]> git.proxmox.com Git - mirror_qemu.git/blob - block/qcow2-refcount.c
qcow2: Don't rely on free_cluster_index in alloc_refcount_block() (CVE-2014-0147)
[mirror_qemu.git] / block / qcow2-refcount.c
1 /*
2 * Block driver for the QCOW version 2 format
3 *
4 * Copyright (c) 2004-2006 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25 #include "qemu-common.h"
26 #include "block/block_int.h"
27 #include "block/qcow2.h"
28 #include "qemu/range.h"
29 #include "qapi/qmp/types.h"
30
31 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size);
32 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
33 int64_t offset, int64_t length,
34 int addend, enum qcow2_discard_type type);
35
36
37 /*********************************************************/
38 /* refcount handling */
39
40 int qcow2_refcount_init(BlockDriverState *bs)
41 {
42 BDRVQcowState *s = bs->opaque;
43 unsigned int refcount_table_size2, i;
44 int ret;
45
46 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
47 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
48 s->refcount_table = g_malloc(refcount_table_size2);
49 if (s->refcount_table_size > 0) {
50 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
51 ret = bdrv_pread(bs->file, s->refcount_table_offset,
52 s->refcount_table, refcount_table_size2);
53 if (ret != refcount_table_size2)
54 goto fail;
55 for(i = 0; i < s->refcount_table_size; i++)
56 be64_to_cpus(&s->refcount_table[i]);
57 }
58 return 0;
59 fail:
60 return -ENOMEM;
61 }
62
63 void qcow2_refcount_close(BlockDriverState *bs)
64 {
65 BDRVQcowState *s = bs->opaque;
66 g_free(s->refcount_table);
67 }
68
69
70 static int load_refcount_block(BlockDriverState *bs,
71 int64_t refcount_block_offset,
72 void **refcount_block)
73 {
74 BDRVQcowState *s = bs->opaque;
75 int ret;
76
77 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
78 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
79 refcount_block);
80
81 return ret;
82 }
83
84 /*
85 * Returns the refcount of the cluster given by its index. Any non-negative
86 * return value is the refcount of the cluster, negative values are -errno
87 * and indicate an error.
88 */
89 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
90 {
91 BDRVQcowState *s = bs->opaque;
92 int refcount_table_index, block_index;
93 int64_t refcount_block_offset;
94 int ret;
95 uint16_t *refcount_block;
96 uint16_t refcount;
97
98 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
99 if (refcount_table_index >= s->refcount_table_size)
100 return 0;
101 refcount_block_offset =
102 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
103 if (!refcount_block_offset)
104 return 0;
105
106 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
107 (void**) &refcount_block);
108 if (ret < 0) {
109 return ret;
110 }
111
112 block_index = cluster_index &
113 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
114 refcount = be16_to_cpu(refcount_block[block_index]);
115
116 ret = qcow2_cache_put(bs, s->refcount_block_cache,
117 (void**) &refcount_block);
118 if (ret < 0) {
119 return ret;
120 }
121
122 return refcount;
123 }
124
125 /*
126 * Rounds the refcount table size up to avoid growing the table for each single
127 * refcount block that is allocated.
128 */
129 static unsigned int next_refcount_table_size(BDRVQcowState *s,
130 unsigned int min_size)
131 {
132 unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
133 unsigned int refcount_table_clusters =
134 MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
135
136 while (min_clusters > refcount_table_clusters) {
137 refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
138 }
139
140 return refcount_table_clusters << (s->cluster_bits - 3);
141 }
142
143
144 /* Checks if two offsets are described by the same refcount block */
145 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
146 uint64_t offset_b)
147 {
148 uint64_t block_a = offset_a >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
149 uint64_t block_b = offset_b >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
150
151 return (block_a == block_b);
152 }
153
154 /*
155 * Loads a refcount block. If it doesn't exist yet, it is allocated first
156 * (including growing the refcount table if needed).
157 *
158 * Returns 0 on success or -errno in error case
159 */
160 static int alloc_refcount_block(BlockDriverState *bs,
161 int64_t cluster_index, uint16_t **refcount_block)
162 {
163 BDRVQcowState *s = bs->opaque;
164 unsigned int refcount_table_index;
165 int ret;
166
167 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
168
169 /* Find the refcount block for the given cluster */
170 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
171
172 if (refcount_table_index < s->refcount_table_size) {
173
174 uint64_t refcount_block_offset =
175 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
176
177 /* If it's already there, we're done */
178 if (refcount_block_offset) {
179 return load_refcount_block(bs, refcount_block_offset,
180 (void**) refcount_block);
181 }
182 }
183
184 /*
185 * If we came here, we need to allocate something. Something is at least
186 * a cluster for the new refcount block. It may also include a new refcount
187 * table if the old refcount table is too small.
188 *
189 * Note that allocating clusters here needs some special care:
190 *
191 * - We can't use the normal qcow2_alloc_clusters(), it would try to
192 * increase the refcount and very likely we would end up with an endless
193 * recursion. Instead we must place the refcount blocks in a way that
194 * they can describe them themselves.
195 *
196 * - We need to consider that at this point we are inside update_refcounts
197 * and potentially doing an initial refcount increase. This means that
198 * some clusters have already been allocated by the caller, but their
199 * refcount isn't accurate yet. If we allocate clusters for metadata, we
200 * need to return -EAGAIN to signal the caller that it needs to restart
201 * the search for free clusters.
202 *
203 * - alloc_clusters_noref and qcow2_free_clusters may load a different
204 * refcount block into the cache
205 */
206
207 *refcount_block = NULL;
208
209 /* We write to the refcount table, so we might depend on L2 tables */
210 ret = qcow2_cache_flush(bs, s->l2_table_cache);
211 if (ret < 0) {
212 return ret;
213 }
214
215 /* Allocate the refcount block itself and mark it as used */
216 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
217 if (new_block < 0) {
218 return new_block;
219 }
220
221 #ifdef DEBUG_ALLOC2
222 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
223 " at %" PRIx64 "\n",
224 refcount_table_index, cluster_index << s->cluster_bits, new_block);
225 #endif
226
227 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
228 /* Zero the new refcount block before updating it */
229 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
230 (void**) refcount_block);
231 if (ret < 0) {
232 goto fail_block;
233 }
234
235 memset(*refcount_block, 0, s->cluster_size);
236
237 /* The block describes itself, need to update the cache */
238 int block_index = (new_block >> s->cluster_bits) &
239 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
240 (*refcount_block)[block_index] = cpu_to_be16(1);
241 } else {
242 /* Described somewhere else. This can recurse at most twice before we
243 * arrive at a block that describes itself. */
244 ret = update_refcount(bs, new_block, s->cluster_size, 1,
245 QCOW2_DISCARD_NEVER);
246 if (ret < 0) {
247 goto fail_block;
248 }
249
250 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
251 if (ret < 0) {
252 goto fail_block;
253 }
254
255 /* Initialize the new refcount block only after updating its refcount,
256 * update_refcount uses the refcount cache itself */
257 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
258 (void**) refcount_block);
259 if (ret < 0) {
260 goto fail_block;
261 }
262
263 memset(*refcount_block, 0, s->cluster_size);
264 }
265
266 /* Now the new refcount block needs to be written to disk */
267 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
268 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
269 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
270 if (ret < 0) {
271 goto fail_block;
272 }
273
274 /* If the refcount table is big enough, just hook the block up there */
275 if (refcount_table_index < s->refcount_table_size) {
276 uint64_t data64 = cpu_to_be64(new_block);
277 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
278 ret = bdrv_pwrite_sync(bs->file,
279 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
280 &data64, sizeof(data64));
281 if (ret < 0) {
282 goto fail_block;
283 }
284
285 s->refcount_table[refcount_table_index] = new_block;
286
287 /* The new refcount block may be where the caller intended to put its
288 * data, so let it restart the search. */
289 return -EAGAIN;
290 }
291
292 ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
293 if (ret < 0) {
294 goto fail_block;
295 }
296
297 /*
298 * If we come here, we need to grow the refcount table. Again, a new
299 * refcount table needs some space and we can't simply allocate to avoid
300 * endless recursion.
301 *
302 * Therefore let's grab new refcount blocks at the end of the image, which
303 * will describe themselves and the new refcount table. This way we can
304 * reference them only in the new table and do the switch to the new
305 * refcount table at once without producing an inconsistent state in
306 * between.
307 */
308 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
309
310 /* Calculate the number of refcount blocks needed so far */
311 uint64_t refcount_block_clusters = 1 << (s->cluster_bits - REFCOUNT_SHIFT);
312 uint64_t blocks_used = DIV_ROUND_UP(cluster_index, refcount_block_clusters);
313
314 /* And now we need at least one block more for the new metadata */
315 uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
316 uint64_t last_table_size;
317 uint64_t blocks_clusters;
318 do {
319 uint64_t table_clusters =
320 size_to_clusters(s, table_size * sizeof(uint64_t));
321 blocks_clusters = 1 +
322 ((table_clusters + refcount_block_clusters - 1)
323 / refcount_block_clusters);
324 uint64_t meta_clusters = table_clusters + blocks_clusters;
325
326 last_table_size = table_size;
327 table_size = next_refcount_table_size(s, blocks_used +
328 ((meta_clusters + refcount_block_clusters - 1)
329 / refcount_block_clusters));
330
331 } while (last_table_size != table_size);
332
333 #ifdef DEBUG_ALLOC2
334 fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
335 s->refcount_table_size, table_size);
336 #endif
337
338 /* Create the new refcount table and blocks */
339 uint64_t meta_offset = (blocks_used * refcount_block_clusters) *
340 s->cluster_size;
341 uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
342 uint16_t *new_blocks = g_malloc0(blocks_clusters * s->cluster_size);
343 uint64_t *new_table = g_malloc0(table_size * sizeof(uint64_t));
344
345 /* Fill the new refcount table */
346 memcpy(new_table, s->refcount_table,
347 s->refcount_table_size * sizeof(uint64_t));
348 new_table[refcount_table_index] = new_block;
349
350 int i;
351 for (i = 0; i < blocks_clusters; i++) {
352 new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
353 }
354
355 /* Fill the refcount blocks */
356 uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
357 int block = 0;
358 for (i = 0; i < table_clusters + blocks_clusters; i++) {
359 new_blocks[block++] = cpu_to_be16(1);
360 }
361
362 /* Write refcount blocks to disk */
363 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
364 ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
365 blocks_clusters * s->cluster_size);
366 g_free(new_blocks);
367 if (ret < 0) {
368 goto fail_table;
369 }
370
371 /* Write refcount table to disk */
372 for(i = 0; i < table_size; i++) {
373 cpu_to_be64s(&new_table[i]);
374 }
375
376 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
377 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
378 table_size * sizeof(uint64_t));
379 if (ret < 0) {
380 goto fail_table;
381 }
382
383 for(i = 0; i < table_size; i++) {
384 be64_to_cpus(&new_table[i]);
385 }
386
387 /* Hook up the new refcount table in the qcow2 header */
388 uint8_t data[12];
389 cpu_to_be64w((uint64_t*)data, table_offset);
390 cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
391 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
392 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
393 data, sizeof(data));
394 if (ret < 0) {
395 goto fail_table;
396 }
397
398 /* And switch it in memory */
399 uint64_t old_table_offset = s->refcount_table_offset;
400 uint64_t old_table_size = s->refcount_table_size;
401
402 g_free(s->refcount_table);
403 s->refcount_table = new_table;
404 s->refcount_table_size = table_size;
405 s->refcount_table_offset = table_offset;
406
407 /* Free old table. */
408 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
409 QCOW2_DISCARD_OTHER);
410
411 ret = load_refcount_block(bs, new_block, (void**) refcount_block);
412 if (ret < 0) {
413 return ret;
414 }
415
416 /* If we were trying to do the initial refcount update for some cluster
417 * allocation, we might have used the same clusters to store newly
418 * allocated metadata. Make the caller search some new space. */
419 return -EAGAIN;
420
421 fail_table:
422 g_free(new_table);
423 fail_block:
424 if (*refcount_block != NULL) {
425 qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
426 }
427 return ret;
428 }
429
430 void qcow2_process_discards(BlockDriverState *bs, int ret)
431 {
432 BDRVQcowState *s = bs->opaque;
433 Qcow2DiscardRegion *d, *next;
434
435 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
436 QTAILQ_REMOVE(&s->discards, d, next);
437
438 /* Discard is optional, ignore the return value */
439 if (ret >= 0) {
440 bdrv_discard(bs->file,
441 d->offset >> BDRV_SECTOR_BITS,
442 d->bytes >> BDRV_SECTOR_BITS);
443 }
444
445 g_free(d);
446 }
447 }
448
449 static void update_refcount_discard(BlockDriverState *bs,
450 uint64_t offset, uint64_t length)
451 {
452 BDRVQcowState *s = bs->opaque;
453 Qcow2DiscardRegion *d, *p, *next;
454
455 QTAILQ_FOREACH(d, &s->discards, next) {
456 uint64_t new_start = MIN(offset, d->offset);
457 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
458
459 if (new_end - new_start <= length + d->bytes) {
460 /* There can't be any overlap, areas ending up here have no
461 * references any more and therefore shouldn't get freed another
462 * time. */
463 assert(d->bytes + length == new_end - new_start);
464 d->offset = new_start;
465 d->bytes = new_end - new_start;
466 goto found;
467 }
468 }
469
470 d = g_malloc(sizeof(*d));
471 *d = (Qcow2DiscardRegion) {
472 .bs = bs,
473 .offset = offset,
474 .bytes = length,
475 };
476 QTAILQ_INSERT_TAIL(&s->discards, d, next);
477
478 found:
479 /* Merge discard requests if they are adjacent now */
480 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
481 if (p == d
482 || p->offset > d->offset + d->bytes
483 || d->offset > p->offset + p->bytes)
484 {
485 continue;
486 }
487
488 /* Still no overlap possible */
489 assert(p->offset == d->offset + d->bytes
490 || d->offset == p->offset + p->bytes);
491
492 QTAILQ_REMOVE(&s->discards, p, next);
493 d->offset = MIN(d->offset, p->offset);
494 d->bytes += p->bytes;
495 }
496 }
497
498 /* XXX: cache several refcount block clusters ? */
499 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
500 int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
501 {
502 BDRVQcowState *s = bs->opaque;
503 int64_t start, last, cluster_offset;
504 uint16_t *refcount_block = NULL;
505 int64_t old_table_index = -1;
506 int ret;
507
508 #ifdef DEBUG_ALLOC2
509 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
510 offset, length, addend);
511 #endif
512 if (length < 0) {
513 return -EINVAL;
514 } else if (length == 0) {
515 return 0;
516 }
517
518 if (addend < 0) {
519 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
520 s->l2_table_cache);
521 }
522
523 start = start_of_cluster(s, offset);
524 last = start_of_cluster(s, offset + length - 1);
525 for(cluster_offset = start; cluster_offset <= last;
526 cluster_offset += s->cluster_size)
527 {
528 int block_index, refcount;
529 int64_t cluster_index = cluster_offset >> s->cluster_bits;
530 int64_t table_index =
531 cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
532
533 /* Load the refcount block and allocate it if needed */
534 if (table_index != old_table_index) {
535 if (refcount_block) {
536 ret = qcow2_cache_put(bs, s->refcount_block_cache,
537 (void**) &refcount_block);
538 if (ret < 0) {
539 goto fail;
540 }
541 }
542
543 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
544 if (ret < 0) {
545 goto fail;
546 }
547 }
548 old_table_index = table_index;
549
550 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
551
552 /* we can update the count and save it */
553 block_index = cluster_index &
554 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
555
556 refcount = be16_to_cpu(refcount_block[block_index]);
557 refcount += addend;
558 if (refcount < 0 || refcount > 0xffff) {
559 ret = -EINVAL;
560 goto fail;
561 }
562 if (refcount == 0 && cluster_index < s->free_cluster_index) {
563 s->free_cluster_index = cluster_index;
564 }
565 refcount_block[block_index] = cpu_to_be16(refcount);
566
567 if (refcount == 0 && s->discard_passthrough[type]) {
568 update_refcount_discard(bs, cluster_offset, s->cluster_size);
569 }
570 }
571
572 ret = 0;
573 fail:
574 if (!s->cache_discards) {
575 qcow2_process_discards(bs, ret);
576 }
577
578 /* Write last changed block to disk */
579 if (refcount_block) {
580 int wret;
581 wret = qcow2_cache_put(bs, s->refcount_block_cache,
582 (void**) &refcount_block);
583 if (wret < 0) {
584 return ret < 0 ? ret : wret;
585 }
586 }
587
588 /*
589 * Try do undo any updates if an error is returned (This may succeed in
590 * some cases like ENOSPC for allocating a new refcount block)
591 */
592 if (ret < 0) {
593 int dummy;
594 dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
595 QCOW2_DISCARD_NEVER);
596 (void)dummy;
597 }
598
599 return ret;
600 }
601
602 /*
603 * Increases or decreases the refcount of a given cluster by one.
604 * addend must be 1 or -1.
605 *
606 * If the return value is non-negative, it is the new refcount of the cluster.
607 * If it is negative, it is -errno and indicates an error.
608 */
609 int qcow2_update_cluster_refcount(BlockDriverState *bs,
610 int64_t cluster_index,
611 int addend,
612 enum qcow2_discard_type type)
613 {
614 BDRVQcowState *s = bs->opaque;
615 int ret;
616
617 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
618 type);
619 if (ret < 0) {
620 return ret;
621 }
622
623 return get_refcount(bs, cluster_index);
624 }
625
626
627
628 /*********************************************************/
629 /* cluster allocation functions */
630
631
632
633 /* return < 0 if error */
634 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size)
635 {
636 BDRVQcowState *s = bs->opaque;
637 int i, nb_clusters, refcount;
638
639 nb_clusters = size_to_clusters(s, size);
640 retry:
641 for(i = 0; i < nb_clusters; i++) {
642 int64_t next_cluster_index = s->free_cluster_index++;
643 refcount = get_refcount(bs, next_cluster_index);
644
645 if (refcount < 0) {
646 return refcount;
647 } else if (refcount != 0) {
648 goto retry;
649 }
650 }
651 #ifdef DEBUG_ALLOC2
652 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
653 size,
654 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
655 #endif
656 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
657 }
658
659 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size)
660 {
661 int64_t offset;
662 int ret;
663
664 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
665 do {
666 offset = alloc_clusters_noref(bs, size);
667 if (offset < 0) {
668 return offset;
669 }
670
671 ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
672 } while (ret == -EAGAIN);
673
674 if (ret < 0) {
675 return ret;
676 }
677
678 return offset;
679 }
680
681 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
682 int nb_clusters)
683 {
684 BDRVQcowState *s = bs->opaque;
685 uint64_t cluster_index;
686 uint64_t i;
687 int refcount, ret;
688
689 assert(nb_clusters >= 0);
690 if (nb_clusters == 0) {
691 return 0;
692 }
693
694 do {
695 /* Check how many clusters there are free */
696 cluster_index = offset >> s->cluster_bits;
697 for(i = 0; i < nb_clusters; i++) {
698 refcount = get_refcount(bs, cluster_index++);
699
700 if (refcount < 0) {
701 return refcount;
702 } else if (refcount != 0) {
703 break;
704 }
705 }
706
707 /* And then allocate them */
708 ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
709 QCOW2_DISCARD_NEVER);
710 } while (ret == -EAGAIN);
711
712 if (ret < 0) {
713 return ret;
714 }
715
716 return i;
717 }
718
719 /* only used to allocate compressed sectors. We try to allocate
720 contiguous sectors. size must be <= cluster_size */
721 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
722 {
723 BDRVQcowState *s = bs->opaque;
724 int64_t offset, cluster_offset;
725 int free_in_cluster;
726
727 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
728 assert(size > 0 && size <= s->cluster_size);
729 if (s->free_byte_offset == 0) {
730 offset = qcow2_alloc_clusters(bs, s->cluster_size);
731 if (offset < 0) {
732 return offset;
733 }
734 s->free_byte_offset = offset;
735 }
736 redo:
737 free_in_cluster = s->cluster_size -
738 offset_into_cluster(s, s->free_byte_offset);
739 if (size <= free_in_cluster) {
740 /* enough space in current cluster */
741 offset = s->free_byte_offset;
742 s->free_byte_offset += size;
743 free_in_cluster -= size;
744 if (free_in_cluster == 0)
745 s->free_byte_offset = 0;
746 if (offset_into_cluster(s, offset) != 0)
747 qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
748 QCOW2_DISCARD_NEVER);
749 } else {
750 offset = qcow2_alloc_clusters(bs, s->cluster_size);
751 if (offset < 0) {
752 return offset;
753 }
754 cluster_offset = start_of_cluster(s, s->free_byte_offset);
755 if ((cluster_offset + s->cluster_size) == offset) {
756 /* we are lucky: contiguous data */
757 offset = s->free_byte_offset;
758 qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
759 QCOW2_DISCARD_NEVER);
760 s->free_byte_offset += size;
761 } else {
762 s->free_byte_offset = offset;
763 goto redo;
764 }
765 }
766
767 /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
768 * or explicitly by qcow2_update_cluster_refcount(). Refcount blocks must
769 * be flushed before the caller's L2 table updates.
770 */
771 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
772 return offset;
773 }
774
775 void qcow2_free_clusters(BlockDriverState *bs,
776 int64_t offset, int64_t size,
777 enum qcow2_discard_type type)
778 {
779 int ret;
780
781 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
782 ret = update_refcount(bs, offset, size, -1, type);
783 if (ret < 0) {
784 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
785 /* TODO Remember the clusters to free them later and avoid leaking */
786 }
787 }
788
789 /*
790 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
791 * normal cluster, compressed cluster, etc.)
792 */
793 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
794 int nb_clusters, enum qcow2_discard_type type)
795 {
796 BDRVQcowState *s = bs->opaque;
797
798 switch (qcow2_get_cluster_type(l2_entry)) {
799 case QCOW2_CLUSTER_COMPRESSED:
800 {
801 int nb_csectors;
802 nb_csectors = ((l2_entry >> s->csize_shift) &
803 s->csize_mask) + 1;
804 qcow2_free_clusters(bs,
805 (l2_entry & s->cluster_offset_mask) & ~511,
806 nb_csectors * 512, type);
807 }
808 break;
809 case QCOW2_CLUSTER_NORMAL:
810 case QCOW2_CLUSTER_ZERO:
811 if (l2_entry & L2E_OFFSET_MASK) {
812 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
813 nb_clusters << s->cluster_bits, type);
814 }
815 break;
816 case QCOW2_CLUSTER_UNALLOCATED:
817 break;
818 default:
819 abort();
820 }
821 }
822
823
824
825 /*********************************************************/
826 /* snapshots and image creation */
827
828
829
830 /* update the refcounts of snapshots and the copied flag */
831 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
832 int64_t l1_table_offset, int l1_size, int addend)
833 {
834 BDRVQcowState *s = bs->opaque;
835 uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, l1_allocated;
836 int64_t old_offset, old_l2_offset;
837 int i, j, l1_modified = 0, nb_csectors, refcount;
838 int ret;
839
840 l2_table = NULL;
841 l1_table = NULL;
842 l1_size2 = l1_size * sizeof(uint64_t);
843
844 s->cache_discards = true;
845
846 /* WARNING: qcow2_snapshot_goto relies on this function not using the
847 * l1_table_offset when it is the current s->l1_table_offset! Be careful
848 * when changing this! */
849 if (l1_table_offset != s->l1_table_offset) {
850 l1_table = g_malloc0(align_offset(l1_size2, 512));
851 l1_allocated = 1;
852
853 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
854 if (ret < 0) {
855 goto fail;
856 }
857
858 for(i = 0;i < l1_size; i++)
859 be64_to_cpus(&l1_table[i]);
860 } else {
861 assert(l1_size == s->l1_size);
862 l1_table = s->l1_table;
863 l1_allocated = 0;
864 }
865
866 for(i = 0; i < l1_size; i++) {
867 l2_offset = l1_table[i];
868 if (l2_offset) {
869 old_l2_offset = l2_offset;
870 l2_offset &= L1E_OFFSET_MASK;
871
872 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
873 (void**) &l2_table);
874 if (ret < 0) {
875 goto fail;
876 }
877
878 for(j = 0; j < s->l2_size; j++) {
879 uint64_t cluster_index;
880
881 offset = be64_to_cpu(l2_table[j]);
882 old_offset = offset;
883 offset &= ~QCOW_OFLAG_COPIED;
884
885 switch (qcow2_get_cluster_type(offset)) {
886 case QCOW2_CLUSTER_COMPRESSED:
887 nb_csectors = ((offset >> s->csize_shift) &
888 s->csize_mask) + 1;
889 if (addend != 0) {
890 ret = update_refcount(bs,
891 (offset & s->cluster_offset_mask) & ~511,
892 nb_csectors * 512, addend,
893 QCOW2_DISCARD_SNAPSHOT);
894 if (ret < 0) {
895 goto fail;
896 }
897 }
898 /* compressed clusters are never modified */
899 refcount = 2;
900 break;
901
902 case QCOW2_CLUSTER_NORMAL:
903 case QCOW2_CLUSTER_ZERO:
904 cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
905 if (!cluster_index) {
906 /* unallocated */
907 refcount = 0;
908 break;
909 }
910 if (addend != 0) {
911 refcount = qcow2_update_cluster_refcount(bs,
912 cluster_index, addend,
913 QCOW2_DISCARD_SNAPSHOT);
914 } else {
915 refcount = get_refcount(bs, cluster_index);
916 }
917
918 if (refcount < 0) {
919 ret = refcount;
920 goto fail;
921 }
922 break;
923
924 case QCOW2_CLUSTER_UNALLOCATED:
925 refcount = 0;
926 break;
927
928 default:
929 abort();
930 }
931
932 if (refcount == 1) {
933 offset |= QCOW_OFLAG_COPIED;
934 }
935 if (offset != old_offset) {
936 if (addend > 0) {
937 qcow2_cache_set_dependency(bs, s->l2_table_cache,
938 s->refcount_block_cache);
939 }
940 l2_table[j] = cpu_to_be64(offset);
941 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
942 }
943 }
944
945 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
946 if (ret < 0) {
947 goto fail;
948 }
949
950
951 if (addend != 0) {
952 refcount = qcow2_update_cluster_refcount(bs, l2_offset >>
953 s->cluster_bits, addend, QCOW2_DISCARD_SNAPSHOT);
954 } else {
955 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
956 }
957 if (refcount < 0) {
958 ret = refcount;
959 goto fail;
960 } else if (refcount == 1) {
961 l2_offset |= QCOW_OFLAG_COPIED;
962 }
963 if (l2_offset != old_l2_offset) {
964 l1_table[i] = l2_offset;
965 l1_modified = 1;
966 }
967 }
968 }
969
970 ret = bdrv_flush(bs);
971 fail:
972 if (l2_table) {
973 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
974 }
975
976 s->cache_discards = false;
977 qcow2_process_discards(bs, ret);
978
979 /* Update L1 only if it isn't deleted anyway (addend = -1) */
980 if (ret == 0 && addend >= 0 && l1_modified) {
981 for (i = 0; i < l1_size; i++) {
982 cpu_to_be64s(&l1_table[i]);
983 }
984
985 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
986
987 for (i = 0; i < l1_size; i++) {
988 be64_to_cpus(&l1_table[i]);
989 }
990 }
991 if (l1_allocated)
992 g_free(l1_table);
993 return ret;
994 }
995
996
997
998
999 /*********************************************************/
1000 /* refcount checking functions */
1001
1002
1003
1004 /*
1005 * Increases the refcount for a range of clusters in a given refcount table.
1006 * This is used to construct a temporary refcount table out of L1 and L2 tables
1007 * which can be compared the the refcount table saved in the image.
1008 *
1009 * Modifies the number of errors in res.
1010 */
1011 static void inc_refcounts(BlockDriverState *bs,
1012 BdrvCheckResult *res,
1013 uint16_t *refcount_table,
1014 int refcount_table_size,
1015 int64_t offset, int64_t size)
1016 {
1017 BDRVQcowState *s = bs->opaque;
1018 int64_t start, last, cluster_offset;
1019 int k;
1020
1021 if (size <= 0)
1022 return;
1023
1024 start = start_of_cluster(s, offset);
1025 last = start_of_cluster(s, offset + size - 1);
1026 for(cluster_offset = start; cluster_offset <= last;
1027 cluster_offset += s->cluster_size) {
1028 k = cluster_offset >> s->cluster_bits;
1029 if (k < 0) {
1030 fprintf(stderr, "ERROR: invalid cluster offset=0x%" PRIx64 "\n",
1031 cluster_offset);
1032 res->corruptions++;
1033 } else if (k >= refcount_table_size) {
1034 fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
1035 "the end of the image file, can't properly check refcounts.\n",
1036 cluster_offset);
1037 res->check_errors++;
1038 } else {
1039 if (++refcount_table[k] == 0) {
1040 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1041 "\n", cluster_offset);
1042 res->corruptions++;
1043 }
1044 }
1045 }
1046 }
1047
1048 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1049 enum {
1050 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1051 };
1052
1053 /*
1054 * Increases the refcount in the given refcount table for the all clusters
1055 * referenced in the L2 table. While doing so, performs some checks on L2
1056 * entries.
1057 *
1058 * Returns the number of errors found by the checks or -errno if an internal
1059 * error occurred.
1060 */
1061 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1062 uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
1063 int flags)
1064 {
1065 BDRVQcowState *s = bs->opaque;
1066 uint64_t *l2_table, l2_entry;
1067 uint64_t next_contiguous_offset = 0;
1068 int i, l2_size, nb_csectors;
1069
1070 /* Read L2 table from disk */
1071 l2_size = s->l2_size * sizeof(uint64_t);
1072 l2_table = g_malloc(l2_size);
1073
1074 if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
1075 goto fail;
1076
1077 /* Do the actual checks */
1078 for(i = 0; i < s->l2_size; i++) {
1079 l2_entry = be64_to_cpu(l2_table[i]);
1080
1081 switch (qcow2_get_cluster_type(l2_entry)) {
1082 case QCOW2_CLUSTER_COMPRESSED:
1083 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1084 if (l2_entry & QCOW_OFLAG_COPIED) {
1085 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1086 "copied flag must never be set for compressed "
1087 "clusters\n", l2_entry >> s->cluster_bits);
1088 l2_entry &= ~QCOW_OFLAG_COPIED;
1089 res->corruptions++;
1090 }
1091
1092 /* Mark cluster as used */
1093 nb_csectors = ((l2_entry >> s->csize_shift) &
1094 s->csize_mask) + 1;
1095 l2_entry &= s->cluster_offset_mask;
1096 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1097 l2_entry & ~511, nb_csectors * 512);
1098
1099 if (flags & CHECK_FRAG_INFO) {
1100 res->bfi.allocated_clusters++;
1101 res->bfi.compressed_clusters++;
1102
1103 /* Compressed clusters are fragmented by nature. Since they
1104 * take up sub-sector space but we only have sector granularity
1105 * I/O we need to re-read the same sectors even for adjacent
1106 * compressed clusters.
1107 */
1108 res->bfi.fragmented_clusters++;
1109 }
1110 break;
1111
1112 case QCOW2_CLUSTER_ZERO:
1113 if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1114 break;
1115 }
1116 /* fall through */
1117
1118 case QCOW2_CLUSTER_NORMAL:
1119 {
1120 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1121
1122 if (flags & CHECK_FRAG_INFO) {
1123 res->bfi.allocated_clusters++;
1124 if (next_contiguous_offset &&
1125 offset != next_contiguous_offset) {
1126 res->bfi.fragmented_clusters++;
1127 }
1128 next_contiguous_offset = offset + s->cluster_size;
1129 }
1130
1131 /* Mark cluster as used */
1132 inc_refcounts(bs, res, refcount_table,refcount_table_size,
1133 offset, s->cluster_size);
1134
1135 /* Correct offsets are cluster aligned */
1136 if (offset_into_cluster(s, offset)) {
1137 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1138 "properly aligned; L2 entry corrupted.\n", offset);
1139 res->corruptions++;
1140 }
1141 break;
1142 }
1143
1144 case QCOW2_CLUSTER_UNALLOCATED:
1145 break;
1146
1147 default:
1148 abort();
1149 }
1150 }
1151
1152 g_free(l2_table);
1153 return 0;
1154
1155 fail:
1156 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1157 g_free(l2_table);
1158 return -EIO;
1159 }
1160
1161 /*
1162 * Increases the refcount for the L1 table, its L2 tables and all referenced
1163 * clusters in the given refcount table. While doing so, performs some checks
1164 * on L1 and L2 entries.
1165 *
1166 * Returns the number of errors found by the checks or -errno if an internal
1167 * error occurred.
1168 */
1169 static int check_refcounts_l1(BlockDriverState *bs,
1170 BdrvCheckResult *res,
1171 uint16_t *refcount_table,
1172 int refcount_table_size,
1173 int64_t l1_table_offset, int l1_size,
1174 int flags)
1175 {
1176 BDRVQcowState *s = bs->opaque;
1177 uint64_t *l1_table, l2_offset, l1_size2;
1178 int i, ret;
1179
1180 l1_size2 = l1_size * sizeof(uint64_t);
1181
1182 /* Mark L1 table as used */
1183 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1184 l1_table_offset, l1_size2);
1185
1186 /* Read L1 table entries from disk */
1187 if (l1_size2 == 0) {
1188 l1_table = NULL;
1189 } else {
1190 l1_table = g_malloc(l1_size2);
1191 if (bdrv_pread(bs->file, l1_table_offset,
1192 l1_table, l1_size2) != l1_size2)
1193 goto fail;
1194 for(i = 0;i < l1_size; i++)
1195 be64_to_cpus(&l1_table[i]);
1196 }
1197
1198 /* Do the actual checks */
1199 for(i = 0; i < l1_size; i++) {
1200 l2_offset = l1_table[i];
1201 if (l2_offset) {
1202 /* Mark L2 table as used */
1203 l2_offset &= L1E_OFFSET_MASK;
1204 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1205 l2_offset, s->cluster_size);
1206
1207 /* L2 tables are cluster aligned */
1208 if (offset_into_cluster(s, l2_offset)) {
1209 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1210 "cluster aligned; L1 entry corrupted\n", l2_offset);
1211 res->corruptions++;
1212 }
1213
1214 /* Process and check L2 entries */
1215 ret = check_refcounts_l2(bs, res, refcount_table,
1216 refcount_table_size, l2_offset, flags);
1217 if (ret < 0) {
1218 goto fail;
1219 }
1220 }
1221 }
1222 g_free(l1_table);
1223 return 0;
1224
1225 fail:
1226 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1227 res->check_errors++;
1228 g_free(l1_table);
1229 return -EIO;
1230 }
1231
1232 /*
1233 * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1234 *
1235 * This function does not print an error message nor does it increment
1236 * check_errors if get_refcount fails (this is because such an error will have
1237 * been already detected and sufficiently signaled by the calling function
1238 * (qcow2_check_refcounts) by the time this function is called).
1239 */
1240 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1241 BdrvCheckMode fix)
1242 {
1243 BDRVQcowState *s = bs->opaque;
1244 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1245 int ret;
1246 int refcount;
1247 int i, j;
1248
1249 for (i = 0; i < s->l1_size; i++) {
1250 uint64_t l1_entry = s->l1_table[i];
1251 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1252 bool l2_dirty = false;
1253
1254 if (!l2_offset) {
1255 continue;
1256 }
1257
1258 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
1259 if (refcount < 0) {
1260 /* don't print message nor increment check_errors */
1261 continue;
1262 }
1263 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1264 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1265 "l1_entry=%" PRIx64 " refcount=%d\n",
1266 fix & BDRV_FIX_ERRORS ? "Repairing" :
1267 "ERROR",
1268 i, l1_entry, refcount);
1269 if (fix & BDRV_FIX_ERRORS) {
1270 s->l1_table[i] = refcount == 1
1271 ? l1_entry | QCOW_OFLAG_COPIED
1272 : l1_entry & ~QCOW_OFLAG_COPIED;
1273 ret = qcow2_write_l1_entry(bs, i);
1274 if (ret < 0) {
1275 res->check_errors++;
1276 goto fail;
1277 }
1278 res->corruptions_fixed++;
1279 } else {
1280 res->corruptions++;
1281 }
1282 }
1283
1284 ret = bdrv_pread(bs->file, l2_offset, l2_table,
1285 s->l2_size * sizeof(uint64_t));
1286 if (ret < 0) {
1287 fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1288 strerror(-ret));
1289 res->check_errors++;
1290 goto fail;
1291 }
1292
1293 for (j = 0; j < s->l2_size; j++) {
1294 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1295 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1296 int cluster_type = qcow2_get_cluster_type(l2_entry);
1297
1298 if ((cluster_type == QCOW2_CLUSTER_NORMAL) ||
1299 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) {
1300 refcount = get_refcount(bs, data_offset >> s->cluster_bits);
1301 if (refcount < 0) {
1302 /* don't print message nor increment check_errors */
1303 continue;
1304 }
1305 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1306 fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1307 "l2_entry=%" PRIx64 " refcount=%d\n",
1308 fix & BDRV_FIX_ERRORS ? "Repairing" :
1309 "ERROR",
1310 l2_entry, refcount);
1311 if (fix & BDRV_FIX_ERRORS) {
1312 l2_table[j] = cpu_to_be64(refcount == 1
1313 ? l2_entry | QCOW_OFLAG_COPIED
1314 : l2_entry & ~QCOW_OFLAG_COPIED);
1315 l2_dirty = true;
1316 res->corruptions_fixed++;
1317 } else {
1318 res->corruptions++;
1319 }
1320 }
1321 }
1322 }
1323
1324 if (l2_dirty) {
1325 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1326 l2_offset, s->cluster_size);
1327 if (ret < 0) {
1328 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1329 "overlap check failed: %s\n", strerror(-ret));
1330 res->check_errors++;
1331 goto fail;
1332 }
1333
1334 ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size);
1335 if (ret < 0) {
1336 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1337 strerror(-ret));
1338 res->check_errors++;
1339 goto fail;
1340 }
1341 }
1342 }
1343
1344 ret = 0;
1345
1346 fail:
1347 qemu_vfree(l2_table);
1348 return ret;
1349 }
1350
1351 /*
1352 * Writes one sector of the refcount table to the disk
1353 */
1354 #define RT_ENTRIES_PER_SECTOR (512 / sizeof(uint64_t))
1355 static int write_reftable_entry(BlockDriverState *bs, int rt_index)
1356 {
1357 BDRVQcowState *s = bs->opaque;
1358 uint64_t buf[RT_ENTRIES_PER_SECTOR];
1359 int rt_start_index;
1360 int i, ret;
1361
1362 rt_start_index = rt_index & ~(RT_ENTRIES_PER_SECTOR - 1);
1363 for (i = 0; i < RT_ENTRIES_PER_SECTOR; i++) {
1364 buf[i] = cpu_to_be64(s->refcount_table[rt_start_index + i]);
1365 }
1366
1367 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_REFCOUNT_TABLE,
1368 s->refcount_table_offset + rt_start_index * sizeof(uint64_t),
1369 sizeof(buf));
1370 if (ret < 0) {
1371 return ret;
1372 }
1373
1374 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_UPDATE);
1375 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset +
1376 rt_start_index * sizeof(uint64_t), buf, sizeof(buf));
1377 if (ret < 0) {
1378 return ret;
1379 }
1380
1381 return 0;
1382 }
1383
1384 /*
1385 * Allocates a new cluster for the given refcount block (represented by its
1386 * offset in the image file) and copies the current content there. This function
1387 * does _not_ decrement the reference count for the currently occupied cluster.
1388 *
1389 * This function prints an informative message to stderr on error (and returns
1390 * -errno); on success, the offset of the newly allocated cluster is returned.
1391 */
1392 static int64_t realloc_refcount_block(BlockDriverState *bs, int reftable_index,
1393 uint64_t offset)
1394 {
1395 BDRVQcowState *s = bs->opaque;
1396 int64_t new_offset = 0;
1397 void *refcount_block = NULL;
1398 int ret;
1399
1400 /* allocate new refcount block */
1401 new_offset = qcow2_alloc_clusters(bs, s->cluster_size);
1402 if (new_offset < 0) {
1403 fprintf(stderr, "Could not allocate new cluster: %s\n",
1404 strerror(-new_offset));
1405 ret = new_offset;
1406 goto done;
1407 }
1408
1409 /* fetch current refcount block content */
1410 ret = qcow2_cache_get(bs, s->refcount_block_cache, offset, &refcount_block);
1411 if (ret < 0) {
1412 fprintf(stderr, "Could not fetch refcount block: %s\n", strerror(-ret));
1413 goto fail_free_cluster;
1414 }
1415
1416 /* new block has not yet been entered into refcount table, therefore it is
1417 * no refcount block yet (regarding this check) */
1418 ret = qcow2_pre_write_overlap_check(bs, 0, new_offset, s->cluster_size);
1419 if (ret < 0) {
1420 fprintf(stderr, "Could not write refcount block; metadata overlap "
1421 "check failed: %s\n", strerror(-ret));
1422 /* the image will be marked corrupt, so don't even attempt on freeing
1423 * the cluster */
1424 goto done;
1425 }
1426
1427 /* write to new block */
1428 ret = bdrv_write(bs->file, new_offset / BDRV_SECTOR_SIZE, refcount_block,
1429 s->cluster_sectors);
1430 if (ret < 0) {
1431 fprintf(stderr, "Could not write refcount block: %s\n", strerror(-ret));
1432 goto fail_free_cluster;
1433 }
1434
1435 /* update refcount table */
1436 assert(!offset_into_cluster(s, new_offset));
1437 s->refcount_table[reftable_index] = new_offset;
1438 ret = write_reftable_entry(bs, reftable_index);
1439 if (ret < 0) {
1440 fprintf(stderr, "Could not update refcount table: %s\n",
1441 strerror(-ret));
1442 goto fail_free_cluster;
1443 }
1444
1445 goto done;
1446
1447 fail_free_cluster:
1448 qcow2_free_clusters(bs, new_offset, s->cluster_size, QCOW2_DISCARD_OTHER);
1449
1450 done:
1451 if (refcount_block) {
1452 /* This should never fail, as it would only do so if the given refcount
1453 * block cannot be found in the cache. As this is impossible as long as
1454 * there are no bugs, assert the success. */
1455 int tmp = qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
1456 assert(tmp == 0);
1457 }
1458
1459 if (ret < 0) {
1460 return ret;
1461 }
1462
1463 return new_offset;
1464 }
1465
1466 /*
1467 * Checks an image for refcount consistency.
1468 *
1469 * Returns 0 if no errors are found, the number of errors in case the image is
1470 * detected as corrupted, and -errno when an internal error occurred.
1471 */
1472 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1473 BdrvCheckMode fix)
1474 {
1475 BDRVQcowState *s = bs->opaque;
1476 int64_t size, i, highest_cluster;
1477 int nb_clusters, refcount1, refcount2;
1478 QCowSnapshot *sn;
1479 uint16_t *refcount_table;
1480 int ret;
1481
1482 size = bdrv_getlength(bs->file);
1483 nb_clusters = size_to_clusters(s, size);
1484 refcount_table = g_malloc0(nb_clusters * sizeof(uint16_t));
1485
1486 res->bfi.total_clusters =
1487 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1488
1489 /* header */
1490 inc_refcounts(bs, res, refcount_table, nb_clusters,
1491 0, s->cluster_size);
1492
1493 /* current L1 table */
1494 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1495 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO);
1496 if (ret < 0) {
1497 goto fail;
1498 }
1499
1500 /* snapshots */
1501 for(i = 0; i < s->nb_snapshots; i++) {
1502 sn = s->snapshots + i;
1503 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1504 sn->l1_table_offset, sn->l1_size, 0);
1505 if (ret < 0) {
1506 goto fail;
1507 }
1508 }
1509 inc_refcounts(bs, res, refcount_table, nb_clusters,
1510 s->snapshots_offset, s->snapshots_size);
1511
1512 /* refcount data */
1513 inc_refcounts(bs, res, refcount_table, nb_clusters,
1514 s->refcount_table_offset,
1515 s->refcount_table_size * sizeof(uint64_t));
1516
1517 for(i = 0; i < s->refcount_table_size; i++) {
1518 uint64_t offset, cluster;
1519 offset = s->refcount_table[i];
1520 cluster = offset >> s->cluster_bits;
1521
1522 /* Refcount blocks are cluster aligned */
1523 if (offset_into_cluster(s, offset)) {
1524 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1525 "cluster aligned; refcount table entry corrupted\n", i);
1526 res->corruptions++;
1527 continue;
1528 }
1529
1530 if (cluster >= nb_clusters) {
1531 fprintf(stderr, "ERROR refcount block %" PRId64
1532 " is outside image\n", i);
1533 res->corruptions++;
1534 continue;
1535 }
1536
1537 if (offset != 0) {
1538 inc_refcounts(bs, res, refcount_table, nb_clusters,
1539 offset, s->cluster_size);
1540 if (refcount_table[cluster] != 1) {
1541 fprintf(stderr, "%s refcount block %" PRId64
1542 " refcount=%d\n",
1543 fix & BDRV_FIX_ERRORS ? "Repairing" :
1544 "ERROR",
1545 i, refcount_table[cluster]);
1546
1547 if (fix & BDRV_FIX_ERRORS) {
1548 int64_t new_offset;
1549
1550 new_offset = realloc_refcount_block(bs, i, offset);
1551 if (new_offset < 0) {
1552 res->corruptions++;
1553 continue;
1554 }
1555
1556 /* update refcounts */
1557 if ((new_offset >> s->cluster_bits) >= nb_clusters) {
1558 /* increase refcount_table size if necessary */
1559 int old_nb_clusters = nb_clusters;
1560 nb_clusters = (new_offset >> s->cluster_bits) + 1;
1561 refcount_table = g_realloc(refcount_table,
1562 nb_clusters * sizeof(uint16_t));
1563 memset(&refcount_table[old_nb_clusters], 0, (nb_clusters
1564 - old_nb_clusters) * sizeof(uint16_t));
1565 }
1566 refcount_table[cluster]--;
1567 inc_refcounts(bs, res, refcount_table, nb_clusters,
1568 new_offset, s->cluster_size);
1569
1570 res->corruptions_fixed++;
1571 } else {
1572 res->corruptions++;
1573 }
1574 }
1575 }
1576 }
1577
1578 /* compare ref counts */
1579 for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1580 refcount1 = get_refcount(bs, i);
1581 if (refcount1 < 0) {
1582 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1583 i, strerror(-refcount1));
1584 res->check_errors++;
1585 continue;
1586 }
1587
1588 refcount2 = refcount_table[i];
1589
1590 if (refcount1 > 0 || refcount2 > 0) {
1591 highest_cluster = i;
1592 }
1593
1594 if (refcount1 != refcount2) {
1595
1596 /* Check if we're allowed to fix the mismatch */
1597 int *num_fixed = NULL;
1598 if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1599 num_fixed = &res->leaks_fixed;
1600 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1601 num_fixed = &res->corruptions_fixed;
1602 }
1603
1604 fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1605 num_fixed != NULL ? "Repairing" :
1606 refcount1 < refcount2 ? "ERROR" :
1607 "Leaked",
1608 i, refcount1, refcount2);
1609
1610 if (num_fixed) {
1611 ret = update_refcount(bs, i << s->cluster_bits, 1,
1612 refcount2 - refcount1,
1613 QCOW2_DISCARD_ALWAYS);
1614 if (ret >= 0) {
1615 (*num_fixed)++;
1616 continue;
1617 }
1618 }
1619
1620 /* And if we couldn't, print an error */
1621 if (refcount1 < refcount2) {
1622 res->corruptions++;
1623 } else {
1624 res->leaks++;
1625 }
1626 }
1627 }
1628
1629 /* check OFLAG_COPIED */
1630 ret = check_oflag_copied(bs, res, fix);
1631 if (ret < 0) {
1632 goto fail;
1633 }
1634
1635 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1636 ret = 0;
1637
1638 fail:
1639 g_free(refcount_table);
1640
1641 return ret;
1642 }
1643
1644 #define overlaps_with(ofs, sz) \
1645 ranges_overlap(offset, size, ofs, sz)
1646
1647 /*
1648 * Checks if the given offset into the image file is actually free to use by
1649 * looking for overlaps with important metadata sections (L1/L2 tables etc.),
1650 * i.e. a sanity check without relying on the refcount tables.
1651 *
1652 * The ign parameter specifies what checks not to perform (being a bitmask of
1653 * QCow2MetadataOverlap values), i.e., what sections to ignore.
1654 *
1655 * Returns:
1656 * - 0 if writing to this offset will not affect the mentioned metadata
1657 * - a positive QCow2MetadataOverlap value indicating one overlapping section
1658 * - a negative value (-errno) indicating an error while performing a check,
1659 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
1660 */
1661 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
1662 int64_t size)
1663 {
1664 BDRVQcowState *s = bs->opaque;
1665 int chk = s->overlap_check & ~ign;
1666 int i, j;
1667
1668 if (!size) {
1669 return 0;
1670 }
1671
1672 if (chk & QCOW2_OL_MAIN_HEADER) {
1673 if (offset < s->cluster_size) {
1674 return QCOW2_OL_MAIN_HEADER;
1675 }
1676 }
1677
1678 /* align range to test to cluster boundaries */
1679 size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size);
1680 offset = start_of_cluster(s, offset);
1681
1682 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
1683 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
1684 return QCOW2_OL_ACTIVE_L1;
1685 }
1686 }
1687
1688 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
1689 if (overlaps_with(s->refcount_table_offset,
1690 s->refcount_table_size * sizeof(uint64_t))) {
1691 return QCOW2_OL_REFCOUNT_TABLE;
1692 }
1693 }
1694
1695 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
1696 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
1697 return QCOW2_OL_SNAPSHOT_TABLE;
1698 }
1699 }
1700
1701 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
1702 for (i = 0; i < s->nb_snapshots; i++) {
1703 if (s->snapshots[i].l1_size &&
1704 overlaps_with(s->snapshots[i].l1_table_offset,
1705 s->snapshots[i].l1_size * sizeof(uint64_t))) {
1706 return QCOW2_OL_INACTIVE_L1;
1707 }
1708 }
1709 }
1710
1711 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
1712 for (i = 0; i < s->l1_size; i++) {
1713 if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
1714 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
1715 s->cluster_size)) {
1716 return QCOW2_OL_ACTIVE_L2;
1717 }
1718 }
1719 }
1720
1721 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
1722 for (i = 0; i < s->refcount_table_size; i++) {
1723 if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
1724 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
1725 s->cluster_size)) {
1726 return QCOW2_OL_REFCOUNT_BLOCK;
1727 }
1728 }
1729 }
1730
1731 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
1732 for (i = 0; i < s->nb_snapshots; i++) {
1733 uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
1734 uint32_t l1_sz = s->snapshots[i].l1_size;
1735 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
1736 uint64_t *l1 = g_malloc(l1_sz2);
1737 int ret;
1738
1739 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
1740 if (ret < 0) {
1741 g_free(l1);
1742 return ret;
1743 }
1744
1745 for (j = 0; j < l1_sz; j++) {
1746 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
1747 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
1748 g_free(l1);
1749 return QCOW2_OL_INACTIVE_L2;
1750 }
1751 }
1752
1753 g_free(l1);
1754 }
1755 }
1756
1757 return 0;
1758 }
1759
1760 static const char *metadata_ol_names[] = {
1761 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header",
1762 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table",
1763 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table",
1764 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
1765 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
1766 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
1767 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table",
1768 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table",
1769 };
1770
1771 /*
1772 * First performs a check for metadata overlaps (through
1773 * qcow2_check_metadata_overlap); if that fails with a negative value (error
1774 * while performing a check), that value is returned. If an impending overlap
1775 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
1776 * and -EIO returned.
1777 *
1778 * Returns 0 if there were neither overlaps nor errors while checking for
1779 * overlaps; or a negative value (-errno) on error.
1780 */
1781 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
1782 int64_t size)
1783 {
1784 int ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
1785
1786 if (ret < 0) {
1787 return ret;
1788 } else if (ret > 0) {
1789 int metadata_ol_bitnr = ffs(ret) - 1;
1790 char *message;
1791 QObject *data;
1792
1793 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
1794
1795 fprintf(stderr, "qcow2: Preventing invalid write on metadata (overlaps "
1796 "with %s); image marked as corrupt.\n",
1797 metadata_ol_names[metadata_ol_bitnr]);
1798 message = g_strdup_printf("Prevented %s overwrite",
1799 metadata_ol_names[metadata_ol_bitnr]);
1800 data = qobject_from_jsonf("{ 'device': %s, 'msg': %s, 'offset': %"
1801 PRId64 ", 'size': %" PRId64 " }", bs->device_name, message,
1802 offset, size);
1803 monitor_protocol_event(QEVENT_BLOCK_IMAGE_CORRUPTED, data);
1804 g_free(message);
1805 qobject_decref(data);
1806
1807 qcow2_mark_corrupt(bs);
1808 bs->drv = NULL; /* make BDS unusable */
1809 return -EIO;
1810 }
1811
1812 return 0;
1813 }