xref: /openbmc/qemu/block/qcow2-refcount.c (revision 69c98726537627e708abb8fcb33e3a2b10e40bf1)
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 
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);
33 
34 
35 /*********************************************************/
36 /* refcount handling */
37 
38 int qcow2_refcount_init(BlockDriverState *bs)
39 {
40     BDRVQcowState *s = bs->opaque;
41     int ret, refcount_table_size2, i;
42 
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]);
53     }
54     return 0;
55  fail:
56     return -ENOMEM;
57 }
58 
59 void qcow2_refcount_close(BlockDriverState *bs)
60 {
61     BDRVQcowState *s = bs->opaque;
62     g_free(s->refcount_table);
63 }
64 
65 
66 static int load_refcount_block(BlockDriverState *bs,
67                                int64_t refcount_block_offset,
68                                void **refcount_block)
69 {
70     BDRVQcowState *s = bs->opaque;
71     int ret;
72 
73     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
74     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
75         refcount_block);
76 
77     return ret;
78 }
79 
80 /*
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.
84  */
85 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
86 {
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;
93 
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;
100 
101     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
102         (void**) &refcount_block);
103     if (ret < 0) {
104         return ret;
105     }
106 
107     block_index = cluster_index &
108         ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
109     refcount = be16_to_cpu(refcount_block[block_index]);
110 
111     ret = qcow2_cache_put(bs, s->refcount_block_cache,
112         (void**) &refcount_block);
113     if (ret < 0) {
114         return ret;
115     }
116 
117     return refcount;
118 }
119 
120 /*
121  * Rounds the refcount table size up to avoid growing the table for each single
122  * refcount block that is allocated.
123  */
124 static unsigned int next_refcount_table_size(BDRVQcowState *s,
125     unsigned int min_size)
126 {
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));
130 
131     while (min_clusters > refcount_table_clusters) {
132         refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
133     }
134 
135     return refcount_table_clusters << (s->cluster_bits - 3);
136 }
137 
138 
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)
142 {
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);
145 
146     return (block_a == block_b);
147 }
148 
149 /*
150  * Loads a refcount block. If it doesn't exist yet, it is allocated first
151  * (including growing the refcount table if needed).
152  *
153  * Returns 0 on success or -errno in error case
154  */
155 static int alloc_refcount_block(BlockDriverState *bs,
156     int64_t cluster_index, uint16_t **refcount_block)
157 {
158     BDRVQcowState *s = bs->opaque;
159     unsigned int refcount_table_index;
160     int ret;
161 
162     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
163 
164     /* Find the refcount block for the given cluster */
165     refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
166 
167     if (refcount_table_index < s->refcount_table_size) {
168 
169         uint64_t refcount_block_offset =
170             s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
171 
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);
176         }
177     }
178 
179     /*
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.
183      *
184      * Note that allocating clusters here needs some special care:
185      *
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.
190      *
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.
196      *
197      * - alloc_clusters_noref and qcow2_free_clusters may load a different
198      *   refcount block into the cache
199      */
200 
201     *refcount_block = NULL;
202 
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;
207     }
208 
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;
213     }
214 
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
220 
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;
227         }
228 
229         memset(*refcount_block, 0, s->cluster_size);
230 
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;
242         }
243 
244         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
245         if (ret < 0) {
246             goto fail_block;
247         }
248 
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;
255         }
256 
257         memset(*refcount_block, 0, s->cluster_size);
258     }
259 
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;
266     }
267 
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;
277         }
278 
279         s->refcount_table[refcount_table_index] = new_block;
280         return 0;
281     }
282 
283     ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
284     if (ret < 0) {
285         goto fail_block;
286     }
287 
288     /*
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.
292      *
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.
298      */
299     BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
300 
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;
305 
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;
317 
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));
322 
323     } while (last_table_size != table_size);
324 
325 #ifdef DEBUG_ALLOC2
326     fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
327         s->refcount_table_size, table_size);
328 #endif
329 
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));
336 
337     assert(meta_offset >= (s->free_cluster_index * s->cluster_size));
338 
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;
343 
344     int i;
345     for (i = 0; i < blocks_clusters; i++) {
346         new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
347     }
348 
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);
354     }
355 
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;
363     }
364 
365     /* Write refcount table to disk */
366     for(i = 0; i < table_size; i++) {
367         cpu_to_be64s(&new_table[i]);
368     }
369 
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;
375     }
376 
377     for(i = 0; i < table_size; i++) {
378         be64_to_cpus(&new_table[i]);
379     }
380 
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;
390     }
391 
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;
395 
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;
400 
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;
406 
407     ret = load_refcount_block(bs, new_block, (void**) refcount_block);
408     if (ret < 0) {
409         return ret;
410     }
411 
412     return 0;
413 
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);
419     }
420     return ret;
421 }
422 
423 void qcow2_process_discards(BlockDriverState *bs, int ret)
424 {
425     BDRVQcowState *s = bs->opaque;
426     Qcow2DiscardRegion *d, *next;
427 
428     QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
429         QTAILQ_REMOVE(&s->discards, d, next);
430 
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);
436         }
437 
438         g_free(d);
439     }
440 }
441 
442 static void update_refcount_discard(BlockDriverState *bs,
443                                     uint64_t offset, uint64_t length)
444 {
445     BDRVQcowState *s = bs->opaque;
446     Qcow2DiscardRegion *d, *p, *next;
447 
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);
451 
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;
460         }
461     }
462 
463     d = g_malloc(sizeof(*d));
464     *d = (Qcow2DiscardRegion) {
465         .bs     = bs,
466         .offset = offset,
467         .bytes  = length,
468     };
469     QTAILQ_INSERT_TAIL(&s->discards, d, next);
470 
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)
477         {
478             continue;
479         }
480 
481         /* Still no overlap possible */
482         assert(p->offset == d->offset + d->bytes
483             || d->offset == p->offset + p->bytes);
484 
485         QTAILQ_REMOVE(&s->discards, p, next);
486         d->offset = MIN(d->offset, p->offset);
487         d->bytes += p->bytes;
488     }
489 }
490 
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)
494 {
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;
500 
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;
509     }
510 
511     if (addend < 0) {
512         qcow2_cache_set_dependency(bs, s->refcount_block_cache,
513             s->l2_table_cache);
514     }
515 
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)
520     {
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);
525 
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;
533                 }
534             }
535 
536             ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
537             if (ret < 0) {
538                 goto fail;
539             }
540         }
541         old_table_index = table_index;
542 
543         qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
544 
545         /* we can update the count and save it */
546         block_index = cluster_index &
547             ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
548 
549         refcount = be16_to_cpu(refcount_block[block_index]);
550         refcount += addend;
551         if (refcount < 0 || refcount > 0xffff) {
552             ret = -EINVAL;
553             goto fail;
554         }
555         if (refcount == 0 && cluster_index < s->free_cluster_index) {
556             s->free_cluster_index = cluster_index;
557         }
558         refcount_block[block_index] = cpu_to_be16(refcount);
559 
560         if (refcount == 0 && s->discard_passthrough[type]) {
561             update_refcount_discard(bs, cluster_offset, s->cluster_size);
562         }
563     }
564 
565     ret = 0;
566 fail:
567     if (!s->cache_discards) {
568         qcow2_process_discards(bs, ret);
569     }
570 
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;
578         }
579     }
580 
581     /*
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)
584      */
585     if (ret < 0) {
586         int dummy;
587         dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
588                                 QCOW2_DISCARD_NEVER);
589         (void)dummy;
590     }
591 
592     return ret;
593 }
594 
595 /*
596  * Increases or decreases the refcount of a given cluster by one.
597  * addend must be 1 or -1.
598  *
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.
601  */
602 static int update_cluster_refcount(BlockDriverState *bs,
603                                    int64_t cluster_index,
604                                    int addend,
605                                    enum qcow2_discard_type type)
606 {
607     BDRVQcowState *s = bs->opaque;
608     int ret;
609 
610     ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
611                           type);
612     if (ret < 0) {
613         return ret;
614     }
615 
616     return get_refcount(bs, cluster_index);
617 }
618 
619 
620 
621 /*********************************************************/
622 /* cluster allocation functions */
623 
624 
625 
626 /* return < 0 if error */
627 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size)
628 {
629     BDRVQcowState *s = bs->opaque;
630     int i, nb_clusters, refcount;
631 
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);
637 
638         if (refcount < 0) {
639             return refcount;
640         } else if (refcount != 0) {
641             goto retry;
642         }
643     }
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;
650 }
651 
652 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size)
653 {
654     int64_t offset;
655     int ret;
656 
657     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
658     offset = alloc_clusters_noref(bs, size);
659     if (offset < 0) {
660         return offset;
661     }
662 
663     ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
664     if (ret < 0) {
665         return ret;
666     }
667 
668     return offset;
669 }
670 
671 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
672     int nb_clusters)
673 {
674     BDRVQcowState *s = bs->opaque;
675     uint64_t cluster_index;
676     uint64_t old_free_cluster_index;
677     int i, refcount, ret;
678 
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++);
683 
684         if (refcount < 0) {
685             return refcount;
686         } else if (refcount != 0) {
687             break;
688         }
689     }
690 
691     /* And then allocate them */
692     old_free_cluster_index = s->free_cluster_index;
693     s->free_cluster_index = cluster_index + i;
694 
695     ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
696                           QCOW2_DISCARD_NEVER);
697     if (ret < 0) {
698         return ret;
699     }
700 
701     s->free_cluster_index = old_free_cluster_index;
702 
703     return i;
704 }
705 
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)
709 {
710     BDRVQcowState *s = bs->opaque;
711     int64_t offset, cluster_offset;
712     int free_in_cluster;
713 
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;
720         }
721         s->free_byte_offset = offset;
722     }
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;
740         }
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;
751         }
752     }
753 
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.
757      */
758     qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
759     return offset;
760 }
761 
762 void qcow2_free_clusters(BlockDriverState *bs,
763                           int64_t offset, int64_t size,
764                           enum qcow2_discard_type type)
765 {
766     int ret;
767 
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 */
773     }
774 }
775 
776 /*
777  * Free a cluster using its L2 entry (handles clusters of all types, e.g.
778  * normal cluster, compressed cluster, etc.)
779  */
780 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
781                              int nb_clusters, enum qcow2_discard_type type)
782 {
783     BDRVQcowState *s = bs->opaque;
784 
785     switch (qcow2_get_cluster_type(l2_entry)) {
786     case QCOW2_CLUSTER_COMPRESSED:
787         {
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);
794         }
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();
805     }
806 }
807 
808 
809 
810 /*********************************************************/
811 /* snapshots and image creation */
812 
813 
814 
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)
818 {
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;
824 
825     l2_table = NULL;
826     l1_table = NULL;
827     l1_size2 = l1_size * sizeof(uint64_t);
828 
829     s->cache_discards = true;
830 
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;
837 
838         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
839         if (ret < 0) {
840             goto fail;
841         }
842 
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;
849     }
850 
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;
856 
857             ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
858                 (void**) &l2_table);
859             if (ret < 0) {
860                 goto fail;
861             }
862 
863             for(j = 0; j < s->l2_size; j++) {
864                 uint64_t cluster_index;
865 
866                 offset = be64_to_cpu(l2_table[j]);
867                 old_offset = offset;
868                 offset &= ~QCOW_OFLAG_COPIED;
869 
870                 switch (qcow2_get_cluster_type(offset)) {
871                     case QCOW2_CLUSTER_COMPRESSED:
872                         nb_csectors = ((offset >> s->csize_shift) &
873                                        s->csize_mask) + 1;
874                         if (addend != 0) {
875                             int ret;
876                             ret = update_refcount(bs,
877                                 (offset & s->cluster_offset_mask) & ~511,
878                                 nb_csectors * 512, addend,
879                                 QCOW2_DISCARD_SNAPSHOT);
880                             if (ret < 0) {
881                                 goto fail;
882                             }
883                         }
884                         /* compressed clusters are never modified */
885                         refcount = 2;
886                         break;
887 
888                     case QCOW2_CLUSTER_NORMAL:
889                     case QCOW2_CLUSTER_ZERO:
890                         cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
891                         if (!cluster_index) {
892                             /* unallocated */
893                             refcount = 0;
894                             break;
895                         }
896                         if (addend != 0) {
897                             refcount = update_cluster_refcount(bs, cluster_index, addend,
898                                                                QCOW2_DISCARD_SNAPSHOT);
899                         } else {
900                             refcount = get_refcount(bs, cluster_index);
901                         }
902 
903                         if (refcount < 0) {
904                             ret = refcount;
905                             goto fail;
906                         }
907                         break;
908 
909                     case QCOW2_CLUSTER_UNALLOCATED:
910                         refcount = 0;
911                         break;
912 
913                     default:
914                         abort();
915                 }
916 
917                 if (refcount == 1) {
918                     offset |= QCOW_OFLAG_COPIED;
919                 }
920                 if (offset != old_offset) {
921                     if (addend > 0) {
922                         qcow2_cache_set_dependency(bs, s->l2_table_cache,
923                             s->refcount_block_cache);
924                     }
925                     l2_table[j] = cpu_to_be64(offset);
926                     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
927                 }
928             }
929 
930             ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
931             if (ret < 0) {
932                 goto fail;
933             }
934 
935 
936             if (addend != 0) {
937                 refcount = update_cluster_refcount(bs, l2_offset >> s->cluster_bits, addend,
938                                                    QCOW2_DISCARD_SNAPSHOT);
939             } else {
940                 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
941             }
942             if (refcount < 0) {
943                 ret = refcount;
944                 goto fail;
945             } else if (refcount == 1) {
946                 l2_offset |= QCOW_OFLAG_COPIED;
947             }
948             if (l2_offset != old_l2_offset) {
949                 l1_table[i] = l2_offset;
950                 l1_modified = 1;
951             }
952         }
953     }
954 
955     ret = bdrv_flush(bs);
956 fail:
957     if (l2_table) {
958         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
959     }
960 
961     s->cache_discards = false;
962     qcow2_process_discards(bs, ret);
963 
964     /* Update L1 only if it isn't deleted anyway (addend = -1) */
965     if (ret == 0 && addend >= 0 && l1_modified) {
966         for (i = 0; i < l1_size; i++) {
967             cpu_to_be64s(&l1_table[i]);
968         }
969 
970         ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
971 
972         for (i = 0; i < l1_size; i++) {
973             be64_to_cpus(&l1_table[i]);
974         }
975     }
976     if (l1_allocated)
977         g_free(l1_table);
978     return ret;
979 }
980 
981 
982 
983 
984 /*********************************************************/
985 /* refcount checking functions */
986 
987 
988 
989 /*
990  * Increases the refcount for a range of clusters in a given refcount table.
991  * This is used to construct a temporary refcount table out of L1 and L2 tables
992  * which can be compared the the refcount table saved in the image.
993  *
994  * Modifies the number of errors in res.
995  */
996 static void inc_refcounts(BlockDriverState *bs,
997                           BdrvCheckResult *res,
998                           uint16_t *refcount_table,
999                           int refcount_table_size,
1000                           int64_t offset, int64_t size)
1001 {
1002     BDRVQcowState *s = bs->opaque;
1003     int64_t start, last, cluster_offset;
1004     int k;
1005 
1006     if (size <= 0)
1007         return;
1008 
1009     start = offset & ~(s->cluster_size - 1);
1010     last = (offset + size - 1) & ~(s->cluster_size - 1);
1011     for(cluster_offset = start; cluster_offset <= last;
1012         cluster_offset += s->cluster_size) {
1013         k = cluster_offset >> s->cluster_bits;
1014         if (k < 0) {
1015             fprintf(stderr, "ERROR: invalid cluster offset=0x%" PRIx64 "\n",
1016                 cluster_offset);
1017             res->corruptions++;
1018         } else if (k >= refcount_table_size) {
1019             fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
1020                 "the end of the image file, can't properly check refcounts.\n",
1021                 cluster_offset);
1022             res->check_errors++;
1023         } else {
1024             if (++refcount_table[k] == 0) {
1025                 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1026                     "\n", cluster_offset);
1027                 res->corruptions++;
1028             }
1029         }
1030     }
1031 }
1032 
1033 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1034 enum {
1035     CHECK_OFLAG_COPIED = 0x1,   /* check QCOW_OFLAG_COPIED matches refcount */
1036     CHECK_FRAG_INFO = 0x2,      /* update BlockFragInfo counters */
1037 };
1038 
1039 /*
1040  * Increases the refcount in the given refcount table for the all clusters
1041  * referenced in the L2 table. While doing so, performs some checks on L2
1042  * entries.
1043  *
1044  * Returns the number of errors found by the checks or -errno if an internal
1045  * error occurred.
1046  */
1047 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1048     uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
1049     int flags)
1050 {
1051     BDRVQcowState *s = bs->opaque;
1052     uint64_t *l2_table, l2_entry;
1053     uint64_t next_contiguous_offset = 0;
1054     int i, l2_size, nb_csectors, refcount;
1055 
1056     /* Read L2 table from disk */
1057     l2_size = s->l2_size * sizeof(uint64_t);
1058     l2_table = g_malloc(l2_size);
1059 
1060     if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
1061         goto fail;
1062 
1063     /* Do the actual checks */
1064     for(i = 0; i < s->l2_size; i++) {
1065         l2_entry = be64_to_cpu(l2_table[i]);
1066 
1067         switch (qcow2_get_cluster_type(l2_entry)) {
1068         case QCOW2_CLUSTER_COMPRESSED:
1069             /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1070             if (l2_entry & QCOW_OFLAG_COPIED) {
1071                 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1072                     "copied flag must never be set for compressed "
1073                     "clusters\n", l2_entry >> s->cluster_bits);
1074                 l2_entry &= ~QCOW_OFLAG_COPIED;
1075                 res->corruptions++;
1076             }
1077 
1078             /* Mark cluster as used */
1079             nb_csectors = ((l2_entry >> s->csize_shift) &
1080                            s->csize_mask) + 1;
1081             l2_entry &= s->cluster_offset_mask;
1082             inc_refcounts(bs, res, refcount_table, refcount_table_size,
1083                 l2_entry & ~511, nb_csectors * 512);
1084 
1085             if (flags & CHECK_FRAG_INFO) {
1086                 res->bfi.allocated_clusters++;
1087                 res->bfi.compressed_clusters++;
1088 
1089                 /* Compressed clusters are fragmented by nature.  Since they
1090                  * take up sub-sector space but we only have sector granularity
1091                  * I/O we need to re-read the same sectors even for adjacent
1092                  * compressed clusters.
1093                  */
1094                 res->bfi.fragmented_clusters++;
1095             }
1096             break;
1097 
1098         case QCOW2_CLUSTER_ZERO:
1099             if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1100                 break;
1101             }
1102             /* fall through */
1103 
1104         case QCOW2_CLUSTER_NORMAL:
1105         {
1106             /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1107             uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1108 
1109             if (flags & CHECK_OFLAG_COPIED) {
1110                 refcount = get_refcount(bs, offset >> s->cluster_bits);
1111                 if (refcount < 0) {
1112                     fprintf(stderr, "Can't get refcount for offset %"
1113                         PRIx64 ": %s\n", l2_entry, strerror(-refcount));
1114                     goto fail;
1115                 }
1116                 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1117                     fprintf(stderr, "ERROR OFLAG_COPIED: offset=%"
1118                         PRIx64 " refcount=%d\n", l2_entry, refcount);
1119                     res->corruptions++;
1120                 }
1121             }
1122 
1123             if (flags & CHECK_FRAG_INFO) {
1124                 res->bfi.allocated_clusters++;
1125                 if (next_contiguous_offset &&
1126                     offset != next_contiguous_offset) {
1127                     res->bfi.fragmented_clusters++;
1128                 }
1129                 next_contiguous_offset = offset + s->cluster_size;
1130             }
1131 
1132             /* Mark cluster as used */
1133             inc_refcounts(bs, res, refcount_table,refcount_table_size,
1134                 offset, s->cluster_size);
1135 
1136             /* Correct offsets are cluster aligned */
1137             if (offset & (s->cluster_size - 1)) {
1138                 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1139                     "properly aligned; L2 entry corrupted.\n", offset);
1140                 res->corruptions++;
1141             }
1142             break;
1143         }
1144 
1145         case QCOW2_CLUSTER_UNALLOCATED:
1146             break;
1147 
1148         default:
1149             abort();
1150         }
1151     }
1152 
1153     g_free(l2_table);
1154     return 0;
1155 
1156 fail:
1157     fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1158     g_free(l2_table);
1159     return -EIO;
1160 }
1161 
1162 /*
1163  * Increases the refcount for the L1 table, its L2 tables and all referenced
1164  * clusters in the given refcount table. While doing so, performs some checks
1165  * on L1 and L2 entries.
1166  *
1167  * Returns the number of errors found by the checks or -errno if an internal
1168  * error occurred.
1169  */
1170 static int check_refcounts_l1(BlockDriverState *bs,
1171                               BdrvCheckResult *res,
1172                               uint16_t *refcount_table,
1173                               int refcount_table_size,
1174                               int64_t l1_table_offset, int l1_size,
1175                               int flags)
1176 {
1177     BDRVQcowState *s = bs->opaque;
1178     uint64_t *l1_table, l2_offset, l1_size2;
1179     int i, refcount, ret;
1180 
1181     l1_size2 = l1_size * sizeof(uint64_t);
1182 
1183     /* Mark L1 table as used */
1184     inc_refcounts(bs, res, refcount_table, refcount_table_size,
1185         l1_table_offset, l1_size2);
1186 
1187     /* Read L1 table entries from disk */
1188     if (l1_size2 == 0) {
1189         l1_table = NULL;
1190     } else {
1191         l1_table = g_malloc(l1_size2);
1192         if (bdrv_pread(bs->file, l1_table_offset,
1193                        l1_table, l1_size2) != l1_size2)
1194             goto fail;
1195         for(i = 0;i < l1_size; i++)
1196             be64_to_cpus(&l1_table[i]);
1197     }
1198 
1199     /* Do the actual checks */
1200     for(i = 0; i < l1_size; i++) {
1201         l2_offset = l1_table[i];
1202         if (l2_offset) {
1203             /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1204             if (flags & CHECK_OFLAG_COPIED) {
1205                 refcount = get_refcount(bs, (l2_offset & ~QCOW_OFLAG_COPIED)
1206                     >> s->cluster_bits);
1207                 if (refcount < 0) {
1208                     fprintf(stderr, "Can't get refcount for l2_offset %"
1209                         PRIx64 ": %s\n", l2_offset, strerror(-refcount));
1210                     goto fail;
1211                 }
1212                 if ((refcount == 1) != ((l2_offset & QCOW_OFLAG_COPIED) != 0)) {
1213                     fprintf(stderr, "ERROR OFLAG_COPIED: l2_offset=%" PRIx64
1214                         " refcount=%d\n", l2_offset, refcount);
1215                     res->corruptions++;
1216                 }
1217             }
1218 
1219             /* Mark L2 table as used */
1220             l2_offset &= L1E_OFFSET_MASK;
1221             inc_refcounts(bs, res, refcount_table, refcount_table_size,
1222                 l2_offset, s->cluster_size);
1223 
1224             /* L2 tables are cluster aligned */
1225             if (l2_offset & (s->cluster_size - 1)) {
1226                 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1227                     "cluster aligned; L1 entry corrupted\n", l2_offset);
1228                 res->corruptions++;
1229             }
1230 
1231             /* Process and check L2 entries */
1232             ret = check_refcounts_l2(bs, res, refcount_table,
1233                                      refcount_table_size, l2_offset, flags);
1234             if (ret < 0) {
1235                 goto fail;
1236             }
1237         }
1238     }
1239     g_free(l1_table);
1240     return 0;
1241 
1242 fail:
1243     fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1244     res->check_errors++;
1245     g_free(l1_table);
1246     return -EIO;
1247 }
1248 
1249 /*
1250  * Checks an image for refcount consistency.
1251  *
1252  * Returns 0 if no errors are found, the number of errors in case the image is
1253  * detected as corrupted, and -errno when an internal error occurred.
1254  */
1255 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1256                           BdrvCheckMode fix)
1257 {
1258     BDRVQcowState *s = bs->opaque;
1259     int64_t size, i, highest_cluster;
1260     int nb_clusters, refcount1, refcount2;
1261     QCowSnapshot *sn;
1262     uint16_t *refcount_table;
1263     int ret;
1264 
1265     size = bdrv_getlength(bs->file);
1266     nb_clusters = size_to_clusters(s, size);
1267     refcount_table = g_malloc0(nb_clusters * sizeof(uint16_t));
1268 
1269     res->bfi.total_clusters =
1270         size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1271 
1272     /* header */
1273     inc_refcounts(bs, res, refcount_table, nb_clusters,
1274         0, s->cluster_size);
1275 
1276     /* current L1 table */
1277     ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1278                              s->l1_table_offset, s->l1_size,
1279                              CHECK_OFLAG_COPIED | CHECK_FRAG_INFO);
1280     if (ret < 0) {
1281         goto fail;
1282     }
1283 
1284     /* snapshots */
1285     for(i = 0; i < s->nb_snapshots; i++) {
1286         sn = s->snapshots + i;
1287         ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1288             sn->l1_table_offset, sn->l1_size, 0);
1289         if (ret < 0) {
1290             goto fail;
1291         }
1292     }
1293     inc_refcounts(bs, res, refcount_table, nb_clusters,
1294         s->snapshots_offset, s->snapshots_size);
1295 
1296     /* refcount data */
1297     inc_refcounts(bs, res, refcount_table, nb_clusters,
1298         s->refcount_table_offset,
1299         s->refcount_table_size * sizeof(uint64_t));
1300 
1301     for(i = 0; i < s->refcount_table_size; i++) {
1302         uint64_t offset, cluster;
1303         offset = s->refcount_table[i];
1304         cluster = offset >> s->cluster_bits;
1305 
1306         /* Refcount blocks are cluster aligned */
1307         if (offset & (s->cluster_size - 1)) {
1308             fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1309                 "cluster aligned; refcount table entry corrupted\n", i);
1310             res->corruptions++;
1311             continue;
1312         }
1313 
1314         if (cluster >= nb_clusters) {
1315             fprintf(stderr, "ERROR refcount block %" PRId64
1316                     " is outside image\n", i);
1317             res->corruptions++;
1318             continue;
1319         }
1320 
1321         if (offset != 0) {
1322             inc_refcounts(bs, res, refcount_table, nb_clusters,
1323                 offset, s->cluster_size);
1324             if (refcount_table[cluster] != 1) {
1325                 fprintf(stderr, "ERROR refcount block %" PRId64
1326                     " refcount=%d\n",
1327                     i, refcount_table[cluster]);
1328                 res->corruptions++;
1329             }
1330         }
1331     }
1332 
1333     /* compare ref counts */
1334     for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1335         refcount1 = get_refcount(bs, i);
1336         if (refcount1 < 0) {
1337             fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1338                 i, strerror(-refcount1));
1339             res->check_errors++;
1340             continue;
1341         }
1342 
1343         refcount2 = refcount_table[i];
1344 
1345         if (refcount1 > 0 || refcount2 > 0) {
1346             highest_cluster = i;
1347         }
1348 
1349         if (refcount1 != refcount2) {
1350 
1351             /* Check if we're allowed to fix the mismatch */
1352             int *num_fixed = NULL;
1353             if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1354                 num_fixed = &res->leaks_fixed;
1355             } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1356                 num_fixed = &res->corruptions_fixed;
1357             }
1358 
1359             fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1360                    num_fixed != NULL     ? "Repairing" :
1361                    refcount1 < refcount2 ? "ERROR" :
1362                                            "Leaked",
1363                    i, refcount1, refcount2);
1364 
1365             if (num_fixed) {
1366                 ret = update_refcount(bs, i << s->cluster_bits, 1,
1367                                       refcount2 - refcount1,
1368                                       QCOW2_DISCARD_ALWAYS);
1369                 if (ret >= 0) {
1370                     (*num_fixed)++;
1371                     continue;
1372                 }
1373             }
1374 
1375             /* And if we couldn't, print an error */
1376             if (refcount1 < refcount2) {
1377                 res->corruptions++;
1378             } else {
1379                 res->leaks++;
1380             }
1381         }
1382     }
1383 
1384     res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1385     ret = 0;
1386 
1387 fail:
1388     g_free(refcount_table);
1389 
1390     return ret;
1391 }
1392 
1393