hw/arm/boot: load DTB as a ROM image
[qemu/ar7.git] / block / qcow2-refcount.c
blob43665b86e73ce029b64f04273941fa7e46d59c2f
1 /*
2 * Block driver for the QCOW version 2 format
4 * Copyright (c) 2004-2006 Fabrice Bellard
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:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
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.
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 #include "qapi-event.h"
32 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size);
33 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
34 int64_t offset, int64_t length,
35 int addend, enum qcow2_discard_type type);
38 /*********************************************************/
39 /* refcount handling */
41 int qcow2_refcount_init(BlockDriverState *bs)
43 BDRVQcowState *s = bs->opaque;
44 unsigned int refcount_table_size2, i;
45 int ret;
47 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
48 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
49 s->refcount_table = g_try_malloc(refcount_table_size2);
51 if (s->refcount_table_size > 0) {
52 if (s->refcount_table == NULL) {
53 ret = -ENOMEM;
54 goto fail;
56 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
57 ret = bdrv_pread(bs->file, s->refcount_table_offset,
58 s->refcount_table, refcount_table_size2);
59 if (ret < 0) {
60 goto fail;
62 for(i = 0; i < s->refcount_table_size; i++)
63 be64_to_cpus(&s->refcount_table[i]);
65 return 0;
66 fail:
67 return ret;
70 void qcow2_refcount_close(BlockDriverState *bs)
72 BDRVQcowState *s = bs->opaque;
73 g_free(s->refcount_table);
77 static int load_refcount_block(BlockDriverState *bs,
78 int64_t refcount_block_offset,
79 void **refcount_block)
81 BDRVQcowState *s = bs->opaque;
82 int ret;
84 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
85 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
86 refcount_block);
88 return ret;
92 * Returns the refcount of the cluster given by its index. Any non-negative
93 * return value is the refcount of the cluster, negative values are -errno
94 * and indicate an error.
96 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
98 BDRVQcowState *s = bs->opaque;
99 uint64_t refcount_table_index, block_index;
100 int64_t refcount_block_offset;
101 int ret;
102 uint16_t *refcount_block;
103 uint16_t refcount;
105 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
106 if (refcount_table_index >= s->refcount_table_size)
107 return 0;
108 refcount_block_offset =
109 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
110 if (!refcount_block_offset)
111 return 0;
113 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
114 (void**) &refcount_block);
115 if (ret < 0) {
116 return ret;
119 block_index = cluster_index &
120 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
121 refcount = be16_to_cpu(refcount_block[block_index]);
123 ret = qcow2_cache_put(bs, s->refcount_block_cache,
124 (void**) &refcount_block);
125 if (ret < 0) {
126 return ret;
129 return refcount;
133 * Rounds the refcount table size up to avoid growing the table for each single
134 * refcount block that is allocated.
136 static unsigned int next_refcount_table_size(BDRVQcowState *s,
137 unsigned int min_size)
139 unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
140 unsigned int refcount_table_clusters =
141 MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
143 while (min_clusters > refcount_table_clusters) {
144 refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
147 return refcount_table_clusters << (s->cluster_bits - 3);
151 /* Checks if two offsets are described by the same refcount block */
152 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
153 uint64_t offset_b)
155 uint64_t block_a = offset_a >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
156 uint64_t block_b = offset_b >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
158 return (block_a == block_b);
162 * Loads a refcount block. If it doesn't exist yet, it is allocated first
163 * (including growing the refcount table if needed).
165 * Returns 0 on success or -errno in error case
167 static int alloc_refcount_block(BlockDriverState *bs,
168 int64_t cluster_index, uint16_t **refcount_block)
170 BDRVQcowState *s = bs->opaque;
171 unsigned int refcount_table_index;
172 int ret;
174 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
176 /* Find the refcount block for the given cluster */
177 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
179 if (refcount_table_index < s->refcount_table_size) {
181 uint64_t refcount_block_offset =
182 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
184 /* If it's already there, we're done */
185 if (refcount_block_offset) {
186 return load_refcount_block(bs, refcount_block_offset,
187 (void**) refcount_block);
192 * If we came here, we need to allocate something. Something is at least
193 * a cluster for the new refcount block. It may also include a new refcount
194 * table if the old refcount table is too small.
196 * Note that allocating clusters here needs some special care:
198 * - We can't use the normal qcow2_alloc_clusters(), it would try to
199 * increase the refcount and very likely we would end up with an endless
200 * recursion. Instead we must place the refcount blocks in a way that
201 * they can describe them themselves.
203 * - We need to consider that at this point we are inside update_refcounts
204 * and potentially doing an initial refcount increase. This means that
205 * some clusters have already been allocated by the caller, but their
206 * refcount isn't accurate yet. If we allocate clusters for metadata, we
207 * need to return -EAGAIN to signal the caller that it needs to restart
208 * the search for free clusters.
210 * - alloc_clusters_noref and qcow2_free_clusters may load a different
211 * refcount block into the cache
214 *refcount_block = NULL;
216 /* We write to the refcount table, so we might depend on L2 tables */
217 ret = qcow2_cache_flush(bs, s->l2_table_cache);
218 if (ret < 0) {
219 return ret;
222 /* Allocate the refcount block itself and mark it as used */
223 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
224 if (new_block < 0) {
225 return new_block;
228 #ifdef DEBUG_ALLOC2
229 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
230 " at %" PRIx64 "\n",
231 refcount_table_index, cluster_index << s->cluster_bits, new_block);
232 #endif
234 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
235 /* Zero the new refcount block before updating it */
236 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
237 (void**) refcount_block);
238 if (ret < 0) {
239 goto fail_block;
242 memset(*refcount_block, 0, s->cluster_size);
244 /* The block describes itself, need to update the cache */
245 int block_index = (new_block >> s->cluster_bits) &
246 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
247 (*refcount_block)[block_index] = cpu_to_be16(1);
248 } else {
249 /* Described somewhere else. This can recurse at most twice before we
250 * arrive at a block that describes itself. */
251 ret = update_refcount(bs, new_block, s->cluster_size, 1,
252 QCOW2_DISCARD_NEVER);
253 if (ret < 0) {
254 goto fail_block;
257 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
258 if (ret < 0) {
259 goto fail_block;
262 /* Initialize the new refcount block only after updating its refcount,
263 * update_refcount uses the refcount cache itself */
264 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
265 (void**) refcount_block);
266 if (ret < 0) {
267 goto fail_block;
270 memset(*refcount_block, 0, s->cluster_size);
273 /* Now the new refcount block needs to be written to disk */
274 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
275 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
276 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
277 if (ret < 0) {
278 goto fail_block;
281 /* If the refcount table is big enough, just hook the block up there */
282 if (refcount_table_index < s->refcount_table_size) {
283 uint64_t data64 = cpu_to_be64(new_block);
284 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
285 ret = bdrv_pwrite_sync(bs->file,
286 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
287 &data64, sizeof(data64));
288 if (ret < 0) {
289 goto fail_block;
292 s->refcount_table[refcount_table_index] = new_block;
294 /* The new refcount block may be where the caller intended to put its
295 * data, so let it restart the search. */
296 return -EAGAIN;
299 ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
300 if (ret < 0) {
301 goto fail_block;
305 * If we come here, we need to grow the refcount table. Again, a new
306 * refcount table needs some space and we can't simply allocate to avoid
307 * endless recursion.
309 * Therefore let's grab new refcount blocks at the end of the image, which
310 * will describe themselves and the new refcount table. This way we can
311 * reference them only in the new table and do the switch to the new
312 * refcount table at once without producing an inconsistent state in
313 * between.
315 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
317 /* Calculate the number of refcount blocks needed so far */
318 uint64_t refcount_block_clusters = 1 << (s->cluster_bits - REFCOUNT_SHIFT);
319 uint64_t blocks_used = DIV_ROUND_UP(cluster_index, refcount_block_clusters);
321 if (blocks_used > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
322 return -EFBIG;
325 /* And now we need at least one block more for the new metadata */
326 uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
327 uint64_t last_table_size;
328 uint64_t blocks_clusters;
329 do {
330 uint64_t table_clusters =
331 size_to_clusters(s, table_size * sizeof(uint64_t));
332 blocks_clusters = 1 +
333 ((table_clusters + refcount_block_clusters - 1)
334 / refcount_block_clusters);
335 uint64_t meta_clusters = table_clusters + blocks_clusters;
337 last_table_size = table_size;
338 table_size = next_refcount_table_size(s, blocks_used +
339 ((meta_clusters + refcount_block_clusters - 1)
340 / refcount_block_clusters));
342 } while (last_table_size != table_size);
344 #ifdef DEBUG_ALLOC2
345 fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
346 s->refcount_table_size, table_size);
347 #endif
349 /* Create the new refcount table and blocks */
350 uint64_t meta_offset = (blocks_used * refcount_block_clusters) *
351 s->cluster_size;
352 uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
353 uint64_t *new_table = g_try_new0(uint64_t, table_size);
354 uint16_t *new_blocks = g_try_malloc0(blocks_clusters * s->cluster_size);
356 assert(table_size > 0 && blocks_clusters > 0);
357 if (new_table == NULL || new_blocks == NULL) {
358 ret = -ENOMEM;
359 goto fail_table;
362 /* Fill the new refcount table */
363 memcpy(new_table, s->refcount_table,
364 s->refcount_table_size * sizeof(uint64_t));
365 new_table[refcount_table_index] = new_block;
367 int i;
368 for (i = 0; i < blocks_clusters; i++) {
369 new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
372 /* Fill the refcount blocks */
373 uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
374 int block = 0;
375 for (i = 0; i < table_clusters + blocks_clusters; i++) {
376 new_blocks[block++] = cpu_to_be16(1);
379 /* Write refcount blocks to disk */
380 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
381 ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
382 blocks_clusters * s->cluster_size);
383 g_free(new_blocks);
384 new_blocks = NULL;
385 if (ret < 0) {
386 goto fail_table;
389 /* Write refcount table to disk */
390 for(i = 0; i < table_size; i++) {
391 cpu_to_be64s(&new_table[i]);
394 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
395 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
396 table_size * sizeof(uint64_t));
397 if (ret < 0) {
398 goto fail_table;
401 for(i = 0; i < table_size; i++) {
402 be64_to_cpus(&new_table[i]);
405 /* Hook up the new refcount table in the qcow2 header */
406 uint8_t data[12];
407 cpu_to_be64w((uint64_t*)data, table_offset);
408 cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
409 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
410 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
411 data, sizeof(data));
412 if (ret < 0) {
413 goto fail_table;
416 /* And switch it in memory */
417 uint64_t old_table_offset = s->refcount_table_offset;
418 uint64_t old_table_size = s->refcount_table_size;
420 g_free(s->refcount_table);
421 s->refcount_table = new_table;
422 s->refcount_table_size = table_size;
423 s->refcount_table_offset = table_offset;
425 /* Free old table. */
426 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
427 QCOW2_DISCARD_OTHER);
429 ret = load_refcount_block(bs, new_block, (void**) refcount_block);
430 if (ret < 0) {
431 return ret;
434 /* If we were trying to do the initial refcount update for some cluster
435 * allocation, we might have used the same clusters to store newly
436 * allocated metadata. Make the caller search some new space. */
437 return -EAGAIN;
439 fail_table:
440 g_free(new_blocks);
441 g_free(new_table);
442 fail_block:
443 if (*refcount_block != NULL) {
444 qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
446 return ret;
449 void qcow2_process_discards(BlockDriverState *bs, int ret)
451 BDRVQcowState *s = bs->opaque;
452 Qcow2DiscardRegion *d, *next;
454 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
455 QTAILQ_REMOVE(&s->discards, d, next);
457 /* Discard is optional, ignore the return value */
458 if (ret >= 0) {
459 bdrv_discard(bs->file,
460 d->offset >> BDRV_SECTOR_BITS,
461 d->bytes >> BDRV_SECTOR_BITS);
464 g_free(d);
468 static void update_refcount_discard(BlockDriverState *bs,
469 uint64_t offset, uint64_t length)
471 BDRVQcowState *s = bs->opaque;
472 Qcow2DiscardRegion *d, *p, *next;
474 QTAILQ_FOREACH(d, &s->discards, next) {
475 uint64_t new_start = MIN(offset, d->offset);
476 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
478 if (new_end - new_start <= length + d->bytes) {
479 /* There can't be any overlap, areas ending up here have no
480 * references any more and therefore shouldn't get freed another
481 * time. */
482 assert(d->bytes + length == new_end - new_start);
483 d->offset = new_start;
484 d->bytes = new_end - new_start;
485 goto found;
489 d = g_malloc(sizeof(*d));
490 *d = (Qcow2DiscardRegion) {
491 .bs = bs,
492 .offset = offset,
493 .bytes = length,
495 QTAILQ_INSERT_TAIL(&s->discards, d, next);
497 found:
498 /* Merge discard requests if they are adjacent now */
499 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
500 if (p == d
501 || p->offset > d->offset + d->bytes
502 || d->offset > p->offset + p->bytes)
504 continue;
507 /* Still no overlap possible */
508 assert(p->offset == d->offset + d->bytes
509 || d->offset == p->offset + p->bytes);
511 QTAILQ_REMOVE(&s->discards, p, next);
512 d->offset = MIN(d->offset, p->offset);
513 d->bytes += p->bytes;
517 /* XXX: cache several refcount block clusters ? */
518 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
519 int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
521 BDRVQcowState *s = bs->opaque;
522 int64_t start, last, cluster_offset;
523 uint16_t *refcount_block = NULL;
524 int64_t old_table_index = -1;
525 int ret;
527 #ifdef DEBUG_ALLOC2
528 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
529 offset, length, addend);
530 #endif
531 if (length < 0) {
532 return -EINVAL;
533 } else if (length == 0) {
534 return 0;
537 if (addend < 0) {
538 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
539 s->l2_table_cache);
542 start = start_of_cluster(s, offset);
543 last = start_of_cluster(s, offset + length - 1);
544 for(cluster_offset = start; cluster_offset <= last;
545 cluster_offset += s->cluster_size)
547 int block_index, refcount;
548 int64_t cluster_index = cluster_offset >> s->cluster_bits;
549 int64_t table_index =
550 cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
552 /* Load the refcount block and allocate it if needed */
553 if (table_index != old_table_index) {
554 if (refcount_block) {
555 ret = qcow2_cache_put(bs, s->refcount_block_cache,
556 (void**) &refcount_block);
557 if (ret < 0) {
558 goto fail;
562 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
563 if (ret < 0) {
564 goto fail;
567 old_table_index = table_index;
569 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
571 /* we can update the count and save it */
572 block_index = cluster_index &
573 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
575 refcount = be16_to_cpu(refcount_block[block_index]);
576 refcount += addend;
577 if (refcount < 0 || refcount > 0xffff) {
578 ret = -EINVAL;
579 goto fail;
581 if (refcount == 0 && cluster_index < s->free_cluster_index) {
582 s->free_cluster_index = cluster_index;
584 refcount_block[block_index] = cpu_to_be16(refcount);
586 if (refcount == 0 && s->discard_passthrough[type]) {
587 update_refcount_discard(bs, cluster_offset, s->cluster_size);
591 ret = 0;
592 fail:
593 if (!s->cache_discards) {
594 qcow2_process_discards(bs, ret);
597 /* Write last changed block to disk */
598 if (refcount_block) {
599 int wret;
600 wret = qcow2_cache_put(bs, s->refcount_block_cache,
601 (void**) &refcount_block);
602 if (wret < 0) {
603 return ret < 0 ? ret : wret;
608 * Try do undo any updates if an error is returned (This may succeed in
609 * some cases like ENOSPC for allocating a new refcount block)
611 if (ret < 0) {
612 int dummy;
613 dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
614 QCOW2_DISCARD_NEVER);
615 (void)dummy;
618 return ret;
622 * Increases or decreases the refcount of a given cluster by one.
623 * addend must be 1 or -1.
625 * If the return value is non-negative, it is the new refcount of the cluster.
626 * If it is negative, it is -errno and indicates an error.
628 int qcow2_update_cluster_refcount(BlockDriverState *bs,
629 int64_t cluster_index,
630 int addend,
631 enum qcow2_discard_type type)
633 BDRVQcowState *s = bs->opaque;
634 int ret;
636 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
637 type);
638 if (ret < 0) {
639 return ret;
642 return get_refcount(bs, cluster_index);
647 /*********************************************************/
648 /* cluster allocation functions */
652 /* return < 0 if error */
653 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size)
655 BDRVQcowState *s = bs->opaque;
656 uint64_t i, nb_clusters;
657 int refcount;
659 nb_clusters = size_to_clusters(s, size);
660 retry:
661 for(i = 0; i < nb_clusters; i++) {
662 uint64_t next_cluster_index = s->free_cluster_index++;
663 refcount = get_refcount(bs, next_cluster_index);
665 if (refcount < 0) {
666 return refcount;
667 } else if (refcount != 0) {
668 goto retry;
672 /* Make sure that all offsets in the "allocated" range are representable
673 * in an int64_t */
674 if (s->free_cluster_index > 0 &&
675 s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits))
677 return -EFBIG;
680 #ifdef DEBUG_ALLOC2
681 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
682 size,
683 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
684 #endif
685 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
688 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
690 int64_t offset;
691 int ret;
693 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
694 do {
695 offset = alloc_clusters_noref(bs, size);
696 if (offset < 0) {
697 return offset;
700 ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
701 } while (ret == -EAGAIN);
703 if (ret < 0) {
704 return ret;
707 return offset;
710 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
711 int nb_clusters)
713 BDRVQcowState *s = bs->opaque;
714 uint64_t cluster_index;
715 uint64_t i;
716 int refcount, ret;
718 assert(nb_clusters >= 0);
719 if (nb_clusters == 0) {
720 return 0;
723 do {
724 /* Check how many clusters there are free */
725 cluster_index = offset >> s->cluster_bits;
726 for(i = 0; i < nb_clusters; i++) {
727 refcount = get_refcount(bs, cluster_index++);
729 if (refcount < 0) {
730 return refcount;
731 } else if (refcount != 0) {
732 break;
736 /* And then allocate them */
737 ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
738 QCOW2_DISCARD_NEVER);
739 } while (ret == -EAGAIN);
741 if (ret < 0) {
742 return ret;
745 return i;
748 /* only used to allocate compressed sectors. We try to allocate
749 contiguous sectors. size must be <= cluster_size */
750 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
752 BDRVQcowState *s = bs->opaque;
753 int64_t offset, cluster_offset;
754 int free_in_cluster;
756 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
757 assert(size > 0 && size <= s->cluster_size);
758 if (s->free_byte_offset == 0) {
759 offset = qcow2_alloc_clusters(bs, s->cluster_size);
760 if (offset < 0) {
761 return offset;
763 s->free_byte_offset = offset;
765 redo:
766 free_in_cluster = s->cluster_size -
767 offset_into_cluster(s, s->free_byte_offset);
768 if (size <= free_in_cluster) {
769 /* enough space in current cluster */
770 offset = s->free_byte_offset;
771 s->free_byte_offset += size;
772 free_in_cluster -= size;
773 if (free_in_cluster == 0)
774 s->free_byte_offset = 0;
775 if (offset_into_cluster(s, offset) != 0)
776 qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
777 QCOW2_DISCARD_NEVER);
778 } else {
779 offset = qcow2_alloc_clusters(bs, s->cluster_size);
780 if (offset < 0) {
781 return offset;
783 cluster_offset = start_of_cluster(s, s->free_byte_offset);
784 if ((cluster_offset + s->cluster_size) == offset) {
785 /* we are lucky: contiguous data */
786 offset = s->free_byte_offset;
787 qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
788 QCOW2_DISCARD_NEVER);
789 s->free_byte_offset += size;
790 } else {
791 s->free_byte_offset = offset;
792 goto redo;
796 /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
797 * or explicitly by qcow2_update_cluster_refcount(). Refcount blocks must
798 * be flushed before the caller's L2 table updates.
800 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
801 return offset;
804 void qcow2_free_clusters(BlockDriverState *bs,
805 int64_t offset, int64_t size,
806 enum qcow2_discard_type type)
808 int ret;
810 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
811 ret = update_refcount(bs, offset, size, -1, type);
812 if (ret < 0) {
813 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
814 /* TODO Remember the clusters to free them later and avoid leaking */
819 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
820 * normal cluster, compressed cluster, etc.)
822 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
823 int nb_clusters, enum qcow2_discard_type type)
825 BDRVQcowState *s = bs->opaque;
827 switch (qcow2_get_cluster_type(l2_entry)) {
828 case QCOW2_CLUSTER_COMPRESSED:
830 int nb_csectors;
831 nb_csectors = ((l2_entry >> s->csize_shift) &
832 s->csize_mask) + 1;
833 qcow2_free_clusters(bs,
834 (l2_entry & s->cluster_offset_mask) & ~511,
835 nb_csectors * 512, type);
837 break;
838 case QCOW2_CLUSTER_NORMAL:
839 case QCOW2_CLUSTER_ZERO:
840 if (l2_entry & L2E_OFFSET_MASK) {
841 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
842 nb_clusters << s->cluster_bits, type);
844 break;
845 case QCOW2_CLUSTER_UNALLOCATED:
846 break;
847 default:
848 abort();
854 /*********************************************************/
855 /* snapshots and image creation */
859 /* update the refcounts of snapshots and the copied flag */
860 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
861 int64_t l1_table_offset, int l1_size, int addend)
863 BDRVQcowState *s = bs->opaque;
864 uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2;
865 bool l1_allocated = false;
866 int64_t old_offset, old_l2_offset;
867 int i, j, l1_modified = 0, nb_csectors, refcount;
868 int ret;
870 l2_table = NULL;
871 l1_table = NULL;
872 l1_size2 = l1_size * sizeof(uint64_t);
874 s->cache_discards = true;
876 /* WARNING: qcow2_snapshot_goto relies on this function not using the
877 * l1_table_offset when it is the current s->l1_table_offset! Be careful
878 * when changing this! */
879 if (l1_table_offset != s->l1_table_offset) {
880 l1_table = g_try_malloc0(align_offset(l1_size2, 512));
881 if (l1_size2 && l1_table == NULL) {
882 ret = -ENOMEM;
883 goto fail;
885 l1_allocated = true;
887 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
888 if (ret < 0) {
889 goto fail;
892 for(i = 0;i < l1_size; i++)
893 be64_to_cpus(&l1_table[i]);
894 } else {
895 assert(l1_size == s->l1_size);
896 l1_table = s->l1_table;
897 l1_allocated = false;
900 for(i = 0; i < l1_size; i++) {
901 l2_offset = l1_table[i];
902 if (l2_offset) {
903 old_l2_offset = l2_offset;
904 l2_offset &= L1E_OFFSET_MASK;
906 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
907 (void**) &l2_table);
908 if (ret < 0) {
909 goto fail;
912 for(j = 0; j < s->l2_size; j++) {
913 uint64_t cluster_index;
915 offset = be64_to_cpu(l2_table[j]);
916 old_offset = offset;
917 offset &= ~QCOW_OFLAG_COPIED;
919 switch (qcow2_get_cluster_type(offset)) {
920 case QCOW2_CLUSTER_COMPRESSED:
921 nb_csectors = ((offset >> s->csize_shift) &
922 s->csize_mask) + 1;
923 if (addend != 0) {
924 ret = update_refcount(bs,
925 (offset & s->cluster_offset_mask) & ~511,
926 nb_csectors * 512, addend,
927 QCOW2_DISCARD_SNAPSHOT);
928 if (ret < 0) {
929 goto fail;
932 /* compressed clusters are never modified */
933 refcount = 2;
934 break;
936 case QCOW2_CLUSTER_NORMAL:
937 case QCOW2_CLUSTER_ZERO:
938 cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
939 if (!cluster_index) {
940 /* unallocated */
941 refcount = 0;
942 break;
944 if (addend != 0) {
945 refcount = qcow2_update_cluster_refcount(bs,
946 cluster_index, addend,
947 QCOW2_DISCARD_SNAPSHOT);
948 } else {
949 refcount = get_refcount(bs, cluster_index);
952 if (refcount < 0) {
953 ret = refcount;
954 goto fail;
956 break;
958 case QCOW2_CLUSTER_UNALLOCATED:
959 refcount = 0;
960 break;
962 default:
963 abort();
966 if (refcount == 1) {
967 offset |= QCOW_OFLAG_COPIED;
969 if (offset != old_offset) {
970 if (addend > 0) {
971 qcow2_cache_set_dependency(bs, s->l2_table_cache,
972 s->refcount_block_cache);
974 l2_table[j] = cpu_to_be64(offset);
975 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
979 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
980 if (ret < 0) {
981 goto fail;
985 if (addend != 0) {
986 refcount = qcow2_update_cluster_refcount(bs, l2_offset >>
987 s->cluster_bits, addend, QCOW2_DISCARD_SNAPSHOT);
988 } else {
989 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
991 if (refcount < 0) {
992 ret = refcount;
993 goto fail;
994 } else if (refcount == 1) {
995 l2_offset |= QCOW_OFLAG_COPIED;
997 if (l2_offset != old_l2_offset) {
998 l1_table[i] = l2_offset;
999 l1_modified = 1;
1004 ret = bdrv_flush(bs);
1005 fail:
1006 if (l2_table) {
1007 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1010 s->cache_discards = false;
1011 qcow2_process_discards(bs, ret);
1013 /* Update L1 only if it isn't deleted anyway (addend = -1) */
1014 if (ret == 0 && addend >= 0 && l1_modified) {
1015 for (i = 0; i < l1_size; i++) {
1016 cpu_to_be64s(&l1_table[i]);
1019 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
1021 for (i = 0; i < l1_size; i++) {
1022 be64_to_cpus(&l1_table[i]);
1025 if (l1_allocated)
1026 g_free(l1_table);
1027 return ret;
1033 /*********************************************************/
1034 /* refcount checking functions */
1039 * Increases the refcount for a range of clusters in a given refcount table.
1040 * This is used to construct a temporary refcount table out of L1 and L2 tables
1041 * which can be compared the the refcount table saved in the image.
1043 * Modifies the number of errors in res.
1045 static void inc_refcounts(BlockDriverState *bs,
1046 BdrvCheckResult *res,
1047 uint16_t *refcount_table,
1048 int refcount_table_size,
1049 int64_t offset, int64_t size)
1051 BDRVQcowState *s = bs->opaque;
1052 uint64_t start, last, cluster_offset, k;
1054 if (size <= 0)
1055 return;
1057 start = start_of_cluster(s, offset);
1058 last = start_of_cluster(s, offset + size - 1);
1059 for(cluster_offset = start; cluster_offset <= last;
1060 cluster_offset += s->cluster_size) {
1061 k = cluster_offset >> s->cluster_bits;
1062 if (k >= refcount_table_size) {
1063 fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
1064 "the end of the image file, can't properly check refcounts.\n",
1065 cluster_offset);
1066 res->check_errors++;
1067 } else {
1068 if (++refcount_table[k] == 0) {
1069 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1070 "\n", cluster_offset);
1071 res->corruptions++;
1077 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1078 enum {
1079 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1083 * Increases the refcount in the given refcount table for the all clusters
1084 * referenced in the L2 table. While doing so, performs some checks on L2
1085 * entries.
1087 * Returns the number of errors found by the checks or -errno if an internal
1088 * error occurred.
1090 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1091 uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
1092 int flags)
1094 BDRVQcowState *s = bs->opaque;
1095 uint64_t *l2_table, l2_entry;
1096 uint64_t next_contiguous_offset = 0;
1097 int i, l2_size, nb_csectors;
1099 /* Read L2 table from disk */
1100 l2_size = s->l2_size * sizeof(uint64_t);
1101 l2_table = g_malloc(l2_size);
1103 if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
1104 goto fail;
1106 /* Do the actual checks */
1107 for(i = 0; i < s->l2_size; i++) {
1108 l2_entry = be64_to_cpu(l2_table[i]);
1110 switch (qcow2_get_cluster_type(l2_entry)) {
1111 case QCOW2_CLUSTER_COMPRESSED:
1112 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1113 if (l2_entry & QCOW_OFLAG_COPIED) {
1114 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1115 "copied flag must never be set for compressed "
1116 "clusters\n", l2_entry >> s->cluster_bits);
1117 l2_entry &= ~QCOW_OFLAG_COPIED;
1118 res->corruptions++;
1121 /* Mark cluster as used */
1122 nb_csectors = ((l2_entry >> s->csize_shift) &
1123 s->csize_mask) + 1;
1124 l2_entry &= s->cluster_offset_mask;
1125 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1126 l2_entry & ~511, nb_csectors * 512);
1128 if (flags & CHECK_FRAG_INFO) {
1129 res->bfi.allocated_clusters++;
1130 res->bfi.compressed_clusters++;
1132 /* Compressed clusters are fragmented by nature. Since they
1133 * take up sub-sector space but we only have sector granularity
1134 * I/O we need to re-read the same sectors even for adjacent
1135 * compressed clusters.
1137 res->bfi.fragmented_clusters++;
1139 break;
1141 case QCOW2_CLUSTER_ZERO:
1142 if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1143 break;
1145 /* fall through */
1147 case QCOW2_CLUSTER_NORMAL:
1149 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1151 if (flags & CHECK_FRAG_INFO) {
1152 res->bfi.allocated_clusters++;
1153 if (next_contiguous_offset &&
1154 offset != next_contiguous_offset) {
1155 res->bfi.fragmented_clusters++;
1157 next_contiguous_offset = offset + s->cluster_size;
1160 /* Mark cluster as used */
1161 inc_refcounts(bs, res, refcount_table,refcount_table_size,
1162 offset, s->cluster_size);
1164 /* Correct offsets are cluster aligned */
1165 if (offset_into_cluster(s, offset)) {
1166 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1167 "properly aligned; L2 entry corrupted.\n", offset);
1168 res->corruptions++;
1170 break;
1173 case QCOW2_CLUSTER_UNALLOCATED:
1174 break;
1176 default:
1177 abort();
1181 g_free(l2_table);
1182 return 0;
1184 fail:
1185 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1186 g_free(l2_table);
1187 return -EIO;
1191 * Increases the refcount for the L1 table, its L2 tables and all referenced
1192 * clusters in the given refcount table. While doing so, performs some checks
1193 * on L1 and L2 entries.
1195 * Returns the number of errors found by the checks or -errno if an internal
1196 * error occurred.
1198 static int check_refcounts_l1(BlockDriverState *bs,
1199 BdrvCheckResult *res,
1200 uint16_t *refcount_table,
1201 int refcount_table_size,
1202 int64_t l1_table_offset, int l1_size,
1203 int flags)
1205 BDRVQcowState *s = bs->opaque;
1206 uint64_t *l1_table, l2_offset, l1_size2;
1207 int i, ret;
1209 l1_size2 = l1_size * sizeof(uint64_t);
1211 /* Mark L1 table as used */
1212 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1213 l1_table_offset, l1_size2);
1215 /* Read L1 table entries from disk */
1216 if (l1_size2 == 0) {
1217 l1_table = NULL;
1218 } else {
1219 l1_table = g_try_malloc(l1_size2);
1220 if (l1_table == NULL) {
1221 ret = -ENOMEM;
1222 goto fail;
1224 if (bdrv_pread(bs->file, l1_table_offset,
1225 l1_table, l1_size2) != l1_size2)
1226 goto fail;
1227 for(i = 0;i < l1_size; i++)
1228 be64_to_cpus(&l1_table[i]);
1231 /* Do the actual checks */
1232 for(i = 0; i < l1_size; i++) {
1233 l2_offset = l1_table[i];
1234 if (l2_offset) {
1235 /* Mark L2 table as used */
1236 l2_offset &= L1E_OFFSET_MASK;
1237 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1238 l2_offset, s->cluster_size);
1240 /* L2 tables are cluster aligned */
1241 if (offset_into_cluster(s, l2_offset)) {
1242 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1243 "cluster aligned; L1 entry corrupted\n", l2_offset);
1244 res->corruptions++;
1247 /* Process and check L2 entries */
1248 ret = check_refcounts_l2(bs, res, refcount_table,
1249 refcount_table_size, l2_offset, flags);
1250 if (ret < 0) {
1251 goto fail;
1255 g_free(l1_table);
1256 return 0;
1258 fail:
1259 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1260 res->check_errors++;
1261 g_free(l1_table);
1262 return -EIO;
1266 * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1268 * This function does not print an error message nor does it increment
1269 * check_errors if get_refcount fails (this is because such an error will have
1270 * been already detected and sufficiently signaled by the calling function
1271 * (qcow2_check_refcounts) by the time this function is called).
1273 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1274 BdrvCheckMode fix)
1276 BDRVQcowState *s = bs->opaque;
1277 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1278 int ret;
1279 int refcount;
1280 int i, j;
1282 for (i = 0; i < s->l1_size; i++) {
1283 uint64_t l1_entry = s->l1_table[i];
1284 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1285 bool l2_dirty = false;
1287 if (!l2_offset) {
1288 continue;
1291 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
1292 if (refcount < 0) {
1293 /* don't print message nor increment check_errors */
1294 continue;
1296 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1297 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1298 "l1_entry=%" PRIx64 " refcount=%d\n",
1299 fix & BDRV_FIX_ERRORS ? "Repairing" :
1300 "ERROR",
1301 i, l1_entry, refcount);
1302 if (fix & BDRV_FIX_ERRORS) {
1303 s->l1_table[i] = refcount == 1
1304 ? l1_entry | QCOW_OFLAG_COPIED
1305 : l1_entry & ~QCOW_OFLAG_COPIED;
1306 ret = qcow2_write_l1_entry(bs, i);
1307 if (ret < 0) {
1308 res->check_errors++;
1309 goto fail;
1311 res->corruptions_fixed++;
1312 } else {
1313 res->corruptions++;
1317 ret = bdrv_pread(bs->file, l2_offset, l2_table,
1318 s->l2_size * sizeof(uint64_t));
1319 if (ret < 0) {
1320 fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1321 strerror(-ret));
1322 res->check_errors++;
1323 goto fail;
1326 for (j = 0; j < s->l2_size; j++) {
1327 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1328 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1329 int cluster_type = qcow2_get_cluster_type(l2_entry);
1331 if ((cluster_type == QCOW2_CLUSTER_NORMAL) ||
1332 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) {
1333 refcount = get_refcount(bs, data_offset >> s->cluster_bits);
1334 if (refcount < 0) {
1335 /* don't print message nor increment check_errors */
1336 continue;
1338 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1339 fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1340 "l2_entry=%" PRIx64 " refcount=%d\n",
1341 fix & BDRV_FIX_ERRORS ? "Repairing" :
1342 "ERROR",
1343 l2_entry, refcount);
1344 if (fix & BDRV_FIX_ERRORS) {
1345 l2_table[j] = cpu_to_be64(refcount == 1
1346 ? l2_entry | QCOW_OFLAG_COPIED
1347 : l2_entry & ~QCOW_OFLAG_COPIED);
1348 l2_dirty = true;
1349 res->corruptions_fixed++;
1350 } else {
1351 res->corruptions++;
1357 if (l2_dirty) {
1358 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1359 l2_offset, s->cluster_size);
1360 if (ret < 0) {
1361 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1362 "overlap check failed: %s\n", strerror(-ret));
1363 res->check_errors++;
1364 goto fail;
1367 ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size);
1368 if (ret < 0) {
1369 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1370 strerror(-ret));
1371 res->check_errors++;
1372 goto fail;
1377 ret = 0;
1379 fail:
1380 qemu_vfree(l2_table);
1381 return ret;
1385 * Writes one sector of the refcount table to the disk
1387 #define RT_ENTRIES_PER_SECTOR (512 / sizeof(uint64_t))
1388 static int write_reftable_entry(BlockDriverState *bs, int rt_index)
1390 BDRVQcowState *s = bs->opaque;
1391 uint64_t buf[RT_ENTRIES_PER_SECTOR];
1392 int rt_start_index;
1393 int i, ret;
1395 rt_start_index = rt_index & ~(RT_ENTRIES_PER_SECTOR - 1);
1396 for (i = 0; i < RT_ENTRIES_PER_SECTOR; i++) {
1397 buf[i] = cpu_to_be64(s->refcount_table[rt_start_index + i]);
1400 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_REFCOUNT_TABLE,
1401 s->refcount_table_offset + rt_start_index * sizeof(uint64_t),
1402 sizeof(buf));
1403 if (ret < 0) {
1404 return ret;
1407 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_UPDATE);
1408 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset +
1409 rt_start_index * sizeof(uint64_t), buf, sizeof(buf));
1410 if (ret < 0) {
1411 return ret;
1414 return 0;
1418 * Allocates a new cluster for the given refcount block (represented by its
1419 * offset in the image file) and copies the current content there. This function
1420 * does _not_ decrement the reference count for the currently occupied cluster.
1422 * This function prints an informative message to stderr on error (and returns
1423 * -errno); on success, the offset of the newly allocated cluster is returned.
1425 static int64_t realloc_refcount_block(BlockDriverState *bs, int reftable_index,
1426 uint64_t offset)
1428 BDRVQcowState *s = bs->opaque;
1429 int64_t new_offset = 0;
1430 void *refcount_block = NULL;
1431 int ret;
1433 /* allocate new refcount block */
1434 new_offset = qcow2_alloc_clusters(bs, s->cluster_size);
1435 if (new_offset < 0) {
1436 fprintf(stderr, "Could not allocate new cluster: %s\n",
1437 strerror(-new_offset));
1438 ret = new_offset;
1439 goto done;
1442 /* fetch current refcount block content */
1443 ret = qcow2_cache_get(bs, s->refcount_block_cache, offset, &refcount_block);
1444 if (ret < 0) {
1445 fprintf(stderr, "Could not fetch refcount block: %s\n", strerror(-ret));
1446 goto fail_free_cluster;
1449 /* new block has not yet been entered into refcount table, therefore it is
1450 * no refcount block yet (regarding this check) */
1451 ret = qcow2_pre_write_overlap_check(bs, 0, new_offset, s->cluster_size);
1452 if (ret < 0) {
1453 fprintf(stderr, "Could not write refcount block; metadata overlap "
1454 "check failed: %s\n", strerror(-ret));
1455 /* the image will be marked corrupt, so don't even attempt on freeing
1456 * the cluster */
1457 goto done;
1460 /* write to new block */
1461 ret = bdrv_write(bs->file, new_offset / BDRV_SECTOR_SIZE, refcount_block,
1462 s->cluster_sectors);
1463 if (ret < 0) {
1464 fprintf(stderr, "Could not write refcount block: %s\n", strerror(-ret));
1465 goto fail_free_cluster;
1468 /* update refcount table */
1469 assert(!offset_into_cluster(s, new_offset));
1470 s->refcount_table[reftable_index] = new_offset;
1471 ret = write_reftable_entry(bs, reftable_index);
1472 if (ret < 0) {
1473 fprintf(stderr, "Could not update refcount table: %s\n",
1474 strerror(-ret));
1475 goto fail_free_cluster;
1478 goto done;
1480 fail_free_cluster:
1481 qcow2_free_clusters(bs, new_offset, s->cluster_size, QCOW2_DISCARD_OTHER);
1483 done:
1484 if (refcount_block) {
1485 /* This should never fail, as it would only do so if the given refcount
1486 * block cannot be found in the cache. As this is impossible as long as
1487 * there are no bugs, assert the success. */
1488 int tmp = qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
1489 assert(tmp == 0);
1492 if (ret < 0) {
1493 return ret;
1496 return new_offset;
1500 * Checks an image for refcount consistency.
1502 * Returns 0 if no errors are found, the number of errors in case the image is
1503 * detected as corrupted, and -errno when an internal error occurred.
1505 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1506 BdrvCheckMode fix)
1508 BDRVQcowState *s = bs->opaque;
1509 int64_t size, i, highest_cluster, nb_clusters;
1510 int refcount1, refcount2;
1511 QCowSnapshot *sn;
1512 uint16_t *refcount_table;
1513 int ret;
1515 size = bdrv_getlength(bs->file);
1516 if (size < 0) {
1517 res->check_errors++;
1518 return size;
1521 nb_clusters = size_to_clusters(s, size);
1522 if (nb_clusters > INT_MAX) {
1523 res->check_errors++;
1524 return -EFBIG;
1527 refcount_table = g_try_new0(uint16_t, nb_clusters);
1528 if (nb_clusters && refcount_table == NULL) {
1529 res->check_errors++;
1530 return -ENOMEM;
1533 res->bfi.total_clusters =
1534 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1536 /* header */
1537 inc_refcounts(bs, res, refcount_table, nb_clusters,
1538 0, s->cluster_size);
1540 /* current L1 table */
1541 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1542 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO);
1543 if (ret < 0) {
1544 goto fail;
1547 /* snapshots */
1548 for(i = 0; i < s->nb_snapshots; i++) {
1549 sn = s->snapshots + i;
1550 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1551 sn->l1_table_offset, sn->l1_size, 0);
1552 if (ret < 0) {
1553 goto fail;
1556 inc_refcounts(bs, res, refcount_table, nb_clusters,
1557 s->snapshots_offset, s->snapshots_size);
1559 /* refcount data */
1560 inc_refcounts(bs, res, refcount_table, nb_clusters,
1561 s->refcount_table_offset,
1562 s->refcount_table_size * sizeof(uint64_t));
1564 for(i = 0; i < s->refcount_table_size; i++) {
1565 uint64_t offset, cluster;
1566 offset = s->refcount_table[i];
1567 cluster = offset >> s->cluster_bits;
1569 /* Refcount blocks are cluster aligned */
1570 if (offset_into_cluster(s, offset)) {
1571 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1572 "cluster aligned; refcount table entry corrupted\n", i);
1573 res->corruptions++;
1574 continue;
1577 if (cluster >= nb_clusters) {
1578 fprintf(stderr, "ERROR refcount block %" PRId64
1579 " is outside image\n", i);
1580 res->corruptions++;
1581 continue;
1584 if (offset != 0) {
1585 inc_refcounts(bs, res, refcount_table, nb_clusters,
1586 offset, s->cluster_size);
1587 if (refcount_table[cluster] != 1) {
1588 fprintf(stderr, "%s refcount block %" PRId64
1589 " refcount=%d\n",
1590 fix & BDRV_FIX_ERRORS ? "Repairing" :
1591 "ERROR",
1592 i, refcount_table[cluster]);
1594 if (fix & BDRV_FIX_ERRORS) {
1595 int64_t new_offset;
1597 new_offset = realloc_refcount_block(bs, i, offset);
1598 if (new_offset < 0) {
1599 res->corruptions++;
1600 continue;
1603 /* update refcounts */
1604 if ((new_offset >> s->cluster_bits) >= nb_clusters) {
1605 /* increase refcount_table size if necessary */
1606 int old_nb_clusters = nb_clusters;
1607 nb_clusters = (new_offset >> s->cluster_bits) + 1;
1608 refcount_table = g_renew(uint16_t, refcount_table,
1609 nb_clusters);
1610 memset(&refcount_table[old_nb_clusters], 0, (nb_clusters
1611 - old_nb_clusters) * sizeof(uint16_t));
1613 refcount_table[cluster]--;
1614 inc_refcounts(bs, res, refcount_table, nb_clusters,
1615 new_offset, s->cluster_size);
1617 res->corruptions_fixed++;
1618 } else {
1619 res->corruptions++;
1625 /* compare ref counts */
1626 for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1627 refcount1 = get_refcount(bs, i);
1628 if (refcount1 < 0) {
1629 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1630 i, strerror(-refcount1));
1631 res->check_errors++;
1632 continue;
1635 refcount2 = refcount_table[i];
1637 if (refcount1 > 0 || refcount2 > 0) {
1638 highest_cluster = i;
1641 if (refcount1 != refcount2) {
1643 /* Check if we're allowed to fix the mismatch */
1644 int *num_fixed = NULL;
1645 if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1646 num_fixed = &res->leaks_fixed;
1647 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1648 num_fixed = &res->corruptions_fixed;
1651 fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1652 num_fixed != NULL ? "Repairing" :
1653 refcount1 < refcount2 ? "ERROR" :
1654 "Leaked",
1655 i, refcount1, refcount2);
1657 if (num_fixed) {
1658 ret = update_refcount(bs, i << s->cluster_bits, 1,
1659 refcount2 - refcount1,
1660 QCOW2_DISCARD_ALWAYS);
1661 if (ret >= 0) {
1662 (*num_fixed)++;
1663 continue;
1667 /* And if we couldn't, print an error */
1668 if (refcount1 < refcount2) {
1669 res->corruptions++;
1670 } else {
1671 res->leaks++;
1676 /* check OFLAG_COPIED */
1677 ret = check_oflag_copied(bs, res, fix);
1678 if (ret < 0) {
1679 goto fail;
1682 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1683 ret = 0;
1685 fail:
1686 g_free(refcount_table);
1688 return ret;
1691 #define overlaps_with(ofs, sz) \
1692 ranges_overlap(offset, size, ofs, sz)
1695 * Checks if the given offset into the image file is actually free to use by
1696 * looking for overlaps with important metadata sections (L1/L2 tables etc.),
1697 * i.e. a sanity check without relying on the refcount tables.
1699 * The ign parameter specifies what checks not to perform (being a bitmask of
1700 * QCow2MetadataOverlap values), i.e., what sections to ignore.
1702 * Returns:
1703 * - 0 if writing to this offset will not affect the mentioned metadata
1704 * - a positive QCow2MetadataOverlap value indicating one overlapping section
1705 * - a negative value (-errno) indicating an error while performing a check,
1706 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
1708 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
1709 int64_t size)
1711 BDRVQcowState *s = bs->opaque;
1712 int chk = s->overlap_check & ~ign;
1713 int i, j;
1715 if (!size) {
1716 return 0;
1719 if (chk & QCOW2_OL_MAIN_HEADER) {
1720 if (offset < s->cluster_size) {
1721 return QCOW2_OL_MAIN_HEADER;
1725 /* align range to test to cluster boundaries */
1726 size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size);
1727 offset = start_of_cluster(s, offset);
1729 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
1730 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
1731 return QCOW2_OL_ACTIVE_L1;
1735 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
1736 if (overlaps_with(s->refcount_table_offset,
1737 s->refcount_table_size * sizeof(uint64_t))) {
1738 return QCOW2_OL_REFCOUNT_TABLE;
1742 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
1743 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
1744 return QCOW2_OL_SNAPSHOT_TABLE;
1748 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
1749 for (i = 0; i < s->nb_snapshots; i++) {
1750 if (s->snapshots[i].l1_size &&
1751 overlaps_with(s->snapshots[i].l1_table_offset,
1752 s->snapshots[i].l1_size * sizeof(uint64_t))) {
1753 return QCOW2_OL_INACTIVE_L1;
1758 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
1759 for (i = 0; i < s->l1_size; i++) {
1760 if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
1761 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
1762 s->cluster_size)) {
1763 return QCOW2_OL_ACTIVE_L2;
1768 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
1769 for (i = 0; i < s->refcount_table_size; i++) {
1770 if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
1771 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
1772 s->cluster_size)) {
1773 return QCOW2_OL_REFCOUNT_BLOCK;
1778 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
1779 for (i = 0; i < s->nb_snapshots; i++) {
1780 uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
1781 uint32_t l1_sz = s->snapshots[i].l1_size;
1782 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
1783 uint64_t *l1 = g_try_malloc(l1_sz2);
1784 int ret;
1786 if (l1_sz2 && l1 == NULL) {
1787 return -ENOMEM;
1790 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
1791 if (ret < 0) {
1792 g_free(l1);
1793 return ret;
1796 for (j = 0; j < l1_sz; j++) {
1797 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
1798 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
1799 g_free(l1);
1800 return QCOW2_OL_INACTIVE_L2;
1804 g_free(l1);
1808 return 0;
1811 static const char *metadata_ol_names[] = {
1812 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header",
1813 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table",
1814 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table",
1815 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
1816 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
1817 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
1818 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table",
1819 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table",
1823 * First performs a check for metadata overlaps (through
1824 * qcow2_check_metadata_overlap); if that fails with a negative value (error
1825 * while performing a check), that value is returned. If an impending overlap
1826 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
1827 * and -EIO returned.
1829 * Returns 0 if there were neither overlaps nor errors while checking for
1830 * overlaps; or a negative value (-errno) on error.
1832 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
1833 int64_t size)
1835 int ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
1837 if (ret < 0) {
1838 return ret;
1839 } else if (ret > 0) {
1840 int metadata_ol_bitnr = ffs(ret) - 1;
1841 char *message;
1843 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
1845 fprintf(stderr, "qcow2: Preventing invalid write on metadata (overlaps "
1846 "with %s); image marked as corrupt.\n",
1847 metadata_ol_names[metadata_ol_bitnr]);
1848 message = g_strdup_printf("Prevented %s overwrite",
1849 metadata_ol_names[metadata_ol_bitnr]);
1850 qapi_event_send_block_image_corrupted(bdrv_get_device_name(bs),
1851 message,
1852 true,
1853 offset,
1854 true,
1855 size,
1856 &error_abort);
1857 g_free(message);
1859 qcow2_mark_corrupt(bs);
1860 bs->drv = NULL; /* make BDS unusable */
1861 return -EIO;
1864 return 0;