]> git.proxmox.com Git - qemu.git/blob - block/qcow2-refcount.c
qcow2: Options to enable discard for freed clusters
[qemu.git] / 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
29 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size);
30 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
31 int64_t offset, int64_t length,
32 int addend, enum qcow2_discard_type type);
33
34
35 /*********************************************************/
36 /* refcount handling */
37
38 int qcow2_refcount_init(BlockDriverState *bs)
39 {
40 BDRVQcowState *s = bs->opaque;
41 int ret, refcount_table_size2, i;
42
43 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
44 s->refcount_table = g_malloc(refcount_table_size2);
45 if (s->refcount_table_size > 0) {
46 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
47 ret = bdrv_pread(bs->file, s->refcount_table_offset,
48 s->refcount_table, refcount_table_size2);
49 if (ret != refcount_table_size2)
50 goto fail;
51 for(i = 0; i < s->refcount_table_size; i++)
52 be64_to_cpus(&s->refcount_table[i]);
53 }
54 return 0;
55 fail:
56 return -ENOMEM;
57 }
58
59 void qcow2_refcount_close(BlockDriverState *bs)
60 {
61 BDRVQcowState *s = bs->opaque;
62 g_free(s->refcount_table);
63 }
64
65
66 static int load_refcount_block(BlockDriverState *bs,
67 int64_t refcount_block_offset,
68 void **refcount_block)
69 {
70 BDRVQcowState *s = bs->opaque;
71 int ret;
72
73 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
74 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
75 refcount_block);
76
77 return ret;
78 }
79
80 /*
81 * Returns the refcount of the cluster given by its index. Any non-negative
82 * return value is the refcount of the cluster, negative values are -errno
83 * and indicate an error.
84 */
85 static int get_refcount(BlockDriverState *bs, int64_t cluster_index)
86 {
87 BDRVQcowState *s = bs->opaque;
88 int refcount_table_index, block_index;
89 int64_t refcount_block_offset;
90 int ret;
91 uint16_t *refcount_block;
92 uint16_t refcount;
93
94 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
95 if (refcount_table_index >= s->refcount_table_size)
96 return 0;
97 refcount_block_offset = s->refcount_table[refcount_table_index];
98 if (!refcount_block_offset)
99 return 0;
100
101 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
102 (void**) &refcount_block);
103 if (ret < 0) {
104 return ret;
105 }
106
107 block_index = cluster_index &
108 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
109 refcount = be16_to_cpu(refcount_block[block_index]);
110
111 ret = qcow2_cache_put(bs, s->refcount_block_cache,
112 (void**) &refcount_block);
113 if (ret < 0) {
114 return ret;
115 }
116
117 return refcount;
118 }
119
120 /*
121 * Rounds the refcount table size up to avoid growing the table for each single
122 * refcount block that is allocated.
123 */
124 static unsigned int next_refcount_table_size(BDRVQcowState *s,
125 unsigned int min_size)
126 {
127 unsigned int min_clusters = (min_size >> (s->cluster_bits - 3)) + 1;
128 unsigned int refcount_table_clusters =
129 MAX(1, s->refcount_table_size >> (s->cluster_bits - 3));
130
131 while (min_clusters > refcount_table_clusters) {
132 refcount_table_clusters = (refcount_table_clusters * 3 + 1) / 2;
133 }
134
135 return refcount_table_clusters << (s->cluster_bits - 3);
136 }
137
138
139 /* Checks if two offsets are described by the same refcount block */
140 static int in_same_refcount_block(BDRVQcowState *s, uint64_t offset_a,
141 uint64_t offset_b)
142 {
143 uint64_t block_a = offset_a >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
144 uint64_t block_b = offset_b >> (2 * s->cluster_bits - REFCOUNT_SHIFT);
145
146 return (block_a == block_b);
147 }
148
149 /*
150 * Loads a refcount block. If it doesn't exist yet, it is allocated first
151 * (including growing the refcount table if needed).
152 *
153 * Returns 0 on success or -errno in error case
154 */
155 static int alloc_refcount_block(BlockDriverState *bs,
156 int64_t cluster_index, uint16_t **refcount_block)
157 {
158 BDRVQcowState *s = bs->opaque;
159 unsigned int refcount_table_index;
160 int ret;
161
162 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
163
164 /* Find the refcount block for the given cluster */
165 refcount_table_index = cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
166
167 if (refcount_table_index < s->refcount_table_size) {
168
169 uint64_t refcount_block_offset =
170 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
171
172 /* If it's already there, we're done */
173 if (refcount_block_offset) {
174 return load_refcount_block(bs, refcount_block_offset,
175 (void**) refcount_block);
176 }
177 }
178
179 /*
180 * If we came here, we need to allocate something. Something is at least
181 * a cluster for the new refcount block. It may also include a new refcount
182 * table if the old refcount table is too small.
183 *
184 * Note that allocating clusters here needs some special care:
185 *
186 * - We can't use the normal qcow2_alloc_clusters(), it would try to
187 * increase the refcount and very likely we would end up with an endless
188 * recursion. Instead we must place the refcount blocks in a way that
189 * they can describe them themselves.
190 *
191 * - We need to consider that at this point we are inside update_refcounts
192 * and doing the initial refcount increase. This means that some clusters
193 * have already been allocated by the caller, but their refcount isn't
194 * accurate yet. free_cluster_index tells us where this allocation ends
195 * as long as we don't overwrite it by freeing clusters.
196 *
197 * - alloc_clusters_noref and qcow2_free_clusters may load a different
198 * refcount block into the cache
199 */
200
201 *refcount_block = NULL;
202
203 /* We write to the refcount table, so we might depend on L2 tables */
204 ret = qcow2_cache_flush(bs, s->l2_table_cache);
205 if (ret < 0) {
206 return ret;
207 }
208
209 /* Allocate the refcount block itself and mark it as used */
210 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size);
211 if (new_block < 0) {
212 return new_block;
213 }
214
215 #ifdef DEBUG_ALLOC2
216 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
217 " at %" PRIx64 "\n",
218 refcount_table_index, cluster_index << s->cluster_bits, new_block);
219 #endif
220
221 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
222 /* Zero the new refcount block before updating it */
223 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
224 (void**) refcount_block);
225 if (ret < 0) {
226 goto fail_block;
227 }
228
229 memset(*refcount_block, 0, s->cluster_size);
230
231 /* The block describes itself, need to update the cache */
232 int block_index = (new_block >> s->cluster_bits) &
233 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
234 (*refcount_block)[block_index] = cpu_to_be16(1);
235 } else {
236 /* Described somewhere else. This can recurse at most twice before we
237 * arrive at a block that describes itself. */
238 ret = update_refcount(bs, new_block, s->cluster_size, 1,
239 QCOW2_DISCARD_NEVER);
240 if (ret < 0) {
241 goto fail_block;
242 }
243
244 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
245 if (ret < 0) {
246 goto fail_block;
247 }
248
249 /* Initialize the new refcount block only after updating its refcount,
250 * update_refcount uses the refcount cache itself */
251 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
252 (void**) refcount_block);
253 if (ret < 0) {
254 goto fail_block;
255 }
256
257 memset(*refcount_block, 0, s->cluster_size);
258 }
259
260 /* Now the new refcount block needs to be written to disk */
261 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
262 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
263 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
264 if (ret < 0) {
265 goto fail_block;
266 }
267
268 /* If the refcount table is big enough, just hook the block up there */
269 if (refcount_table_index < s->refcount_table_size) {
270 uint64_t data64 = cpu_to_be64(new_block);
271 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
272 ret = bdrv_pwrite_sync(bs->file,
273 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
274 &data64, sizeof(data64));
275 if (ret < 0) {
276 goto fail_block;
277 }
278
279 s->refcount_table[refcount_table_index] = new_block;
280 return 0;
281 }
282
283 ret = qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
284 if (ret < 0) {
285 goto fail_block;
286 }
287
288 /*
289 * If we come here, we need to grow the refcount table. Again, a new
290 * refcount table needs some space and we can't simply allocate to avoid
291 * endless recursion.
292 *
293 * Therefore let's grab new refcount blocks at the end of the image, which
294 * will describe themselves and the new refcount table. This way we can
295 * reference them only in the new table and do the switch to the new
296 * refcount table at once without producing an inconsistent state in
297 * between.
298 */
299 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
300
301 /* Calculate the number of refcount blocks needed so far */
302 uint64_t refcount_block_clusters = 1 << (s->cluster_bits - REFCOUNT_SHIFT);
303 uint64_t blocks_used = (s->free_cluster_index +
304 refcount_block_clusters - 1) / refcount_block_clusters;
305
306 /* And now we need at least one block more for the new metadata */
307 uint64_t table_size = next_refcount_table_size(s, blocks_used + 1);
308 uint64_t last_table_size;
309 uint64_t blocks_clusters;
310 do {
311 uint64_t table_clusters =
312 size_to_clusters(s, table_size * sizeof(uint64_t));
313 blocks_clusters = 1 +
314 ((table_clusters + refcount_block_clusters - 1)
315 / refcount_block_clusters);
316 uint64_t meta_clusters = table_clusters + blocks_clusters;
317
318 last_table_size = table_size;
319 table_size = next_refcount_table_size(s, blocks_used +
320 ((meta_clusters + refcount_block_clusters - 1)
321 / refcount_block_clusters));
322
323 } while (last_table_size != table_size);
324
325 #ifdef DEBUG_ALLOC2
326 fprintf(stderr, "qcow2: Grow refcount table %" PRId32 " => %" PRId64 "\n",
327 s->refcount_table_size, table_size);
328 #endif
329
330 /* Create the new refcount table and blocks */
331 uint64_t meta_offset = (blocks_used * refcount_block_clusters) *
332 s->cluster_size;
333 uint64_t table_offset = meta_offset + blocks_clusters * s->cluster_size;
334 uint16_t *new_blocks = g_malloc0(blocks_clusters * s->cluster_size);
335 uint64_t *new_table = g_malloc0(table_size * sizeof(uint64_t));
336
337 assert(meta_offset >= (s->free_cluster_index * s->cluster_size));
338
339 /* Fill the new refcount table */
340 memcpy(new_table, s->refcount_table,
341 s->refcount_table_size * sizeof(uint64_t));
342 new_table[refcount_table_index] = new_block;
343
344 int i;
345 for (i = 0; i < blocks_clusters; i++) {
346 new_table[blocks_used + i] = meta_offset + (i * s->cluster_size);
347 }
348
349 /* Fill the refcount blocks */
350 uint64_t table_clusters = size_to_clusters(s, table_size * sizeof(uint64_t));
351 int block = 0;
352 for (i = 0; i < table_clusters + blocks_clusters; i++) {
353 new_blocks[block++] = cpu_to_be16(1);
354 }
355
356 /* Write refcount blocks to disk */
357 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
358 ret = bdrv_pwrite_sync(bs->file, meta_offset, new_blocks,
359 blocks_clusters * s->cluster_size);
360 g_free(new_blocks);
361 if (ret < 0) {
362 goto fail_table;
363 }
364
365 /* Write refcount table to disk */
366 for(i = 0; i < table_size; i++) {
367 cpu_to_be64s(&new_table[i]);
368 }
369
370 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
371 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
372 table_size * sizeof(uint64_t));
373 if (ret < 0) {
374 goto fail_table;
375 }
376
377 for(i = 0; i < table_size; i++) {
378 be64_to_cpus(&new_table[i]);
379 }
380
381 /* Hook up the new refcount table in the qcow2 header */
382 uint8_t data[12];
383 cpu_to_be64w((uint64_t*)data, table_offset);
384 cpu_to_be32w((uint32_t*)(data + 8), table_clusters);
385 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
386 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, refcount_table_offset),
387 data, sizeof(data));
388 if (ret < 0) {
389 goto fail_table;
390 }
391
392 /* And switch it in memory */
393 uint64_t old_table_offset = s->refcount_table_offset;
394 uint64_t old_table_size = s->refcount_table_size;
395
396 g_free(s->refcount_table);
397 s->refcount_table = new_table;
398 s->refcount_table_size = table_size;
399 s->refcount_table_offset = table_offset;
400
401 /* Free old table. Remember, we must not change free_cluster_index */
402 uint64_t old_free_cluster_index = s->free_cluster_index;
403 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
404 QCOW2_DISCARD_OTHER);
405 s->free_cluster_index = old_free_cluster_index;
406
407 ret = load_refcount_block(bs, new_block, (void**) refcount_block);
408 if (ret < 0) {
409 return ret;
410 }
411
412 return 0;
413
414 fail_table:
415 g_free(new_table);
416 fail_block:
417 if (*refcount_block != NULL) {
418 qcow2_cache_put(bs, s->refcount_block_cache, (void**) refcount_block);
419 }
420 return ret;
421 }
422
423 /* XXX: cache several refcount block clusters ? */
424 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
425 int64_t offset, int64_t length, int addend, enum qcow2_discard_type type)
426 {
427 BDRVQcowState *s = bs->opaque;
428 int64_t start, last, cluster_offset;
429 uint16_t *refcount_block = NULL;
430 int64_t old_table_index = -1;
431 int ret;
432
433 #ifdef DEBUG_ALLOC2
434 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 " addend=%d\n",
435 offset, length, addend);
436 #endif
437 if (length < 0) {
438 return -EINVAL;
439 } else if (length == 0) {
440 return 0;
441 }
442
443 if (addend < 0) {
444 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
445 s->l2_table_cache);
446 }
447
448 start = offset & ~(s->cluster_size - 1);
449 last = (offset + length - 1) & ~(s->cluster_size - 1);
450 for(cluster_offset = start; cluster_offset <= last;
451 cluster_offset += s->cluster_size)
452 {
453 int block_index, refcount;
454 int64_t cluster_index = cluster_offset >> s->cluster_bits;
455 int64_t table_index =
456 cluster_index >> (s->cluster_bits - REFCOUNT_SHIFT);
457
458 /* Load the refcount block and allocate it if needed */
459 if (table_index != old_table_index) {
460 if (refcount_block) {
461 ret = qcow2_cache_put(bs, s->refcount_block_cache,
462 (void**) &refcount_block);
463 if (ret < 0) {
464 goto fail;
465 }
466 }
467
468 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
469 if (ret < 0) {
470 goto fail;
471 }
472 }
473 old_table_index = table_index;
474
475 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
476
477 /* we can update the count and save it */
478 block_index = cluster_index &
479 ((1 << (s->cluster_bits - REFCOUNT_SHIFT)) - 1);
480
481 refcount = be16_to_cpu(refcount_block[block_index]);
482 refcount += addend;
483 if (refcount < 0 || refcount > 0xffff) {
484 ret = -EINVAL;
485 goto fail;
486 }
487 if (refcount == 0 && cluster_index < s->free_cluster_index) {
488 s->free_cluster_index = cluster_index;
489 }
490 refcount_block[block_index] = cpu_to_be16(refcount);
491 if (refcount == 0 && s->discard_passthrough[type]) {
492 /* Try discarding, ignore errors */
493 /* FIXME Doing this cluster by cluster will be painfully slow */
494 bdrv_discard(bs->file, cluster_offset, 1);
495 }
496 }
497
498 ret = 0;
499 fail:
500 /* Write last changed block to disk */
501 if (refcount_block) {
502 int wret;
503 wret = qcow2_cache_put(bs, s->refcount_block_cache,
504 (void**) &refcount_block);
505 if (wret < 0) {
506 return ret < 0 ? ret : wret;
507 }
508 }
509
510 /*
511 * Try do undo any updates if an error is returned (This may succeed in
512 * some cases like ENOSPC for allocating a new refcount block)
513 */
514 if (ret < 0) {
515 int dummy;
516 dummy = update_refcount(bs, offset, cluster_offset - offset, -addend,
517 QCOW2_DISCARD_NEVER);
518 (void)dummy;
519 }
520
521 return ret;
522 }
523
524 /*
525 * Increases or decreases the refcount of a given cluster by one.
526 * addend must be 1 or -1.
527 *
528 * If the return value is non-negative, it is the new refcount of the cluster.
529 * If it is negative, it is -errno and indicates an error.
530 */
531 static int update_cluster_refcount(BlockDriverState *bs,
532 int64_t cluster_index,
533 int addend,
534 enum qcow2_discard_type type)
535 {
536 BDRVQcowState *s = bs->opaque;
537 int ret;
538
539 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
540 type);
541 if (ret < 0) {
542 return ret;
543 }
544
545 return get_refcount(bs, cluster_index);
546 }
547
548
549
550 /*********************************************************/
551 /* cluster allocation functions */
552
553
554
555 /* return < 0 if error */
556 static int64_t alloc_clusters_noref(BlockDriverState *bs, int64_t size)
557 {
558 BDRVQcowState *s = bs->opaque;
559 int i, nb_clusters, refcount;
560
561 nb_clusters = size_to_clusters(s, size);
562 retry:
563 for(i = 0; i < nb_clusters; i++) {
564 int64_t next_cluster_index = s->free_cluster_index++;
565 refcount = get_refcount(bs, next_cluster_index);
566
567 if (refcount < 0) {
568 return refcount;
569 } else if (refcount != 0) {
570 goto retry;
571 }
572 }
573 #ifdef DEBUG_ALLOC2
574 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
575 size,
576 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
577 #endif
578 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
579 }
580
581 int64_t qcow2_alloc_clusters(BlockDriverState *bs, int64_t size)
582 {
583 int64_t offset;
584 int ret;
585
586 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
587 offset = alloc_clusters_noref(bs, size);
588 if (offset < 0) {
589 return offset;
590 }
591
592 ret = update_refcount(bs, offset, size, 1, QCOW2_DISCARD_NEVER);
593 if (ret < 0) {
594 return ret;
595 }
596
597 return offset;
598 }
599
600 int qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
601 int nb_clusters)
602 {
603 BDRVQcowState *s = bs->opaque;
604 uint64_t cluster_index;
605 uint64_t old_free_cluster_index;
606 int i, refcount, ret;
607
608 /* Check how many clusters there are free */
609 cluster_index = offset >> s->cluster_bits;
610 for(i = 0; i < nb_clusters; i++) {
611 refcount = get_refcount(bs, cluster_index++);
612
613 if (refcount < 0) {
614 return refcount;
615 } else if (refcount != 0) {
616 break;
617 }
618 }
619
620 /* And then allocate them */
621 old_free_cluster_index = s->free_cluster_index;
622 s->free_cluster_index = cluster_index + i;
623
624 ret = update_refcount(bs, offset, i << s->cluster_bits, 1,
625 QCOW2_DISCARD_NEVER);
626 if (ret < 0) {
627 return ret;
628 }
629
630 s->free_cluster_index = old_free_cluster_index;
631
632 return i;
633 }
634
635 /* only used to allocate compressed sectors. We try to allocate
636 contiguous sectors. size must be <= cluster_size */
637 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
638 {
639 BDRVQcowState *s = bs->opaque;
640 int64_t offset, cluster_offset;
641 int free_in_cluster;
642
643 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
644 assert(size > 0 && size <= s->cluster_size);
645 if (s->free_byte_offset == 0) {
646 offset = qcow2_alloc_clusters(bs, s->cluster_size);
647 if (offset < 0) {
648 return offset;
649 }
650 s->free_byte_offset = offset;
651 }
652 redo:
653 free_in_cluster = s->cluster_size -
654 (s->free_byte_offset & (s->cluster_size - 1));
655 if (size <= free_in_cluster) {
656 /* enough space in current cluster */
657 offset = s->free_byte_offset;
658 s->free_byte_offset += size;
659 free_in_cluster -= size;
660 if (free_in_cluster == 0)
661 s->free_byte_offset = 0;
662 if ((offset & (s->cluster_size - 1)) != 0)
663 update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
664 QCOW2_DISCARD_NEVER);
665 } else {
666 offset = qcow2_alloc_clusters(bs, s->cluster_size);
667 if (offset < 0) {
668 return offset;
669 }
670 cluster_offset = s->free_byte_offset & ~(s->cluster_size - 1);
671 if ((cluster_offset + s->cluster_size) == offset) {
672 /* we are lucky: contiguous data */
673 offset = s->free_byte_offset;
674 update_cluster_refcount(bs, offset >> s->cluster_bits, 1,
675 QCOW2_DISCARD_NEVER);
676 s->free_byte_offset += size;
677 } else {
678 s->free_byte_offset = offset;
679 goto redo;
680 }
681 }
682
683 /* The cluster refcount was incremented, either by qcow2_alloc_clusters()
684 * or explicitly by update_cluster_refcount(). Refcount blocks must be
685 * flushed before the caller's L2 table updates.
686 */
687 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
688 return offset;
689 }
690
691 void qcow2_free_clusters(BlockDriverState *bs,
692 int64_t offset, int64_t size,
693 enum qcow2_discard_type type)
694 {
695 int ret;
696
697 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
698 ret = update_refcount(bs, offset, size, -1, type);
699 if (ret < 0) {
700 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
701 /* TODO Remember the clusters to free them later and avoid leaking */
702 }
703 }
704
705 /*
706 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
707 * normal cluster, compressed cluster, etc.)
708 */
709 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
710 int nb_clusters, enum qcow2_discard_type type)
711 {
712 BDRVQcowState *s = bs->opaque;
713
714 switch (qcow2_get_cluster_type(l2_entry)) {
715 case QCOW2_CLUSTER_COMPRESSED:
716 {
717 int nb_csectors;
718 nb_csectors = ((l2_entry >> s->csize_shift) &
719 s->csize_mask) + 1;
720 qcow2_free_clusters(bs,
721 (l2_entry & s->cluster_offset_mask) & ~511,
722 nb_csectors * 512, type);
723 }
724 break;
725 case QCOW2_CLUSTER_NORMAL:
726 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
727 nb_clusters << s->cluster_bits, type);
728 break;
729 case QCOW2_CLUSTER_UNALLOCATED:
730 case QCOW2_CLUSTER_ZERO:
731 break;
732 default:
733 abort();
734 }
735 }
736
737
738
739 /*********************************************************/
740 /* snapshots and image creation */
741
742
743
744 /* update the refcounts of snapshots and the copied flag */
745 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
746 int64_t l1_table_offset, int l1_size, int addend)
747 {
748 BDRVQcowState *s = bs->opaque;
749 uint64_t *l1_table, *l2_table, l2_offset, offset, l1_size2, l1_allocated;
750 int64_t old_offset, old_l2_offset;
751 int i, j, l1_modified = 0, nb_csectors, refcount;
752 int ret;
753
754 l2_table = NULL;
755 l1_table = NULL;
756 l1_size2 = l1_size * sizeof(uint64_t);
757
758 /* WARNING: qcow2_snapshot_goto relies on this function not using the
759 * l1_table_offset when it is the current s->l1_table_offset! Be careful
760 * when changing this! */
761 if (l1_table_offset != s->l1_table_offset) {
762 l1_table = g_malloc0(align_offset(l1_size2, 512));
763 l1_allocated = 1;
764
765 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
766 if (ret < 0) {
767 goto fail;
768 }
769
770 for(i = 0;i < l1_size; i++)
771 be64_to_cpus(&l1_table[i]);
772 } else {
773 assert(l1_size == s->l1_size);
774 l1_table = s->l1_table;
775 l1_allocated = 0;
776 }
777
778 for(i = 0; i < l1_size; i++) {
779 l2_offset = l1_table[i];
780 if (l2_offset) {
781 old_l2_offset = l2_offset;
782 l2_offset &= L1E_OFFSET_MASK;
783
784 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
785 (void**) &l2_table);
786 if (ret < 0) {
787 goto fail;
788 }
789
790 for(j = 0; j < s->l2_size; j++) {
791 offset = be64_to_cpu(l2_table[j]);
792 if (offset != 0) {
793 old_offset = offset;
794 offset &= ~QCOW_OFLAG_COPIED;
795 if (offset & QCOW_OFLAG_COMPRESSED) {
796 nb_csectors = ((offset >> s->csize_shift) &
797 s->csize_mask) + 1;
798 if (addend != 0) {
799 int ret;
800 ret = update_refcount(bs,
801 (offset & s->cluster_offset_mask) & ~511,
802 nb_csectors * 512, addend,
803 QCOW2_DISCARD_SNAPSHOT);
804 if (ret < 0) {
805 goto fail;
806 }
807 }
808 /* compressed clusters are never modified */
809 refcount = 2;
810 } else {
811 uint64_t cluster_index = (offset & L2E_OFFSET_MASK) >> s->cluster_bits;
812 if (addend != 0) {
813 refcount = update_cluster_refcount(bs, cluster_index, addend,
814 QCOW2_DISCARD_SNAPSHOT);
815 } else {
816 refcount = get_refcount(bs, cluster_index);
817 }
818
819 if (refcount < 0) {
820 ret = refcount;
821 goto fail;
822 }
823 }
824
825 if (refcount == 1) {
826 offset |= QCOW_OFLAG_COPIED;
827 }
828 if (offset != old_offset) {
829 if (addend > 0) {
830 qcow2_cache_set_dependency(bs, s->l2_table_cache,
831 s->refcount_block_cache);
832 }
833 l2_table[j] = cpu_to_be64(offset);
834 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
835 }
836 }
837 }
838
839 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
840 if (ret < 0) {
841 goto fail;
842 }
843
844
845 if (addend != 0) {
846 refcount = update_cluster_refcount(bs, l2_offset >> s->cluster_bits, addend,
847 QCOW2_DISCARD_SNAPSHOT);
848 } else {
849 refcount = get_refcount(bs, l2_offset >> s->cluster_bits);
850 }
851 if (refcount < 0) {
852 ret = refcount;
853 goto fail;
854 } else if (refcount == 1) {
855 l2_offset |= QCOW_OFLAG_COPIED;
856 }
857 if (l2_offset != old_l2_offset) {
858 l1_table[i] = l2_offset;
859 l1_modified = 1;
860 }
861 }
862 }
863
864 ret = bdrv_flush(bs);
865 fail:
866 if (l2_table) {
867 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
868 }
869
870 /* Update L1 only if it isn't deleted anyway (addend = -1) */
871 if (ret == 0 && addend >= 0 && l1_modified) {
872 for (i = 0; i < l1_size; i++) {
873 cpu_to_be64s(&l1_table[i]);
874 }
875
876 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table, l1_size2);
877
878 for (i = 0; i < l1_size; i++) {
879 be64_to_cpus(&l1_table[i]);
880 }
881 }
882 if (l1_allocated)
883 g_free(l1_table);
884 return ret;
885 }
886
887
888
889
890 /*********************************************************/
891 /* refcount checking functions */
892
893
894
895 /*
896 * Increases the refcount for a range of clusters in a given refcount table.
897 * This is used to construct a temporary refcount table out of L1 and L2 tables
898 * which can be compared the the refcount table saved in the image.
899 *
900 * Modifies the number of errors in res.
901 */
902 static void inc_refcounts(BlockDriverState *bs,
903 BdrvCheckResult *res,
904 uint16_t *refcount_table,
905 int refcount_table_size,
906 int64_t offset, int64_t size)
907 {
908 BDRVQcowState *s = bs->opaque;
909 int64_t start, last, cluster_offset;
910 int k;
911
912 if (size <= 0)
913 return;
914
915 start = offset & ~(s->cluster_size - 1);
916 last = (offset + size - 1) & ~(s->cluster_size - 1);
917 for(cluster_offset = start; cluster_offset <= last;
918 cluster_offset += s->cluster_size) {
919 k = cluster_offset >> s->cluster_bits;
920 if (k < 0) {
921 fprintf(stderr, "ERROR: invalid cluster offset=0x%" PRIx64 "\n",
922 cluster_offset);
923 res->corruptions++;
924 } else if (k >= refcount_table_size) {
925 fprintf(stderr, "Warning: cluster offset=0x%" PRIx64 " is after "
926 "the end of the image file, can't properly check refcounts.\n",
927 cluster_offset);
928 res->check_errors++;
929 } else {
930 if (++refcount_table[k] == 0) {
931 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
932 "\n", cluster_offset);
933 res->corruptions++;
934 }
935 }
936 }
937 }
938
939 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
940 enum {
941 CHECK_OFLAG_COPIED = 0x1, /* check QCOW_OFLAG_COPIED matches refcount */
942 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
943 };
944
945 /*
946 * Increases the refcount in the given refcount table for the all clusters
947 * referenced in the L2 table. While doing so, performs some checks on L2
948 * entries.
949 *
950 * Returns the number of errors found by the checks or -errno if an internal
951 * error occurred.
952 */
953 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
954 uint16_t *refcount_table, int refcount_table_size, int64_t l2_offset,
955 int flags)
956 {
957 BDRVQcowState *s = bs->opaque;
958 uint64_t *l2_table, l2_entry;
959 uint64_t next_contiguous_offset = 0;
960 int i, l2_size, nb_csectors, refcount;
961
962 /* Read L2 table from disk */
963 l2_size = s->l2_size * sizeof(uint64_t);
964 l2_table = g_malloc(l2_size);
965
966 if (bdrv_pread(bs->file, l2_offset, l2_table, l2_size) != l2_size)
967 goto fail;
968
969 /* Do the actual checks */
970 for(i = 0; i < s->l2_size; i++) {
971 l2_entry = be64_to_cpu(l2_table[i]);
972
973 switch (qcow2_get_cluster_type(l2_entry)) {
974 case QCOW2_CLUSTER_COMPRESSED:
975 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
976 if (l2_entry & QCOW_OFLAG_COPIED) {
977 fprintf(stderr, "ERROR: cluster %" PRId64 ": "
978 "copied flag must never be set for compressed "
979 "clusters\n", l2_entry >> s->cluster_bits);
980 l2_entry &= ~QCOW_OFLAG_COPIED;
981 res->corruptions++;
982 }
983
984 /* Mark cluster as used */
985 nb_csectors = ((l2_entry >> s->csize_shift) &
986 s->csize_mask) + 1;
987 l2_entry &= s->cluster_offset_mask;
988 inc_refcounts(bs, res, refcount_table, refcount_table_size,
989 l2_entry & ~511, nb_csectors * 512);
990
991 if (flags & CHECK_FRAG_INFO) {
992 res->bfi.allocated_clusters++;
993 res->bfi.compressed_clusters++;
994
995 /* Compressed clusters are fragmented by nature. Since they
996 * take up sub-sector space but we only have sector granularity
997 * I/O we need to re-read the same sectors even for adjacent
998 * compressed clusters.
999 */
1000 res->bfi.fragmented_clusters++;
1001 }
1002 break;
1003
1004 case QCOW2_CLUSTER_ZERO:
1005 if ((l2_entry & L2E_OFFSET_MASK) == 0) {
1006 break;
1007 }
1008 /* fall through */
1009
1010 case QCOW2_CLUSTER_NORMAL:
1011 {
1012 /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1013 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1014
1015 if (flags & CHECK_OFLAG_COPIED) {
1016 refcount = get_refcount(bs, offset >> s->cluster_bits);
1017 if (refcount < 0) {
1018 fprintf(stderr, "Can't get refcount for offset %"
1019 PRIx64 ": %s\n", l2_entry, strerror(-refcount));
1020 goto fail;
1021 }
1022 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1023 fprintf(stderr, "ERROR OFLAG_COPIED: offset=%"
1024 PRIx64 " refcount=%d\n", l2_entry, refcount);
1025 res->corruptions++;
1026 }
1027 }
1028
1029 if (flags & CHECK_FRAG_INFO) {
1030 res->bfi.allocated_clusters++;
1031 if (next_contiguous_offset &&
1032 offset != next_contiguous_offset) {
1033 res->bfi.fragmented_clusters++;
1034 }
1035 next_contiguous_offset = offset + s->cluster_size;
1036 }
1037
1038 /* Mark cluster as used */
1039 inc_refcounts(bs, res, refcount_table,refcount_table_size,
1040 offset, s->cluster_size);
1041
1042 /* Correct offsets are cluster aligned */
1043 if (offset & (s->cluster_size - 1)) {
1044 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not "
1045 "properly aligned; L2 entry corrupted.\n", offset);
1046 res->corruptions++;
1047 }
1048 break;
1049 }
1050
1051 case QCOW2_CLUSTER_UNALLOCATED:
1052 break;
1053
1054 default:
1055 abort();
1056 }
1057 }
1058
1059 g_free(l2_table);
1060 return 0;
1061
1062 fail:
1063 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1064 g_free(l2_table);
1065 return -EIO;
1066 }
1067
1068 /*
1069 * Increases the refcount for the L1 table, its L2 tables and all referenced
1070 * clusters in the given refcount table. While doing so, performs some checks
1071 * on L1 and L2 entries.
1072 *
1073 * Returns the number of errors found by the checks or -errno if an internal
1074 * error occurred.
1075 */
1076 static int check_refcounts_l1(BlockDriverState *bs,
1077 BdrvCheckResult *res,
1078 uint16_t *refcount_table,
1079 int refcount_table_size,
1080 int64_t l1_table_offset, int l1_size,
1081 int flags)
1082 {
1083 BDRVQcowState *s = bs->opaque;
1084 uint64_t *l1_table, l2_offset, l1_size2;
1085 int i, refcount, ret;
1086
1087 l1_size2 = l1_size * sizeof(uint64_t);
1088
1089 /* Mark L1 table as used */
1090 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1091 l1_table_offset, l1_size2);
1092
1093 /* Read L1 table entries from disk */
1094 if (l1_size2 == 0) {
1095 l1_table = NULL;
1096 } else {
1097 l1_table = g_malloc(l1_size2);
1098 if (bdrv_pread(bs->file, l1_table_offset,
1099 l1_table, l1_size2) != l1_size2)
1100 goto fail;
1101 for(i = 0;i < l1_size; i++)
1102 be64_to_cpus(&l1_table[i]);
1103 }
1104
1105 /* Do the actual checks */
1106 for(i = 0; i < l1_size; i++) {
1107 l2_offset = l1_table[i];
1108 if (l2_offset) {
1109 /* QCOW_OFLAG_COPIED must be set iff refcount == 1 */
1110 if (flags & CHECK_OFLAG_COPIED) {
1111 refcount = get_refcount(bs, (l2_offset & ~QCOW_OFLAG_COPIED)
1112 >> s->cluster_bits);
1113 if (refcount < 0) {
1114 fprintf(stderr, "Can't get refcount for l2_offset %"
1115 PRIx64 ": %s\n", l2_offset, strerror(-refcount));
1116 goto fail;
1117 }
1118 if ((refcount == 1) != ((l2_offset & QCOW_OFLAG_COPIED) != 0)) {
1119 fprintf(stderr, "ERROR OFLAG_COPIED: l2_offset=%" PRIx64
1120 " refcount=%d\n", l2_offset, refcount);
1121 res->corruptions++;
1122 }
1123 }
1124
1125 /* Mark L2 table as used */
1126 l2_offset &= L1E_OFFSET_MASK;
1127 inc_refcounts(bs, res, refcount_table, refcount_table_size,
1128 l2_offset, s->cluster_size);
1129
1130 /* L2 tables are cluster aligned */
1131 if (l2_offset & (s->cluster_size - 1)) {
1132 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1133 "cluster aligned; L1 entry corrupted\n", l2_offset);
1134 res->corruptions++;
1135 }
1136
1137 /* Process and check L2 entries */
1138 ret = check_refcounts_l2(bs, res, refcount_table,
1139 refcount_table_size, l2_offset, flags);
1140 if (ret < 0) {
1141 goto fail;
1142 }
1143 }
1144 }
1145 g_free(l1_table);
1146 return 0;
1147
1148 fail:
1149 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1150 res->check_errors++;
1151 g_free(l1_table);
1152 return -EIO;
1153 }
1154
1155 /*
1156 * Checks an image for refcount consistency.
1157 *
1158 * Returns 0 if no errors are found, the number of errors in case the image is
1159 * detected as corrupted, and -errno when an internal error occurred.
1160 */
1161 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
1162 BdrvCheckMode fix)
1163 {
1164 BDRVQcowState *s = bs->opaque;
1165 int64_t size, i, highest_cluster;
1166 int nb_clusters, refcount1, refcount2;
1167 QCowSnapshot *sn;
1168 uint16_t *refcount_table;
1169 int ret;
1170
1171 size = bdrv_getlength(bs->file);
1172 nb_clusters = size_to_clusters(s, size);
1173 refcount_table = g_malloc0(nb_clusters * sizeof(uint16_t));
1174
1175 res->bfi.total_clusters =
1176 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
1177
1178 /* header */
1179 inc_refcounts(bs, res, refcount_table, nb_clusters,
1180 0, s->cluster_size);
1181
1182 /* current L1 table */
1183 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1184 s->l1_table_offset, s->l1_size,
1185 CHECK_OFLAG_COPIED | CHECK_FRAG_INFO);
1186 if (ret < 0) {
1187 goto fail;
1188 }
1189
1190 /* snapshots */
1191 for(i = 0; i < s->nb_snapshots; i++) {
1192 sn = s->snapshots + i;
1193 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
1194 sn->l1_table_offset, sn->l1_size, 0);
1195 if (ret < 0) {
1196 goto fail;
1197 }
1198 }
1199 inc_refcounts(bs, res, refcount_table, nb_clusters,
1200 s->snapshots_offset, s->snapshots_size);
1201
1202 /* refcount data */
1203 inc_refcounts(bs, res, refcount_table, nb_clusters,
1204 s->refcount_table_offset,
1205 s->refcount_table_size * sizeof(uint64_t));
1206
1207 for(i = 0; i < s->refcount_table_size; i++) {
1208 uint64_t offset, cluster;
1209 offset = s->refcount_table[i];
1210 cluster = offset >> s->cluster_bits;
1211
1212 /* Refcount blocks are cluster aligned */
1213 if (offset & (s->cluster_size - 1)) {
1214 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1215 "cluster aligned; refcount table entry corrupted\n", i);
1216 res->corruptions++;
1217 continue;
1218 }
1219
1220 if (cluster >= nb_clusters) {
1221 fprintf(stderr, "ERROR refcount block %" PRId64
1222 " is outside image\n", i);
1223 res->corruptions++;
1224 continue;
1225 }
1226
1227 if (offset != 0) {
1228 inc_refcounts(bs, res, refcount_table, nb_clusters,
1229 offset, s->cluster_size);
1230 if (refcount_table[cluster] != 1) {
1231 fprintf(stderr, "ERROR refcount block %" PRId64
1232 " refcount=%d\n",
1233 i, refcount_table[cluster]);
1234 res->corruptions++;
1235 }
1236 }
1237 }
1238
1239 /* compare ref counts */
1240 for (i = 0, highest_cluster = 0; i < nb_clusters; i++) {
1241 refcount1 = get_refcount(bs, i);
1242 if (refcount1 < 0) {
1243 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
1244 i, strerror(-refcount1));
1245 res->check_errors++;
1246 continue;
1247 }
1248
1249 refcount2 = refcount_table[i];
1250
1251 if (refcount1 > 0 || refcount2 > 0) {
1252 highest_cluster = i;
1253 }
1254
1255 if (refcount1 != refcount2) {
1256
1257 /* Check if we're allowed to fix the mismatch */
1258 int *num_fixed = NULL;
1259 if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
1260 num_fixed = &res->leaks_fixed;
1261 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
1262 num_fixed = &res->corruptions_fixed;
1263 }
1264
1265 fprintf(stderr, "%s cluster %" PRId64 " refcount=%d reference=%d\n",
1266 num_fixed != NULL ? "Repairing" :
1267 refcount1 < refcount2 ? "ERROR" :
1268 "Leaked",
1269 i, refcount1, refcount2);
1270
1271 if (num_fixed) {
1272 ret = update_refcount(bs, i << s->cluster_bits, 1,
1273 refcount2 - refcount1,
1274 QCOW2_DISCARD_ALWAYS);
1275 if (ret >= 0) {
1276 (*num_fixed)++;
1277 continue;
1278 }
1279 }
1280
1281 /* And if we couldn't, print an error */
1282 if (refcount1 < refcount2) {
1283 res->corruptions++;
1284 } else {
1285 res->leaks++;
1286 }
1287 }
1288 }
1289
1290 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
1291 ret = 0;
1292
1293 fail:
1294 g_free(refcount_table);
1295
1296 return ret;
1297 }
1298