Add qemu 2.4.0
[kvmfornfv.git] / qemu / block / qcow2-refcount.c
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, uint64_t addend,
33                             bool decrease, enum qcow2_discard_type type);
34
35 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index);
36 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index);
37 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index);
38 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index);
39 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index);
40 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index);
41 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index);
42
43 static void set_refcount_ro0(void *refcount_array, uint64_t index,
44                              uint64_t value);
45 static void set_refcount_ro1(void *refcount_array, uint64_t index,
46                              uint64_t value);
47 static void set_refcount_ro2(void *refcount_array, uint64_t index,
48                              uint64_t value);
49 static void set_refcount_ro3(void *refcount_array, uint64_t index,
50                              uint64_t value);
51 static void set_refcount_ro4(void *refcount_array, uint64_t index,
52                              uint64_t value);
53 static void set_refcount_ro5(void *refcount_array, uint64_t index,
54                              uint64_t value);
55 static void set_refcount_ro6(void *refcount_array, uint64_t index,
56                              uint64_t value);
57
58
59 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = {
60     &get_refcount_ro0,
61     &get_refcount_ro1,
62     &get_refcount_ro2,
63     &get_refcount_ro3,
64     &get_refcount_ro4,
65     &get_refcount_ro5,
66     &get_refcount_ro6
67 };
68
69 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = {
70     &set_refcount_ro0,
71     &set_refcount_ro1,
72     &set_refcount_ro2,
73     &set_refcount_ro3,
74     &set_refcount_ro4,
75     &set_refcount_ro5,
76     &set_refcount_ro6
77 };
78
79
80 /*********************************************************/
81 /* refcount handling */
82
83 int qcow2_refcount_init(BlockDriverState *bs)
84 {
85     BDRVQcowState *s = bs->opaque;
86     unsigned int refcount_table_size2, i;
87     int ret;
88
89     assert(s->refcount_order >= 0 && s->refcount_order <= 6);
90
91     s->get_refcount = get_refcount_funcs[s->refcount_order];
92     s->set_refcount = set_refcount_funcs[s->refcount_order];
93
94     assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
95     refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
96     s->refcount_table = g_try_malloc(refcount_table_size2);
97
98     if (s->refcount_table_size > 0) {
99         if (s->refcount_table == NULL) {
100             ret = -ENOMEM;
101             goto fail;
102         }
103         BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
104         ret = bdrv_pread(bs->file, s->refcount_table_offset,
105                          s->refcount_table, refcount_table_size2);
106         if (ret < 0) {
107             goto fail;
108         }
109         for(i = 0; i < s->refcount_table_size; i++)
110             be64_to_cpus(&s->refcount_table[i]);
111     }
112     return 0;
113  fail:
114     return ret;
115 }
116
117 void qcow2_refcount_close(BlockDriverState *bs)
118 {
119     BDRVQcowState *s = bs->opaque;
120     g_free(s->refcount_table);
121 }
122
123
124 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index)
125 {
126     return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1;
127 }
128
129 static void set_refcount_ro0(void *refcount_array, uint64_t index,
130                              uint64_t value)
131 {
132     assert(!(value >> 1));
133     ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8));
134     ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8);
135 }
136
137 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index)
138 {
139     return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4)))
140            & 0x3;
141 }
142
143 static void set_refcount_ro1(void *refcount_array, uint64_t index,
144                              uint64_t value)
145 {
146     assert(!(value >> 2));
147     ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4)));
148     ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4));
149 }
150
151 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index)
152 {
153     return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2)))
154            & 0xf;
155 }
156
157 static void set_refcount_ro2(void *refcount_array, uint64_t index,
158                              uint64_t value)
159 {
160     assert(!(value >> 4));
161     ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2)));
162     ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2));
163 }
164
165 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index)
166 {
167     return ((const uint8_t *)refcount_array)[index];
168 }
169
170 static void set_refcount_ro3(void *refcount_array, uint64_t index,
171                              uint64_t value)
172 {
173     assert(!(value >> 8));
174     ((uint8_t *)refcount_array)[index] = value;
175 }
176
177 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index)
178 {
179     return be16_to_cpu(((const uint16_t *)refcount_array)[index]);
180 }
181
182 static void set_refcount_ro4(void *refcount_array, uint64_t index,
183                              uint64_t value)
184 {
185     assert(!(value >> 16));
186     ((uint16_t *)refcount_array)[index] = cpu_to_be16(value);
187 }
188
189 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index)
190 {
191     return be32_to_cpu(((const uint32_t *)refcount_array)[index]);
192 }
193
194 static void set_refcount_ro5(void *refcount_array, uint64_t index,
195                              uint64_t value)
196 {
197     assert(!(value >> 32));
198     ((uint32_t *)refcount_array)[index] = cpu_to_be32(value);
199 }
200
201 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index)
202 {
203     return be64_to_cpu(((const uint64_t *)refcount_array)[index]);
204 }
205
206 static void set_refcount_ro6(void *refcount_array, uint64_t index,
207                              uint64_t value)
208 {
209     ((uint64_t *)refcount_array)[index] = cpu_to_be64(value);
210 }
211
212
213 static int load_refcount_block(BlockDriverState *bs,
214                                int64_t refcount_block_offset,
215                                void **refcount_block)
216 {
217     BDRVQcowState *s = bs->opaque;
218     int ret;
219
220     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
221     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
222         refcount_block);
223
224     return ret;
225 }
226
227 /*
228  * Retrieves the refcount of the cluster given by its index and stores it in
229  * *refcount. Returns 0 on success and -errno on failure.
230  */
231 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index,
232                        uint64_t *refcount)
233 {
234     BDRVQcowState *s = bs->opaque;
235     uint64_t refcount_table_index, block_index;
236     int64_t refcount_block_offset;
237     int ret;
238     void *refcount_block;
239
240     refcount_table_index = cluster_index >> s->refcount_block_bits;
241     if (refcount_table_index >= s->refcount_table_size) {
242         *refcount = 0;
243         return 0;
244     }
245     refcount_block_offset =
246         s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
247     if (!refcount_block_offset) {
248         *refcount = 0;
249         return 0;
250     }
251
252     if (offset_into_cluster(s, refcount_block_offset)) {
253         qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64
254                                 " unaligned (reftable index: %#" PRIx64 ")",
255                                 refcount_block_offset, refcount_table_index);
256         return -EIO;
257     }
258
259     ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
260                           &refcount_block);
261     if (ret < 0) {
262         return ret;
263     }
264
265     block_index = cluster_index & (s->refcount_block_size - 1);
266     *refcount = s->get_refcount(refcount_block, block_index);
267
268     qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
269
270     return 0;
271 }
272
273 /*
274  * Rounds the refcount table size up to avoid growing the table for each single
275  * refcount block that is allocated.
276  */
277 static unsigned int next_refcount_table_size(BDRVQcowState *s,
278     unsigned int min_size)
279 {
280     unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
281     unsigned int refcount_table_clusters =
282         MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
283
284     while (min_clusters > refcount_table_clusters) {
285         refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
286     }
287
288     return refcount_table_clusters << (s->cluster_bits - 3);
289 }
290
291
292 /* Checks if two offsets are described by the same refcount block */
293 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
294     uint64_t offset_b)
295 {
296     uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits);
297     uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits);
298
299     return (block_a == block_b);
300 }
301
302 /*
303  * Loads a refcount block. If it doesn't exist yet, it is allocated first
304  * (including growing the refcount table if needed).
305  *
306  * Returns 0 on success or -errno in error case
307  */
308 static int alloc_refcount_block(BlockDriverState *bs,
309                                 int64_t cluster_index, void **refcount_block)
310 {
311     BDRVQcowState *s = bs->opaque;
312     unsigned int refcount_table_index;
313     int ret;
314
315     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
316
317     /* Find the refcount block for the given cluster */
318     refcount_table_index = cluster_index >> s->refcount_block_bits;
319
320     if (refcount_table_index < s->refcount_table_size) {
321
322         uint64_t refcount_block_offset =
323             s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
324
325         /* If it's already there, we're done */
326         if (refcount_block_offset) {
327             if (offset_into_cluster(s, refcount_block_offset)) {
328                 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
329                                         PRIx64 " unaligned (reftable index: "
330                                         "%#x)", refcount_block_offset,
331                                         refcount_table_index);
332                 return -EIO;
333             }
334
335              return load_refcount_block(bs, refcount_block_offset,
336                                         refcount_block);
337         }
338     }
339
340     /*
341      * If we came here, we need to allocate something. Something is at least
342      * a cluster for the new refcount block. It may also include a new refcount
343      * table if the old refcount table is too small.
344      *
345      * Note that allocating clusters here needs some special care:
346      *
347      * - We can't use the normal qcow2_alloc_clusters(), it would try to
348      *   increase the refcount and very likely we would end up with an endless
349      *   recursion. Instead we must place the refcount blocks in a way that
350      *   they can describe them themselves.
351      *
352      * - We need to consider that at this point we are inside update_refcounts
353      *   and potentially doing an initial refcount increase. This means that
354      *   some clusters have already been allocated by the caller, but their
355      *   refcount isn't accurate yet. If we allocate clusters for metadata, we
356      *   need to return -EAGAIN to signal the caller that it needs to restart
357      *   the search for free clusters.
358      *
359      * - alloc_clusters_noref and qcow2_free_clusters may load a different
360      *   refcount block into the cache
361      */
362
363     *refcount_block = NULL;
364
365     /* We write to the refcount table, so we might depend on L2 tables */
366     ret = qcow2_cache_flush(bs, s->l2_table_cache);
367     if (ret < 0) {
368         return ret;
369     }
370
371     /* Allocate the refcount block itself and mark it as used */
372     int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
373     if (new_block < 0) {
374         return new_block;
375     }
376
377 #ifdef DEBUG_ALLOC2
378     fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
379         " at %" PRIx64 "\n",
380         refcount_table_index, cluster_index << s->cluster_bits, new_block);
381 #endif
382
383     if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
384         /* Zero the new refcount block before updating it */
385         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
386                                     refcount_block);
387         if (ret < 0) {
388             goto fail_block;
389         }
390
391         memset(*refcount_block, 0, s->cluster_size);
392
393         /* The block describes itself, need to update the cache */
394         int block_index = (new_block >> s->cluster_bits) &
395             (s->refcount_block_size - 1);
396         s->set_refcount(*refcount_block, block_index, 1);
397     } else {
398         /* Described somewhere else. This can recurse at most twice before we
399          * arrive at a block that describes itself. */
400         ret = update_refcount(bs, new_block, s->cluster_size, 1, false,
401                               QCOW2_DISCARD_NEVER);
402         if (ret < 0) {
403             goto fail_block;
404         }
405
406         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
407         if (ret < 0) {
408             goto fail_block;
409         }
410
411         /* Initialize the new refcount block only after updating its refcount,
412          * update_refcount uses the refcount cache itself */
413         ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
414                                     refcount_block);
415         if (ret < 0) {
416             goto fail_block;
417         }
418
419         memset(*refcount_block, 0, s->cluster_size);
420     }
421
422     /* Now the new refcount block needs to be written to disk */
423     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
424     qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, *refcount_block);
425     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
426     if (ret < 0) {
427         goto fail_block;
428     }
429
430     /* If the refcount table is big enough, just hook the block up there */
431     if (refcount_table_index < s->refcount_table_size) {
432         uint64_t data64 = cpu_to_be64(new_block);
433         BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
434         ret = bdrv_pwrite_sync(bs->file,
435             s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
436             &data64, sizeof(data64));
437         if (ret < 0) {
438             goto fail_block;
439         }
440
441         s->refcount_table[refcount_table_index] = new_block;
442
443         /* The new refcount block may be where the caller intended to put its
444          * data, so let it restart the search. */
445         return -EAGAIN;
446     }
447
448     qcow2_cache_put(bs, s->refcount_block_cache, refcount_block);
449
450     /*
451      * If we come here, we need to grow the refcount table. Again, a new
452      * refcount table needs some space and we can't simply allocate to avoid
453      * endless recursion.
454      *
455      * Therefore let's grab new refcount blocks at the end of the image, which
456      * will describe themselves and the new refcount table. This way we can
457      * reference them only in the new table and do the switch to the new
458      * refcount table at once without producing an inconsistent state in
459      * between.
460      */
461     BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
462
463     /* Calculate the number of refcount blocks needed so far; this will be the
464      * basis for calculating the index of the first cluster used for the
465      * self-describing refcount structures which we are about to create.
466      *
467      * Because we reached this point, there cannot be any refcount entries for
468      * cluster_index or higher indices yet. However, because new_block has been
469      * allocated to describe that cluster (and it will assume this role later
470      * on), we cannot use that index; also, new_block may actually have a higher
471      * cluster index than cluster_index, so it needs to be taken into account
472      * here (and 1 needs to be added to its value because that cluster is used).
473      */
474     uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1,
475                                             (new_block >> s->cluster_bits) + 1),
476                                         s->refcount_block_size);
477
478     if (blocks_used > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
479         return -EFBIG;
480     }
481
482     /* And now we need at least one block more for the new metadata */
483     uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
484     uint64_t last_table_size;
485     uint64_t blocks_clusters;
486     do {
487         uint64_t table_clusters =
488             size_to_clusters(s, table_size * sizeof(uint64_t));
489         blocks_clusters = 1 +
490             ((table_clusters + s->refcount_block_size - 1)
491             / s->refcount_block_size);
492         uint64_t meta_clusters = table_clusters + blocks_clusters;
493
494         last_table_size = table_size;
495         table_size = next_refcount_table_size(s, blocks_used +
496             ((meta_clusters + s->refcount_block_size - 1)
497             / s->refcount_block_size));
498
499     } while (last_table_size != table_size);
500
501 #ifdef DEBUG_ALLOC2
502     fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
503         s->refcount_table_size, table_size);
504 #endif
505
506     /* Create the new refcount table and blocks */
507     uint64_t meta_offset = (blocks_used * s->refcount_block_size) *
508         s->cluster_size;
509     uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
510     uint64_t *new_table = g_try_new0(uint64_t, table_size);
511     void *new_blocks = g_try_malloc0(blocks_clusters * s->cluster_size);
512
513     assert(table_size > 0 && blocks_clusters > 0);
514     if (new_table == NULL || new_blocks == NULL) {
515         ret = -ENOMEM;
516         goto fail_table;
517     }
518
519     /* Fill the new refcount table */
520     memcpy(new_table, s->refcount_table,
521         s->refcount_table_size * sizeof(uint64_t));
522     new_table[refcount_table_index] = new_block;
523
524     int i;
525     for (i = 0; i < blocks_clusters; i++) {
526         new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
527     }
528
529     /* Fill the refcount blocks */
530     uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
531     int block = 0;
532     for (i = 0; i < table_clusters + blocks_clusters; i++) {
533         s->set_refcount(new_blocks, block++, 1);
534     }
535
536     /* Write refcount blocks to disk */
537     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
538     ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
539         blocks_clusters * s->cluster_size);
540     g_free(new_blocks);
541     new_blocks = NULL;
542     if (ret < 0) {
543         goto fail_table;
544     }
545
546     /* Write refcount table to disk */
547     for(i = 0; i < table_size; i++) {
548         cpu_to_be64s(&new_table[i]);
549     }
550
551     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
552     ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
553         table_size * sizeof(uint64_t));
554     if (ret < 0) {
555         goto fail_table;
556     }
557
558     for(i = 0; i < table_size; i++) {
559         be64_to_cpus(&new_table[i]);
560     }
561
562     /* Hook up the new refcount table in the qcow2 header */
563     uint8_t data[12];
564     cpu_to_be64w((uint64_t*)data, table_offset);
565     cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
566     BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
567     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
568         data, sizeof(data));
569     if (ret < 0) {
570         goto fail_table;
571     }
572
573     /* And switch it in memory */
574     uint64_t old_table_offset = s->refcount_table_offset;
575     uint64_t old_table_size = s->refcount_table_size;
576
577     g_free(s->refcount_table);
578     s->refcount_table = new_table;
579     s->refcount_table_size = table_size;
580     s->refcount_table_offset = table_offset;
581
582     /* Free old table. */
583     qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
584                         QCOW2_DISCARD_OTHER);
585
586     ret = load_refcount_block(bs, new_block, refcount_block);
587     if (ret < 0) {
588         return ret;
589     }
590
591     /* If we were trying to do the initial refcount update for some cluster
592      * allocation, we might have used the same clusters to store newly
593      * allocated metadata. Make the caller search some new space. */
594     return -EAGAIN;
595
596 fail_table:
597     g_free(new_blocks);
598     g_free(new_table);
599 fail_block:
600     if (*refcount_block != NULL) {
601         qcow2_cache_put(bs, s->refcount_block_cache, refcount_block);
602     }
603     return ret;
604 }
605
606 void qcow2_process_discards(BlockDriverState *bs, int ret)
607 {
608     BDRVQcowState *s = bs->opaque;
609     Qcow2DiscardRegion *d, *next;
610
611     QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
612         QTAILQ_REMOVE(&s->discards, d, next);
613
614         /* Discard is optional, ignore the return value */
615         if (ret >= 0) {
616             bdrv_discard(bs->file,
617                          d->offset >> BDRV_SECTOR_BITS,
618                          d->bytes >> BDRV_SECTOR_BITS);
619         }
620
621         g_free(d);
622     }
623 }
624
625 static void update_refcount_discard(BlockDriverState *bs,
626                                     uint64_t offset, uint64_t length)
627 {
628     BDRVQcowState *s = bs->opaque;
629     Qcow2DiscardRegion *d, *p, *next;
630
631     QTAILQ_FOREACH(d, &s->discards, next) {
632         uint64_t new_start = MIN(offset, d->offset);
633         uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
634
635         if (new_end - new_start <= length + d->bytes) {
636             /* There can't be any overlap, areas ending up here have no
637              * references any more and therefore shouldn't get freed another
638              * time. */
639             assert(d->bytes + length == new_end - new_start);
640             d->offset = new_start;
641             d->bytes = new_end - new_start;
642             goto found;
643         }
644     }
645
646     d = g_malloc(sizeof(*d));
647     *d = (Qcow2DiscardRegion) {
648         .bs     = bs,
649         .offset = offset,
650         .bytes  = length,
651     };
652     QTAILQ_INSERT_TAIL(&s->discards, d, next);
653
654 found:
655     /* Merge discard requests if they are adjacent now */
656     QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
657         if (p == d
658             || p->offset > d->offset + d->bytes
659             || d->offset > p->offset + p->bytes)
660         {
661             continue;
662         }
663
664         /* Still no overlap possible */
665         assert(p->offset == d->offset + d->bytes
666             || d->offset == p->offset + p->bytes);
667
668         QTAILQ_REMOVE(&s->discards, p, next);
669         d->offset = MIN(d->offset, p->offset);
670         d->bytes += p->bytes;
671         g_free(p);
672     }
673 }
674
675 /* XXX: cache several refcount block clusters ? */
676 /* @addend is the absolute value of the addend; if @decrease is set, @addend
677  * will be subtracted from the current refcount, otherwise it will be added */
678 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
679                                                    int64_t offset,
680                                                    int64_t length,
681                                                    uint64_t addend,
682                                                    bool decrease,
683                                                    enum qcow2_discard_type type)
684 {
685     BDRVQcowState *s = bs->opaque;
686     int64_t start, last, cluster_offset;
687     void *refcount_block = NULL;
688     int64_t old_table_index = -1;
689     int ret;
690
691 #ifdef DEBUG_ALLOC2
692     fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64
693             " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "",
694             addend);
695 #endif
696     if (length < 0) {
697         return -EINVAL;
698     } else if (length == 0) {
699         return 0;
700     }
701
702     if (decrease) {
703         qcow2_cache_set_dependency(bs, s->refcount_block_cache,
704             s->l2_table_cache);
705     }
706
707     start = start_of_cluster(s, offset);
708     last = start_of_cluster(s, offset + length - 1);
709     for(cluster_offset = start; cluster_offset <= last;
710         cluster_offset += s->cluster_size)
711     {
712         int block_index;
713         uint64_t refcount;
714         int64_t cluster_index = cluster_offset >> s->cluster_bits;
715         int64_t table_index = cluster_index >> s->refcount_block_bits;
716
717         /* Load the refcount block and allocate it if needed */
718         if (table_index != old_table_index) {
719             if (refcount_block) {
720                 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
721             }
722             ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
723             if (ret < 0) {
724                 goto fail;
725             }
726         }
727         old_table_index = table_index;
728
729         qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache,
730                                      refcount_block);
731
732         /* we can update the count and save it */
733         block_index = cluster_index & (s->refcount_block_size - 1);
734
735         refcount = s->get_refcount(refcount_block, block_index);
736         if (decrease ? (refcount - addend > refcount)
737                      : (refcount + addend < refcount ||
738                         refcount + addend > s->refcount_max))
739         {
740             ret = -EINVAL;
741             goto fail;
742         }
743         if (decrease) {
744             refcount -= addend;
745         } else {
746             refcount += addend;
747         }
748         if (refcount == 0 && cluster_index < s->free_cluster_index) {
749             s->free_cluster_index = cluster_index;
750         }
751         s->set_refcount(refcount_block, block_index, refcount);
752
753         if (refcount == 0 && s->discard_passthrough[type]) {
754             update_refcount_discard(bs, cluster_offset, s->cluster_size);
755         }
756     }
757
758     ret = 0;
759 fail:
760     if (!s->cache_discards) {
761         qcow2_process_discards(bs, ret);
762     }
763
764     /* Write last changed block to disk */
765     if (refcount_block) {
766         qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block);
767     }
768
769     /*
770      * Try do undo any updates if an error is returned (This may succeed in
771      * some cases like ENOSPC for allocating a new refcount block)
772      */
773     if (ret < 0) {
774         int dummy;
775         dummy = update_refcount(bs, offset, cluster_offset - offset, addend,
776                                 !decrease, QCOW2_DISCARD_NEVER);
777         (void)dummy;
778     }
779
780     return ret;
781 }
782
783 /*
784  * Increases or decreases the refcount of a given cluster.
785  *
786  * @addend is the absolute value of the addend; if @decrease is set, @addend
787  * will be subtracted from the current refcount, otherwise it will be added.
788  *
789  * On success 0 is returned; on failure -errno is returned.
790  */
791 int qcow2_update_cluster_refcount(BlockDriverState *bs,
792                                   int64_t cluster_index,
793                                   uint64_t addend, bool decrease,
794                                   enum qcow2_discard_type type)
795 {
796     BDRVQcowState *s = bs->opaque;
797     int ret;
798
799     ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
800                           decrease, type);
801     if (ret < 0) {
802         return ret;
803     }
804
805     return 0;
806 }
807
808
809
810 /*********************************************************/
811 /* cluster allocation functions */
812
813
814
815 /* return < 0 if error */
816 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size)
817 {
818     BDRVQcowState *s = bs->opaque;
819     uint64_t i, nb_clusters, refcount;
820     int ret;
821
822     /* We can't allocate clusters if they may still be queued for discard. */
823     if (s->cache_discards) {
824         qcow2_process_discards(bs, 0);
825     }
826
827     nb_clusters = size_to_clusters(s, size);
828 retry:
829     for(i = 0; i < nb_clusters; i++) {
830         uint64_t next_cluster_index = s->free_cluster_index++;
831         ret = qcow2_get_refcount(bs, next_cluster_index, &refcount);
832
833         if (ret < 0) {
834             return ret;
835         } else if (refcount != 0) {
836             goto retry;
837         }
838     }
839
840     /* Make sure that all offsets in the "allocated" range are representable
841      * in an int64_t */
842     if (s->free_cluster_index > 0 &&
843         s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits))
844     {
845         return -EFBIG;
846     }
847
848 #ifdef DEBUG_ALLOC2
849     fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
850             size,
851             (s->free_cluster_index - nb_clusters) << s->cluster_bits);
852 #endif
853     return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
854 }
855
856 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
857 {
858     int64_t offset;
859     int ret;
860
861     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
862     do {
863         offset = alloc_clusters_noref(bs, size);
864         if (offset < 0) {
865             return offset;
866         }
867
868         ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
869     } while (ret == -EAGAIN);
870
871     if (ret < 0) {
872         return ret;
873     }
874
875     return offset;
876 }
877
878 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
879     int nb_clusters)
880 {
881     BDRVQcowState *s = bs->opaque;
882     uint64_t cluster_index, refcount;
883     uint64_t i;
884     int ret;
885
886     assert(nb_clusters >= 0);
887     if (nb_clusters == 0) {
888         return 0;
889     }
890
891     do {
892         /* Check how many clusters there are free */
893         cluster_index = offset >> s->cluster_bits;
894         for(i = 0; i < nb_clusters; i++) {
895             ret = qcow2_get_refcount(bs, cluster_index++, &refcount);
896             if (ret < 0) {
897                 return ret;
898             } else if (refcount != 0) {
899                 break;
900             }
901         }
902
903         /* And then allocate them */
904         ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false,
905                               QCOW2_DISCARD_NEVER);
906     } while (ret == -EAGAIN);
907
908     if (ret < 0) {
909         return ret;
910     }
911
912     return i;
913 }
914
915 /* only used to allocate compressed sectors. We try to allocate
916    contiguous sectors. size must be <= cluster_size */
917 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
918 {
919     BDRVQcowState *s = bs->opaque;
920     int64_t offset;
921     size_t free_in_cluster;
922     int ret;
923
924     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
925     assert(size > 0 && size <= s->cluster_size);
926     assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset));
927
928     offset = s->free_byte_offset;
929
930     if (offset) {
931         uint64_t refcount;
932         ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount);
933         if (ret < 0) {
934             return ret;
935         }
936
937         if (refcount == s->refcount_max) {
938             offset = 0;
939         }
940     }
941
942     free_in_cluster = s->cluster_size - offset_into_cluster(s, offset);
943     do {
944         if (!offset || free_in_cluster < size) {
945             int64_t new_cluster = alloc_clusters_noref(bs, s->cluster_size);
946             if (new_cluster < 0) {
947                 return new_cluster;
948             }
949
950             if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) {
951                 offset = new_cluster;
952             }
953         }
954
955         assert(offset);
956         ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
957     } while (ret == -EAGAIN);
958     if (ret < 0) {
959         return ret;
960     }
961
962     /* The cluster refcount was incremented; refcount blocks must be flushed
963      * before the caller's L2 table updates. */
964     qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
965
966     s->free_byte_offset = offset + size;
967     if (!offset_into_cluster(s, s->free_byte_offset)) {
968         s->free_byte_offset = 0;
969     }
970
971     return offset;
972 }
973
974 void qcow2_free_clusters(BlockDriverState *bs,
975                           int64_t offset, int64_t size,
976                           enum qcow2_discard_type type)
977 {
978     int ret;
979
980     BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
981     ret = update_refcount(bs, offset, size, 1, true, type);
982     if (ret < 0) {
983         fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
984         /* TODO Remember the clusters to free them later and avoid leaking */
985     }
986 }
987
988 /*
989  * Free a cluster using its L2 entry (handles clusters of all types, e.g.
990  * normal cluster, compressed cluster, etc.)
991  */
992 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
993                              int nb_clusters, enum qcow2_discard_type type)
994 {
995     BDRVQcowState *s = bs->opaque;
996
997     switch (qcow2_get_cluster_type(l2_entry)) {
998     case QCOW2_CLUSTER_COMPRESSED:
999         {
1000             int nb_csectors;
1001             nb_csectors = ((l2_entry >> s->csize_shift) &
1002                            s->csize_mask) + 1;
1003             qcow2_free_clusters(bs,
1004                 (l2_entry & s->cluster_offset_mask) & ~511,
1005                 nb_csectors * 512, type);
1006         }
1007         break;
1008     case QCOW2_CLUSTER_NORMAL:
1009     case QCOW2_CLUSTER_ZERO:
1010         if (l2_entry & L2E_OFFSET_MASK) {
1011             if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) {
1012                 qcow2_signal_corruption(bs, false, -1, -1,
1013                                         "Cannot free unaligned cluster %#llx",
1014                                         l2_entry & L2E_OFFSET_MASK);
1015             } else {
1016                 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
1017                                     nb_clusters << s->cluster_bits, type);
1018             }
1019         }
1020         break;
1021     case QCOW2_CLUSTER_UNALLOCATED:
1022         break;
1023     default:
1024         abort();
1025     }
1026 }
1027
1028
1029
1030 /*********************************************************/
1031 /* snapshots and image creation */
1032
1033
1034
1035 /* update the refcounts of snapshots and the copied flag */
1036 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
1037     int64_t l1_table_offset, int l1_size, int addend)
1038 {
1039     BDRVQcowState *s = bs->opaque;
1040     uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, refcount;
1041     bool l1_allocated = false;
1042     int64_t old_offset, old_l2_offset;
1043     int i, j, l1_modified = 0, nb_csectors;
1044     int ret;
1045
1046     assert(addend >= -1 && addend <= 1);
1047
1048     l2_table = NULL;
1049     l1_table = NULL;
1050     l1_size2 = l1_size * sizeof(uint64_t);
1051
1052     s->cache_discards = true;
1053
1054     /* WARNING: qcow2_snapshot_goto relies on this function not using the
1055      * l1_table_offset when it is the current s->l1_table_offset! Be careful
1056      * when changing this! */
1057     if (l1_table_offset != s->l1_table_offset) {
1058         l1_table = g_try_malloc0(align_offset(l1_size2, 512));
1059         if (l1_size2 && l1_table == NULL) {
1060             ret = -ENOMEM;
1061             goto fail;
1062         }
1063         l1_allocated = true;
1064
1065         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1066         if (ret < 0) {
1067             goto fail;
1068         }
1069
1070         for(i = 0;i < l1_size; i++)
1071             be64_to_cpus(&l1_table[i]);
1072     } else {
1073         assert(l1_size == s->l1_size);
1074         l1_table = s->l1_table;
1075         l1_allocated = false;
1076     }
1077
1078     for(i = 0; i < l1_size; i++) {
1079         l2_offset = l1_table[i];
1080         if (l2_offset) {
1081             old_l2_offset = l2_offset;
1082             l2_offset &= L1E_OFFSET_MASK;
1083
1084             if (offset_into_cluster(s, l2_offset)) {
1085                 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1086                                         PRIx64 " unaligned (L1 index: %#x)",
1087                                         l2_offset, i);
1088                 ret = -EIO;
1089                 goto fail;
1090             }
1091
1092             ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
1093                 (void**) &l2_table);
1094             if (ret < 0) {
1095                 goto fail;
1096             }
1097
1098             for(j = 0; j < s->l2_size; j++) {
1099                 uint64_t cluster_index;
1100
1101                 offset = be64_to_cpu(l2_table[j]);
1102                 old_offset = offset;
1103                 offset &= ~QCOW_OFLAG_COPIED;
1104
1105                 switch (qcow2_get_cluster_type(offset)) {
1106                     case QCOW2_CLUSTER_COMPRESSED:
1107                         nb_csectors = ((offset >> s->csize_shift) &
1108                                        s->csize_mask) + 1;
1109                         if (addend != 0) {
1110                             ret = update_refcount(bs,
1111                                 (offset & s->cluster_offset_mask) & ~511,
1112                                 nb_csectors * 512, abs(addend), addend < 0,
1113                                 QCOW2_DISCARD_SNAPSHOT);
1114                             if (ret < 0) {
1115                                 goto fail;
1116                             }
1117                         }
1118                         /* compressed clusters are never modified */
1119                         refcount = 2;
1120                         break;
1121
1122                     case QCOW2_CLUSTER_NORMAL:
1123                     case QCOW2_CLUSTER_ZERO:
1124                         if (offset_into_cluster(s, offset & L2E_OFFSET_MASK)) {
1125                             qcow2_signal_corruption(bs, true, -1, -1, "Data "
1126                                                     "cluster offset %#llx "
1127                                                     "unaligned (L2 offset: %#"
1128                                                     PRIx64 ", L2 index: %#x)",
1129                                                     offset & L2E_OFFSET_MASK,
1130                                                     l2_offset, j);
1131                             ret = -EIO;
1132                             goto fail;
1133                         }
1134
1135                         cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
1136                         if (!cluster_index) {
1137                             /* unallocated */
1138                             refcount = 0;
1139                             break;
1140                         }
1141                         if (addend != 0) {
1142                             ret = qcow2_update_cluster_refcount(bs,
1143                                     cluster_index, abs(addend), addend < 0,
1144                                     QCOW2_DISCARD_SNAPSHOT);
1145                             if (ret < 0) {
1146                                 goto fail;
1147                             }
1148                         }
1149
1150                         ret = qcow2_get_refcount(bs, cluster_index, &refcount);
1151                         if (ret < 0) {
1152                             goto fail;
1153                         }
1154                         break;
1155
1156                     case QCOW2_CLUSTER_UNALLOCATED:
1157                         refcount = 0;
1158                         break;
1159
1160                     default:
1161                         abort();
1162                 }
1163
1164                 if (refcount == 1) {
1165                     offset |= QCOW_OFLAG_COPIED;
1166                 }
1167                 if (offset != old_offset) {
1168                     if (addend > 0) {
1169                         qcow2_cache_set_dependency(bs, s->l2_table_cache,
1170                             s->refcount_block_cache);
1171                     }
1172                     l2_table[j] = cpu_to_be64(offset);
1173                     qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache,
1174                                                  l2_table);
1175                 }
1176             }
1177
1178             qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1179
1180             if (addend != 0) {
1181                 ret = qcow2_update_cluster_refcount(bs, l2_offset >>
1182                                                         s->cluster_bits,
1183                                                     abs(addend), addend < 0,
1184                                                     QCOW2_DISCARD_SNAPSHOT);
1185                 if (ret < 0) {
1186                     goto fail;
1187                 }
1188             }
1189             ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1190                                      &refcount);
1191             if (ret < 0) {
1192                 goto fail;
1193             } else if (refcount == 1) {
1194                 l2_offset |= QCOW_OFLAG_COPIED;
1195             }
1196             if (l2_offset != old_l2_offset) {
1197                 l1_table[i] = l2_offset;
1198                 l1_modified = 1;
1199             }
1200         }
1201     }
1202
1203     ret = bdrv_flush(bs);
1204 fail:
1205     if (l2_table) {
1206         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1207     }
1208
1209     s->cache_discards = false;
1210     qcow2_process_discards(bs, ret);
1211
1212     /* Update L1 only if it isn't deleted anyway (addend = -1) */
1213     if (ret == 0 && addend >= 0 && l1_modified) {
1214         for (i = 0; i < l1_size; i++) {
1215             cpu_to_be64s(&l1_table[i]);
1216         }
1217
1218         ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
1219
1220         for (i = 0; i < l1_size; i++) {
1221             be64_to_cpus(&l1_table[i]);
1222         }
1223     }
1224     if (l1_allocated)
1225         g_free(l1_table);
1226     return ret;
1227 }
1228
1229
1230
1231
1232 /*********************************************************/
1233 /* refcount checking functions */
1234
1235
1236 static size_t refcount_array_byte_size(BDRVQcowState *s, uint64_t entries)
1237 {
1238     /* This assertion holds because there is no way we can address more than
1239      * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because
1240      * offsets have to be representable in bytes); due to every cluster
1241      * corresponding to one refcount entry, we are well below that limit */
1242     assert(entries < (UINT64_C(1) << (64 - 9)));
1243
1244     /* Thanks to the assertion this will not overflow, because
1245      * s->refcount_order < 7.
1246      * (note: x << s->refcount_order == x * s->refcount_bits) */
1247     return DIV_ROUND_UP(entries << s->refcount_order, 8);
1248 }
1249
1250 /**
1251  * Reallocates *array so that it can hold new_size entries. *size must contain
1252  * the current number of entries in *array. If the reallocation fails, *array
1253  * and *size will not be modified and -errno will be returned. If the
1254  * reallocation is successful, *array will be set to the new buffer, *size
1255  * will be set to new_size and 0 will be returned. The size of the reallocated
1256  * refcount array buffer will be aligned to a cluster boundary, and the newly
1257  * allocated area will be zeroed.
1258  */
1259 static int realloc_refcount_array(BDRVQcowState *s, void **array,
1260                                   int64_t *size, int64_t new_size)
1261 {
1262     size_t old_byte_size, new_byte_size;
1263     void *new_ptr;
1264
1265     /* Round to clusters so the array can be directly written to disk */
1266     old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size))
1267                     * s->cluster_size;
1268     new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size))
1269                     * s->cluster_size;
1270
1271     if (new_byte_size == old_byte_size) {
1272         *size = new_size;
1273         return 0;
1274     }
1275
1276     assert(new_byte_size > 0);
1277
1278     new_ptr = g_try_realloc(*array, new_byte_size);
1279     if (!new_ptr) {
1280         return -ENOMEM;
1281     }
1282
1283     if (new_byte_size > old_byte_size) {
1284         memset((void *)((uintptr_t)new_ptr + old_byte_size), 0,
1285                new_byte_size - old_byte_size);
1286     }
1287
1288     *array = new_ptr;
1289     *size  = new_size;
1290
1291     return 0;
1292 }
1293
1294 /*
1295  * Increases the refcount for a range of clusters in a given refcount table.
1296  * This is used to construct a temporary refcount table out of L1 and L2 tables
1297  * which can be compared the the refcount table saved in the image.
1298  *
1299  * Modifies the number of errors in res.
1300  */
1301 static int inc_refcounts(BlockDriverState *bs,
1302                          BdrvCheckResult *res,
1303                          void **refcount_table,
1304                          int64_t *refcount_table_size,
1305                          int64_t offset, int64_t size)
1306 {
1307     BDRVQcowState *s = bs->opaque;
1308     uint64_t start, last, cluster_offset, k, refcount;
1309     int ret;
1310
1311     if (size <= 0) {
1312         return 0;
1313     }
1314
1315     start = start_of_cluster(s, offset);
1316     last = start_of_cluster(s, offset + size - 1);
1317     for(cluster_offset = start; cluster_offset <= last;
1318         cluster_offset += s->cluster_size) {
1319         k = cluster_offset >> s->cluster_bits;
1320         if (k >= *refcount_table_size) {
1321             ret = realloc_refcount_array(s, refcount_table,
1322                                          refcount_table_size, k + 1);
1323             if (ret < 0) {
1324                 res->check_errors++;
1325                 return ret;
1326             }
1327         }
1328
1329         refcount = s->get_refcount(*refcount_table, k);
1330         if (refcount == s->refcount_max) {
1331             fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1332                     "\n", cluster_offset);
1333             res->corruptions++;
1334             continue;
1335         }
1336         s->set_refcount(*refcount_table, k, refcount + 1);
1337     }
1338
1339     return 0;
1340 }
1341
1342 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1343 enum {
1344     CHECK_FRAG_INFO = 0x2,      /* update BlockFragInfo counters */
1345 };
1346
1347 /*
1348  * Increases the refcount in the given refcount table for the all clusters
1349  * referenced in the L2 table. While doing so, performs some checks on L2
1350  * entries.
1351  *
1352  * Returns the number of errors found by the checks or -errno if an internal
1353  * error occurred.
1354  */
1355 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1356                               void **refcount_table,
1357                               int64_t *refcount_table_size, int64_t l2_offset,
1358                               int flags)
1359 {
1360     BDRVQcowState *s = bs->opaque;
1361     uint64_t *l2_table, l2_entry;
1362     uint64_t next_contiguous_offset = 0;
1363     int i, l2_size, nb_csectors, ret;
1364
1365     /* Read L2 table from disk */
1366     l2_size = s->l2_size * sizeof(uint64_t);
1367     l2_table = g_malloc(l2_size);
1368
1369     ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size);
1370     if (ret < 0) {
1371         fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1372         res->check_errors++;
1373         goto fail;
1374     }
1375
1376     /* Do the actual checks */
1377     for(i = 0; i < s->l2_size; i++) {
1378         l2_entry = be64_to_cpu(l2_table[i]);
1379
1380         switch (qcow2_get_cluster_type(l2_entry)) {
1381         case QCOW2_CLUSTER_COMPRESSED:
1382             /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1383             if (l2_entry & QCOW_OFLAG_COPIED) {
1384                 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
1385                     "copied flag must never be set for compressed "
1386                     "clusters\n", l2_entry >> s->cluster_bits);
1387                 l2_entry &= ~QCOW_OFLAG_COPIED;
1388                 res->corruptions++;
1389             }
1390
1391             /* Mark cluster as used */
1392             nb_csectors = ((l2_entry >> s->csize_shift) &
1393                            s->csize_mask) + 1;
1394             l2_entry &= s->cluster_offset_mask;
1395             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1396                                 l2_entry & ~511, nb_csectors * 512);
1397             if (ret < 0) {
1398                 goto fail;
1399             }
1400
1401             if (flags & CHECK_FRAG_INFO) {
1402                 res->bfi.allocated_clusters++;
1403                 res->bfi.compressed_clusters++;
1404
1405                 /* Compressed clusters are fragmented by nature.  Since they
1406                  * take up sub-sector space but we only have sector granularity
1407                  * I/O we need to re-read the same sectors even for adjacent
1408                  * compressed clusters.
1409                  */
1410                 res->bfi.fragmented_clusters++;
1411             }
1412             break;
1413
1414         case QCOW2_CLUSTER_ZERO:
1415             if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1416                 break;
1417             }
1418             /* fall through */
1419
1420         case QCOW2_CLUSTER_NORMAL:
1421         {
1422             uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1423
1424             if (flags & CHECK_FRAG_INFO) {
1425                 res->bfi.allocated_clusters++;
1426                 if (next_contiguous_offset &&
1427                     offset != next_contiguous_offset) {
1428                     res->bfi.fragmented_clusters++;
1429                 }
1430                 next_contiguous_offset = offset + s->cluster_size;
1431             }
1432
1433             /* Mark cluster as used */
1434             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1435                                 offset, s->cluster_size);
1436             if (ret < 0) {
1437                 goto fail;
1438             }
1439
1440             /* Correct offsets are cluster aligned */
1441             if (offset_into_cluster(s, offset)) {
1442                 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1443                     "properly aligned; L2 entry corrupted.\n", offset);
1444                 res->corruptions++;
1445             }
1446             break;
1447         }
1448
1449         case QCOW2_CLUSTER_UNALLOCATED:
1450             break;
1451
1452         default:
1453             abort();
1454         }
1455     }
1456
1457     g_free(l2_table);
1458     return 0;
1459
1460 fail:
1461     g_free(l2_table);
1462     return ret;
1463 }
1464
1465 /*
1466  * Increases the refcount for the L1 table, its L2 tables and all referenced
1467  * clusters in the given refcount table. While doing so, performs some checks
1468  * on L1 and L2 entries.
1469  *
1470  * Returns the number of errors found by the checks or -errno if an internal
1471  * error occurred.
1472  */
1473 static int check_refcounts_l1(BlockDriverState *bs,
1474                               BdrvCheckResult *res,
1475                               void **refcount_table,
1476                               int64_t *refcount_table_size,
1477                               int64_t l1_table_offset, int l1_size,
1478                               int flags)
1479 {
1480     BDRVQcowState *s = bs->opaque;
1481     uint64_t *l1_table = NULL, l2_offset, l1_size2;
1482     int i, ret;
1483
1484     l1_size2 = l1_size * sizeof(uint64_t);
1485
1486     /* Mark L1 table as used */
1487     ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1488                         l1_table_offset, l1_size2);
1489     if (ret < 0) {
1490         goto fail;
1491     }
1492
1493     /* Read L1 table entries from disk */
1494     if (l1_size2 > 0) {
1495         l1_table = g_try_malloc(l1_size2);
1496         if (l1_table == NULL) {
1497             ret = -ENOMEM;
1498             res->check_errors++;
1499             goto fail;
1500         }
1501         ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1502         if (ret < 0) {
1503             fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1504             res->check_errors++;
1505             goto fail;
1506         }
1507         for(i = 0;i < l1_size; i++)
1508             be64_to_cpus(&l1_table[i]);
1509     }
1510
1511     /* Do the actual checks */
1512     for(i = 0; i < l1_size; i++) {
1513         l2_offset = l1_table[i];
1514         if (l2_offset) {
1515             /* Mark L2 table as used */
1516             l2_offset &= L1E_OFFSET_MASK;
1517             ret = inc_refcounts(bs, res, refcount_table, refcount_table_size,
1518                                 l2_offset, s->cluster_size);
1519             if (ret < 0) {
1520                 goto fail;
1521             }
1522
1523             /* L2 tables are cluster aligned */
1524             if (offset_into_cluster(s, l2_offset)) {
1525                 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1526                     "cluster aligned; L1 entry corrupted\n", l2_offset);
1527                 res->corruptions++;
1528             }
1529
1530             /* Process and check L2 entries */
1531             ret = check_refcounts_l2(bs, res, refcount_table,
1532                                      refcount_table_size, l2_offset, flags);
1533             if (ret < 0) {
1534                 goto fail;
1535             }
1536         }
1537     }
1538     g_free(l1_table);
1539     return 0;
1540
1541 fail:
1542     g_free(l1_table);
1543     return ret;
1544 }
1545
1546 /*
1547  * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1548  *
1549  * This function does not print an error message nor does it increment
1550  * check_errors if qcow2_get_refcount fails (this is because such an error will
1551  * have been already detected and sufficiently signaled by the calling function
1552  * (qcow2_check_refcounts) by the time this function is called).
1553  */
1554 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1555                               BdrvCheckMode fix)
1556 {
1557     BDRVQcowState *s = bs->opaque;
1558     uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1559     int ret;
1560     uint64_t refcount;
1561     int i, j;
1562
1563     for (i = 0; i < s->l1_size; i++) {
1564         uint64_t l1_entry = s->l1_table[i];
1565         uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1566         bool l2_dirty = false;
1567
1568         if (!l2_offset) {
1569             continue;
1570         }
1571
1572         ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1573                                  &refcount);
1574         if (ret < 0) {
1575             /* don't print message nor increment check_errors */
1576             continue;
1577         }
1578         if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1579             fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1580                     "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1581                     fix & BDRV_FIX_ERRORS ? "Repairing" :
1582                                             "ERROR",
1583                     i, l1_entry, refcount);
1584             if (fix & BDRV_FIX_ERRORS) {
1585                 s->l1_table[i] = refcount == 1
1586                                ? l1_entry |  QCOW_OFLAG_COPIED
1587                                : l1_entry & ~QCOW_OFLAG_COPIED;
1588                 ret = qcow2_write_l1_entry(bs, i);
1589                 if (ret < 0) {
1590                     res->check_errors++;
1591                     goto fail;
1592                 }
1593                 res->corruptions_fixed++;
1594             } else {
1595                 res->corruptions++;
1596             }
1597         }
1598
1599         ret = bdrv_pread(bs->file, l2_offset, l2_table,
1600                          s->l2_size * sizeof(uint64_t));
1601         if (ret < 0) {
1602             fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1603                     strerror(-ret));
1604             res->check_errors++;
1605             goto fail;
1606         }
1607
1608         for (j = 0; j < s->l2_size; j++) {
1609             uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1610             uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1611             int cluster_type = qcow2_get_cluster_type(l2_entry);
1612
1613             if ((cluster_type == QCOW2_CLUSTER_NORMAL) ||
1614                 ((cluster_type == QCOW2_CLUSTER_ZERO) && (data_offset != 0))) {
1615                 ret = qcow2_get_refcount(bs,
1616                                          data_offset >> s->cluster_bits,
1617                                          &refcount);
1618                 if (ret < 0) {
1619                     /* don't print message nor increment check_errors */
1620                     continue;
1621                 }
1622                 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1623                     fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1624                             "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1625                             fix & BDRV_FIX_ERRORS ? "Repairing" :
1626                                                     "ERROR",
1627                             l2_entry, refcount);
1628                     if (fix & BDRV_FIX_ERRORS) {
1629                         l2_table[j] = cpu_to_be64(refcount == 1
1630                                     ? l2_entry |  QCOW_OFLAG_COPIED
1631                                     : l2_entry & ~QCOW_OFLAG_COPIED);
1632                         l2_dirty = true;
1633                         res->corruptions_fixed++;
1634                     } else {
1635                         res->corruptions++;
1636                     }
1637                 }
1638             }
1639         }
1640
1641         if (l2_dirty) {
1642             ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1643                                                 l2_offset, s->cluster_size);
1644             if (ret < 0) {
1645                 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1646                         "overlap check failed: %s\n", strerror(-ret));
1647                 res->check_errors++;
1648                 goto fail;
1649             }
1650
1651             ret = bdrv_pwrite(bs->file, l2_offset, l2_table, s->cluster_size);
1652             if (ret < 0) {
1653                 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1654                         strerror(-ret));
1655                 res->check_errors++;
1656                 goto fail;
1657             }
1658         }
1659     }
1660
1661     ret = 0;
1662
1663 fail:
1664     qemu_vfree(l2_table);
1665     return ret;
1666 }
1667
1668 /*
1669  * Checks consistency of refblocks and accounts for each refblock in
1670  * *refcount_table.
1671  */
1672 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res,
1673                            BdrvCheckMode fix, bool *rebuild,
1674                            void **refcount_table, int64_t *nb_clusters)
1675 {
1676     BDRVQcowState *s = bs->opaque;
1677     int64_t i, size;
1678     int ret;
1679
1680     for(i = 0; i < s->refcount_table_size; i++) {
1681         uint64_t offset, cluster;
1682         offset = s->refcount_table[i];
1683         cluster = offset >> s->cluster_bits;
1684
1685         /* Refcount blocks are cluster aligned */
1686         if (offset_into_cluster(s, offset)) {
1687             fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1688                 "cluster aligned; refcount table entry corrupted\n", i);
1689             res->corruptions++;
1690             *rebuild = true;
1691             continue;
1692         }
1693
1694         if (cluster >= *nb_clusters) {
1695             fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n",
1696                     fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i);
1697
1698             if (fix & BDRV_FIX_ERRORS) {
1699                 int64_t new_nb_clusters;
1700
1701                 if (offset > INT64_MAX - s->cluster_size) {
1702                     ret = -EINVAL;
1703                     goto resize_fail;
1704                 }
1705
1706                 ret = bdrv_truncate(bs->file, offset + s->cluster_size);
1707                 if (ret < 0) {
1708                     goto resize_fail;
1709                 }
1710                 size = bdrv_getlength(bs->file);
1711                 if (size < 0) {
1712                     ret = size;
1713                     goto resize_fail;
1714                 }
1715
1716                 new_nb_clusters = size_to_clusters(s, size);
1717                 assert(new_nb_clusters >= *nb_clusters);
1718
1719                 ret = realloc_refcount_array(s, refcount_table,
1720                                              nb_clusters, new_nb_clusters);
1721                 if (ret < 0) {
1722                     res->check_errors++;
1723                     return ret;
1724                 }
1725
1726                 if (cluster >= *nb_clusters) {
1727                     ret = -EINVAL;
1728                     goto resize_fail;
1729                 }
1730
1731                 res->corruptions_fixed++;
1732                 ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1733                                     offset, s->cluster_size);
1734                 if (ret < 0) {
1735                     return ret;
1736                 }
1737                 /* No need to check whether the refcount is now greater than 1:
1738                  * This area was just allocated and zeroed, so it can only be
1739                  * exactly 1 after inc_refcounts() */
1740                 continue;
1741
1742 resize_fail:
1743                 res->corruptions++;
1744                 *rebuild = true;
1745                 fprintf(stderr, "ERROR could not resize image: %s\n",
1746                         strerror(-ret));
1747             } else {
1748                 res->corruptions++;
1749             }
1750             continue;
1751         }
1752
1753         if (offset != 0) {
1754             ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1755                                 offset, s->cluster_size);
1756             if (ret < 0) {
1757                 return ret;
1758             }
1759             if (s->get_refcount(*refcount_table, cluster) != 1) {
1760                 fprintf(stderr, "ERROR refcount block %" PRId64
1761                         " refcount=%" PRIu64 "\n", i,
1762                         s->get_refcount(*refcount_table, cluster));
1763                 res->corruptions++;
1764                 *rebuild = true;
1765             }
1766         }
1767     }
1768
1769     return 0;
1770 }
1771
1772 /*
1773  * Calculates an in-memory refcount table.
1774  */
1775 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1776                                BdrvCheckMode fix, bool *rebuild,
1777                                void **refcount_table, int64_t *nb_clusters)
1778 {
1779     BDRVQcowState *s = bs->opaque;
1780     int64_t i;
1781     QCowSnapshot *sn;
1782     int ret;
1783
1784     if (!*refcount_table) {
1785         int64_t old_size = 0;
1786         ret = realloc_refcount_array(s, refcount_table,
1787                                      &old_size, *nb_clusters);
1788         if (ret < 0) {
1789             res->check_errors++;
1790             return ret;
1791         }
1792     }
1793
1794     /* header */
1795     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1796                         0, s->cluster_size);
1797     if (ret < 0) {
1798         return ret;
1799     }
1800
1801     /* current L1 table */
1802     ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1803                              s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO);
1804     if (ret < 0) {
1805         return ret;
1806     }
1807
1808     /* snapshots */
1809     for (i = 0; i < s->nb_snapshots; i++) {
1810         sn = s->snapshots + i;
1811         ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1812                                  sn->l1_table_offset, sn->l1_size, 0);
1813         if (ret < 0) {
1814             return ret;
1815         }
1816     }
1817     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1818                         s->snapshots_offset, s->snapshots_size);
1819     if (ret < 0) {
1820         return ret;
1821     }
1822
1823     /* refcount data */
1824     ret = inc_refcounts(bs, res, refcount_table, nb_clusters,
1825                         s->refcount_table_offset,
1826                         s->refcount_table_size * sizeof(uint64_t));
1827     if (ret < 0) {
1828         return ret;
1829     }
1830
1831     return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters);
1832 }
1833
1834 /*
1835  * Compares the actual reference count for each cluster in the image against the
1836  * refcount as reported by the refcount structures on-disk.
1837  */
1838 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1839                               BdrvCheckMode fix, bool *rebuild,
1840                               int64_t *highest_cluster,
1841                               void *refcount_table, int64_t nb_clusters)
1842 {
1843     BDRVQcowState *s = bs->opaque;
1844     int64_t i;
1845     uint64_t refcount1, refcount2;
1846     int ret;
1847
1848     for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) {
1849         ret = qcow2_get_refcount(bs, i, &refcount1);
1850         if (ret < 0) {
1851             fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1852                     i, strerror(-ret));
1853             res->check_errors++;
1854             continue;
1855         }
1856
1857         refcount2 = s->get_refcount(refcount_table, i);
1858
1859         if (refcount1 > 0 || refcount2 > 0) {
1860             *highest_cluster = i;
1861         }
1862
1863         if (refcount1 != refcount2) {
1864             /* Check if we're allowed to fix the mismatch */
1865             int *num_fixed = NULL;
1866             if (refcount1 == 0) {
1867                 *rebuild = true;
1868             } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1869                 num_fixed = &res->leaks_fixed;
1870             } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1871                 num_fixed = &res->corruptions_fixed;
1872             }
1873
1874             fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64
1875                     " reference=%" PRIu64 "\n",
1876                    num_fixed != NULL     ? "Repairing" :
1877                    refcount1 < refcount2 ? "ERROR" :
1878                                            "Leaked",
1879                    i, refcount1, refcount2);
1880
1881             if (num_fixed) {
1882                 ret = update_refcount(bs, i << s->cluster_bits, 1,
1883                                       refcount_diff(refcount1, refcount2),
1884                                       refcount1 > refcount2,
1885                                       QCOW2_DISCARD_ALWAYS);
1886                 if (ret >= 0) {
1887                     (*num_fixed)++;
1888                     continue;
1889                 }
1890             }
1891
1892             /* And if we couldn't, print an error */
1893             if (refcount1 < refcount2) {
1894                 res->corruptions++;
1895             } else {
1896                 res->leaks++;
1897             }
1898         }
1899     }
1900 }
1901
1902 /*
1903  * Allocates clusters using an in-memory refcount table (IMRT) in contrast to
1904  * the on-disk refcount structures.
1905  *
1906  * On input, *first_free_cluster tells where to start looking, and need not
1907  * actually be a free cluster; the returned offset will not be before that
1908  * cluster.  On output, *first_free_cluster points to the first gap found, even
1909  * if that gap was too small to be used as the returned offset.
1910  *
1911  * Note that *first_free_cluster is a cluster index whereas the return value is
1912  * an offset.
1913  */
1914 static int64_t alloc_clusters_imrt(BlockDriverState *bs,
1915                                    int cluster_count,
1916                                    void **refcount_table,
1917                                    int64_t *imrt_nb_clusters,
1918                                    int64_t *first_free_cluster)
1919 {
1920     BDRVQcowState *s = bs->opaque;
1921     int64_t cluster = *first_free_cluster, i;
1922     bool first_gap = true;
1923     int contiguous_free_clusters;
1924     int ret;
1925
1926     /* Starting at *first_free_cluster, find a range of at least cluster_count
1927      * continuously free clusters */
1928     for (contiguous_free_clusters = 0;
1929          cluster < *imrt_nb_clusters &&
1930          contiguous_free_clusters < cluster_count;
1931          cluster++)
1932     {
1933         if (!s->get_refcount(*refcount_table, cluster)) {
1934             contiguous_free_clusters++;
1935             if (first_gap) {
1936                 /* If this is the first free cluster found, update
1937                  * *first_free_cluster accordingly */
1938                 *first_free_cluster = cluster;
1939                 first_gap = false;
1940             }
1941         } else if (contiguous_free_clusters) {
1942             contiguous_free_clusters = 0;
1943         }
1944     }
1945
1946     /* If contiguous_free_clusters is greater than zero, it contains the number
1947      * of continuously free clusters until the current cluster; the first free
1948      * cluster in the current "gap" is therefore
1949      * cluster - contiguous_free_clusters */
1950
1951     /* If no such range could be found, grow the in-memory refcount table
1952      * accordingly to append free clusters at the end of the image */
1953     if (contiguous_free_clusters < cluster_count) {
1954         /* contiguous_free_clusters clusters are already empty at the image end;
1955          * we need cluster_count clusters; therefore, we have to allocate
1956          * cluster_count - contiguous_free_clusters new clusters at the end of
1957          * the image (which is the current value of cluster; note that cluster
1958          * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond
1959          * the image end) */
1960         ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters,
1961                                      cluster + cluster_count
1962                                      - contiguous_free_clusters);
1963         if (ret < 0) {
1964             return ret;
1965         }
1966     }
1967
1968     /* Go back to the first free cluster */
1969     cluster -= contiguous_free_clusters;
1970     for (i = 0; i < cluster_count; i++) {
1971         s->set_refcount(*refcount_table, cluster + i, 1);
1972     }
1973
1974     return cluster << s->cluster_bits;
1975 }
1976
1977 /*
1978  * Creates a new refcount structure based solely on the in-memory information
1979  * given through *refcount_table. All necessary allocations will be reflected
1980  * in that array.
1981  *
1982  * On success, the old refcount structure is leaked (it will be covered by the
1983  * new refcount structure).
1984  */
1985 static int rebuild_refcount_structure(BlockDriverState *bs,
1986                                       BdrvCheckResult *res,
1987                                       void **refcount_table,
1988                                       int64_t *nb_clusters)
1989 {
1990     BDRVQcowState *s = bs->opaque;
1991     int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0;
1992     int64_t refblock_offset, refblock_start, refblock_index;
1993     uint32_t reftable_size = 0;
1994     uint64_t *on_disk_reftable = NULL;
1995     void *on_disk_refblock;
1996     int ret = 0;
1997     struct {
1998         uint64_t reftable_offset;
1999         uint32_t reftable_clusters;
2000     } QEMU_PACKED reftable_offset_and_clusters;
2001
2002     qcow2_cache_empty(bs, s->refcount_block_cache);
2003
2004 write_refblocks:
2005     for (; cluster < *nb_clusters; cluster++) {
2006         if (!s->get_refcount(*refcount_table, cluster)) {
2007             continue;
2008         }
2009
2010         refblock_index = cluster >> s->refcount_block_bits;
2011         refblock_start = refblock_index << s->refcount_block_bits;
2012
2013         /* Don't allocate a cluster in a refblock already written to disk */
2014         if (first_free_cluster < refblock_start) {
2015             first_free_cluster = refblock_start;
2016         }
2017         refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table,
2018                                               nb_clusters, &first_free_cluster);
2019         if (refblock_offset < 0) {
2020             fprintf(stderr, "ERROR allocating refblock: %s\n",
2021                     strerror(-refblock_offset));
2022             res->check_errors++;
2023             ret = refblock_offset;
2024             goto fail;
2025         }
2026
2027         if (reftable_size <= refblock_index) {
2028             uint32_t old_reftable_size = reftable_size;
2029             uint64_t *new_on_disk_reftable;
2030
2031             reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t),
2032                                      s->cluster_size) / sizeof(uint64_t);
2033             new_on_disk_reftable = g_try_realloc(on_disk_reftable,
2034                                                  reftable_size *
2035                                                  sizeof(uint64_t));
2036             if (!new_on_disk_reftable) {
2037                 res->check_errors++;
2038                 ret = -ENOMEM;
2039                 goto fail;
2040             }
2041             on_disk_reftable = new_on_disk_reftable;
2042
2043             memset(on_disk_reftable + old_reftable_size, 0,
2044                    (reftable_size - old_reftable_size) * sizeof(uint64_t));
2045
2046             /* The offset we have for the reftable is now no longer valid;
2047              * this will leak that range, but we can easily fix that by running
2048              * a leak-fixing check after this rebuild operation */
2049             reftable_offset = -1;
2050         }
2051         on_disk_reftable[refblock_index] = refblock_offset;
2052
2053         /* If this is apparently the last refblock (for now), try to squeeze the
2054          * reftable in */
2055         if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits &&
2056             reftable_offset < 0)
2057         {
2058             uint64_t reftable_clusters = size_to_clusters(s, reftable_size *
2059                                                           sizeof(uint64_t));
2060             reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2061                                                   refcount_table, nb_clusters,
2062                                                   &first_free_cluster);
2063             if (reftable_offset < 0) {
2064                 fprintf(stderr, "ERROR allocating reftable: %s\n",
2065                         strerror(-reftable_offset));
2066                 res->check_errors++;
2067                 ret = reftable_offset;
2068                 goto fail;
2069             }
2070         }
2071
2072         ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset,
2073                                             s->cluster_size);
2074         if (ret < 0) {
2075             fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2076             goto fail;
2077         }
2078
2079         /* The size of *refcount_table is always cluster-aligned, therefore the
2080          * write operation will not overflow */
2081         on_disk_refblock = (void *)((char *) *refcount_table +
2082                                     refblock_index * s->cluster_size);
2083
2084         ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE,
2085                          on_disk_refblock, s->cluster_sectors);
2086         if (ret < 0) {
2087             fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2088             goto fail;
2089         }
2090
2091         /* Go to the end of this refblock */
2092         cluster = refblock_start + s->refcount_block_size - 1;
2093     }
2094
2095     if (reftable_offset < 0) {
2096         uint64_t post_refblock_start, reftable_clusters;
2097
2098         post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size);
2099         reftable_clusters = size_to_clusters(s,
2100                                              reftable_size * sizeof(uint64_t));
2101         /* Not pretty but simple */
2102         if (first_free_cluster < post_refblock_start) {
2103             first_free_cluster = post_refblock_start;
2104         }
2105         reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2106                                               refcount_table, nb_clusters,
2107                                               &first_free_cluster);
2108         if (reftable_offset < 0) {
2109             fprintf(stderr, "ERROR allocating reftable: %s\n",
2110                     strerror(-reftable_offset));
2111             res->check_errors++;
2112             ret = reftable_offset;
2113             goto fail;
2114         }
2115
2116         goto write_refblocks;
2117     }
2118
2119     assert(on_disk_reftable);
2120
2121     for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2122         cpu_to_be64s(&on_disk_reftable[refblock_index]);
2123     }
2124
2125     ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset,
2126                                         reftable_size * sizeof(uint64_t));
2127     if (ret < 0) {
2128         fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2129         goto fail;
2130     }
2131
2132     assert(reftable_size < INT_MAX / sizeof(uint64_t));
2133     ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable,
2134                       reftable_size * sizeof(uint64_t));
2135     if (ret < 0) {
2136         fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2137         goto fail;
2138     }
2139
2140     /* Enter new reftable into the image header */
2141     cpu_to_be64w(&reftable_offset_and_clusters.reftable_offset,
2142                  reftable_offset);
2143     cpu_to_be32w(&reftable_offset_and_clusters.reftable_clusters,
2144                  size_to_clusters(s, reftable_size * sizeof(uint64_t)));
2145     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader,
2146                                               refcount_table_offset),
2147                            &reftable_offset_and_clusters,
2148                            sizeof(reftable_offset_and_clusters));
2149     if (ret < 0) {
2150         fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret));
2151         goto fail;
2152     }
2153
2154     for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2155         be64_to_cpus(&on_disk_reftable[refblock_index]);
2156     }
2157     s->refcount_table = on_disk_reftable;
2158     s->refcount_table_offset = reftable_offset;
2159     s->refcount_table_size = reftable_size;
2160
2161     return 0;
2162
2163 fail:
2164     g_free(on_disk_reftable);
2165     return ret;
2166 }
2167
2168 /*
2169  * Checks an image for refcount consistency.
2170  *
2171  * Returns 0 if no errors are found, the number of errors in case the image is
2172  * detected as corrupted, and -errno when an internal error occurred.
2173  */
2174 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2175                           BdrvCheckMode fix)
2176 {
2177     BDRVQcowState *s = bs->opaque;
2178     BdrvCheckResult pre_compare_res;
2179     int64_t size, highest_cluster, nb_clusters;
2180     void *refcount_table = NULL;
2181     bool rebuild = false;
2182     int ret;
2183
2184     size = bdrv_getlength(bs->file);
2185     if (size < 0) {
2186         res->check_errors++;
2187         return size;
2188     }
2189
2190     nb_clusters = size_to_clusters(s, size);
2191     if (nb_clusters > INT_MAX) {
2192         res->check_errors++;
2193         return -EFBIG;
2194     }
2195
2196     res->bfi.total_clusters =
2197         size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
2198
2199     ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table,
2200                               &nb_clusters);
2201     if (ret < 0) {
2202         goto fail;
2203     }
2204
2205     /* In case we don't need to rebuild the refcount structure (but want to fix
2206      * something), this function is immediately called again, in which case the
2207      * result should be ignored */
2208     pre_compare_res = *res;
2209     compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table,
2210                       nb_clusters);
2211
2212     if (rebuild && (fix & BDRV_FIX_ERRORS)) {
2213         BdrvCheckResult old_res = *res;
2214         int fresh_leaks = 0;
2215
2216         fprintf(stderr, "Rebuilding refcount structure\n");
2217         ret = rebuild_refcount_structure(bs, res, &refcount_table,
2218                                          &nb_clusters);
2219         if (ret < 0) {
2220             goto fail;
2221         }
2222
2223         res->corruptions = 0;
2224         res->leaks = 0;
2225
2226         /* Because the old reftable has been exchanged for a new one the
2227          * references have to be recalculated */
2228         rebuild = false;
2229         memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters));
2230         ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table,
2231                                   &nb_clusters);
2232         if (ret < 0) {
2233             goto fail;
2234         }
2235
2236         if (fix & BDRV_FIX_LEAKS) {
2237             /* The old refcount structures are now leaked, fix it; the result
2238              * can be ignored, aside from leaks which were introduced by
2239              * rebuild_refcount_structure() that could not be fixed */
2240             BdrvCheckResult saved_res = *res;
2241             *res = (BdrvCheckResult){ 0 };
2242
2243             compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild,
2244                               &highest_cluster, refcount_table, nb_clusters);
2245             if (rebuild) {
2246                 fprintf(stderr, "ERROR rebuilt refcount structure is still "
2247                         "broken\n");
2248             }
2249
2250             /* Any leaks accounted for here were introduced by
2251              * rebuild_refcount_structure() because that function has created a
2252              * new refcount structure from scratch */
2253             fresh_leaks = res->leaks;
2254             *res = saved_res;
2255         }
2256
2257         if (res->corruptions < old_res.corruptions) {
2258             res->corruptions_fixed += old_res.corruptions - res->corruptions;
2259         }
2260         if (res->leaks < old_res.leaks) {
2261             res->leaks_fixed += old_res.leaks - res->leaks;
2262         }
2263         res->leaks += fresh_leaks;
2264     } else if (fix) {
2265         if (rebuild) {
2266             fprintf(stderr, "ERROR need to rebuild refcount structures\n");
2267             res->check_errors++;
2268             ret = -EIO;
2269             goto fail;
2270         }
2271
2272         if (res->leaks || res->corruptions) {
2273             *res = pre_compare_res;
2274             compare_refcounts(bs, res, fix, &rebuild, &highest_cluster,
2275                               refcount_table, nb_clusters);
2276         }
2277     }
2278
2279     /* check OFLAG_COPIED */
2280     ret = check_oflag_copied(bs, res, fix);
2281     if (ret < 0) {
2282         goto fail;
2283     }
2284
2285     res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
2286     ret = 0;
2287
2288 fail:
2289     g_free(refcount_table);
2290
2291     return ret;
2292 }
2293
2294 #define overlaps_with(ofs, sz) \
2295     ranges_overlap(offset, size, ofs, sz)
2296
2297 /*
2298  * Checks if the given offset into the image file is actually free to use by
2299  * looking for overlaps with important metadata sections (L1/L2 tables etc.),
2300  * i.e. a sanity check without relying on the refcount tables.
2301  *
2302  * The ign parameter specifies what checks not to perform (being a bitmask of
2303  * QCow2MetadataOverlap values), i.e., what sections to ignore.
2304  *
2305  * Returns:
2306  * - 0 if writing to this offset will not affect the mentioned metadata
2307  * - a positive QCow2MetadataOverlap value indicating one overlapping section
2308  * - a negative value (-errno) indicating an error while performing a check,
2309  *   e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
2310  */
2311 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
2312                                  int64_t size)
2313 {
2314     BDRVQcowState *s = bs->opaque;
2315     int chk = s->overlap_check & ~ign;
2316     int i, j;
2317
2318     if (!size) {
2319         return 0;
2320     }
2321
2322     if (chk & QCOW2_OL_MAIN_HEADER) {
2323         if (offset < s->cluster_size) {
2324             return QCOW2_OL_MAIN_HEADER;
2325         }
2326     }
2327
2328     /* align range to test to cluster boundaries */
2329     size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size);
2330     offset = start_of_cluster(s, offset);
2331
2332     if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
2333         if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
2334             return QCOW2_OL_ACTIVE_L1;
2335         }
2336     }
2337
2338     if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
2339         if (overlaps_with(s->refcount_table_offset,
2340             s->refcount_table_size * sizeof(uint64_t))) {
2341             return QCOW2_OL_REFCOUNT_TABLE;
2342         }
2343     }
2344
2345     if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
2346         if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
2347             return QCOW2_OL_SNAPSHOT_TABLE;
2348         }
2349     }
2350
2351     if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
2352         for (i = 0; i < s->nb_snapshots; i++) {
2353             if (s->snapshots[i].l1_size &&
2354                 overlaps_with(s->snapshots[i].l1_table_offset,
2355                 s->snapshots[i].l1_size * sizeof(uint64_t))) {
2356                 return QCOW2_OL_INACTIVE_L1;
2357             }
2358         }
2359     }
2360
2361     if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
2362         for (i = 0; i < s->l1_size; i++) {
2363             if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
2364                 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
2365                 s->cluster_size)) {
2366                 return QCOW2_OL_ACTIVE_L2;
2367             }
2368         }
2369     }
2370
2371     if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
2372         for (i = 0; i < s->refcount_table_size; i++) {
2373             if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
2374                 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
2375                 s->cluster_size)) {
2376                 return QCOW2_OL_REFCOUNT_BLOCK;
2377             }
2378         }
2379     }
2380
2381     if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
2382         for (i = 0; i < s->nb_snapshots; i++) {
2383             uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
2384             uint32_t l1_sz  = s->snapshots[i].l1_size;
2385             uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
2386             uint64_t *l1 = g_try_malloc(l1_sz2);
2387             int ret;
2388
2389             if (l1_sz2 && l1 == NULL) {
2390                 return -ENOMEM;
2391             }
2392
2393             ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
2394             if (ret < 0) {
2395                 g_free(l1);
2396                 return ret;
2397             }
2398
2399             for (j = 0; j < l1_sz; j++) {
2400                 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
2401                 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
2402                     g_free(l1);
2403                     return QCOW2_OL_INACTIVE_L2;
2404                 }
2405             }
2406
2407             g_free(l1);
2408         }
2409     }
2410
2411     return 0;
2412 }
2413
2414 static const char *metadata_ol_names[] = {
2415     [QCOW2_OL_MAIN_HEADER_BITNR]    = "qcow2_header",
2416     [QCOW2_OL_ACTIVE_L1_BITNR]      = "active L1 table",
2417     [QCOW2_OL_ACTIVE_L2_BITNR]      = "active L2 table",
2418     [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
2419     [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
2420     [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
2421     [QCOW2_OL_INACTIVE_L1_BITNR]    = "inactive L1 table",
2422     [QCOW2_OL_INACTIVE_L2_BITNR]    = "inactive L2 table",
2423 };
2424
2425 /*
2426  * First performs a check for metadata overlaps (through
2427  * qcow2_check_metadata_overlap); if that fails with a negative value (error
2428  * while performing a check), that value is returned. If an impending overlap
2429  * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
2430  * and -EIO returned.
2431  *
2432  * Returns 0 if there were neither overlaps nor errors while checking for
2433  * overlaps; or a negative value (-errno) on error.
2434  */
2435 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
2436                                   int64_t size)
2437 {
2438     int ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
2439
2440     if (ret < 0) {
2441         return ret;
2442     } else if (ret > 0) {
2443         int metadata_ol_bitnr = ctz32(ret);
2444         assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
2445
2446         qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid "
2447                                 "write on metadata (overlaps with %s)",
2448                                 metadata_ol_names[metadata_ol_bitnr]);
2449         return -EIO;
2450     }
2451
2452     return 0;
2453 }