Merge remote-tracking branch 'afaerber/tags/qom-cpu-for-anthony' into staging
[qemu/ar7.git] / block / qcow2-refcount.c
blob1244693f39e73e31d765feef34968ae0728735ef
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"
29 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size);
30 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
31 int64_t offset, int64_t length,
32 int addend, enum qcow2_discard_type type);
35 /*********************************************************/
36 /* refcount handling */
38 int qcow2_refcount_init(BlockDriverState *bs)
40 BDRVQcowState *s = bs->opaque;
41 int ret, refcount_table_size2, i;
43 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
44 s->refcount_table = g_malloc(refcount_table_size2);
45 if (s->refcount_table_size > 0) {
46 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
47 ret = bdrv_pread(bs->file, s->refcount_table_offset,
48 s->refcount_table, refcount_table_size2);
49 if (ret != refcount_table_size2)
50 goto fail;
51 for(i = 0; i < s->refcount_table_size; i++)
52 be64_to_cpus(&s->refcount_table[i]);
54 return 0;
55 fail:
56 return -ENOMEM;
59 void qcow2_refcount_close(BlockDriverState *bs)
61 BDRVQcowState *s = bs->opaque;
62 g_free(s->refcount_table);
66 static int load_refcount_block(BlockDriverState *bs,
67 int64_t refcount_block_offset,
68 void **refcount_block)
70 BDRVQcowState *s = bs->opaque;
71 int ret;
73 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
74 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
75 refcount_block);
77 return ret;
81 * Returns the refcount of the cluster given by its index. Any non-negative
82 * return value is the refcount of the cluster, negative values are -errno
83 * and indicate an error.
85 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
87 BDRVQcowState *s = bs->opaque;
88 int refcount_table_index, block_index;
89 int64_t refcount_block_offset;
90 int ret;
91 uint16_t *refcount_block;
92 uint16_t refcount;
94 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
95 if (refcount_table_index >= s->refcount_table_size)
96 return 0;
97 refcount_block_offset = s->refcount_table[refcount_table_index];
98 if (!refcount_block_offset)
99 return 0;
101 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
102 (void**) &refcount_block);
103 if (ret < 0) {
104 return ret;
107 block_index = cluster_index &
108 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
109 refcount = be16_to_cpu(refcount_block[block_index]);
111 ret = qcow2_cache_put(bs, s->refcount_block_cache,
112 (void**) &refcount_block);
113 if (ret < 0) {
114 return ret;
117 return refcount;
121 * Rounds the refcount table size up to avoid growing the table for each single
122 * refcount block that is allocated.
124 static unsigned int next_refcount_table_size(BDRVQcowState *s,
125 unsigned int min_size)
127 unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
128 unsigned int refcount_table_clusters =
129 MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
131 while (min_clusters > refcount_table_clusters) {
132 refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
135 return refcount_table_clusters << (s->cluster_bits - 3);
139 /* Checks if two offsets are described by the same refcount block */
140 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
141 uint64_t offset_b)
143 uint64_t block_a = offset_a >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
144 uint64_t block_b = offset_b >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
146 return (block_a == block_b);
150 * Loads a refcount block. If it doesn't exist yet, it is allocated first
151 * (including growing the refcount table if needed).
153 * Returns 0 on success or -errno in error case
155 static int alloc_refcount_block(BlockDriverState *bs,
156 int64_t cluster_index, uint16_t **refcount_block)
158 BDRVQcowState *s = bs->opaque;
159 unsigned int refcount_table_index;
160 int ret;
162 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
164 /* Find the refcount block for the given cluster */
165 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
167 if (refcount_table_index < s->refcount_table_size) {
169 uint64_t refcount_block_offset =
170 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
172 /* If it's already there, we're done */
173 if (refcount_block_offset) {
174 return load_refcount_block(bs, refcount_block_offset,
175 (void**) refcount_block);
180 * If we came here, we need to allocate something. Something is at least
181 * a cluster for the new refcount block. It may also include a new refcount
182 * table if the old refcount table is too small.
184 * Note that allocating clusters here needs some special care:
186 * - We can't use the normal qcow2_alloc_clusters(), it would try to
187 * increase the refcount and very likely we would end up with an endless
188 * recursion. Instead we must place the refcount blocks in a way that
189 * they can describe them themselves.
191 * - We need to consider that at this point we are inside update_refcounts
192 * and doing the initial refcount increase. This means that some clusters
193 * have already been allocated by the caller, but their refcount isn't
194 * accurate yet. free_cluster_index tells us where this allocation ends
195 * as long as we don't overwrite it by freeing clusters.
197 * - alloc_clusters_noref and qcow2_free_clusters may load a different
198 * refcount block into the cache
201 *refcount_block = NULL;
203 /* We write to the refcount table, so we might depend on L2 tables */
204 ret = qcow2_cache_flush(bs, s->l2_table_cache);
205 if (ret < 0) {
206 return ret;
209 /* Allocate the refcount block itself and mark it as used */
210 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
211 if (new_block < 0) {
212 return new_block;
215 #ifdef DEBUG_ALLOC2
216 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
217 " at %" PRIx64 "\n",
218 refcount_table_index, cluster_index << s->cluster_bits, new_block);
219 #endif
221 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
222 /* Zero the new refcount block before updating it */
223 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
224 (void**) refcount_block);
225 if (ret < 0) {
226 goto fail_block;
229 memset(*refcount_block, 0, s->cluster_size);
231 /* The block describes itself, need to update the cache */
232 int block_index = (new_block >> s->cluster_bits) &
233 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
234 (*refcount_block)[block_index] = cpu_to_be16(1);
235 } else {
236 /* Described somewhere else. This can recurse at most twice before we
237 * arrive at a block that describes itself. */
238 ret = update_refcount(bs, new_block, s->cluster_size, 1,
239 QCOW2_DISCARD_NEVER);
240 if (ret < 0) {
241 goto fail_block;
244 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
245 if (ret < 0) {
246 goto fail_block;
249 /* Initialize the new refcount block only after updating its refcount,
250 * update_refcount uses the refcount cache itself */
251 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
252 (void**) refcount_block);
253 if (ret < 0) {
254 goto fail_block;
257 memset(*refcount_block, 0, s->cluster_size);
260 /* Now the new refcount block needs to be written to disk */
261 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
262 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
263 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
264 if (ret < 0) {
265 goto fail_block;
268 /* If the refcount table is big enough, just hook the block up there */
269 if (refcount_table_index < s->refcount_table_size) {
270 uint64_t data64 = cpu_to_be64(new_block);
271 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
272 ret = bdrv_pwrite_sync(bs->file,
273 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
274 &data64, sizeof(data64));
275 if (ret < 0) {
276 goto fail_block;
279 s->refcount_table[refcount_table_index] = new_block;
280 return 0;
283 ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
284 if (ret < 0) {
285 goto fail_block;
289 * If we come here, we need to grow the refcount table. Again, a new
290 * refcount table needs some space and we can't simply allocate to avoid
291 * endless recursion.
293 * Therefore let's grab new refcount blocks at the end of the image, which
294 * will describe themselves and the new refcount table. This way we can
295 * reference them only in the new table and do the switch to the new
296 * refcount table at once without producing an inconsistent state in
297 * between.
299 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
301 /* Calculate the number of refcount blocks needed so far */
302 uint64_t refcount_block_clusters = 1 << (s->cluster_bits - REFCOUNT_SHIFT);
303 uint64_t blocks_used = (s->free_cluster_index +
304 refcount_block_clusters - 1) / refcount_block_clusters;
306 /* And now we need at least one block more for the new metadata */
307 uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
308 uint64_t last_table_size;
309 uint64_t blocks_clusters;
310 do {
311 uint64_t table_clusters =
312 size_to_clusters(s, table_size * sizeof(uint64_t));
313 blocks_clusters = 1 +
314 ((table_clusters + refcount_block_clusters - 1)
315 / refcount_block_clusters);
316 uint64_t meta_clusters = table_clusters + blocks_clusters;
318 last_table_size = table_size;
319 table_size = next_refcount_table_size(s, blocks_used +
320 ((meta_clusters + refcount_block_clusters - 1)
321 / refcount_block_clusters));
323 } while (last_table_size != table_size);
325 #ifdef DEBUG_ALLOC2
326 fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
327 s->refcount_table_size, table_size);
328 #endif
330 /* Create the new refcount table and blocks */
331 uint64_t meta_offset = (blocks_used * refcount_block_clusters) *
332 s->cluster_size;
333 uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
334 uint16_t *new_blocks = g_malloc0(blocks_clusters * s->cluster_size);
335 uint64_t *new_table = g_malloc0(table_size * sizeof(uint64_t));
337 assert(meta_offset >= (s->free_cluster_index * s->cluster_size));
339 /* Fill the new refcount table */
340 memcpy(new_table, s->refcount_table,
341 s->refcount_table_size * sizeof(uint64_t));
342 new_table[refcount_table_index] = new_block;
344 int i;
345 for (i = 0; i < blocks_clusters; i++) {
346 new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
349 /* Fill the refcount blocks */
350 uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
351 int block = 0;
352 for (i = 0; i < table_clusters + blocks_clusters; i++) {
353 new_blocks[block++] = cpu_to_be16(1);
356 /* Write refcount blocks to disk */
357 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
358 ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
359 blocks_clusters * s->cluster_size);
360 g_free(new_blocks);
361 if (ret < 0) {
362 goto fail_table;
365 /* Write refcount table to disk */
366 for(i = 0; i < table_size; i++) {
367 cpu_to_be64s(&new_table[i]);
370 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
371 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
372 table_size * sizeof(uint64_t));
373 if (ret < 0) {
374 goto fail_table;
377 for(i = 0; i < table_size; i++) {
378 be64_to_cpus(&new_table[i]);
381 /* Hook up the new refcount table in the qcow2 header */
382 uint8_t data[12];
383 cpu_to_be64w((uint64_t*)data, table_offset);
384 cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
385 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
386 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
387 data, sizeof(data));
388 if (ret < 0) {
389 goto fail_table;
392 /* And switch it in memory */
393 uint64_t old_table_offset = s->refcount_table_offset;
394 uint64_t old_table_size = s->refcount_table_size;
396 g_free(s->refcount_table);
397 s->refcount_table = new_table;
398 s->refcount_table_size = table_size;
399 s->refcount_table_offset = table_offset;
401 /* Free old table. Remember, we must not change free_cluster_index */
402 uint64_t old_free_cluster_index = s->free_cluster_index;
403 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
404 QCOW2_DISCARD_OTHER);
405 s->free_cluster_index = old_free_cluster_index;
407 ret = load_refcount_block(bs, new_block, (void**) refcount_block);
408 if (ret < 0) {
409 return ret;
412 return 0;
414 fail_table:
415 g_free(new_table);
416 fail_block:
417 if (*refcount_block != NULL) {
418 qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
420 return ret;
423 void qcow2_process_discards(BlockDriverState *bs, int ret)
425 BDRVQcowState *s = bs->opaque;
426 Qcow2DiscardRegion *d, *next;
428 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
429 QTAILQ_REMOVE(&s->discards, d, next);
431 /* Discard is optional, ignore the return value */
432 if (ret >= 0) {
433 bdrv_discard(bs->file,
434 d->offset >> BDRV_SECTOR_BITS,
435 d->bytes >> BDRV_SECTOR_BITS);
438 g_free(d);
442 static void update_refcount_discard(BlockDriverState *bs,
443 uint64_t offset, uint64_t length)
445 BDRVQcowState *s = bs->opaque;
446 Qcow2DiscardRegion *d, *p, *next;
448 QTAILQ_FOREACH(d, &s->discards, next) {
449 uint64_t new_start = MIN(offset, d->offset);
450 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
452 if (new_end - new_start <= length + d->bytes) {
453 /* There can't be any overlap, areas ending up here have no
454 * references any more and therefore shouldn't get freed another
455 * time. */
456 assert(d->bytes + length == new_end - new_start);
457 d->offset = new_start;
458 d->bytes = new_end - new_start;
459 goto found;
463 d = g_malloc(sizeof(*d));
464 *d = (Qcow2DiscardRegion) {
465 .bs = bs,
466 .offset = offset,
467 .bytes = length,
469 QTAILQ_INSERT_TAIL(&s->discards, d, next);
471 found:
472 /* Merge discard requests if they are adjacent now */
473 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
474 if (p == d
475 || p->offset > d->offset + d->bytes
476 || d->offset > p->offset + p->bytes)
478 continue;
481 /* Still no overlap possible */
482 assert(p->offset == d->offset + d->bytes
483 || d->offset == p->offset + p->bytes);
485 QTAILQ_REMOVE(&s->discards, p, next);
486 d->offset = MIN(d->offset, p->offset);
487 d->bytes += p->bytes;
491 /* XXX: cache several refcount block clusters ? */
492 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
493 int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
495 BDRVQcowState *s = bs->opaque;
496 int64_t start, last, cluster_offset;
497 uint16_t *refcount_block = NULL;
498 int64_t old_table_index = -1;
499 int ret;
501 #ifdef DEBUG_ALLOC2
502 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
503 offset, length, addend);
504 #endif
505 if (length < 0) {
506 return -EINVAL;
507 } else if (length == 0) {
508 return 0;
511 if (addend < 0) {
512 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
513 s->l2_table_cache);
516 start = offset & ~(s->cluster_size - 1);
517 last = (offset + length - 1) & ~(s->cluster_size - 1);
518 for(cluster_offset = start; cluster_offset <= last;
519 cluster_offset += s->cluster_size)
521 int block_index, refcount;
522 int64_t cluster_index = cluster_offset >> s->cluster_bits;
523 int64_t table_index =
524 cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
526 /* Load the refcount block and allocate it if needed */
527 if (table_index != old_table_index) {
528 if (refcount_block) {
529 ret = qcow2_cache_put(bs, s->refcount_block_cache,
530 (void**) &refcount_block);
531 if (ret < 0) {
532 goto fail;
536 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
537 if (ret < 0) {
538 goto fail;
541 old_table_index = table_index;
543 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
545 /* we can update the count and save it */
546 block_index = cluster_index &
547 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
549 refcount = be16_to_cpu(refcount_block[block_index]);
550 refcount += addend;
551 if (refcount < 0 || refcount > 0xffff) {
552 ret = -EINVAL;
553 goto fail;
555 if (refcount == 0 && cluster_index < s->free_cluster_index) {
556 s->free_cluster_index = cluster_index;
558 refcount_block[block_index] = cpu_to_be16(refcount);
560 if (refcount == 0 && s->discard_passthrough[type]) {
561 update_refcount_discard(bs, cluster_offset, s->cluster_size);
565 ret = 0;
566 fail:
567 if (!s->cache_discards) {
568 qcow2_process_discards(bs, ret);
571 /* Write last changed block to disk */
572 if (refcount_block) {
573 int wret;
574 wret = qcow2_cache_put(bs, s->refcount_block_cache,
575 (void**) &refcount_block);
576 if (wret < 0) {
577 return ret < 0 ? ret : wret;
582 * Try do undo any updates if an error is returned (This may succeed in
583 * some cases like ENOSPC for allocating a new refcount block)
585 if (ret < 0) {
586 int dummy;
587 dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
588 QCOW2_DISCARD_NEVER);
589 (void)dummy;
592 return ret;
596 * Increases or decreases the refcount of a given cluster by one.
597 * addend must be 1 or -1.
599 * If the return value is non-negative, it is the new refcount of the cluster.
600 * If it is negative, it is -errno and indicates an error.
602 static int update_cluster_refcount(BlockDriverState *bs,
603 int64_t cluster_index,
604 int addend,
605 enum qcow2_discard_type type)
607 BDRVQcowState *s = bs->opaque;
608 int ret;
610 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
611 type);
612 if (ret < 0) {
613 return ret;
616 return get_refcount(bs, cluster_index);
621 /*********************************************************/
622 /* cluster allocation functions */
626 /* return < 0 if error */
627 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size)
629 BDRVQcowState *s = bs->opaque;
630 int i, nb_clusters, refcount;
632 nb_clusters = size_to_clusters(s, size);
633 retry:
634 for(i = 0; i < nb_clusters; i++) {
635 int64_t next_cluster_index = s->free_cluster_index++;
636 refcount = get_refcount(bs, next_cluster_index);
638 if (refcount < 0) {
639 return refcount;
640 } else if (refcount != 0) {
641 goto retry;
644 #ifdef DEBUG_ALLOC2
645 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
646 size,
647 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
648 #endif
649 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
652 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size)
654 int64_t offset;
655 int ret;
657 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
658 offset = alloc_clusters_noref(bs, size);
659 if (offset < 0) {
660 return offset;
663 ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
664 if (ret < 0) {
665 return ret;
668 return offset;
671 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
672 int nb_clusters)
674 BDRVQcowState *s = bs->opaque;
675 uint64_t cluster_index;
676 uint64_t old_free_cluster_index;
677 int i, refcount, ret;
679 /* Check how many clusters there are free */
680 cluster_index = offset >> s->cluster_bits;
681 for(i = 0; i < nb_clusters; i++) {
682 refcount = get_refcount(bs, cluster_index++);
684 if (refcount < 0) {
685 return refcount;
686 } else if (refcount != 0) {
687 break;
691 /* And then allocate them */
692 old_free_cluster_index = s->free_cluster_index;
693 s->free_cluster_index = cluster_index + i;
695 ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
696 QCOW2_DISCARD_NEVER);
697 if (ret < 0) {
698 return ret;
701 s->free_cluster_index = old_free_cluster_index;
703 return i;
706 /* only used to allocate compressed sectors. We try to allocate
707 contiguous sectors. size must be <= cluster_size */
708 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
710 BDRVQcowState *s = bs->opaque;
711 int64_t offset, cluster_offset;
712 int free_in_cluster;
714 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
715 assert(size > 0 && size <= s->cluster_size);
716 if (s->free_byte_offset == 0) {
717 offset = qcow2_alloc_clusters(bs, s->cluster_size);
718 if (offset < 0) {
719 return offset;
721 s->free_byte_offset = offset;
723 redo:
724 free_in_cluster = s->cluster_size -
725 (s->free_byte_offset & (s->cluster_size - 1));
726 if (size <= free_in_cluster) {
727 /* enough space in current cluster */
728 offset = s->free_byte_offset;
729 s->free_byte_offset += size;
730 free_in_cluster -= size;
731 if (free_in_cluster == 0)
732 s->free_byte_offset = 0;
733 if ((offset & (s->cluster_size - 1)) != 0)
734 update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
735 QCOW2_DISCARD_NEVER);
736 } else {
737 offset = qcow2_alloc_clusters(bs, s->cluster_size);
738 if (offset < 0) {
739 return offset;
741 cluster_offset = s->free_byte_offset & ~(s->cluster_size - 1);
742 if ((cluster_offset + s->cluster_size) == offset) {
743 /* we are lucky: contiguous data */
744 offset = s->free_byte_offset;
745 update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
746 QCOW2_DISCARD_NEVER);
747 s->free_byte_offset += size;
748 } else {
749 s->free_byte_offset = offset;
750 goto redo;
754 /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
755 * or explicitly by update_cluster_refcount(). Refcount blocks must be
756 * flushed before the caller's L2 table updates.
758 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
759 return offset;
762 void qcow2_free_clusters(BlockDriverState *bs,
763 int64_t offset, int64_t size,
764 enum qcow2_discard_type type)
766 int ret;
768 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
769 ret = update_refcount(bs, offset, size, -1, type);
770 if (ret < 0) {
771 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
772 /* TODO Remember the clusters to free them later and avoid leaking */
777 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
778 * normal cluster, compressed cluster, etc.)
780 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
781 int nb_clusters, enum qcow2_discard_type type)
783 BDRVQcowState *s = bs->opaque;
785 switch (qcow2_get_cluster_type(l2_entry)) {
786 case QCOW2_CLUSTER_COMPRESSED:
788 int nb_csectors;
789 nb_csectors = ((l2_entry >> s->csize_shift) &
790 s->csize_mask) + 1;
791 qcow2_free_clusters(bs,
792 (l2_entry & s->cluster_offset_mask) & ~511,
793 nb_csectors * 512, type);
795 break;
796 case QCOW2_CLUSTER_NORMAL:
797 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
798 nb_clusters << s->cluster_bits, type);
799 break;
800 case QCOW2_CLUSTER_UNALLOCATED:
801 case QCOW2_CLUSTER_ZERO:
802 break;
803 default:
804 abort();
810 /*********************************************************/
811 /* snapshots and image creation */
815 /* update the refcounts of snapshots and the copied flag */
816 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
817 int64_t l1_table_offset, int l1_size, int addend)
819 BDRVQcowState *s = bs->opaque;
820 uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, l1_allocated;
821 int64_t old_offset, old_l2_offset;
822 int i, j, l1_modified = 0, nb_csectors, refcount;
823 int ret;
825 l2_table = NULL;
826 l1_table = NULL;
827 l1_size2 = l1_size * sizeof(uint64_t);
829 s->cache_discards = true;
831 /* WARNING: qcow2_snapshot_goto relies on this function not using the
832 * l1_table_offset when it is the current s->l1_table_offset! Be careful
833 * when changing this! */
834 if (l1_table_offset != s->l1_table_offset) {
835 l1_table = g_malloc0(align_offset(l1_size2, 512));
836 l1_allocated = 1;
838 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
839 if (ret < 0) {
840 goto fail;
843 for(i = 0;i < l1_size; i++)
844 be64_to_cpus(&l1_table[i]);
845 } else {
846 assert(l1_size == s->l1_size);
847 l1_table = s->l1_table;
848 l1_allocated = 0;
851 for(i = 0; i < l1_size; i++) {
852 l2_offset = l1_table[i];
853 if (l2_offset) {
854 old_l2_offset = l2_offset;
855 l2_offset &= L1E_OFFSET_MASK;
857 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
858 (void**) &l2_table);
859 if (ret < 0) {
860 goto fail;
863 for(j = 0; j < s->l2_size; j++) {
864 offset = be64_to_cpu(l2_table[j]);
865 if (offset != 0) {
866 old_offset = offset;
867 offset &= ~QCOW_OFLAG_COPIED;
868 if (offset & QCOW_OFLAG_COMPRESSED) {
869 nb_csectors = ((offset >> s->csize_shift) &
870 s->csize_mask) + 1;
871 if (addend != 0) {
872 int ret;
873 ret = update_refcount(bs,
874 (offset & s->cluster_offset_mask) & ~511,
875 nb_csectors * 512, addend,
876 QCOW2_DISCARD_SNAPSHOT);
877 if (ret < 0) {
878 goto fail;
881 /* compressed clusters are never modified */
882 refcount = 2;
883 } else {
884 uint64_t cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
885 if (addend != 0) {
886 refcount = update_cluster_refcount(bs, cluster_index, addend,
887 QCOW2_DISCARD_SNAPSHOT);
888 } else {
889 refcount = get_refcount(bs, cluster_index);
892 if (refcount < 0) {
893 ret = refcount;
894 goto fail;
898 if (refcount == 1) {
899 offset |= QCOW_OFLAG_COPIED;
901 if (offset != old_offset) {
902 if (addend > 0) {
903 qcow2_cache_set_dependency(bs, s->l2_table_cache,
904 s->refcount_block_cache);
906 l2_table[j] = cpu_to_be64(offset);
907 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
912 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
913 if (ret < 0) {
914 goto fail;
918 if (addend != 0) {
919 refcount = update_cluster_refcount(bs, l2_offset >> s->cluster_bits, addend,
920 QCOW2_DISCARD_SNAPSHOT);
921 } else {
922 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
924 if (refcount < 0) {
925 ret = refcount;
926 goto fail;
927 } else if (refcount == 1) {
928 l2_offset |= QCOW_OFLAG_COPIED;
930 if (l2_offset != old_l2_offset) {
931 l1_table[i] = l2_offset;
932 l1_modified = 1;
937 ret = bdrv_flush(bs);
938 fail:
939 if (l2_table) {
940 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
943 s->cache_discards = false;
944 qcow2_process_discards(bs, ret);
946 /* Update L1 only if it isn't deleted anyway (addend = -1) */
947 if (ret == 0 && addend >= 0 && l1_modified) {
948 for (i = 0; i < l1_size; i++) {
949 cpu_to_be64s(&l1_table[i]);
952 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
954 for (i = 0; i < l1_size; i++) {
955 be64_to_cpus(&l1_table[i]);
958 if (l1_allocated)
959 g_free(l1_table);
960 return ret;
966 /*********************************************************/
967 /* refcount checking functions */
972 * Increases the refcount for a range of clusters in a given refcount table.
973 * This is used to construct a temporary refcount table out of L1 and L2 tables
974 * which can be compared the the refcount table saved in the image.
976 * Modifies the number of errors in res.
978 static void inc_refcounts(BlockDriverState *bs,
979 BdrvCheckResult *res,
980 uint16_t *refcount_table,
981 int refcount_table_size,
982 int64_t offset, int64_t size)
984 BDRVQcowState *s = bs->opaque;
985 int64_t start, last, cluster_offset;
986 int k;
988 if (size <= 0)
989 return;
991 start = offset & ~(s->cluster_size - 1);
992 last = (offset + size - 1) & ~(s->cluster_size - 1);
993 for(cluster_offset = start; cluster_offset <= last;
994 cluster_offset += s->cluster_size) {
995 k = cluster_offset >> s->cluster_bits;
996 if (k < 0) {
997 fprintf(stderr, "ERROR: invalid cluster offset=0x%" PRIx64 "\n",
998 cluster_offset);
999 res->corruptions++;
1000 } else if (k >= refcount_table_size) {
1001 fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
1002 "the end of the image file, can't properly check refcounts.\n",
1003 cluster_offset);
1004 res->check_errors++;
1005 } else {
1006 if (++refcount_table[k] == 0) {
1007 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1008 "\n", cluster_offset);
1009 res->corruptions++;
1015 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1016 enum {
1017 CHECK_OFLAG_COPIED = 0x1, /* check QCOW_OFLAG_COPIED matches refcount */
1018 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1022 * Increases the refcount in the given refcount table for the all clusters
1023 * referenced in the L2 table. While doing so, performs some checks on L2
1024 * entries.
1026 * Returns the number of errors found by the checks or -errno if an internal
1027 * error occurred.
1029 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1030 uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
1031 int flags)
1033 BDRVQcowState *s = bs->opaque;
1034 uint64_t *l2_table, l2_entry;
1035 uint64_t next_contiguous_offset = 0;
1036 int i, l2_size, nb_csectors, refcount;
1038 /* Read L2 table from disk */
1039 l2_size = s->l2_size * sizeof(uint64_t);
1040 l2_table = g_malloc(l2_size);
1042 if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
1043 goto fail;
1045 /* Do the actual checks */
1046 for(i = 0; i < s->l2_size; i++) {
1047 l2_entry = be64_to_cpu(l2_table[i]);
1049 switch (qcow2_get_cluster_type(l2_entry)) {
1050 case QCOW2_CLUSTER_COMPRESSED:
1051 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1052 if (l2_entry & QCOW_OFLAG_COPIED) {
1053 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1054 "copied flag must never be set for compressed "
1055 "clusters\n", l2_entry >> s->cluster_bits);
1056 l2_entry &= ~QCOW_OFLAG_COPIED;
1057 res->corruptions++;
1060 /* Mark cluster as used */
1061 nb_csectors = ((l2_entry >> s->csize_shift) &
1062 s->csize_mask) + 1;
1063 l2_entry &= s->cluster_offset_mask;
1064 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1065 l2_entry & ~511, nb_csectors * 512);
1067 if (flags & CHECK_FRAG_INFO) {
1068 res->bfi.allocated_clusters++;
1069 res->bfi.compressed_clusters++;
1071 /* Compressed clusters are fragmented by nature. Since they
1072 * take up sub-sector space but we only have sector granularity
1073 * I/O we need to re-read the same sectors even for adjacent
1074 * compressed clusters.
1076 res->bfi.fragmented_clusters++;
1078 break;
1080 case QCOW2_CLUSTER_ZERO:
1081 if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1082 break;
1084 /* fall through */
1086 case QCOW2_CLUSTER_NORMAL:
1088 /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1089 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1091 if (flags & CHECK_OFLAG_COPIED) {
1092 refcount = get_refcount(bs, offset >> s->cluster_bits);
1093 if (refcount < 0) {
1094 fprintf(stderr, "Can't get refcount for offset %"
1095 PRIx64 ": %s\n", l2_entry, strerror(-refcount));
1096 goto fail;
1098 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1099 fprintf(stderr, "ERROR OFLAG_COPIED: offset=%"
1100 PRIx64 " refcount=%d\n", l2_entry, refcount);
1101 res->corruptions++;
1105 if (flags & CHECK_FRAG_INFO) {
1106 res->bfi.allocated_clusters++;
1107 if (next_contiguous_offset &&
1108 offset != next_contiguous_offset) {
1109 res->bfi.fragmented_clusters++;
1111 next_contiguous_offset = offset + s->cluster_size;
1114 /* Mark cluster as used */
1115 inc_refcounts(bs, res, refcount_table,refcount_table_size,
1116 offset, s->cluster_size);
1118 /* Correct offsets are cluster aligned */
1119 if (offset & (s->cluster_size - 1)) {
1120 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1121 "properly aligned; L2 entry corrupted.\n", offset);
1122 res->corruptions++;
1124 break;
1127 case QCOW2_CLUSTER_UNALLOCATED:
1128 break;
1130 default:
1131 abort();
1135 g_free(l2_table);
1136 return 0;
1138 fail:
1139 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1140 g_free(l2_table);
1141 return -EIO;
1145 * Increases the refcount for the L1 table, its L2 tables and all referenced
1146 * clusters in the given refcount table. While doing so, performs some checks
1147 * on L1 and L2 entries.
1149 * Returns the number of errors found by the checks or -errno if an internal
1150 * error occurred.
1152 static int check_refcounts_l1(BlockDriverState *bs,
1153 BdrvCheckResult *res,
1154 uint16_t *refcount_table,
1155 int refcount_table_size,
1156 int64_t l1_table_offset, int l1_size,
1157 int flags)
1159 BDRVQcowState *s = bs->opaque;
1160 uint64_t *l1_table, l2_offset, l1_size2;
1161 int i, refcount, ret;
1163 l1_size2 = l1_size * sizeof(uint64_t);
1165 /* Mark L1 table as used */
1166 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1167 l1_table_offset, l1_size2);
1169 /* Read L1 table entries from disk */
1170 if (l1_size2 == 0) {
1171 l1_table = NULL;
1172 } else {
1173 l1_table = g_malloc(l1_size2);
1174 if (bdrv_pread(bs->file, l1_table_offset,
1175 l1_table, l1_size2) != l1_size2)
1176 goto fail;
1177 for(i = 0;i < l1_size; i++)
1178 be64_to_cpus(&l1_table[i]);
1181 /* Do the actual checks */
1182 for(i = 0; i < l1_size; i++) {
1183 l2_offset = l1_table[i];
1184 if (l2_offset) {
1185 /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1186 if (flags & CHECK_OFLAG_COPIED) {
1187 refcount = get_refcount(bs, (l2_offset & ~QCOW_OFLAG_COPIED)
1188 >> s->cluster_bits);
1189 if (refcount < 0) {
1190 fprintf(stderr, "Can't get refcount for l2_offset %"
1191 PRIx64 ": %s\n", l2_offset, strerror(-refcount));
1192 goto fail;
1194 if ((refcount == 1) != ((l2_offset & QCOW_OFLAG_COPIED) != 0)) {
1195 fprintf(stderr, "ERROR OFLAG_COPIED: l2_offset=%" PRIx64
1196 " refcount=%d\n", l2_offset, refcount);
1197 res->corruptions++;
1201 /* Mark L2 table as used */
1202 l2_offset &= L1E_OFFSET_MASK;
1203 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1204 l2_offset, s->cluster_size);
1206 /* L2 tables are cluster aligned */
1207 if (l2_offset & (s->cluster_size - 1)) {
1208 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1209 "cluster aligned; L1 entry corrupted\n", l2_offset);
1210 res->corruptions++;
1213 /* Process and check L2 entries */
1214 ret = check_refcounts_l2(bs, res, refcount_table,
1215 refcount_table_size, l2_offset, flags);
1216 if (ret < 0) {
1217 goto fail;
1221 g_free(l1_table);
1222 return 0;
1224 fail:
1225 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1226 res->check_errors++;
1227 g_free(l1_table);
1228 return -EIO;
1232 * Checks an image for refcount consistency.
1234 * Returns 0 if no errors are found, the number of errors in case the image is
1235 * detected as corrupted, and -errno when an internal error occurred.
1237 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1238 BdrvCheckMode fix)
1240 BDRVQcowState *s = bs->opaque;
1241 int64_t size, i, highest_cluster;
1242 int nb_clusters, refcount1, refcount2;
1243 QCowSnapshot *sn;
1244 uint16_t *refcount_table;
1245 int ret;
1247 size = bdrv_getlength(bs->file);
1248 nb_clusters = size_to_clusters(s, size);
1249 refcount_table = g_malloc0(nb_clusters * sizeof(uint16_t));
1251 res->bfi.total_clusters =
1252 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1254 /* header */
1255 inc_refcounts(bs, res, refcount_table, nb_clusters,
1256 0, s->cluster_size);
1258 /* current L1 table */
1259 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1260 s->l1_table_offset, s->l1_size,
1261 CHECK_OFLAG_COPIED | CHECK_FRAG_INFO);
1262 if (ret < 0) {
1263 goto fail;
1266 /* snapshots */
1267 for(i = 0; i < s->nb_snapshots; i++) {
1268 sn = s->snapshots + i;
1269 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1270 sn->l1_table_offset, sn->l1_size, 0);
1271 if (ret < 0) {
1272 goto fail;
1275 inc_refcounts(bs, res, refcount_table, nb_clusters,
1276 s->snapshots_offset, s->snapshots_size);
1278 /* refcount data */
1279 inc_refcounts(bs, res, refcount_table, nb_clusters,
1280 s->refcount_table_offset,
1281 s->refcount_table_size * sizeof(uint64_t));
1283 for(i = 0; i < s->refcount_table_size; i++) {
1284 uint64_t offset, cluster;
1285 offset = s->refcount_table[i];
1286 cluster = offset >> s->cluster_bits;
1288 /* Refcount blocks are cluster aligned */
1289 if (offset & (s->cluster_size - 1)) {
1290 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1291 "cluster aligned; refcount table entry corrupted\n", i);
1292 res->corruptions++;
1293 continue;
1296 if (cluster >= nb_clusters) {
1297 fprintf(stderr, "ERROR refcount block %" PRId64
1298 " is outside image\n", i);
1299 res->corruptions++;
1300 continue;
1303 if (offset != 0) {
1304 inc_refcounts(bs, res, refcount_table, nb_clusters,
1305 offset, s->cluster_size);
1306 if (refcount_table[cluster] != 1) {
1307 fprintf(stderr, "ERROR refcount block %" PRId64
1308 " refcount=%d\n",
1309 i, refcount_table[cluster]);
1310 res->corruptions++;
1315 /* compare ref counts */
1316 for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1317 refcount1 = get_refcount(bs, i);
1318 if (refcount1 < 0) {
1319 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1320 i, strerror(-refcount1));
1321 res->check_errors++;
1322 continue;
1325 refcount2 = refcount_table[i];
1327 if (refcount1 > 0 || refcount2 > 0) {
1328 highest_cluster = i;
1331 if (refcount1 != refcount2) {
1333 /* Check if we're allowed to fix the mismatch */
1334 int *num_fixed = NULL;
1335 if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1336 num_fixed = &res->leaks_fixed;
1337 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1338 num_fixed = &res->corruptions_fixed;
1341 fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1342 num_fixed != NULL ? "Repairing" :
1343 refcount1 < refcount2 ? "ERROR" :
1344 "Leaked",
1345 i, refcount1, refcount2);
1347 if (num_fixed) {
1348 ret = update_refcount(bs, i << s->cluster_bits, 1,
1349 refcount2 - refcount1,
1350 QCOW2_DISCARD_ALWAYS);
1351 if (ret >= 0) {
1352 (*num_fixed)++;
1353 continue;
1357 /* And if we couldn't, print an error */
1358 if (refcount1 < refcount2) {
1359 res->corruptions++;
1360 } else {
1361 res->leaks++;
1366 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1367 ret = 0;
1369 fail:
1370 g_free(refcount_table);
1372 return ret;