Loading...
Note: File does not exist in v3.5.6.
1/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 *
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.
7 *
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.
12 */
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
26#include "disasm.h"
27
28static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29#define BPF_PROG_TYPE(_id, _name) \
30 [_id] = & _name ## _verifier_ops,
31#define BPF_MAP_TYPE(_id, _ops)
32#include <linux/bpf_types.h>
33#undef BPF_PROG_TYPE
34#undef BPF_MAP_TYPE
35};
36
37/* bpf_check() is a static code analyzer that walks eBPF program
38 * instruction by instruction and updates register/stack state.
39 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
40 *
41 * The first pass is depth-first-search to check that the program is a DAG.
42 * It rejects the following programs:
43 * - larger than BPF_MAXINSNS insns
44 * - if loop is present (detected via back-edge)
45 * - unreachable insns exist (shouldn't be a forest. program = one function)
46 * - out of bounds or malformed jumps
47 * The second pass is all possible path descent from the 1st insn.
48 * Since it's analyzing all pathes through the program, the length of the
49 * analysis is limited to 64k insn, which may be hit even if total number of
50 * insn is less then 4K, but there are too many branches that change stack/regs.
51 * Number of 'branches to be analyzed' is limited to 1k
52 *
53 * On entry to each instruction, each register has a type, and the instruction
54 * changes the types of the registers depending on instruction semantics.
55 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
56 * copied to R1.
57 *
58 * All registers are 64-bit.
59 * R0 - return register
60 * R1-R5 argument passing registers
61 * R6-R9 callee saved registers
62 * R10 - frame pointer read-only
63 *
64 * At the start of BPF program the register R1 contains a pointer to bpf_context
65 * and has type PTR_TO_CTX.
66 *
67 * Verifier tracks arithmetic operations on pointers in case:
68 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
69 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
70 * 1st insn copies R10 (which has FRAME_PTR) type into R1
71 * and 2nd arithmetic instruction is pattern matched to recognize
72 * that it wants to construct a pointer to some element within stack.
73 * So after 2nd insn, the register R1 has type PTR_TO_STACK
74 * (and -20 constant is saved for further stack bounds checking).
75 * Meaning that this reg is a pointer to stack plus known immediate constant.
76 *
77 * Most of the time the registers have SCALAR_VALUE type, which
78 * means the register has some value, but it's not a valid pointer.
79 * (like pointer plus pointer becomes SCALAR_VALUE type)
80 *
81 * When verifier sees load or store instructions the type of base register
82 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
83 * types recognized by check_mem_access() function.
84 *
85 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
86 * and the range of [ptr, ptr + map's value_size) is accessible.
87 *
88 * registers used to pass values to function calls are checked against
89 * function argument constraints.
90 *
91 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
92 * It means that the register type passed to this function must be
93 * PTR_TO_STACK and it will be used inside the function as
94 * 'pointer to map element key'
95 *
96 * For example the argument constraints for bpf_map_lookup_elem():
97 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
98 * .arg1_type = ARG_CONST_MAP_PTR,
99 * .arg2_type = ARG_PTR_TO_MAP_KEY,
100 *
101 * ret_type says that this function returns 'pointer to map elem value or null'
102 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
103 * 2nd argument should be a pointer to stack, which will be used inside
104 * the helper function as a pointer to map element key.
105 *
106 * On the kernel side the helper function looks like:
107 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
108 * {
109 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
110 * void *key = (void *) (unsigned long) r2;
111 * void *value;
112 *
113 * here kernel can access 'key' and 'map' pointers safely, knowing that
114 * [key, key + map->key_size) bytes are valid and were initialized on
115 * the stack of eBPF program.
116 * }
117 *
118 * Corresponding eBPF program may look like:
119 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
120 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
121 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
122 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
123 * here verifier looks at prototype of map_lookup_elem() and sees:
124 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
125 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
126 *
127 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
128 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
129 * and were initialized prior to this call.
130 * If it's ok, then verifier allows this BPF_CALL insn and looks at
131 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
132 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
133 * returns ether pointer to map value or NULL.
134 *
135 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
136 * insn, the register holding that pointer in the true branch changes state to
137 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
138 * branch. See check_cond_jmp_op().
139 *
140 * After the call R0 is set to return type of the function and registers R1-R5
141 * are set to NOT_INIT to indicate that they are no longer readable.
142 */
143
144/* verifier_state + insn_idx are pushed to stack when branch is encountered */
145struct bpf_verifier_stack_elem {
146 /* verifer state is 'st'
147 * before processing instruction 'insn_idx'
148 * and after processing instruction 'prev_insn_idx'
149 */
150 struct bpf_verifier_state st;
151 int insn_idx;
152 int prev_insn_idx;
153 struct bpf_verifier_stack_elem *next;
154};
155
156#define BPF_COMPLEXITY_LIMIT_INSNS 131072
157#define BPF_COMPLEXITY_LIMIT_STACK 1024
158
159#define BPF_MAP_PTR_UNPRIV 1UL
160#define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
161 POISON_POINTER_DELTA))
162#define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
163
164static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
165{
166 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
167}
168
169static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
170{
171 return aux->map_state & BPF_MAP_PTR_UNPRIV;
172}
173
174static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
175 const struct bpf_map *map, bool unpriv)
176{
177 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
178 unpriv |= bpf_map_ptr_unpriv(aux);
179 aux->map_state = (unsigned long)map |
180 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
181}
182
183struct bpf_call_arg_meta {
184 struct bpf_map *map_ptr;
185 bool raw_mode;
186 bool pkt_access;
187 int regno;
188 int access_size;
189};
190
191static DEFINE_MUTEX(bpf_verifier_lock);
192
193void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
194 va_list args)
195{
196 unsigned int n;
197
198 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
199
200 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
201 "verifier log line truncated - local buffer too short\n");
202
203 n = min(log->len_total - log->len_used - 1, n);
204 log->kbuf[n] = '\0';
205
206 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
207 log->len_used += n;
208 else
209 log->ubuf = NULL;
210}
211
212/* log_level controls verbosity level of eBPF verifier.
213 * bpf_verifier_log_write() is used to dump the verification trace to the log,
214 * so the user can figure out what's wrong with the program
215 */
216__printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
217 const char *fmt, ...)
218{
219 va_list args;
220
221 if (!bpf_verifier_log_needed(&env->log))
222 return;
223
224 va_start(args, fmt);
225 bpf_verifier_vlog(&env->log, fmt, args);
226 va_end(args);
227}
228EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
229
230__printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
231{
232 struct bpf_verifier_env *env = private_data;
233 va_list args;
234
235 if (!bpf_verifier_log_needed(&env->log))
236 return;
237
238 va_start(args, fmt);
239 bpf_verifier_vlog(&env->log, fmt, args);
240 va_end(args);
241}
242
243static bool type_is_pkt_pointer(enum bpf_reg_type type)
244{
245 return type == PTR_TO_PACKET ||
246 type == PTR_TO_PACKET_META;
247}
248
249/* string representation of 'enum bpf_reg_type' */
250static const char * const reg_type_str[] = {
251 [NOT_INIT] = "?",
252 [SCALAR_VALUE] = "inv",
253 [PTR_TO_CTX] = "ctx",
254 [CONST_PTR_TO_MAP] = "map_ptr",
255 [PTR_TO_MAP_VALUE] = "map_value",
256 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
257 [PTR_TO_STACK] = "fp",
258 [PTR_TO_PACKET] = "pkt",
259 [PTR_TO_PACKET_META] = "pkt_meta",
260 [PTR_TO_PACKET_END] = "pkt_end",
261};
262
263static void print_liveness(struct bpf_verifier_env *env,
264 enum bpf_reg_liveness live)
265{
266 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
267 verbose(env, "_");
268 if (live & REG_LIVE_READ)
269 verbose(env, "r");
270 if (live & REG_LIVE_WRITTEN)
271 verbose(env, "w");
272}
273
274static struct bpf_func_state *func(struct bpf_verifier_env *env,
275 const struct bpf_reg_state *reg)
276{
277 struct bpf_verifier_state *cur = env->cur_state;
278
279 return cur->frame[reg->frameno];
280}
281
282static void print_verifier_state(struct bpf_verifier_env *env,
283 const struct bpf_func_state *state)
284{
285 const struct bpf_reg_state *reg;
286 enum bpf_reg_type t;
287 int i;
288
289 if (state->frameno)
290 verbose(env, " frame%d:", state->frameno);
291 for (i = 0; i < MAX_BPF_REG; i++) {
292 reg = &state->regs[i];
293 t = reg->type;
294 if (t == NOT_INIT)
295 continue;
296 verbose(env, " R%d", i);
297 print_liveness(env, reg->live);
298 verbose(env, "=%s", reg_type_str[t]);
299 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
300 tnum_is_const(reg->var_off)) {
301 /* reg->off should be 0 for SCALAR_VALUE */
302 verbose(env, "%lld", reg->var_off.value + reg->off);
303 if (t == PTR_TO_STACK)
304 verbose(env, ",call_%d", func(env, reg)->callsite);
305 } else {
306 verbose(env, "(id=%d", reg->id);
307 if (t != SCALAR_VALUE)
308 verbose(env, ",off=%d", reg->off);
309 if (type_is_pkt_pointer(t))
310 verbose(env, ",r=%d", reg->range);
311 else if (t == CONST_PTR_TO_MAP ||
312 t == PTR_TO_MAP_VALUE ||
313 t == PTR_TO_MAP_VALUE_OR_NULL)
314 verbose(env, ",ks=%d,vs=%d",
315 reg->map_ptr->key_size,
316 reg->map_ptr->value_size);
317 if (tnum_is_const(reg->var_off)) {
318 /* Typically an immediate SCALAR_VALUE, but
319 * could be a pointer whose offset is too big
320 * for reg->off
321 */
322 verbose(env, ",imm=%llx", reg->var_off.value);
323 } else {
324 if (reg->smin_value != reg->umin_value &&
325 reg->smin_value != S64_MIN)
326 verbose(env, ",smin_value=%lld",
327 (long long)reg->smin_value);
328 if (reg->smax_value != reg->umax_value &&
329 reg->smax_value != S64_MAX)
330 verbose(env, ",smax_value=%lld",
331 (long long)reg->smax_value);
332 if (reg->umin_value != 0)
333 verbose(env, ",umin_value=%llu",
334 (unsigned long long)reg->umin_value);
335 if (reg->umax_value != U64_MAX)
336 verbose(env, ",umax_value=%llu",
337 (unsigned long long)reg->umax_value);
338 if (!tnum_is_unknown(reg->var_off)) {
339 char tn_buf[48];
340
341 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
342 verbose(env, ",var_off=%s", tn_buf);
343 }
344 }
345 verbose(env, ")");
346 }
347 }
348 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
349 if (state->stack[i].slot_type[0] == STACK_SPILL) {
350 verbose(env, " fp%d",
351 (-i - 1) * BPF_REG_SIZE);
352 print_liveness(env, state->stack[i].spilled_ptr.live);
353 verbose(env, "=%s",
354 reg_type_str[state->stack[i].spilled_ptr.type]);
355 }
356 if (state->stack[i].slot_type[0] == STACK_ZERO)
357 verbose(env, " fp%d=0", (-i - 1) * BPF_REG_SIZE);
358 }
359 verbose(env, "\n");
360}
361
362static int copy_stack_state(struct bpf_func_state *dst,
363 const struct bpf_func_state *src)
364{
365 if (!src->stack)
366 return 0;
367 if (WARN_ON_ONCE(dst->allocated_stack < src->allocated_stack)) {
368 /* internal bug, make state invalid to reject the program */
369 memset(dst, 0, sizeof(*dst));
370 return -EFAULT;
371 }
372 memcpy(dst->stack, src->stack,
373 sizeof(*src->stack) * (src->allocated_stack / BPF_REG_SIZE));
374 return 0;
375}
376
377/* do_check() starts with zero-sized stack in struct bpf_verifier_state to
378 * make it consume minimal amount of memory. check_stack_write() access from
379 * the program calls into realloc_func_state() to grow the stack size.
380 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
381 * which this function copies over. It points to previous bpf_verifier_state
382 * which is never reallocated
383 */
384static int realloc_func_state(struct bpf_func_state *state, int size,
385 bool copy_old)
386{
387 u32 old_size = state->allocated_stack;
388 struct bpf_stack_state *new_stack;
389 int slot = size / BPF_REG_SIZE;
390
391 if (size <= old_size || !size) {
392 if (copy_old)
393 return 0;
394 state->allocated_stack = slot * BPF_REG_SIZE;
395 if (!size && old_size) {
396 kfree(state->stack);
397 state->stack = NULL;
398 }
399 return 0;
400 }
401 new_stack = kmalloc_array(slot, sizeof(struct bpf_stack_state),
402 GFP_KERNEL);
403 if (!new_stack)
404 return -ENOMEM;
405 if (copy_old) {
406 if (state->stack)
407 memcpy(new_stack, state->stack,
408 sizeof(*new_stack) * (old_size / BPF_REG_SIZE));
409 memset(new_stack + old_size / BPF_REG_SIZE, 0,
410 sizeof(*new_stack) * (size - old_size) / BPF_REG_SIZE);
411 }
412 state->allocated_stack = slot * BPF_REG_SIZE;
413 kfree(state->stack);
414 state->stack = new_stack;
415 return 0;
416}
417
418static void free_func_state(struct bpf_func_state *state)
419{
420 if (!state)
421 return;
422 kfree(state->stack);
423 kfree(state);
424}
425
426static void free_verifier_state(struct bpf_verifier_state *state,
427 bool free_self)
428{
429 int i;
430
431 for (i = 0; i <= state->curframe; i++) {
432 free_func_state(state->frame[i]);
433 state->frame[i] = NULL;
434 }
435 if (free_self)
436 kfree(state);
437}
438
439/* copy verifier state from src to dst growing dst stack space
440 * when necessary to accommodate larger src stack
441 */
442static int copy_func_state(struct bpf_func_state *dst,
443 const struct bpf_func_state *src)
444{
445 int err;
446
447 err = realloc_func_state(dst, src->allocated_stack, false);
448 if (err)
449 return err;
450 memcpy(dst, src, offsetof(struct bpf_func_state, allocated_stack));
451 return copy_stack_state(dst, src);
452}
453
454static int copy_verifier_state(struct bpf_verifier_state *dst_state,
455 const struct bpf_verifier_state *src)
456{
457 struct bpf_func_state *dst;
458 int i, err;
459
460 /* if dst has more stack frames then src frame, free them */
461 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
462 free_func_state(dst_state->frame[i]);
463 dst_state->frame[i] = NULL;
464 }
465 dst_state->curframe = src->curframe;
466 dst_state->parent = src->parent;
467 for (i = 0; i <= src->curframe; i++) {
468 dst = dst_state->frame[i];
469 if (!dst) {
470 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
471 if (!dst)
472 return -ENOMEM;
473 dst_state->frame[i] = dst;
474 }
475 err = copy_func_state(dst, src->frame[i]);
476 if (err)
477 return err;
478 }
479 return 0;
480}
481
482static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
483 int *insn_idx)
484{
485 struct bpf_verifier_state *cur = env->cur_state;
486 struct bpf_verifier_stack_elem *elem, *head = env->head;
487 int err;
488
489 if (env->head == NULL)
490 return -ENOENT;
491
492 if (cur) {
493 err = copy_verifier_state(cur, &head->st);
494 if (err)
495 return err;
496 }
497 if (insn_idx)
498 *insn_idx = head->insn_idx;
499 if (prev_insn_idx)
500 *prev_insn_idx = head->prev_insn_idx;
501 elem = head->next;
502 free_verifier_state(&head->st, false);
503 kfree(head);
504 env->head = elem;
505 env->stack_size--;
506 return 0;
507}
508
509static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
510 int insn_idx, int prev_insn_idx)
511{
512 struct bpf_verifier_state *cur = env->cur_state;
513 struct bpf_verifier_stack_elem *elem;
514 int err;
515
516 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
517 if (!elem)
518 goto err;
519
520 elem->insn_idx = insn_idx;
521 elem->prev_insn_idx = prev_insn_idx;
522 elem->next = env->head;
523 env->head = elem;
524 env->stack_size++;
525 err = copy_verifier_state(&elem->st, cur);
526 if (err)
527 goto err;
528 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
529 verbose(env, "BPF program is too complex\n");
530 goto err;
531 }
532 return &elem->st;
533err:
534 free_verifier_state(env->cur_state, true);
535 env->cur_state = NULL;
536 /* pop all elements and return */
537 while (!pop_stack(env, NULL, NULL));
538 return NULL;
539}
540
541#define CALLER_SAVED_REGS 6
542static const int caller_saved[CALLER_SAVED_REGS] = {
543 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
544};
545
546static void __mark_reg_not_init(struct bpf_reg_state *reg);
547
548/* Mark the unknown part of a register (variable offset or scalar value) as
549 * known to have the value @imm.
550 */
551static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
552{
553 reg->id = 0;
554 reg->var_off = tnum_const(imm);
555 reg->smin_value = (s64)imm;
556 reg->smax_value = (s64)imm;
557 reg->umin_value = imm;
558 reg->umax_value = imm;
559}
560
561/* Mark the 'variable offset' part of a register as zero. This should be
562 * used only on registers holding a pointer type.
563 */
564static void __mark_reg_known_zero(struct bpf_reg_state *reg)
565{
566 __mark_reg_known(reg, 0);
567}
568
569static void __mark_reg_const_zero(struct bpf_reg_state *reg)
570{
571 __mark_reg_known(reg, 0);
572 reg->off = 0;
573 reg->type = SCALAR_VALUE;
574}
575
576static void mark_reg_known_zero(struct bpf_verifier_env *env,
577 struct bpf_reg_state *regs, u32 regno)
578{
579 if (WARN_ON(regno >= MAX_BPF_REG)) {
580 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
581 /* Something bad happened, let's kill all regs */
582 for (regno = 0; regno < MAX_BPF_REG; regno++)
583 __mark_reg_not_init(regs + regno);
584 return;
585 }
586 __mark_reg_known_zero(regs + regno);
587}
588
589static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
590{
591 return type_is_pkt_pointer(reg->type);
592}
593
594static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
595{
596 return reg_is_pkt_pointer(reg) ||
597 reg->type == PTR_TO_PACKET_END;
598}
599
600/* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
601static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
602 enum bpf_reg_type which)
603{
604 /* The register can already have a range from prior markings.
605 * This is fine as long as it hasn't been advanced from its
606 * origin.
607 */
608 return reg->type == which &&
609 reg->id == 0 &&
610 reg->off == 0 &&
611 tnum_equals_const(reg->var_off, 0);
612}
613
614/* Attempts to improve min/max values based on var_off information */
615static void __update_reg_bounds(struct bpf_reg_state *reg)
616{
617 /* min signed is max(sign bit) | min(other bits) */
618 reg->smin_value = max_t(s64, reg->smin_value,
619 reg->var_off.value | (reg->var_off.mask & S64_MIN));
620 /* max signed is min(sign bit) | max(other bits) */
621 reg->smax_value = min_t(s64, reg->smax_value,
622 reg->var_off.value | (reg->var_off.mask & S64_MAX));
623 reg->umin_value = max(reg->umin_value, reg->var_off.value);
624 reg->umax_value = min(reg->umax_value,
625 reg->var_off.value | reg->var_off.mask);
626}
627
628/* Uses signed min/max values to inform unsigned, and vice-versa */
629static void __reg_deduce_bounds(struct bpf_reg_state *reg)
630{
631 /* Learn sign from signed bounds.
632 * If we cannot cross the sign boundary, then signed and unsigned bounds
633 * are the same, so combine. This works even in the negative case, e.g.
634 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
635 */
636 if (reg->smin_value >= 0 || reg->smax_value < 0) {
637 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
638 reg->umin_value);
639 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
640 reg->umax_value);
641 return;
642 }
643 /* Learn sign from unsigned bounds. Signed bounds cross the sign
644 * boundary, so we must be careful.
645 */
646 if ((s64)reg->umax_value >= 0) {
647 /* Positive. We can't learn anything from the smin, but smax
648 * is positive, hence safe.
649 */
650 reg->smin_value = reg->umin_value;
651 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
652 reg->umax_value);
653 } else if ((s64)reg->umin_value < 0) {
654 /* Negative. We can't learn anything from the smax, but smin
655 * is negative, hence safe.
656 */
657 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
658 reg->umin_value);
659 reg->smax_value = reg->umax_value;
660 }
661}
662
663/* Attempts to improve var_off based on unsigned min/max information */
664static void __reg_bound_offset(struct bpf_reg_state *reg)
665{
666 reg->var_off = tnum_intersect(reg->var_off,
667 tnum_range(reg->umin_value,
668 reg->umax_value));
669}
670
671/* Reset the min/max bounds of a register */
672static void __mark_reg_unbounded(struct bpf_reg_state *reg)
673{
674 reg->smin_value = S64_MIN;
675 reg->smax_value = S64_MAX;
676 reg->umin_value = 0;
677 reg->umax_value = U64_MAX;
678}
679
680/* Mark a register as having a completely unknown (scalar) value. */
681static void __mark_reg_unknown(struct bpf_reg_state *reg)
682{
683 reg->type = SCALAR_VALUE;
684 reg->id = 0;
685 reg->off = 0;
686 reg->var_off = tnum_unknown;
687 reg->frameno = 0;
688 __mark_reg_unbounded(reg);
689}
690
691static void mark_reg_unknown(struct bpf_verifier_env *env,
692 struct bpf_reg_state *regs, u32 regno)
693{
694 if (WARN_ON(regno >= MAX_BPF_REG)) {
695 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
696 /* Something bad happened, let's kill all regs except FP */
697 for (regno = 0; regno < BPF_REG_FP; regno++)
698 __mark_reg_not_init(regs + regno);
699 return;
700 }
701 __mark_reg_unknown(regs + regno);
702}
703
704static void __mark_reg_not_init(struct bpf_reg_state *reg)
705{
706 __mark_reg_unknown(reg);
707 reg->type = NOT_INIT;
708}
709
710static void mark_reg_not_init(struct bpf_verifier_env *env,
711 struct bpf_reg_state *regs, u32 regno)
712{
713 if (WARN_ON(regno >= MAX_BPF_REG)) {
714 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
715 /* Something bad happened, let's kill all regs except FP */
716 for (regno = 0; regno < BPF_REG_FP; regno++)
717 __mark_reg_not_init(regs + regno);
718 return;
719 }
720 __mark_reg_not_init(regs + regno);
721}
722
723static void init_reg_state(struct bpf_verifier_env *env,
724 struct bpf_func_state *state)
725{
726 struct bpf_reg_state *regs = state->regs;
727 int i;
728
729 for (i = 0; i < MAX_BPF_REG; i++) {
730 mark_reg_not_init(env, regs, i);
731 regs[i].live = REG_LIVE_NONE;
732 }
733
734 /* frame pointer */
735 regs[BPF_REG_FP].type = PTR_TO_STACK;
736 mark_reg_known_zero(env, regs, BPF_REG_FP);
737 regs[BPF_REG_FP].frameno = state->frameno;
738
739 /* 1st arg to a function */
740 regs[BPF_REG_1].type = PTR_TO_CTX;
741 mark_reg_known_zero(env, regs, BPF_REG_1);
742}
743
744#define BPF_MAIN_FUNC (-1)
745static void init_func_state(struct bpf_verifier_env *env,
746 struct bpf_func_state *state,
747 int callsite, int frameno, int subprogno)
748{
749 state->callsite = callsite;
750 state->frameno = frameno;
751 state->subprogno = subprogno;
752 init_reg_state(env, state);
753}
754
755enum reg_arg_type {
756 SRC_OP, /* register is used as source operand */
757 DST_OP, /* register is used as destination operand */
758 DST_OP_NO_MARK /* same as above, check only, don't mark */
759};
760
761static int cmp_subprogs(const void *a, const void *b)
762{
763 return *(int *)a - *(int *)b;
764}
765
766static int find_subprog(struct bpf_verifier_env *env, int off)
767{
768 u32 *p;
769
770 p = bsearch(&off, env->subprog_starts, env->subprog_cnt,
771 sizeof(env->subprog_starts[0]), cmp_subprogs);
772 if (!p)
773 return -ENOENT;
774 return p - env->subprog_starts;
775
776}
777
778static int add_subprog(struct bpf_verifier_env *env, int off)
779{
780 int insn_cnt = env->prog->len;
781 int ret;
782
783 if (off >= insn_cnt || off < 0) {
784 verbose(env, "call to invalid destination\n");
785 return -EINVAL;
786 }
787 ret = find_subprog(env, off);
788 if (ret >= 0)
789 return 0;
790 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
791 verbose(env, "too many subprograms\n");
792 return -E2BIG;
793 }
794 env->subprog_starts[env->subprog_cnt++] = off;
795 sort(env->subprog_starts, env->subprog_cnt,
796 sizeof(env->subprog_starts[0]), cmp_subprogs, NULL);
797 return 0;
798}
799
800static int check_subprogs(struct bpf_verifier_env *env)
801{
802 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
803 struct bpf_insn *insn = env->prog->insnsi;
804 int insn_cnt = env->prog->len;
805
806 /* determine subprog starts. The end is one before the next starts */
807 for (i = 0; i < insn_cnt; i++) {
808 if (insn[i].code != (BPF_JMP | BPF_CALL))
809 continue;
810 if (insn[i].src_reg != BPF_PSEUDO_CALL)
811 continue;
812 if (!env->allow_ptr_leaks) {
813 verbose(env, "function calls to other bpf functions are allowed for root only\n");
814 return -EPERM;
815 }
816 if (bpf_prog_is_dev_bound(env->prog->aux)) {
817 verbose(env, "function calls in offloaded programs are not supported yet\n");
818 return -EINVAL;
819 }
820 ret = add_subprog(env, i + insn[i].imm + 1);
821 if (ret < 0)
822 return ret;
823 }
824
825 if (env->log.level > 1)
826 for (i = 0; i < env->subprog_cnt; i++)
827 verbose(env, "func#%d @%d\n", i, env->subprog_starts[i]);
828
829 /* now check that all jumps are within the same subprog */
830 subprog_start = 0;
831 if (env->subprog_cnt == cur_subprog)
832 subprog_end = insn_cnt;
833 else
834 subprog_end = env->subprog_starts[cur_subprog++];
835 for (i = 0; i < insn_cnt; i++) {
836 u8 code = insn[i].code;
837
838 if (BPF_CLASS(code) != BPF_JMP)
839 goto next;
840 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
841 goto next;
842 off = i + insn[i].off + 1;
843 if (off < subprog_start || off >= subprog_end) {
844 verbose(env, "jump out of range from insn %d to %d\n", i, off);
845 return -EINVAL;
846 }
847next:
848 if (i == subprog_end - 1) {
849 /* to avoid fall-through from one subprog into another
850 * the last insn of the subprog should be either exit
851 * or unconditional jump back
852 */
853 if (code != (BPF_JMP | BPF_EXIT) &&
854 code != (BPF_JMP | BPF_JA)) {
855 verbose(env, "last insn is not an exit or jmp\n");
856 return -EINVAL;
857 }
858 subprog_start = subprog_end;
859 if (env->subprog_cnt == cur_subprog)
860 subprog_end = insn_cnt;
861 else
862 subprog_end = env->subprog_starts[cur_subprog++];
863 }
864 }
865 return 0;
866}
867
868static
869struct bpf_verifier_state *skip_callee(struct bpf_verifier_env *env,
870 const struct bpf_verifier_state *state,
871 struct bpf_verifier_state *parent,
872 u32 regno)
873{
874 struct bpf_verifier_state *tmp = NULL;
875
876 /* 'parent' could be a state of caller and
877 * 'state' could be a state of callee. In such case
878 * parent->curframe < state->curframe
879 * and it's ok for r1 - r5 registers
880 *
881 * 'parent' could be a callee's state after it bpf_exit-ed.
882 * In such case parent->curframe > state->curframe
883 * and it's ok for r0 only
884 */
885 if (parent->curframe == state->curframe ||
886 (parent->curframe < state->curframe &&
887 regno >= BPF_REG_1 && regno <= BPF_REG_5) ||
888 (parent->curframe > state->curframe &&
889 regno == BPF_REG_0))
890 return parent;
891
892 if (parent->curframe > state->curframe &&
893 regno >= BPF_REG_6) {
894 /* for callee saved regs we have to skip the whole chain
895 * of states that belong to callee and mark as LIVE_READ
896 * the registers before the call
897 */
898 tmp = parent;
899 while (tmp && tmp->curframe != state->curframe) {
900 tmp = tmp->parent;
901 }
902 if (!tmp)
903 goto bug;
904 parent = tmp;
905 } else {
906 goto bug;
907 }
908 return parent;
909bug:
910 verbose(env, "verifier bug regno %d tmp %p\n", regno, tmp);
911 verbose(env, "regno %d parent frame %d current frame %d\n",
912 regno, parent->curframe, state->curframe);
913 return NULL;
914}
915
916static int mark_reg_read(struct bpf_verifier_env *env,
917 const struct bpf_verifier_state *state,
918 struct bpf_verifier_state *parent,
919 u32 regno)
920{
921 bool writes = parent == state->parent; /* Observe write marks */
922
923 if (regno == BPF_REG_FP)
924 /* We don't need to worry about FP liveness because it's read-only */
925 return 0;
926
927 while (parent) {
928 /* if read wasn't screened by an earlier write ... */
929 if (writes && state->frame[state->curframe]->regs[regno].live & REG_LIVE_WRITTEN)
930 break;
931 parent = skip_callee(env, state, parent, regno);
932 if (!parent)
933 return -EFAULT;
934 /* ... then we depend on parent's value */
935 parent->frame[parent->curframe]->regs[regno].live |= REG_LIVE_READ;
936 state = parent;
937 parent = state->parent;
938 writes = true;
939 }
940 return 0;
941}
942
943static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
944 enum reg_arg_type t)
945{
946 struct bpf_verifier_state *vstate = env->cur_state;
947 struct bpf_func_state *state = vstate->frame[vstate->curframe];
948 struct bpf_reg_state *regs = state->regs;
949
950 if (regno >= MAX_BPF_REG) {
951 verbose(env, "R%d is invalid\n", regno);
952 return -EINVAL;
953 }
954
955 if (t == SRC_OP) {
956 /* check whether register used as source operand can be read */
957 if (regs[regno].type == NOT_INIT) {
958 verbose(env, "R%d !read_ok\n", regno);
959 return -EACCES;
960 }
961 return mark_reg_read(env, vstate, vstate->parent, regno);
962 } else {
963 /* check whether register used as dest operand can be written to */
964 if (regno == BPF_REG_FP) {
965 verbose(env, "frame pointer is read only\n");
966 return -EACCES;
967 }
968 regs[regno].live |= REG_LIVE_WRITTEN;
969 if (t == DST_OP)
970 mark_reg_unknown(env, regs, regno);
971 }
972 return 0;
973}
974
975static bool is_spillable_regtype(enum bpf_reg_type type)
976{
977 switch (type) {
978 case PTR_TO_MAP_VALUE:
979 case PTR_TO_MAP_VALUE_OR_NULL:
980 case PTR_TO_STACK:
981 case PTR_TO_CTX:
982 case PTR_TO_PACKET:
983 case PTR_TO_PACKET_META:
984 case PTR_TO_PACKET_END:
985 case CONST_PTR_TO_MAP:
986 return true;
987 default:
988 return false;
989 }
990}
991
992/* Does this register contain a constant zero? */
993static bool register_is_null(struct bpf_reg_state *reg)
994{
995 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
996}
997
998/* check_stack_read/write functions track spill/fill of registers,
999 * stack boundary and alignment are checked in check_mem_access()
1000 */
1001static int check_stack_write(struct bpf_verifier_env *env,
1002 struct bpf_func_state *state, /* func where register points to */
1003 int off, int size, int value_regno, int insn_idx)
1004{
1005 struct bpf_func_state *cur; /* state of the current function */
1006 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1007 enum bpf_reg_type type;
1008
1009 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1010 true);
1011 if (err)
1012 return err;
1013 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1014 * so it's aligned access and [off, off + size) are within stack limits
1015 */
1016 if (!env->allow_ptr_leaks &&
1017 state->stack[spi].slot_type[0] == STACK_SPILL &&
1018 size != BPF_REG_SIZE) {
1019 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1020 return -EACCES;
1021 }
1022
1023 cur = env->cur_state->frame[env->cur_state->curframe];
1024 if (value_regno >= 0 &&
1025 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1026
1027 /* register containing pointer is being spilled into stack */
1028 if (size != BPF_REG_SIZE) {
1029 verbose(env, "invalid size of register spill\n");
1030 return -EACCES;
1031 }
1032
1033 if (state != cur && type == PTR_TO_STACK) {
1034 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1035 return -EINVAL;
1036 }
1037
1038 /* save register state */
1039 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1040 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1041
1042 for (i = 0; i < BPF_REG_SIZE; i++) {
1043 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1044 !env->allow_ptr_leaks) {
1045 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1046 int soff = (-spi - 1) * BPF_REG_SIZE;
1047
1048 /* detected reuse of integer stack slot with a pointer
1049 * which means either llvm is reusing stack slot or
1050 * an attacker is trying to exploit CVE-2018-3639
1051 * (speculative store bypass)
1052 * Have to sanitize that slot with preemptive
1053 * store of zero.
1054 */
1055 if (*poff && *poff != soff) {
1056 /* disallow programs where single insn stores
1057 * into two different stack slots, since verifier
1058 * cannot sanitize them
1059 */
1060 verbose(env,
1061 "insn %d cannot access two stack slots fp%d and fp%d",
1062 insn_idx, *poff, soff);
1063 return -EINVAL;
1064 }
1065 *poff = soff;
1066 }
1067 state->stack[spi].slot_type[i] = STACK_SPILL;
1068 }
1069 } else {
1070 u8 type = STACK_MISC;
1071
1072 /* regular write of data into stack */
1073 state->stack[spi].spilled_ptr = (struct bpf_reg_state) {};
1074
1075 /* only mark the slot as written if all 8 bytes were written
1076 * otherwise read propagation may incorrectly stop too soon
1077 * when stack slots are partially written.
1078 * This heuristic means that read propagation will be
1079 * conservative, since it will add reg_live_read marks
1080 * to stack slots all the way to first state when programs
1081 * writes+reads less than 8 bytes
1082 */
1083 if (size == BPF_REG_SIZE)
1084 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1085
1086 /* when we zero initialize stack slots mark them as such */
1087 if (value_regno >= 0 &&
1088 register_is_null(&cur->regs[value_regno]))
1089 type = STACK_ZERO;
1090
1091 for (i = 0; i < size; i++)
1092 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1093 type;
1094 }
1095 return 0;
1096}
1097
1098/* registers of every function are unique and mark_reg_read() propagates
1099 * the liveness in the following cases:
1100 * - from callee into caller for R1 - R5 that were used as arguments
1101 * - from caller into callee for R0 that used as result of the call
1102 * - from caller to the same caller skipping states of the callee for R6 - R9,
1103 * since R6 - R9 are callee saved by implicit function prologue and
1104 * caller's R6 != callee's R6, so when we propagate liveness up to
1105 * parent states we need to skip callee states for R6 - R9.
1106 *
1107 * stack slot marking is different, since stacks of caller and callee are
1108 * accessible in both (since caller can pass a pointer to caller's stack to
1109 * callee which can pass it to another function), hence mark_stack_slot_read()
1110 * has to propagate the stack liveness to all parent states at given frame number.
1111 * Consider code:
1112 * f1() {
1113 * ptr = fp - 8;
1114 * *ptr = ctx;
1115 * call f2 {
1116 * .. = *ptr;
1117 * }
1118 * .. = *ptr;
1119 * }
1120 * First *ptr is reading from f1's stack and mark_stack_slot_read() has
1121 * to mark liveness at the f1's frame and not f2's frame.
1122 * Second *ptr is also reading from f1's stack and mark_stack_slot_read() has
1123 * to propagate liveness to f2 states at f1's frame level and further into
1124 * f1 states at f1's frame level until write into that stack slot
1125 */
1126static void mark_stack_slot_read(struct bpf_verifier_env *env,
1127 const struct bpf_verifier_state *state,
1128 struct bpf_verifier_state *parent,
1129 int slot, int frameno)
1130{
1131 bool writes = parent == state->parent; /* Observe write marks */
1132
1133 while (parent) {
1134 if (parent->frame[frameno]->allocated_stack <= slot * BPF_REG_SIZE)
1135 /* since LIVE_WRITTEN mark is only done for full 8-byte
1136 * write the read marks are conservative and parent
1137 * state may not even have the stack allocated. In such case
1138 * end the propagation, since the loop reached beginning
1139 * of the function
1140 */
1141 break;
1142 /* if read wasn't screened by an earlier write ... */
1143 if (writes && state->frame[frameno]->stack[slot].spilled_ptr.live & REG_LIVE_WRITTEN)
1144 break;
1145 /* ... then we depend on parent's value */
1146 parent->frame[frameno]->stack[slot].spilled_ptr.live |= REG_LIVE_READ;
1147 state = parent;
1148 parent = state->parent;
1149 writes = true;
1150 }
1151}
1152
1153static int check_stack_read(struct bpf_verifier_env *env,
1154 struct bpf_func_state *reg_state /* func where register points to */,
1155 int off, int size, int value_regno)
1156{
1157 struct bpf_verifier_state *vstate = env->cur_state;
1158 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1159 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1160 u8 *stype;
1161
1162 if (reg_state->allocated_stack <= slot) {
1163 verbose(env, "invalid read from stack off %d+0 size %d\n",
1164 off, size);
1165 return -EACCES;
1166 }
1167 stype = reg_state->stack[spi].slot_type;
1168
1169 if (stype[0] == STACK_SPILL) {
1170 if (size != BPF_REG_SIZE) {
1171 verbose(env, "invalid size of register spill\n");
1172 return -EACCES;
1173 }
1174 for (i = 1; i < BPF_REG_SIZE; i++) {
1175 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1176 verbose(env, "corrupted spill memory\n");
1177 return -EACCES;
1178 }
1179 }
1180
1181 if (value_regno >= 0) {
1182 /* restore register state from stack */
1183 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1184 /* mark reg as written since spilled pointer state likely
1185 * has its liveness marks cleared by is_state_visited()
1186 * which resets stack/reg liveness for state transitions
1187 */
1188 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1189 }
1190 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1191 reg_state->frameno);
1192 return 0;
1193 } else {
1194 int zeros = 0;
1195
1196 for (i = 0; i < size; i++) {
1197 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1198 continue;
1199 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1200 zeros++;
1201 continue;
1202 }
1203 verbose(env, "invalid read from stack off %d+%d size %d\n",
1204 off, i, size);
1205 return -EACCES;
1206 }
1207 mark_stack_slot_read(env, vstate, vstate->parent, spi,
1208 reg_state->frameno);
1209 if (value_regno >= 0) {
1210 if (zeros == size) {
1211 /* any size read into register is zero extended,
1212 * so the whole register == const_zero
1213 */
1214 __mark_reg_const_zero(&state->regs[value_regno]);
1215 } else {
1216 /* have read misc data from the stack */
1217 mark_reg_unknown(env, state->regs, value_regno);
1218 }
1219 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1220 }
1221 return 0;
1222 }
1223}
1224
1225/* check read/write into map element returned by bpf_map_lookup_elem() */
1226static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1227 int size, bool zero_size_allowed)
1228{
1229 struct bpf_reg_state *regs = cur_regs(env);
1230 struct bpf_map *map = regs[regno].map_ptr;
1231
1232 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1233 off + size > map->value_size) {
1234 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1235 map->value_size, off, size);
1236 return -EACCES;
1237 }
1238 return 0;
1239}
1240
1241/* check read/write into a map element with possible variable offset */
1242static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1243 int off, int size, bool zero_size_allowed)
1244{
1245 struct bpf_verifier_state *vstate = env->cur_state;
1246 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1247 struct bpf_reg_state *reg = &state->regs[regno];
1248 int err;
1249
1250 /* We may have adjusted the register to this map value, so we
1251 * need to try adding each of min_value and max_value to off
1252 * to make sure our theoretical access will be safe.
1253 */
1254 if (env->log.level)
1255 print_verifier_state(env, state);
1256 /* The minimum value is only important with signed
1257 * comparisons where we can't assume the floor of a
1258 * value is 0. If we are using signed variables for our
1259 * index'es we need to make sure that whatever we use
1260 * will have a set floor within our range.
1261 */
1262 if (reg->smin_value < 0) {
1263 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1264 regno);
1265 return -EACCES;
1266 }
1267 err = __check_map_access(env, regno, reg->smin_value + off, size,
1268 zero_size_allowed);
1269 if (err) {
1270 verbose(env, "R%d min value is outside of the array range\n",
1271 regno);
1272 return err;
1273 }
1274
1275 /* If we haven't set a max value then we need to bail since we can't be
1276 * sure we won't do bad things.
1277 * If reg->umax_value + off could overflow, treat that as unbounded too.
1278 */
1279 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1280 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1281 regno);
1282 return -EACCES;
1283 }
1284 err = __check_map_access(env, regno, reg->umax_value + off, size,
1285 zero_size_allowed);
1286 if (err)
1287 verbose(env, "R%d max value is outside of the array range\n",
1288 regno);
1289 return err;
1290}
1291
1292#define MAX_PACKET_OFF 0xffff
1293
1294static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1295 const struct bpf_call_arg_meta *meta,
1296 enum bpf_access_type t)
1297{
1298 switch (env->prog->type) {
1299 case BPF_PROG_TYPE_LWT_IN:
1300 case BPF_PROG_TYPE_LWT_OUT:
1301 /* dst_input() and dst_output() can't write for now */
1302 if (t == BPF_WRITE)
1303 return false;
1304 /* fallthrough */
1305 case BPF_PROG_TYPE_SCHED_CLS:
1306 case BPF_PROG_TYPE_SCHED_ACT:
1307 case BPF_PROG_TYPE_XDP:
1308 case BPF_PROG_TYPE_LWT_XMIT:
1309 case BPF_PROG_TYPE_SK_SKB:
1310 case BPF_PROG_TYPE_SK_MSG:
1311 if (meta)
1312 return meta->pkt_access;
1313
1314 env->seen_direct_write = true;
1315 return true;
1316 default:
1317 return false;
1318 }
1319}
1320
1321static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1322 int off, int size, bool zero_size_allowed)
1323{
1324 struct bpf_reg_state *regs = cur_regs(env);
1325 struct bpf_reg_state *reg = ®s[regno];
1326
1327 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1328 (u64)off + size > reg->range) {
1329 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1330 off, size, regno, reg->id, reg->off, reg->range);
1331 return -EACCES;
1332 }
1333 return 0;
1334}
1335
1336static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1337 int size, bool zero_size_allowed)
1338{
1339 struct bpf_reg_state *regs = cur_regs(env);
1340 struct bpf_reg_state *reg = ®s[regno];
1341 int err;
1342
1343 /* We may have added a variable offset to the packet pointer; but any
1344 * reg->range we have comes after that. We are only checking the fixed
1345 * offset.
1346 */
1347
1348 /* We don't allow negative numbers, because we aren't tracking enough
1349 * detail to prove they're safe.
1350 */
1351 if (reg->smin_value < 0) {
1352 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1353 regno);
1354 return -EACCES;
1355 }
1356 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1357 if (err) {
1358 verbose(env, "R%d offset is outside of the packet\n", regno);
1359 return err;
1360 }
1361 return err;
1362}
1363
1364/* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1365static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1366 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1367{
1368 struct bpf_insn_access_aux info = {
1369 .reg_type = *reg_type,
1370 };
1371
1372 if (env->ops->is_valid_access &&
1373 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1374 /* A non zero info.ctx_field_size indicates that this field is a
1375 * candidate for later verifier transformation to load the whole
1376 * field and then apply a mask when accessed with a narrower
1377 * access than actual ctx access size. A zero info.ctx_field_size
1378 * will only allow for whole field access and rejects any other
1379 * type of narrower access.
1380 */
1381 *reg_type = info.reg_type;
1382
1383 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1384 /* remember the offset of last byte accessed in ctx */
1385 if (env->prog->aux->max_ctx_offset < off + size)
1386 env->prog->aux->max_ctx_offset = off + size;
1387 return 0;
1388 }
1389
1390 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1391 return -EACCES;
1392}
1393
1394static bool __is_pointer_value(bool allow_ptr_leaks,
1395 const struct bpf_reg_state *reg)
1396{
1397 if (allow_ptr_leaks)
1398 return false;
1399
1400 return reg->type != SCALAR_VALUE;
1401}
1402
1403static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1404{
1405 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1406}
1407
1408static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1409{
1410 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1411
1412 return reg->type == PTR_TO_CTX;
1413}
1414
1415static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1416{
1417 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1418
1419 return type_is_pkt_pointer(reg->type);
1420}
1421
1422static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1423 const struct bpf_reg_state *reg,
1424 int off, int size, bool strict)
1425{
1426 struct tnum reg_off;
1427 int ip_align;
1428
1429 /* Byte size accesses are always allowed. */
1430 if (!strict || size == 1)
1431 return 0;
1432
1433 /* For platforms that do not have a Kconfig enabling
1434 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1435 * NET_IP_ALIGN is universally set to '2'. And on platforms
1436 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1437 * to this code only in strict mode where we want to emulate
1438 * the NET_IP_ALIGN==2 checking. Therefore use an
1439 * unconditional IP align value of '2'.
1440 */
1441 ip_align = 2;
1442
1443 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1444 if (!tnum_is_aligned(reg_off, size)) {
1445 char tn_buf[48];
1446
1447 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1448 verbose(env,
1449 "misaligned packet access off %d+%s+%d+%d size %d\n",
1450 ip_align, tn_buf, reg->off, off, size);
1451 return -EACCES;
1452 }
1453
1454 return 0;
1455}
1456
1457static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1458 const struct bpf_reg_state *reg,
1459 const char *pointer_desc,
1460 int off, int size, bool strict)
1461{
1462 struct tnum reg_off;
1463
1464 /* Byte size accesses are always allowed. */
1465 if (!strict || size == 1)
1466 return 0;
1467
1468 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1469 if (!tnum_is_aligned(reg_off, size)) {
1470 char tn_buf[48];
1471
1472 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1473 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1474 pointer_desc, tn_buf, reg->off, off, size);
1475 return -EACCES;
1476 }
1477
1478 return 0;
1479}
1480
1481static int check_ptr_alignment(struct bpf_verifier_env *env,
1482 const struct bpf_reg_state *reg, int off,
1483 int size, bool strict_alignment_once)
1484{
1485 bool strict = env->strict_alignment || strict_alignment_once;
1486 const char *pointer_desc = "";
1487
1488 switch (reg->type) {
1489 case PTR_TO_PACKET:
1490 case PTR_TO_PACKET_META:
1491 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1492 * right in front, treat it the very same way.
1493 */
1494 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1495 case PTR_TO_MAP_VALUE:
1496 pointer_desc = "value ";
1497 break;
1498 case PTR_TO_CTX:
1499 pointer_desc = "context ";
1500 break;
1501 case PTR_TO_STACK:
1502 pointer_desc = "stack ";
1503 /* The stack spill tracking logic in check_stack_write()
1504 * and check_stack_read() relies on stack accesses being
1505 * aligned.
1506 */
1507 strict = true;
1508 break;
1509 default:
1510 break;
1511 }
1512 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1513 strict);
1514}
1515
1516static int update_stack_depth(struct bpf_verifier_env *env,
1517 const struct bpf_func_state *func,
1518 int off)
1519{
1520 u16 stack = env->subprog_stack_depth[func->subprogno];
1521
1522 if (stack >= -off)
1523 return 0;
1524
1525 /* update known max for given subprogram */
1526 env->subprog_stack_depth[func->subprogno] = -off;
1527 return 0;
1528}
1529
1530/* starting from main bpf function walk all instructions of the function
1531 * and recursively walk all callees that given function can call.
1532 * Ignore jump and exit insns.
1533 * Since recursion is prevented by check_cfg() this algorithm
1534 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1535 */
1536static int check_max_stack_depth(struct bpf_verifier_env *env)
1537{
1538 int depth = 0, frame = 0, subprog = 0, i = 0, subprog_end;
1539 struct bpf_insn *insn = env->prog->insnsi;
1540 int insn_cnt = env->prog->len;
1541 int ret_insn[MAX_CALL_FRAMES];
1542 int ret_prog[MAX_CALL_FRAMES];
1543
1544process_func:
1545 /* round up to 32-bytes, since this is granularity
1546 * of interpreter stack size
1547 */
1548 depth += round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1549 if (depth > MAX_BPF_STACK) {
1550 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1551 frame + 1, depth);
1552 return -EACCES;
1553 }
1554continue_func:
1555 if (env->subprog_cnt == subprog)
1556 subprog_end = insn_cnt;
1557 else
1558 subprog_end = env->subprog_starts[subprog];
1559 for (; i < subprog_end; i++) {
1560 if (insn[i].code != (BPF_JMP | BPF_CALL))
1561 continue;
1562 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1563 continue;
1564 /* remember insn and function to return to */
1565 ret_insn[frame] = i + 1;
1566 ret_prog[frame] = subprog;
1567
1568 /* find the callee */
1569 i = i + insn[i].imm + 1;
1570 subprog = find_subprog(env, i);
1571 if (subprog < 0) {
1572 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1573 i);
1574 return -EFAULT;
1575 }
1576 subprog++;
1577 frame++;
1578 if (frame >= MAX_CALL_FRAMES) {
1579 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1580 return -EFAULT;
1581 }
1582 goto process_func;
1583 }
1584 /* end of for() loop means the last insn of the 'subprog'
1585 * was reached. Doesn't matter whether it was JA or EXIT
1586 */
1587 if (frame == 0)
1588 return 0;
1589 depth -= round_up(max_t(u32, env->subprog_stack_depth[subprog], 1), 32);
1590 frame--;
1591 i = ret_insn[frame];
1592 subprog = ret_prog[frame];
1593 goto continue_func;
1594}
1595
1596#ifndef CONFIG_BPF_JIT_ALWAYS_ON
1597static int get_callee_stack_depth(struct bpf_verifier_env *env,
1598 const struct bpf_insn *insn, int idx)
1599{
1600 int start = idx + insn->imm + 1, subprog;
1601
1602 subprog = find_subprog(env, start);
1603 if (subprog < 0) {
1604 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1605 start);
1606 return -EFAULT;
1607 }
1608 subprog++;
1609 return env->subprog_stack_depth[subprog];
1610}
1611#endif
1612
1613/* truncate register to smaller size (in bytes)
1614 * must be called with size < BPF_REG_SIZE
1615 */
1616static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1617{
1618 u64 mask;
1619
1620 /* clear high bits in bit representation */
1621 reg->var_off = tnum_cast(reg->var_off, size);
1622
1623 /* fix arithmetic bounds */
1624 mask = ((u64)1 << (size * 8)) - 1;
1625 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1626 reg->umin_value &= mask;
1627 reg->umax_value &= mask;
1628 } else {
1629 reg->umin_value = 0;
1630 reg->umax_value = mask;
1631 }
1632 reg->smin_value = reg->umin_value;
1633 reg->smax_value = reg->umax_value;
1634}
1635
1636/* check whether memory at (regno + off) is accessible for t = (read | write)
1637 * if t==write, value_regno is a register which value is stored into memory
1638 * if t==read, value_regno is a register which will receive the value from memory
1639 * if t==write && value_regno==-1, some unknown value is stored into memory
1640 * if t==read && value_regno==-1, don't care what we read from memory
1641 */
1642static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1643 int off, int bpf_size, enum bpf_access_type t,
1644 int value_regno, bool strict_alignment_once)
1645{
1646 struct bpf_reg_state *regs = cur_regs(env);
1647 struct bpf_reg_state *reg = regs + regno;
1648 struct bpf_func_state *state;
1649 int size, err = 0;
1650
1651 size = bpf_size_to_bytes(bpf_size);
1652 if (size < 0)
1653 return size;
1654
1655 /* alignment checks will add in reg->off themselves */
1656 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1657 if (err)
1658 return err;
1659
1660 /* for access checks, reg->off is just part of off */
1661 off += reg->off;
1662
1663 if (reg->type == PTR_TO_MAP_VALUE) {
1664 if (t == BPF_WRITE && value_regno >= 0 &&
1665 is_pointer_value(env, value_regno)) {
1666 verbose(env, "R%d leaks addr into map\n", value_regno);
1667 return -EACCES;
1668 }
1669
1670 err = check_map_access(env, regno, off, size, false);
1671 if (!err && t == BPF_READ && value_regno >= 0)
1672 mark_reg_unknown(env, regs, value_regno);
1673
1674 } else if (reg->type == PTR_TO_CTX) {
1675 enum bpf_reg_type reg_type = SCALAR_VALUE;
1676
1677 if (t == BPF_WRITE && value_regno >= 0 &&
1678 is_pointer_value(env, value_regno)) {
1679 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1680 return -EACCES;
1681 }
1682 /* ctx accesses must be at a fixed offset, so that we can
1683 * determine what type of data were returned.
1684 */
1685 if (reg->off) {
1686 verbose(env,
1687 "dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
1688 regno, reg->off, off - reg->off);
1689 return -EACCES;
1690 }
1691 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1692 char tn_buf[48];
1693
1694 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1695 verbose(env,
1696 "variable ctx access var_off=%s off=%d size=%d",
1697 tn_buf, off, size);
1698 return -EACCES;
1699 }
1700 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1701 if (!err && t == BPF_READ && value_regno >= 0) {
1702 /* ctx access returns either a scalar, or a
1703 * PTR_TO_PACKET[_META,_END]. In the latter
1704 * case, we know the offset is zero.
1705 */
1706 if (reg_type == SCALAR_VALUE)
1707 mark_reg_unknown(env, regs, value_regno);
1708 else
1709 mark_reg_known_zero(env, regs,
1710 value_regno);
1711 regs[value_regno].id = 0;
1712 regs[value_regno].off = 0;
1713 regs[value_regno].range = 0;
1714 regs[value_regno].type = reg_type;
1715 }
1716
1717 } else if (reg->type == PTR_TO_STACK) {
1718 /* stack accesses must be at a fixed offset, so that we can
1719 * determine what type of data were returned.
1720 * See check_stack_read().
1721 */
1722 if (!tnum_is_const(reg->var_off)) {
1723 char tn_buf[48];
1724
1725 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1726 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1727 tn_buf, off, size);
1728 return -EACCES;
1729 }
1730 off += reg->var_off.value;
1731 if (off >= 0 || off < -MAX_BPF_STACK) {
1732 verbose(env, "invalid stack off=%d size=%d\n", off,
1733 size);
1734 return -EACCES;
1735 }
1736
1737 state = func(env, reg);
1738 err = update_stack_depth(env, state, off);
1739 if (err)
1740 return err;
1741
1742 if (t == BPF_WRITE)
1743 err = check_stack_write(env, state, off, size,
1744 value_regno, insn_idx);
1745 else
1746 err = check_stack_read(env, state, off, size,
1747 value_regno);
1748 } else if (reg_is_pkt_pointer(reg)) {
1749 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1750 verbose(env, "cannot write into packet\n");
1751 return -EACCES;
1752 }
1753 if (t == BPF_WRITE && value_regno >= 0 &&
1754 is_pointer_value(env, value_regno)) {
1755 verbose(env, "R%d leaks addr into packet\n",
1756 value_regno);
1757 return -EACCES;
1758 }
1759 err = check_packet_access(env, regno, off, size, false);
1760 if (!err && t == BPF_READ && value_regno >= 0)
1761 mark_reg_unknown(env, regs, value_regno);
1762 } else {
1763 verbose(env, "R%d invalid mem access '%s'\n", regno,
1764 reg_type_str[reg->type]);
1765 return -EACCES;
1766 }
1767
1768 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1769 regs[value_regno].type == SCALAR_VALUE) {
1770 /* b/h/w load zero-extends, mark upper bits as known 0 */
1771 coerce_reg_to_size(®s[value_regno], size);
1772 }
1773 return err;
1774}
1775
1776static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1777{
1778 int err;
1779
1780 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1781 insn->imm != 0) {
1782 verbose(env, "BPF_XADD uses reserved fields\n");
1783 return -EINVAL;
1784 }
1785
1786 /* check src1 operand */
1787 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1788 if (err)
1789 return err;
1790
1791 /* check src2 operand */
1792 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1793 if (err)
1794 return err;
1795
1796 if (is_pointer_value(env, insn->src_reg)) {
1797 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1798 return -EACCES;
1799 }
1800
1801 if (is_ctx_reg(env, insn->dst_reg) ||
1802 is_pkt_reg(env, insn->dst_reg)) {
1803 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1804 insn->dst_reg, is_ctx_reg(env, insn->dst_reg) ?
1805 "context" : "packet");
1806 return -EACCES;
1807 }
1808
1809 /* check whether atomic_add can read the memory */
1810 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1811 BPF_SIZE(insn->code), BPF_READ, -1, true);
1812 if (err)
1813 return err;
1814
1815 /* check whether atomic_add can write into the same memory */
1816 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1817 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1818}
1819
1820/* when register 'regno' is passed into function that will read 'access_size'
1821 * bytes from that pointer, make sure that it's within stack boundary
1822 * and all elements of stack are initialized.
1823 * Unlike most pointer bounds-checking functions, this one doesn't take an
1824 * 'off' argument, so it has to add in reg->off itself.
1825 */
1826static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1827 int access_size, bool zero_size_allowed,
1828 struct bpf_call_arg_meta *meta)
1829{
1830 struct bpf_reg_state *reg = cur_regs(env) + regno;
1831 struct bpf_func_state *state = func(env, reg);
1832 int off, i, slot, spi;
1833
1834 if (reg->type != PTR_TO_STACK) {
1835 /* Allow zero-byte read from NULL, regardless of pointer type */
1836 if (zero_size_allowed && access_size == 0 &&
1837 register_is_null(reg))
1838 return 0;
1839
1840 verbose(env, "R%d type=%s expected=%s\n", regno,
1841 reg_type_str[reg->type],
1842 reg_type_str[PTR_TO_STACK]);
1843 return -EACCES;
1844 }
1845
1846 /* Only allow fixed-offset stack reads */
1847 if (!tnum_is_const(reg->var_off)) {
1848 char tn_buf[48];
1849
1850 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1851 verbose(env, "invalid variable stack read R%d var_off=%s\n",
1852 regno, tn_buf);
1853 return -EACCES;
1854 }
1855 off = reg->off + reg->var_off.value;
1856 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
1857 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
1858 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
1859 regno, off, access_size);
1860 return -EACCES;
1861 }
1862
1863 if (meta && meta->raw_mode) {
1864 meta->access_size = access_size;
1865 meta->regno = regno;
1866 return 0;
1867 }
1868
1869 for (i = 0; i < access_size; i++) {
1870 u8 *stype;
1871
1872 slot = -(off + i) - 1;
1873 spi = slot / BPF_REG_SIZE;
1874 if (state->allocated_stack <= slot)
1875 goto err;
1876 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
1877 if (*stype == STACK_MISC)
1878 goto mark;
1879 if (*stype == STACK_ZERO) {
1880 /* helper can write anything into the stack */
1881 *stype = STACK_MISC;
1882 goto mark;
1883 }
1884err:
1885 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
1886 off, i, access_size);
1887 return -EACCES;
1888mark:
1889 /* reading any byte out of 8-byte 'spill_slot' will cause
1890 * the whole slot to be marked as 'read'
1891 */
1892 mark_stack_slot_read(env, env->cur_state, env->cur_state->parent,
1893 spi, state->frameno);
1894 }
1895 return update_stack_depth(env, state, off);
1896}
1897
1898static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
1899 int access_size, bool zero_size_allowed,
1900 struct bpf_call_arg_meta *meta)
1901{
1902 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1903
1904 switch (reg->type) {
1905 case PTR_TO_PACKET:
1906 case PTR_TO_PACKET_META:
1907 return check_packet_access(env, regno, reg->off, access_size,
1908 zero_size_allowed);
1909 case PTR_TO_MAP_VALUE:
1910 return check_map_access(env, regno, reg->off, access_size,
1911 zero_size_allowed);
1912 default: /* scalar_value|ptr_to_stack or invalid ptr */
1913 return check_stack_boundary(env, regno, access_size,
1914 zero_size_allowed, meta);
1915 }
1916}
1917
1918static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
1919{
1920 return type == ARG_PTR_TO_MEM ||
1921 type == ARG_PTR_TO_MEM_OR_NULL ||
1922 type == ARG_PTR_TO_UNINIT_MEM;
1923}
1924
1925static bool arg_type_is_mem_size(enum bpf_arg_type type)
1926{
1927 return type == ARG_CONST_SIZE ||
1928 type == ARG_CONST_SIZE_OR_ZERO;
1929}
1930
1931static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
1932 enum bpf_arg_type arg_type,
1933 struct bpf_call_arg_meta *meta)
1934{
1935 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
1936 enum bpf_reg_type expected_type, type = reg->type;
1937 int err = 0;
1938
1939 if (arg_type == ARG_DONTCARE)
1940 return 0;
1941
1942 err = check_reg_arg(env, regno, SRC_OP);
1943 if (err)
1944 return err;
1945
1946 if (arg_type == ARG_ANYTHING) {
1947 if (is_pointer_value(env, regno)) {
1948 verbose(env, "R%d leaks addr into helper function\n",
1949 regno);
1950 return -EACCES;
1951 }
1952 return 0;
1953 }
1954
1955 if (type_is_pkt_pointer(type) &&
1956 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
1957 verbose(env, "helper access to the packet is not allowed\n");
1958 return -EACCES;
1959 }
1960
1961 if (arg_type == ARG_PTR_TO_MAP_KEY ||
1962 arg_type == ARG_PTR_TO_MAP_VALUE) {
1963 expected_type = PTR_TO_STACK;
1964 if (!type_is_pkt_pointer(type) &&
1965 type != expected_type)
1966 goto err_type;
1967 } else if (arg_type == ARG_CONST_SIZE ||
1968 arg_type == ARG_CONST_SIZE_OR_ZERO) {
1969 expected_type = SCALAR_VALUE;
1970 if (type != expected_type)
1971 goto err_type;
1972 } else if (arg_type == ARG_CONST_MAP_PTR) {
1973 expected_type = CONST_PTR_TO_MAP;
1974 if (type != expected_type)
1975 goto err_type;
1976 } else if (arg_type == ARG_PTR_TO_CTX) {
1977 expected_type = PTR_TO_CTX;
1978 if (type != expected_type)
1979 goto err_type;
1980 } else if (arg_type_is_mem_ptr(arg_type)) {
1981 expected_type = PTR_TO_STACK;
1982 /* One exception here. In case function allows for NULL to be
1983 * passed in as argument, it's a SCALAR_VALUE type. Final test
1984 * happens during stack boundary checking.
1985 */
1986 if (register_is_null(reg) &&
1987 arg_type == ARG_PTR_TO_MEM_OR_NULL)
1988 /* final test in check_stack_boundary() */;
1989 else if (!type_is_pkt_pointer(type) &&
1990 type != PTR_TO_MAP_VALUE &&
1991 type != expected_type)
1992 goto err_type;
1993 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
1994 } else {
1995 verbose(env, "unsupported arg_type %d\n", arg_type);
1996 return -EFAULT;
1997 }
1998
1999 if (arg_type == ARG_CONST_MAP_PTR) {
2000 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2001 meta->map_ptr = reg->map_ptr;
2002 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2003 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2004 * check that [key, key + map->key_size) are within
2005 * stack limits and initialized
2006 */
2007 if (!meta->map_ptr) {
2008 /* in function declaration map_ptr must come before
2009 * map_key, so that it's verified and known before
2010 * we have to check map_key here. Otherwise it means
2011 * that kernel subsystem misconfigured verifier
2012 */
2013 verbose(env, "invalid map_ptr to access map->key\n");
2014 return -EACCES;
2015 }
2016 if (type_is_pkt_pointer(type))
2017 err = check_packet_access(env, regno, reg->off,
2018 meta->map_ptr->key_size,
2019 false);
2020 else
2021 err = check_stack_boundary(env, regno,
2022 meta->map_ptr->key_size,
2023 false, NULL);
2024 } else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
2025 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2026 * check [value, value + map->value_size) validity
2027 */
2028 if (!meta->map_ptr) {
2029 /* kernel subsystem misconfigured verifier */
2030 verbose(env, "invalid map_ptr to access map->value\n");
2031 return -EACCES;
2032 }
2033 if (type_is_pkt_pointer(type))
2034 err = check_packet_access(env, regno, reg->off,
2035 meta->map_ptr->value_size,
2036 false);
2037 else
2038 err = check_stack_boundary(env, regno,
2039 meta->map_ptr->value_size,
2040 false, NULL);
2041 } else if (arg_type_is_mem_size(arg_type)) {
2042 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2043
2044 /* The register is SCALAR_VALUE; the access check
2045 * happens using its boundaries.
2046 */
2047 if (!tnum_is_const(reg->var_off))
2048 /* For unprivileged variable accesses, disable raw
2049 * mode so that the program is required to
2050 * initialize all the memory that the helper could
2051 * just partially fill up.
2052 */
2053 meta = NULL;
2054
2055 if (reg->smin_value < 0) {
2056 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2057 regno);
2058 return -EACCES;
2059 }
2060
2061 if (reg->umin_value == 0) {
2062 err = check_helper_mem_access(env, regno - 1, 0,
2063 zero_size_allowed,
2064 meta);
2065 if (err)
2066 return err;
2067 }
2068
2069 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2070 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2071 regno);
2072 return -EACCES;
2073 }
2074 err = check_helper_mem_access(env, regno - 1,
2075 reg->umax_value,
2076 zero_size_allowed, meta);
2077 }
2078
2079 return err;
2080err_type:
2081 verbose(env, "R%d type=%s expected=%s\n", regno,
2082 reg_type_str[type], reg_type_str[expected_type]);
2083 return -EACCES;
2084}
2085
2086static int check_map_func_compatibility(struct bpf_verifier_env *env,
2087 struct bpf_map *map, int func_id)
2088{
2089 if (!map)
2090 return 0;
2091
2092 /* We need a two way check, first is from map perspective ... */
2093 switch (map->map_type) {
2094 case BPF_MAP_TYPE_PROG_ARRAY:
2095 if (func_id != BPF_FUNC_tail_call)
2096 goto error;
2097 break;
2098 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2099 if (func_id != BPF_FUNC_perf_event_read &&
2100 func_id != BPF_FUNC_perf_event_output &&
2101 func_id != BPF_FUNC_perf_event_read_value)
2102 goto error;
2103 break;
2104 case BPF_MAP_TYPE_STACK_TRACE:
2105 if (func_id != BPF_FUNC_get_stackid)
2106 goto error;
2107 break;
2108 case BPF_MAP_TYPE_CGROUP_ARRAY:
2109 if (func_id != BPF_FUNC_skb_under_cgroup &&
2110 func_id != BPF_FUNC_current_task_under_cgroup)
2111 goto error;
2112 break;
2113 /* devmap returns a pointer to a live net_device ifindex that we cannot
2114 * allow to be modified from bpf side. So do not allow lookup elements
2115 * for now.
2116 */
2117 case BPF_MAP_TYPE_DEVMAP:
2118 if (func_id != BPF_FUNC_redirect_map)
2119 goto error;
2120 break;
2121 /* Restrict bpf side of cpumap, open when use-cases appear */
2122 case BPF_MAP_TYPE_CPUMAP:
2123 if (func_id != BPF_FUNC_redirect_map)
2124 goto error;
2125 break;
2126 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2127 case BPF_MAP_TYPE_HASH_OF_MAPS:
2128 if (func_id != BPF_FUNC_map_lookup_elem)
2129 goto error;
2130 break;
2131 case BPF_MAP_TYPE_SOCKMAP:
2132 if (func_id != BPF_FUNC_sk_redirect_map &&
2133 func_id != BPF_FUNC_sock_map_update &&
2134 func_id != BPF_FUNC_map_delete_elem &&
2135 func_id != BPF_FUNC_msg_redirect_map)
2136 goto error;
2137 break;
2138 default:
2139 break;
2140 }
2141
2142 /* ... and second from the function itself. */
2143 switch (func_id) {
2144 case BPF_FUNC_tail_call:
2145 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2146 goto error;
2147 if (env->subprog_cnt) {
2148 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2149 return -EINVAL;
2150 }
2151 break;
2152 case BPF_FUNC_perf_event_read:
2153 case BPF_FUNC_perf_event_output:
2154 case BPF_FUNC_perf_event_read_value:
2155 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2156 goto error;
2157 break;
2158 case BPF_FUNC_get_stackid:
2159 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2160 goto error;
2161 break;
2162 case BPF_FUNC_current_task_under_cgroup:
2163 case BPF_FUNC_skb_under_cgroup:
2164 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2165 goto error;
2166 break;
2167 case BPF_FUNC_redirect_map:
2168 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2169 map->map_type != BPF_MAP_TYPE_CPUMAP)
2170 goto error;
2171 break;
2172 case BPF_FUNC_sk_redirect_map:
2173 case BPF_FUNC_msg_redirect_map:
2174 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2175 goto error;
2176 break;
2177 case BPF_FUNC_sock_map_update:
2178 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2179 goto error;
2180 break;
2181 default:
2182 break;
2183 }
2184
2185 return 0;
2186error:
2187 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2188 map->map_type, func_id_name(func_id), func_id);
2189 return -EINVAL;
2190}
2191
2192static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2193{
2194 int count = 0;
2195
2196 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2197 count++;
2198 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2199 count++;
2200 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2201 count++;
2202 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2203 count++;
2204 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2205 count++;
2206
2207 /* We only support one arg being in raw mode at the moment,
2208 * which is sufficient for the helper functions we have
2209 * right now.
2210 */
2211 return count <= 1;
2212}
2213
2214static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2215 enum bpf_arg_type arg_next)
2216{
2217 return (arg_type_is_mem_ptr(arg_curr) &&
2218 !arg_type_is_mem_size(arg_next)) ||
2219 (!arg_type_is_mem_ptr(arg_curr) &&
2220 arg_type_is_mem_size(arg_next));
2221}
2222
2223static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2224{
2225 /* bpf_xxx(..., buf, len) call will access 'len'
2226 * bytes from memory 'buf'. Both arg types need
2227 * to be paired, so make sure there's no buggy
2228 * helper function specification.
2229 */
2230 if (arg_type_is_mem_size(fn->arg1_type) ||
2231 arg_type_is_mem_ptr(fn->arg5_type) ||
2232 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2233 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2234 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2235 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2236 return false;
2237
2238 return true;
2239}
2240
2241static int check_func_proto(const struct bpf_func_proto *fn)
2242{
2243 return check_raw_mode_ok(fn) &&
2244 check_arg_pair_ok(fn) ? 0 : -EINVAL;
2245}
2246
2247/* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2248 * are now invalid, so turn them into unknown SCALAR_VALUE.
2249 */
2250static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2251 struct bpf_func_state *state)
2252{
2253 struct bpf_reg_state *regs = state->regs, *reg;
2254 int i;
2255
2256 for (i = 0; i < MAX_BPF_REG; i++)
2257 if (reg_is_pkt_pointer_any(®s[i]))
2258 mark_reg_unknown(env, regs, i);
2259
2260 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
2261 if (state->stack[i].slot_type[0] != STACK_SPILL)
2262 continue;
2263 reg = &state->stack[i].spilled_ptr;
2264 if (reg_is_pkt_pointer_any(reg))
2265 __mark_reg_unknown(reg);
2266 }
2267}
2268
2269static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2270{
2271 struct bpf_verifier_state *vstate = env->cur_state;
2272 int i;
2273
2274 for (i = 0; i <= vstate->curframe; i++)
2275 __clear_all_pkt_pointers(env, vstate->frame[i]);
2276}
2277
2278static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2279 int *insn_idx)
2280{
2281 struct bpf_verifier_state *state = env->cur_state;
2282 struct bpf_func_state *caller, *callee;
2283 int i, subprog, target_insn;
2284
2285 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2286 verbose(env, "the call stack of %d frames is too deep\n",
2287 state->curframe + 2);
2288 return -E2BIG;
2289 }
2290
2291 target_insn = *insn_idx + insn->imm;
2292 subprog = find_subprog(env, target_insn + 1);
2293 if (subprog < 0) {
2294 verbose(env, "verifier bug. No program starts at insn %d\n",
2295 target_insn + 1);
2296 return -EFAULT;
2297 }
2298
2299 caller = state->frame[state->curframe];
2300 if (state->frame[state->curframe + 1]) {
2301 verbose(env, "verifier bug. Frame %d already allocated\n",
2302 state->curframe + 1);
2303 return -EFAULT;
2304 }
2305
2306 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2307 if (!callee)
2308 return -ENOMEM;
2309 state->frame[state->curframe + 1] = callee;
2310
2311 /* callee cannot access r0, r6 - r9 for reading and has to write
2312 * into its own stack before reading from it.
2313 * callee can read/write into caller's stack
2314 */
2315 init_func_state(env, callee,
2316 /* remember the callsite, it will be used by bpf_exit */
2317 *insn_idx /* callsite */,
2318 state->curframe + 1 /* frameno within this callchain */,
2319 subprog + 1 /* subprog number within this prog */);
2320
2321 /* copy r1 - r5 args that callee can access */
2322 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2323 callee->regs[i] = caller->regs[i];
2324
2325 /* after the call regsiters r0 - r5 were scratched */
2326 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2327 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2328 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2329 }
2330
2331 /* only increment it after check_reg_arg() finished */
2332 state->curframe++;
2333
2334 /* and go analyze first insn of the callee */
2335 *insn_idx = target_insn;
2336
2337 if (env->log.level) {
2338 verbose(env, "caller:\n");
2339 print_verifier_state(env, caller);
2340 verbose(env, "callee:\n");
2341 print_verifier_state(env, callee);
2342 }
2343 return 0;
2344}
2345
2346static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2347{
2348 struct bpf_verifier_state *state = env->cur_state;
2349 struct bpf_func_state *caller, *callee;
2350 struct bpf_reg_state *r0;
2351
2352 callee = state->frame[state->curframe];
2353 r0 = &callee->regs[BPF_REG_0];
2354 if (r0->type == PTR_TO_STACK) {
2355 /* technically it's ok to return caller's stack pointer
2356 * (or caller's caller's pointer) back to the caller,
2357 * since these pointers are valid. Only current stack
2358 * pointer will be invalid as soon as function exits,
2359 * but let's be conservative
2360 */
2361 verbose(env, "cannot return stack pointer to the caller\n");
2362 return -EINVAL;
2363 }
2364
2365 state->curframe--;
2366 caller = state->frame[state->curframe];
2367 /* return to the caller whatever r0 had in the callee */
2368 caller->regs[BPF_REG_0] = *r0;
2369
2370 *insn_idx = callee->callsite + 1;
2371 if (env->log.level) {
2372 verbose(env, "returning from callee:\n");
2373 print_verifier_state(env, callee);
2374 verbose(env, "to caller at %d:\n", *insn_idx);
2375 print_verifier_state(env, caller);
2376 }
2377 /* clear everything in the callee */
2378 free_func_state(callee);
2379 state->frame[state->curframe + 1] = NULL;
2380 return 0;
2381}
2382
2383static int
2384record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2385 int func_id, int insn_idx)
2386{
2387 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2388
2389 if (func_id != BPF_FUNC_tail_call &&
2390 func_id != BPF_FUNC_map_lookup_elem)
2391 return 0;
2392 if (meta->map_ptr == NULL) {
2393 verbose(env, "kernel subsystem misconfigured verifier\n");
2394 return -EINVAL;
2395 }
2396
2397 if (!BPF_MAP_PTR(aux->map_state))
2398 bpf_map_ptr_store(aux, meta->map_ptr,
2399 meta->map_ptr->unpriv_array);
2400 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2401 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2402 meta->map_ptr->unpriv_array);
2403 return 0;
2404}
2405
2406static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2407{
2408 const struct bpf_func_proto *fn = NULL;
2409 struct bpf_reg_state *regs;
2410 struct bpf_call_arg_meta meta;
2411 bool changes_data;
2412 int i, err;
2413
2414 /* find function prototype */
2415 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2416 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2417 func_id);
2418 return -EINVAL;
2419 }
2420
2421 if (env->ops->get_func_proto)
2422 fn = env->ops->get_func_proto(func_id, env->prog);
2423 if (!fn) {
2424 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2425 func_id);
2426 return -EINVAL;
2427 }
2428
2429 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2430 if (!env->prog->gpl_compatible && fn->gpl_only) {
2431 verbose(env, "cannot call GPL only function from proprietary program\n");
2432 return -EINVAL;
2433 }
2434
2435 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2436 changes_data = bpf_helper_changes_pkt_data(fn->func);
2437 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2438 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2439 func_id_name(func_id), func_id);
2440 return -EINVAL;
2441 }
2442
2443 memset(&meta, 0, sizeof(meta));
2444 meta.pkt_access = fn->pkt_access;
2445
2446 err = check_func_proto(fn);
2447 if (err) {
2448 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2449 func_id_name(func_id), func_id);
2450 return err;
2451 }
2452
2453 /* check args */
2454 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2455 if (err)
2456 return err;
2457 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2458 if (err)
2459 return err;
2460 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2461 if (err)
2462 return err;
2463 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2464 if (err)
2465 return err;
2466 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2467 if (err)
2468 return err;
2469
2470 err = record_func_map(env, &meta, func_id, insn_idx);
2471 if (err)
2472 return err;
2473
2474 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2475 * is inferred from register state.
2476 */
2477 for (i = 0; i < meta.access_size; i++) {
2478 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2479 BPF_WRITE, -1, false);
2480 if (err)
2481 return err;
2482 }
2483
2484 regs = cur_regs(env);
2485 /* reset caller saved regs */
2486 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2487 mark_reg_not_init(env, regs, caller_saved[i]);
2488 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2489 }
2490
2491 /* update return register (already marked as written above) */
2492 if (fn->ret_type == RET_INTEGER) {
2493 /* sets type to SCALAR_VALUE */
2494 mark_reg_unknown(env, regs, BPF_REG_0);
2495 } else if (fn->ret_type == RET_VOID) {
2496 regs[BPF_REG_0].type = NOT_INIT;
2497 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
2498 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2499 /* There is no offset yet applied, variable or fixed */
2500 mark_reg_known_zero(env, regs, BPF_REG_0);
2501 regs[BPF_REG_0].off = 0;
2502 /* remember map_ptr, so that check_map_access()
2503 * can check 'value_size' boundary of memory access
2504 * to map element returned from bpf_map_lookup_elem()
2505 */
2506 if (meta.map_ptr == NULL) {
2507 verbose(env,
2508 "kernel subsystem misconfigured verifier\n");
2509 return -EINVAL;
2510 }
2511 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2512 regs[BPF_REG_0].id = ++env->id_gen;
2513 } else {
2514 verbose(env, "unknown return type %d of func %s#%d\n",
2515 fn->ret_type, func_id_name(func_id), func_id);
2516 return -EINVAL;
2517 }
2518
2519 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2520 if (err)
2521 return err;
2522
2523 if (changes_data)
2524 clear_all_pkt_pointers(env);
2525 return 0;
2526}
2527
2528static bool signed_add_overflows(s64 a, s64 b)
2529{
2530 /* Do the add in u64, where overflow is well-defined */
2531 s64 res = (s64)((u64)a + (u64)b);
2532
2533 if (b < 0)
2534 return res > a;
2535 return res < a;
2536}
2537
2538static bool signed_sub_overflows(s64 a, s64 b)
2539{
2540 /* Do the sub in u64, where overflow is well-defined */
2541 s64 res = (s64)((u64)a - (u64)b);
2542
2543 if (b < 0)
2544 return res < a;
2545 return res > a;
2546}
2547
2548static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2549 const struct bpf_reg_state *reg,
2550 enum bpf_reg_type type)
2551{
2552 bool known = tnum_is_const(reg->var_off);
2553 s64 val = reg->var_off.value;
2554 s64 smin = reg->smin_value;
2555
2556 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2557 verbose(env, "math between %s pointer and %lld is not allowed\n",
2558 reg_type_str[type], val);
2559 return false;
2560 }
2561
2562 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2563 verbose(env, "%s pointer offset %d is not allowed\n",
2564 reg_type_str[type], reg->off);
2565 return false;
2566 }
2567
2568 if (smin == S64_MIN) {
2569 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2570 reg_type_str[type]);
2571 return false;
2572 }
2573
2574 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2575 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2576 smin, reg_type_str[type]);
2577 return false;
2578 }
2579
2580 return true;
2581}
2582
2583/* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2584 * Caller should also handle BPF_MOV case separately.
2585 * If we return -EACCES, caller may want to try again treating pointer as a
2586 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2587 */
2588static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2589 struct bpf_insn *insn,
2590 const struct bpf_reg_state *ptr_reg,
2591 const struct bpf_reg_state *off_reg)
2592{
2593 struct bpf_verifier_state *vstate = env->cur_state;
2594 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2595 struct bpf_reg_state *regs = state->regs, *dst_reg;
2596 bool known = tnum_is_const(off_reg->var_off);
2597 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2598 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2599 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2600 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2601 u8 opcode = BPF_OP(insn->code);
2602 u32 dst = insn->dst_reg;
2603
2604 dst_reg = ®s[dst];
2605
2606 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2607 smin_val > smax_val || umin_val > umax_val) {
2608 /* Taint dst register if offset had invalid bounds derived from
2609 * e.g. dead branches.
2610 */
2611 __mark_reg_unknown(dst_reg);
2612 return 0;
2613 }
2614
2615 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2616 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2617 verbose(env,
2618 "R%d 32-bit pointer arithmetic prohibited\n",
2619 dst);
2620 return -EACCES;
2621 }
2622
2623 if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
2624 verbose(env, "R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
2625 dst);
2626 return -EACCES;
2627 }
2628 if (ptr_reg->type == CONST_PTR_TO_MAP) {
2629 verbose(env, "R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
2630 dst);
2631 return -EACCES;
2632 }
2633 if (ptr_reg->type == PTR_TO_PACKET_END) {
2634 verbose(env, "R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
2635 dst);
2636 return -EACCES;
2637 }
2638
2639 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
2640 * The id may be overwritten later if we create a new variable offset.
2641 */
2642 dst_reg->type = ptr_reg->type;
2643 dst_reg->id = ptr_reg->id;
2644
2645 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
2646 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
2647 return -EINVAL;
2648
2649 switch (opcode) {
2650 case BPF_ADD:
2651 /* We can take a fixed offset as long as it doesn't overflow
2652 * the s32 'off' field
2653 */
2654 if (known && (ptr_reg->off + smin_val ==
2655 (s64)(s32)(ptr_reg->off + smin_val))) {
2656 /* pointer += K. Accumulate it into fixed offset */
2657 dst_reg->smin_value = smin_ptr;
2658 dst_reg->smax_value = smax_ptr;
2659 dst_reg->umin_value = umin_ptr;
2660 dst_reg->umax_value = umax_ptr;
2661 dst_reg->var_off = ptr_reg->var_off;
2662 dst_reg->off = ptr_reg->off + smin_val;
2663 dst_reg->range = ptr_reg->range;
2664 break;
2665 }
2666 /* A new variable offset is created. Note that off_reg->off
2667 * == 0, since it's a scalar.
2668 * dst_reg gets the pointer type and since some positive
2669 * integer value was added to the pointer, give it a new 'id'
2670 * if it's a PTR_TO_PACKET.
2671 * this creates a new 'base' pointer, off_reg (variable) gets
2672 * added into the variable offset, and we copy the fixed offset
2673 * from ptr_reg.
2674 */
2675 if (signed_add_overflows(smin_ptr, smin_val) ||
2676 signed_add_overflows(smax_ptr, smax_val)) {
2677 dst_reg->smin_value = S64_MIN;
2678 dst_reg->smax_value = S64_MAX;
2679 } else {
2680 dst_reg->smin_value = smin_ptr + smin_val;
2681 dst_reg->smax_value = smax_ptr + smax_val;
2682 }
2683 if (umin_ptr + umin_val < umin_ptr ||
2684 umax_ptr + umax_val < umax_ptr) {
2685 dst_reg->umin_value = 0;
2686 dst_reg->umax_value = U64_MAX;
2687 } else {
2688 dst_reg->umin_value = umin_ptr + umin_val;
2689 dst_reg->umax_value = umax_ptr + umax_val;
2690 }
2691 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
2692 dst_reg->off = ptr_reg->off;
2693 if (reg_is_pkt_pointer(ptr_reg)) {
2694 dst_reg->id = ++env->id_gen;
2695 /* something was added to pkt_ptr, set range to zero */
2696 dst_reg->range = 0;
2697 }
2698 break;
2699 case BPF_SUB:
2700 if (dst_reg == off_reg) {
2701 /* scalar -= pointer. Creates an unknown scalar */
2702 verbose(env, "R%d tried to subtract pointer from scalar\n",
2703 dst);
2704 return -EACCES;
2705 }
2706 /* We don't allow subtraction from FP, because (according to
2707 * test_verifier.c test "invalid fp arithmetic", JITs might not
2708 * be able to deal with it.
2709 */
2710 if (ptr_reg->type == PTR_TO_STACK) {
2711 verbose(env, "R%d subtraction from stack pointer prohibited\n",
2712 dst);
2713 return -EACCES;
2714 }
2715 if (known && (ptr_reg->off - smin_val ==
2716 (s64)(s32)(ptr_reg->off - smin_val))) {
2717 /* pointer -= K. Subtract it from fixed offset */
2718 dst_reg->smin_value = smin_ptr;
2719 dst_reg->smax_value = smax_ptr;
2720 dst_reg->umin_value = umin_ptr;
2721 dst_reg->umax_value = umax_ptr;
2722 dst_reg->var_off = ptr_reg->var_off;
2723 dst_reg->id = ptr_reg->id;
2724 dst_reg->off = ptr_reg->off - smin_val;
2725 dst_reg->range = ptr_reg->range;
2726 break;
2727 }
2728 /* A new variable offset is created. If the subtrahend is known
2729 * nonnegative, then any reg->range we had before is still good.
2730 */
2731 if (signed_sub_overflows(smin_ptr, smax_val) ||
2732 signed_sub_overflows(smax_ptr, smin_val)) {
2733 /* Overflow possible, we know nothing */
2734 dst_reg->smin_value = S64_MIN;
2735 dst_reg->smax_value = S64_MAX;
2736 } else {
2737 dst_reg->smin_value = smin_ptr - smax_val;
2738 dst_reg->smax_value = smax_ptr - smin_val;
2739 }
2740 if (umin_ptr < umax_val) {
2741 /* Overflow possible, we know nothing */
2742 dst_reg->umin_value = 0;
2743 dst_reg->umax_value = U64_MAX;
2744 } else {
2745 /* Cannot overflow (as long as bounds are consistent) */
2746 dst_reg->umin_value = umin_ptr - umax_val;
2747 dst_reg->umax_value = umax_ptr - umin_val;
2748 }
2749 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
2750 dst_reg->off = ptr_reg->off;
2751 if (reg_is_pkt_pointer(ptr_reg)) {
2752 dst_reg->id = ++env->id_gen;
2753 /* something was added to pkt_ptr, set range to zero */
2754 if (smin_val < 0)
2755 dst_reg->range = 0;
2756 }
2757 break;
2758 case BPF_AND:
2759 case BPF_OR:
2760 case BPF_XOR:
2761 /* bitwise ops on pointers are troublesome, prohibit. */
2762 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
2763 dst, bpf_alu_string[opcode >> 4]);
2764 return -EACCES;
2765 default:
2766 /* other operators (e.g. MUL,LSH) produce non-pointer results */
2767 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
2768 dst, bpf_alu_string[opcode >> 4]);
2769 return -EACCES;
2770 }
2771
2772 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
2773 return -EINVAL;
2774
2775 __update_reg_bounds(dst_reg);
2776 __reg_deduce_bounds(dst_reg);
2777 __reg_bound_offset(dst_reg);
2778 return 0;
2779}
2780
2781/* WARNING: This function does calculations on 64-bit values, but the actual
2782 * execution may occur on 32-bit values. Therefore, things like bitshifts
2783 * need extra checks in the 32-bit case.
2784 */
2785static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
2786 struct bpf_insn *insn,
2787 struct bpf_reg_state *dst_reg,
2788 struct bpf_reg_state src_reg)
2789{
2790 struct bpf_reg_state *regs = cur_regs(env);
2791 u8 opcode = BPF_OP(insn->code);
2792 bool src_known, dst_known;
2793 s64 smin_val, smax_val;
2794 u64 umin_val, umax_val;
2795 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
2796
2797 smin_val = src_reg.smin_value;
2798 smax_val = src_reg.smax_value;
2799 umin_val = src_reg.umin_value;
2800 umax_val = src_reg.umax_value;
2801 src_known = tnum_is_const(src_reg.var_off);
2802 dst_known = tnum_is_const(dst_reg->var_off);
2803
2804 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
2805 smin_val > smax_val || umin_val > umax_val) {
2806 /* Taint dst register if offset had invalid bounds derived from
2807 * e.g. dead branches.
2808 */
2809 __mark_reg_unknown(dst_reg);
2810 return 0;
2811 }
2812
2813 if (!src_known &&
2814 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
2815 __mark_reg_unknown(dst_reg);
2816 return 0;
2817 }
2818
2819 switch (opcode) {
2820 case BPF_ADD:
2821 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
2822 signed_add_overflows(dst_reg->smax_value, smax_val)) {
2823 dst_reg->smin_value = S64_MIN;
2824 dst_reg->smax_value = S64_MAX;
2825 } else {
2826 dst_reg->smin_value += smin_val;
2827 dst_reg->smax_value += smax_val;
2828 }
2829 if (dst_reg->umin_value + umin_val < umin_val ||
2830 dst_reg->umax_value + umax_val < umax_val) {
2831 dst_reg->umin_value = 0;
2832 dst_reg->umax_value = U64_MAX;
2833 } else {
2834 dst_reg->umin_value += umin_val;
2835 dst_reg->umax_value += umax_val;
2836 }
2837 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
2838 break;
2839 case BPF_SUB:
2840 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
2841 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
2842 /* Overflow possible, we know nothing */
2843 dst_reg->smin_value = S64_MIN;
2844 dst_reg->smax_value = S64_MAX;
2845 } else {
2846 dst_reg->smin_value -= smax_val;
2847 dst_reg->smax_value -= smin_val;
2848 }
2849 if (dst_reg->umin_value < umax_val) {
2850 /* Overflow possible, we know nothing */
2851 dst_reg->umin_value = 0;
2852 dst_reg->umax_value = U64_MAX;
2853 } else {
2854 /* Cannot overflow (as long as bounds are consistent) */
2855 dst_reg->umin_value -= umax_val;
2856 dst_reg->umax_value -= umin_val;
2857 }
2858 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
2859 break;
2860 case BPF_MUL:
2861 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
2862 if (smin_val < 0 || dst_reg->smin_value < 0) {
2863 /* Ain't nobody got time to multiply that sign */
2864 __mark_reg_unbounded(dst_reg);
2865 __update_reg_bounds(dst_reg);
2866 break;
2867 }
2868 /* Both values are positive, so we can work with unsigned and
2869 * copy the result to signed (unless it exceeds S64_MAX).
2870 */
2871 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
2872 /* Potential overflow, we know nothing */
2873 __mark_reg_unbounded(dst_reg);
2874 /* (except what we can learn from the var_off) */
2875 __update_reg_bounds(dst_reg);
2876 break;
2877 }
2878 dst_reg->umin_value *= umin_val;
2879 dst_reg->umax_value *= umax_val;
2880 if (dst_reg->umax_value > S64_MAX) {
2881 /* Overflow possible, we know nothing */
2882 dst_reg->smin_value = S64_MIN;
2883 dst_reg->smax_value = S64_MAX;
2884 } else {
2885 dst_reg->smin_value = dst_reg->umin_value;
2886 dst_reg->smax_value = dst_reg->umax_value;
2887 }
2888 break;
2889 case BPF_AND:
2890 if (src_known && dst_known) {
2891 __mark_reg_known(dst_reg, dst_reg->var_off.value &
2892 src_reg.var_off.value);
2893 break;
2894 }
2895 /* We get our minimum from the var_off, since that's inherently
2896 * bitwise. Our maximum is the minimum of the operands' maxima.
2897 */
2898 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
2899 dst_reg->umin_value = dst_reg->var_off.value;
2900 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
2901 if (dst_reg->smin_value < 0 || smin_val < 0) {
2902 /* Lose signed bounds when ANDing negative numbers,
2903 * ain't nobody got time for that.
2904 */
2905 dst_reg->smin_value = S64_MIN;
2906 dst_reg->smax_value = S64_MAX;
2907 } else {
2908 /* ANDing two positives gives a positive, so safe to
2909 * cast result into s64.
2910 */
2911 dst_reg->smin_value = dst_reg->umin_value;
2912 dst_reg->smax_value = dst_reg->umax_value;
2913 }
2914 /* We may learn something more from the var_off */
2915 __update_reg_bounds(dst_reg);
2916 break;
2917 case BPF_OR:
2918 if (src_known && dst_known) {
2919 __mark_reg_known(dst_reg, dst_reg->var_off.value |
2920 src_reg.var_off.value);
2921 break;
2922 }
2923 /* We get our maximum from the var_off, and our minimum is the
2924 * maximum of the operands' minima
2925 */
2926 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
2927 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
2928 dst_reg->umax_value = dst_reg->var_off.value |
2929 dst_reg->var_off.mask;
2930 if (dst_reg->smin_value < 0 || smin_val < 0) {
2931 /* Lose signed bounds when ORing negative numbers,
2932 * ain't nobody got time for that.
2933 */
2934 dst_reg->smin_value = S64_MIN;
2935 dst_reg->smax_value = S64_MAX;
2936 } else {
2937 /* ORing two positives gives a positive, so safe to
2938 * cast result into s64.
2939 */
2940 dst_reg->smin_value = dst_reg->umin_value;
2941 dst_reg->smax_value = dst_reg->umax_value;
2942 }
2943 /* We may learn something more from the var_off */
2944 __update_reg_bounds(dst_reg);
2945 break;
2946 case BPF_LSH:
2947 if (umax_val >= insn_bitness) {
2948 /* Shifts greater than 31 or 63 are undefined.
2949 * This includes shifts by a negative number.
2950 */
2951 mark_reg_unknown(env, regs, insn->dst_reg);
2952 break;
2953 }
2954 /* We lose all sign bit information (except what we can pick
2955 * up from var_off)
2956 */
2957 dst_reg->smin_value = S64_MIN;
2958 dst_reg->smax_value = S64_MAX;
2959 /* If we might shift our top bit out, then we know nothing */
2960 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
2961 dst_reg->umin_value = 0;
2962 dst_reg->umax_value = U64_MAX;
2963 } else {
2964 dst_reg->umin_value <<= umin_val;
2965 dst_reg->umax_value <<= umax_val;
2966 }
2967 if (src_known)
2968 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
2969 else
2970 dst_reg->var_off = tnum_lshift(tnum_unknown, umin_val);
2971 /* We may learn something more from the var_off */
2972 __update_reg_bounds(dst_reg);
2973 break;
2974 case BPF_RSH:
2975 if (umax_val >= insn_bitness) {
2976 /* Shifts greater than 31 or 63 are undefined.
2977 * This includes shifts by a negative number.
2978 */
2979 mark_reg_unknown(env, regs, insn->dst_reg);
2980 break;
2981 }
2982 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
2983 * be negative, then either:
2984 * 1) src_reg might be zero, so the sign bit of the result is
2985 * unknown, so we lose our signed bounds
2986 * 2) it's known negative, thus the unsigned bounds capture the
2987 * signed bounds
2988 * 3) the signed bounds cross zero, so they tell us nothing
2989 * about the result
2990 * If the value in dst_reg is known nonnegative, then again the
2991 * unsigned bounts capture the signed bounds.
2992 * Thus, in all cases it suffices to blow away our signed bounds
2993 * and rely on inferring new ones from the unsigned bounds and
2994 * var_off of the result.
2995 */
2996 dst_reg->smin_value = S64_MIN;
2997 dst_reg->smax_value = S64_MAX;
2998 if (src_known)
2999 dst_reg->var_off = tnum_rshift(dst_reg->var_off,
3000 umin_val);
3001 else
3002 dst_reg->var_off = tnum_rshift(tnum_unknown, umin_val);
3003 dst_reg->umin_value >>= umax_val;
3004 dst_reg->umax_value >>= umin_val;
3005 /* We may learn something more from the var_off */
3006 __update_reg_bounds(dst_reg);
3007 break;
3008 default:
3009 mark_reg_unknown(env, regs, insn->dst_reg);
3010 break;
3011 }
3012
3013 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3014 /* 32-bit ALU ops are (32,32)->32 */
3015 coerce_reg_to_size(dst_reg, 4);
3016 coerce_reg_to_size(&src_reg, 4);
3017 }
3018
3019 __reg_deduce_bounds(dst_reg);
3020 __reg_bound_offset(dst_reg);
3021 return 0;
3022}
3023
3024/* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3025 * and var_off.
3026 */
3027static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3028 struct bpf_insn *insn)
3029{
3030 struct bpf_verifier_state *vstate = env->cur_state;
3031 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3032 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3033 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3034 u8 opcode = BPF_OP(insn->code);
3035
3036 dst_reg = ®s[insn->dst_reg];
3037 src_reg = NULL;
3038 if (dst_reg->type != SCALAR_VALUE)
3039 ptr_reg = dst_reg;
3040 if (BPF_SRC(insn->code) == BPF_X) {
3041 src_reg = ®s[insn->src_reg];
3042 if (src_reg->type != SCALAR_VALUE) {
3043 if (dst_reg->type != SCALAR_VALUE) {
3044 /* Combining two pointers by any ALU op yields
3045 * an arbitrary scalar. Disallow all math except
3046 * pointer subtraction
3047 */
3048 if (opcode == BPF_SUB){
3049 mark_reg_unknown(env, regs, insn->dst_reg);
3050 return 0;
3051 }
3052 verbose(env, "R%d pointer %s pointer prohibited\n",
3053 insn->dst_reg,
3054 bpf_alu_string[opcode >> 4]);
3055 return -EACCES;
3056 } else {
3057 /* scalar += pointer
3058 * This is legal, but we have to reverse our
3059 * src/dest handling in computing the range
3060 */
3061 return adjust_ptr_min_max_vals(env, insn,
3062 src_reg, dst_reg);
3063 }
3064 } else if (ptr_reg) {
3065 /* pointer += scalar */
3066 return adjust_ptr_min_max_vals(env, insn,
3067 dst_reg, src_reg);
3068 }
3069 } else {
3070 /* Pretend the src is a reg with a known value, since we only
3071 * need to be able to read from this state.
3072 */
3073 off_reg.type = SCALAR_VALUE;
3074 __mark_reg_known(&off_reg, insn->imm);
3075 src_reg = &off_reg;
3076 if (ptr_reg) /* pointer += K */
3077 return adjust_ptr_min_max_vals(env, insn,
3078 ptr_reg, src_reg);
3079 }
3080
3081 /* Got here implies adding two SCALAR_VALUEs */
3082 if (WARN_ON_ONCE(ptr_reg)) {
3083 print_verifier_state(env, state);
3084 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3085 return -EINVAL;
3086 }
3087 if (WARN_ON(!src_reg)) {
3088 print_verifier_state(env, state);
3089 verbose(env, "verifier internal error: no src_reg\n");
3090 return -EINVAL;
3091 }
3092 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3093}
3094
3095/* check validity of 32-bit and 64-bit arithmetic operations */
3096static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3097{
3098 struct bpf_reg_state *regs = cur_regs(env);
3099 u8 opcode = BPF_OP(insn->code);
3100 int err;
3101
3102 if (opcode == BPF_END || opcode == BPF_NEG) {
3103 if (opcode == BPF_NEG) {
3104 if (BPF_SRC(insn->code) != 0 ||
3105 insn->src_reg != BPF_REG_0 ||
3106 insn->off != 0 || insn->imm != 0) {
3107 verbose(env, "BPF_NEG uses reserved fields\n");
3108 return -EINVAL;
3109 }
3110 } else {
3111 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3112 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3113 BPF_CLASS(insn->code) == BPF_ALU64) {
3114 verbose(env, "BPF_END uses reserved fields\n");
3115 return -EINVAL;
3116 }
3117 }
3118
3119 /* check src operand */
3120 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3121 if (err)
3122 return err;
3123
3124 if (is_pointer_value(env, insn->dst_reg)) {
3125 verbose(env, "R%d pointer arithmetic prohibited\n",
3126 insn->dst_reg);
3127 return -EACCES;
3128 }
3129
3130 /* check dest operand */
3131 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3132 if (err)
3133 return err;
3134
3135 } else if (opcode == BPF_MOV) {
3136
3137 if (BPF_SRC(insn->code) == BPF_X) {
3138 if (insn->imm != 0 || insn->off != 0) {
3139 verbose(env, "BPF_MOV uses reserved fields\n");
3140 return -EINVAL;
3141 }
3142
3143 /* check src operand */
3144 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3145 if (err)
3146 return err;
3147 } else {
3148 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3149 verbose(env, "BPF_MOV uses reserved fields\n");
3150 return -EINVAL;
3151 }
3152 }
3153
3154 /* check dest operand */
3155 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3156 if (err)
3157 return err;
3158
3159 if (BPF_SRC(insn->code) == BPF_X) {
3160 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3161 /* case: R1 = R2
3162 * copy register state to dest reg
3163 */
3164 regs[insn->dst_reg] = regs[insn->src_reg];
3165 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3166 } else {
3167 /* R1 = (u32) R2 */
3168 if (is_pointer_value(env, insn->src_reg)) {
3169 verbose(env,
3170 "R%d partial copy of pointer\n",
3171 insn->src_reg);
3172 return -EACCES;
3173 }
3174 mark_reg_unknown(env, regs, insn->dst_reg);
3175 coerce_reg_to_size(®s[insn->dst_reg], 4);
3176 }
3177 } else {
3178 /* case: R = imm
3179 * remember the value we stored into this reg
3180 */
3181 regs[insn->dst_reg].type = SCALAR_VALUE;
3182 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3183 __mark_reg_known(regs + insn->dst_reg,
3184 insn->imm);
3185 } else {
3186 __mark_reg_known(regs + insn->dst_reg,
3187 (u32)insn->imm);
3188 }
3189 }
3190
3191 } else if (opcode > BPF_END) {
3192 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3193 return -EINVAL;
3194
3195 } else { /* all other ALU ops: and, sub, xor, add, ... */
3196
3197 if (BPF_SRC(insn->code) == BPF_X) {
3198 if (insn->imm != 0 || insn->off != 0) {
3199 verbose(env, "BPF_ALU uses reserved fields\n");
3200 return -EINVAL;
3201 }
3202 /* check src1 operand */
3203 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3204 if (err)
3205 return err;
3206 } else {
3207 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3208 verbose(env, "BPF_ALU uses reserved fields\n");
3209 return -EINVAL;
3210 }
3211 }
3212
3213 /* check src2 operand */
3214 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3215 if (err)
3216 return err;
3217
3218 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3219 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3220 verbose(env, "div by zero\n");
3221 return -EINVAL;
3222 }
3223
3224 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3225 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3226 return -EINVAL;
3227 }
3228
3229 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3230 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3231 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3232
3233 if (insn->imm < 0 || insn->imm >= size) {
3234 verbose(env, "invalid shift %d\n", insn->imm);
3235 return -EINVAL;
3236 }
3237 }
3238
3239 /* check dest operand */
3240 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3241 if (err)
3242 return err;
3243
3244 return adjust_reg_min_max_vals(env, insn);
3245 }
3246
3247 return 0;
3248}
3249
3250static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3251 struct bpf_reg_state *dst_reg,
3252 enum bpf_reg_type type,
3253 bool range_right_open)
3254{
3255 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3256 struct bpf_reg_state *regs = state->regs, *reg;
3257 u16 new_range;
3258 int i, j;
3259
3260 if (dst_reg->off < 0 ||
3261 (dst_reg->off == 0 && range_right_open))
3262 /* This doesn't give us any range */
3263 return;
3264
3265 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3266 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3267 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3268 * than pkt_end, but that's because it's also less than pkt.
3269 */
3270 return;
3271
3272 new_range = dst_reg->off;
3273 if (range_right_open)
3274 new_range--;
3275
3276 /* Examples for register markings:
3277 *
3278 * pkt_data in dst register:
3279 *
3280 * r2 = r3;
3281 * r2 += 8;
3282 * if (r2 > pkt_end) goto <handle exception>
3283 * <access okay>
3284 *
3285 * r2 = r3;
3286 * r2 += 8;
3287 * if (r2 < pkt_end) goto <access okay>
3288 * <handle exception>
3289 *
3290 * Where:
3291 * r2 == dst_reg, pkt_end == src_reg
3292 * r2=pkt(id=n,off=8,r=0)
3293 * r3=pkt(id=n,off=0,r=0)
3294 *
3295 * pkt_data in src register:
3296 *
3297 * r2 = r3;
3298 * r2 += 8;
3299 * if (pkt_end >= r2) goto <access okay>
3300 * <handle exception>
3301 *
3302 * r2 = r3;
3303 * r2 += 8;
3304 * if (pkt_end <= r2) goto <handle exception>
3305 * <access okay>
3306 *
3307 * Where:
3308 * pkt_end == dst_reg, r2 == src_reg
3309 * r2=pkt(id=n,off=8,r=0)
3310 * r3=pkt(id=n,off=0,r=0)
3311 *
3312 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3313 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3314 * and [r3, r3 + 8-1) respectively is safe to access depending on
3315 * the check.
3316 */
3317
3318 /* If our ids match, then we must have the same max_value. And we
3319 * don't care about the other reg's fixed offset, since if it's too big
3320 * the range won't allow anything.
3321 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3322 */
3323 for (i = 0; i < MAX_BPF_REG; i++)
3324 if (regs[i].type == type && regs[i].id == dst_reg->id)
3325 /* keep the maximum range already checked */
3326 regs[i].range = max(regs[i].range, new_range);
3327
3328 for (j = 0; j <= vstate->curframe; j++) {
3329 state = vstate->frame[j];
3330 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3331 if (state->stack[i].slot_type[0] != STACK_SPILL)
3332 continue;
3333 reg = &state->stack[i].spilled_ptr;
3334 if (reg->type == type && reg->id == dst_reg->id)
3335 reg->range = max(reg->range, new_range);
3336 }
3337 }
3338}
3339
3340/* Adjusts the register min/max values in the case that the dst_reg is the
3341 * variable register that we are working on, and src_reg is a constant or we're
3342 * simply doing a BPF_K check.
3343 * In JEQ/JNE cases we also adjust the var_off values.
3344 */
3345static void reg_set_min_max(struct bpf_reg_state *true_reg,
3346 struct bpf_reg_state *false_reg, u64 val,
3347 u8 opcode)
3348{
3349 /* If the dst_reg is a pointer, we can't learn anything about its
3350 * variable offset from the compare (unless src_reg were a pointer into
3351 * the same object, but we don't bother with that.
3352 * Since false_reg and true_reg have the same type by construction, we
3353 * only need to check one of them for pointerness.
3354 */
3355 if (__is_pointer_value(false, false_reg))
3356 return;
3357
3358 switch (opcode) {
3359 case BPF_JEQ:
3360 /* If this is false then we know nothing Jon Snow, but if it is
3361 * true then we know for sure.
3362 */
3363 __mark_reg_known(true_reg, val);
3364 break;
3365 case BPF_JNE:
3366 /* If this is true we know nothing Jon Snow, but if it is false
3367 * we know the value for sure;
3368 */
3369 __mark_reg_known(false_reg, val);
3370 break;
3371 case BPF_JGT:
3372 false_reg->umax_value = min(false_reg->umax_value, val);
3373 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3374 break;
3375 case BPF_JSGT:
3376 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3377 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3378 break;
3379 case BPF_JLT:
3380 false_reg->umin_value = max(false_reg->umin_value, val);
3381 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3382 break;
3383 case BPF_JSLT:
3384 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3385 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3386 break;
3387 case BPF_JGE:
3388 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3389 true_reg->umin_value = max(true_reg->umin_value, val);
3390 break;
3391 case BPF_JSGE:
3392 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3393 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3394 break;
3395 case BPF_JLE:
3396 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3397 true_reg->umax_value = min(true_reg->umax_value, val);
3398 break;
3399 case BPF_JSLE:
3400 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3401 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3402 break;
3403 default:
3404 break;
3405 }
3406
3407 __reg_deduce_bounds(false_reg);
3408 __reg_deduce_bounds(true_reg);
3409 /* We might have learned some bits from the bounds. */
3410 __reg_bound_offset(false_reg);
3411 __reg_bound_offset(true_reg);
3412 /* Intersecting with the old var_off might have improved our bounds
3413 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3414 * then new var_off is (0; 0x7f...fc) which improves our umax.
3415 */
3416 __update_reg_bounds(false_reg);
3417 __update_reg_bounds(true_reg);
3418}
3419
3420/* Same as above, but for the case that dst_reg holds a constant and src_reg is
3421 * the variable reg.
3422 */
3423static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3424 struct bpf_reg_state *false_reg, u64 val,
3425 u8 opcode)
3426{
3427 if (__is_pointer_value(false, false_reg))
3428 return;
3429
3430 switch (opcode) {
3431 case BPF_JEQ:
3432 /* If this is false then we know nothing Jon Snow, but if it is
3433 * true then we know for sure.
3434 */
3435 __mark_reg_known(true_reg, val);
3436 break;
3437 case BPF_JNE:
3438 /* If this is true we know nothing Jon Snow, but if it is false
3439 * we know the value for sure;
3440 */
3441 __mark_reg_known(false_reg, val);
3442 break;
3443 case BPF_JGT:
3444 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3445 false_reg->umin_value = max(false_reg->umin_value, val);
3446 break;
3447 case BPF_JSGT:
3448 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3449 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3450 break;
3451 case BPF_JLT:
3452 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3453 false_reg->umax_value = min(false_reg->umax_value, val);
3454 break;
3455 case BPF_JSLT:
3456 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3457 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3458 break;
3459 case BPF_JGE:
3460 true_reg->umax_value = min(true_reg->umax_value, val);
3461 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3462 break;
3463 case BPF_JSGE:
3464 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3465 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3466 break;
3467 case BPF_JLE:
3468 true_reg->umin_value = max(true_reg->umin_value, val);
3469 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3470 break;
3471 case BPF_JSLE:
3472 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3473 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3474 break;
3475 default:
3476 break;
3477 }
3478
3479 __reg_deduce_bounds(false_reg);
3480 __reg_deduce_bounds(true_reg);
3481 /* We might have learned some bits from the bounds. */
3482 __reg_bound_offset(false_reg);
3483 __reg_bound_offset(true_reg);
3484 /* Intersecting with the old var_off might have improved our bounds
3485 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3486 * then new var_off is (0; 0x7f...fc) which improves our umax.
3487 */
3488 __update_reg_bounds(false_reg);
3489 __update_reg_bounds(true_reg);
3490}
3491
3492/* Regs are known to be equal, so intersect their min/max/var_off */
3493static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3494 struct bpf_reg_state *dst_reg)
3495{
3496 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3497 dst_reg->umin_value);
3498 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3499 dst_reg->umax_value);
3500 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3501 dst_reg->smin_value);
3502 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3503 dst_reg->smax_value);
3504 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3505 dst_reg->var_off);
3506 /* We might have learned new bounds from the var_off. */
3507 __update_reg_bounds(src_reg);
3508 __update_reg_bounds(dst_reg);
3509 /* We might have learned something about the sign bit. */
3510 __reg_deduce_bounds(src_reg);
3511 __reg_deduce_bounds(dst_reg);
3512 /* We might have learned some bits from the bounds. */
3513 __reg_bound_offset(src_reg);
3514 __reg_bound_offset(dst_reg);
3515 /* Intersecting with the old var_off might have improved our bounds
3516 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3517 * then new var_off is (0; 0x7f...fc) which improves our umax.
3518 */
3519 __update_reg_bounds(src_reg);
3520 __update_reg_bounds(dst_reg);
3521}
3522
3523static void reg_combine_min_max(struct bpf_reg_state *true_src,
3524 struct bpf_reg_state *true_dst,
3525 struct bpf_reg_state *false_src,
3526 struct bpf_reg_state *false_dst,
3527 u8 opcode)
3528{
3529 switch (opcode) {
3530 case BPF_JEQ:
3531 __reg_combine_min_max(true_src, true_dst);
3532 break;
3533 case BPF_JNE:
3534 __reg_combine_min_max(false_src, false_dst);
3535 break;
3536 }
3537}
3538
3539static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
3540 bool is_null)
3541{
3542 struct bpf_reg_state *reg = ®s[regno];
3543
3544 if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
3545 /* Old offset (both fixed and variable parts) should
3546 * have been known-zero, because we don't allow pointer
3547 * arithmetic on pointers that might be NULL.
3548 */
3549 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3550 !tnum_equals_const(reg->var_off, 0) ||
3551 reg->off)) {
3552 __mark_reg_known_zero(reg);
3553 reg->off = 0;
3554 }
3555 if (is_null) {
3556 reg->type = SCALAR_VALUE;
3557 } else if (reg->map_ptr->inner_map_meta) {
3558 reg->type = CONST_PTR_TO_MAP;
3559 reg->map_ptr = reg->map_ptr->inner_map_meta;
3560 } else {
3561 reg->type = PTR_TO_MAP_VALUE;
3562 }
3563 /* We don't need id from this point onwards anymore, thus we
3564 * should better reset it, so that state pruning has chances
3565 * to take effect.
3566 */
3567 reg->id = 0;
3568 }
3569}
3570
3571/* The logic is similar to find_good_pkt_pointers(), both could eventually
3572 * be folded together at some point.
3573 */
3574static void mark_map_regs(struct bpf_verifier_state *vstate, u32 regno,
3575 bool is_null)
3576{
3577 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3578 struct bpf_reg_state *regs = state->regs;
3579 u32 id = regs[regno].id;
3580 int i, j;
3581
3582 for (i = 0; i < MAX_BPF_REG; i++)
3583 mark_map_reg(regs, i, id, is_null);
3584
3585 for (j = 0; j <= vstate->curframe; j++) {
3586 state = vstate->frame[j];
3587 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
3588 if (state->stack[i].slot_type[0] != STACK_SPILL)
3589 continue;
3590 mark_map_reg(&state->stack[i].spilled_ptr, 0, id, is_null);
3591 }
3592 }
3593}
3594
3595static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3596 struct bpf_reg_state *dst_reg,
3597 struct bpf_reg_state *src_reg,
3598 struct bpf_verifier_state *this_branch,
3599 struct bpf_verifier_state *other_branch)
3600{
3601 if (BPF_SRC(insn->code) != BPF_X)
3602 return false;
3603
3604 switch (BPF_OP(insn->code)) {
3605 case BPF_JGT:
3606 if ((dst_reg->type == PTR_TO_PACKET &&
3607 src_reg->type == PTR_TO_PACKET_END) ||
3608 (dst_reg->type == PTR_TO_PACKET_META &&
3609 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3610 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
3611 find_good_pkt_pointers(this_branch, dst_reg,
3612 dst_reg->type, false);
3613 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3614 src_reg->type == PTR_TO_PACKET) ||
3615 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3616 src_reg->type == PTR_TO_PACKET_META)) {
3617 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
3618 find_good_pkt_pointers(other_branch, src_reg,
3619 src_reg->type, true);
3620 } else {
3621 return false;
3622 }
3623 break;
3624 case BPF_JLT:
3625 if ((dst_reg->type == PTR_TO_PACKET &&
3626 src_reg->type == PTR_TO_PACKET_END) ||
3627 (dst_reg->type == PTR_TO_PACKET_META &&
3628 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3629 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
3630 find_good_pkt_pointers(other_branch, dst_reg,
3631 dst_reg->type, true);
3632 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3633 src_reg->type == PTR_TO_PACKET) ||
3634 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3635 src_reg->type == PTR_TO_PACKET_META)) {
3636 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
3637 find_good_pkt_pointers(this_branch, src_reg,
3638 src_reg->type, false);
3639 } else {
3640 return false;
3641 }
3642 break;
3643 case BPF_JGE:
3644 if ((dst_reg->type == PTR_TO_PACKET &&
3645 src_reg->type == PTR_TO_PACKET_END) ||
3646 (dst_reg->type == PTR_TO_PACKET_META &&
3647 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3648 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
3649 find_good_pkt_pointers(this_branch, dst_reg,
3650 dst_reg->type, true);
3651 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3652 src_reg->type == PTR_TO_PACKET) ||
3653 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3654 src_reg->type == PTR_TO_PACKET_META)) {
3655 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
3656 find_good_pkt_pointers(other_branch, src_reg,
3657 src_reg->type, false);
3658 } else {
3659 return false;
3660 }
3661 break;
3662 case BPF_JLE:
3663 if ((dst_reg->type == PTR_TO_PACKET &&
3664 src_reg->type == PTR_TO_PACKET_END) ||
3665 (dst_reg->type == PTR_TO_PACKET_META &&
3666 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
3667 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
3668 find_good_pkt_pointers(other_branch, dst_reg,
3669 dst_reg->type, false);
3670 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
3671 src_reg->type == PTR_TO_PACKET) ||
3672 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
3673 src_reg->type == PTR_TO_PACKET_META)) {
3674 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
3675 find_good_pkt_pointers(this_branch, src_reg,
3676 src_reg->type, true);
3677 } else {
3678 return false;
3679 }
3680 break;
3681 default:
3682 return false;
3683 }
3684
3685 return true;
3686}
3687
3688static int check_cond_jmp_op(struct bpf_verifier_env *env,
3689 struct bpf_insn *insn, int *insn_idx)
3690{
3691 struct bpf_verifier_state *this_branch = env->cur_state;
3692 struct bpf_verifier_state *other_branch;
3693 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
3694 struct bpf_reg_state *dst_reg, *other_branch_regs;
3695 u8 opcode = BPF_OP(insn->code);
3696 int err;
3697
3698 if (opcode > BPF_JSLE) {
3699 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
3700 return -EINVAL;
3701 }
3702
3703 if (BPF_SRC(insn->code) == BPF_X) {
3704 if (insn->imm != 0) {
3705 verbose(env, "BPF_JMP uses reserved fields\n");
3706 return -EINVAL;
3707 }
3708
3709 /* check src1 operand */
3710 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3711 if (err)
3712 return err;
3713
3714 if (is_pointer_value(env, insn->src_reg)) {
3715 verbose(env, "R%d pointer comparison prohibited\n",
3716 insn->src_reg);
3717 return -EACCES;
3718 }
3719 } else {
3720 if (insn->src_reg != BPF_REG_0) {
3721 verbose(env, "BPF_JMP uses reserved fields\n");
3722 return -EINVAL;
3723 }
3724 }
3725
3726 /* check src2 operand */
3727 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3728 if (err)
3729 return err;
3730
3731 dst_reg = ®s[insn->dst_reg];
3732
3733 /* detect if R == 0 where R was initialized to zero earlier */
3734 if (BPF_SRC(insn->code) == BPF_K &&
3735 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3736 dst_reg->type == SCALAR_VALUE &&
3737 tnum_is_const(dst_reg->var_off)) {
3738 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
3739 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
3740 /* if (imm == imm) goto pc+off;
3741 * only follow the goto, ignore fall-through
3742 */
3743 *insn_idx += insn->off;
3744 return 0;
3745 } else {
3746 /* if (imm != imm) goto pc+off;
3747 * only follow fall-through branch, since
3748 * that's where the program will go
3749 */
3750 return 0;
3751 }
3752 }
3753
3754 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
3755 if (!other_branch)
3756 return -EFAULT;
3757 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
3758
3759 /* detect if we are comparing against a constant value so we can adjust
3760 * our min/max values for our dst register.
3761 * this is only legit if both are scalars (or pointers to the same
3762 * object, I suppose, but we don't support that right now), because
3763 * otherwise the different base pointers mean the offsets aren't
3764 * comparable.
3765 */
3766 if (BPF_SRC(insn->code) == BPF_X) {
3767 if (dst_reg->type == SCALAR_VALUE &&
3768 regs[insn->src_reg].type == SCALAR_VALUE) {
3769 if (tnum_is_const(regs[insn->src_reg].var_off))
3770 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3771 dst_reg, regs[insn->src_reg].var_off.value,
3772 opcode);
3773 else if (tnum_is_const(dst_reg->var_off))
3774 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
3775 ®s[insn->src_reg],
3776 dst_reg->var_off.value, opcode);
3777 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
3778 /* Comparing for equality, we can combine knowledge */
3779 reg_combine_min_max(&other_branch_regs[insn->src_reg],
3780 &other_branch_regs[insn->dst_reg],
3781 ®s[insn->src_reg],
3782 ®s[insn->dst_reg], opcode);
3783 }
3784 } else if (dst_reg->type == SCALAR_VALUE) {
3785 reg_set_min_max(&other_branch_regs[insn->dst_reg],
3786 dst_reg, insn->imm, opcode);
3787 }
3788
3789 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
3790 if (BPF_SRC(insn->code) == BPF_K &&
3791 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
3792 dst_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3793 /* Mark all identical map registers in each branch as either
3794 * safe or unknown depending R == 0 or R != 0 conditional.
3795 */
3796 mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
3797 mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
3798 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
3799 this_branch, other_branch) &&
3800 is_pointer_value(env, insn->dst_reg)) {
3801 verbose(env, "R%d pointer comparison prohibited\n",
3802 insn->dst_reg);
3803 return -EACCES;
3804 }
3805 if (env->log.level)
3806 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
3807 return 0;
3808}
3809
3810/* return the map pointer stored inside BPF_LD_IMM64 instruction */
3811static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
3812{
3813 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
3814
3815 return (struct bpf_map *) (unsigned long) imm64;
3816}
3817
3818/* verify BPF_LD_IMM64 instruction */
3819static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
3820{
3821 struct bpf_reg_state *regs = cur_regs(env);
3822 int err;
3823
3824 if (BPF_SIZE(insn->code) != BPF_DW) {
3825 verbose(env, "invalid BPF_LD_IMM insn\n");
3826 return -EINVAL;
3827 }
3828 if (insn->off != 0) {
3829 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
3830 return -EINVAL;
3831 }
3832
3833 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3834 if (err)
3835 return err;
3836
3837 if (insn->src_reg == 0) {
3838 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
3839
3840 regs[insn->dst_reg].type = SCALAR_VALUE;
3841 __mark_reg_known(®s[insn->dst_reg], imm);
3842 return 0;
3843 }
3844
3845 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
3846 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
3847
3848 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
3849 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
3850 return 0;
3851}
3852
3853static bool may_access_skb(enum bpf_prog_type type)
3854{
3855 switch (type) {
3856 case BPF_PROG_TYPE_SOCKET_FILTER:
3857 case BPF_PROG_TYPE_SCHED_CLS:
3858 case BPF_PROG_TYPE_SCHED_ACT:
3859 return true;
3860 default:
3861 return false;
3862 }
3863}
3864
3865/* verify safety of LD_ABS|LD_IND instructions:
3866 * - they can only appear in the programs where ctx == skb
3867 * - since they are wrappers of function calls, they scratch R1-R5 registers,
3868 * preserve R6-R9, and store return value into R0
3869 *
3870 * Implicit input:
3871 * ctx == skb == R6 == CTX
3872 *
3873 * Explicit input:
3874 * SRC == any register
3875 * IMM == 32-bit immediate
3876 *
3877 * Output:
3878 * R0 - 8/16/32-bit skb data converted to cpu endianness
3879 */
3880static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
3881{
3882 struct bpf_reg_state *regs = cur_regs(env);
3883 u8 mode = BPF_MODE(insn->code);
3884 int i, err;
3885
3886 if (!may_access_skb(env->prog->type)) {
3887 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
3888 return -EINVAL;
3889 }
3890
3891 if (env->subprog_cnt) {
3892 /* when program has LD_ABS insn JITs and interpreter assume
3893 * that r1 == ctx == skb which is not the case for callees
3894 * that can have arbitrary arguments. It's problematic
3895 * for main prog as well since JITs would need to analyze
3896 * all functions in order to make proper register save/restore
3897 * decisions in the main prog. Hence disallow LD_ABS with calls
3898 */
3899 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
3900 return -EINVAL;
3901 }
3902
3903 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
3904 BPF_SIZE(insn->code) == BPF_DW ||
3905 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
3906 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
3907 return -EINVAL;
3908 }
3909
3910 /* check whether implicit source operand (register R6) is readable */
3911 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
3912 if (err)
3913 return err;
3914
3915 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
3916 verbose(env,
3917 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
3918 return -EINVAL;
3919 }
3920
3921 if (mode == BPF_IND) {
3922 /* check explicit source operand */
3923 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3924 if (err)
3925 return err;
3926 }
3927
3928 /* reset caller saved regs to unreadable */
3929 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3930 mark_reg_not_init(env, regs, caller_saved[i]);
3931 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3932 }
3933
3934 /* mark destination R0 register as readable, since it contains
3935 * the value fetched from the packet.
3936 * Already marked as written above.
3937 */
3938 mark_reg_unknown(env, regs, BPF_REG_0);
3939 return 0;
3940}
3941
3942static int check_return_code(struct bpf_verifier_env *env)
3943{
3944 struct bpf_reg_state *reg;
3945 struct tnum range = tnum_range(0, 1);
3946
3947 switch (env->prog->type) {
3948 case BPF_PROG_TYPE_CGROUP_SKB:
3949 case BPF_PROG_TYPE_CGROUP_SOCK:
3950 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
3951 case BPF_PROG_TYPE_SOCK_OPS:
3952 case BPF_PROG_TYPE_CGROUP_DEVICE:
3953 break;
3954 default:
3955 return 0;
3956 }
3957
3958 reg = cur_regs(env) + BPF_REG_0;
3959 if (reg->type != SCALAR_VALUE) {
3960 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
3961 reg_type_str[reg->type]);
3962 return -EINVAL;
3963 }
3964
3965 if (!tnum_in(range, reg->var_off)) {
3966 verbose(env, "At program exit the register R0 ");
3967 if (!tnum_is_unknown(reg->var_off)) {
3968 char tn_buf[48];
3969
3970 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3971 verbose(env, "has value %s", tn_buf);
3972 } else {
3973 verbose(env, "has unknown scalar value");
3974 }
3975 verbose(env, " should have been 0 or 1\n");
3976 return -EINVAL;
3977 }
3978 return 0;
3979}
3980
3981/* non-recursive DFS pseudo code
3982 * 1 procedure DFS-iterative(G,v):
3983 * 2 label v as discovered
3984 * 3 let S be a stack
3985 * 4 S.push(v)
3986 * 5 while S is not empty
3987 * 6 t <- S.pop()
3988 * 7 if t is what we're looking for:
3989 * 8 return t
3990 * 9 for all edges e in G.adjacentEdges(t) do
3991 * 10 if edge e is already labelled
3992 * 11 continue with the next edge
3993 * 12 w <- G.adjacentVertex(t,e)
3994 * 13 if vertex w is not discovered and not explored
3995 * 14 label e as tree-edge
3996 * 15 label w as discovered
3997 * 16 S.push(w)
3998 * 17 continue at 5
3999 * 18 else if vertex w is discovered
4000 * 19 label e as back-edge
4001 * 20 else
4002 * 21 // vertex w is explored
4003 * 22 label e as forward- or cross-edge
4004 * 23 label t as explored
4005 * 24 S.pop()
4006 *
4007 * convention:
4008 * 0x10 - discovered
4009 * 0x11 - discovered and fall-through edge labelled
4010 * 0x12 - discovered and fall-through and branch edges labelled
4011 * 0x20 - explored
4012 */
4013
4014enum {
4015 DISCOVERED = 0x10,
4016 EXPLORED = 0x20,
4017 FALLTHROUGH = 1,
4018 BRANCH = 2,
4019};
4020
4021#define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4022
4023static int *insn_stack; /* stack of insns to process */
4024static int cur_stack; /* current stack index */
4025static int *insn_state;
4026
4027/* t, w, e - match pseudo-code above:
4028 * t - index of current instruction
4029 * w - next instruction
4030 * e - edge
4031 */
4032static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4033{
4034 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4035 return 0;
4036
4037 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4038 return 0;
4039
4040 if (w < 0 || w >= env->prog->len) {
4041 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4042 return -EINVAL;
4043 }
4044
4045 if (e == BRANCH)
4046 /* mark branch target for state pruning */
4047 env->explored_states[w] = STATE_LIST_MARK;
4048
4049 if (insn_state[w] == 0) {
4050 /* tree-edge */
4051 insn_state[t] = DISCOVERED | e;
4052 insn_state[w] = DISCOVERED;
4053 if (cur_stack >= env->prog->len)
4054 return -E2BIG;
4055 insn_stack[cur_stack++] = w;
4056 return 1;
4057 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4058 verbose(env, "back-edge from insn %d to %d\n", t, w);
4059 return -EINVAL;
4060 } else if (insn_state[w] == EXPLORED) {
4061 /* forward- or cross-edge */
4062 insn_state[t] = DISCOVERED | e;
4063 } else {
4064 verbose(env, "insn state internal bug\n");
4065 return -EFAULT;
4066 }
4067 return 0;
4068}
4069
4070/* non-recursive depth-first-search to detect loops in BPF program
4071 * loop == back-edge in directed graph
4072 */
4073static int check_cfg(struct bpf_verifier_env *env)
4074{
4075 struct bpf_insn *insns = env->prog->insnsi;
4076 int insn_cnt = env->prog->len;
4077 int ret = 0;
4078 int i, t;
4079
4080 ret = check_subprogs(env);
4081 if (ret < 0)
4082 return ret;
4083
4084 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4085 if (!insn_state)
4086 return -ENOMEM;
4087
4088 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4089 if (!insn_stack) {
4090 kfree(insn_state);
4091 return -ENOMEM;
4092 }
4093
4094 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4095 insn_stack[0] = 0; /* 0 is the first instruction */
4096 cur_stack = 1;
4097
4098peek_stack:
4099 if (cur_stack == 0)
4100 goto check_state;
4101 t = insn_stack[cur_stack - 1];
4102
4103 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4104 u8 opcode = BPF_OP(insns[t].code);
4105
4106 if (opcode == BPF_EXIT) {
4107 goto mark_explored;
4108 } else if (opcode == BPF_CALL) {
4109 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4110 if (ret == 1)
4111 goto peek_stack;
4112 else if (ret < 0)
4113 goto err_free;
4114 if (t + 1 < insn_cnt)
4115 env->explored_states[t + 1] = STATE_LIST_MARK;
4116 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4117 env->explored_states[t] = STATE_LIST_MARK;
4118 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4119 if (ret == 1)
4120 goto peek_stack;
4121 else if (ret < 0)
4122 goto err_free;
4123 }
4124 } else if (opcode == BPF_JA) {
4125 if (BPF_SRC(insns[t].code) != BPF_K) {
4126 ret = -EINVAL;
4127 goto err_free;
4128 }
4129 /* unconditional jump with single edge */
4130 ret = push_insn(t, t + insns[t].off + 1,
4131 FALLTHROUGH, env);
4132 if (ret == 1)
4133 goto peek_stack;
4134 else if (ret < 0)
4135 goto err_free;
4136 /* tell verifier to check for equivalent states
4137 * after every call and jump
4138 */
4139 if (t + 1 < insn_cnt)
4140 env->explored_states[t + 1] = STATE_LIST_MARK;
4141 } else {
4142 /* conditional jump with two edges */
4143 env->explored_states[t] = STATE_LIST_MARK;
4144 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4145 if (ret == 1)
4146 goto peek_stack;
4147 else if (ret < 0)
4148 goto err_free;
4149
4150 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4151 if (ret == 1)
4152 goto peek_stack;
4153 else if (ret < 0)
4154 goto err_free;
4155 }
4156 } else {
4157 /* all other non-branch instructions with single
4158 * fall-through edge
4159 */
4160 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4161 if (ret == 1)
4162 goto peek_stack;
4163 else if (ret < 0)
4164 goto err_free;
4165 }
4166
4167mark_explored:
4168 insn_state[t] = EXPLORED;
4169 if (cur_stack-- <= 0) {
4170 verbose(env, "pop stack internal bug\n");
4171 ret = -EFAULT;
4172 goto err_free;
4173 }
4174 goto peek_stack;
4175
4176check_state:
4177 for (i = 0; i < insn_cnt; i++) {
4178 if (insn_state[i] != EXPLORED) {
4179 verbose(env, "unreachable insn %d\n", i);
4180 ret = -EINVAL;
4181 goto err_free;
4182 }
4183 }
4184 ret = 0; /* cfg looks good */
4185
4186err_free:
4187 kfree(insn_state);
4188 kfree(insn_stack);
4189 return ret;
4190}
4191
4192/* check %cur's range satisfies %old's */
4193static bool range_within(struct bpf_reg_state *old,
4194 struct bpf_reg_state *cur)
4195{
4196 return old->umin_value <= cur->umin_value &&
4197 old->umax_value >= cur->umax_value &&
4198 old->smin_value <= cur->smin_value &&
4199 old->smax_value >= cur->smax_value;
4200}
4201
4202/* Maximum number of register states that can exist at once */
4203#define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4204struct idpair {
4205 u32 old;
4206 u32 cur;
4207};
4208
4209/* If in the old state two registers had the same id, then they need to have
4210 * the same id in the new state as well. But that id could be different from
4211 * the old state, so we need to track the mapping from old to new ids.
4212 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4213 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4214 * regs with a different old id could still have new id 9, we don't care about
4215 * that.
4216 * So we look through our idmap to see if this old id has been seen before. If
4217 * so, we require the new id to match; otherwise, we add the id pair to the map.
4218 */
4219static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4220{
4221 unsigned int i;
4222
4223 for (i = 0; i < ID_MAP_SIZE; i++) {
4224 if (!idmap[i].old) {
4225 /* Reached an empty slot; haven't seen this id before */
4226 idmap[i].old = old_id;
4227 idmap[i].cur = cur_id;
4228 return true;
4229 }
4230 if (idmap[i].old == old_id)
4231 return idmap[i].cur == cur_id;
4232 }
4233 /* We ran out of idmap slots, which should be impossible */
4234 WARN_ON_ONCE(1);
4235 return false;
4236}
4237
4238/* Returns true if (rold safe implies rcur safe) */
4239static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4240 struct idpair *idmap)
4241{
4242 bool equal;
4243
4244 if (!(rold->live & REG_LIVE_READ))
4245 /* explored state didn't use this */
4246 return true;
4247
4248 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, frameno)) == 0;
4249
4250 if (rold->type == PTR_TO_STACK)
4251 /* two stack pointers are equal only if they're pointing to
4252 * the same stack frame, since fp-8 in foo != fp-8 in bar
4253 */
4254 return equal && rold->frameno == rcur->frameno;
4255
4256 if (equal)
4257 return true;
4258
4259 if (rold->type == NOT_INIT)
4260 /* explored state can't have used this */
4261 return true;
4262 if (rcur->type == NOT_INIT)
4263 return false;
4264 switch (rold->type) {
4265 case SCALAR_VALUE:
4266 if (rcur->type == SCALAR_VALUE) {
4267 /* new val must satisfy old val knowledge */
4268 return range_within(rold, rcur) &&
4269 tnum_in(rold->var_off, rcur->var_off);
4270 } else {
4271 /* We're trying to use a pointer in place of a scalar.
4272 * Even if the scalar was unbounded, this could lead to
4273 * pointer leaks because scalars are allowed to leak
4274 * while pointers are not. We could make this safe in
4275 * special cases if root is calling us, but it's
4276 * probably not worth the hassle.
4277 */
4278 return false;
4279 }
4280 case PTR_TO_MAP_VALUE:
4281 /* If the new min/max/var_off satisfy the old ones and
4282 * everything else matches, we are OK.
4283 * We don't care about the 'id' value, because nothing
4284 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4285 */
4286 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4287 range_within(rold, rcur) &&
4288 tnum_in(rold->var_off, rcur->var_off);
4289 case PTR_TO_MAP_VALUE_OR_NULL:
4290 /* a PTR_TO_MAP_VALUE could be safe to use as a
4291 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4292 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4293 * checked, doing so could have affected others with the same
4294 * id, and we can't check for that because we lost the id when
4295 * we converted to a PTR_TO_MAP_VALUE.
4296 */
4297 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4298 return false;
4299 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4300 return false;
4301 /* Check our ids match any regs they're supposed to */
4302 return check_ids(rold->id, rcur->id, idmap);
4303 case PTR_TO_PACKET_META:
4304 case PTR_TO_PACKET:
4305 if (rcur->type != rold->type)
4306 return false;
4307 /* We must have at least as much range as the old ptr
4308 * did, so that any accesses which were safe before are
4309 * still safe. This is true even if old range < old off,
4310 * since someone could have accessed through (ptr - k), or
4311 * even done ptr -= k in a register, to get a safe access.
4312 */
4313 if (rold->range > rcur->range)
4314 return false;
4315 /* If the offsets don't match, we can't trust our alignment;
4316 * nor can we be sure that we won't fall out of range.
4317 */
4318 if (rold->off != rcur->off)
4319 return false;
4320 /* id relations must be preserved */
4321 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4322 return false;
4323 /* new val must satisfy old val knowledge */
4324 return range_within(rold, rcur) &&
4325 tnum_in(rold->var_off, rcur->var_off);
4326 case PTR_TO_CTX:
4327 case CONST_PTR_TO_MAP:
4328 case PTR_TO_PACKET_END:
4329 /* Only valid matches are exact, which memcmp() above
4330 * would have accepted
4331 */
4332 default:
4333 /* Don't know what's going on, just say it's not safe */
4334 return false;
4335 }
4336
4337 /* Shouldn't get here; if we do, say it's not safe */
4338 WARN_ON_ONCE(1);
4339 return false;
4340}
4341
4342static bool stacksafe(struct bpf_func_state *old,
4343 struct bpf_func_state *cur,
4344 struct idpair *idmap)
4345{
4346 int i, spi;
4347
4348 /* if explored stack has more populated slots than current stack
4349 * such stacks are not equivalent
4350 */
4351 if (old->allocated_stack > cur->allocated_stack)
4352 return false;
4353
4354 /* walk slots of the explored stack and ignore any additional
4355 * slots in the current stack, since explored(safe) state
4356 * didn't use them
4357 */
4358 for (i = 0; i < old->allocated_stack; i++) {
4359 spi = i / BPF_REG_SIZE;
4360
4361 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4362 /* explored state didn't use this */
4363 continue;
4364
4365 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4366 continue;
4367 /* if old state was safe with misc data in the stack
4368 * it will be safe with zero-initialized stack.
4369 * The opposite is not true
4370 */
4371 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4372 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4373 continue;
4374 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4375 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4376 /* Ex: old explored (safe) state has STACK_SPILL in
4377 * this stack slot, but current has has STACK_MISC ->
4378 * this verifier states are not equivalent,
4379 * return false to continue verification of this path
4380 */
4381 return false;
4382 if (i % BPF_REG_SIZE)
4383 continue;
4384 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4385 continue;
4386 if (!regsafe(&old->stack[spi].spilled_ptr,
4387 &cur->stack[spi].spilled_ptr,
4388 idmap))
4389 /* when explored and current stack slot are both storing
4390 * spilled registers, check that stored pointers types
4391 * are the same as well.
4392 * Ex: explored safe path could have stored
4393 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4394 * but current path has stored:
4395 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4396 * such verifier states are not equivalent.
4397 * return false to continue verification of this path
4398 */
4399 return false;
4400 }
4401 return true;
4402}
4403
4404/* compare two verifier states
4405 *
4406 * all states stored in state_list are known to be valid, since
4407 * verifier reached 'bpf_exit' instruction through them
4408 *
4409 * this function is called when verifier exploring different branches of
4410 * execution popped from the state stack. If it sees an old state that has
4411 * more strict register state and more strict stack state then this execution
4412 * branch doesn't need to be explored further, since verifier already
4413 * concluded that more strict state leads to valid finish.
4414 *
4415 * Therefore two states are equivalent if register state is more conservative
4416 * and explored stack state is more conservative than the current one.
4417 * Example:
4418 * explored current
4419 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4420 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4421 *
4422 * In other words if current stack state (one being explored) has more
4423 * valid slots than old one that already passed validation, it means
4424 * the verifier can stop exploring and conclude that current state is valid too
4425 *
4426 * Similarly with registers. If explored state has register type as invalid
4427 * whereas register type in current state is meaningful, it means that
4428 * the current state will reach 'bpf_exit' instruction safely
4429 */
4430static bool func_states_equal(struct bpf_func_state *old,
4431 struct bpf_func_state *cur)
4432{
4433 struct idpair *idmap;
4434 bool ret = false;
4435 int i;
4436
4437 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4438 /* If we failed to allocate the idmap, just say it's not safe */
4439 if (!idmap)
4440 return false;
4441
4442 for (i = 0; i < MAX_BPF_REG; i++) {
4443 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4444 goto out_free;
4445 }
4446
4447 if (!stacksafe(old, cur, idmap))
4448 goto out_free;
4449 ret = true;
4450out_free:
4451 kfree(idmap);
4452 return ret;
4453}
4454
4455static bool states_equal(struct bpf_verifier_env *env,
4456 struct bpf_verifier_state *old,
4457 struct bpf_verifier_state *cur)
4458{
4459 int i;
4460
4461 if (old->curframe != cur->curframe)
4462 return false;
4463
4464 /* for states to be equal callsites have to be the same
4465 * and all frame states need to be equivalent
4466 */
4467 for (i = 0; i <= old->curframe; i++) {
4468 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4469 return false;
4470 if (!func_states_equal(old->frame[i], cur->frame[i]))
4471 return false;
4472 }
4473 return true;
4474}
4475
4476/* A write screens off any subsequent reads; but write marks come from the
4477 * straight-line code between a state and its parent. When we arrive at an
4478 * equivalent state (jump target or such) we didn't arrive by the straight-line
4479 * code, so read marks in the state must propagate to the parent regardless
4480 * of the state's write marks. That's what 'parent == state->parent' comparison
4481 * in mark_reg_read() and mark_stack_slot_read() is for.
4482 */
4483static int propagate_liveness(struct bpf_verifier_env *env,
4484 const struct bpf_verifier_state *vstate,
4485 struct bpf_verifier_state *vparent)
4486{
4487 int i, frame, err = 0;
4488 struct bpf_func_state *state, *parent;
4489
4490 if (vparent->curframe != vstate->curframe) {
4491 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4492 vparent->curframe, vstate->curframe);
4493 return -EFAULT;
4494 }
4495 /* Propagate read liveness of registers... */
4496 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4497 /* We don't need to worry about FP liveness because it's read-only */
4498 for (i = 0; i < BPF_REG_FP; i++) {
4499 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4500 continue;
4501 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4502 err = mark_reg_read(env, vstate, vparent, i);
4503 if (err)
4504 return err;
4505 }
4506 }
4507
4508 /* ... and stack slots */
4509 for (frame = 0; frame <= vstate->curframe; frame++) {
4510 state = vstate->frame[frame];
4511 parent = vparent->frame[frame];
4512 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4513 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4514 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4515 continue;
4516 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4517 mark_stack_slot_read(env, vstate, vparent, i, frame);
4518 }
4519 }
4520 return err;
4521}
4522
4523static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4524{
4525 struct bpf_verifier_state_list *new_sl;
4526 struct bpf_verifier_state_list *sl;
4527 struct bpf_verifier_state *cur = env->cur_state;
4528 int i, j, err;
4529
4530 sl = env->explored_states[insn_idx];
4531 if (!sl)
4532 /* this 'insn_idx' instruction wasn't marked, so we will not
4533 * be doing state search here
4534 */
4535 return 0;
4536
4537 while (sl != STATE_LIST_MARK) {
4538 if (states_equal(env, &sl->state, cur)) {
4539 /* reached equivalent register/stack state,
4540 * prune the search.
4541 * Registers read by the continuation are read by us.
4542 * If we have any write marks in env->cur_state, they
4543 * will prevent corresponding reads in the continuation
4544 * from reaching our parent (an explored_state). Our
4545 * own state will get the read marks recorded, but
4546 * they'll be immediately forgotten as we're pruning
4547 * this state and will pop a new one.
4548 */
4549 err = propagate_liveness(env, &sl->state, cur);
4550 if (err)
4551 return err;
4552 return 1;
4553 }
4554 sl = sl->next;
4555 }
4556
4557 /* there were no equivalent states, remember current one.
4558 * technically the current state is not proven to be safe yet,
4559 * but it will either reach outer most bpf_exit (which means it's safe)
4560 * or it will be rejected. Since there are no loops, we won't be
4561 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4562 * again on the way to bpf_exit
4563 */
4564 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
4565 if (!new_sl)
4566 return -ENOMEM;
4567
4568 /* add new state to the head of linked list */
4569 err = copy_verifier_state(&new_sl->state, cur);
4570 if (err) {
4571 free_verifier_state(&new_sl->state, false);
4572 kfree(new_sl);
4573 return err;
4574 }
4575 new_sl->next = env->explored_states[insn_idx];
4576 env->explored_states[insn_idx] = new_sl;
4577 /* connect new state to parentage chain */
4578 cur->parent = &new_sl->state;
4579 /* clear write marks in current state: the writes we did are not writes
4580 * our child did, so they don't screen off its reads from us.
4581 * (There are no read marks in current state, because reads always mark
4582 * their parent and current state never has children yet. Only
4583 * explored_states can get read marks.)
4584 */
4585 for (i = 0; i < BPF_REG_FP; i++)
4586 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
4587
4588 /* all stack frames are accessible from callee, clear them all */
4589 for (j = 0; j <= cur->curframe; j++) {
4590 struct bpf_func_state *frame = cur->frame[j];
4591
4592 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++)
4593 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
4594 }
4595 return 0;
4596}
4597
4598static int do_check(struct bpf_verifier_env *env)
4599{
4600 struct bpf_verifier_state *state;
4601 struct bpf_insn *insns = env->prog->insnsi;
4602 struct bpf_reg_state *regs;
4603 int insn_cnt = env->prog->len, i;
4604 int insn_idx, prev_insn_idx = 0;
4605 int insn_processed = 0;
4606 bool do_print_state = false;
4607
4608 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
4609 if (!state)
4610 return -ENOMEM;
4611 state->curframe = 0;
4612 state->parent = NULL;
4613 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
4614 if (!state->frame[0]) {
4615 kfree(state);
4616 return -ENOMEM;
4617 }
4618 env->cur_state = state;
4619 init_func_state(env, state->frame[0],
4620 BPF_MAIN_FUNC /* callsite */,
4621 0 /* frameno */,
4622 0 /* subprogno, zero == main subprog */);
4623 insn_idx = 0;
4624 for (;;) {
4625 struct bpf_insn *insn;
4626 u8 class;
4627 int err;
4628
4629 if (insn_idx >= insn_cnt) {
4630 verbose(env, "invalid insn idx %d insn_cnt %d\n",
4631 insn_idx, insn_cnt);
4632 return -EFAULT;
4633 }
4634
4635 insn = &insns[insn_idx];
4636 class = BPF_CLASS(insn->code);
4637
4638 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
4639 verbose(env,
4640 "BPF program is too large. Processed %d insn\n",
4641 insn_processed);
4642 return -E2BIG;
4643 }
4644
4645 err = is_state_visited(env, insn_idx);
4646 if (err < 0)
4647 return err;
4648 if (err == 1) {
4649 /* found equivalent state, can prune the search */
4650 if (env->log.level) {
4651 if (do_print_state)
4652 verbose(env, "\nfrom %d to %d: safe\n",
4653 prev_insn_idx, insn_idx);
4654 else
4655 verbose(env, "%d: safe\n", insn_idx);
4656 }
4657 goto process_bpf_exit;
4658 }
4659
4660 if (need_resched())
4661 cond_resched();
4662
4663 if (env->log.level > 1 || (env->log.level && do_print_state)) {
4664 if (env->log.level > 1)
4665 verbose(env, "%d:", insn_idx);
4666 else
4667 verbose(env, "\nfrom %d to %d:",
4668 prev_insn_idx, insn_idx);
4669 print_verifier_state(env, state->frame[state->curframe]);
4670 do_print_state = false;
4671 }
4672
4673 if (env->log.level) {
4674 const struct bpf_insn_cbs cbs = {
4675 .cb_print = verbose,
4676 .private_data = env,
4677 };
4678
4679 verbose(env, "%d: ", insn_idx);
4680 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4681 }
4682
4683 if (bpf_prog_is_dev_bound(env->prog->aux)) {
4684 err = bpf_prog_offload_verify_insn(env, insn_idx,
4685 prev_insn_idx);
4686 if (err)
4687 return err;
4688 }
4689
4690 regs = cur_regs(env);
4691 env->insn_aux_data[insn_idx].seen = true;
4692 if (class == BPF_ALU || class == BPF_ALU64) {
4693 err = check_alu_op(env, insn);
4694 if (err)
4695 return err;
4696
4697 } else if (class == BPF_LDX) {
4698 enum bpf_reg_type *prev_src_type, src_reg_type;
4699
4700 /* check for reserved fields is already done */
4701
4702 /* check src operand */
4703 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4704 if (err)
4705 return err;
4706
4707 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4708 if (err)
4709 return err;
4710
4711 src_reg_type = regs[insn->src_reg].type;
4712
4713 /* check that memory (src_reg + off) is readable,
4714 * the state of dst_reg will be updated by this func
4715 */
4716 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
4717 BPF_SIZE(insn->code), BPF_READ,
4718 insn->dst_reg, false);
4719 if (err)
4720 return err;
4721
4722 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
4723
4724 if (*prev_src_type == NOT_INIT) {
4725 /* saw a valid insn
4726 * dst_reg = *(u32 *)(src_reg + off)
4727 * save type to validate intersecting paths
4728 */
4729 *prev_src_type = src_reg_type;
4730
4731 } else if (src_reg_type != *prev_src_type &&
4732 (src_reg_type == PTR_TO_CTX ||
4733 *prev_src_type == PTR_TO_CTX)) {
4734 /* ABuser program is trying to use the same insn
4735 * dst_reg = *(u32*) (src_reg + off)
4736 * with different pointer types:
4737 * src_reg == ctx in one branch and
4738 * src_reg == stack|map in some other branch.
4739 * Reject it.
4740 */
4741 verbose(env, "same insn cannot be used with different pointers\n");
4742 return -EINVAL;
4743 }
4744
4745 } else if (class == BPF_STX) {
4746 enum bpf_reg_type *prev_dst_type, dst_reg_type;
4747
4748 if (BPF_MODE(insn->code) == BPF_XADD) {
4749 err = check_xadd(env, insn_idx, insn);
4750 if (err)
4751 return err;
4752 insn_idx++;
4753 continue;
4754 }
4755
4756 /* check src1 operand */
4757 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4758 if (err)
4759 return err;
4760 /* check src2 operand */
4761 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4762 if (err)
4763 return err;
4764
4765 dst_reg_type = regs[insn->dst_reg].type;
4766
4767 /* check that memory (dst_reg + off) is writeable */
4768 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4769 BPF_SIZE(insn->code), BPF_WRITE,
4770 insn->src_reg, false);
4771 if (err)
4772 return err;
4773
4774 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
4775
4776 if (*prev_dst_type == NOT_INIT) {
4777 *prev_dst_type = dst_reg_type;
4778 } else if (dst_reg_type != *prev_dst_type &&
4779 (dst_reg_type == PTR_TO_CTX ||
4780 *prev_dst_type == PTR_TO_CTX)) {
4781 verbose(env, "same insn cannot be used with different pointers\n");
4782 return -EINVAL;
4783 }
4784
4785 } else if (class == BPF_ST) {
4786 if (BPF_MODE(insn->code) != BPF_MEM ||
4787 insn->src_reg != BPF_REG_0) {
4788 verbose(env, "BPF_ST uses reserved fields\n");
4789 return -EINVAL;
4790 }
4791 /* check src operand */
4792 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4793 if (err)
4794 return err;
4795
4796 if (is_ctx_reg(env, insn->dst_reg)) {
4797 verbose(env, "BPF_ST stores into R%d context is not allowed\n",
4798 insn->dst_reg);
4799 return -EACCES;
4800 }
4801
4802 /* check that memory (dst_reg + off) is writeable */
4803 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4804 BPF_SIZE(insn->code), BPF_WRITE,
4805 -1, false);
4806 if (err)
4807 return err;
4808
4809 } else if (class == BPF_JMP) {
4810 u8 opcode = BPF_OP(insn->code);
4811
4812 if (opcode == BPF_CALL) {
4813 if (BPF_SRC(insn->code) != BPF_K ||
4814 insn->off != 0 ||
4815 (insn->src_reg != BPF_REG_0 &&
4816 insn->src_reg != BPF_PSEUDO_CALL) ||
4817 insn->dst_reg != BPF_REG_0) {
4818 verbose(env, "BPF_CALL uses reserved fields\n");
4819 return -EINVAL;
4820 }
4821
4822 if (insn->src_reg == BPF_PSEUDO_CALL)
4823 err = check_func_call(env, insn, &insn_idx);
4824 else
4825 err = check_helper_call(env, insn->imm, insn_idx);
4826 if (err)
4827 return err;
4828
4829 } else if (opcode == BPF_JA) {
4830 if (BPF_SRC(insn->code) != BPF_K ||
4831 insn->imm != 0 ||
4832 insn->src_reg != BPF_REG_0 ||
4833 insn->dst_reg != BPF_REG_0) {
4834 verbose(env, "BPF_JA uses reserved fields\n");
4835 return -EINVAL;
4836 }
4837
4838 insn_idx += insn->off + 1;
4839 continue;
4840
4841 } else if (opcode == BPF_EXIT) {
4842 if (BPF_SRC(insn->code) != BPF_K ||
4843 insn->imm != 0 ||
4844 insn->src_reg != BPF_REG_0 ||
4845 insn->dst_reg != BPF_REG_0) {
4846 verbose(env, "BPF_EXIT uses reserved fields\n");
4847 return -EINVAL;
4848 }
4849
4850 if (state->curframe) {
4851 /* exit from nested function */
4852 prev_insn_idx = insn_idx;
4853 err = prepare_func_exit(env, &insn_idx);
4854 if (err)
4855 return err;
4856 do_print_state = true;
4857 continue;
4858 }
4859
4860 /* eBPF calling convetion is such that R0 is used
4861 * to return the value from eBPF program.
4862 * Make sure that it's readable at this time
4863 * of bpf_exit, which means that program wrote
4864 * something into it earlier
4865 */
4866 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4867 if (err)
4868 return err;
4869
4870 if (is_pointer_value(env, BPF_REG_0)) {
4871 verbose(env, "R0 leaks addr as return value\n");
4872 return -EACCES;
4873 }
4874
4875 err = check_return_code(env);
4876 if (err)
4877 return err;
4878process_bpf_exit:
4879 err = pop_stack(env, &prev_insn_idx, &insn_idx);
4880 if (err < 0) {
4881 if (err != -ENOENT)
4882 return err;
4883 break;
4884 } else {
4885 do_print_state = true;
4886 continue;
4887 }
4888 } else {
4889 err = check_cond_jmp_op(env, insn, &insn_idx);
4890 if (err)
4891 return err;
4892 }
4893 } else if (class == BPF_LD) {
4894 u8 mode = BPF_MODE(insn->code);
4895
4896 if (mode == BPF_ABS || mode == BPF_IND) {
4897 err = check_ld_abs(env, insn);
4898 if (err)
4899 return err;
4900
4901 } else if (mode == BPF_IMM) {
4902 err = check_ld_imm(env, insn);
4903 if (err)
4904 return err;
4905
4906 insn_idx++;
4907 env->insn_aux_data[insn_idx].seen = true;
4908 } else {
4909 verbose(env, "invalid BPF_LD mode\n");
4910 return -EINVAL;
4911 }
4912 } else {
4913 verbose(env, "unknown insn class %d\n", class);
4914 return -EINVAL;
4915 }
4916
4917 insn_idx++;
4918 }
4919
4920 verbose(env, "processed %d insns (limit %d), stack depth ",
4921 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
4922 for (i = 0; i < env->subprog_cnt + 1; i++) {
4923 u32 depth = env->subprog_stack_depth[i];
4924
4925 verbose(env, "%d", depth);
4926 if (i + 1 < env->subprog_cnt + 1)
4927 verbose(env, "+");
4928 }
4929 verbose(env, "\n");
4930 env->prog->aux->stack_depth = env->subprog_stack_depth[0];
4931 return 0;
4932}
4933
4934static int check_map_prealloc(struct bpf_map *map)
4935{
4936 return (map->map_type != BPF_MAP_TYPE_HASH &&
4937 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
4938 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
4939 !(map->map_flags & BPF_F_NO_PREALLOC);
4940}
4941
4942static int check_map_prog_compatibility(struct bpf_verifier_env *env,
4943 struct bpf_map *map,
4944 struct bpf_prog *prog)
4945
4946{
4947 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
4948 * preallocated hash maps, since doing memory allocation
4949 * in overflow_handler can crash depending on where nmi got
4950 * triggered.
4951 */
4952 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
4953 if (!check_map_prealloc(map)) {
4954 verbose(env, "perf_event programs can only use preallocated hash map\n");
4955 return -EINVAL;
4956 }
4957 if (map->inner_map_meta &&
4958 !check_map_prealloc(map->inner_map_meta)) {
4959 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
4960 return -EINVAL;
4961 }
4962 }
4963
4964 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
4965 !bpf_offload_dev_match(prog, map)) {
4966 verbose(env, "offload device mismatch between prog and map\n");
4967 return -EINVAL;
4968 }
4969
4970 return 0;
4971}
4972
4973/* look for pseudo eBPF instructions that access map FDs and
4974 * replace them with actual map pointers
4975 */
4976static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
4977{
4978 struct bpf_insn *insn = env->prog->insnsi;
4979 int insn_cnt = env->prog->len;
4980 int i, j, err;
4981
4982 err = bpf_prog_calc_tag(env->prog);
4983 if (err)
4984 return err;
4985
4986 for (i = 0; i < insn_cnt; i++, insn++) {
4987 if (BPF_CLASS(insn->code) == BPF_LDX &&
4988 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
4989 verbose(env, "BPF_LDX uses reserved fields\n");
4990 return -EINVAL;
4991 }
4992
4993 if (BPF_CLASS(insn->code) == BPF_STX &&
4994 ((BPF_MODE(insn->code) != BPF_MEM &&
4995 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
4996 verbose(env, "BPF_STX uses reserved fields\n");
4997 return -EINVAL;
4998 }
4999
5000 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5001 struct bpf_map *map;
5002 struct fd f;
5003
5004 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5005 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5006 insn[1].off != 0) {
5007 verbose(env, "invalid bpf_ld_imm64 insn\n");
5008 return -EINVAL;
5009 }
5010
5011 if (insn->src_reg == 0)
5012 /* valid generic load 64-bit imm */
5013 goto next_insn;
5014
5015 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5016 verbose(env,
5017 "unrecognized bpf_ld_imm64 insn\n");
5018 return -EINVAL;
5019 }
5020
5021 f = fdget(insn->imm);
5022 map = __bpf_map_get(f);
5023 if (IS_ERR(map)) {
5024 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5025 insn->imm);
5026 return PTR_ERR(map);
5027 }
5028
5029 err = check_map_prog_compatibility(env, map, env->prog);
5030 if (err) {
5031 fdput(f);
5032 return err;
5033 }
5034
5035 /* store map pointer inside BPF_LD_IMM64 instruction */
5036 insn[0].imm = (u32) (unsigned long) map;
5037 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5038
5039 /* check whether we recorded this map already */
5040 for (j = 0; j < env->used_map_cnt; j++)
5041 if (env->used_maps[j] == map) {
5042 fdput(f);
5043 goto next_insn;
5044 }
5045
5046 if (env->used_map_cnt >= MAX_USED_MAPS) {
5047 fdput(f);
5048 return -E2BIG;
5049 }
5050
5051 /* hold the map. If the program is rejected by verifier,
5052 * the map will be released by release_maps() or it
5053 * will be used by the valid program until it's unloaded
5054 * and all maps are released in free_bpf_prog_info()
5055 */
5056 map = bpf_map_inc(map, false);
5057 if (IS_ERR(map)) {
5058 fdput(f);
5059 return PTR_ERR(map);
5060 }
5061 env->used_maps[env->used_map_cnt++] = map;
5062
5063 fdput(f);
5064next_insn:
5065 insn++;
5066 i++;
5067 continue;
5068 }
5069
5070 /* Basic sanity check before we invest more work here. */
5071 if (!bpf_opcode_in_insntable(insn->code)) {
5072 verbose(env, "unknown opcode %02x\n", insn->code);
5073 return -EINVAL;
5074 }
5075 }
5076
5077 /* now all pseudo BPF_LD_IMM64 instructions load valid
5078 * 'struct bpf_map *' into a register instead of user map_fd.
5079 * These pointers will be used later by verifier to validate map access.
5080 */
5081 return 0;
5082}
5083
5084/* drop refcnt of maps used by the rejected program */
5085static void release_maps(struct bpf_verifier_env *env)
5086{
5087 int i;
5088
5089 for (i = 0; i < env->used_map_cnt; i++)
5090 bpf_map_put(env->used_maps[i]);
5091}
5092
5093/* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5094static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5095{
5096 struct bpf_insn *insn = env->prog->insnsi;
5097 int insn_cnt = env->prog->len;
5098 int i;
5099
5100 for (i = 0; i < insn_cnt; i++, insn++)
5101 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5102 insn->src_reg = 0;
5103}
5104
5105/* single env->prog->insni[off] instruction was replaced with the range
5106 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5107 * [0, off) and [off, end) to new locations, so the patched range stays zero
5108 */
5109static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5110 u32 off, u32 cnt)
5111{
5112 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5113 int i;
5114
5115 if (cnt == 1)
5116 return 0;
5117 new_data = vzalloc(sizeof(struct bpf_insn_aux_data) * prog_len);
5118 if (!new_data)
5119 return -ENOMEM;
5120 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5121 memcpy(new_data + off + cnt - 1, old_data + off,
5122 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5123 for (i = off; i < off + cnt - 1; i++)
5124 new_data[i].seen = true;
5125 env->insn_aux_data = new_data;
5126 vfree(old_data);
5127 return 0;
5128}
5129
5130static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5131{
5132 int i;
5133
5134 if (len == 1)
5135 return;
5136 for (i = 0; i < env->subprog_cnt; i++) {
5137 if (env->subprog_starts[i] < off)
5138 continue;
5139 env->subprog_starts[i] += len - 1;
5140 }
5141}
5142
5143static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5144 const struct bpf_insn *patch, u32 len)
5145{
5146 struct bpf_prog *new_prog;
5147
5148 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5149 if (!new_prog)
5150 return NULL;
5151 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5152 return NULL;
5153 adjust_subprog_starts(env, off, len);
5154 return new_prog;
5155}
5156
5157/* The verifier does more data flow analysis than llvm and will not
5158 * explore branches that are dead at run time. Malicious programs can
5159 * have dead code too. Therefore replace all dead at-run-time code
5160 * with 'ja -1'.
5161 *
5162 * Just nops are not optimal, e.g. if they would sit at the end of the
5163 * program and through another bug we would manage to jump there, then
5164 * we'd execute beyond program memory otherwise. Returning exception
5165 * code also wouldn't work since we can have subprogs where the dead
5166 * code could be located.
5167 */
5168static void sanitize_dead_code(struct bpf_verifier_env *env)
5169{
5170 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5171 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5172 struct bpf_insn *insn = env->prog->insnsi;
5173 const int insn_cnt = env->prog->len;
5174 int i;
5175
5176 for (i = 0; i < insn_cnt; i++) {
5177 if (aux_data[i].seen)
5178 continue;
5179 memcpy(insn + i, &trap, sizeof(trap));
5180 }
5181}
5182
5183/* convert load instructions that access fields of 'struct __sk_buff'
5184 * into sequence of instructions that access fields of 'struct sk_buff'
5185 */
5186static int convert_ctx_accesses(struct bpf_verifier_env *env)
5187{
5188 const struct bpf_verifier_ops *ops = env->ops;
5189 int i, cnt, size, ctx_field_size, delta = 0;
5190 const int insn_cnt = env->prog->len;
5191 struct bpf_insn insn_buf[16], *insn;
5192 struct bpf_prog *new_prog;
5193 enum bpf_access_type type;
5194 bool is_narrower_load;
5195 u32 target_size;
5196
5197 if (ops->gen_prologue) {
5198 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5199 env->prog);
5200 if (cnt >= ARRAY_SIZE(insn_buf)) {
5201 verbose(env, "bpf verifier is misconfigured\n");
5202 return -EINVAL;
5203 } else if (cnt) {
5204 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5205 if (!new_prog)
5206 return -ENOMEM;
5207
5208 env->prog = new_prog;
5209 delta += cnt - 1;
5210 }
5211 }
5212
5213 if (!ops->convert_ctx_access)
5214 return 0;
5215
5216 insn = env->prog->insnsi + delta;
5217
5218 for (i = 0; i < insn_cnt; i++, insn++) {
5219 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5220 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5221 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5222 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5223 type = BPF_READ;
5224 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5225 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5226 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5227 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5228 type = BPF_WRITE;
5229 else
5230 continue;
5231
5232 if (type == BPF_WRITE &&
5233 env->insn_aux_data[i + delta].sanitize_stack_off) {
5234 struct bpf_insn patch[] = {
5235 /* Sanitize suspicious stack slot with zero.
5236 * There are no memory dependencies for this store,
5237 * since it's only using frame pointer and immediate
5238 * constant of zero
5239 */
5240 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5241 env->insn_aux_data[i + delta].sanitize_stack_off,
5242 0),
5243 /* the original STX instruction will immediately
5244 * overwrite the same stack slot with appropriate value
5245 */
5246 *insn,
5247 };
5248
5249 cnt = ARRAY_SIZE(patch);
5250 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5251 if (!new_prog)
5252 return -ENOMEM;
5253
5254 delta += cnt - 1;
5255 env->prog = new_prog;
5256 insn = new_prog->insnsi + i + delta;
5257 continue;
5258 }
5259
5260 if (env->insn_aux_data[i + delta].ptr_type != PTR_TO_CTX)
5261 continue;
5262
5263 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5264 size = BPF_LDST_BYTES(insn);
5265
5266 /* If the read access is a narrower load of the field,
5267 * convert to a 4/8-byte load, to minimum program type specific
5268 * convert_ctx_access changes. If conversion is successful,
5269 * we will apply proper mask to the result.
5270 */
5271 is_narrower_load = size < ctx_field_size;
5272 if (is_narrower_load) {
5273 u32 off = insn->off;
5274 u8 size_code;
5275
5276 if (type == BPF_WRITE) {
5277 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5278 return -EINVAL;
5279 }
5280
5281 size_code = BPF_H;
5282 if (ctx_field_size == 4)
5283 size_code = BPF_W;
5284 else if (ctx_field_size == 8)
5285 size_code = BPF_DW;
5286
5287 insn->off = off & ~(ctx_field_size - 1);
5288 insn->code = BPF_LDX | BPF_MEM | size_code;
5289 }
5290
5291 target_size = 0;
5292 cnt = ops->convert_ctx_access(type, insn, insn_buf, env->prog,
5293 &target_size);
5294 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5295 (ctx_field_size && !target_size)) {
5296 verbose(env, "bpf verifier is misconfigured\n");
5297 return -EINVAL;
5298 }
5299
5300 if (is_narrower_load && size < target_size) {
5301 if (ctx_field_size <= 4)
5302 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5303 (1 << size * 8) - 1);
5304 else
5305 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5306 (1 << size * 8) - 1);
5307 }
5308
5309 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5310 if (!new_prog)
5311 return -ENOMEM;
5312
5313 delta += cnt - 1;
5314
5315 /* keep walking new program and skip insns we just inserted */
5316 env->prog = new_prog;
5317 insn = new_prog->insnsi + i + delta;
5318 }
5319
5320 return 0;
5321}
5322
5323static int jit_subprogs(struct bpf_verifier_env *env)
5324{
5325 struct bpf_prog *prog = env->prog, **func, *tmp;
5326 int i, j, subprog_start, subprog_end = 0, len, subprog;
5327 struct bpf_insn *insn;
5328 void *old_bpf_func;
5329 int err = -ENOMEM;
5330
5331 if (env->subprog_cnt == 0)
5332 return 0;
5333
5334 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5335 if (insn->code != (BPF_JMP | BPF_CALL) ||
5336 insn->src_reg != BPF_PSEUDO_CALL)
5337 continue;
5338 subprog = find_subprog(env, i + insn->imm + 1);
5339 if (subprog < 0) {
5340 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5341 i + insn->imm + 1);
5342 return -EFAULT;
5343 }
5344 /* temporarily remember subprog id inside insn instead of
5345 * aux_data, since next loop will split up all insns into funcs
5346 */
5347 insn->off = subprog + 1;
5348 /* remember original imm in case JIT fails and fallback
5349 * to interpreter will be needed
5350 */
5351 env->insn_aux_data[i].call_imm = insn->imm;
5352 /* point imm to __bpf_call_base+1 from JITs point of view */
5353 insn->imm = 1;
5354 }
5355
5356 func = kzalloc(sizeof(prog) * (env->subprog_cnt + 1), GFP_KERNEL);
5357 if (!func)
5358 return -ENOMEM;
5359
5360 for (i = 0; i <= env->subprog_cnt; i++) {
5361 subprog_start = subprog_end;
5362 if (env->subprog_cnt == i)
5363 subprog_end = prog->len;
5364 else
5365 subprog_end = env->subprog_starts[i];
5366
5367 len = subprog_end - subprog_start;
5368 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5369 if (!func[i])
5370 goto out_free;
5371 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5372 len * sizeof(struct bpf_insn));
5373 func[i]->type = prog->type;
5374 func[i]->len = len;
5375 if (bpf_prog_calc_tag(func[i]))
5376 goto out_free;
5377 func[i]->is_func = 1;
5378 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5379 * Long term would need debug info to populate names
5380 */
5381 func[i]->aux->name[0] = 'F';
5382 func[i]->aux->stack_depth = env->subprog_stack_depth[i];
5383 func[i]->jit_requested = 1;
5384 func[i] = bpf_int_jit_compile(func[i]);
5385 if (!func[i]->jited) {
5386 err = -ENOTSUPP;
5387 goto out_free;
5388 }
5389 cond_resched();
5390 }
5391 /* at this point all bpf functions were successfully JITed
5392 * now populate all bpf_calls with correct addresses and
5393 * run last pass of JIT
5394 */
5395 for (i = 0; i <= env->subprog_cnt; i++) {
5396 insn = func[i]->insnsi;
5397 for (j = 0; j < func[i]->len; j++, insn++) {
5398 if (insn->code != (BPF_JMP | BPF_CALL) ||
5399 insn->src_reg != BPF_PSEUDO_CALL)
5400 continue;
5401 subprog = insn->off;
5402 insn->off = 0;
5403 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5404 func[subprog]->bpf_func -
5405 __bpf_call_base;
5406 }
5407 }
5408 for (i = 0; i <= env->subprog_cnt; i++) {
5409 old_bpf_func = func[i]->bpf_func;
5410 tmp = bpf_int_jit_compile(func[i]);
5411 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5412 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5413 err = -EFAULT;
5414 goto out_free;
5415 }
5416 cond_resched();
5417 }
5418
5419 /* finally lock prog and jit images for all functions and
5420 * populate kallsysm
5421 */
5422 for (i = 0; i <= env->subprog_cnt; i++) {
5423 bpf_prog_lock_ro(func[i]);
5424 bpf_prog_kallsyms_add(func[i]);
5425 }
5426
5427 /* Last step: make now unused interpreter insns from main
5428 * prog consistent for later dump requests, so they can
5429 * later look the same as if they were interpreted only.
5430 */
5431 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5432 unsigned long addr;
5433
5434 if (insn->code != (BPF_JMP | BPF_CALL) ||
5435 insn->src_reg != BPF_PSEUDO_CALL)
5436 continue;
5437 insn->off = env->insn_aux_data[i].call_imm;
5438 subprog = find_subprog(env, i + insn->off + 1);
5439 addr = (unsigned long)func[subprog + 1]->bpf_func;
5440 addr &= PAGE_MASK;
5441 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5442 addr - __bpf_call_base;
5443 }
5444
5445 prog->jited = 1;
5446 prog->bpf_func = func[0]->bpf_func;
5447 prog->aux->func = func;
5448 prog->aux->func_cnt = env->subprog_cnt + 1;
5449 return 0;
5450out_free:
5451 for (i = 0; i <= env->subprog_cnt; i++)
5452 if (func[i])
5453 bpf_jit_free(func[i]);
5454 kfree(func);
5455 /* cleanup main prog to be interpreted */
5456 prog->jit_requested = 0;
5457 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5458 if (insn->code != (BPF_JMP | BPF_CALL) ||
5459 insn->src_reg != BPF_PSEUDO_CALL)
5460 continue;
5461 insn->off = 0;
5462 insn->imm = env->insn_aux_data[i].call_imm;
5463 }
5464 return err;
5465}
5466
5467static int fixup_call_args(struct bpf_verifier_env *env)
5468{
5469#ifndef CONFIG_BPF_JIT_ALWAYS_ON
5470 struct bpf_prog *prog = env->prog;
5471 struct bpf_insn *insn = prog->insnsi;
5472 int i, depth;
5473#endif
5474 int err;
5475
5476 err = 0;
5477 if (env->prog->jit_requested) {
5478 err = jit_subprogs(env);
5479 if (err == 0)
5480 return 0;
5481 }
5482#ifndef CONFIG_BPF_JIT_ALWAYS_ON
5483 for (i = 0; i < prog->len; i++, insn++) {
5484 if (insn->code != (BPF_JMP | BPF_CALL) ||
5485 insn->src_reg != BPF_PSEUDO_CALL)
5486 continue;
5487 depth = get_callee_stack_depth(env, insn, i);
5488 if (depth < 0)
5489 return depth;
5490 bpf_patch_call_args(insn, depth);
5491 }
5492 err = 0;
5493#endif
5494 return err;
5495}
5496
5497/* fixup insn->imm field of bpf_call instructions
5498 * and inline eligible helpers as explicit sequence of BPF instructions
5499 *
5500 * this function is called after eBPF program passed verification
5501 */
5502static int fixup_bpf_calls(struct bpf_verifier_env *env)
5503{
5504 struct bpf_prog *prog = env->prog;
5505 struct bpf_insn *insn = prog->insnsi;
5506 const struct bpf_func_proto *fn;
5507 const int insn_cnt = prog->len;
5508 struct bpf_insn_aux_data *aux;
5509 struct bpf_insn insn_buf[16];
5510 struct bpf_prog *new_prog;
5511 struct bpf_map *map_ptr;
5512 int i, cnt, delta = 0;
5513
5514 for (i = 0; i < insn_cnt; i++, insn++) {
5515 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
5516 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5517 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
5518 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5519 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
5520 struct bpf_insn mask_and_div[] = {
5521 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5522 /* Rx div 0 -> 0 */
5523 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
5524 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
5525 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
5526 *insn,
5527 };
5528 struct bpf_insn mask_and_mod[] = {
5529 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
5530 /* Rx mod 0 -> Rx */
5531 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
5532 *insn,
5533 };
5534 struct bpf_insn *patchlet;
5535
5536 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
5537 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
5538 patchlet = mask_and_div + (is64 ? 1 : 0);
5539 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
5540 } else {
5541 patchlet = mask_and_mod + (is64 ? 1 : 0);
5542 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
5543 }
5544
5545 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
5546 if (!new_prog)
5547 return -ENOMEM;
5548
5549 delta += cnt - 1;
5550 env->prog = prog = new_prog;
5551 insn = new_prog->insnsi + i + delta;
5552 continue;
5553 }
5554
5555 if (insn->code != (BPF_JMP | BPF_CALL))
5556 continue;
5557 if (insn->src_reg == BPF_PSEUDO_CALL)
5558 continue;
5559
5560 if (insn->imm == BPF_FUNC_get_route_realm)
5561 prog->dst_needed = 1;
5562 if (insn->imm == BPF_FUNC_get_prandom_u32)
5563 bpf_user_rnd_init_once();
5564 if (insn->imm == BPF_FUNC_override_return)
5565 prog->kprobe_override = 1;
5566 if (insn->imm == BPF_FUNC_tail_call) {
5567 /* If we tail call into other programs, we
5568 * cannot make any assumptions since they can
5569 * be replaced dynamically during runtime in
5570 * the program array.
5571 */
5572 prog->cb_access = 1;
5573 env->prog->aux->stack_depth = MAX_BPF_STACK;
5574
5575 /* mark bpf_tail_call as different opcode to avoid
5576 * conditional branch in the interpeter for every normal
5577 * call and to prevent accidental JITing by JIT compiler
5578 * that doesn't support bpf_tail_call yet
5579 */
5580 insn->imm = 0;
5581 insn->code = BPF_JMP | BPF_TAIL_CALL;
5582
5583 aux = &env->insn_aux_data[i + delta];
5584 if (!bpf_map_ptr_unpriv(aux))
5585 continue;
5586
5587 /* instead of changing every JIT dealing with tail_call
5588 * emit two extra insns:
5589 * if (index >= max_entries) goto out;
5590 * index &= array->index_mask;
5591 * to avoid out-of-bounds cpu speculation
5592 */
5593 if (bpf_map_ptr_poisoned(aux)) {
5594 verbose(env, "tail_call abusing map_ptr\n");
5595 return -EINVAL;
5596 }
5597
5598 map_ptr = BPF_MAP_PTR(aux->map_state);
5599 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
5600 map_ptr->max_entries, 2);
5601 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
5602 container_of(map_ptr,
5603 struct bpf_array,
5604 map)->index_mask);
5605 insn_buf[2] = *insn;
5606 cnt = 3;
5607 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5608 if (!new_prog)
5609 return -ENOMEM;
5610
5611 delta += cnt - 1;
5612 env->prog = prog = new_prog;
5613 insn = new_prog->insnsi + i + delta;
5614 continue;
5615 }
5616
5617 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
5618 * handlers are currently limited to 64 bit only.
5619 */
5620 if (prog->jit_requested && BITS_PER_LONG == 64 &&
5621 insn->imm == BPF_FUNC_map_lookup_elem) {
5622 aux = &env->insn_aux_data[i + delta];
5623 if (bpf_map_ptr_poisoned(aux))
5624 goto patch_call_imm;
5625
5626 map_ptr = BPF_MAP_PTR(aux->map_state);
5627 if (!map_ptr->ops->map_gen_lookup)
5628 goto patch_call_imm;
5629
5630 cnt = map_ptr->ops->map_gen_lookup(map_ptr, insn_buf);
5631 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
5632 verbose(env, "bpf verifier is misconfigured\n");
5633 return -EINVAL;
5634 }
5635
5636 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
5637 cnt);
5638 if (!new_prog)
5639 return -ENOMEM;
5640
5641 delta += cnt - 1;
5642
5643 /* keep walking new program and skip insns we just inserted */
5644 env->prog = prog = new_prog;
5645 insn = new_prog->insnsi + i + delta;
5646 continue;
5647 }
5648
5649 if (insn->imm == BPF_FUNC_redirect_map) {
5650 /* Note, we cannot use prog directly as imm as subsequent
5651 * rewrites would still change the prog pointer. The only
5652 * stable address we can use is aux, which also works with
5653 * prog clones during blinding.
5654 */
5655 u64 addr = (unsigned long)prog->aux;
5656 struct bpf_insn r4_ld[] = {
5657 BPF_LD_IMM64(BPF_REG_4, addr),
5658 *insn,
5659 };
5660 cnt = ARRAY_SIZE(r4_ld);
5661
5662 new_prog = bpf_patch_insn_data(env, i + delta, r4_ld, cnt);
5663 if (!new_prog)
5664 return -ENOMEM;
5665
5666 delta += cnt - 1;
5667 env->prog = prog = new_prog;
5668 insn = new_prog->insnsi + i + delta;
5669 }
5670patch_call_imm:
5671 fn = env->ops->get_func_proto(insn->imm, env->prog);
5672 /* all functions that have prototype and verifier allowed
5673 * programs to call them, must be real in-kernel functions
5674 */
5675 if (!fn->func) {
5676 verbose(env,
5677 "kernel subsystem misconfigured func %s#%d\n",
5678 func_id_name(insn->imm), insn->imm);
5679 return -EFAULT;
5680 }
5681 insn->imm = fn->func - __bpf_call_base;
5682 }
5683
5684 return 0;
5685}
5686
5687static void free_states(struct bpf_verifier_env *env)
5688{
5689 struct bpf_verifier_state_list *sl, *sln;
5690 int i;
5691
5692 if (!env->explored_states)
5693 return;
5694
5695 for (i = 0; i < env->prog->len; i++) {
5696 sl = env->explored_states[i];
5697
5698 if (sl)
5699 while (sl != STATE_LIST_MARK) {
5700 sln = sl->next;
5701 free_verifier_state(&sl->state, false);
5702 kfree(sl);
5703 sl = sln;
5704 }
5705 }
5706
5707 kfree(env->explored_states);
5708}
5709
5710int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
5711{
5712 struct bpf_verifier_env *env;
5713 struct bpf_verifier_log *log;
5714 int ret = -EINVAL;
5715
5716 /* no program is valid */
5717 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
5718 return -EINVAL;
5719
5720 /* 'struct bpf_verifier_env' can be global, but since it's not small,
5721 * allocate/free it every time bpf_check() is called
5722 */
5723 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
5724 if (!env)
5725 return -ENOMEM;
5726 log = &env->log;
5727
5728 env->insn_aux_data = vzalloc(sizeof(struct bpf_insn_aux_data) *
5729 (*prog)->len);
5730 ret = -ENOMEM;
5731 if (!env->insn_aux_data)
5732 goto err_free_env;
5733 env->prog = *prog;
5734 env->ops = bpf_verifier_ops[env->prog->type];
5735
5736 /* grab the mutex to protect few globals used by verifier */
5737 mutex_lock(&bpf_verifier_lock);
5738
5739 if (attr->log_level || attr->log_buf || attr->log_size) {
5740 /* user requested verbose verifier output
5741 * and supplied buffer to store the verification trace
5742 */
5743 log->level = attr->log_level;
5744 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
5745 log->len_total = attr->log_size;
5746
5747 ret = -EINVAL;
5748 /* log attributes have to be sane */
5749 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
5750 !log->level || !log->ubuf)
5751 goto err_unlock;
5752 }
5753
5754 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
5755 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
5756 env->strict_alignment = true;
5757
5758 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5759 ret = bpf_prog_offload_verifier_prep(env);
5760 if (ret)
5761 goto err_unlock;
5762 }
5763
5764 ret = replace_map_fd_with_map_ptr(env);
5765 if (ret < 0)
5766 goto skip_full_check;
5767
5768 env->explored_states = kcalloc(env->prog->len,
5769 sizeof(struct bpf_verifier_state_list *),
5770 GFP_USER);
5771 ret = -ENOMEM;
5772 if (!env->explored_states)
5773 goto skip_full_check;
5774
5775 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
5776
5777 ret = check_cfg(env);
5778 if (ret < 0)
5779 goto skip_full_check;
5780
5781 ret = do_check(env);
5782 if (env->cur_state) {
5783 free_verifier_state(env->cur_state, true);
5784 env->cur_state = NULL;
5785 }
5786
5787skip_full_check:
5788 while (!pop_stack(env, NULL, NULL));
5789 free_states(env);
5790
5791 if (ret == 0)
5792 sanitize_dead_code(env);
5793
5794 if (ret == 0)
5795 ret = check_max_stack_depth(env);
5796
5797 if (ret == 0)
5798 /* program is valid, convert *(u32*)(ctx + off) accesses */
5799 ret = convert_ctx_accesses(env);
5800
5801 if (ret == 0)
5802 ret = fixup_bpf_calls(env);
5803
5804 if (ret == 0)
5805 ret = fixup_call_args(env);
5806
5807 if (log->level && bpf_verifier_log_full(log))
5808 ret = -ENOSPC;
5809 if (log->level && !log->ubuf) {
5810 ret = -EFAULT;
5811 goto err_release_maps;
5812 }
5813
5814 if (ret == 0 && env->used_map_cnt) {
5815 /* if program passed verifier, update used_maps in bpf_prog_info */
5816 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
5817 sizeof(env->used_maps[0]),
5818 GFP_KERNEL);
5819
5820 if (!env->prog->aux->used_maps) {
5821 ret = -ENOMEM;
5822 goto err_release_maps;
5823 }
5824
5825 memcpy(env->prog->aux->used_maps, env->used_maps,
5826 sizeof(env->used_maps[0]) * env->used_map_cnt);
5827 env->prog->aux->used_map_cnt = env->used_map_cnt;
5828
5829 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
5830 * bpf_ld_imm64 instructions
5831 */
5832 convert_pseudo_ld_imm64(env);
5833 }
5834
5835err_release_maps:
5836 if (!env->prog->aux->used_maps)
5837 /* if we didn't copy map pointers into bpf_prog_info, release
5838 * them now. Otherwise free_bpf_prog_info() will release them.
5839 */
5840 release_maps(env);
5841 *prog = env->prog;
5842err_unlock:
5843 mutex_unlock(&bpf_verifier_lock);
5844 vfree(env->insn_aux_data);
5845err_free_env:
5846 kfree(env);
5847 return ret;
5848}