent state, can prune the search */
4626 if (env->log.level) {
4628 verbose(env, "\nfrom %d to %d: safe\n",
4629 prev_insn_idx, insn_idx);
4631 verbose(env, "%d: safe\n", insn_idx);
4633 goto process_bpf_exit;
4639 if (env->log.level > 1 || (env->log.level && do_print_state)) {
4640 if (env->log.level > 1)
4641 verbose(env, "%d:", insn_idx);
4643 verbose(env, "\nfrom %d to %d:",
4644 prev_insn_idx, insn_idx);
4645 print_verifier_state(env, state->frame[state->curframe]);
4646 do_print_state = false;
4649 if (env->log.level) {
4650 const struct bpf_insn_cbs cbs = {
4651 .cb_print = verbose,
4652 .private_data = env,
4655 verbose(env, "%d: ", insn_idx);
4656 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4659 if (bpf_prog_is_dev_bound(env->prog->aux)) {
4660 err = bpf_prog_offload_verify_insn(env, insn_idx,
4666 regs = cur_regs(env);
4667 env->insn_aux_data[insn_idx].seen = true;
4668 if (class == BPF_ALU || class == BPF_ALU64) {
4669 err = check_alu_op(env, insn);
4673 } else if (class == BPF_LDX) {
4674 enum bpf_reg_type *prev_src_type, src_reg_type;
4676 /* check for reserved fields is already done */
4678 /* check src operand */
4679 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4683 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4687 src_reg_type = regs[insn->src_reg].type;
4689 /* check that memory (src_reg + off) is readable,
4690 * the state of dst_reg will be updated by this func
4692 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
4693 BPF_SIZE(insn->code), BPF_READ,
4694 insn->dst_reg, false);
4698 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
4700 if (*prev_src_type == NOT_INIT) {
4702 * dst_reg = *(u32 *)(src_reg + off)
4703 * save type to validate intersecting paths
4705 *prev_src_type = src_reg_type;
4707 } else if (src_reg_type != *prev_src_type &&
4708 (src_reg_type == PTR_TO_CTX ||
4709 *prev_src_type == PTR_TO_CTX)) {
4710 /* ABuser program is trying to use the same insn
4711 * dst_reg = *(u32*) (src_reg + off)
4712 * with different pointer types:
4713 * src_reg == ctx in one branch and
4714 * src_reg == stack|map in some other branch.
4717 verbose(env, "same insn cannot be used with different pointers\n");
4721 } else if (class == BPF_STX) {
4722 enum bpf_reg_type *prev_dst_type, dst_reg_type;
4724 if (BPF_MODE(insn->code) == BPF_XADD) {
4725 err = check_xadd(env, insn_idx, insn);
4732 /* check src1 operand */
4733 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4736 /* check src2 operand */
4737 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4741 dst_reg_type = regs[insn->dst_reg].type;
4743 /* check that memory (dst_reg + off) is writeable */
4744 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4745 BPF_SIZE(insn->code), BPF_WRITE,
4746 insn->src_reg, false);
4750 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
4752 if (*prev_dst_type == NOT_INIT) {
4753 *prev_dst_type = dst_reg_type;
4754 } else if (dst_reg_type != *prev_dst_type &&
4755 (dst_reg_type == PTR_TO_CTX ||
4756 *prev_dst_type == PTR_TO_CTX)) {
4757 verbose(env, "same insn cannot be used with different pointers\n");
4761 } else if (class == BPF_ST) {
4762 if (BPF_MODE(insn->code) != BPF_MEM ||
4763 insn->src_reg != BPF_REG_0) {
4764 verbose(env, "BPF_ST uses reserved fields\n");
4767 /* check src operand */
4768 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4772 if (is_ctx_reg(env, insn->dst_reg)) {
4773 verbose(env, "BPF_ST stores into R%d context is not allowed\n",
4778 /* check that memory (dst_reg + off) is writeable */
4779 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4780 BPF_SIZE(insn->code), BPF_WRITE,
4785 } else if (class == BPF_JMP) {
4786 u8 opcode = BPF_OP(insn->code);
4788 if (opcode == BPF_CALL) {
4789 if (BPF_SRC(insn->code) != BPF_K ||
4791 (insn->src_reg != BPF_REG_0 &&
4792 insn->src_reg != BPF_PSEUDO_CALL) ||
4793 insn->dst_reg != BPF_REG_0) {
4794 verbose(env, "BPF_CALL uses reserved fields\n");
4798 if (insn->src_reg == BPF_PSEUDO_CALL)
4799 err = check_func_call(env, insn, &insn_idx);
4801 err = check_helper_call(env, insn->imm, insn_idx);
4805 } else if (opcode == BPF_JA) {
4806 if (BPF_SRC(insn->code) != BPF_K ||
4808 insn->src_reg != BPF_REG_0 ||
4809 insn->dst_reg != BPF_REG_0) {
4810 verbose(env, "BPF_JA uses reserved fields\n");
4814 insn_idx += insn->off + 1;
4817 } else if (opcode == BPF_EXIT) {
4818 if (BPF_SRC(insn->code) != BPF_K ||
4820 insn->src_reg != BPF_REG_0 ||
4821 insn->dst_reg != BPF_REG_0) {
4822 verbose(env, "BPF_EXIT uses reserved fields\n");
4826 if (state->curframe) {
4827 /* exit from nested function */
4828 prev_insn_idx = insn_idx;
4829 err = prepare_func_exit(env, &insn_idx);
4832 do_print_state = true;
4836 /* eBPF calling convetion is such that R0 is used
4837 * to return the value from eBPF program.
4838 * Make sure that it's readable at this time
4839 * of bpf_exit, which means that program wrote
4840 * something into it earlier
4842 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4846 if (is_pointer_value(env, BPF_REG_0)) {
4847 verbose(env, "R0 leaks addr as return value\n");
4851 err = check_return_code(env);
4855 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4861 do_print_state = true;
4865 err = check_cond_jmp_op(env, insn, &insn_idx);
4869 } else if (class == BPF_LD) {
4870 u8 mode = BPF_MODE(insn->code);
4872 if (mode == BPF_ABS || mode == BPF_IND) {
4873 err = check_ld_abs(env, insn);
4877 } else if (mode == BPF_IMM) {
4878 err = check_ld_imm(env, insn);
4883 env->insn_aux_data[insn_idx].seen = true;
4885 verbose(env, "invalid BPF_LD mode\n");
4889 verbose(env, "unknown insn class %d\n", class);
4896 verbose(env, "processed %d insns (limit %d), stack depth ",
4897 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
4898 for (i = 0; i < env->subprog_cnt + 1; i++) {
4899 u32 depth = env->subprog_stack_depth[i];
4901 verbose(env, "%d", depth);
4902 if (i + 1 < env->subprog_cnt + 1)
4906 env->prog->aux->stack_depth = env->subprog_stack_depth[0];
4910 static int check_map_prealloc(struct bpf_map *map)
4912 return (map->map_type != BPF_MAP_TYPE_HASH &&
4913 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
4914 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
4915 !(map->map_flags & BPF_F_NO_PREALLOC);
4918 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
4919 struct bpf_map *map,
4920 struct bpf_prog *prog)
4923 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4924 * preallocated hash maps, since doing memory allocation
4925 * in overflow_handler can crash depending on where nmi got
4928 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4929 if (!check_map_prealloc(map)) {
4930 verbose(env, "perf_event programs can only use preallocated hash map\n");
4933 if (map->inner_map_meta &&
4934 !check_map_prealloc(map->inner_map_meta)) {
4935 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
4940 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
4941 !bpf_offload_dev_match(prog, map)) {
4942 verbose(env, "offload device mismatch between prog and map\n");
4949 /* look for pseudo eBPF instructions that access map FDs and
4950 * replace them with actual map pointers
4952 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4954 struct bpf_insn *insn = env->prog->insnsi;
4955 int insn_cnt = env->prog->len;
4958 err = bpf_prog_calc_tag(env->prog);
4962 for (i = 0; i < insn_cnt; i++, insn++) {
4963 if (BPF_CLASS(insn->code) == BPF_LDX &&
4964 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4965 verbose(env, "BPF_LDX uses reserved fields\n");
4969 if (BPF_CLASS(insn->code) == BPF_STX &&
4970 ((BPF_MODE(insn->code) != BPF_MEM &&
4971 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4972 verbose(env, "BPF_STX uses reserved fields\n");
4976 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
4977 struct bpf_map *map;
4980 if (i == insn_cnt - 1 || insn[1].code != 0 ||
4981 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
4983 verbose(env, "invalid bpf_ld_imm64 insn\n");
4987 if (insn->src_reg == 0)
4988 /* valid generic load 64-bit imm */
4991 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
4993 "unrecognized bpf_ld_imm64 insn\n");
4997 f = fdget(insn->imm);
4998 map = __bpf_map_get(f);
5000 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5002 return PTR_ERR(map);
5005 err = check_map_prog_compatibility(env, map, env->prog);
5011 /* store map pointer inside BPF_LD_IMM64 instruction */
5012 insn[0].imm = (u32) (unsigned long) map;
5013 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5015 /* check whether we recorded this map already */
5016 for (j = 0; j < env->used_map_cnt; j++)
5017 if (env->used_maps[j] == map) {
5022 if (env->used_map_cnt >= MAX_USED_MAPS) {
5027 /* hold the map. If the program is rejected by verifier,
5028 * the map will be released by release_maps() or it
5029 * will be used by the valid program until it's unloaded
5030 * and all maps are released in free_bpf_prog_info()
5032 map = bpf_map_inc(map, false);
5035 return PTR_ERR(map);
5037 env->used_maps[env->used_map_cnt++] = map;
5046 /* Basic sanity check before we invest more work here. */
5047 if (!bpf_opcode_in_insntable(insn->code)) {
5048 verbose(env, "unknown opcode %02x\n", insn->code);
5053 /* now all pseudo BPF_LD_IMM64 instructions load valid
5054 * 'struct bpf_map *' into a register instead of user map_fd.
5055 * These pointers will be used later by verifier to validate map access.
5060 /* drop refcnt of maps used by the rejected program */
5061 static void release_maps(struct bpf_verifier_env *env)
5065 for (i = 0; i < env->used_map_cnt; i++)
5066 bpf_map_put(env->used_maps[i]);
5069 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5070 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5072 struct bpf_insn *insn = env->prog->insnsi;
5073 int insn_cnt = env->prog->len;
5076 for (i = 0; i < insn_cnt; i++, insn++)
5077 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5081 /* single env->prog->insni[off] instruction was replaced with the range
5082 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5083 * [0, off) and [off, end) to new locations, so the patched range stays zero
5085 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5088 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5093 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
5096 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5097 memcpy(new_data + off + cnt - 1, old_data + off,
5098 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5099 for (i = off; i < off + cnt - 1; i++)
5100 new_data[i].seen = true;
5101 env->insn_aux_data = new_data;
5106 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5112 for (i = 0; i < env->subprog_cnt; i++) {
5113 if (env->subprog_starts[i] < off)
5115 env->subprog_starts[i] += len - 1;
5119 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5120 const struct bpf_insn *patch, u32 len)
5122 struct bpf_prog *new_prog;
5124 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5127 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5129 adjust_subprog_starts(env, off, len);
5133 /* The verifier does more data flow analysis than llvm and will not
5134 * explore branches that are dead at run time. Malicious programs can
5135 * have dead code too. Therefore replace all dead at-run-time code
5138 * Just nops are not optimal, e.g. if they would sit at the end of the
5139 * program and through another bug we would manage to jump there, then
5140 * we'd execute beyond program memory otherwise. Returning exception
5141 * code also wouldn't work since we can have subprogs where the dead
5142 * code could be located.
5144 static void sanitize_dead_code(struct bpf_verifier_env *env)
5146 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5147 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5148 struct bpf_insn *insn = env->prog->insnsi;
5149 const int insn_cnt = env->prog->len;
5152 for (i = 0; i < insn_cnt; i++) {
5153 if (aux_data[i].seen)
5155 memcpy(insn + i, &trap, sizeof(trap));
5159 /* convert load instructions that access fields of 'struct __sk_buff'
5160 * into sequence of instructions that access fields of 'struct sk_buff'
5162 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5164 const struct bpf_verifier_ops *ops = env->ops;
5165 int i, cnt, size, ctx_field_size, delta = 0;
5166 const int insn_cnt = env->prog->len;
5167 struct bpf_insn insn_buf[16], *insn;
5168 struct bpf_prog *new_prog;
5169 enum bpf_access_type type;
5170 bool is_narrower_load;
5173 if (ops->gen_prologue) {
5174 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5176 if (cnt >= ARRAY_SIZE(insn_buf)) {
5177 verbose(env, "bpf verifier is misconfigured\n");
5180 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5184 env->prog = new_prog;
5189 if (!ops->convert_ctx_access)
5192 insn = env->prog->insnsi + delta;
5194 for (i = 0; i < insn_cnt; i++, insn++) {
5195 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5196 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5197 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5198 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5200 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5201 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5202 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5203 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5208 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
5211 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5212 size = BPF_LDST_BYTES(insn);
5214 /* If the read access is a narrower load of the field,
5215 * convert to a 4/8-byte load, to minimum program type specific
5216 * convert_ctx_access changes. If conversion is successful,
5217 * we will apply proper mask to the result.
5219 is_narrower_load = size < ctx_field_size;
5220 if (is_narrower_load) {
5221 u32 off = insn->off;
5224 if (type == BPF_WRITE) {
5225 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5230 if (ctx_field_size == 4)
5232 else if (ctx_field_size == 8)
5235 insn->off = off & ~(ctx_field_size - 1);
5236 insn->code = BPF_LDX | BPF_MEM | size_code;
5240 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
5242 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5243 (ctx_field_size && !target_size)) {
5244 verbose(env, "bpf verifier is misconfigured\n");
5248 if (is_narrower_load && size < target_size) {
5249 if (ctx_field_size <= 4)
5250 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5251 (1 << size * 8) - 1);
5253 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5254 (1 << size * 8) - 1);
5257 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5263 /* keep walking new program and skip insns we just inserted */
5264 env->prog = new_prog;
5265 insn = new_prog->insnsi + i + delta;
5271 static int jit_subprogs(struct bpf_verifier_env *env)
5273 struct bpf_prog *prog = env->prog, **func, *tmp;
5274 int i, j, subprog_start, subprog_end = 0, len, subprog;
5275 struct bpf_insn *insn;
5279 if (env->subprog_cnt == 0)
5282 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5283 if (insn->code != (BPF_JMP | BPF_CALL) ||
5284 insn->src_reg != BPF_PSEUDO_CALL)
5286 subprog = find_subprog(env, i + insn->imm + 1);
5288 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5292 /* temporarily remember subprog id inside insn instead of
5293 * aux_data, since next loop will split up all insns into funcs
5295 insn->off = subprog + 1;
5296 /* remember original imm in case JIT fails and fallback
5297 * to interpreter will be needed
5299 env->insn_aux_data[i].call_imm = insn->imm;
5300 /* point imm to __bpf_call_base+1 from JITs point of view */
5304 func = kzalloc(sizeof(prog) * (env->subprog_cnt + 1), GFP_KERNEL);
5308 for (i = 0; i <= env->subprog_cnt; i++) {
5309 subprog_start = subprog_end;
5310 if (env->subprog_cnt == i)
5311 subprog_end = prog->len;
5313 subprog_end = env->subprog_starts[i];
5315 len = subprog_end - subprog_start;
5316 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5319 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5320 len * sizeof(struct bpf_insn));
5321 func[i]->type = prog->type;
5323 if (bpf_prog_calc_tag(func[i]))
5325 func[i]->is_func = 1;
5326 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5327 * Long term would need debug info to populate names
5329 func[i]->aux->name[0] = 'F';
5330 func[i]->aux->stack_depth = env->subprog_stack_depth[i];
5331 func[i]->jit_requested = 1;
5332 func[i] = bpf_int_jit_compile(func[i]);
5333 if (!func[i]->jited) {
5339 /* at this point all bpf functions were successfully JITed
5340 * now populate all bpf_calls with correct addresses and
5341 * run last pass of JIT
5343 for (i = 0; i <= env->subprog_cnt; i++) {
5344 insn = func[i]->insnsi;
5345 for (j = 0; j < func[i]->len; j++, insn++) {
5346 if (insn->code != (BPF_JMP | BPF_CALL) ||
5347 insn->src_reg != BPF_PSEUDO_CALL)
5349 subprog = insn->off;
5351 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5352 func[subprog]->bpf_func -
5356 for (i = 0; i <= env->subprog_cnt; i++) {
5357 old_bpf_func = func[i]->bpf_func;
5358 tmp = bpf_int_jit_compile(func[i]);
5359 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5360 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5367 /* finally lock prog and jit images for all functions and
5370 for (i = 0; i <= env->subprog_cnt; i++) {
5371 bpf_prog_lock_ro(func[i]);
5372 bpf_prog_kallsyms_add(func[i]);
5375 /* Last step: make now unused interpreter insns from main
5376 * prog consistent for later dump requests, so they can
5377 * later look the same as if they were interpreted only.
5379 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5382 if (insn->code != (BPF_JMP | BPF_CALL) ||
5383 insn->src_reg != BPF_PSEUDO_CALL)
5385 insn->off = env->insn_aux_data[i].call_imm;
5386 subprog = find_subprog(env, i + insn->off + 1);
5387 addr = (unsigned long)func[subprog + 1]->bpf_func;
5389 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5390 addr - __bpf_call_base;
5394 prog->bpf_func = func[0]->bpf_func;
5395 prog->aux->func = func;
5396 prog->aux->func_cnt = env->subprog_cnt + 1;
5399 for (i = 0; i <= env->subprog_cnt; i++)
5401 bpf_jit_free(func[i]);
5403 /* cleanup main prog to be interpreted */
5404 prog->jit_requested = 0;
5405 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5406 if (insn->code != (BPF_JMP | BPF_CALL) ||
5407 insn->src_reg != BPF_PSEUDO_CALL)
5410 insn->imm = env->insn_aux_data[i].call_imm;
5415 static int fixup_call_args(struct bpf_verifier_env *env)
5417 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5418 struct bpf_prog *prog = env->prog;
5419 struct bpf_insn *insn = prog->insnsi;
5425 if (env->prog->jit_requested) {
5426 err = jit_subprogs(env);
5430 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5431 for (i = 0; i < prog->len; i++, insn++) {
5432 if (insn->code != (BPF_JMP | BPF_CALL) ||
5433 insn->src_reg != BPF_PSEUDO_CALL)
5435 depth = get_callee_stack_depth(env, insn, i);
5438 bpf_patch_call_args(insn, depth);
5445 /* fixup insn->imm field of bpf_call instructions
5446 * and inline eligible helpers as explicit sequence of BPF instructions
5448 * this function is called after eBPF program passed verification
5450 static int fixup_bpf_calls(struct bpf_verifier_env *env)
5452 struct bpf_prog *prog = env->prog;
5453 struct bpf_insn *insn = prog->insnsi;
5454 const struct bpf_func_proto *fn;
5455 const int insn_cnt = prog->len;
5456 struct bpf_insn insn_buf[16];
5457 struct bpf_prog *new_prog;
5458 struct bpf_map *map_ptr;
5459 int i, cnt, delta = 0;
5461 for (i = 0; i < insn_cnt; i++, insn++) {
5462 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
5463 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5464 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
5465 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5466 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
5467 struct bpf_insn mask_and_div[] = {
5468 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5470 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
5471 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
5472 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
5475 struct bpf_insn mask_and_mod[] = {
5476 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5477 /* Rx mod 0 -> Rx */
5478 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
5481 struct bpf_insn *patchlet;
5483 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5484 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5485 patchlet = mask_and_div + (is64 ? 1 : 0);
5486 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
5488 patchlet = mask_and_mod + (is64 ? 1 : 0);
5489 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
5492 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
5497 env->prog = prog = new_prog;
5498 insn = new_prog->insnsi + i + delta;
5502 if (insn->code != (BPF_JMP | BPF_CALL))
5504 if (insn->src_reg == BPF_PSEUDO_CALL)
5507 if (insn->imm == BPF_FUNC_get_route_realm)
5508 prog->dst_needed = 1;
5509 if (insn->imm == BPF_FUNC_get_prandom_u32)
5510 bpf_user_rnd_init_once();
5511 if (insn->imm == BPF_FUNC_override_return)
5512 prog->kprobe_override = 1;
5513 if (insn->imm == BPF_FUNC_tail_call) {
5514 /* If we tail call into other programs, we
5515 * cannot make any assumptions since they can
5516 * be replaced dynamically during runtime in
5517 * the program array.
5519 prog->cb_access = 1;
5520 env->prog->aux->stack_depth = MAX_BPF_STACK;
5522 /* mark bpf_tail_call as different opcode to avoid
5523 * conditional branch in the interpeter for every normal
5524 * call and to prevent accidental JITing by JIT compiler
5525 * that doesn't support bpf_tail_call yet
5528 insn->code = BPF_JMP | BPF_TAIL_CALL;
5530 /* instead of changing every JIT dealing with tail_call
5531 * emit two extra insns:
5532 * if (index >= max_entries) goto out;
5533 * index &= array->index_mask;
5534 * to avoid out-of-bounds cpu speculation
5536 map_ptr = env->insn_aux_data[i + delta].map_ptr;
5537 if (map_ptr == BPF_MAP_PTR_POISON) {
5538 verbose(env, "tail_call abusing map_ptr\n");
5541 if (!map_ptr->unpriv_array)
5543 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
5544 map_ptr->max_entries, 2);
5545 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
5546 container_of(map_ptr,
5549 insn_buf[2] = *insn;
5551 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5556 env->prog = prog = new_prog;
5557 insn = new_prog->insnsi + i + delta;
5561 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5562 * handlers are currently limited to 64 bit only.
5564 if (prog->jit_requested && BITS_PER_LONG == 64 &&
5565 insn->imm == BPF_FUNC_map_lookup_elem) {
5566 map_ptr = env->insn_aux_data[i + delta].map_ptr;
5567 if (map_ptr == BPF_MAP_PTR_POISON ||
5568 !map_ptr->ops->map_gen_lookup)
5569 goto patch_call_imm;
5571 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
5572 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5573 verbose(env, "bpf verifier is misconfigured\n");
5577 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
5584 /* keep walking new program and skip insns we just inserted */
5585 env->prog = prog = new_prog;
5586 insn = new_prog->insnsi + i + delta;
5590 if (insn->imm == BPF_FUNC_redirect_map) {
5591 /* Note, we cannot use prog directly as imm as subsequent
5592 * rewrites would still change the prog pointer. The only
5593 * stable address we can use is aux, which also works with
5594 * prog clones during blinding.
5596 u64 addr = (unsigned long)prog->aux;
5597 struct bpf_insn r4_ld[] = {
5598 BPF_LD_IMM64(BPF_REG_4, addr),
5601 cnt = ARRAY_SIZE(r4_ld);
5603 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
5608 env->prog = prog = new_prog;
5609 insn = new_prog->insnsi + i + delta;
5612 fn = env->ops->get_func_proto(insn->imm, env->prog);
5613 /* all functions that have prototype and verifier allowed
5614 * programs to call them, must be real in-kernel functions
5618 "kernel subsystem misconfigured func %s#%d\n",
5619 func_id_name(insn->imm), insn->imm);
5622 insn->imm = fn->func - __bpf_call_base;
5628 static void free_states(struct bpf_verifier_env *env)
5630 struct bpf_verifier_state_list *sl, *sln;
5633 if (!env->explored_states)
5636 for (i = 0; i < env->prog->len; i++) {
5637 sl = env->explored_states[i];
5640 while (sl != STATE_LIST_MARK) {
5642 free_verifier_state(&sl->state, false);
5648 kfree(env->explored_states);
5651 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
5653 struct bpf_verifier_env *env;
5654 struct bpf_verifier_log *log;
5657 /* no program is valid */
5658 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
5661 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5662 * allocate/free it every time bpf_check() is called
5664 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
5669 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
5672 if (!env->insn_aux_data)
5675 env->ops = bpf_verifier_ops[env->prog->type];
5677 /* grab the mutex to protect few globals used by verifier */
5678 mutex_lock(&bpf_verifier_lock);
5680 if (attr->log_level || attr->log_buf || attr->log_size) {
5681 /* user requested verbose verifier output
5682 * and supplied buffer to store the verification trace
5684 log->level = attr->log_level;
5685 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
5686 log->len_total = attr->log_size;
5689 /* log attributes have to be sane */
5690 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
5691 !log->level || !log->ubuf)
5695 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
5696 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
5697 env->strict_alignment = true;
5699 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5700 ret = bpf_prog_offload_verifier_prep(env);
5705 ret = replace_map_fd_with_map_ptr(env);
5707 goto skip_full_check;
5709 env->explored_states = kcalloc(env->prog->len,
5710 sizeof(struct bpf_verifier_state_list *),
5713 if (!env->explored_states)
5714 goto skip_full_check;
5716 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
5718 ret = check_cfg(env);
5720 goto skip_full_check;
5722 ret = do_check(env);
5723 if (env->cur_state) {
5724 free_verifier_state(env->cur_state, true);
5725 env->cur_state = NULL;
5729 while (!pop_stack(env, NULL, NULL));
5733 sanitize_dead_code(env);
5736 ret = check_max_stack_depth(env);
5739 /* program is valid, convert *(u32*)(ctx + off) accesses */
5740 ret = convert_ctx_accesses(env);
5743 ret = fixup_bpf_calls(env);
5746 ret = fixup_call_args(env);
5748 if (log->level && bpf_verifier_log_full(log))
5750 if (log->level && !log->ubuf) {
5752 goto err_release_maps;
5755 if (ret == 0 && env->used_map_cnt) {
5756 /* if program passed verifier, update used_maps in bpf_prog_info */
5757 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
5758 sizeof(env->used_maps[0]),
5761 if (!env->prog->aux->used_maps) {
5763 goto err_release_maps;
5766 memcpy(env->prog->aux->used_maps, env->used_maps,
5767 sizeof(env->used_maps[0]) * env->used_map_cnt);
5768 env->prog->aux->used_map_cnt = env->used_map_cnt;
5770 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
5771 * bpf_ld_imm64 instructions
5773 convert_pseudo_ld_imm64(env);
5777 if (!env->prog->aux->used_maps)
5778 /* if we didn't copy map pointers into bpf_prog_info, release
5779 * them now. Otherwise free_bpf_prog_info() will release them.
5784 mutex_unlock(&bpf_verifier_lock);
5785 vfree(env->insn_aux_data);