xref: /openbmc/qemu/block/qcow2-refcount.c (revision a42e9c41)
1 /*
2  * Block driver for the QCOW version 2 format
3  *
4  * Copyright (c) 2004-2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu-common.h"
26 #include "block/block_int.h"
27 #include "block/qcow2.h"
28 #include "qemu/range.h"
29 #include "qapi/qmp/types.h"
30 
31 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size);
32 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
33                             int64_t offset, int64_t length,
34                             int addend, enum qcow2_discard_type type);
35 
36 
37 /*********************************************************/
38 /* refcount handling */
39 
40 int qcow2_refcount_init(BlockDriverState *bs)
41 {
42     BDRVQcowState *s = bs->opaque;
43     int ret, refcount_table_size2, i;
44 
45     refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
46     s->refcount_table = g_malloc(refcount_table_size2);
47     if (s->refcount_table_size > 0) {
48         BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
49         ret = bdrv_pread(bs->file, s->refcount_table_offset,
50                          s->refcount_table, refcount_table_size2);
51         if (ret != refcount_table_size2)
52             goto fail;
53         for(i = 0; i < s->refcount_table_size; i++)
54             be64_to_cpus(&s->refcount_table[i]);
55     }
56     return 0;
57  fail:
58     return -ENOMEM;
59 }
60 
61 void qcow2_refcount_close(BlockDriverState *bs)
62 {
63     BDRVQcowState *s = bs->opaque;
64     g_free(s->refcount_table);
65 }
66 
67 
68 static int load_refcount_block(BlockDriverState *bs,
69                                int64_t refcount_block_offset,
70                                void **refcount_block)
71 {
72     BDRVQcowState *s = bs->opaque;
73     int ret;
74 
75     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
76     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
77         refcount_block);
78 
79     return ret;
80 }
81 
82 /*
83  * Returns the refcount of the cluster given by its index. Any non-negative
84  * return value is the refcount of the cluster, negative values are -errno
85  * and indicate an error.
86  */
87 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
88 {
89     BDRVQcowState *s = bs->opaque;
90     int refcount_table_index, block_index;
91     int64_t refcount_block_offset;
92     int ret;
93     uint16_t *refcount_block;
94     uint16_t refcount;
95 
96     refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
97     if (refcount_table_index >= s->refcount_table_size)
98         return 0;
99     refcount_block_offset = s->refcount_table[refcount_table_index];
100     if (!refcount_block_offset)
101         return 0;
102 
103     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
104         (void**) &refcount_block);
105     if (ret < 0) {
106         return ret;
107     }
108 
109     block_index = cluster_index &
110         ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
111     refcount = be16_to_cpu(refcount_block[block_index]);
112 
113     ret = qcow2_cache_put(bs, s->refcount_block_cache,
114         (void**) &refcount_block);
115     if (ret < 0) {
116         return ret;
117     }
118 
119     return refcount;
120 }
121 
122 /*
123  * Rounds the refcount table size up to avoid growing the table for each single
124  * refcount block that is allocated.
125  */
126 static unsigned int next_refcount_table_size(BDRVQcowState *s,
127     unsigned int min_size)
128 {
129     unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
130     unsigned int refcount_table_clusters =
131         MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
132 
133     while (min_clusters > refcount_table_clusters) {
134         refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
135     }
136 
137     return refcount_table_clusters << (s->cluster_bits - 3);
138 }
139 
140 
141 /* Checks if two offsets are described by the same refcount block */
142 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
143     uint64_t offset_b)
144 {
145     uint64_t block_a = offset_a >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
146     uint64_t block_b = offset_b >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
147 
148     return (block_a == block_b);
149 }
150 
151 /*
152  * Loads a refcount block. If it doesn't exist yet, it is allocated first
153  * (including growing the refcount table if needed).
154  *
155  * Returns 0 on success or -errno in error case
156  */
157 static int alloc_refcount_block(BlockDriverState *bs,
158     int64_t cluster_index, uint16_t **refcount_block)
159 {
160     BDRVQcowState *s = bs->opaque;
161     unsigned int refcount_table_index;
162     int ret;
163 
164     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
165 
166     /* Find the refcount block for the given cluster */
167     refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
168 
169     if (refcount_table_index < s->refcount_table_size) {
170 
171         uint64_t refcount_block_offset =
172             s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
173 
174         /* If it's already there, we're done */
175         if (refcount_block_offset) {
176              return load_refcount_block(bs, refcount_block_offset,
177                  (void**) refcount_block);
178         }
179     }
180 
181     /*
182      * If we came here, we need to allocate something. Something is at least
183      * a cluster for the new refcount block. It may also include a new refcount
184      * table if the old refcount table is too small.
185      *
186      * Note that allocating clusters here needs some special care:
187      *
188      * - We can't use the normal qcow2_alloc_clusters(), it would try to
189      *   increase the refcount and very likely we would end up with an endless
190      *   recursion. Instead we must place the refcount blocks in a way that
191      *   they can describe them themselves.
192      *
193      * - We need to consider that at this point we are inside update_refcounts
194      *   and doing the initial refcount increase. This means that some clusters
195      *   have already been allocated by the caller, but their refcount isn't
196      *   accurate yet. free_cluster_index tells us where this allocation ends
197      *   as long as we don't overwrite it by freeing clusters.
198      *
199      * - alloc_clusters_noref and qcow2_free_clusters may load a different
200      *   refcount block into the cache
201      */
202 
203     *refcount_block = NULL;
204 
205     /* We write to the refcount table, so we might depend on L2 tables */
206     ret = qcow2_cache_flush(bs, s->l2_table_cache);
207     if (ret < 0) {
208         return ret;
209     }
210 
211     /* Allocate the refcount block itself and mark it as used */
212     int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
213     if (new_block < 0) {
214         return new_block;
215     }
216 
217 #ifdef DEBUG_ALLOC2
218     fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
219         " at %" PRIx64 "\n",
220         refcount_table_index, cluster_index << s->cluster_bits, new_block);
221 #endif
222 
223     if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
224         /* Zero the new refcount block before updating it */
225         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
226             (void**) refcount_block);
227         if (ret < 0) {
228             goto fail_block;
229         }
230 
231         memset(*refcount_block, 0, s->cluster_size);
232 
233         /* The block describes itself, need to update the cache */
234         int block_index = (new_block >> s->cluster_bits) &
235             ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
236         (*refcount_block)[block_index] = cpu_to_be16(1);
237     } else {
238         /* Described somewhere else. This can recurse at most twice before we
239          * arrive at a block that describes itself. */
240         ret = update_refcount(bs, new_block, s->cluster_size, 1,
241                               QCOW2_DISCARD_NEVER);
242         if (ret < 0) {
243             goto fail_block;
244         }
245 
246         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
247         if (ret < 0) {
248             goto fail_block;
249         }
250 
251         /* Initialize the new refcount block only after updating its refcount,
252          * update_refcount uses the refcount cache itself */
253         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
254             (void**) refcount_block);
255         if (ret < 0) {
256             goto fail_block;
257         }
258 
259         memset(*refcount_block, 0, s->cluster_size);
260     }
261 
262     /* Now the new refcount block needs to be written to disk */
263     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
264     qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
265     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
266     if (ret < 0) {
267         goto fail_block;
268     }
269 
270     /* If the refcount table is big enough, just hook the block up there */
271     if (refcount_table_index < s->refcount_table_size) {
272         uint64_t data64 = cpu_to_be64(new_block);
273         BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
274         ret = bdrv_pwrite_sync(bs->file,
275             s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
276             &data64, sizeof(data64));
277         if (ret < 0) {
278             goto fail_block;
279         }
280 
281         s->refcount_table[refcount_table_index] = new_block;
282         return 0;
283     }
284 
285     ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
286     if (ret < 0) {
287         goto fail_block;
288     }
289 
290     /*
291      * If we come here, we need to grow the refcount table. Again, a new
292      * refcount table needs some space and we can't simply allocate to avoid
293      * endless recursion.
294      *
295      * Therefore let's grab new refcount blocks at the end of the image, which
296      * will describe themselves and the new refcount table. This way we can
297      * reference them only in the new table and do the switch to the new
298      * refcount table at once without producing an inconsistent state in
299      * between.
300      */
301     BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
302 
303     /* Calculate the number of refcount blocks needed so far */
304     uint64_t refcount_block_clusters = 1 << (s->cluster_bits - REFCOUNT_SHIFT);
305     uint64_t blocks_used = (s->free_cluster_index +
306         refcount_block_clusters - 1) / refcount_block_clusters;
307 
308     /* And now we need at least one block more for the new metadata */
309     uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
310     uint64_t last_table_size;
311     uint64_t blocks_clusters;
312     do {
313         uint64_t table_clusters =
314             size_to_clusters(s, table_size * sizeof(uint64_t));
315         blocks_clusters = 1 +
316             ((table_clusters + refcount_block_clusters - 1)
317             / refcount_block_clusters);
318         uint64_t meta_clusters = table_clusters + blocks_clusters;
319 
320         last_table_size = table_size;
321         table_size = next_refcount_table_size(s, blocks_used +
322             ((meta_clusters + refcount_block_clusters - 1)
323             / refcount_block_clusters));
324 
325     } while (last_table_size != table_size);
326 
327 #ifdef DEBUG_ALLOC2
328     fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
329         s->refcount_table_size, table_size);
330 #endif
331 
332     /* Create the new refcount table and blocks */
333     uint64_t meta_offset = (blocks_used * refcount_block_clusters) *
334         s->cluster_size;
335     uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
336     uint16_t *new_blocks = g_malloc0(blocks_clusters * s->cluster_size);
337     uint64_t *new_table = g_malloc0(table_size * sizeof(uint64_t));
338 
339     assert(meta_offset >= (s->free_cluster_index * s->cluster_size));
340 
341     /* Fill the new refcount table */
342     memcpy(new_table, s->refcount_table,
343         s->refcount_table_size * sizeof(uint64_t));
344     new_table[refcount_table_index] = new_block;
345 
346     int i;
347     for (i = 0; i < blocks_clusters; i++) {
348         new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
349     }
350 
351     /* Fill the refcount blocks */
352     uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
353     int block = 0;
354     for (i = 0; i < table_clusters + blocks_clusters; i++) {
355         new_blocks[block++] = cpu_to_be16(1);
356     }
357 
358     /* Write refcount blocks to disk */
359     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
360     ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
361         blocks_clusters * s->cluster_size);
362     g_free(new_blocks);
363     if (ret < 0) {
364         goto fail_table;
365     }
366 
367     /* Write refcount table to disk */
368     for(i = 0; i < table_size; i++) {
369         cpu_to_be64s(&new_table[i]);
370     }
371 
372     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
373     ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
374         table_size * sizeof(uint64_t));
375     if (ret < 0) {
376         goto fail_table;
377     }
378 
379     for(i = 0; i < table_size; i++) {
380         be64_to_cpus(&new_table[i]);
381     }
382 
383     /* Hook up the new refcount table in the qcow2 header */
384     uint8_t data[12];
385     cpu_to_be64w((uint64_t*)data, table_offset);
386     cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
387     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
388     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
389         data, sizeof(data));
390     if (ret < 0) {
391         goto fail_table;
392     }
393 
394     /* And switch it in memory */
395     uint64_t old_table_offset = s->refcount_table_offset;
396     uint64_t old_table_size = s->refcount_table_size;
397 
398     g_free(s->refcount_table);
399     s->refcount_table = new_table;
400     s->refcount_table_size = table_size;
401     s->refcount_table_offset = table_offset;
402 
403     /* Free old table. Remember, we must not change free_cluster_index */
404     uint64_t old_free_cluster_index = s->free_cluster_index;
405     qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
406                         QCOW2_DISCARD_OTHER);
407     s->free_cluster_index = old_free_cluster_index;
408 
409     ret = load_refcount_block(bs, new_block, (void**) refcount_block);
410     if (ret < 0) {
411         return ret;
412     }
413 
414     return 0;
415 
416 fail_table:
417     g_free(new_table);
418 fail_block:
419     if (*refcount_block != NULL) {
420         qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
421     }
422     return ret;
423 }
424 
425 void qcow2_process_discards(BlockDriverState *bs, int ret)
426 {
427     BDRVQcowState *s = bs->opaque;
428     Qcow2DiscardRegion *d, *next;
429 
430     QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
431         QTAILQ_REMOVE(&s->discards, d, next);
432 
433         /* Discard is optional, ignore the return value */
434         if (ret >= 0) {
435             bdrv_discard(bs->file,
436                          d->offset >> BDRV_SECTOR_BITS,
437                          d->bytes >> BDRV_SECTOR_BITS);
438         }
439 
440         g_free(d);
441     }
442 }
443 
444 static void update_refcount_discard(BlockDriverState *bs,
445                                     uint64_t offset, uint64_t length)
446 {
447     BDRVQcowState *s = bs->opaque;
448     Qcow2DiscardRegion *d, *p, *next;
449 
450     QTAILQ_FOREACH(d, &s->discards, next) {
451         uint64_t new_start = MIN(offset, d->offset);
452         uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
453 
454         if (new_end - new_start <= length + d->bytes) {
455             /* There can't be any overlap, areas ending up here have no
456              * references any more and therefore shouldn't get freed another
457              * time. */
458             assert(d->bytes + length == new_end - new_start);
459             d->offset = new_start;
460             d->bytes = new_end - new_start;
461             goto found;
462         }
463     }
464 
465     d = g_malloc(sizeof(*d));
466     *d = (Qcow2DiscardRegion) {
467         .bs     = bs,
468         .offset = offset,
469         .bytes  = length,
470     };
471     QTAILQ_INSERT_TAIL(&s->discards, d, next);
472 
473 found:
474     /* Merge discard requests if they are adjacent now */
475     QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
476         if (p == d
477             || p->offset > d->offset + d->bytes
478             || d->offset > p->offset + p->bytes)
479         {
480             continue;
481         }
482 
483         /* Still no overlap possible */
484         assert(p->offset == d->offset + d->bytes
485             || d->offset == p->offset + p->bytes);
486 
487         QTAILQ_REMOVE(&s->discards, p, next);
488         d->offset = MIN(d->offset, p->offset);
489         d->bytes += p->bytes;
490     }
491 }
492 
493 /* XXX: cache several refcount block clusters ? */
494 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
495     int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
496 {
497     BDRVQcowState *s = bs->opaque;
498     int64_t start, last, cluster_offset;
499     uint16_t *refcount_block = NULL;
500     int64_t old_table_index = -1;
501     int ret;
502 
503 #ifdef DEBUG_ALLOC2
504     fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
505            offset, length, addend);
506 #endif
507     if (length < 0) {
508         return -EINVAL;
509     } else if (length == 0) {
510         return 0;
511     }
512 
513     if (addend < 0) {
514         qcow2_cache_set_dependency(bs, s->refcount_block_cache,
515             s->l2_table_cache);
516     }
517 
518     start = offset & ~(s->cluster_size - 1);
519     last = (offset + length - 1) & ~(s->cluster_size - 1);
520     for(cluster_offset = start; cluster_offset <= last;
521         cluster_offset += s->cluster_size)
522     {
523         int block_index, refcount;
524         int64_t cluster_index = cluster_offset >> s->cluster_bits;
525         int64_t table_index =
526             cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
527 
528         /* Load the refcount block and allocate it if needed */
529         if (table_index != old_table_index) {
530             if (refcount_block) {
531                 ret = qcow2_cache_put(bs, s->refcount_block_cache,
532                     (void**) &refcount_block);
533                 if (ret < 0) {
534                     goto fail;
535                 }
536             }
537 
538             ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
539             if (ret < 0) {
540                 goto fail;
541             }
542         }
543         old_table_index = table_index;
544 
545         qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
546 
547         /* we can update the count and save it */
548         block_index = cluster_index &
549             ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
550 
551         refcount = be16_to_cpu(refcount_block[block_index]);
552         refcount += addend;
553         if (refcount < 0 || refcount > 0xffff) {
554             ret = -EINVAL;
555             goto fail;
556         }
557         if (refcount == 0 && cluster_index < s->free_cluster_index) {
558             s->free_cluster_index = cluster_index;
559         }
560         refcount_block[block_index] = cpu_to_be16(refcount);
561 
562         if (refcount == 0 && s->discard_passthrough[type]) {
563             update_refcount_discard(bs, cluster_offset, s->cluster_size);
564         }
565     }
566 
567     ret = 0;
568 fail:
569     if (!s->cache_discards) {
570         qcow2_process_discards(bs, ret);
571     }
572 
573     /* Write last changed block to disk */
574     if (refcount_block) {
575         int wret;
576         wret = qcow2_cache_put(bs, s->refcount_block_cache,
577             (void**) &refcount_block);
578         if (wret < 0) {
579             return ret < 0 ? ret : wret;
580         }
581     }
582 
583     /*
584      * Try do undo any updates if an error is returned (This may succeed in
585      * some cases like ENOSPC for allocating a new refcount block)
586      */
587     if (ret < 0) {
588         int dummy;
589         dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
590                                 QCOW2_DISCARD_NEVER);
591         (void)dummy;
592     }
593 
594     return ret;
595 }
596 
597 /*
598  * Increases or decreases the refcount of a given cluster by one.
599  * addend must be 1 or -1.
600  *
601  * If the return value is non-negative, it is the new refcount of the cluster.
602  * If it is negative, it is -errno and indicates an error.
603  */
604 static int update_cluster_refcount(BlockDriverState *bs,
605                                    int64_t cluster_index,
606                                    int addend,
607                                    enum qcow2_discard_type type)
608 {
609     BDRVQcowState *s = bs->opaque;
610     int ret;
611 
612     ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
613                           type);
614     if (ret < 0) {
615         return ret;
616     }
617 
618     return get_refcount(bs, cluster_index);
619 }
620 
621 
622 
623 /*********************************************************/
624 /* cluster allocation functions */
625 
626 
627 
628 /* return < 0 if error */
629 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size)
630 {
631     BDRVQcowState *s = bs->opaque;
632     int i, nb_clusters, refcount;
633 
634     nb_clusters = size_to_clusters(s, size);
635 retry:
636     for(i = 0; i < nb_clusters; i++) {
637         int64_t next_cluster_index = s->free_cluster_index++;
638         refcount = get_refcount(bs, next_cluster_index);
639 
640         if (refcount < 0) {
641             return refcount;
642         } else if (refcount != 0) {
643             goto retry;
644         }
645     }
646 #ifdef DEBUG_ALLOC2
647     fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
648             size,
649             (s->free_cluster_index - nb_clusters) << s->cluster_bits);
650 #endif
651     return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
652 }
653 
654 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size)
655 {
656     int64_t offset;
657     int ret;
658 
659     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
660     offset = alloc_clusters_noref(bs, size);
661     if (offset < 0) {
662         return offset;
663     }
664 
665     ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
666     if (ret < 0) {
667         return ret;
668     }
669 
670     return offset;
671 }
672 
673 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
674     int nb_clusters)
675 {
676     BDRVQcowState *s = bs->opaque;
677     uint64_t cluster_index;
678     uint64_t old_free_cluster_index;
679     int i, refcount, ret;
680 
681     /* Check how many clusters there are free */
682     cluster_index = offset >> s->cluster_bits;
683     for(i = 0; i < nb_clusters; i++) {
684         refcount = get_refcount(bs, cluster_index++);
685 
686         if (refcount < 0) {
687             return refcount;
688         } else if (refcount != 0) {
689             break;
690         }
691     }
692 
693     /* And then allocate them */
694     old_free_cluster_index = s->free_cluster_index;
695     s->free_cluster_index = cluster_index + i;
696 
697     ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
698                           QCOW2_DISCARD_NEVER);
699     if (ret < 0) {
700         return ret;
701     }
702 
703     s->free_cluster_index = old_free_cluster_index;
704 
705     return i;
706 }
707 
708 /* only used to allocate compressed sectors. We try to allocate
709    contiguous sectors. size must be <= cluster_size */
710 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
711 {
712     BDRVQcowState *s = bs->opaque;
713     int64_t offset, cluster_offset;
714     int free_in_cluster;
715 
716     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
717     assert(size > 0 && size <= s->cluster_size);
718     if (s->free_byte_offset == 0) {
719         offset = qcow2_alloc_clusters(bs, s->cluster_size);
720         if (offset < 0) {
721             return offset;
722         }
723         s->free_byte_offset = offset;
724     }
725  redo:
726     free_in_cluster = s->cluster_size -
727         (s->free_byte_offset & (s->cluster_size - 1));
728     if (size <= free_in_cluster) {
729         /* enough space in current cluster */
730         offset = s->free_byte_offset;
731         s->free_byte_offset += size;
732         free_in_cluster -= size;
733         if (free_in_cluster == 0)
734             s->free_byte_offset = 0;
735         if ((offset & (s->cluster_size - 1)) != 0)
736             update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
737                                     QCOW2_DISCARD_NEVER);
738     } else {
739         offset = qcow2_alloc_clusters(bs, s->cluster_size);
740         if (offset < 0) {
741             return offset;
742         }
743         cluster_offset = s->free_byte_offset & ~(s->cluster_size - 1);
744         if ((cluster_offset + s->cluster_size) == offset) {
745             /* we are lucky: contiguous data */
746             offset = s->free_byte_offset;
747             update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
748                                     QCOW2_DISCARD_NEVER);
749             s->free_byte_offset += size;
750         } else {
751             s->free_byte_offset = offset;
752             goto redo;
753         }
754     }
755 
756     /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
757      * or explicitly by update_cluster_refcount().  Refcount blocks must be
758      * flushed before the caller's L2 table updates.
759      */
760     qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
761     return offset;
762 }
763 
764 void qcow2_free_clusters(BlockDriverState *bs,
765                           int64_t offset, int64_t size,
766                           enum qcow2_discard_type type)
767 {
768     int ret;
769 
770     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
771     ret = update_refcount(bs, offset, size, -1, type);
772     if (ret < 0) {
773         fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
774         /* TODO Remember the clusters to free them later and avoid leaking */
775     }
776 }
777 
778 /*
779  * Free a cluster using its L2 entry (handles clusters of all types, e.g.
780  * normal cluster, compressed cluster, etc.)
781  */
782 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
783                              int nb_clusters, enum qcow2_discard_type type)
784 {
785     BDRVQcowState *s = bs->opaque;
786 
787     switch (qcow2_get_cluster_type(l2_entry)) {
788     case QCOW2_CLUSTER_COMPRESSED:
789         {
790             int nb_csectors;
791             nb_csectors = ((l2_entry >> s->csize_shift) &
792                            s->csize_mask) + 1;
793             qcow2_free_clusters(bs,
794                 (l2_entry & s->cluster_offset_mask) & ~511,
795                 nb_csectors * 512, type);
796         }
797         break;
798     case QCOW2_CLUSTER_NORMAL:
799         qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
800                             nb_clusters << s->cluster_bits, type);
801         break;
802     case QCOW2_CLUSTER_UNALLOCATED:
803     case QCOW2_CLUSTER_ZERO:
804         break;
805     default:
806         abort();
807     }
808 }
809 
810 
811 
812 /*********************************************************/
813 /* snapshots and image creation */
814 
815 
816 
817 /* update the refcounts of snapshots and the copied flag */
818 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
819     int64_t l1_table_offset, int l1_size, int addend)
820 {
821     BDRVQcowState *s = bs->opaque;
822     uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, l1_allocated;
823     int64_t old_offset, old_l2_offset;
824     int i, j, l1_modified = 0, nb_csectors, refcount;
825     int ret;
826 
827     l2_table = NULL;
828     l1_table = NULL;
829     l1_size2 = l1_size * sizeof(uint64_t);
830 
831     s->cache_discards = true;
832 
833     /* WARNING: qcow2_snapshot_goto relies on this function not using the
834      * l1_table_offset when it is the current s->l1_table_offset! Be careful
835      * when changing this! */
836     if (l1_table_offset != s->l1_table_offset) {
837         l1_table = g_malloc0(align_offset(l1_size2, 512));
838         l1_allocated = 1;
839 
840         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
841         if (ret < 0) {
842             goto fail;
843         }
844 
845         for(i = 0;i < l1_size; i++)
846             be64_to_cpus(&l1_table[i]);
847     } else {
848         assert(l1_size == s->l1_size);
849         l1_table = s->l1_table;
850         l1_allocated = 0;
851     }
852 
853     for(i = 0; i < l1_size; i++) {
854         l2_offset = l1_table[i];
855         if (l2_offset) {
856             old_l2_offset = l2_offset;
857             l2_offset &= L1E_OFFSET_MASK;
858 
859             ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
860                 (void**) &l2_table);
861             if (ret < 0) {
862                 goto fail;
863             }
864 
865             for(j = 0; j < s->l2_size; j++) {
866                 uint64_t cluster_index;
867 
868                 offset = be64_to_cpu(l2_table[j]);
869                 old_offset = offset;
870                 offset &= ~QCOW_OFLAG_COPIED;
871 
872                 switch (qcow2_get_cluster_type(offset)) {
873                     case QCOW2_CLUSTER_COMPRESSED:
874                         nb_csectors = ((offset >> s->csize_shift) &
875                                        s->csize_mask) + 1;
876                         if (addend != 0) {
877                             int ret;
878                             ret = update_refcount(bs,
879                                 (offset & s->cluster_offset_mask) & ~511,
880                                 nb_csectors * 512, addend,
881                                 QCOW2_DISCARD_SNAPSHOT);
882                             if (ret < 0) {
883                                 goto fail;
884                             }
885                         }
886                         /* compressed clusters are never modified */
887                         refcount = 2;
888                         break;
889 
890                     case QCOW2_CLUSTER_NORMAL:
891                     case QCOW2_CLUSTER_ZERO:
892                         cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
893                         if (!cluster_index) {
894                             /* unallocated */
895                             refcount = 0;
896                             break;
897                         }
898                         if (addend != 0) {
899                             refcount = update_cluster_refcount(bs, cluster_index, addend,
900                                                                QCOW2_DISCARD_SNAPSHOT);
901                         } else {
902                             refcount = get_refcount(bs, cluster_index);
903                         }
904 
905                         if (refcount < 0) {
906                             ret = refcount;
907                             goto fail;
908                         }
909                         break;
910 
911                     case QCOW2_CLUSTER_UNALLOCATED:
912                         refcount = 0;
913                         break;
914 
915                     default:
916                         abort();
917                 }
918 
919                 if (refcount == 1) {
920                     offset |= QCOW_OFLAG_COPIED;
921                 }
922                 if (offset != old_offset) {
923                     if (addend > 0) {
924                         qcow2_cache_set_dependency(bs, s->l2_table_cache,
925                             s->refcount_block_cache);
926                     }
927                     l2_table[j] = cpu_to_be64(offset);
928                     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
929                 }
930             }
931 
932             ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
933             if (ret < 0) {
934                 goto fail;
935             }
936 
937 
938             if (addend != 0) {
939                 refcount = update_cluster_refcount(bs, l2_offset >> s->cluster_bits, addend,
940                                                    QCOW2_DISCARD_SNAPSHOT);
941             } else {
942                 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
943             }
944             if (refcount < 0) {
945                 ret = refcount;
946                 goto fail;
947             } else if (refcount == 1) {
948                 l2_offset |= QCOW_OFLAG_COPIED;
949             }
950             if (l2_offset != old_l2_offset) {
951                 l1_table[i] = l2_offset;
952                 l1_modified = 1;
953             }
954         }
955     }
956 
957     ret = bdrv_flush(bs);
958 fail:
959     if (l2_table) {
960         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
961     }
962 
963     s->cache_discards = false;
964     qcow2_process_discards(bs, ret);
965 
966     /* Update L1 only if it isn't deleted anyway (addend = -1) */
967     if (ret == 0 && addend >= 0 && l1_modified) {
968         for (i = 0; i < l1_size; i++) {
969             cpu_to_be64s(&l1_table[i]);
970         }
971 
972         ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
973 
974         for (i = 0; i < l1_size; i++) {
975             be64_to_cpus(&l1_table[i]);
976         }
977     }
978     if (l1_allocated)
979         g_free(l1_table);
980     return ret;
981 }
982 
983 
984 
985 
986 /*********************************************************/
987 /* refcount checking functions */
988 
989 
990 
991 /*
992  * Increases the refcount for a range of clusters in a given refcount table.
993  * This is used to construct a temporary refcount table out of L1 and L2 tables
994  * which can be compared the the refcount table saved in the image.
995  *
996  * Modifies the number of errors in res.
997  */
998 static void inc_refcounts(BlockDriverState *bs,
999                           BdrvCheckResult *res,
1000                           uint16_t *refcount_table,
1001                           int refcount_table_size,
1002                           int64_t offset, int64_t size)
1003 {
1004     BDRVQcowState *s = bs->opaque;
1005     int64_t start, last, cluster_offset;
1006     int k;
1007 
1008     if (size <= 0)
1009         return;
1010 
1011     start = offset & ~(s->cluster_size - 1);
1012     last = (offset + size - 1) & ~(s->cluster_size - 1);
1013     for(cluster_offset = start; cluster_offset <= last;
1014         cluster_offset += s->cluster_size) {
1015         k = cluster_offset >> s->cluster_bits;
1016         if (k < 0) {
1017             fprintf(stderr, "ERROR: invalid cluster offset=0x%" PRIx64 "\n",
1018                 cluster_offset);
1019             res->corruptions++;
1020         } else if (k >= refcount_table_size) {
1021             fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
1022                 "the end of the image file, can't properly check refcounts.\n",
1023                 cluster_offset);
1024             res->check_errors++;
1025         } else {
1026             if (++refcount_table[k] == 0) {
1027                 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1028                     "\n", cluster_offset);
1029                 res->corruptions++;
1030             }
1031         }
1032     }
1033 }
1034 
1035 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1036 enum {
1037     CHECK_OFLAG_COPIED = 0x1,   /* check QCOW_OFLAG_COPIED matches refcount */
1038     CHECK_FRAG_INFO = 0x2,      /* update BlockFragInfo counters */
1039 };
1040 
1041 /*
1042  * Increases the refcount in the given refcount table for the all clusters
1043  * referenced in the L2 table. While doing so, performs some checks on L2
1044  * entries.
1045  *
1046  * Returns the number of errors found by the checks or -errno if an internal
1047  * error occurred.
1048  */
1049 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1050     uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
1051     int flags)
1052 {
1053     BDRVQcowState *s = bs->opaque;
1054     uint64_t *l2_table, l2_entry;
1055     uint64_t next_contiguous_offset = 0;
1056     int i, l2_size, nb_csectors;
1057 
1058     /* Read L2 table from disk */
1059     l2_size = s->l2_size * sizeof(uint64_t);
1060     l2_table = g_malloc(l2_size);
1061 
1062     if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
1063         goto fail;
1064 
1065     /* Do the actual checks */
1066     for(i = 0; i < s->l2_size; i++) {
1067         l2_entry = be64_to_cpu(l2_table[i]);
1068 
1069         switch (qcow2_get_cluster_type(l2_entry)) {
1070         case QCOW2_CLUSTER_COMPRESSED:
1071             /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1072             if (l2_entry & QCOW_OFLAG_COPIED) {
1073                 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1074                     "copied flag must never be set for compressed "
1075                     "clusters\n", l2_entry >> s->cluster_bits);
1076                 l2_entry &= ~QCOW_OFLAG_COPIED;
1077                 res->corruptions++;
1078             }
1079 
1080             /* Mark cluster as used */
1081             nb_csectors = ((l2_entry >> s->csize_shift) &
1082                            s->csize_mask) + 1;
1083             l2_entry &= s->cluster_offset_mask;
1084             inc_refcounts(bs, res, refcount_table, refcount_table_size,
1085                 l2_entry & ~511, nb_csectors * 512);
1086 
1087             if (flags & CHECK_FRAG_INFO) {
1088                 res->bfi.allocated_clusters++;
1089                 res->bfi.compressed_clusters++;
1090 
1091                 /* Compressed clusters are fragmented by nature.  Since they
1092                  * take up sub-sector space but we only have sector granularity
1093                  * I/O we need to re-read the same sectors even for adjacent
1094                  * compressed clusters.
1095                  */
1096                 res->bfi.fragmented_clusters++;
1097             }
1098             break;
1099 
1100         case QCOW2_CLUSTER_ZERO:
1101             if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1102                 break;
1103             }
1104             /* fall through */
1105 
1106         case QCOW2_CLUSTER_NORMAL:
1107         {
1108             uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1109 
1110             if (flags & CHECK_FRAG_INFO) {
1111                 res->bfi.allocated_clusters++;
1112                 if (next_contiguous_offset &&
1113                     offset != next_contiguous_offset) {
1114                     res->bfi.fragmented_clusters++;
1115                 }
1116                 next_contiguous_offset = offset + s->cluster_size;
1117             }
1118 
1119             /* Mark cluster as used */
1120             inc_refcounts(bs, res, refcount_table,refcount_table_size,
1121                 offset, s->cluster_size);
1122 
1123             /* Correct offsets are cluster aligned */
1124             if (offset & (s->cluster_size - 1)) {
1125                 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1126                     "properly aligned; L2 entry corrupted.\n", offset);
1127                 res->corruptions++;
1128             }
1129             break;
1130         }
1131 
1132         case QCOW2_CLUSTER_UNALLOCATED:
1133             break;
1134 
1135         default:
1136             abort();
1137         }
1138     }
1139 
1140     g_free(l2_table);
1141     return 0;
1142 
1143 fail:
1144     fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1145     g_free(l2_table);
1146     return -EIO;
1147 }
1148 
1149 /*
1150  * Increases the refcount for the L1 table, its L2 tables and all referenced
1151  * clusters in the given refcount table. While doing so, performs some checks
1152  * on L1 and L2 entries.
1153  *
1154  * Returns the number of errors found by the checks or -errno if an internal
1155  * error occurred.
1156  */
1157 static int check_refcounts_l1(BlockDriverState *bs,
1158                               BdrvCheckResult *res,
1159                               uint16_t *refcount_table,
1160                               int refcount_table_size,
1161                               int64_t l1_table_offset, int l1_size,
1162                               int flags)
1163 {
1164     BDRVQcowState *s = bs->opaque;
1165     uint64_t *l1_table, l2_offset, l1_size2;
1166     int i, ret;
1167 
1168     l1_size2 = l1_size * sizeof(uint64_t);
1169 
1170     /* Mark L1 table as used */
1171     inc_refcounts(bs, res, refcount_table, refcount_table_size,
1172         l1_table_offset, l1_size2);
1173 
1174     /* Read L1 table entries from disk */
1175     if (l1_size2 == 0) {
1176         l1_table = NULL;
1177     } else {
1178         l1_table = g_malloc(l1_size2);
1179         if (bdrv_pread(bs->file, l1_table_offset,
1180                        l1_table, l1_size2) != l1_size2)
1181             goto fail;
1182         for(i = 0;i < l1_size; i++)
1183             be64_to_cpus(&l1_table[i]);
1184     }
1185 
1186     /* Do the actual checks */
1187     for(i = 0; i < l1_size; i++) {
1188         l2_offset = l1_table[i];
1189         if (l2_offset) {
1190             /* Mark L2 table as used */
1191             l2_offset &= L1E_OFFSET_MASK;
1192             inc_refcounts(bs, res, refcount_table, refcount_table_size,
1193                 l2_offset, s->cluster_size);
1194 
1195             /* L2 tables are cluster aligned */
1196             if (l2_offset & (s->cluster_size - 1)) {
1197                 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1198                     "cluster aligned; L1 entry corrupted\n", l2_offset);
1199                 res->corruptions++;
1200             }
1201 
1202             /* Process and check L2 entries */
1203             ret = check_refcounts_l2(bs, res, refcount_table,
1204                                      refcount_table_size, l2_offset, flags);
1205             if (ret < 0) {
1206                 goto fail;
1207             }
1208         }
1209     }
1210     g_free(l1_table);
1211     return 0;
1212 
1213 fail:
1214     fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1215     res->check_errors++;
1216     g_free(l1_table);
1217     return -EIO;
1218 }
1219 
1220 /*
1221  * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1222  *
1223  * This function does not print an error message nor does it increment
1224  * check_errors if get_refcount fails (this is because such an error will have
1225  * been already detected and sufficiently signaled by the calling function
1226  * (qcow2_check_refcounts) by the time this function is called).
1227  */
1228 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1229                               BdrvCheckMode fix)
1230 {
1231     BDRVQcowState *s = bs->opaque;
1232     uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1233     int ret;
1234     int refcount;
1235     int i, j;
1236 
1237     for (i = 0; i < s->l1_size; i++) {
1238         uint64_t l1_entry = s->l1_table[i];
1239         uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1240         bool l2_dirty = false;
1241 
1242         if (!l2_offset) {
1243             continue;
1244         }
1245 
1246         refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
1247         if (refcount < 0) {
1248             /* don't print message nor increment check_errors */
1249             continue;
1250         }
1251         if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1252             fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1253                     "l1_entry=%" PRIx64 " refcount=%d\n",
1254                     fix & BDRV_FIX_ERRORS ? "Repairing" :
1255                                             "ERROR",
1256                     i, l1_entry, refcount);
1257             if (fix & BDRV_FIX_ERRORS) {
1258                 s->l1_table[i] = refcount == 1
1259                                ? l1_entry |  QCOW_OFLAG_COPIED
1260                                : l1_entry & ~QCOW_OFLAG_COPIED;
1261                 ret = qcow2_write_l1_entry(bs, i);
1262                 if (ret < 0) {
1263                     res->check_errors++;
1264                     goto fail;
1265                 }
1266                 res->corruptions_fixed++;
1267             } else {
1268                 res->corruptions++;
1269             }
1270         }
1271 
1272         ret = bdrv_pread(bs->file, l2_offset, l2_table,
1273                          s->l2_size * sizeof(uint64_t));
1274         if (ret < 0) {
1275             fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1276                     strerror(-ret));
1277             res->check_errors++;
1278             goto fail;
1279         }
1280 
1281         for (j = 0; j < s->l2_size; j++) {
1282             uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1283             uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1284             int cluster_type = qcow2_get_cluster_type(l2_entry);
1285 
1286             if ((cluster_type == QCOW2_CLUSTER_NORMAL) ||
1287                 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) {
1288                 refcount = get_refcount(bs, data_offset >> s->cluster_bits);
1289                 if (refcount < 0) {
1290                     /* don't print message nor increment check_errors */
1291                     continue;
1292                 }
1293                 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1294                     fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1295                             "l2_entry=%" PRIx64 " refcount=%d\n",
1296                             fix & BDRV_FIX_ERRORS ? "Repairing" :
1297                                                     "ERROR",
1298                             l2_entry, refcount);
1299                     if (fix & BDRV_FIX_ERRORS) {
1300                         l2_table[j] = cpu_to_be64(refcount == 1
1301                                     ? l2_entry |  QCOW_OFLAG_COPIED
1302                                     : l2_entry & ~QCOW_OFLAG_COPIED);
1303                         l2_dirty = true;
1304                         res->corruptions_fixed++;
1305                     } else {
1306                         res->corruptions++;
1307                     }
1308                 }
1309             }
1310         }
1311 
1312         if (l2_dirty) {
1313             ret = qcow2_pre_write_overlap_check(bs,
1314                     QCOW2_OL_DEFAULT & ~QCOW2_OL_ACTIVE_L2, l2_offset,
1315                     s->cluster_size);
1316             if (ret < 0) {
1317                 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1318                         "overlap check failed: %s\n", strerror(-ret));
1319                 res->check_errors++;
1320                 goto fail;
1321             }
1322 
1323             ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size);
1324             if (ret < 0) {
1325                 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1326                         strerror(-ret));
1327                 res->check_errors++;
1328                 goto fail;
1329             }
1330         }
1331     }
1332 
1333     ret = 0;
1334 
1335 fail:
1336     qemu_vfree(l2_table);
1337     return ret;
1338 }
1339 
1340 /*
1341  * Writes one sector of the refcount table to the disk
1342  */
1343 #define RT_ENTRIES_PER_SECTOR (512 / sizeof(uint64_t))
1344 static int write_reftable_entry(BlockDriverState *bs, int rt_index)
1345 {
1346     BDRVQcowState *s = bs->opaque;
1347     uint64_t buf[RT_ENTRIES_PER_SECTOR];
1348     int rt_start_index;
1349     int i, ret;
1350 
1351     rt_start_index = rt_index & ~(RT_ENTRIES_PER_SECTOR - 1);
1352     for (i = 0; i < RT_ENTRIES_PER_SECTOR; i++) {
1353         buf[i] = cpu_to_be64(s->refcount_table[rt_start_index + i]);
1354     }
1355 
1356     ret = qcow2_pre_write_overlap_check(bs,
1357             QCOW2_OL_DEFAULT & ~QCOW2_OL_REFCOUNT_TABLE,
1358             s->refcount_table_offset + rt_start_index * sizeof(uint64_t),
1359             sizeof(buf));
1360     if (ret < 0) {
1361         return ret;
1362     }
1363 
1364     BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_UPDATE);
1365     ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset +
1366             rt_start_index * sizeof(uint64_t), buf, sizeof(buf));
1367     if (ret < 0) {
1368         return ret;
1369     }
1370 
1371     return 0;
1372 }
1373 
1374 /*
1375  * Allocates a new cluster for the given refcount block (represented by its
1376  * offset in the image file) and copies the current content there. This function
1377  * does _not_ decrement the reference count for the currently occupied cluster.
1378  *
1379  * This function prints an informative message to stderr on error (and returns
1380  * -errno); on success, 0 is returned.
1381  */
1382 static int64_t realloc_refcount_block(BlockDriverState *bs, int reftable_index,
1383                                       uint64_t offset)
1384 {
1385     BDRVQcowState *s = bs->opaque;
1386     int64_t new_offset = 0;
1387     void *refcount_block = NULL;
1388     int ret;
1389 
1390     /* allocate new refcount block */
1391     new_offset = qcow2_alloc_clusters(bs, s->cluster_size);
1392     if (new_offset < 0) {
1393         fprintf(stderr, "Could not allocate new cluster: %s\n",
1394                 strerror(-new_offset));
1395         ret = new_offset;
1396         goto fail;
1397     }
1398 
1399     /* fetch current refcount block content */
1400     ret = qcow2_cache_get(bs, s->refcount_block_cache, offset, &refcount_block);
1401     if (ret < 0) {
1402         fprintf(stderr, "Could not fetch refcount block: %s\n", strerror(-ret));
1403         goto fail;
1404     }
1405 
1406     /* new block has not yet been entered into refcount table, therefore it is
1407      * no refcount block yet (regarding this check) */
1408     ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_DEFAULT, new_offset,
1409             s->cluster_size);
1410     if (ret < 0) {
1411         fprintf(stderr, "Could not write refcount block; metadata overlap "
1412                 "check failed: %s\n", strerror(-ret));
1413         /* the image will be marked corrupt, so don't even attempt on freeing
1414          * the cluster */
1415         new_offset = 0;
1416         goto fail;
1417     }
1418 
1419     /* write to new block */
1420     ret = bdrv_write(bs->file, new_offset / BDRV_SECTOR_SIZE, refcount_block,
1421             s->cluster_sectors);
1422     if (ret < 0) {
1423         fprintf(stderr, "Could not write refcount block: %s\n", strerror(-ret));
1424         goto fail;
1425     }
1426 
1427     /* update refcount table */
1428     assert(!(new_offset & (s->cluster_size - 1)));
1429     s->refcount_table[reftable_index] = new_offset;
1430     ret = write_reftable_entry(bs, reftable_index);
1431     if (ret < 0) {
1432         fprintf(stderr, "Could not update refcount table: %s\n",
1433                 strerror(-ret));
1434         goto fail;
1435     }
1436 
1437 fail:
1438     if (new_offset && (ret < 0)) {
1439         qcow2_free_clusters(bs, new_offset, s->cluster_size,
1440                 QCOW2_DISCARD_ALWAYS);
1441     }
1442     if (refcount_block) {
1443         if (ret < 0) {
1444             qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
1445         } else {
1446             ret = qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
1447         }
1448     }
1449     if (ret < 0) {
1450         return ret;
1451     }
1452     return new_offset;
1453 }
1454 
1455 /*
1456  * Checks an image for refcount consistency.
1457  *
1458  * Returns 0 if no errors are found, the number of errors in case the image is
1459  * detected as corrupted, and -errno when an internal error occurred.
1460  */
1461 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1462                           BdrvCheckMode fix)
1463 {
1464     BDRVQcowState *s = bs->opaque;
1465     int64_t size, i, highest_cluster;
1466     int nb_clusters, refcount1, refcount2;
1467     QCowSnapshot *sn;
1468     uint16_t *refcount_table;
1469     int ret;
1470 
1471     size = bdrv_getlength(bs->file);
1472     nb_clusters = size_to_clusters(s, size);
1473     refcount_table = g_malloc0(nb_clusters * sizeof(uint16_t));
1474 
1475     res->bfi.total_clusters =
1476         size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1477 
1478     /* header */
1479     inc_refcounts(bs, res, refcount_table, nb_clusters,
1480         0, s->cluster_size);
1481 
1482     /* current L1 table */
1483     ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1484                              s->l1_table_offset, s->l1_size,
1485                              CHECK_OFLAG_COPIED | CHECK_FRAG_INFO);
1486     if (ret < 0) {
1487         goto fail;
1488     }
1489 
1490     /* snapshots */
1491     for(i = 0; i < s->nb_snapshots; i++) {
1492         sn = s->snapshots + i;
1493         ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1494             sn->l1_table_offset, sn->l1_size, 0);
1495         if (ret < 0) {
1496             goto fail;
1497         }
1498     }
1499     inc_refcounts(bs, res, refcount_table, nb_clusters,
1500         s->snapshots_offset, s->snapshots_size);
1501 
1502     /* refcount data */
1503     inc_refcounts(bs, res, refcount_table, nb_clusters,
1504         s->refcount_table_offset,
1505         s->refcount_table_size * sizeof(uint64_t));
1506 
1507     for(i = 0; i < s->refcount_table_size; i++) {
1508         uint64_t offset, cluster;
1509         offset = s->refcount_table[i];
1510         cluster = offset >> s->cluster_bits;
1511 
1512         /* Refcount blocks are cluster aligned */
1513         if (offset & (s->cluster_size - 1)) {
1514             fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1515                 "cluster aligned; refcount table entry corrupted\n", i);
1516             res->corruptions++;
1517             continue;
1518         }
1519 
1520         if (cluster >= nb_clusters) {
1521             fprintf(stderr, "ERROR refcount block %" PRId64
1522                     " is outside image\n", i);
1523             res->corruptions++;
1524             continue;
1525         }
1526 
1527         if (offset != 0) {
1528             inc_refcounts(bs, res, refcount_table, nb_clusters,
1529                 offset, s->cluster_size);
1530             if (refcount_table[cluster] != 1) {
1531                 fprintf(stderr, "%s refcount block %" PRId64
1532                     " refcount=%d\n",
1533                     fix & BDRV_FIX_ERRORS ? "Repairing" :
1534                                             "ERROR",
1535                     i, refcount_table[cluster]);
1536 
1537                 if (fix & BDRV_FIX_ERRORS) {
1538                     int64_t new_offset;
1539 
1540                     new_offset = realloc_refcount_block(bs, i, offset);
1541                     if (new_offset < 0) {
1542                         res->corruptions++;
1543                         continue;
1544                     }
1545 
1546                     /* update refcounts */
1547                     if ((new_offset >> s->cluster_bits) >= nb_clusters) {
1548                         /* increase refcount_table size if necessary */
1549                         int old_nb_clusters = nb_clusters;
1550                         nb_clusters = (new_offset >> s->cluster_bits) + 1;
1551                         refcount_table = g_realloc(refcount_table,
1552                                 nb_clusters * sizeof(uint16_t));
1553                         memset(&refcount_table[old_nb_clusters], 0, (nb_clusters
1554                                 - old_nb_clusters) * sizeof(uint16_t));
1555                     }
1556                     refcount_table[cluster]--;
1557                     inc_refcounts(bs, res, refcount_table, nb_clusters,
1558                             new_offset, s->cluster_size);
1559 
1560                     res->corruptions_fixed++;
1561                 } else {
1562                     res->corruptions++;
1563                 }
1564             }
1565         }
1566     }
1567 
1568     /* compare ref counts */
1569     for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1570         refcount1 = get_refcount(bs, i);
1571         if (refcount1 < 0) {
1572             fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1573                 i, strerror(-refcount1));
1574             res->check_errors++;
1575             continue;
1576         }
1577 
1578         refcount2 = refcount_table[i];
1579 
1580         if (refcount1 > 0 || refcount2 > 0) {
1581             highest_cluster = i;
1582         }
1583 
1584         if (refcount1 != refcount2) {
1585 
1586             /* Check if we're allowed to fix the mismatch */
1587             int *num_fixed = NULL;
1588             if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1589                 num_fixed = &res->leaks_fixed;
1590             } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1591                 num_fixed = &res->corruptions_fixed;
1592             }
1593 
1594             fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1595                    num_fixed != NULL     ? "Repairing" :
1596                    refcount1 < refcount2 ? "ERROR" :
1597                                            "Leaked",
1598                    i, refcount1, refcount2);
1599 
1600             if (num_fixed) {
1601                 ret = update_refcount(bs, i << s->cluster_bits, 1,
1602                                       refcount2 - refcount1,
1603                                       QCOW2_DISCARD_ALWAYS);
1604                 if (ret >= 0) {
1605                     (*num_fixed)++;
1606                     continue;
1607                 }
1608             }
1609 
1610             /* And if we couldn't, print an error */
1611             if (refcount1 < refcount2) {
1612                 res->corruptions++;
1613             } else {
1614                 res->leaks++;
1615             }
1616         }
1617     }
1618 
1619     /* check OFLAG_COPIED */
1620     ret = check_oflag_copied(bs, res, fix);
1621     if (ret < 0) {
1622         goto fail;
1623     }
1624 
1625     res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1626     ret = 0;
1627 
1628 fail:
1629     g_free(refcount_table);
1630 
1631     return ret;
1632 }
1633 
1634 #define overlaps_with(ofs, sz) \
1635     ranges_overlap(offset, size, ofs, sz)
1636 
1637 /*
1638  * Checks if the given offset into the image file is actually free to use by
1639  * looking for overlaps with important metadata sections (L1/L2 tables etc.),
1640  * i.e. a sanity check without relying on the refcount tables.
1641  *
1642  * The chk parameter specifies exactly what checks to perform (being a bitmask
1643  * of QCow2MetadataOverlap values).
1644  *
1645  * Returns:
1646  * - 0 if writing to this offset will not affect the mentioned metadata
1647  * - a positive QCow2MetadataOverlap value indicating one overlapping section
1648  * - a negative value (-errno) indicating an error while performing a check,
1649  *   e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
1650  */
1651 int qcow2_check_metadata_overlap(BlockDriverState *bs, int chk, int64_t offset,
1652                                  int64_t size)
1653 {
1654     BDRVQcowState *s = bs->opaque;
1655     int i, j;
1656 
1657     if (!size) {
1658         return 0;
1659     }
1660 
1661     if (chk & QCOW2_OL_MAIN_HEADER) {
1662         if (offset < s->cluster_size) {
1663             return QCOW2_OL_MAIN_HEADER;
1664         }
1665     }
1666 
1667     /* align range to test to cluster boundaries */
1668     size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size);
1669     offset = start_of_cluster(s, offset);
1670 
1671     if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
1672         if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
1673             return QCOW2_OL_ACTIVE_L1;
1674         }
1675     }
1676 
1677     if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
1678         if (overlaps_with(s->refcount_table_offset,
1679             s->refcount_table_size * sizeof(uint64_t))) {
1680             return QCOW2_OL_REFCOUNT_TABLE;
1681         }
1682     }
1683 
1684     if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
1685         if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
1686             return QCOW2_OL_SNAPSHOT_TABLE;
1687         }
1688     }
1689 
1690     if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
1691         for (i = 0; i < s->nb_snapshots; i++) {
1692             if (s->snapshots[i].l1_size &&
1693                 overlaps_with(s->snapshots[i].l1_table_offset,
1694                 s->snapshots[i].l1_size * sizeof(uint64_t))) {
1695                 return QCOW2_OL_INACTIVE_L1;
1696             }
1697         }
1698     }
1699 
1700     if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
1701         for (i = 0; i < s->l1_size; i++) {
1702             if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
1703                 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
1704                 s->cluster_size)) {
1705                 return QCOW2_OL_ACTIVE_L2;
1706             }
1707         }
1708     }
1709 
1710     if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
1711         for (i = 0; i < s->refcount_table_size; i++) {
1712             if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
1713                 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
1714                 s->cluster_size)) {
1715                 return QCOW2_OL_REFCOUNT_BLOCK;
1716             }
1717         }
1718     }
1719 
1720     if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
1721         for (i = 0; i < s->nb_snapshots; i++) {
1722             uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
1723             uint32_t l1_sz  = s->snapshots[i].l1_size;
1724             uint64_t *l1 = g_malloc(l1_sz * sizeof(uint64_t));
1725             int ret;
1726 
1727             ret = bdrv_read(bs->file, l1_ofs / BDRV_SECTOR_SIZE, (uint8_t *)l1,
1728                             l1_sz * sizeof(uint64_t) / BDRV_SECTOR_SIZE);
1729 
1730             if (ret < 0) {
1731                 g_free(l1);
1732                 return ret;
1733             }
1734 
1735             for (j = 0; j < l1_sz; j++) {
1736                 if ((l1[j] & L1E_OFFSET_MASK) &&
1737                     overlaps_with(l1[j] & L1E_OFFSET_MASK, s->cluster_size)) {
1738                     g_free(l1);
1739                     return QCOW2_OL_INACTIVE_L2;
1740                 }
1741             }
1742 
1743             g_free(l1);
1744         }
1745     }
1746 
1747     return 0;
1748 }
1749 
1750 static const char *metadata_ol_names[] = {
1751     [QCOW2_OL_MAIN_HEADER_BITNR]    = "qcow2_header",
1752     [QCOW2_OL_ACTIVE_L1_BITNR]      = "active L1 table",
1753     [QCOW2_OL_ACTIVE_L2_BITNR]      = "active L2 table",
1754     [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
1755     [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
1756     [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
1757     [QCOW2_OL_INACTIVE_L1_BITNR]    = "inactive L1 table",
1758     [QCOW2_OL_INACTIVE_L2_BITNR]    = "inactive L2 table",
1759 };
1760 
1761 /*
1762  * First performs a check for metadata overlaps (through
1763  * qcow2_check_metadata_overlap); if that fails with a negative value (error
1764  * while performing a check), that value is returned. If an impending overlap
1765  * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
1766  * and -EIO returned.
1767  *
1768  * Returns 0 if there were neither overlaps nor errors while checking for
1769  * overlaps; or a negative value (-errno) on error.
1770  */
1771 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int chk, int64_t offset,
1772                                   int64_t size)
1773 {
1774     int ret = qcow2_check_metadata_overlap(bs, chk, offset, size);
1775 
1776     if (ret < 0) {
1777         return ret;
1778     } else if (ret > 0) {
1779         int metadata_ol_bitnr = ffs(ret) - 1;
1780         char *message;
1781         QObject *data;
1782 
1783         assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
1784 
1785         fprintf(stderr, "qcow2: Preventing invalid write on metadata (overlaps "
1786                 "with %s); image marked as corrupt.\n",
1787                 metadata_ol_names[metadata_ol_bitnr]);
1788         message = g_strdup_printf("Prevented %s overwrite",
1789                 metadata_ol_names[metadata_ol_bitnr]);
1790         data = qobject_from_jsonf("{ 'device': %s, 'msg': %s, 'offset': %"
1791                 PRId64 ", 'size': %" PRId64 " }", bs->device_name, message,
1792                 offset, size);
1793         monitor_protocol_event(QEVENT_BLOCK_IMAGE_CORRUPTED, data);
1794         g_free(message);
1795         qobject_decref(data);
1796 
1797         qcow2_mark_corrupt(bs);
1798         bs->drv = NULL; /* make BDS unusable */
1799         return -EIO;
1800     }
1801 
1802     return 0;
1803 }
1804