rebuild-expected-aml.sh: remind about the process
[qemu/kevin.git] / block / qcow2-refcount.c
blob7ef1c0e42ad018c3bef2be7497e9e3cb441593c3
1 /*
2 * Block driver for the QCOW version 2 format
4 * Copyright (c) 2004-2006 Fabrice Bellard
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:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
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.
25 #include "qemu/osdep.h"
26 #include "qapi/error.h"
27 #include "qcow2.h"
28 #include "qemu/range.h"
29 #include "qemu/bswap.h"
30 #include "qemu/cutils.h"
31 #include "trace.h"
33 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
34 uint64_t max);
35 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
36 int64_t offset, int64_t length, uint64_t addend,
37 bool decrease, enum qcow2_discard_type type);
39 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index);
40 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index);
41 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index);
42 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index);
43 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index);
44 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index);
45 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index);
47 static void set_refcount_ro0(void *refcount_array, uint64_t index,
48 uint64_t value);
49 static void set_refcount_ro1(void *refcount_array, uint64_t index,
50 uint64_t value);
51 static void set_refcount_ro2(void *refcount_array, uint64_t index,
52 uint64_t value);
53 static void set_refcount_ro3(void *refcount_array, uint64_t index,
54 uint64_t value);
55 static void set_refcount_ro4(void *refcount_array, uint64_t index,
56 uint64_t value);
57 static void set_refcount_ro5(void *refcount_array, uint64_t index,
58 uint64_t value);
59 static void set_refcount_ro6(void *refcount_array, uint64_t index,
60 uint64_t value);
63 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = {
64 &get_refcount_ro0,
65 &get_refcount_ro1,
66 &get_refcount_ro2,
67 &get_refcount_ro3,
68 &get_refcount_ro4,
69 &get_refcount_ro5,
70 &get_refcount_ro6
73 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = {
74 &set_refcount_ro0,
75 &set_refcount_ro1,
76 &set_refcount_ro2,
77 &set_refcount_ro3,
78 &set_refcount_ro4,
79 &set_refcount_ro5,
80 &set_refcount_ro6
84 /*********************************************************/
85 /* refcount handling */
87 static void update_max_refcount_table_index(BDRVQcow2State *s)
89 unsigned i = s->refcount_table_size - 1;
90 while (i > 0 && (s->refcount_table[i] & REFT_OFFSET_MASK) == 0) {
91 i--;
93 /* Set s->max_refcount_table_index to the index of the last used entry */
94 s->max_refcount_table_index = i;
97 int qcow2_refcount_init(BlockDriverState *bs)
99 BDRVQcow2State *s = bs->opaque;
100 unsigned int refcount_table_size2, i;
101 int ret;
103 assert(s->refcount_order >= 0 && s->refcount_order <= 6);
105 s->get_refcount = get_refcount_funcs[s->refcount_order];
106 s->set_refcount = set_refcount_funcs[s->refcount_order];
108 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
109 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
110 s->refcount_table = g_try_malloc(refcount_table_size2);
112 if (s->refcount_table_size > 0) {
113 if (s->refcount_table == NULL) {
114 ret = -ENOMEM;
115 goto fail;
117 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
118 ret = bdrv_pread(bs->file, s->refcount_table_offset,
119 s->refcount_table, refcount_table_size2);
120 if (ret < 0) {
121 goto fail;
123 for(i = 0; i < s->refcount_table_size; i++)
124 be64_to_cpus(&s->refcount_table[i]);
125 update_max_refcount_table_index(s);
127 return 0;
128 fail:
129 return ret;
132 void qcow2_refcount_close(BlockDriverState *bs)
134 BDRVQcow2State *s = bs->opaque;
135 g_free(s->refcount_table);
139 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index)
141 return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1;
144 static void set_refcount_ro0(void *refcount_array, uint64_t index,
145 uint64_t value)
147 assert(!(value >> 1));
148 ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8));
149 ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8);
152 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index)
154 return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4)))
155 & 0x3;
158 static void set_refcount_ro1(void *refcount_array, uint64_t index,
159 uint64_t value)
161 assert(!(value >> 2));
162 ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4)));
163 ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4));
166 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index)
168 return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2)))
169 & 0xf;
172 static void set_refcount_ro2(void *refcount_array, uint64_t index,
173 uint64_t value)
175 assert(!(value >> 4));
176 ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2)));
177 ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2));
180 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index)
182 return ((const uint8_t *)refcount_array)[index];
185 static void set_refcount_ro3(void *refcount_array, uint64_t index,
186 uint64_t value)
188 assert(!(value >> 8));
189 ((uint8_t *)refcount_array)[index] = value;
192 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index)
194 return be16_to_cpu(((const uint16_t *)refcount_array)[index]);
197 static void set_refcount_ro4(void *refcount_array, uint64_t index,
198 uint64_t value)
200 assert(!(value >> 16));
201 ((uint16_t *)refcount_array)[index] = cpu_to_be16(value);
204 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index)
206 return be32_to_cpu(((const uint32_t *)refcount_array)[index]);
209 static void set_refcount_ro5(void *refcount_array, uint64_t index,
210 uint64_t value)
212 assert(!(value >> 32));
213 ((uint32_t *)refcount_array)[index] = cpu_to_be32(value);
216 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index)
218 return be64_to_cpu(((const uint64_t *)refcount_array)[index]);
221 static void set_refcount_ro6(void *refcount_array, uint64_t index,
222 uint64_t value)
224 ((uint64_t *)refcount_array)[index] = cpu_to_be64(value);
228 static int load_refcount_block(BlockDriverState *bs,
229 int64_t refcount_block_offset,
230 void **refcount_block)
232 BDRVQcow2State *s = bs->opaque;
234 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
235 return qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
236 refcount_block);
240 * Retrieves the refcount of the cluster given by its index and stores it in
241 * *refcount. Returns 0 on success and -errno on failure.
243 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index,
244 uint64_t *refcount)
246 BDRVQcow2State *s = bs->opaque;
247 uint64_t refcount_table_index, block_index;
248 int64_t refcount_block_offset;
249 int ret;
250 void *refcount_block;
252 refcount_table_index = cluster_index >> s->refcount_block_bits;
253 if (refcount_table_index >= s->refcount_table_size) {
254 *refcount = 0;
255 return 0;
257 refcount_block_offset =
258 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
259 if (!refcount_block_offset) {
260 *refcount = 0;
261 return 0;
264 if (offset_into_cluster(s, refcount_block_offset)) {
265 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64
266 " unaligned (reftable index: %#" PRIx64 ")",
267 refcount_block_offset, refcount_table_index);
268 return -EIO;
271 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
272 &refcount_block);
273 if (ret < 0) {
274 return ret;
277 block_index = cluster_index & (s->refcount_block_size - 1);
278 *refcount = s->get_refcount(refcount_block, block_index);
280 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
282 return 0;
285 /* Checks if two offsets are described by the same refcount block */
286 static int in_same_refcount_block(BDRVQcow2State *s, uint64_t offset_a,
287 uint64_t offset_b)
289 uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits);
290 uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits);
292 return (block_a == block_b);
296 * Loads a refcount block. If it doesn't exist yet, it is allocated first
297 * (including growing the refcount table if needed).
299 * Returns 0 on success or -errno in error case
301 static int alloc_refcount_block(BlockDriverState *bs,
302 int64_t cluster_index, void **refcount_block)
304 BDRVQcow2State *s = bs->opaque;
305 unsigned int refcount_table_index;
306 int64_t ret;
308 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
310 /* Find the refcount block for the given cluster */
311 refcount_table_index = cluster_index >> s->refcount_block_bits;
313 if (refcount_table_index < s->refcount_table_size) {
315 uint64_t refcount_block_offset =
316 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
318 /* If it's already there, we're done */
319 if (refcount_block_offset) {
320 if (offset_into_cluster(s, refcount_block_offset)) {
321 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
322 PRIx64 " unaligned (reftable index: "
323 "%#x)", refcount_block_offset,
324 refcount_table_index);
325 return -EIO;
328 return load_refcount_block(bs, refcount_block_offset,
329 refcount_block);
334 * If we came here, we need to allocate something. Something is at least
335 * a cluster for the new refcount block. It may also include a new refcount
336 * table if the old refcount table is too small.
338 * Note that allocating clusters here needs some special care:
340 * - We can't use the normal qcow2_alloc_clusters(), it would try to
341 * increase the refcount and very likely we would end up with an endless
342 * recursion. Instead we must place the refcount blocks in a way that
343 * they can describe them themselves.
345 * - We need to consider that at this point we are inside update_refcounts
346 * and potentially doing an initial refcount increase. This means that
347 * some clusters have already been allocated by the caller, but their
348 * refcount isn't accurate yet. If we allocate clusters for metadata, we
349 * need to return -EAGAIN to signal the caller that it needs to restart
350 * the search for free clusters.
352 * - alloc_clusters_noref and qcow2_free_clusters may load a different
353 * refcount block into the cache
356 *refcount_block = NULL;
358 /* We write to the refcount table, so we might depend on L2 tables */
359 ret = qcow2_cache_flush(bs, s->l2_table_cache);
360 if (ret < 0) {
361 return ret;
364 /* Allocate the refcount block itself and mark it as used */
365 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size, INT64_MAX);
366 if (new_block < 0) {
367 return new_block;
370 /* The offset must fit in the offset field of the refcount table entry */
371 assert((new_block & REFT_OFFSET_MASK) == new_block);
373 /* If we're allocating the block at offset 0 then something is wrong */
374 if (new_block == 0) {
375 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
376 "allocation of refcount block at offset 0");
377 return -EIO;
380 #ifdef DEBUG_ALLOC2
381 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
382 " at %" PRIx64 "\n",
383 refcount_table_index, cluster_index << s->cluster_bits, new_block);
384 #endif
386 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
387 /* Zero the new refcount block before updating it */
388 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
389 refcount_block);
390 if (ret < 0) {
391 goto fail;
394 memset(*refcount_block, 0, s->cluster_size);
396 /* The block describes itself, need to update the cache */
397 int block_index = (new_block >> s->cluster_bits) &
398 (s->refcount_block_size - 1);
399 s->set_refcount(*refcount_block, block_index, 1);
400 } else {
401 /* Described somewhere else. This can recurse at most twice before we
402 * arrive at a block that describes itself. */
403 ret = update_refcount(bs, new_block, s->cluster_size, 1, false,
404 QCOW2_DISCARD_NEVER);
405 if (ret < 0) {
406 goto fail;
409 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
410 if (ret < 0) {
411 goto fail;
414 /* Initialize the new refcount block only after updating its refcount,
415 * update_refcount uses the refcount cache itself */
416 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
417 refcount_block);
418 if (ret < 0) {
419 goto fail;
422 memset(*refcount_block, 0, s->cluster_size);
425 /* Now the new refcount block needs to be written to disk */
426 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
427 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
428 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
429 if (ret < 0) {
430 goto fail;
433 /* If the refcount table is big enough, just hook the block up there */
434 if (refcount_table_index < s->refcount_table_size) {
435 uint64_t data64 = cpu_to_be64(new_block);
436 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
437 ret = bdrv_pwrite_sync(bs->file,
438 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
439 &data64, sizeof(data64));
440 if (ret < 0) {
441 goto fail;
444 s->refcount_table[refcount_table_index] = new_block;
445 /* If there's a hole in s->refcount_table then it can happen
446 * that refcount_table_index < s->max_refcount_table_index */
447 s->max_refcount_table_index =
448 MAX(s->max_refcount_table_index, refcount_table_index);
450 /* The new refcount block may be where the caller intended to put its
451 * data, so let it restart the search. */
452 return -EAGAIN;
455 qcow2_cache_put(s->refcount_block_cache, refcount_block);
458 * If we come here, we need to grow the refcount table. Again, a new
459 * refcount table needs some space and we can't simply allocate to avoid
460 * endless recursion.
462 * Therefore let's grab new refcount blocks at the end of the image, which
463 * will describe themselves and the new refcount table. This way we can
464 * reference them only in the new table and do the switch to the new
465 * refcount table at once without producing an inconsistent state in
466 * between.
468 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
470 /* Calculate the number of refcount blocks needed so far; this will be the
471 * basis for calculating the index of the first cluster used for the
472 * self-describing refcount structures which we are about to create.
474 * Because we reached this point, there cannot be any refcount entries for
475 * cluster_index or higher indices yet. However, because new_block has been
476 * allocated to describe that cluster (and it will assume this role later
477 * on), we cannot use that index; also, new_block may actually have a higher
478 * cluster index than cluster_index, so it needs to be taken into account
479 * here (and 1 needs to be added to its value because that cluster is used).
481 uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1,
482 (new_block >> s->cluster_bits) + 1),
483 s->refcount_block_size);
485 /* Create the new refcount table and blocks */
486 uint64_t meta_offset = (blocks_used * s->refcount_block_size) *
487 s->cluster_size;
489 ret = qcow2_refcount_area(bs, meta_offset, 0, false,
490 refcount_table_index, new_block);
491 if (ret < 0) {
492 return ret;
495 ret = load_refcount_block(bs, new_block, refcount_block);
496 if (ret < 0) {
497 return ret;
500 /* If we were trying to do the initial refcount update for some cluster
501 * allocation, we might have used the same clusters to store newly
502 * allocated metadata. Make the caller search some new space. */
503 return -EAGAIN;
505 fail:
506 if (*refcount_block != NULL) {
507 qcow2_cache_put(s->refcount_block_cache, refcount_block);
509 return ret;
513 * Starting at @start_offset, this function creates new self-covering refcount
514 * structures: A new refcount table and refcount blocks which cover all of
515 * themselves, and a number of @additional_clusters beyond their end.
516 * @start_offset must be at the end of the image file, that is, there must be
517 * only empty space beyond it.
518 * If @exact_size is false, the refcount table will have 50 % more entries than
519 * necessary so it will not need to grow again soon.
520 * If @new_refblock_offset is not zero, it contains the offset of a refcount
521 * block that should be entered into the new refcount table at index
522 * @new_refblock_index.
524 * Returns: The offset after the new refcount structures (i.e. where the
525 * @additional_clusters may be placed) on success, -errno on error.
527 int64_t qcow2_refcount_area(BlockDriverState *bs, uint64_t start_offset,
528 uint64_t additional_clusters, bool exact_size,
529 int new_refblock_index,
530 uint64_t new_refblock_offset)
532 BDRVQcow2State *s = bs->opaque;
533 uint64_t total_refblock_count_u64, additional_refblock_count;
534 int total_refblock_count, table_size, area_reftable_index, table_clusters;
535 int i;
536 uint64_t table_offset, block_offset, end_offset;
537 int ret;
538 uint64_t *new_table;
540 assert(!(start_offset % s->cluster_size));
542 qcow2_refcount_metadata_size(start_offset / s->cluster_size +
543 additional_clusters,
544 s->cluster_size, s->refcount_order,
545 !exact_size, &total_refblock_count_u64);
546 if (total_refblock_count_u64 > QCOW_MAX_REFTABLE_SIZE) {
547 return -EFBIG;
549 total_refblock_count = total_refblock_count_u64;
551 /* Index in the refcount table of the first refcount block to cover the area
552 * of refcount structures we are about to create; we know that
553 * @total_refblock_count can cover @start_offset, so this will definitely
554 * fit into an int. */
555 area_reftable_index = (start_offset / s->cluster_size) /
556 s->refcount_block_size;
558 if (exact_size) {
559 table_size = total_refblock_count;
560 } else {
561 table_size = total_refblock_count +
562 DIV_ROUND_UP(total_refblock_count, 2);
564 /* The qcow2 file can only store the reftable size in number of clusters */
565 table_size = ROUND_UP(table_size, s->cluster_size / sizeof(uint64_t));
566 table_clusters = (table_size * sizeof(uint64_t)) / s->cluster_size;
568 if (table_size > QCOW_MAX_REFTABLE_SIZE) {
569 return -EFBIG;
572 new_table = g_try_new0(uint64_t, table_size);
574 assert(table_size > 0);
575 if (new_table == NULL) {
576 ret = -ENOMEM;
577 goto fail;
580 /* Fill the new refcount table */
581 if (table_size > s->max_refcount_table_index) {
582 /* We're actually growing the reftable */
583 memcpy(new_table, s->refcount_table,
584 (s->max_refcount_table_index + 1) * sizeof(uint64_t));
585 } else {
586 /* Improbable case: We're shrinking the reftable. However, the caller
587 * has assured us that there is only empty space beyond @start_offset,
588 * so we can simply drop all of the refblocks that won't fit into the
589 * new reftable. */
590 memcpy(new_table, s->refcount_table, table_size * sizeof(uint64_t));
593 if (new_refblock_offset) {
594 assert(new_refblock_index < total_refblock_count);
595 new_table[new_refblock_index] = new_refblock_offset;
598 /* Count how many new refblocks we have to create */
599 additional_refblock_count = 0;
600 for (i = area_reftable_index; i < total_refblock_count; i++) {
601 if (!new_table[i]) {
602 additional_refblock_count++;
606 table_offset = start_offset + additional_refblock_count * s->cluster_size;
607 end_offset = table_offset + table_clusters * s->cluster_size;
609 /* Fill the refcount blocks, and create new ones, if necessary */
610 block_offset = start_offset;
611 for (i = area_reftable_index; i < total_refblock_count; i++) {
612 void *refblock_data;
613 uint64_t first_offset_covered;
615 /* Reuse an existing refblock if possible, create a new one otherwise */
616 if (new_table[i]) {
617 ret = qcow2_cache_get(bs, s->refcount_block_cache, new_table[i],
618 &refblock_data);
619 if (ret < 0) {
620 goto fail;
622 } else {
623 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache,
624 block_offset, &refblock_data);
625 if (ret < 0) {
626 goto fail;
628 memset(refblock_data, 0, s->cluster_size);
629 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
630 refblock_data);
632 new_table[i] = block_offset;
633 block_offset += s->cluster_size;
636 /* First host offset covered by this refblock */
637 first_offset_covered = (uint64_t)i * s->refcount_block_size *
638 s->cluster_size;
639 if (first_offset_covered < end_offset) {
640 int j, end_index;
642 /* Set the refcount of all of the new refcount structures to 1 */
644 if (first_offset_covered < start_offset) {
645 assert(i == area_reftable_index);
646 j = (start_offset - first_offset_covered) / s->cluster_size;
647 assert(j < s->refcount_block_size);
648 } else {
649 j = 0;
652 end_index = MIN((end_offset - first_offset_covered) /
653 s->cluster_size,
654 s->refcount_block_size);
656 for (; j < end_index; j++) {
657 /* The caller guaranteed us this space would be empty */
658 assert(s->get_refcount(refblock_data, j) == 0);
659 s->set_refcount(refblock_data, j, 1);
662 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
663 refblock_data);
666 qcow2_cache_put(s->refcount_block_cache, &refblock_data);
669 assert(block_offset == table_offset);
671 /* Write refcount blocks to disk */
672 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
673 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
674 if (ret < 0) {
675 goto fail;
678 /* Write refcount table to disk */
679 for (i = 0; i < total_refblock_count; i++) {
680 cpu_to_be64s(&new_table[i]);
683 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
684 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
685 table_size * sizeof(uint64_t));
686 if (ret < 0) {
687 goto fail;
690 for (i = 0; i < total_refblock_count; i++) {
691 be64_to_cpus(&new_table[i]);
694 /* Hook up the new refcount table in the qcow2 header */
695 struct QEMU_PACKED {
696 uint64_t d64;
697 uint32_t d32;
698 } data;
699 data.d64 = cpu_to_be64(table_offset);
700 data.d32 = cpu_to_be32(table_clusters);
701 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
702 ret = bdrv_pwrite_sync(bs->file,
703 offsetof(QCowHeader, refcount_table_offset),
704 &data, sizeof(data));
705 if (ret < 0) {
706 goto fail;
709 /* And switch it in memory */
710 uint64_t old_table_offset = s->refcount_table_offset;
711 uint64_t old_table_size = s->refcount_table_size;
713 g_free(s->refcount_table);
714 s->refcount_table = new_table;
715 s->refcount_table_size = table_size;
716 s->refcount_table_offset = table_offset;
717 update_max_refcount_table_index(s);
719 /* Free old table. */
720 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
721 QCOW2_DISCARD_OTHER);
723 return end_offset;
725 fail:
726 g_free(new_table);
727 return ret;
730 void qcow2_process_discards(BlockDriverState *bs, int ret)
732 BDRVQcow2State *s = bs->opaque;
733 Qcow2DiscardRegion *d, *next;
735 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
736 QTAILQ_REMOVE(&s->discards, d, next);
738 /* Discard is optional, ignore the return value */
739 if (ret >= 0) {
740 int r2 = bdrv_pdiscard(bs->file, d->offset, d->bytes);
741 if (r2 < 0) {
742 trace_qcow2_process_discards_failed_region(d->offset, d->bytes,
743 r2);
747 g_free(d);
751 static void update_refcount_discard(BlockDriverState *bs,
752 uint64_t offset, uint64_t length)
754 BDRVQcow2State *s = bs->opaque;
755 Qcow2DiscardRegion *d, *p, *next;
757 QTAILQ_FOREACH(d, &s->discards, next) {
758 uint64_t new_start = MIN(offset, d->offset);
759 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
761 if (new_end - new_start <= length + d->bytes) {
762 /* There can't be any overlap, areas ending up here have no
763 * references any more and therefore shouldn't get freed another
764 * time. */
765 assert(d->bytes + length == new_end - new_start);
766 d->offset = new_start;
767 d->bytes = new_end - new_start;
768 goto found;
772 d = g_malloc(sizeof(*d));
773 *d = (Qcow2DiscardRegion) {
774 .bs = bs,
775 .offset = offset,
776 .bytes = length,
778 QTAILQ_INSERT_TAIL(&s->discards, d, next);
780 found:
781 /* Merge discard requests if they are adjacent now */
782 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
783 if (p == d
784 || p->offset > d->offset + d->bytes
785 || d->offset > p->offset + p->bytes)
787 continue;
790 /* Still no overlap possible */
791 assert(p->offset == d->offset + d->bytes
792 || d->offset == p->offset + p->bytes);
794 QTAILQ_REMOVE(&s->discards, p, next);
795 d->offset = MIN(d->offset, p->offset);
796 d->bytes += p->bytes;
797 g_free(p);
801 /* XXX: cache several refcount block clusters ? */
802 /* @addend is the absolute value of the addend; if @decrease is set, @addend
803 * will be subtracted from the current refcount, otherwise it will be added */
804 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
805 int64_t offset,
806 int64_t length,
807 uint64_t addend,
808 bool decrease,
809 enum qcow2_discard_type type)
811 BDRVQcow2State *s = bs->opaque;
812 int64_t start, last, cluster_offset;
813 void *refcount_block = NULL;
814 int64_t old_table_index = -1;
815 int ret;
817 #ifdef DEBUG_ALLOC2
818 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64
819 " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "",
820 addend);
821 #endif
822 if (length < 0) {
823 return -EINVAL;
824 } else if (length == 0) {
825 return 0;
828 if (decrease) {
829 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
830 s->l2_table_cache);
833 start = start_of_cluster(s, offset);
834 last = start_of_cluster(s, offset + length - 1);
835 for(cluster_offset = start; cluster_offset <= last;
836 cluster_offset += s->cluster_size)
838 int block_index;
839 uint64_t refcount;
840 int64_t cluster_index = cluster_offset >> s->cluster_bits;
841 int64_t table_index = cluster_index >> s->refcount_block_bits;
843 /* Load the refcount block and allocate it if needed */
844 if (table_index != old_table_index) {
845 if (refcount_block) {
846 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
848 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
849 /* If the caller needs to restart the search for free clusters,
850 * try the same ones first to see if they're still free. */
851 if (ret == -EAGAIN) {
852 if (s->free_cluster_index > (start >> s->cluster_bits)) {
853 s->free_cluster_index = (start >> s->cluster_bits);
856 if (ret < 0) {
857 goto fail;
860 old_table_index = table_index;
862 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
864 /* we can update the count and save it */
865 block_index = cluster_index & (s->refcount_block_size - 1);
867 refcount = s->get_refcount(refcount_block, block_index);
868 if (decrease ? (refcount - addend > refcount)
869 : (refcount + addend < refcount ||
870 refcount + addend > s->refcount_max))
872 ret = -EINVAL;
873 goto fail;
875 if (decrease) {
876 refcount -= addend;
877 } else {
878 refcount += addend;
880 if (refcount == 0 && cluster_index < s->free_cluster_index) {
881 s->free_cluster_index = cluster_index;
883 s->set_refcount(refcount_block, block_index, refcount);
885 if (refcount == 0) {
886 void *table;
888 table = qcow2_cache_is_table_offset(s->refcount_block_cache,
889 offset);
890 if (table != NULL) {
891 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
892 old_table_index = -1;
893 qcow2_cache_discard(s->refcount_block_cache, table);
896 table = qcow2_cache_is_table_offset(s->l2_table_cache, offset);
897 if (table != NULL) {
898 qcow2_cache_discard(s->l2_table_cache, table);
901 if (s->discard_passthrough[type]) {
902 update_refcount_discard(bs, cluster_offset, s->cluster_size);
907 ret = 0;
908 fail:
909 if (!s->cache_discards) {
910 qcow2_process_discards(bs, ret);
913 /* Write last changed block to disk */
914 if (refcount_block) {
915 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
919 * Try do undo any updates if an error is returned (This may succeed in
920 * some cases like ENOSPC for allocating a new refcount block)
922 if (ret < 0) {
923 int dummy;
924 dummy = update_refcount(bs, offset, cluster_offset - offset, addend,
925 !decrease, QCOW2_DISCARD_NEVER);
926 (void)dummy;
929 return ret;
933 * Increases or decreases the refcount of a given cluster.
935 * @addend is the absolute value of the addend; if @decrease is set, @addend
936 * will be subtracted from the current refcount, otherwise it will be added.
938 * On success 0 is returned; on failure -errno is returned.
940 int qcow2_update_cluster_refcount(BlockDriverState *bs,
941 int64_t cluster_index,
942 uint64_t addend, bool decrease,
943 enum qcow2_discard_type type)
945 BDRVQcow2State *s = bs->opaque;
946 int ret;
948 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
949 decrease, type);
950 if (ret < 0) {
951 return ret;
954 return 0;
959 /*********************************************************/
960 /* cluster allocation functions */
964 /* return < 0 if error */
965 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
966 uint64_t max)
968 BDRVQcow2State *s = bs->opaque;
969 uint64_t i, nb_clusters, refcount;
970 int ret;
972 /* We can't allocate clusters if they may still be queued for discard. */
973 if (s->cache_discards) {
974 qcow2_process_discards(bs, 0);
977 nb_clusters = size_to_clusters(s, size);
978 retry:
979 for(i = 0; i < nb_clusters; i++) {
980 uint64_t next_cluster_index = s->free_cluster_index++;
981 ret = qcow2_get_refcount(bs, next_cluster_index, &refcount);
983 if (ret < 0) {
984 return ret;
985 } else if (refcount != 0) {
986 goto retry;
990 /* Make sure that all offsets in the "allocated" range are representable
991 * in the requested max */
992 if (s->free_cluster_index > 0 &&
993 s->free_cluster_index - 1 > (max >> s->cluster_bits))
995 return -EFBIG;
998 #ifdef DEBUG_ALLOC2
999 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
1000 size,
1001 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
1002 #endif
1003 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
1006 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
1008 int64_t offset;
1009 int ret;
1011 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
1012 do {
1013 offset = alloc_clusters_noref(bs, size, QCOW_MAX_CLUSTER_OFFSET);
1014 if (offset < 0) {
1015 return offset;
1018 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1019 } while (ret == -EAGAIN);
1021 if (ret < 0) {
1022 return ret;
1025 return offset;
1028 int64_t qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
1029 int64_t nb_clusters)
1031 BDRVQcow2State *s = bs->opaque;
1032 uint64_t cluster_index, refcount;
1033 uint64_t i;
1034 int ret;
1036 assert(nb_clusters >= 0);
1037 if (nb_clusters == 0) {
1038 return 0;
1041 do {
1042 /* Check how many clusters there are free */
1043 cluster_index = offset >> s->cluster_bits;
1044 for(i = 0; i < nb_clusters; i++) {
1045 ret = qcow2_get_refcount(bs, cluster_index++, &refcount);
1046 if (ret < 0) {
1047 return ret;
1048 } else if (refcount != 0) {
1049 break;
1053 /* And then allocate them */
1054 ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false,
1055 QCOW2_DISCARD_NEVER);
1056 } while (ret == -EAGAIN);
1058 if (ret < 0) {
1059 return ret;
1062 return i;
1065 /* only used to allocate compressed sectors. We try to allocate
1066 contiguous sectors. size must be <= cluster_size */
1067 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
1069 BDRVQcow2State *s = bs->opaque;
1070 int64_t offset;
1071 size_t free_in_cluster;
1072 int ret;
1074 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
1075 assert(size > 0 && size <= s->cluster_size);
1076 assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset));
1078 offset = s->free_byte_offset;
1080 if (offset) {
1081 uint64_t refcount;
1082 ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount);
1083 if (ret < 0) {
1084 return ret;
1087 if (refcount == s->refcount_max) {
1088 offset = 0;
1092 free_in_cluster = s->cluster_size - offset_into_cluster(s, offset);
1093 do {
1094 if (!offset || free_in_cluster < size) {
1095 int64_t new_cluster;
1097 new_cluster = alloc_clusters_noref(bs, s->cluster_size,
1098 MIN(s->cluster_offset_mask,
1099 QCOW_MAX_CLUSTER_OFFSET));
1100 if (new_cluster < 0) {
1101 return new_cluster;
1104 if (new_cluster == 0) {
1105 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
1106 "allocation of compressed cluster "
1107 "at offset 0");
1108 return -EIO;
1111 if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) {
1112 offset = new_cluster;
1113 free_in_cluster = s->cluster_size;
1114 } else {
1115 free_in_cluster += s->cluster_size;
1119 assert(offset);
1120 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1121 if (ret < 0) {
1122 offset = 0;
1124 } while (ret == -EAGAIN);
1125 if (ret < 0) {
1126 return ret;
1129 /* The cluster refcount was incremented; refcount blocks must be flushed
1130 * before the caller's L2 table updates. */
1131 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
1133 s->free_byte_offset = offset + size;
1134 if (!offset_into_cluster(s, s->free_byte_offset)) {
1135 s->free_byte_offset = 0;
1138 return offset;
1141 void qcow2_free_clusters(BlockDriverState *bs,
1142 int64_t offset, int64_t size,
1143 enum qcow2_discard_type type)
1145 int ret;
1147 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
1148 ret = update_refcount(bs, offset, size, 1, true, type);
1149 if (ret < 0) {
1150 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
1151 /* TODO Remember the clusters to free them later and avoid leaking */
1156 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
1157 * normal cluster, compressed cluster, etc.)
1159 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
1160 int nb_clusters, enum qcow2_discard_type type)
1162 BDRVQcow2State *s = bs->opaque;
1163 QCow2ClusterType ctype = qcow2_get_cluster_type(bs, l2_entry);
1165 if (has_data_file(bs)) {
1166 if (s->discard_passthrough[type] &&
1167 (ctype == QCOW2_CLUSTER_NORMAL ||
1168 ctype == QCOW2_CLUSTER_ZERO_ALLOC))
1170 bdrv_pdiscard(s->data_file, l2_entry & L2E_OFFSET_MASK,
1171 nb_clusters << s->cluster_bits);
1173 return;
1176 switch (ctype) {
1177 case QCOW2_CLUSTER_COMPRESSED:
1179 int64_t offset = (l2_entry & s->cluster_offset_mask)
1180 & QCOW2_COMPRESSED_SECTOR_MASK;
1181 int size = QCOW2_COMPRESSED_SECTOR_SIZE *
1182 (((l2_entry >> s->csize_shift) & s->csize_mask) + 1);
1183 qcow2_free_clusters(bs, offset, size, type);
1185 break;
1186 case QCOW2_CLUSTER_NORMAL:
1187 case QCOW2_CLUSTER_ZERO_ALLOC:
1188 if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) {
1189 qcow2_signal_corruption(bs, false, -1, -1,
1190 "Cannot free unaligned cluster %#llx",
1191 l2_entry & L2E_OFFSET_MASK);
1192 } else {
1193 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
1194 nb_clusters << s->cluster_bits, type);
1196 break;
1197 case QCOW2_CLUSTER_ZERO_PLAIN:
1198 case QCOW2_CLUSTER_UNALLOCATED:
1199 break;
1200 default:
1201 abort();
1205 int coroutine_fn qcow2_write_caches(BlockDriverState *bs)
1207 BDRVQcow2State *s = bs->opaque;
1208 int ret;
1210 ret = qcow2_cache_write(bs, s->l2_table_cache);
1211 if (ret < 0) {
1212 return ret;
1215 if (qcow2_need_accurate_refcounts(s)) {
1216 ret = qcow2_cache_write(bs, s->refcount_block_cache);
1217 if (ret < 0) {
1218 return ret;
1222 return 0;
1225 int coroutine_fn qcow2_flush_caches(BlockDriverState *bs)
1227 int ret = qcow2_write_caches(bs);
1228 if (ret < 0) {
1229 return ret;
1232 return bdrv_flush(bs->file->bs);
1235 /*********************************************************/
1236 /* snapshots and image creation */
1240 /* update the refcounts of snapshots and the copied flag */
1241 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
1242 int64_t l1_table_offset, int l1_size, int addend)
1244 BDRVQcow2State *s = bs->opaque;
1245 uint64_t *l1_table, *l2_slice, l2_offset, entry, l1_size2, refcount;
1246 bool l1_allocated = false;
1247 int64_t old_entry, old_l2_offset;
1248 unsigned slice, slice_size2, n_slices;
1249 int i, j, l1_modified = 0, nb_csectors;
1250 int ret;
1252 assert(addend >= -1 && addend <= 1);
1254 l2_slice = NULL;
1255 l1_table = NULL;
1256 l1_size2 = l1_size * sizeof(uint64_t);
1257 slice_size2 = s->l2_slice_size * sizeof(uint64_t);
1258 n_slices = s->cluster_size / slice_size2;
1260 s->cache_discards = true;
1262 /* WARNING: qcow2_snapshot_goto relies on this function not using the
1263 * l1_table_offset when it is the current s->l1_table_offset! Be careful
1264 * when changing this! */
1265 if (l1_table_offset != s->l1_table_offset) {
1266 l1_table = g_try_malloc0(l1_size2);
1267 if (l1_size2 && l1_table == NULL) {
1268 ret = -ENOMEM;
1269 goto fail;
1271 l1_allocated = true;
1273 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1274 if (ret < 0) {
1275 goto fail;
1278 for (i = 0; i < l1_size; i++) {
1279 be64_to_cpus(&l1_table[i]);
1281 } else {
1282 assert(l1_size == s->l1_size);
1283 l1_table = s->l1_table;
1284 l1_allocated = false;
1287 for (i = 0; i < l1_size; i++) {
1288 l2_offset = l1_table[i];
1289 if (l2_offset) {
1290 old_l2_offset = l2_offset;
1291 l2_offset &= L1E_OFFSET_MASK;
1293 if (offset_into_cluster(s, l2_offset)) {
1294 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1295 PRIx64 " unaligned (L1 index: %#x)",
1296 l2_offset, i);
1297 ret = -EIO;
1298 goto fail;
1301 for (slice = 0; slice < n_slices; slice++) {
1302 ret = qcow2_cache_get(bs, s->l2_table_cache,
1303 l2_offset + slice * slice_size2,
1304 (void **) &l2_slice);
1305 if (ret < 0) {
1306 goto fail;
1309 for (j = 0; j < s->l2_slice_size; j++) {
1310 uint64_t cluster_index;
1311 uint64_t offset;
1313 entry = be64_to_cpu(l2_slice[j]);
1314 old_entry = entry;
1315 entry &= ~QCOW_OFLAG_COPIED;
1316 offset = entry & L2E_OFFSET_MASK;
1318 switch (qcow2_get_cluster_type(bs, entry)) {
1319 case QCOW2_CLUSTER_COMPRESSED:
1320 nb_csectors = ((entry >> s->csize_shift) &
1321 s->csize_mask) + 1;
1322 if (addend != 0) {
1323 uint64_t coffset = (entry & s->cluster_offset_mask)
1324 & QCOW2_COMPRESSED_SECTOR_MASK;
1325 ret = update_refcount(
1326 bs, coffset,
1327 nb_csectors * QCOW2_COMPRESSED_SECTOR_SIZE,
1328 abs(addend), addend < 0,
1329 QCOW2_DISCARD_SNAPSHOT);
1330 if (ret < 0) {
1331 goto fail;
1334 /* compressed clusters are never modified */
1335 refcount = 2;
1336 break;
1338 case QCOW2_CLUSTER_NORMAL:
1339 case QCOW2_CLUSTER_ZERO_ALLOC:
1340 if (offset_into_cluster(s, offset)) {
1341 /* Here l2_index means table (not slice) index */
1342 int l2_index = slice * s->l2_slice_size + j;
1343 qcow2_signal_corruption(
1344 bs, true, -1, -1, "Cluster "
1345 "allocation offset %#" PRIx64
1346 " unaligned (L2 offset: %#"
1347 PRIx64 ", L2 index: %#x)",
1348 offset, l2_offset, l2_index);
1349 ret = -EIO;
1350 goto fail;
1353 cluster_index = offset >> s->cluster_bits;
1354 assert(cluster_index);
1355 if (addend != 0) {
1356 ret = qcow2_update_cluster_refcount(
1357 bs, cluster_index, abs(addend), addend < 0,
1358 QCOW2_DISCARD_SNAPSHOT);
1359 if (ret < 0) {
1360 goto fail;
1364 ret = qcow2_get_refcount(bs, cluster_index, &refcount);
1365 if (ret < 0) {
1366 goto fail;
1368 break;
1370 case QCOW2_CLUSTER_ZERO_PLAIN:
1371 case QCOW2_CLUSTER_UNALLOCATED:
1372 refcount = 0;
1373 break;
1375 default:
1376 abort();
1379 if (refcount == 1) {
1380 entry |= QCOW_OFLAG_COPIED;
1382 if (entry != old_entry) {
1383 if (addend > 0) {
1384 qcow2_cache_set_dependency(bs, s->l2_table_cache,
1385 s->refcount_block_cache);
1387 l2_slice[j] = cpu_to_be64(entry);
1388 qcow2_cache_entry_mark_dirty(s->l2_table_cache,
1389 l2_slice);
1393 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1396 if (addend != 0) {
1397 ret = qcow2_update_cluster_refcount(bs, l2_offset >>
1398 s->cluster_bits,
1399 abs(addend), addend < 0,
1400 QCOW2_DISCARD_SNAPSHOT);
1401 if (ret < 0) {
1402 goto fail;
1405 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1406 &refcount);
1407 if (ret < 0) {
1408 goto fail;
1409 } else if (refcount == 1) {
1410 l2_offset |= QCOW_OFLAG_COPIED;
1412 if (l2_offset != old_l2_offset) {
1413 l1_table[i] = l2_offset;
1414 l1_modified = 1;
1419 ret = bdrv_flush(bs);
1420 fail:
1421 if (l2_slice) {
1422 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1425 s->cache_discards = false;
1426 qcow2_process_discards(bs, ret);
1428 /* Update L1 only if it isn't deleted anyway (addend = -1) */
1429 if (ret == 0 && addend >= 0 && l1_modified) {
1430 for (i = 0; i < l1_size; i++) {
1431 cpu_to_be64s(&l1_table[i]);
1434 ret = bdrv_pwrite_sync(bs->file, l1_table_offset,
1435 l1_table, l1_size2);
1437 for (i = 0; i < l1_size; i++) {
1438 be64_to_cpus(&l1_table[i]);
1441 if (l1_allocated)
1442 g_free(l1_table);
1443 return ret;
1449 /*********************************************************/
1450 /* refcount checking functions */
1453 static uint64_t refcount_array_byte_size(BDRVQcow2State *s, uint64_t entries)
1455 /* This assertion holds because there is no way we can address more than
1456 * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because
1457 * offsets have to be representable in bytes); due to every cluster
1458 * corresponding to one refcount entry, we are well below that limit */
1459 assert(entries < (UINT64_C(1) << (64 - 9)));
1461 /* Thanks to the assertion this will not overflow, because
1462 * s->refcount_order < 7.
1463 * (note: x << s->refcount_order == x * s->refcount_bits) */
1464 return DIV_ROUND_UP(entries << s->refcount_order, 8);
1468 * Reallocates *array so that it can hold new_size entries. *size must contain
1469 * the current number of entries in *array. If the reallocation fails, *array
1470 * and *size will not be modified and -errno will be returned. If the
1471 * reallocation is successful, *array will be set to the new buffer, *size
1472 * will be set to new_size and 0 will be returned. The size of the reallocated
1473 * refcount array buffer will be aligned to a cluster boundary, and the newly
1474 * allocated area will be zeroed.
1476 static int realloc_refcount_array(BDRVQcow2State *s, void **array,
1477 int64_t *size, int64_t new_size)
1479 int64_t old_byte_size, new_byte_size;
1480 void *new_ptr;
1482 /* Round to clusters so the array can be directly written to disk */
1483 old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size))
1484 * s->cluster_size;
1485 new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size))
1486 * s->cluster_size;
1488 if (new_byte_size == old_byte_size) {
1489 *size = new_size;
1490 return 0;
1493 assert(new_byte_size > 0);
1495 if (new_byte_size > SIZE_MAX) {
1496 return -ENOMEM;
1499 new_ptr = g_try_realloc(*array, new_byte_size);
1500 if (!new_ptr) {
1501 return -ENOMEM;
1504 if (new_byte_size > old_byte_size) {
1505 memset((char *)new_ptr + old_byte_size, 0,
1506 new_byte_size - old_byte_size);
1509 *array = new_ptr;
1510 *size = new_size;
1512 return 0;
1516 * Increases the refcount for a range of clusters in a given refcount table.
1517 * This is used to construct a temporary refcount table out of L1 and L2 tables
1518 * which can be compared to the refcount table saved in the image.
1520 * Modifies the number of errors in res.
1522 int qcow2_inc_refcounts_imrt(BlockDriverState *bs, BdrvCheckResult *res,
1523 void **refcount_table,
1524 int64_t *refcount_table_size,
1525 int64_t offset, int64_t size)
1527 BDRVQcow2State *s = bs->opaque;
1528 uint64_t start, last, cluster_offset, k, refcount;
1529 int64_t file_len;
1530 int ret;
1532 if (size <= 0) {
1533 return 0;
1536 file_len = bdrv_getlength(bs->file->bs);
1537 if (file_len < 0) {
1538 return file_len;
1542 * Last cluster of qcow2 image may be semi-allocated, so it may be OK to
1543 * reference some space after file end but it should be less than one
1544 * cluster.
1546 if (offset + size - file_len >= s->cluster_size) {
1547 fprintf(stderr, "ERROR: counting reference for region exceeding the "
1548 "end of the file by one cluster or more: offset 0x%" PRIx64
1549 " size 0x%" PRIx64 "\n", offset, size);
1550 res->corruptions++;
1551 return 0;
1554 start = start_of_cluster(s, offset);
1555 last = start_of_cluster(s, offset + size - 1);
1556 for(cluster_offset = start; cluster_offset <= last;
1557 cluster_offset += s->cluster_size) {
1558 k = cluster_offset >> s->cluster_bits;
1559 if (k >= *refcount_table_size) {
1560 ret = realloc_refcount_array(s, refcount_table,
1561 refcount_table_size, k + 1);
1562 if (ret < 0) {
1563 res->check_errors++;
1564 return ret;
1568 refcount = s->get_refcount(*refcount_table, k);
1569 if (refcount == s->refcount_max) {
1570 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1571 "\n", cluster_offset);
1572 fprintf(stderr, "Use qemu-img amend to increase the refcount entry "
1573 "width or qemu-img convert to create a clean copy if the "
1574 "image cannot be opened for writing\n");
1575 res->corruptions++;
1576 continue;
1578 s->set_refcount(*refcount_table, k, refcount + 1);
1581 return 0;
1584 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1585 enum {
1586 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1590 * Increases the refcount in the given refcount table for the all clusters
1591 * referenced in the L2 table. While doing so, performs some checks on L2
1592 * entries.
1594 * Returns the number of errors found by the checks or -errno if an internal
1595 * error occurred.
1597 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1598 void **refcount_table,
1599 int64_t *refcount_table_size, int64_t l2_offset,
1600 int flags, BdrvCheckMode fix, bool active)
1602 BDRVQcow2State *s = bs->opaque;
1603 uint64_t *l2_table, l2_entry;
1604 uint64_t next_contiguous_offset = 0;
1605 int i, l2_size, nb_csectors, ret;
1607 /* Read L2 table from disk */
1608 l2_size = s->l2_size * sizeof(uint64_t);
1609 l2_table = g_malloc(l2_size);
1611 ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size);
1612 if (ret < 0) {
1613 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1614 res->check_errors++;
1615 goto fail;
1618 /* Do the actual checks */
1619 for(i = 0; i < s->l2_size; i++) {
1620 l2_entry = be64_to_cpu(l2_table[i]);
1622 switch (qcow2_get_cluster_type(bs, l2_entry)) {
1623 case QCOW2_CLUSTER_COMPRESSED:
1624 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1625 if (l2_entry & QCOW_OFLAG_COPIED) {
1626 fprintf(stderr, "ERROR: coffset=0x%" PRIx64 ": "
1627 "copied flag must never be set for compressed "
1628 "clusters\n", l2_entry & s->cluster_offset_mask);
1629 l2_entry &= ~QCOW_OFLAG_COPIED;
1630 res->corruptions++;
1633 if (has_data_file(bs)) {
1634 fprintf(stderr, "ERROR compressed cluster %d with data file, "
1635 "entry=0x%" PRIx64 "\n", i, l2_entry);
1636 res->corruptions++;
1637 break;
1640 /* Mark cluster as used */
1641 nb_csectors = ((l2_entry >> s->csize_shift) &
1642 s->csize_mask) + 1;
1643 l2_entry &= s->cluster_offset_mask;
1644 ret = qcow2_inc_refcounts_imrt(
1645 bs, res, refcount_table, refcount_table_size,
1646 l2_entry & QCOW2_COMPRESSED_SECTOR_MASK,
1647 nb_csectors * QCOW2_COMPRESSED_SECTOR_SIZE);
1648 if (ret < 0) {
1649 goto fail;
1652 if (flags & CHECK_FRAG_INFO) {
1653 res->bfi.allocated_clusters++;
1654 res->bfi.compressed_clusters++;
1656 /* Compressed clusters are fragmented by nature. Since they
1657 * take up sub-sector space but we only have sector granularity
1658 * I/O we need to re-read the same sectors even for adjacent
1659 * compressed clusters.
1661 res->bfi.fragmented_clusters++;
1663 break;
1665 case QCOW2_CLUSTER_ZERO_ALLOC:
1666 case QCOW2_CLUSTER_NORMAL:
1668 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1670 /* Correct offsets are cluster aligned */
1671 if (offset_into_cluster(s, offset)) {
1672 res->corruptions++;
1674 if (qcow2_get_cluster_type(bs, l2_entry) ==
1675 QCOW2_CLUSTER_ZERO_ALLOC)
1677 fprintf(stderr, "%s offset=%" PRIx64 ": Preallocated zero "
1678 "cluster is not properly aligned; L2 entry "
1679 "corrupted.\n",
1680 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR",
1681 offset);
1682 if (fix & BDRV_FIX_ERRORS) {
1683 uint64_t l2e_offset =
1684 l2_offset + (uint64_t)i * sizeof(uint64_t);
1685 int ign = active ? QCOW2_OL_ACTIVE_L2 :
1686 QCOW2_OL_INACTIVE_L2;
1688 l2_entry = QCOW_OFLAG_ZERO;
1689 l2_table[i] = cpu_to_be64(l2_entry);
1690 ret = qcow2_pre_write_overlap_check(bs, ign,
1691 l2e_offset, sizeof(uint64_t), false);
1692 if (ret < 0) {
1693 fprintf(stderr, "ERROR: Overlap check failed\n");
1694 res->check_errors++;
1695 /* Something is seriously wrong, so abort checking
1696 * this L2 table */
1697 goto fail;
1700 ret = bdrv_pwrite_sync(bs->file, l2e_offset,
1701 &l2_table[i], sizeof(uint64_t));
1702 if (ret < 0) {
1703 fprintf(stderr, "ERROR: Failed to overwrite L2 "
1704 "table entry: %s\n", strerror(-ret));
1705 res->check_errors++;
1706 /* Do not abort, continue checking the rest of this
1707 * L2 table's entries */
1708 } else {
1709 res->corruptions--;
1710 res->corruptions_fixed++;
1711 /* Skip marking the cluster as used
1712 * (it is unused now) */
1713 continue;
1716 } else {
1717 fprintf(stderr, "ERROR offset=%" PRIx64 ": Data cluster is "
1718 "not properly aligned; L2 entry corrupted.\n", offset);
1722 if (flags & CHECK_FRAG_INFO) {
1723 res->bfi.allocated_clusters++;
1724 if (next_contiguous_offset &&
1725 offset != next_contiguous_offset) {
1726 res->bfi.fragmented_clusters++;
1728 next_contiguous_offset = offset + s->cluster_size;
1731 /* Mark cluster as used */
1732 if (!has_data_file(bs)) {
1733 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table,
1734 refcount_table_size,
1735 offset, s->cluster_size);
1736 if (ret < 0) {
1737 goto fail;
1740 break;
1743 case QCOW2_CLUSTER_ZERO_PLAIN:
1744 case QCOW2_CLUSTER_UNALLOCATED:
1745 break;
1747 default:
1748 abort();
1752 g_free(l2_table);
1753 return 0;
1755 fail:
1756 g_free(l2_table);
1757 return ret;
1761 * Increases the refcount for the L1 table, its L2 tables and all referenced
1762 * clusters in the given refcount table. While doing so, performs some checks
1763 * on L1 and L2 entries.
1765 * Returns the number of errors found by the checks or -errno if an internal
1766 * error occurred.
1768 static int check_refcounts_l1(BlockDriverState *bs,
1769 BdrvCheckResult *res,
1770 void **refcount_table,
1771 int64_t *refcount_table_size,
1772 int64_t l1_table_offset, int l1_size,
1773 int flags, BdrvCheckMode fix, bool active)
1775 BDRVQcow2State *s = bs->opaque;
1776 uint64_t *l1_table = NULL, l2_offset, l1_size2;
1777 int i, ret;
1779 l1_size2 = l1_size * sizeof(uint64_t);
1781 /* Mark L1 table as used */
1782 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, refcount_table_size,
1783 l1_table_offset, l1_size2);
1784 if (ret < 0) {
1785 goto fail;
1788 /* Read L1 table entries from disk */
1789 if (l1_size2 > 0) {
1790 l1_table = g_try_malloc(l1_size2);
1791 if (l1_table == NULL) {
1792 ret = -ENOMEM;
1793 res->check_errors++;
1794 goto fail;
1796 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1797 if (ret < 0) {
1798 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1799 res->check_errors++;
1800 goto fail;
1802 for(i = 0;i < l1_size; i++)
1803 be64_to_cpus(&l1_table[i]);
1806 /* Do the actual checks */
1807 for(i = 0; i < l1_size; i++) {
1808 l2_offset = l1_table[i];
1809 if (l2_offset) {
1810 /* Mark L2 table as used */
1811 l2_offset &= L1E_OFFSET_MASK;
1812 ret = qcow2_inc_refcounts_imrt(bs, res,
1813 refcount_table, refcount_table_size,
1814 l2_offset, s->cluster_size);
1815 if (ret < 0) {
1816 goto fail;
1819 /* L2 tables are cluster aligned */
1820 if (offset_into_cluster(s, l2_offset)) {
1821 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1822 "cluster aligned; L1 entry corrupted\n", l2_offset);
1823 res->corruptions++;
1826 /* Process and check L2 entries */
1827 ret = check_refcounts_l2(bs, res, refcount_table,
1828 refcount_table_size, l2_offset, flags,
1829 fix, active);
1830 if (ret < 0) {
1831 goto fail;
1835 g_free(l1_table);
1836 return 0;
1838 fail:
1839 g_free(l1_table);
1840 return ret;
1844 * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1846 * This function does not print an error message nor does it increment
1847 * check_errors if qcow2_get_refcount fails (this is because such an error will
1848 * have been already detected and sufficiently signaled by the calling function
1849 * (qcow2_check_refcounts) by the time this function is called).
1851 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1852 BdrvCheckMode fix)
1854 BDRVQcow2State *s = bs->opaque;
1855 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1856 int ret;
1857 uint64_t refcount;
1858 int i, j;
1859 bool repair;
1861 if (fix & BDRV_FIX_ERRORS) {
1862 /* Always repair */
1863 repair = true;
1864 } else if (fix & BDRV_FIX_LEAKS) {
1865 /* Repair only if that seems safe: This function is always
1866 * called after the refcounts have been fixed, so the refcount
1867 * is accurate if that repair was successful */
1868 repair = !res->check_errors && !res->corruptions && !res->leaks;
1869 } else {
1870 repair = false;
1873 for (i = 0; i < s->l1_size; i++) {
1874 uint64_t l1_entry = s->l1_table[i];
1875 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1876 int l2_dirty = 0;
1878 if (!l2_offset) {
1879 continue;
1882 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1883 &refcount);
1884 if (ret < 0) {
1885 /* don't print message nor increment check_errors */
1886 continue;
1888 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1889 res->corruptions++;
1890 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1891 "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1892 repair ? "Repairing" : "ERROR", i, l1_entry, refcount);
1893 if (repair) {
1894 s->l1_table[i] = refcount == 1
1895 ? l1_entry | QCOW_OFLAG_COPIED
1896 : l1_entry & ~QCOW_OFLAG_COPIED;
1897 ret = qcow2_write_l1_entry(bs, i);
1898 if (ret < 0) {
1899 res->check_errors++;
1900 goto fail;
1902 res->corruptions--;
1903 res->corruptions_fixed++;
1907 ret = bdrv_pread(bs->file, l2_offset, l2_table,
1908 s->l2_size * sizeof(uint64_t));
1909 if (ret < 0) {
1910 fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1911 strerror(-ret));
1912 res->check_errors++;
1913 goto fail;
1916 for (j = 0; j < s->l2_size; j++) {
1917 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1918 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1919 QCow2ClusterType cluster_type = qcow2_get_cluster_type(bs, l2_entry);
1921 if (cluster_type == QCOW2_CLUSTER_NORMAL ||
1922 cluster_type == QCOW2_CLUSTER_ZERO_ALLOC) {
1923 if (has_data_file(bs)) {
1924 refcount = 1;
1925 } else {
1926 ret = qcow2_get_refcount(bs,
1927 data_offset >> s->cluster_bits,
1928 &refcount);
1929 if (ret < 0) {
1930 /* don't print message nor increment check_errors */
1931 continue;
1934 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1935 res->corruptions++;
1936 fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1937 "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1938 repair ? "Repairing" : "ERROR", l2_entry, refcount);
1939 if (repair) {
1940 l2_table[j] = cpu_to_be64(refcount == 1
1941 ? l2_entry | QCOW_OFLAG_COPIED
1942 : l2_entry & ~QCOW_OFLAG_COPIED);
1943 l2_dirty++;
1949 if (l2_dirty > 0) {
1950 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1951 l2_offset, s->cluster_size,
1952 false);
1953 if (ret < 0) {
1954 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1955 "overlap check failed: %s\n", strerror(-ret));
1956 res->check_errors++;
1957 goto fail;
1960 ret = bdrv_pwrite(bs->file, l2_offset, l2_table,
1961 s->cluster_size);
1962 if (ret < 0) {
1963 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1964 strerror(-ret));
1965 res->check_errors++;
1966 goto fail;
1968 res->corruptions -= l2_dirty;
1969 res->corruptions_fixed += l2_dirty;
1973 ret = 0;
1975 fail:
1976 qemu_vfree(l2_table);
1977 return ret;
1981 * Checks consistency of refblocks and accounts for each refblock in
1982 * *refcount_table.
1984 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res,
1985 BdrvCheckMode fix, bool *rebuild,
1986 void **refcount_table, int64_t *nb_clusters)
1988 BDRVQcow2State *s = bs->opaque;
1989 int64_t i, size;
1990 int ret;
1992 for(i = 0; i < s->refcount_table_size; i++) {
1993 uint64_t offset, cluster;
1994 offset = s->refcount_table[i];
1995 cluster = offset >> s->cluster_bits;
1997 /* Refcount blocks are cluster aligned */
1998 if (offset_into_cluster(s, offset)) {
1999 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
2000 "cluster aligned; refcount table entry corrupted\n", i);
2001 res->corruptions++;
2002 *rebuild = true;
2003 continue;
2006 if (cluster >= *nb_clusters) {
2007 res->corruptions++;
2008 fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n",
2009 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i);
2011 if (fix & BDRV_FIX_ERRORS) {
2012 int64_t new_nb_clusters;
2013 Error *local_err = NULL;
2015 if (offset > INT64_MAX - s->cluster_size) {
2016 ret = -EINVAL;
2017 goto resize_fail;
2020 ret = bdrv_truncate(bs->file, offset + s->cluster_size, false,
2021 PREALLOC_MODE_OFF, &local_err);
2022 if (ret < 0) {
2023 error_report_err(local_err);
2024 goto resize_fail;
2026 size = bdrv_getlength(bs->file->bs);
2027 if (size < 0) {
2028 ret = size;
2029 goto resize_fail;
2032 new_nb_clusters = size_to_clusters(s, size);
2033 assert(new_nb_clusters >= *nb_clusters);
2035 ret = realloc_refcount_array(s, refcount_table,
2036 nb_clusters, new_nb_clusters);
2037 if (ret < 0) {
2038 res->check_errors++;
2039 return ret;
2042 if (cluster >= *nb_clusters) {
2043 ret = -EINVAL;
2044 goto resize_fail;
2047 res->corruptions--;
2048 res->corruptions_fixed++;
2049 ret = qcow2_inc_refcounts_imrt(bs, res,
2050 refcount_table, nb_clusters,
2051 offset, s->cluster_size);
2052 if (ret < 0) {
2053 return ret;
2055 /* No need to check whether the refcount is now greater than 1:
2056 * This area was just allocated and zeroed, so it can only be
2057 * exactly 1 after qcow2_inc_refcounts_imrt() */
2058 continue;
2060 resize_fail:
2061 *rebuild = true;
2062 fprintf(stderr, "ERROR could not resize image: %s\n",
2063 strerror(-ret));
2065 continue;
2068 if (offset != 0) {
2069 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2070 offset, s->cluster_size);
2071 if (ret < 0) {
2072 return ret;
2074 if (s->get_refcount(*refcount_table, cluster) != 1) {
2075 fprintf(stderr, "ERROR refcount block %" PRId64
2076 " refcount=%" PRIu64 "\n", i,
2077 s->get_refcount(*refcount_table, cluster));
2078 res->corruptions++;
2079 *rebuild = true;
2084 return 0;
2088 * Calculates an in-memory refcount table.
2090 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2091 BdrvCheckMode fix, bool *rebuild,
2092 void **refcount_table, int64_t *nb_clusters)
2094 BDRVQcow2State *s = bs->opaque;
2095 int64_t i;
2096 QCowSnapshot *sn;
2097 int ret;
2099 if (!*refcount_table) {
2100 int64_t old_size = 0;
2101 ret = realloc_refcount_array(s, refcount_table,
2102 &old_size, *nb_clusters);
2103 if (ret < 0) {
2104 res->check_errors++;
2105 return ret;
2109 /* header */
2110 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2111 0, s->cluster_size);
2112 if (ret < 0) {
2113 return ret;
2116 /* current L1 table */
2117 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2118 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO,
2119 fix, true);
2120 if (ret < 0) {
2121 return ret;
2124 /* snapshots */
2125 if (has_data_file(bs) && s->nb_snapshots) {
2126 fprintf(stderr, "ERROR %d snapshots in image with data file\n",
2127 s->nb_snapshots);
2128 res->corruptions++;
2131 for (i = 0; i < s->nb_snapshots; i++) {
2132 sn = s->snapshots + i;
2133 if (offset_into_cluster(s, sn->l1_table_offset)) {
2134 fprintf(stderr, "ERROR snapshot %s (%s) l1_offset=%#" PRIx64 ": "
2135 "L1 table is not cluster aligned; snapshot table entry "
2136 "corrupted\n", sn->id_str, sn->name, sn->l1_table_offset);
2137 res->corruptions++;
2138 continue;
2140 if (sn->l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) {
2141 fprintf(stderr, "ERROR snapshot %s (%s) l1_size=%#" PRIx32 ": "
2142 "L1 table is too large; snapshot table entry corrupted\n",
2143 sn->id_str, sn->name, sn->l1_size);
2144 res->corruptions++;
2145 continue;
2147 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2148 sn->l1_table_offset, sn->l1_size, 0, fix,
2149 false);
2150 if (ret < 0) {
2151 return ret;
2154 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2155 s->snapshots_offset, s->snapshots_size);
2156 if (ret < 0) {
2157 return ret;
2160 /* refcount data */
2161 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2162 s->refcount_table_offset,
2163 s->refcount_table_size * sizeof(uint64_t));
2164 if (ret < 0) {
2165 return ret;
2168 /* encryption */
2169 if (s->crypto_header.length) {
2170 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2171 s->crypto_header.offset,
2172 s->crypto_header.length);
2173 if (ret < 0) {
2174 return ret;
2178 /* bitmaps */
2179 ret = qcow2_check_bitmaps_refcounts(bs, res, refcount_table, nb_clusters);
2180 if (ret < 0) {
2181 return ret;
2184 return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters);
2188 * Compares the actual reference count for each cluster in the image against the
2189 * refcount as reported by the refcount structures on-disk.
2191 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2192 BdrvCheckMode fix, bool *rebuild,
2193 int64_t *highest_cluster,
2194 void *refcount_table, int64_t nb_clusters)
2196 BDRVQcow2State *s = bs->opaque;
2197 int64_t i;
2198 uint64_t refcount1, refcount2;
2199 int ret;
2201 for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) {
2202 ret = qcow2_get_refcount(bs, i, &refcount1);
2203 if (ret < 0) {
2204 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
2205 i, strerror(-ret));
2206 res->check_errors++;
2207 continue;
2210 refcount2 = s->get_refcount(refcount_table, i);
2212 if (refcount1 > 0 || refcount2 > 0) {
2213 *highest_cluster = i;
2216 if (refcount1 != refcount2) {
2217 /* Check if we're allowed to fix the mismatch */
2218 int *num_fixed = NULL;
2219 if (refcount1 == 0) {
2220 *rebuild = true;
2221 } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
2222 num_fixed = &res->leaks_fixed;
2223 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
2224 num_fixed = &res->corruptions_fixed;
2227 fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64
2228 " reference=%" PRIu64 "\n",
2229 num_fixed != NULL ? "Repairing" :
2230 refcount1 < refcount2 ? "ERROR" :
2231 "Leaked",
2232 i, refcount1, refcount2);
2234 if (num_fixed) {
2235 ret = update_refcount(bs, i << s->cluster_bits, 1,
2236 refcount_diff(refcount1, refcount2),
2237 refcount1 > refcount2,
2238 QCOW2_DISCARD_ALWAYS);
2239 if (ret >= 0) {
2240 (*num_fixed)++;
2241 continue;
2245 /* And if we couldn't, print an error */
2246 if (refcount1 < refcount2) {
2247 res->corruptions++;
2248 } else {
2249 res->leaks++;
2256 * Allocates clusters using an in-memory refcount table (IMRT) in contrast to
2257 * the on-disk refcount structures.
2259 * On input, *first_free_cluster tells where to start looking, and need not
2260 * actually be a free cluster; the returned offset will not be before that
2261 * cluster. On output, *first_free_cluster points to the first gap found, even
2262 * if that gap was too small to be used as the returned offset.
2264 * Note that *first_free_cluster is a cluster index whereas the return value is
2265 * an offset.
2267 static int64_t alloc_clusters_imrt(BlockDriverState *bs,
2268 int cluster_count,
2269 void **refcount_table,
2270 int64_t *imrt_nb_clusters,
2271 int64_t *first_free_cluster)
2273 BDRVQcow2State *s = bs->opaque;
2274 int64_t cluster = *first_free_cluster, i;
2275 bool first_gap = true;
2276 int contiguous_free_clusters;
2277 int ret;
2279 /* Starting at *first_free_cluster, find a range of at least cluster_count
2280 * continuously free clusters */
2281 for (contiguous_free_clusters = 0;
2282 cluster < *imrt_nb_clusters &&
2283 contiguous_free_clusters < cluster_count;
2284 cluster++)
2286 if (!s->get_refcount(*refcount_table, cluster)) {
2287 contiguous_free_clusters++;
2288 if (first_gap) {
2289 /* If this is the first free cluster found, update
2290 * *first_free_cluster accordingly */
2291 *first_free_cluster = cluster;
2292 first_gap = false;
2294 } else if (contiguous_free_clusters) {
2295 contiguous_free_clusters = 0;
2299 /* If contiguous_free_clusters is greater than zero, it contains the number
2300 * of continuously free clusters until the current cluster; the first free
2301 * cluster in the current "gap" is therefore
2302 * cluster - contiguous_free_clusters */
2304 /* If no such range could be found, grow the in-memory refcount table
2305 * accordingly to append free clusters at the end of the image */
2306 if (contiguous_free_clusters < cluster_count) {
2307 /* contiguous_free_clusters clusters are already empty at the image end;
2308 * we need cluster_count clusters; therefore, we have to allocate
2309 * cluster_count - contiguous_free_clusters new clusters at the end of
2310 * the image (which is the current value of cluster; note that cluster
2311 * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond
2312 * the image end) */
2313 ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters,
2314 cluster + cluster_count
2315 - contiguous_free_clusters);
2316 if (ret < 0) {
2317 return ret;
2321 /* Go back to the first free cluster */
2322 cluster -= contiguous_free_clusters;
2323 for (i = 0; i < cluster_count; i++) {
2324 s->set_refcount(*refcount_table, cluster + i, 1);
2327 return cluster << s->cluster_bits;
2331 * Creates a new refcount structure based solely on the in-memory information
2332 * given through *refcount_table. All necessary allocations will be reflected
2333 * in that array.
2335 * On success, the old refcount structure is leaked (it will be covered by the
2336 * new refcount structure).
2338 static int rebuild_refcount_structure(BlockDriverState *bs,
2339 BdrvCheckResult *res,
2340 void **refcount_table,
2341 int64_t *nb_clusters)
2343 BDRVQcow2State *s = bs->opaque;
2344 int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0;
2345 int64_t refblock_offset, refblock_start, refblock_index;
2346 uint32_t reftable_size = 0;
2347 uint64_t *on_disk_reftable = NULL;
2348 void *on_disk_refblock;
2349 int ret = 0;
2350 struct {
2351 uint64_t reftable_offset;
2352 uint32_t reftable_clusters;
2353 } QEMU_PACKED reftable_offset_and_clusters;
2355 qcow2_cache_empty(bs, s->refcount_block_cache);
2357 write_refblocks:
2358 for (; cluster < *nb_clusters; cluster++) {
2359 if (!s->get_refcount(*refcount_table, cluster)) {
2360 continue;
2363 refblock_index = cluster >> s->refcount_block_bits;
2364 refblock_start = refblock_index << s->refcount_block_bits;
2366 /* Don't allocate a cluster in a refblock already written to disk */
2367 if (first_free_cluster < refblock_start) {
2368 first_free_cluster = refblock_start;
2370 refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table,
2371 nb_clusters, &first_free_cluster);
2372 if (refblock_offset < 0) {
2373 fprintf(stderr, "ERROR allocating refblock: %s\n",
2374 strerror(-refblock_offset));
2375 res->check_errors++;
2376 ret = refblock_offset;
2377 goto fail;
2380 if (reftable_size <= refblock_index) {
2381 uint32_t old_reftable_size = reftable_size;
2382 uint64_t *new_on_disk_reftable;
2384 reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t),
2385 s->cluster_size) / sizeof(uint64_t);
2386 new_on_disk_reftable = g_try_realloc(on_disk_reftable,
2387 reftable_size *
2388 sizeof(uint64_t));
2389 if (!new_on_disk_reftable) {
2390 res->check_errors++;
2391 ret = -ENOMEM;
2392 goto fail;
2394 on_disk_reftable = new_on_disk_reftable;
2396 memset(on_disk_reftable + old_reftable_size, 0,
2397 (reftable_size - old_reftable_size) * sizeof(uint64_t));
2399 /* The offset we have for the reftable is now no longer valid;
2400 * this will leak that range, but we can easily fix that by running
2401 * a leak-fixing check after this rebuild operation */
2402 reftable_offset = -1;
2403 } else {
2404 assert(on_disk_reftable);
2406 on_disk_reftable[refblock_index] = refblock_offset;
2408 /* If this is apparently the last refblock (for now), try to squeeze the
2409 * reftable in */
2410 if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits &&
2411 reftable_offset < 0)
2413 uint64_t reftable_clusters = size_to_clusters(s, reftable_size *
2414 sizeof(uint64_t));
2415 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2416 refcount_table, nb_clusters,
2417 &first_free_cluster);
2418 if (reftable_offset < 0) {
2419 fprintf(stderr, "ERROR allocating reftable: %s\n",
2420 strerror(-reftable_offset));
2421 res->check_errors++;
2422 ret = reftable_offset;
2423 goto fail;
2427 ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset,
2428 s->cluster_size, false);
2429 if (ret < 0) {
2430 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2431 goto fail;
2434 /* The size of *refcount_table is always cluster-aligned, therefore the
2435 * write operation will not overflow */
2436 on_disk_refblock = (void *)((char *) *refcount_table +
2437 refblock_index * s->cluster_size);
2439 ret = bdrv_pwrite(bs->file, refblock_offset, on_disk_refblock,
2440 s->cluster_size);
2441 if (ret < 0) {
2442 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2443 goto fail;
2446 /* Go to the end of this refblock */
2447 cluster = refblock_start + s->refcount_block_size - 1;
2450 if (reftable_offset < 0) {
2451 uint64_t post_refblock_start, reftable_clusters;
2453 post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size);
2454 reftable_clusters = size_to_clusters(s,
2455 reftable_size * sizeof(uint64_t));
2456 /* Not pretty but simple */
2457 if (first_free_cluster < post_refblock_start) {
2458 first_free_cluster = post_refblock_start;
2460 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2461 refcount_table, nb_clusters,
2462 &first_free_cluster);
2463 if (reftable_offset < 0) {
2464 fprintf(stderr, "ERROR allocating reftable: %s\n",
2465 strerror(-reftable_offset));
2466 res->check_errors++;
2467 ret = reftable_offset;
2468 goto fail;
2471 goto write_refblocks;
2474 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2475 cpu_to_be64s(&on_disk_reftable[refblock_index]);
2478 ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset,
2479 reftable_size * sizeof(uint64_t),
2480 false);
2481 if (ret < 0) {
2482 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2483 goto fail;
2486 assert(reftable_size < INT_MAX / sizeof(uint64_t));
2487 ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable,
2488 reftable_size * sizeof(uint64_t));
2489 if (ret < 0) {
2490 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2491 goto fail;
2494 /* Enter new reftable into the image header */
2495 reftable_offset_and_clusters.reftable_offset = cpu_to_be64(reftable_offset);
2496 reftable_offset_and_clusters.reftable_clusters =
2497 cpu_to_be32(size_to_clusters(s, reftable_size * sizeof(uint64_t)));
2498 ret = bdrv_pwrite_sync(bs->file,
2499 offsetof(QCowHeader, refcount_table_offset),
2500 &reftable_offset_and_clusters,
2501 sizeof(reftable_offset_and_clusters));
2502 if (ret < 0) {
2503 fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret));
2504 goto fail;
2507 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2508 be64_to_cpus(&on_disk_reftable[refblock_index]);
2510 s->refcount_table = on_disk_reftable;
2511 s->refcount_table_offset = reftable_offset;
2512 s->refcount_table_size = reftable_size;
2513 update_max_refcount_table_index(s);
2515 return 0;
2517 fail:
2518 g_free(on_disk_reftable);
2519 return ret;
2523 * Checks an image for refcount consistency.
2525 * Returns 0 if no errors are found, the number of errors in case the image is
2526 * detected as corrupted, and -errno when an internal error occurred.
2528 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2529 BdrvCheckMode fix)
2531 BDRVQcow2State *s = bs->opaque;
2532 BdrvCheckResult pre_compare_res;
2533 int64_t size, highest_cluster, nb_clusters;
2534 void *refcount_table = NULL;
2535 bool rebuild = false;
2536 int ret;
2538 size = bdrv_getlength(bs->file->bs);
2539 if (size < 0) {
2540 res->check_errors++;
2541 return size;
2544 nb_clusters = size_to_clusters(s, size);
2545 if (nb_clusters > INT_MAX) {
2546 res->check_errors++;
2547 return -EFBIG;
2550 res->bfi.total_clusters =
2551 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
2553 ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table,
2554 &nb_clusters);
2555 if (ret < 0) {
2556 goto fail;
2559 /* In case we don't need to rebuild the refcount structure (but want to fix
2560 * something), this function is immediately called again, in which case the
2561 * result should be ignored */
2562 pre_compare_res = *res;
2563 compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table,
2564 nb_clusters);
2566 if (rebuild && (fix & BDRV_FIX_ERRORS)) {
2567 BdrvCheckResult old_res = *res;
2568 int fresh_leaks = 0;
2570 fprintf(stderr, "Rebuilding refcount structure\n");
2571 ret = rebuild_refcount_structure(bs, res, &refcount_table,
2572 &nb_clusters);
2573 if (ret < 0) {
2574 goto fail;
2577 res->corruptions = 0;
2578 res->leaks = 0;
2580 /* Because the old reftable has been exchanged for a new one the
2581 * references have to be recalculated */
2582 rebuild = false;
2583 memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters));
2584 ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table,
2585 &nb_clusters);
2586 if (ret < 0) {
2587 goto fail;
2590 if (fix & BDRV_FIX_LEAKS) {
2591 /* The old refcount structures are now leaked, fix it; the result
2592 * can be ignored, aside from leaks which were introduced by
2593 * rebuild_refcount_structure() that could not be fixed */
2594 BdrvCheckResult saved_res = *res;
2595 *res = (BdrvCheckResult){ 0 };
2597 compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild,
2598 &highest_cluster, refcount_table, nb_clusters);
2599 if (rebuild) {
2600 fprintf(stderr, "ERROR rebuilt refcount structure is still "
2601 "broken\n");
2604 /* Any leaks accounted for here were introduced by
2605 * rebuild_refcount_structure() because that function has created a
2606 * new refcount structure from scratch */
2607 fresh_leaks = res->leaks;
2608 *res = saved_res;
2611 if (res->corruptions < old_res.corruptions) {
2612 res->corruptions_fixed += old_res.corruptions - res->corruptions;
2614 if (res->leaks < old_res.leaks) {
2615 res->leaks_fixed += old_res.leaks - res->leaks;
2617 res->leaks += fresh_leaks;
2618 } else if (fix) {
2619 if (rebuild) {
2620 fprintf(stderr, "ERROR need to rebuild refcount structures\n");
2621 res->check_errors++;
2622 ret = -EIO;
2623 goto fail;
2626 if (res->leaks || res->corruptions) {
2627 *res = pre_compare_res;
2628 compare_refcounts(bs, res, fix, &rebuild, &highest_cluster,
2629 refcount_table, nb_clusters);
2633 /* check OFLAG_COPIED */
2634 ret = check_oflag_copied(bs, res, fix);
2635 if (ret < 0) {
2636 goto fail;
2639 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
2640 ret = 0;
2642 fail:
2643 g_free(refcount_table);
2645 return ret;
2648 #define overlaps_with(ofs, sz) \
2649 ranges_overlap(offset, size, ofs, sz)
2652 * Checks if the given offset into the image file is actually free to use by
2653 * looking for overlaps with important metadata sections (L1/L2 tables etc.),
2654 * i.e. a sanity check without relying on the refcount tables.
2656 * The ign parameter specifies what checks not to perform (being a bitmask of
2657 * QCow2MetadataOverlap values), i.e., what sections to ignore.
2659 * Returns:
2660 * - 0 if writing to this offset will not affect the mentioned metadata
2661 * - a positive QCow2MetadataOverlap value indicating one overlapping section
2662 * - a negative value (-errno) indicating an error while performing a check,
2663 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
2665 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
2666 int64_t size)
2668 BDRVQcow2State *s = bs->opaque;
2669 int chk = s->overlap_check & ~ign;
2670 int i, j;
2672 if (!size) {
2673 return 0;
2676 if (chk & QCOW2_OL_MAIN_HEADER) {
2677 if (offset < s->cluster_size) {
2678 return QCOW2_OL_MAIN_HEADER;
2682 /* align range to test to cluster boundaries */
2683 size = ROUND_UP(offset_into_cluster(s, offset) + size, s->cluster_size);
2684 offset = start_of_cluster(s, offset);
2686 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
2687 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
2688 return QCOW2_OL_ACTIVE_L1;
2692 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
2693 if (overlaps_with(s->refcount_table_offset,
2694 s->refcount_table_size * sizeof(uint64_t))) {
2695 return QCOW2_OL_REFCOUNT_TABLE;
2699 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
2700 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
2701 return QCOW2_OL_SNAPSHOT_TABLE;
2705 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
2706 for (i = 0; i < s->nb_snapshots; i++) {
2707 if (s->snapshots[i].l1_size &&
2708 overlaps_with(s->snapshots[i].l1_table_offset,
2709 s->snapshots[i].l1_size * sizeof(uint64_t))) {
2710 return QCOW2_OL_INACTIVE_L1;
2715 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
2716 for (i = 0; i < s->l1_size; i++) {
2717 if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
2718 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
2719 s->cluster_size)) {
2720 return QCOW2_OL_ACTIVE_L2;
2725 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
2726 unsigned last_entry = s->max_refcount_table_index;
2727 assert(last_entry < s->refcount_table_size);
2728 assert(last_entry + 1 == s->refcount_table_size ||
2729 (s->refcount_table[last_entry + 1] & REFT_OFFSET_MASK) == 0);
2730 for (i = 0; i <= last_entry; i++) {
2731 if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
2732 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
2733 s->cluster_size)) {
2734 return QCOW2_OL_REFCOUNT_BLOCK;
2739 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
2740 for (i = 0; i < s->nb_snapshots; i++) {
2741 uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
2742 uint32_t l1_sz = s->snapshots[i].l1_size;
2743 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
2744 uint64_t *l1;
2745 int ret;
2747 ret = qcow2_validate_table(bs, l1_ofs, l1_sz, sizeof(uint64_t),
2748 QCOW_MAX_L1_SIZE, "", NULL);
2749 if (ret < 0) {
2750 return ret;
2753 l1 = g_try_malloc(l1_sz2);
2755 if (l1_sz2 && l1 == NULL) {
2756 return -ENOMEM;
2759 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
2760 if (ret < 0) {
2761 g_free(l1);
2762 return ret;
2765 for (j = 0; j < l1_sz; j++) {
2766 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
2767 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
2768 g_free(l1);
2769 return QCOW2_OL_INACTIVE_L2;
2773 g_free(l1);
2777 if ((chk & QCOW2_OL_BITMAP_DIRECTORY) &&
2778 (s->autoclear_features & QCOW2_AUTOCLEAR_BITMAPS))
2780 if (overlaps_with(s->bitmap_directory_offset,
2781 s->bitmap_directory_size))
2783 return QCOW2_OL_BITMAP_DIRECTORY;
2787 return 0;
2790 static const char *metadata_ol_names[] = {
2791 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header",
2792 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table",
2793 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table",
2794 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
2795 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
2796 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
2797 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table",
2798 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table",
2799 [QCOW2_OL_BITMAP_DIRECTORY_BITNR] = "bitmap directory",
2801 QEMU_BUILD_BUG_ON(QCOW2_OL_MAX_BITNR != ARRAY_SIZE(metadata_ol_names));
2804 * First performs a check for metadata overlaps (through
2805 * qcow2_check_metadata_overlap); if that fails with a negative value (error
2806 * while performing a check), that value is returned. If an impending overlap
2807 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
2808 * and -EIO returned.
2810 * Returns 0 if there were neither overlaps nor errors while checking for
2811 * overlaps; or a negative value (-errno) on error.
2813 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
2814 int64_t size, bool data_file)
2816 int ret;
2818 if (data_file && has_data_file(bs)) {
2819 return 0;
2822 ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
2823 if (ret < 0) {
2824 return ret;
2825 } else if (ret > 0) {
2826 int metadata_ol_bitnr = ctz32(ret);
2827 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
2829 qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid "
2830 "write on metadata (overlaps with %s)",
2831 metadata_ol_names[metadata_ol_bitnr]);
2832 return -EIO;
2835 return 0;
2838 /* A pointer to a function of this type is given to walk_over_reftable(). That
2839 * function will create refblocks and pass them to a RefblockFinishOp once they
2840 * are completed (@refblock). @refblock_empty is set if the refblock is
2841 * completely empty.
2843 * Along with the refblock, a corresponding reftable entry is passed, in the
2844 * reftable @reftable (which may be reallocated) at @reftable_index.
2846 * @allocated should be set to true if a new cluster has been allocated.
2848 typedef int (RefblockFinishOp)(BlockDriverState *bs, uint64_t **reftable,
2849 uint64_t reftable_index, uint64_t *reftable_size,
2850 void *refblock, bool refblock_empty,
2851 bool *allocated, Error **errp);
2854 * This "operation" for walk_over_reftable() allocates the refblock on disk (if
2855 * it is not empty) and inserts its offset into the new reftable. The size of
2856 * this new reftable is increased as required.
2858 static int alloc_refblock(BlockDriverState *bs, uint64_t **reftable,
2859 uint64_t reftable_index, uint64_t *reftable_size,
2860 void *refblock, bool refblock_empty, bool *allocated,
2861 Error **errp)
2863 BDRVQcow2State *s = bs->opaque;
2864 int64_t offset;
2866 if (!refblock_empty && reftable_index >= *reftable_size) {
2867 uint64_t *new_reftable;
2868 uint64_t new_reftable_size;
2870 new_reftable_size = ROUND_UP(reftable_index + 1,
2871 s->cluster_size / sizeof(uint64_t));
2872 if (new_reftable_size > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
2873 error_setg(errp,
2874 "This operation would make the refcount table grow "
2875 "beyond the maximum size supported by QEMU, aborting");
2876 return -ENOTSUP;
2879 new_reftable = g_try_realloc(*reftable, new_reftable_size *
2880 sizeof(uint64_t));
2881 if (!new_reftable) {
2882 error_setg(errp, "Failed to increase reftable buffer size");
2883 return -ENOMEM;
2886 memset(new_reftable + *reftable_size, 0,
2887 (new_reftable_size - *reftable_size) * sizeof(uint64_t));
2889 *reftable = new_reftable;
2890 *reftable_size = new_reftable_size;
2893 if (!refblock_empty && !(*reftable)[reftable_index]) {
2894 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2895 if (offset < 0) {
2896 error_setg_errno(errp, -offset, "Failed to allocate refblock");
2897 return offset;
2899 (*reftable)[reftable_index] = offset;
2900 *allocated = true;
2903 return 0;
2907 * This "operation" for walk_over_reftable() writes the refblock to disk at the
2908 * offset specified by the new reftable's entry. It does not modify the new
2909 * reftable or change any refcounts.
2911 static int flush_refblock(BlockDriverState *bs, uint64_t **reftable,
2912 uint64_t reftable_index, uint64_t *reftable_size,
2913 void *refblock, bool refblock_empty, bool *allocated,
2914 Error **errp)
2916 BDRVQcow2State *s = bs->opaque;
2917 int64_t offset;
2918 int ret;
2920 if (reftable_index < *reftable_size && (*reftable)[reftable_index]) {
2921 offset = (*reftable)[reftable_index];
2923 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size,
2924 false);
2925 if (ret < 0) {
2926 error_setg_errno(errp, -ret, "Overlap check failed");
2927 return ret;
2930 ret = bdrv_pwrite(bs->file, offset, refblock, s->cluster_size);
2931 if (ret < 0) {
2932 error_setg_errno(errp, -ret, "Failed to write refblock");
2933 return ret;
2935 } else {
2936 assert(refblock_empty);
2939 return 0;
2943 * This function walks over the existing reftable and every referenced refblock;
2944 * if @new_set_refcount is non-NULL, it is called for every refcount entry to
2945 * create an equal new entry in the passed @new_refblock. Once that
2946 * @new_refblock is completely filled, @operation will be called.
2948 * @status_cb and @cb_opaque are used for the amend operation's status callback.
2949 * @index is the index of the walk_over_reftable() calls and @total is the total
2950 * number of walk_over_reftable() calls per amend operation. Both are used for
2951 * calculating the parameters for the status callback.
2953 * @allocated is set to true if a new cluster has been allocated.
2955 static int walk_over_reftable(BlockDriverState *bs, uint64_t **new_reftable,
2956 uint64_t *new_reftable_index,
2957 uint64_t *new_reftable_size,
2958 void *new_refblock, int new_refblock_size,
2959 int new_refcount_bits,
2960 RefblockFinishOp *operation, bool *allocated,
2961 Qcow2SetRefcountFunc *new_set_refcount,
2962 BlockDriverAmendStatusCB *status_cb,
2963 void *cb_opaque, int index, int total,
2964 Error **errp)
2966 BDRVQcow2State *s = bs->opaque;
2967 uint64_t reftable_index;
2968 bool new_refblock_empty = true;
2969 int refblock_index;
2970 int new_refblock_index = 0;
2971 int ret;
2973 for (reftable_index = 0; reftable_index < s->refcount_table_size;
2974 reftable_index++)
2976 uint64_t refblock_offset = s->refcount_table[reftable_index]
2977 & REFT_OFFSET_MASK;
2979 status_cb(bs, (uint64_t)index * s->refcount_table_size + reftable_index,
2980 (uint64_t)total * s->refcount_table_size, cb_opaque);
2982 if (refblock_offset) {
2983 void *refblock;
2985 if (offset_into_cluster(s, refblock_offset)) {
2986 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
2987 PRIx64 " unaligned (reftable index: %#"
2988 PRIx64 ")", refblock_offset,
2989 reftable_index);
2990 error_setg(errp,
2991 "Image is corrupt (unaligned refblock offset)");
2992 return -EIO;
2995 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offset,
2996 &refblock);
2997 if (ret < 0) {
2998 error_setg_errno(errp, -ret, "Failed to retrieve refblock");
2999 return ret;
3002 for (refblock_index = 0; refblock_index < s->refcount_block_size;
3003 refblock_index++)
3005 uint64_t refcount;
3007 if (new_refblock_index >= new_refblock_size) {
3008 /* new_refblock is now complete */
3009 ret = operation(bs, new_reftable, *new_reftable_index,
3010 new_reftable_size, new_refblock,
3011 new_refblock_empty, allocated, errp);
3012 if (ret < 0) {
3013 qcow2_cache_put(s->refcount_block_cache, &refblock);
3014 return ret;
3017 (*new_reftable_index)++;
3018 new_refblock_index = 0;
3019 new_refblock_empty = true;
3022 refcount = s->get_refcount(refblock, refblock_index);
3023 if (new_refcount_bits < 64 && refcount >> new_refcount_bits) {
3024 uint64_t offset;
3026 qcow2_cache_put(s->refcount_block_cache, &refblock);
3028 offset = ((reftable_index << s->refcount_block_bits)
3029 + refblock_index) << s->cluster_bits;
3031 error_setg(errp, "Cannot decrease refcount entry width to "
3032 "%i bits: Cluster at offset %#" PRIx64 " has a "
3033 "refcount of %" PRIu64, new_refcount_bits,
3034 offset, refcount);
3035 return -EINVAL;
3038 if (new_set_refcount) {
3039 new_set_refcount(new_refblock, new_refblock_index++,
3040 refcount);
3041 } else {
3042 new_refblock_index++;
3044 new_refblock_empty = new_refblock_empty && refcount == 0;
3047 qcow2_cache_put(s->refcount_block_cache, &refblock);
3048 } else {
3049 /* No refblock means every refcount is 0 */
3050 for (refblock_index = 0; refblock_index < s->refcount_block_size;
3051 refblock_index++)
3053 if (new_refblock_index >= new_refblock_size) {
3054 /* new_refblock is now complete */
3055 ret = operation(bs, new_reftable, *new_reftable_index,
3056 new_reftable_size, new_refblock,
3057 new_refblock_empty, allocated, errp);
3058 if (ret < 0) {
3059 return ret;
3062 (*new_reftable_index)++;
3063 new_refblock_index = 0;
3064 new_refblock_empty = true;
3067 if (new_set_refcount) {
3068 new_set_refcount(new_refblock, new_refblock_index++, 0);
3069 } else {
3070 new_refblock_index++;
3076 if (new_refblock_index > 0) {
3077 /* Complete the potentially existing partially filled final refblock */
3078 if (new_set_refcount) {
3079 for (; new_refblock_index < new_refblock_size;
3080 new_refblock_index++)
3082 new_set_refcount(new_refblock, new_refblock_index, 0);
3086 ret = operation(bs, new_reftable, *new_reftable_index,
3087 new_reftable_size, new_refblock, new_refblock_empty,
3088 allocated, errp);
3089 if (ret < 0) {
3090 return ret;
3093 (*new_reftable_index)++;
3096 status_cb(bs, (uint64_t)(index + 1) * s->refcount_table_size,
3097 (uint64_t)total * s->refcount_table_size, cb_opaque);
3099 return 0;
3102 int qcow2_change_refcount_order(BlockDriverState *bs, int refcount_order,
3103 BlockDriverAmendStatusCB *status_cb,
3104 void *cb_opaque, Error **errp)
3106 BDRVQcow2State *s = bs->opaque;
3107 Qcow2GetRefcountFunc *new_get_refcount;
3108 Qcow2SetRefcountFunc *new_set_refcount;
3109 void *new_refblock = qemu_blockalign(bs->file->bs, s->cluster_size);
3110 uint64_t *new_reftable = NULL, new_reftable_size = 0;
3111 uint64_t *old_reftable, old_reftable_size, old_reftable_offset;
3112 uint64_t new_reftable_index = 0;
3113 uint64_t i;
3114 int64_t new_reftable_offset = 0, allocated_reftable_size = 0;
3115 int new_refblock_size, new_refcount_bits = 1 << refcount_order;
3116 int old_refcount_order;
3117 int walk_index = 0;
3118 int ret;
3119 bool new_allocation;
3121 assert(s->qcow_version >= 3);
3122 assert(refcount_order >= 0 && refcount_order <= 6);
3124 /* see qcow2_open() */
3125 new_refblock_size = 1 << (s->cluster_bits - (refcount_order - 3));
3127 new_get_refcount = get_refcount_funcs[refcount_order];
3128 new_set_refcount = set_refcount_funcs[refcount_order];
3131 do {
3132 int total_walks;
3134 new_allocation = false;
3136 /* At least we have to do this walk and the one which writes the
3137 * refblocks; also, at least we have to do this loop here at least
3138 * twice (normally), first to do the allocations, and second to
3139 * determine that everything is correctly allocated, this then makes
3140 * three walks in total */
3141 total_walks = MAX(walk_index + 2, 3);
3143 /* First, allocate the structures so they are present in the refcount
3144 * structures */
3145 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3146 &new_reftable_size, NULL, new_refblock_size,
3147 new_refcount_bits, &alloc_refblock,
3148 &new_allocation, NULL, status_cb, cb_opaque,
3149 walk_index++, total_walks, errp);
3150 if (ret < 0) {
3151 goto done;
3154 new_reftable_index = 0;
3156 if (new_allocation) {
3157 if (new_reftable_offset) {
3158 qcow2_free_clusters(bs, new_reftable_offset,
3159 allocated_reftable_size * sizeof(uint64_t),
3160 QCOW2_DISCARD_NEVER);
3163 new_reftable_offset = qcow2_alloc_clusters(bs, new_reftable_size *
3164 sizeof(uint64_t));
3165 if (new_reftable_offset < 0) {
3166 error_setg_errno(errp, -new_reftable_offset,
3167 "Failed to allocate the new reftable");
3168 ret = new_reftable_offset;
3169 goto done;
3171 allocated_reftable_size = new_reftable_size;
3173 } while (new_allocation);
3175 /* Second, write the new refblocks */
3176 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3177 &new_reftable_size, new_refblock,
3178 new_refblock_size, new_refcount_bits,
3179 &flush_refblock, &new_allocation, new_set_refcount,
3180 status_cb, cb_opaque, walk_index, walk_index + 1,
3181 errp);
3182 if (ret < 0) {
3183 goto done;
3185 assert(!new_allocation);
3188 /* Write the new reftable */
3189 ret = qcow2_pre_write_overlap_check(bs, 0, new_reftable_offset,
3190 new_reftable_size * sizeof(uint64_t),
3191 false);
3192 if (ret < 0) {
3193 error_setg_errno(errp, -ret, "Overlap check failed");
3194 goto done;
3197 for (i = 0; i < new_reftable_size; i++) {
3198 cpu_to_be64s(&new_reftable[i]);
3201 ret = bdrv_pwrite(bs->file, new_reftable_offset, new_reftable,
3202 new_reftable_size * sizeof(uint64_t));
3204 for (i = 0; i < new_reftable_size; i++) {
3205 be64_to_cpus(&new_reftable[i]);
3208 if (ret < 0) {
3209 error_setg_errno(errp, -ret, "Failed to write the new reftable");
3210 goto done;
3214 /* Empty the refcount cache */
3215 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
3216 if (ret < 0) {
3217 error_setg_errno(errp, -ret, "Failed to flush the refblock cache");
3218 goto done;
3221 /* Update the image header to point to the new reftable; this only updates
3222 * the fields which are relevant to qcow2_update_header(); other fields
3223 * such as s->refcount_table or s->refcount_bits stay stale for now
3224 * (because we have to restore everything if qcow2_update_header() fails) */
3225 old_refcount_order = s->refcount_order;
3226 old_reftable_size = s->refcount_table_size;
3227 old_reftable_offset = s->refcount_table_offset;
3229 s->refcount_order = refcount_order;
3230 s->refcount_table_size = new_reftable_size;
3231 s->refcount_table_offset = new_reftable_offset;
3233 ret = qcow2_update_header(bs);
3234 if (ret < 0) {
3235 s->refcount_order = old_refcount_order;
3236 s->refcount_table_size = old_reftable_size;
3237 s->refcount_table_offset = old_reftable_offset;
3238 error_setg_errno(errp, -ret, "Failed to update the qcow2 header");
3239 goto done;
3242 /* Now update the rest of the in-memory information */
3243 old_reftable = s->refcount_table;
3244 s->refcount_table = new_reftable;
3245 update_max_refcount_table_index(s);
3247 s->refcount_bits = 1 << refcount_order;
3248 s->refcount_max = UINT64_C(1) << (s->refcount_bits - 1);
3249 s->refcount_max += s->refcount_max - 1;
3251 s->refcount_block_bits = s->cluster_bits - (refcount_order - 3);
3252 s->refcount_block_size = 1 << s->refcount_block_bits;
3254 s->get_refcount = new_get_refcount;
3255 s->set_refcount = new_set_refcount;
3257 /* For cleaning up all old refblocks and the old reftable below the "done"
3258 * label */
3259 new_reftable = old_reftable;
3260 new_reftable_size = old_reftable_size;
3261 new_reftable_offset = old_reftable_offset;
3263 done:
3264 if (new_reftable) {
3265 /* On success, new_reftable actually points to the old reftable (and
3266 * new_reftable_size is the old reftable's size); but that is just
3267 * fine */
3268 for (i = 0; i < new_reftable_size; i++) {
3269 uint64_t offset = new_reftable[i] & REFT_OFFSET_MASK;
3270 if (offset) {
3271 qcow2_free_clusters(bs, offset, s->cluster_size,
3272 QCOW2_DISCARD_OTHER);
3275 g_free(new_reftable);
3277 if (new_reftable_offset > 0) {
3278 qcow2_free_clusters(bs, new_reftable_offset,
3279 new_reftable_size * sizeof(uint64_t),
3280 QCOW2_DISCARD_OTHER);
3284 qemu_vfree(new_refblock);
3285 return ret;
3288 static int64_t get_refblock_offset(BlockDriverState *bs, uint64_t offset)
3290 BDRVQcow2State *s = bs->opaque;
3291 uint32_t index = offset_to_reftable_index(s, offset);
3292 int64_t covering_refblock_offset = 0;
3294 if (index < s->refcount_table_size) {
3295 covering_refblock_offset = s->refcount_table[index] & REFT_OFFSET_MASK;
3297 if (!covering_refblock_offset) {
3298 qcow2_signal_corruption(bs, true, -1, -1, "Refblock at %#" PRIx64 " is "
3299 "not covered by the refcount structures",
3300 offset);
3301 return -EIO;
3304 return covering_refblock_offset;
3307 static int qcow2_discard_refcount_block(BlockDriverState *bs,
3308 uint64_t discard_block_offs)
3310 BDRVQcow2State *s = bs->opaque;
3311 int64_t refblock_offs;
3312 uint64_t cluster_index = discard_block_offs >> s->cluster_bits;
3313 uint32_t block_index = cluster_index & (s->refcount_block_size - 1);
3314 void *refblock;
3315 int ret;
3317 refblock_offs = get_refblock_offset(bs, discard_block_offs);
3318 if (refblock_offs < 0) {
3319 return refblock_offs;
3322 assert(discard_block_offs != 0);
3324 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3325 &refblock);
3326 if (ret < 0) {
3327 return ret;
3330 if (s->get_refcount(refblock, block_index) != 1) {
3331 qcow2_signal_corruption(bs, true, -1, -1, "Invalid refcount:"
3332 " refblock offset %#" PRIx64
3333 ", reftable index %u"
3334 ", block offset %#" PRIx64
3335 ", refcount %#" PRIx64,
3336 refblock_offs,
3337 offset_to_reftable_index(s, discard_block_offs),
3338 discard_block_offs,
3339 s->get_refcount(refblock, block_index));
3340 qcow2_cache_put(s->refcount_block_cache, &refblock);
3341 return -EINVAL;
3343 s->set_refcount(refblock, block_index, 0);
3345 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refblock);
3347 qcow2_cache_put(s->refcount_block_cache, &refblock);
3349 if (cluster_index < s->free_cluster_index) {
3350 s->free_cluster_index = cluster_index;
3353 refblock = qcow2_cache_is_table_offset(s->refcount_block_cache,
3354 discard_block_offs);
3355 if (refblock) {
3356 /* discard refblock from the cache if refblock is cached */
3357 qcow2_cache_discard(s->refcount_block_cache, refblock);
3359 update_refcount_discard(bs, discard_block_offs, s->cluster_size);
3361 return 0;
3364 int qcow2_shrink_reftable(BlockDriverState *bs)
3366 BDRVQcow2State *s = bs->opaque;
3367 uint64_t *reftable_tmp =
3368 g_malloc(s->refcount_table_size * sizeof(uint64_t));
3369 int i, ret;
3371 for (i = 0; i < s->refcount_table_size; i++) {
3372 int64_t refblock_offs = s->refcount_table[i] & REFT_OFFSET_MASK;
3373 void *refblock;
3374 bool unused_block;
3376 if (refblock_offs == 0) {
3377 reftable_tmp[i] = 0;
3378 continue;
3380 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3381 &refblock);
3382 if (ret < 0) {
3383 goto out;
3386 /* the refblock has own reference */
3387 if (i == offset_to_reftable_index(s, refblock_offs)) {
3388 uint64_t block_index = (refblock_offs >> s->cluster_bits) &
3389 (s->refcount_block_size - 1);
3390 uint64_t refcount = s->get_refcount(refblock, block_index);
3392 s->set_refcount(refblock, block_index, 0);
3394 unused_block = buffer_is_zero(refblock, s->cluster_size);
3396 s->set_refcount(refblock, block_index, refcount);
3397 } else {
3398 unused_block = buffer_is_zero(refblock, s->cluster_size);
3400 qcow2_cache_put(s->refcount_block_cache, &refblock);
3402 reftable_tmp[i] = unused_block ? 0 : cpu_to_be64(s->refcount_table[i]);
3405 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset, reftable_tmp,
3406 s->refcount_table_size * sizeof(uint64_t));
3408 * If the write in the reftable failed the image may contain a partially
3409 * overwritten reftable. In this case it would be better to clear the
3410 * reftable in memory to avoid possible image corruption.
3412 for (i = 0; i < s->refcount_table_size; i++) {
3413 if (s->refcount_table[i] && !reftable_tmp[i]) {
3414 if (ret == 0) {
3415 ret = qcow2_discard_refcount_block(bs, s->refcount_table[i] &
3416 REFT_OFFSET_MASK);
3418 s->refcount_table[i] = 0;
3422 if (!s->cache_discards) {
3423 qcow2_process_discards(bs, ret);
3426 out:
3427 g_free(reftable_tmp);
3428 return ret;
3431 int64_t qcow2_get_last_cluster(BlockDriverState *bs, int64_t size)
3433 BDRVQcow2State *s = bs->opaque;
3434 int64_t i;
3436 for (i = size_to_clusters(s, size) - 1; i >= 0; i--) {
3437 uint64_t refcount;
3438 int ret = qcow2_get_refcount(bs, i, &refcount);
3439 if (ret < 0) {
3440 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
3441 i, strerror(-ret));
3442 return ret;
3444 if (refcount > 0) {
3445 return i;
3448 qcow2_signal_corruption(bs, true, -1, -1,
3449 "There are no references in the refcount table.");
3450 return -EIO;
3453 int qcow2_detect_metadata_preallocation(BlockDriverState *bs)
3455 BDRVQcow2State *s = bs->opaque;
3456 int64_t i, end_cluster, cluster_count = 0, threshold;
3457 int64_t file_length, real_allocation, real_clusters;
3459 qemu_co_mutex_assert_locked(&s->lock);
3461 file_length = bdrv_getlength(bs->file->bs);
3462 if (file_length < 0) {
3463 return file_length;
3466 real_allocation = bdrv_get_allocated_file_size(bs->file->bs);
3467 if (real_allocation < 0) {
3468 return real_allocation;
3471 real_clusters = real_allocation / s->cluster_size;
3472 threshold = MAX(real_clusters * 10 / 9, real_clusters + 2);
3474 end_cluster = size_to_clusters(s, file_length);
3475 for (i = 0; i < end_cluster && cluster_count < threshold; i++) {
3476 uint64_t refcount;
3477 int ret = qcow2_get_refcount(bs, i, &refcount);
3478 if (ret < 0) {
3479 return ret;
3481 cluster_count += !!refcount;
3484 return cluster_count >= threshold;