xref: /openbmc/qemu/block/qcow2-refcount.c (revision 228aa992)
1 /*
2  * Block driver for the QCOW version 2 format
3  *
4  * Copyright (c) 2004-2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu-common.h"
26 #include "block/block_int.h"
27 #include "block/qcow2.h"
28 #include "qemu/range.h"
29 
30 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size);
31 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
32                             int64_t offset, int64_t length,
33                             int addend, enum qcow2_discard_type type);
34 
35 
36 /*********************************************************/
37 /* refcount handling */
38 
39 int qcow2_refcount_init(BlockDriverState *bs)
40 {
41     BDRVQcowState *s = bs->opaque;
42     unsigned int refcount_table_size2, i;
43     int ret;
44 
45     assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
46     refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
47     s->refcount_table = g_try_malloc(refcount_table_size2);
48 
49     if (s->refcount_table_size > 0) {
50         if (s->refcount_table == NULL) {
51             ret = -ENOMEM;
52             goto fail;
53         }
54         BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
55         ret = bdrv_pread(bs->file, s->refcount_table_offset,
56                          s->refcount_table, refcount_table_size2);
57         if (ret < 0) {
58             goto fail;
59         }
60         for(i = 0; i < s->refcount_table_size; i++)
61             be64_to_cpus(&s->refcount_table[i]);
62     }
63     return 0;
64  fail:
65     return ret;
66 }
67 
68 void qcow2_refcount_close(BlockDriverState *bs)
69 {
70     BDRVQcowState *s = bs->opaque;
71     g_free(s->refcount_table);
72 }
73 
74 
75 static int load_refcount_block(BlockDriverState *bs,
76                                int64_t refcount_block_offset,
77                                void **refcount_block)
78 {
79     BDRVQcowState *s = bs->opaque;
80     int ret;
81 
82     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
83     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
84         refcount_block);
85 
86     return ret;
87 }
88 
89 /*
90  * Returns the refcount of the cluster given by its index. Any non-negative
91  * return value is the refcount of the cluster, negative values are -errno
92  * and indicate an error.
93  */
94 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index)
95 {
96     BDRVQcowState *s = bs->opaque;
97     uint64_t refcount_table_index, block_index;
98     int64_t refcount_block_offset;
99     int ret;
100     uint16_t *refcount_block;
101     uint16_t refcount;
102 
103     refcount_table_index = cluster_index >> s->refcount_block_bits;
104     if (refcount_table_index >= s->refcount_table_size)
105         return 0;
106     refcount_block_offset =
107         s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
108     if (!refcount_block_offset)
109         return 0;
110 
111     if (offset_into_cluster(s, refcount_block_offset)) {
112         qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64
113                                 " unaligned (reftable index: %#" PRIx64 ")",
114                                 refcount_block_offset, refcount_table_index);
115         return -EIO;
116     }
117 
118     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
119         (void**) &refcount_block);
120     if (ret < 0) {
121         return ret;
122     }
123 
124     block_index = cluster_index & (s->refcount_block_size - 1);
125     refcount = be16_to_cpu(refcount_block[block_index]);
126 
127     ret = qcow2_cache_put(bs, s->refcount_block_cache,
128         (void**) &refcount_block);
129     if (ret < 0) {
130         return ret;
131     }
132 
133     return refcount;
134 }
135 
136 /*
137  * Rounds the refcount table size up to avoid growing the table for each single
138  * refcount block that is allocated.
139  */
140 static unsigned int next_refcount_table_size(BDRVQcowState *s,
141     unsigned int min_size)
142 {
143     unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
144     unsigned int refcount_table_clusters =
145         MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
146 
147     while (min_clusters > refcount_table_clusters) {
148         refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
149     }
150 
151     return refcount_table_clusters << (s->cluster_bits - 3);
152 }
153 
154 
155 /* Checks if two offsets are described by the same refcount block */
156 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
157     uint64_t offset_b)
158 {
159     uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits);
160     uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits);
161 
162     return (block_a == block_b);
163 }
164 
165 /*
166  * Loads a refcount block. If it doesn't exist yet, it is allocated first
167  * (including growing the refcount table if needed).
168  *
169  * Returns 0 on success or -errno in error case
170  */
171 static int alloc_refcount_block(BlockDriverState *bs,
172     int64_t cluster_index, uint16_t **refcount_block)
173 {
174     BDRVQcowState *s = bs->opaque;
175     unsigned int refcount_table_index;
176     int ret;
177 
178     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
179 
180     /* Find the refcount block for the given cluster */
181     refcount_table_index = cluster_index >> s->refcount_block_bits;
182 
183     if (refcount_table_index < s->refcount_table_size) {
184 
185         uint64_t refcount_block_offset =
186             s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
187 
188         /* If it's already there, we're done */
189         if (refcount_block_offset) {
190             if (offset_into_cluster(s, refcount_block_offset)) {
191                 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
192                                         PRIx64 " unaligned (reftable index: "
193                                         "%#x)", refcount_block_offset,
194                                         refcount_table_index);
195                 return -EIO;
196             }
197 
198              return load_refcount_block(bs, refcount_block_offset,
199                  (void**) refcount_block);
200         }
201     }
202 
203     /*
204      * If we came here, we need to allocate something. Something is at least
205      * a cluster for the new refcount block. It may also include a new refcount
206      * table if the old refcount table is too small.
207      *
208      * Note that allocating clusters here needs some special care:
209      *
210      * - We can't use the normal qcow2_alloc_clusters(), it would try to
211      *   increase the refcount and very likely we would end up with an endless
212      *   recursion. Instead we must place the refcount blocks in a way that
213      *   they can describe them themselves.
214      *
215      * - We need to consider that at this point we are inside update_refcounts
216      *   and potentially doing an initial refcount increase. This means that
217      *   some clusters have already been allocated by the caller, but their
218      *   refcount isn't accurate yet. If we allocate clusters for metadata, we
219      *   need to return -EAGAIN to signal the caller that it needs to restart
220      *   the search for free clusters.
221      *
222      * - alloc_clusters_noref and qcow2_free_clusters may load a different
223      *   refcount block into the cache
224      */
225 
226     *refcount_block = NULL;
227 
228     /* We write to the refcount table, so we might depend on L2 tables */
229     ret = qcow2_cache_flush(bs, s->l2_table_cache);
230     if (ret < 0) {
231         return ret;
232     }
233 
234     /* Allocate the refcount block itself and mark it as used */
235     int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
236     if (new_block < 0) {
237         return new_block;
238     }
239 
240 #ifdef DEBUG_ALLOC2
241     fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
242         " at %" PRIx64 "\n",
243         refcount_table_index, cluster_index << s->cluster_bits, new_block);
244 #endif
245 
246     if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
247         /* Zero the new refcount block before updating it */
248         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
249             (void**) refcount_block);
250         if (ret < 0) {
251             goto fail_block;
252         }
253 
254         memset(*refcount_block, 0, s->cluster_size);
255 
256         /* The block describes itself, need to update the cache */
257         int block_index = (new_block >> s->cluster_bits) &
258             (s->refcount_block_size - 1);
259         (*refcount_block)[block_index] = cpu_to_be16(1);
260     } else {
261         /* Described somewhere else. This can recurse at most twice before we
262          * arrive at a block that describes itself. */
263         ret = update_refcount(bs, new_block, s->cluster_size, 1,
264                               QCOW2_DISCARD_NEVER);
265         if (ret < 0) {
266             goto fail_block;
267         }
268 
269         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
270         if (ret < 0) {
271             goto fail_block;
272         }
273 
274         /* Initialize the new refcount block only after updating its refcount,
275          * update_refcount uses the refcount cache itself */
276         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
277             (void**) refcount_block);
278         if (ret < 0) {
279             goto fail_block;
280         }
281 
282         memset(*refcount_block, 0, s->cluster_size);
283     }
284 
285     /* Now the new refcount block needs to be written to disk */
286     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
287     qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
288     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
289     if (ret < 0) {
290         goto fail_block;
291     }
292 
293     /* If the refcount table is big enough, just hook the block up there */
294     if (refcount_table_index < s->refcount_table_size) {
295         uint64_t data64 = cpu_to_be64(new_block);
296         BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
297         ret = bdrv_pwrite_sync(bs->file,
298             s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
299             &data64, sizeof(data64));
300         if (ret < 0) {
301             goto fail_block;
302         }
303 
304         s->refcount_table[refcount_table_index] = new_block;
305 
306         /* The new refcount block may be where the caller intended to put its
307          * data, so let it restart the search. */
308         return -EAGAIN;
309     }
310 
311     ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
312     if (ret < 0) {
313         goto fail_block;
314     }
315 
316     /*
317      * If we come here, we need to grow the refcount table. Again, a new
318      * refcount table needs some space and we can't simply allocate to avoid
319      * endless recursion.
320      *
321      * Therefore let's grab new refcount blocks at the end of the image, which
322      * will describe themselves and the new refcount table. This way we can
323      * reference them only in the new table and do the switch to the new
324      * refcount table at once without producing an inconsistent state in
325      * between.
326      */
327     BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
328 
329     /* Calculate the number of refcount blocks needed so far */
330     uint64_t blocks_used = DIV_ROUND_UP(cluster_index, s->refcount_block_size);
331 
332     if (blocks_used > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
333         return -EFBIG;
334     }
335 
336     /* And now we need at least one block more for the new metadata */
337     uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
338     uint64_t last_table_size;
339     uint64_t blocks_clusters;
340     do {
341         uint64_t table_clusters =
342             size_to_clusters(s, table_size * sizeof(uint64_t));
343         blocks_clusters = 1 +
344             ((table_clusters + s->refcount_block_size - 1)
345             / s->refcount_block_size);
346         uint64_t meta_clusters = table_clusters + blocks_clusters;
347 
348         last_table_size = table_size;
349         table_size = next_refcount_table_size(s, blocks_used +
350             ((meta_clusters + s->refcount_block_size - 1)
351             / s->refcount_block_size));
352 
353     } while (last_table_size != table_size);
354 
355 #ifdef DEBUG_ALLOC2
356     fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
357         s->refcount_table_size, table_size);
358 #endif
359 
360     /* Create the new refcount table and blocks */
361     uint64_t meta_offset = (blocks_used * s->refcount_block_size) *
362         s->cluster_size;
363     uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
364     uint64_t *new_table = g_try_new0(uint64_t, table_size);
365     uint16_t *new_blocks = g_try_malloc0(blocks_clusters * s->cluster_size);
366 
367     assert(table_size > 0 && blocks_clusters > 0);
368     if (new_table == NULL || new_blocks == NULL) {
369         ret = -ENOMEM;
370         goto fail_table;
371     }
372 
373     /* Fill the new refcount table */
374     memcpy(new_table, s->refcount_table,
375         s->refcount_table_size * sizeof(uint64_t));
376     new_table[refcount_table_index] = new_block;
377 
378     int i;
379     for (i = 0; i < blocks_clusters; i++) {
380         new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
381     }
382 
383     /* Fill the refcount blocks */
384     uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
385     int block = 0;
386     for (i = 0; i < table_clusters + blocks_clusters; i++) {
387         new_blocks[block++] = cpu_to_be16(1);
388     }
389 
390     /* Write refcount blocks to disk */
391     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
392     ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
393         blocks_clusters * s->cluster_size);
394     g_free(new_blocks);
395     new_blocks = NULL;
396     if (ret < 0) {
397         goto fail_table;
398     }
399 
400     /* Write refcount table to disk */
401     for(i = 0; i < table_size; i++) {
402         cpu_to_be64s(&new_table[i]);
403     }
404 
405     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
406     ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
407         table_size * sizeof(uint64_t));
408     if (ret < 0) {
409         goto fail_table;
410     }
411 
412     for(i = 0; i < table_size; i++) {
413         be64_to_cpus(&new_table[i]);
414     }
415 
416     /* Hook up the new refcount table in the qcow2 header */
417     uint8_t data[12];
418     cpu_to_be64w((uint64_t*)data, table_offset);
419     cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
420     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
421     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
422         data, sizeof(data));
423     if (ret < 0) {
424         goto fail_table;
425     }
426 
427     /* And switch it in memory */
428     uint64_t old_table_offset = s->refcount_table_offset;
429     uint64_t old_table_size = s->refcount_table_size;
430 
431     g_free(s->refcount_table);
432     s->refcount_table = new_table;
433     s->refcount_table_size = table_size;
434     s->refcount_table_offset = table_offset;
435 
436     /* Free old table. */
437     qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
438                         QCOW2_DISCARD_OTHER);
439 
440     ret = load_refcount_block(bs, new_block, (void**) refcount_block);
441     if (ret < 0) {
442         return ret;
443     }
444 
445     /* If we were trying to do the initial refcount update for some cluster
446      * allocation, we might have used the same clusters to store newly
447      * allocated metadata. Make the caller search some new space. */
448     return -EAGAIN;
449 
450 fail_table:
451     g_free(new_blocks);
452     g_free(new_table);
453 fail_block:
454     if (*refcount_block != NULL) {
455         qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
456     }
457     return ret;
458 }
459 
460 void qcow2_process_discards(BlockDriverState *bs, int ret)
461 {
462     BDRVQcowState *s = bs->opaque;
463     Qcow2DiscardRegion *d, *next;
464 
465     QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
466         QTAILQ_REMOVE(&s->discards, d, next);
467 
468         /* Discard is optional, ignore the return value */
469         if (ret >= 0) {
470             bdrv_discard(bs->file,
471                          d->offset >> BDRV_SECTOR_BITS,
472                          d->bytes >> BDRV_SECTOR_BITS);
473         }
474 
475         g_free(d);
476     }
477 }
478 
479 static void update_refcount_discard(BlockDriverState *bs,
480                                     uint64_t offset, uint64_t length)
481 {
482     BDRVQcowState *s = bs->opaque;
483     Qcow2DiscardRegion *d, *p, *next;
484 
485     QTAILQ_FOREACH(d, &s->discards, next) {
486         uint64_t new_start = MIN(offset, d->offset);
487         uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
488 
489         if (new_end - new_start <= length + d->bytes) {
490             /* There can't be any overlap, areas ending up here have no
491              * references any more and therefore shouldn't get freed another
492              * time. */
493             assert(d->bytes + length == new_end - new_start);
494             d->offset = new_start;
495             d->bytes = new_end - new_start;
496             goto found;
497         }
498     }
499 
500     d = g_malloc(sizeof(*d));
501     *d = (Qcow2DiscardRegion) {
502         .bs     = bs,
503         .offset = offset,
504         .bytes  = length,
505     };
506     QTAILQ_INSERT_TAIL(&s->discards, d, next);
507 
508 found:
509     /* Merge discard requests if they are adjacent now */
510     QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
511         if (p == d
512             || p->offset > d->offset + d->bytes
513             || d->offset > p->offset + p->bytes)
514         {
515             continue;
516         }
517 
518         /* Still no overlap possible */
519         assert(p->offset == d->offset + d->bytes
520             || d->offset == p->offset + p->bytes);
521 
522         QTAILQ_REMOVE(&s->discards, p, next);
523         d->offset = MIN(d->offset, p->offset);
524         d->bytes += p->bytes;
525         g_free(p);
526     }
527 }
528 
529 /* XXX: cache several refcount block clusters ? */
530 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
531     int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
532 {
533     BDRVQcowState *s = bs->opaque;
534     int64_t start, last, cluster_offset;
535     uint16_t *refcount_block = NULL;
536     int64_t old_table_index = -1;
537     int ret;
538 
539 #ifdef DEBUG_ALLOC2
540     fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
541            offset, length, addend);
542 #endif
543     if (length < 0) {
544         return -EINVAL;
545     } else if (length == 0) {
546         return 0;
547     }
548 
549     if (addend < 0) {
550         qcow2_cache_set_dependency(bs, s->refcount_block_cache,
551             s->l2_table_cache);
552     }
553 
554     start = start_of_cluster(s, offset);
555     last = start_of_cluster(s, offset + length - 1);
556     for(cluster_offset = start; cluster_offset <= last;
557         cluster_offset += s->cluster_size)
558     {
559         int block_index, refcount;
560         int64_t cluster_index = cluster_offset >> s->cluster_bits;
561         int64_t table_index = cluster_index >> s->refcount_block_bits;
562 
563         /* Load the refcount block and allocate it if needed */
564         if (table_index != old_table_index) {
565             if (refcount_block) {
566                 ret = qcow2_cache_put(bs, s->refcount_block_cache,
567                     (void**) &refcount_block);
568                 if (ret < 0) {
569                     goto fail;
570                 }
571             }
572 
573             ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
574             if (ret < 0) {
575                 goto fail;
576             }
577         }
578         old_table_index = table_index;
579 
580         qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
581 
582         /* we can update the count and save it */
583         block_index = cluster_index & (s->refcount_block_size - 1);
584 
585         refcount = be16_to_cpu(refcount_block[block_index]);
586         refcount += addend;
587         if (refcount < 0 || refcount > 0xffff) {
588             ret = -EINVAL;
589             goto fail;
590         }
591         if (refcount == 0 && cluster_index < s->free_cluster_index) {
592             s->free_cluster_index = cluster_index;
593         }
594         refcount_block[block_index] = cpu_to_be16(refcount);
595 
596         if (refcount == 0 && s->discard_passthrough[type]) {
597             update_refcount_discard(bs, cluster_offset, s->cluster_size);
598         }
599     }
600 
601     ret = 0;
602 fail:
603     if (!s->cache_discards) {
604         qcow2_process_discards(bs, ret);
605     }
606 
607     /* Write last changed block to disk */
608     if (refcount_block) {
609         int wret;
610         wret = qcow2_cache_put(bs, s->refcount_block_cache,
611             (void**) &refcount_block);
612         if (wret < 0) {
613             return ret < 0 ? ret : wret;
614         }
615     }
616 
617     /*
618      * Try do undo any updates if an error is returned (This may succeed in
619      * some cases like ENOSPC for allocating a new refcount block)
620      */
621     if (ret < 0) {
622         int dummy;
623         dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
624                                 QCOW2_DISCARD_NEVER);
625         (void)dummy;
626     }
627 
628     return ret;
629 }
630 
631 /*
632  * Increases or decreases the refcount of a given cluster.
633  *
634  * If the return value is non-negative, it is the new refcount of the cluster.
635  * If it is negative, it is -errno and indicates an error.
636  */
637 int qcow2_update_cluster_refcount(BlockDriverState *bs,
638                                   int64_t cluster_index,
639                                   int addend,
640                                   enum qcow2_discard_type type)
641 {
642     BDRVQcowState *s = bs->opaque;
643     int ret;
644 
645     ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
646                           type);
647     if (ret < 0) {
648         return ret;
649     }
650 
651     return qcow2_get_refcount(bs, cluster_index);
652 }
653 
654 
655 
656 /*********************************************************/
657 /* cluster allocation functions */
658 
659 
660 
661 /* return < 0 if error */
662 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size)
663 {
664     BDRVQcowState *s = bs->opaque;
665     uint64_t i, nb_clusters;
666     int refcount;
667 
668     nb_clusters = size_to_clusters(s, size);
669 retry:
670     for(i = 0; i < nb_clusters; i++) {
671         uint64_t next_cluster_index = s->free_cluster_index++;
672         refcount = qcow2_get_refcount(bs, next_cluster_index);
673 
674         if (refcount < 0) {
675             return refcount;
676         } else if (refcount != 0) {
677             goto retry;
678         }
679     }
680 
681     /* Make sure that all offsets in the "allocated" range are representable
682      * in an int64_t */
683     if (s->free_cluster_index > 0 &&
684         s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits))
685     {
686         return -EFBIG;
687     }
688 
689 #ifdef DEBUG_ALLOC2
690     fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
691             size,
692             (s->free_cluster_index - nb_clusters) << s->cluster_bits);
693 #endif
694     return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
695 }
696 
697 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
698 {
699     int64_t offset;
700     int ret;
701 
702     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
703     do {
704         offset = alloc_clusters_noref(bs, size);
705         if (offset < 0) {
706             return offset;
707         }
708 
709         ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
710     } while (ret == -EAGAIN);
711 
712     if (ret < 0) {
713         return ret;
714     }
715 
716     return offset;
717 }
718 
719 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
720     int nb_clusters)
721 {
722     BDRVQcowState *s = bs->opaque;
723     uint64_t cluster_index;
724     uint64_t i;
725     int refcount, ret;
726 
727     assert(nb_clusters >= 0);
728     if (nb_clusters == 0) {
729         return 0;
730     }
731 
732     do {
733         /* Check how many clusters there are free */
734         cluster_index = offset >> s->cluster_bits;
735         for(i = 0; i < nb_clusters; i++) {
736             refcount = qcow2_get_refcount(bs, cluster_index++);
737 
738             if (refcount < 0) {
739                 return refcount;
740             } else if (refcount != 0) {
741                 break;
742             }
743         }
744 
745         /* And then allocate them */
746         ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
747                               QCOW2_DISCARD_NEVER);
748     } while (ret == -EAGAIN);
749 
750     if (ret < 0) {
751         return ret;
752     }
753 
754     return i;
755 }
756 
757 /* only used to allocate compressed sectors. We try to allocate
758    contiguous sectors. size must be <= cluster_size */
759 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
760 {
761     BDRVQcowState *s = bs->opaque;
762     int64_t offset, cluster_offset;
763     int free_in_cluster;
764 
765     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
766     assert(size > 0 && size <= s->cluster_size);
767     if (s->free_byte_offset == 0) {
768         offset = qcow2_alloc_clusters(bs, s->cluster_size);
769         if (offset < 0) {
770             return offset;
771         }
772         s->free_byte_offset = offset;
773     }
774  redo:
775     free_in_cluster = s->cluster_size -
776         offset_into_cluster(s, s->free_byte_offset);
777     if (size <= free_in_cluster) {
778         /* enough space in current cluster */
779         offset = s->free_byte_offset;
780         s->free_byte_offset += size;
781         free_in_cluster -= size;
782         if (free_in_cluster == 0)
783             s->free_byte_offset = 0;
784         if (offset_into_cluster(s, offset) != 0)
785             qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
786                                           QCOW2_DISCARD_NEVER);
787     } else {
788         offset = qcow2_alloc_clusters(bs, s->cluster_size);
789         if (offset < 0) {
790             return offset;
791         }
792         cluster_offset = start_of_cluster(s, s->free_byte_offset);
793         if ((cluster_offset + s->cluster_size) == offset) {
794             /* we are lucky: contiguous data */
795             offset = s->free_byte_offset;
796             qcow2_update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
797                                           QCOW2_DISCARD_NEVER);
798             s->free_byte_offset += size;
799         } else {
800             s->free_byte_offset = offset;
801             goto redo;
802         }
803     }
804 
805     /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
806      * or explicitly by qcow2_update_cluster_refcount().  Refcount blocks must
807      * be flushed before the caller's L2 table updates.
808      */
809     qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
810     return offset;
811 }
812 
813 void qcow2_free_clusters(BlockDriverState *bs,
814                           int64_t offset, int64_t size,
815                           enum qcow2_discard_type type)
816 {
817     int ret;
818 
819     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
820     ret = update_refcount(bs, offset, size, -1, type);
821     if (ret < 0) {
822         fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
823         /* TODO Remember the clusters to free them later and avoid leaking */
824     }
825 }
826 
827 /*
828  * Free a cluster using its L2 entry (handles clusters of all types, e.g.
829  * normal cluster, compressed cluster, etc.)
830  */
831 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
832                              int nb_clusters, enum qcow2_discard_type type)
833 {
834     BDRVQcowState *s = bs->opaque;
835 
836     switch (qcow2_get_cluster_type(l2_entry)) {
837     case QCOW2_CLUSTER_COMPRESSED:
838         {
839             int nb_csectors;
840             nb_csectors = ((l2_entry >> s->csize_shift) &
841                            s->csize_mask) + 1;
842             qcow2_free_clusters(bs,
843                 (l2_entry & s->cluster_offset_mask) & ~511,
844                 nb_csectors * 512, type);
845         }
846         break;
847     case QCOW2_CLUSTER_NORMAL:
848     case QCOW2_CLUSTER_ZERO:
849         if (l2_entry & L2E_OFFSET_MASK) {
850             if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) {
851                 qcow2_signal_corruption(bs, false, -1, -1,
852                                         "Cannot free unaligned cluster %#llx",
853                                         l2_entry & L2E_OFFSET_MASK);
854             } else {
855                 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
856                                     nb_clusters << s->cluster_bits, type);
857             }
858         }
859         break;
860     case QCOW2_CLUSTER_UNALLOCATED:
861         break;
862     default:
863         abort();
864     }
865 }
866 
867 
868 
869 /*********************************************************/
870 /* snapshots and image creation */
871 
872 
873 
874 /* update the refcounts of snapshots and the copied flag */
875 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
876     int64_t l1_table_offset, int l1_size, int addend)
877 {
878     BDRVQcowState *s = bs->opaque;
879     uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2;
880     bool l1_allocated = false;
881     int64_t old_offset, old_l2_offset;
882     int i, j, l1_modified = 0, nb_csectors, refcount;
883     int ret;
884 
885     l2_table = NULL;
886     l1_table = NULL;
887     l1_size2 = l1_size * sizeof(uint64_t);
888 
889     s->cache_discards = true;
890 
891     /* WARNING: qcow2_snapshot_goto relies on this function not using the
892      * l1_table_offset when it is the current s->l1_table_offset! Be careful
893      * when changing this! */
894     if (l1_table_offset != s->l1_table_offset) {
895         l1_table = g_try_malloc0(align_offset(l1_size2, 512));
896         if (l1_size2 && l1_table == NULL) {
897             ret = -ENOMEM;
898             goto fail;
899         }
900         l1_allocated = true;
901 
902         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
903         if (ret < 0) {
904             goto fail;
905         }
906 
907         for(i = 0;i < l1_size; i++)
908             be64_to_cpus(&l1_table[i]);
909     } else {
910         assert(l1_size == s->l1_size);
911         l1_table = s->l1_table;
912         l1_allocated = false;
913     }
914 
915     for(i = 0; i < l1_size; i++) {
916         l2_offset = l1_table[i];
917         if (l2_offset) {
918             old_l2_offset = l2_offset;
919             l2_offset &= L1E_OFFSET_MASK;
920 
921             if (offset_into_cluster(s, l2_offset)) {
922                 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
923                                         PRIx64 " unaligned (L1 index: %#x)",
924                                         l2_offset, i);
925                 ret = -EIO;
926                 goto fail;
927             }
928 
929             ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
930                 (void**) &l2_table);
931             if (ret < 0) {
932                 goto fail;
933             }
934 
935             for(j = 0; j < s->l2_size; j++) {
936                 uint64_t cluster_index;
937 
938                 offset = be64_to_cpu(l2_table[j]);
939                 old_offset = offset;
940                 offset &= ~QCOW_OFLAG_COPIED;
941 
942                 switch (qcow2_get_cluster_type(offset)) {
943                     case QCOW2_CLUSTER_COMPRESSED:
944                         nb_csectors = ((offset >> s->csize_shift) &
945                                        s->csize_mask) + 1;
946                         if (addend != 0) {
947                             ret = update_refcount(bs,
948                                 (offset & s->cluster_offset_mask) & ~511,
949                                 nb_csectors * 512, addend,
950                                 QCOW2_DISCARD_SNAPSHOT);
951                             if (ret < 0) {
952                                 goto fail;
953                             }
954                         }
955                         /* compressed clusters are never modified */
956                         refcount = 2;
957                         break;
958 
959                     case QCOW2_CLUSTER_NORMAL:
960                     case QCOW2_CLUSTER_ZERO:
961                         if (offset_into_cluster(s, offset & L2E_OFFSET_MASK)) {
962                             qcow2_signal_corruption(bs, true, -1, -1, "Data "
963                                                     "cluster offset %#llx "
964                                                     "unaligned (L2 offset: %#"
965                                                     PRIx64 ", L2 index: %#x)",
966                                                     offset & L2E_OFFSET_MASK,
967                                                     l2_offset, j);
968                             ret = -EIO;
969                             goto fail;
970                         }
971 
972                         cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
973                         if (!cluster_index) {
974                             /* unallocated */
975                             refcount = 0;
976                             break;
977                         }
978                         if (addend != 0) {
979                             refcount = qcow2_update_cluster_refcount(bs,
980                                     cluster_index, addend,
981                                     QCOW2_DISCARD_SNAPSHOT);
982                         } else {
983                             refcount = qcow2_get_refcount(bs, cluster_index);
984                         }
985 
986                         if (refcount < 0) {
987                             ret = refcount;
988                             goto fail;
989                         }
990                         break;
991 
992                     case QCOW2_CLUSTER_UNALLOCATED:
993                         refcount = 0;
994                         break;
995 
996                     default:
997                         abort();
998                 }
999 
1000                 if (refcount == 1) {
1001                     offset |= QCOW_OFLAG_COPIED;
1002                 }
1003                 if (offset != old_offset) {
1004                     if (addend > 0) {
1005                         qcow2_cache_set_dependency(bs, s->l2_table_cache,
1006                             s->refcount_block_cache);
1007                     }
1008                     l2_table[j] = cpu_to_be64(offset);
1009                     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1010                 }
1011             }
1012 
1013             ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1014             if (ret < 0) {
1015                 goto fail;
1016             }
1017 
1018 
1019             if (addend != 0) {
1020                 refcount = qcow2_update_cluster_refcount(bs, l2_offset >>
1021                         s->cluster_bits, addend, QCOW2_DISCARD_SNAPSHOT);
1022             } else {
1023                 refcount = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits);
1024             }
1025             if (refcount < 0) {
1026                 ret = refcount;
1027                 goto fail;
1028             } else if (refcount == 1) {
1029                 l2_offset |= QCOW_OFLAG_COPIED;
1030             }
1031             if (l2_offset != old_l2_offset) {
1032                 l1_table[i] = l2_offset;
1033                 l1_modified = 1;
1034             }
1035         }
1036     }
1037 
1038     ret = bdrv_flush(bs);
1039 fail:
1040     if (l2_table) {
1041         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1042     }
1043 
1044     s->cache_discards = false;
1045     qcow2_process_discards(bs, ret);
1046 
1047     /* Update L1 only if it isn't deleted anyway (addend = -1) */
1048     if (ret == 0 && addend >= 0 && l1_modified) {
1049         for (i = 0; i < l1_size; i++) {
1050             cpu_to_be64s(&l1_table[i]);
1051         }
1052 
1053         ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
1054 
1055         for (i = 0; i < l1_size; i++) {
1056             be64_to_cpus(&l1_table[i]);
1057         }
1058     }
1059     if (l1_allocated)
1060         g_free(l1_table);
1061     return ret;
1062 }
1063 
1064 
1065 
1066 
1067 /*********************************************************/
1068 /* refcount checking functions */
1069 
1070 
1071 
1072 /*
1073  * Increases the refcount for a range of clusters in a given refcount table.
1074  * This is used to construct a temporary refcount table out of L1 and L2 tables
1075  * which can be compared the the refcount table saved in the image.
1076  *
1077  * Modifies the number of errors in res.
1078  */
1079 static int inc_refcounts(BlockDriverState *bs,
1080                          BdrvCheckResult *res,
1081                          uint16_t **refcount_table,
1082                          int64_t *refcount_table_size,
1083                          int64_t offset, int64_t size)
1084 {
1085     BDRVQcowState *s = bs->opaque;
1086     uint64_t start, last, cluster_offset, k;
1087 
1088     if (size <= 0) {
1089         return 0;
1090     }
1091 
1092     start = start_of_cluster(s, offset);
1093     last = start_of_cluster(s, offset + size - 1);
1094     for(cluster_offset = start; cluster_offset <= last;
1095         cluster_offset += s->cluster_size) {
1096         k = cluster_offset >> s->cluster_bits;
1097         if (k >= *refcount_table_size) {
1098             int64_t old_refcount_table_size = *refcount_table_size;
1099             uint16_t *new_refcount_table;
1100 
1101             *refcount_table_size = k + 1;
1102             new_refcount_table = g_try_realloc(*refcount_table,
1103                                                *refcount_table_size *
1104                                                sizeof(**refcount_table));
1105             if (!new_refcount_table) {
1106                 *refcount_table_size = old_refcount_table_size;
1107                 res->check_errors++;
1108                 return -ENOMEM;
1109             }
1110             *refcount_table = new_refcount_table;
1111 
1112             memset(*refcount_table + old_refcount_table_size, 0,
1113                    (*refcount_table_size - old_refcount_table_size) *
1114                    sizeof(**refcount_table));
1115         }
1116 
1117         if (++(*refcount_table)[k] == 0) {
1118             fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1119                     "\n", cluster_offset);
1120             res->corruptions++;
1121         }
1122     }
1123 
1124     return 0;
1125 }
1126 
1127 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1128 enum {
1129     CHECK_FRAG_INFO = 0x2,      /* update BlockFragInfo counters */
1130 };
1131 
1132 /*
1133  * Increases the refcount in the given refcount table for the all clusters
1134  * referenced in the L2 table. While doing so, performs some checks on L2
1135  * entries.
1136  *
1137  * Returns the number of errors found by the checks or -errno if an internal
1138  * error occurred.
1139  */
1140 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1141     uint16_t **refcount_table, int64_t *refcount_table_size, int64_t l2_offset,
1142     int flags)
1143 {
1144     BDRVQcowState *s = bs->opaque;
1145     uint64_t *l2_table, l2_entry;
1146     uint64_t next_contiguous_offset = 0;
1147     int i, l2_size, nb_csectors, ret;
1148 
1149     /* Read L2 table from disk */
1150     l2_size = s->l2_size * sizeof(uint64_t);
1151     l2_table = g_malloc(l2_size);
1152 
1153     ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size);
1154     if (ret < 0) {
1155         fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1156         res->check_errors++;
1157         goto fail;
1158     }
1159 
1160     /* Do the actual checks */
1161     for(i = 0; i < s->l2_size; i++) {
1162         l2_entry = be64_to_cpu(l2_table[i]);
1163 
1164         switch (qcow2_get_cluster_type(l2_entry)) {
1165         case QCOW2_CLUSTER_COMPRESSED:
1166             /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1167             if (l2_entry & QCOW_OFLAG_COPIED) {
1168                 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1169                     "copied flag must never be set for compressed "
1170                     "clusters\n", l2_entry >> s->cluster_bits);
1171                 l2_entry &= ~QCOW_OFLAG_COPIED;
1172                 res->corruptions++;
1173             }
1174 
1175             /* Mark cluster as used */
1176             nb_csectors = ((l2_entry >> s->csize_shift) &
1177                            s->csize_mask) + 1;
1178             l2_entry &= s->cluster_offset_mask;
1179             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1180                                 l2_entry & ~511, nb_csectors * 512);
1181             if (ret < 0) {
1182                 goto fail;
1183             }
1184 
1185             if (flags & CHECK_FRAG_INFO) {
1186                 res->bfi.allocated_clusters++;
1187                 res->bfi.compressed_clusters++;
1188 
1189                 /* Compressed clusters are fragmented by nature.  Since they
1190                  * take up sub-sector space but we only have sector granularity
1191                  * I/O we need to re-read the same sectors even for adjacent
1192                  * compressed clusters.
1193                  */
1194                 res->bfi.fragmented_clusters++;
1195             }
1196             break;
1197 
1198         case QCOW2_CLUSTER_ZERO:
1199             if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1200                 break;
1201             }
1202             /* fall through */
1203 
1204         case QCOW2_CLUSTER_NORMAL:
1205         {
1206             uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1207 
1208             if (flags & CHECK_FRAG_INFO) {
1209                 res->bfi.allocated_clusters++;
1210                 if (next_contiguous_offset &&
1211                     offset != next_contiguous_offset) {
1212                     res->bfi.fragmented_clusters++;
1213                 }
1214                 next_contiguous_offset = offset + s->cluster_size;
1215             }
1216 
1217             /* Mark cluster as used */
1218             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1219                                 offset, s->cluster_size);
1220             if (ret < 0) {
1221                 goto fail;
1222             }
1223 
1224             /* Correct offsets are cluster aligned */
1225             if (offset_into_cluster(s, offset)) {
1226                 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1227                     "properly aligned; L2 entry corrupted.\n", offset);
1228                 res->corruptions++;
1229             }
1230             break;
1231         }
1232 
1233         case QCOW2_CLUSTER_UNALLOCATED:
1234             break;
1235 
1236         default:
1237             abort();
1238         }
1239     }
1240 
1241     g_free(l2_table);
1242     return 0;
1243 
1244 fail:
1245     g_free(l2_table);
1246     return ret;
1247 }
1248 
1249 /*
1250  * Increases the refcount for the L1 table, its L2 tables and all referenced
1251  * clusters in the given refcount table. While doing so, performs some checks
1252  * on L1 and L2 entries.
1253  *
1254  * Returns the number of errors found by the checks or -errno if an internal
1255  * error occurred.
1256  */
1257 static int check_refcounts_l1(BlockDriverState *bs,
1258                               BdrvCheckResult *res,
1259                               uint16_t **refcount_table,
1260                               int64_t *refcount_table_size,
1261                               int64_t l1_table_offset, int l1_size,
1262                               int flags)
1263 {
1264     BDRVQcowState *s = bs->opaque;
1265     uint64_t *l1_table = NULL, l2_offset, l1_size2;
1266     int i, ret;
1267 
1268     l1_size2 = l1_size * sizeof(uint64_t);
1269 
1270     /* Mark L1 table as used */
1271     ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1272                         l1_table_offset, l1_size2);
1273     if (ret < 0) {
1274         goto fail;
1275     }
1276 
1277     /* Read L1 table entries from disk */
1278     if (l1_size2 > 0) {
1279         l1_table = g_try_malloc(l1_size2);
1280         if (l1_table == NULL) {
1281             ret = -ENOMEM;
1282             res->check_errors++;
1283             goto fail;
1284         }
1285         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1286         if (ret < 0) {
1287             fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1288             res->check_errors++;
1289             goto fail;
1290         }
1291         for(i = 0;i < l1_size; i++)
1292             be64_to_cpus(&l1_table[i]);
1293     }
1294 
1295     /* Do the actual checks */
1296     for(i = 0; i < l1_size; i++) {
1297         l2_offset = l1_table[i];
1298         if (l2_offset) {
1299             /* Mark L2 table as used */
1300             l2_offset &= L1E_OFFSET_MASK;
1301             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1302                                 l2_offset, s->cluster_size);
1303             if (ret < 0) {
1304                 goto fail;
1305             }
1306 
1307             /* L2 tables are cluster aligned */
1308             if (offset_into_cluster(s, l2_offset)) {
1309                 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1310                     "cluster aligned; L1 entry corrupted\n", l2_offset);
1311                 res->corruptions++;
1312             }
1313 
1314             /* Process and check L2 entries */
1315             ret = check_refcounts_l2(bs, res, refcount_table,
1316                                      refcount_table_size, l2_offset, flags);
1317             if (ret < 0) {
1318                 goto fail;
1319             }
1320         }
1321     }
1322     g_free(l1_table);
1323     return 0;
1324 
1325 fail:
1326     g_free(l1_table);
1327     return ret;
1328 }
1329 
1330 /*
1331  * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1332  *
1333  * This function does not print an error message nor does it increment
1334  * check_errors if qcow2_get_refcount fails (this is because such an error will
1335  * have been already detected and sufficiently signaled by the calling function
1336  * (qcow2_check_refcounts) by the time this function is called).
1337  */
1338 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1339                               BdrvCheckMode fix)
1340 {
1341     BDRVQcowState *s = bs->opaque;
1342     uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1343     int ret;
1344     int refcount;
1345     int i, j;
1346 
1347     for (i = 0; i < s->l1_size; i++) {
1348         uint64_t l1_entry = s->l1_table[i];
1349         uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1350         bool l2_dirty = false;
1351 
1352         if (!l2_offset) {
1353             continue;
1354         }
1355 
1356         refcount = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits);
1357         if (refcount < 0) {
1358             /* don't print message nor increment check_errors */
1359             continue;
1360         }
1361         if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1362             fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1363                     "l1_entry=%" PRIx64 " refcount=%d\n",
1364                     fix & BDRV_FIX_ERRORS ? "Repairing" :
1365                                             "ERROR",
1366                     i, l1_entry, refcount);
1367             if (fix & BDRV_FIX_ERRORS) {
1368                 s->l1_table[i] = refcount == 1
1369                                ? l1_entry |  QCOW_OFLAG_COPIED
1370                                : l1_entry & ~QCOW_OFLAG_COPIED;
1371                 ret = qcow2_write_l1_entry(bs, i);
1372                 if (ret < 0) {
1373                     res->check_errors++;
1374                     goto fail;
1375                 }
1376                 res->corruptions_fixed++;
1377             } else {
1378                 res->corruptions++;
1379             }
1380         }
1381 
1382         ret = bdrv_pread(bs->file, l2_offset, l2_table,
1383                          s->l2_size * sizeof(uint64_t));
1384         if (ret < 0) {
1385             fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1386                     strerror(-ret));
1387             res->check_errors++;
1388             goto fail;
1389         }
1390 
1391         for (j = 0; j < s->l2_size; j++) {
1392             uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1393             uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1394             int cluster_type = qcow2_get_cluster_type(l2_entry);
1395 
1396             if ((cluster_type == QCOW2_CLUSTER_NORMAL) ||
1397                 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) {
1398                 refcount = qcow2_get_refcount(bs,
1399                                               data_offset >> s->cluster_bits);
1400                 if (refcount < 0) {
1401                     /* don't print message nor increment check_errors */
1402                     continue;
1403                 }
1404                 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1405                     fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1406                             "l2_entry=%" PRIx64 " refcount=%d\n",
1407                             fix & BDRV_FIX_ERRORS ? "Repairing" :
1408                                                     "ERROR",
1409                             l2_entry, refcount);
1410                     if (fix & BDRV_FIX_ERRORS) {
1411                         l2_table[j] = cpu_to_be64(refcount == 1
1412                                     ? l2_entry |  QCOW_OFLAG_COPIED
1413                                     : l2_entry & ~QCOW_OFLAG_COPIED);
1414                         l2_dirty = true;
1415                         res->corruptions_fixed++;
1416                     } else {
1417                         res->corruptions++;
1418                     }
1419                 }
1420             }
1421         }
1422 
1423         if (l2_dirty) {
1424             ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1425                                                 l2_offset, s->cluster_size);
1426             if (ret < 0) {
1427                 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1428                         "overlap check failed: %s\n", strerror(-ret));
1429                 res->check_errors++;
1430                 goto fail;
1431             }
1432 
1433             ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size);
1434             if (ret < 0) {
1435                 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1436                         strerror(-ret));
1437                 res->check_errors++;
1438                 goto fail;
1439             }
1440         }
1441     }
1442 
1443     ret = 0;
1444 
1445 fail:
1446     qemu_vfree(l2_table);
1447     return ret;
1448 }
1449 
1450 /*
1451  * Checks consistency of refblocks and accounts for each refblock in
1452  * *refcount_table.
1453  */
1454 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res,
1455                            BdrvCheckMode fix, bool *rebuild,
1456                            uint16_t **refcount_table, int64_t *nb_clusters)
1457 {
1458     BDRVQcowState *s = bs->opaque;
1459     int64_t i, size;
1460     int ret;
1461 
1462     for(i = 0; i < s->refcount_table_size; i++) {
1463         uint64_t offset, cluster;
1464         offset = s->refcount_table[i];
1465         cluster = offset >> s->cluster_bits;
1466 
1467         /* Refcount blocks are cluster aligned */
1468         if (offset_into_cluster(s, offset)) {
1469             fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1470                 "cluster aligned; refcount table entry corrupted\n", i);
1471             res->corruptions++;
1472             *rebuild = true;
1473             continue;
1474         }
1475 
1476         if (cluster >= *nb_clusters) {
1477             fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n",
1478                     fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i);
1479 
1480             if (fix & BDRV_FIX_ERRORS) {
1481                 int64_t old_nb_clusters = *nb_clusters;
1482                 uint16_t *new_refcount_table;
1483 
1484                 if (offset > INT64_MAX - s->cluster_size) {
1485                     ret = -EINVAL;
1486                     goto resize_fail;
1487                 }
1488 
1489                 ret = bdrv_truncate(bs->file, offset + s->cluster_size);
1490                 if (ret < 0) {
1491                     goto resize_fail;
1492                 }
1493                 size = bdrv_getlength(bs->file);
1494                 if (size < 0) {
1495                     ret = size;
1496                     goto resize_fail;
1497                 }
1498 
1499                 *nb_clusters = size_to_clusters(s, size);
1500                 assert(*nb_clusters >= old_nb_clusters);
1501 
1502                 new_refcount_table = g_try_realloc(*refcount_table,
1503                                                    *nb_clusters *
1504                                                    sizeof(**refcount_table));
1505                 if (!new_refcount_table) {
1506                     *nb_clusters = old_nb_clusters;
1507                     res->check_errors++;
1508                     return -ENOMEM;
1509                 }
1510                 *refcount_table = new_refcount_table;
1511 
1512                 memset(*refcount_table + old_nb_clusters, 0,
1513                        (*nb_clusters - old_nb_clusters) *
1514                        sizeof(**refcount_table));
1515 
1516                 if (cluster >= *nb_clusters) {
1517                     ret = -EINVAL;
1518                     goto resize_fail;
1519                 }
1520 
1521                 res->corruptions_fixed++;
1522                 ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1523                                     offset, s->cluster_size);
1524                 if (ret < 0) {
1525                     return ret;
1526                 }
1527                 /* No need to check whether the refcount is now greater than 1:
1528                  * This area was just allocated and zeroed, so it can only be
1529                  * exactly 1 after inc_refcounts() */
1530                 continue;
1531 
1532 resize_fail:
1533                 res->corruptions++;
1534                 *rebuild = true;
1535                 fprintf(stderr, "ERROR could not resize image: %s\n",
1536                         strerror(-ret));
1537             } else {
1538                 res->corruptions++;
1539             }
1540             continue;
1541         }
1542 
1543         if (offset != 0) {
1544             ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1545                                 offset, s->cluster_size);
1546             if (ret < 0) {
1547                 return ret;
1548             }
1549             if ((*refcount_table)[cluster] != 1) {
1550                 fprintf(stderr, "ERROR refcount block %" PRId64
1551                         " refcount=%d\n", i, (*refcount_table)[cluster]);
1552                 res->corruptions++;
1553                 *rebuild = true;
1554             }
1555         }
1556     }
1557 
1558     return 0;
1559 }
1560 
1561 /*
1562  * Calculates an in-memory refcount table.
1563  */
1564 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1565                                BdrvCheckMode fix, bool *rebuild,
1566                                uint16_t **refcount_table, int64_t *nb_clusters)
1567 {
1568     BDRVQcowState *s = bs->opaque;
1569     int64_t i;
1570     QCowSnapshot *sn;
1571     int ret;
1572 
1573     if (!*refcount_table) {
1574         *refcount_table = g_try_new0(uint16_t, *nb_clusters);
1575         if (*nb_clusters && *refcount_table == NULL) {
1576             res->check_errors++;
1577             return -ENOMEM;
1578         }
1579     }
1580 
1581     /* header */
1582     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1583                         0, s->cluster_size);
1584     if (ret < 0) {
1585         return ret;
1586     }
1587 
1588     /* current L1 table */
1589     ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1590                              s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO);
1591     if (ret < 0) {
1592         return ret;
1593     }
1594 
1595     /* snapshots */
1596     for (i = 0; i < s->nb_snapshots; i++) {
1597         sn = s->snapshots + i;
1598         ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1599                                  sn->l1_table_offset, sn->l1_size, 0);
1600         if (ret < 0) {
1601             return ret;
1602         }
1603     }
1604     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1605                         s->snapshots_offset, s->snapshots_size);
1606     if (ret < 0) {
1607         return ret;
1608     }
1609 
1610     /* refcount data */
1611     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1612                         s->refcount_table_offset,
1613                         s->refcount_table_size * sizeof(uint64_t));
1614     if (ret < 0) {
1615         return ret;
1616     }
1617 
1618     return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters);
1619 }
1620 
1621 /*
1622  * Compares the actual reference count for each cluster in the image against the
1623  * refcount as reported by the refcount structures on-disk.
1624  */
1625 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1626                               BdrvCheckMode fix, bool *rebuild,
1627                               int64_t *highest_cluster,
1628                               uint16_t *refcount_table, int64_t nb_clusters)
1629 {
1630     BDRVQcowState *s = bs->opaque;
1631     int64_t i;
1632     int refcount1, refcount2, ret;
1633 
1634     for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) {
1635         refcount1 = qcow2_get_refcount(bs, i);
1636         if (refcount1 < 0) {
1637             fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1638                 i, strerror(-refcount1));
1639             res->check_errors++;
1640             continue;
1641         }
1642 
1643         refcount2 = refcount_table[i];
1644 
1645         if (refcount1 > 0 || refcount2 > 0) {
1646             *highest_cluster = i;
1647         }
1648 
1649         if (refcount1 != refcount2) {
1650             /* Check if we're allowed to fix the mismatch */
1651             int *num_fixed = NULL;
1652             if (refcount1 == 0) {
1653                 *rebuild = true;
1654             } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1655                 num_fixed = &res->leaks_fixed;
1656             } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1657                 num_fixed = &res->corruptions_fixed;
1658             }
1659 
1660             fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1661                    num_fixed != NULL     ? "Repairing" :
1662                    refcount1 < refcount2 ? "ERROR" :
1663                                            "Leaked",
1664                    i, refcount1, refcount2);
1665 
1666             if (num_fixed) {
1667                 ret = update_refcount(bs, i << s->cluster_bits, 1,
1668                                       refcount2 - refcount1,
1669                                       QCOW2_DISCARD_ALWAYS);
1670                 if (ret >= 0) {
1671                     (*num_fixed)++;
1672                     continue;
1673                 }
1674             }
1675 
1676             /* And if we couldn't, print an error */
1677             if (refcount1 < refcount2) {
1678                 res->corruptions++;
1679             } else {
1680                 res->leaks++;
1681             }
1682         }
1683     }
1684 }
1685 
1686 /*
1687  * Allocates clusters using an in-memory refcount table (IMRT) in contrast to
1688  * the on-disk refcount structures.
1689  *
1690  * On input, *first_free_cluster tells where to start looking, and need not
1691  * actually be a free cluster; the returned offset will not be before that
1692  * cluster.  On output, *first_free_cluster points to the first gap found, even
1693  * if that gap was too small to be used as the returned offset.
1694  *
1695  * Note that *first_free_cluster is a cluster index whereas the return value is
1696  * an offset.
1697  */
1698 static int64_t alloc_clusters_imrt(BlockDriverState *bs,
1699                                    int cluster_count,
1700                                    uint16_t **refcount_table,
1701                                    int64_t *imrt_nb_clusters,
1702                                    int64_t *first_free_cluster)
1703 {
1704     BDRVQcowState *s = bs->opaque;
1705     int64_t cluster = *first_free_cluster, i;
1706     bool first_gap = true;
1707     int contiguous_free_clusters;
1708 
1709     /* Starting at *first_free_cluster, find a range of at least cluster_count
1710      * continuously free clusters */
1711     for (contiguous_free_clusters = 0;
1712          cluster < *imrt_nb_clusters &&
1713          contiguous_free_clusters < cluster_count;
1714          cluster++)
1715     {
1716         if (!(*refcount_table)[cluster]) {
1717             contiguous_free_clusters++;
1718             if (first_gap) {
1719                 /* If this is the first free cluster found, update
1720                  * *first_free_cluster accordingly */
1721                 *first_free_cluster = cluster;
1722                 first_gap = false;
1723             }
1724         } else if (contiguous_free_clusters) {
1725             contiguous_free_clusters = 0;
1726         }
1727     }
1728 
1729     /* If contiguous_free_clusters is greater than zero, it contains the number
1730      * of continuously free clusters until the current cluster; the first free
1731      * cluster in the current "gap" is therefore
1732      * cluster - contiguous_free_clusters */
1733 
1734     /* If no such range could be found, grow the in-memory refcount table
1735      * accordingly to append free clusters at the end of the image */
1736     if (contiguous_free_clusters < cluster_count) {
1737         int64_t old_imrt_nb_clusters = *imrt_nb_clusters;
1738         uint16_t *new_refcount_table;
1739 
1740         /* contiguous_free_clusters clusters are already empty at the image end;
1741          * we need cluster_count clusters; therefore, we have to allocate
1742          * cluster_count - contiguous_free_clusters new clusters at the end of
1743          * the image (which is the current value of cluster; note that cluster
1744          * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond
1745          * the image end) */
1746         *imrt_nb_clusters = cluster + cluster_count - contiguous_free_clusters;
1747         new_refcount_table = g_try_realloc(*refcount_table,
1748                                            *imrt_nb_clusters *
1749                                            sizeof(**refcount_table));
1750         if (!new_refcount_table) {
1751             *imrt_nb_clusters = old_imrt_nb_clusters;
1752             return -ENOMEM;
1753         }
1754         *refcount_table = new_refcount_table;
1755 
1756         memset(*refcount_table + old_imrt_nb_clusters, 0,
1757                (*imrt_nb_clusters - old_imrt_nb_clusters) *
1758                sizeof(**refcount_table));
1759     }
1760 
1761     /* Go back to the first free cluster */
1762     cluster -= contiguous_free_clusters;
1763     for (i = 0; i < cluster_count; i++) {
1764         (*refcount_table)[cluster + i] = 1;
1765     }
1766 
1767     return cluster << s->cluster_bits;
1768 }
1769 
1770 /*
1771  * Creates a new refcount structure based solely on the in-memory information
1772  * given through *refcount_table. All necessary allocations will be reflected
1773  * in that array.
1774  *
1775  * On success, the old refcount structure is leaked (it will be covered by the
1776  * new refcount structure).
1777  */
1778 static int rebuild_refcount_structure(BlockDriverState *bs,
1779                                       BdrvCheckResult *res,
1780                                       uint16_t **refcount_table,
1781                                       int64_t *nb_clusters)
1782 {
1783     BDRVQcowState *s = bs->opaque;
1784     int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0;
1785     int64_t refblock_offset, refblock_start, refblock_index;
1786     uint32_t reftable_size = 0;
1787     uint64_t *on_disk_reftable = NULL;
1788     uint16_t *on_disk_refblock;
1789     int i, ret = 0;
1790     struct {
1791         uint64_t reftable_offset;
1792         uint32_t reftable_clusters;
1793     } QEMU_PACKED reftable_offset_and_clusters;
1794 
1795     qcow2_cache_empty(bs, s->refcount_block_cache);
1796 
1797 write_refblocks:
1798     for (; cluster < *nb_clusters; cluster++) {
1799         if (!(*refcount_table)[cluster]) {
1800             continue;
1801         }
1802 
1803         refblock_index = cluster >> s->refcount_block_bits;
1804         refblock_start = refblock_index << s->refcount_block_bits;
1805 
1806         /* Don't allocate a cluster in a refblock already written to disk */
1807         if (first_free_cluster < refblock_start) {
1808             first_free_cluster = refblock_start;
1809         }
1810         refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table,
1811                                               nb_clusters, &first_free_cluster);
1812         if (refblock_offset < 0) {
1813             fprintf(stderr, "ERROR allocating refblock: %s\n",
1814                     strerror(-refblock_offset));
1815             res->check_errors++;
1816             ret = refblock_offset;
1817             goto fail;
1818         }
1819 
1820         if (reftable_size <= refblock_index) {
1821             uint32_t old_reftable_size = reftable_size;
1822             uint64_t *new_on_disk_reftable;
1823 
1824             reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t),
1825                                      s->cluster_size) / sizeof(uint64_t);
1826             new_on_disk_reftable = g_try_realloc(on_disk_reftable,
1827                                                  reftable_size *
1828                                                  sizeof(uint64_t));
1829             if (!new_on_disk_reftable) {
1830                 res->check_errors++;
1831                 ret = -ENOMEM;
1832                 goto fail;
1833             }
1834             on_disk_reftable = new_on_disk_reftable;
1835 
1836             memset(on_disk_reftable + old_reftable_size, 0,
1837                    (reftable_size - old_reftable_size) * sizeof(uint64_t));
1838 
1839             /* The offset we have for the reftable is now no longer valid;
1840              * this will leak that range, but we can easily fix that by running
1841              * a leak-fixing check after this rebuild operation */
1842             reftable_offset = -1;
1843         }
1844         on_disk_reftable[refblock_index] = refblock_offset;
1845 
1846         /* If this is apparently the last refblock (for now), try to squeeze the
1847          * reftable in */
1848         if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits &&
1849             reftable_offset < 0)
1850         {
1851             uint64_t reftable_clusters = size_to_clusters(s, reftable_size *
1852                                                           sizeof(uint64_t));
1853             reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
1854                                                   refcount_table, nb_clusters,
1855                                                   &first_free_cluster);
1856             if (reftable_offset < 0) {
1857                 fprintf(stderr, "ERROR allocating reftable: %s\n",
1858                         strerror(-reftable_offset));
1859                 res->check_errors++;
1860                 ret = reftable_offset;
1861                 goto fail;
1862             }
1863         }
1864 
1865         ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset,
1866                                             s->cluster_size);
1867         if (ret < 0) {
1868             fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
1869             goto fail;
1870         }
1871 
1872         on_disk_refblock = qemu_blockalign0(bs->file, s->cluster_size);
1873         for (i = 0; i < s->refcount_block_size &&
1874                     refblock_start + i < *nb_clusters; i++)
1875         {
1876             on_disk_refblock[i] =
1877                 cpu_to_be16((*refcount_table)[refblock_start + i]);
1878         }
1879 
1880         ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE,
1881                          (void *)on_disk_refblock, s->cluster_sectors);
1882         qemu_vfree(on_disk_refblock);
1883         if (ret < 0) {
1884             fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
1885             goto fail;
1886         }
1887 
1888         /* Go to the end of this refblock */
1889         cluster = refblock_start + s->refcount_block_size - 1;
1890     }
1891 
1892     if (reftable_offset < 0) {
1893         uint64_t post_refblock_start, reftable_clusters;
1894 
1895         post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size);
1896         reftable_clusters = size_to_clusters(s,
1897                                              reftable_size * sizeof(uint64_t));
1898         /* Not pretty but simple */
1899         if (first_free_cluster < post_refblock_start) {
1900             first_free_cluster = post_refblock_start;
1901         }
1902         reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
1903                                               refcount_table, nb_clusters,
1904                                               &first_free_cluster);
1905         if (reftable_offset < 0) {
1906             fprintf(stderr, "ERROR allocating reftable: %s\n",
1907                     strerror(-reftable_offset));
1908             res->check_errors++;
1909             ret = reftable_offset;
1910             goto fail;
1911         }
1912 
1913         goto write_refblocks;
1914     }
1915 
1916     assert(on_disk_reftable);
1917 
1918     for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
1919         cpu_to_be64s(&on_disk_reftable[refblock_index]);
1920     }
1921 
1922     ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset,
1923                                         reftable_size * sizeof(uint64_t));
1924     if (ret < 0) {
1925         fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
1926         goto fail;
1927     }
1928 
1929     assert(reftable_size < INT_MAX / sizeof(uint64_t));
1930     ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable,
1931                       reftable_size * sizeof(uint64_t));
1932     if (ret < 0) {
1933         fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
1934         goto fail;
1935     }
1936 
1937     /* Enter new reftable into the image header */
1938     cpu_to_be64w(&reftable_offset_and_clusters.reftable_offset,
1939                  reftable_offset);
1940     cpu_to_be32w(&reftable_offset_and_clusters.reftable_clusters,
1941                  size_to_clusters(s, reftable_size * sizeof(uint64_t)));
1942     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader,
1943                                               refcount_table_offset),
1944                            &reftable_offset_and_clusters,
1945                            sizeof(reftable_offset_and_clusters));
1946     if (ret < 0) {
1947         fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret));
1948         goto fail;
1949     }
1950 
1951     for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
1952         be64_to_cpus(&on_disk_reftable[refblock_index]);
1953     }
1954     s->refcount_table = on_disk_reftable;
1955     s->refcount_table_offset = reftable_offset;
1956     s->refcount_table_size = reftable_size;
1957 
1958     return 0;
1959 
1960 fail:
1961     g_free(on_disk_reftable);
1962     return ret;
1963 }
1964 
1965 /*
1966  * Checks an image for refcount consistency.
1967  *
1968  * Returns 0 if no errors are found, the number of errors in case the image is
1969  * detected as corrupted, and -errno when an internal error occurred.
1970  */
1971 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1972                           BdrvCheckMode fix)
1973 {
1974     BDRVQcowState *s = bs->opaque;
1975     BdrvCheckResult pre_compare_res;
1976     int64_t size, highest_cluster, nb_clusters;
1977     uint16_t *refcount_table = NULL;
1978     bool rebuild = false;
1979     int ret;
1980 
1981     size = bdrv_getlength(bs->file);
1982     if (size < 0) {
1983         res->check_errors++;
1984         return size;
1985     }
1986 
1987     nb_clusters = size_to_clusters(s, size);
1988     if (nb_clusters > INT_MAX) {
1989         res->check_errors++;
1990         return -EFBIG;
1991     }
1992 
1993     res->bfi.total_clusters =
1994         size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1995 
1996     ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table,
1997                               &nb_clusters);
1998     if (ret < 0) {
1999         goto fail;
2000     }
2001 
2002     /* In case we don't need to rebuild the refcount structure (but want to fix
2003      * something), this function is immediately called again, in which case the
2004      * result should be ignored */
2005     pre_compare_res = *res;
2006     compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table,
2007                       nb_clusters);
2008 
2009     if (rebuild && (fix & BDRV_FIX_ERRORS)) {
2010         BdrvCheckResult old_res = *res;
2011         int fresh_leaks = 0;
2012 
2013         fprintf(stderr, "Rebuilding refcount structure\n");
2014         ret = rebuild_refcount_structure(bs, res, &refcount_table,
2015                                          &nb_clusters);
2016         if (ret < 0) {
2017             goto fail;
2018         }
2019 
2020         res->corruptions = 0;
2021         res->leaks = 0;
2022 
2023         /* Because the old reftable has been exchanged for a new one the
2024          * references have to be recalculated */
2025         rebuild = false;
2026         memset(refcount_table, 0, nb_clusters * sizeof(uint16_t));
2027         ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table,
2028                                   &nb_clusters);
2029         if (ret < 0) {
2030             goto fail;
2031         }
2032 
2033         if (fix & BDRV_FIX_LEAKS) {
2034             /* The old refcount structures are now leaked, fix it; the result
2035              * can be ignored, aside from leaks which were introduced by
2036              * rebuild_refcount_structure() that could not be fixed */
2037             BdrvCheckResult saved_res = *res;
2038             *res = (BdrvCheckResult){ 0 };
2039 
2040             compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild,
2041                               &highest_cluster, refcount_table, nb_clusters);
2042             if (rebuild) {
2043                 fprintf(stderr, "ERROR rebuilt refcount structure is still "
2044                         "broken\n");
2045             }
2046 
2047             /* Any leaks accounted for here were introduced by
2048              * rebuild_refcount_structure() because that function has created a
2049              * new refcount structure from scratch */
2050             fresh_leaks = res->leaks;
2051             *res = saved_res;
2052         }
2053 
2054         if (res->corruptions < old_res.corruptions) {
2055             res->corruptions_fixed += old_res.corruptions - res->corruptions;
2056         }
2057         if (res->leaks < old_res.leaks) {
2058             res->leaks_fixed += old_res.leaks - res->leaks;
2059         }
2060         res->leaks += fresh_leaks;
2061     } else if (fix) {
2062         if (rebuild) {
2063             fprintf(stderr, "ERROR need to rebuild refcount structures\n");
2064             res->check_errors++;
2065             ret = -EIO;
2066             goto fail;
2067         }
2068 
2069         if (res->leaks || res->corruptions) {
2070             *res = pre_compare_res;
2071             compare_refcounts(bs, res, fix, &rebuild, &highest_cluster,
2072                               refcount_table, nb_clusters);
2073         }
2074     }
2075 
2076     /* check OFLAG_COPIED */
2077     ret = check_oflag_copied(bs, res, fix);
2078     if (ret < 0) {
2079         goto fail;
2080     }
2081 
2082     res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
2083     ret = 0;
2084 
2085 fail:
2086     g_free(refcount_table);
2087 
2088     return ret;
2089 }
2090 
2091 #define overlaps_with(ofs, sz) \
2092     ranges_overlap(offset, size, ofs, sz)
2093 
2094 /*
2095  * Checks if the given offset into the image file is actually free to use by
2096  * looking for overlaps with important metadata sections (L1/L2 tables etc.),
2097  * i.e. a sanity check without relying on the refcount tables.
2098  *
2099  * The ign parameter specifies what checks not to perform (being a bitmask of
2100  * QCow2MetadataOverlap values), i.e., what sections to ignore.
2101  *
2102  * Returns:
2103  * - 0 if writing to this offset will not affect the mentioned metadata
2104  * - a positive QCow2MetadataOverlap value indicating one overlapping section
2105  * - a negative value (-errno) indicating an error while performing a check,
2106  *   e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
2107  */
2108 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
2109                                  int64_t size)
2110 {
2111     BDRVQcowState *s = bs->opaque;
2112     int chk = s->overlap_check & ~ign;
2113     int i, j;
2114 
2115     if (!size) {
2116         return 0;
2117     }
2118 
2119     if (chk & QCOW2_OL_MAIN_HEADER) {
2120         if (offset < s->cluster_size) {
2121             return QCOW2_OL_MAIN_HEADER;
2122         }
2123     }
2124 
2125     /* align range to test to cluster boundaries */
2126     size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size);
2127     offset = start_of_cluster(s, offset);
2128 
2129     if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
2130         if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
2131             return QCOW2_OL_ACTIVE_L1;
2132         }
2133     }
2134 
2135     if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
2136         if (overlaps_with(s->refcount_table_offset,
2137             s->refcount_table_size * sizeof(uint64_t))) {
2138             return QCOW2_OL_REFCOUNT_TABLE;
2139         }
2140     }
2141 
2142     if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
2143         if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
2144             return QCOW2_OL_SNAPSHOT_TABLE;
2145         }
2146     }
2147 
2148     if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
2149         for (i = 0; i < s->nb_snapshots; i++) {
2150             if (s->snapshots[i].l1_size &&
2151                 overlaps_with(s->snapshots[i].l1_table_offset,
2152                 s->snapshots[i].l1_size * sizeof(uint64_t))) {
2153                 return QCOW2_OL_INACTIVE_L1;
2154             }
2155         }
2156     }
2157 
2158     if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
2159         for (i = 0; i < s->l1_size; i++) {
2160             if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
2161                 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
2162                 s->cluster_size)) {
2163                 return QCOW2_OL_ACTIVE_L2;
2164             }
2165         }
2166     }
2167 
2168     if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
2169         for (i = 0; i < s->refcount_table_size; i++) {
2170             if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
2171                 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
2172                 s->cluster_size)) {
2173                 return QCOW2_OL_REFCOUNT_BLOCK;
2174             }
2175         }
2176     }
2177 
2178     if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
2179         for (i = 0; i < s->nb_snapshots; i++) {
2180             uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
2181             uint32_t l1_sz  = s->snapshots[i].l1_size;
2182             uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
2183             uint64_t *l1 = g_try_malloc(l1_sz2);
2184             int ret;
2185 
2186             if (l1_sz2 && l1 == NULL) {
2187                 return -ENOMEM;
2188             }
2189 
2190             ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
2191             if (ret < 0) {
2192                 g_free(l1);
2193                 return ret;
2194             }
2195 
2196             for (j = 0; j < l1_sz; j++) {
2197                 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
2198                 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
2199                     g_free(l1);
2200                     return QCOW2_OL_INACTIVE_L2;
2201                 }
2202             }
2203 
2204             g_free(l1);
2205         }
2206     }
2207 
2208     return 0;
2209 }
2210 
2211 static const char *metadata_ol_names[] = {
2212     [QCOW2_OL_MAIN_HEADER_BITNR]    = "qcow2_header",
2213     [QCOW2_OL_ACTIVE_L1_BITNR]      = "active L1 table",
2214     [QCOW2_OL_ACTIVE_L2_BITNR]      = "active L2 table",
2215     [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
2216     [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
2217     [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
2218     [QCOW2_OL_INACTIVE_L1_BITNR]    = "inactive L1 table",
2219     [QCOW2_OL_INACTIVE_L2_BITNR]    = "inactive L2 table",
2220 };
2221 
2222 /*
2223  * First performs a check for metadata overlaps (through
2224  * qcow2_check_metadata_overlap); if that fails with a negative value (error
2225  * while performing a check), that value is returned. If an impending overlap
2226  * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
2227  * and -EIO returned.
2228  *
2229  * Returns 0 if there were neither overlaps nor errors while checking for
2230  * overlaps; or a negative value (-errno) on error.
2231  */
2232 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
2233                                   int64_t size)
2234 {
2235     int ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
2236 
2237     if (ret < 0) {
2238         return ret;
2239     } else if (ret > 0) {
2240         int metadata_ol_bitnr = ffs(ret) - 1;
2241         assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
2242 
2243         qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid "
2244                                 "write on metadata (overlaps with %s)",
2245                                 metadata_ol_names[metadata_ol_bitnr]);
2246         return -EIO;
2247     }
2248 
2249     return 0;
2250 }
2251