hw: virtio-pci: drop DO_UPCAST
[qemu/ar7.git] / block / qcow2-refcount.c
blob6f13d470d3659bef3db2ccaebde24b601987f7f4
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 "qemu-common.h"
28 #include "block/block_int.h"
29 #include "qcow2.h"
30 #include "qemu/range.h"
31 #include "qemu/bswap.h"
32 #include "qemu/cutils.h"
34 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
35 uint64_t max);
36 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
37 int64_t offset, int64_t length, uint64_t addend,
38 bool decrease, enum qcow2_discard_type type);
40 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index);
41 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index);
42 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index);
43 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index);
44 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index);
45 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index);
46 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index);
48 static void set_refcount_ro0(void *refcount_array, uint64_t index,
49 uint64_t value);
50 static void set_refcount_ro1(void *refcount_array, uint64_t index,
51 uint64_t value);
52 static void set_refcount_ro2(void *refcount_array, uint64_t index,
53 uint64_t value);
54 static void set_refcount_ro3(void *refcount_array, uint64_t index,
55 uint64_t value);
56 static void set_refcount_ro4(void *refcount_array, uint64_t index,
57 uint64_t value);
58 static void set_refcount_ro5(void *refcount_array, uint64_t index,
59 uint64_t value);
60 static void set_refcount_ro6(void *refcount_array, uint64_t index,
61 uint64_t value);
64 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = {
65 &get_refcount_ro0,
66 &get_refcount_ro1,
67 &get_refcount_ro2,
68 &get_refcount_ro3,
69 &get_refcount_ro4,
70 &get_refcount_ro5,
71 &get_refcount_ro6
74 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = {
75 &set_refcount_ro0,
76 &set_refcount_ro1,
77 &set_refcount_ro2,
78 &set_refcount_ro3,
79 &set_refcount_ro4,
80 &set_refcount_ro5,
81 &set_refcount_ro6
85 /*********************************************************/
86 /* refcount handling */
88 static void update_max_refcount_table_index(BDRVQcow2State *s)
90 unsigned i = s->refcount_table_size - 1;
91 while (i > 0 && (s->refcount_table[i] & REFT_OFFSET_MASK) == 0) {
92 i--;
94 /* Set s->max_refcount_table_index to the index of the last used entry */
95 s->max_refcount_table_index = i;
98 int qcow2_refcount_init(BlockDriverState *bs)
100 BDRVQcow2State *s = bs->opaque;
101 unsigned int refcount_table_size2, i;
102 int ret;
104 assert(s->refcount_order >= 0 && s->refcount_order <= 6);
106 s->get_refcount = get_refcount_funcs[s->refcount_order];
107 s->set_refcount = set_refcount_funcs[s->refcount_order];
109 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t));
110 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t);
111 s->refcount_table = g_try_malloc(refcount_table_size2);
113 if (s->refcount_table_size > 0) {
114 if (s->refcount_table == NULL) {
115 ret = -ENOMEM;
116 goto fail;
118 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD);
119 ret = bdrv_pread(bs->file, s->refcount_table_offset,
120 s->refcount_table, refcount_table_size2);
121 if (ret < 0) {
122 goto fail;
124 for(i = 0; i < s->refcount_table_size; i++)
125 be64_to_cpus(&s->refcount_table[i]);
126 update_max_refcount_table_index(s);
128 return 0;
129 fail:
130 return ret;
133 void qcow2_refcount_close(BlockDriverState *bs)
135 BDRVQcow2State *s = bs->opaque;
136 g_free(s->refcount_table);
140 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index)
142 return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1;
145 static void set_refcount_ro0(void *refcount_array, uint64_t index,
146 uint64_t value)
148 assert(!(value >> 1));
149 ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8));
150 ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8);
153 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index)
155 return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4)))
156 & 0x3;
159 static void set_refcount_ro1(void *refcount_array, uint64_t index,
160 uint64_t value)
162 assert(!(value >> 2));
163 ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4)));
164 ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4));
167 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index)
169 return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2)))
170 & 0xf;
173 static void set_refcount_ro2(void *refcount_array, uint64_t index,
174 uint64_t value)
176 assert(!(value >> 4));
177 ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2)));
178 ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2));
181 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index)
183 return ((const uint8_t *)refcount_array)[index];
186 static void set_refcount_ro3(void *refcount_array, uint64_t index,
187 uint64_t value)
189 assert(!(value >> 8));
190 ((uint8_t *)refcount_array)[index] = value;
193 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index)
195 return be16_to_cpu(((const uint16_t *)refcount_array)[index]);
198 static void set_refcount_ro4(void *refcount_array, uint64_t index,
199 uint64_t value)
201 assert(!(value >> 16));
202 ((uint16_t *)refcount_array)[index] = cpu_to_be16(value);
205 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index)
207 return be32_to_cpu(((const uint32_t *)refcount_array)[index]);
210 static void set_refcount_ro5(void *refcount_array, uint64_t index,
211 uint64_t value)
213 assert(!(value >> 32));
214 ((uint32_t *)refcount_array)[index] = cpu_to_be32(value);
217 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index)
219 return be64_to_cpu(((const uint64_t *)refcount_array)[index]);
222 static void set_refcount_ro6(void *refcount_array, uint64_t index,
223 uint64_t value)
225 ((uint64_t *)refcount_array)[index] = cpu_to_be64(value);
229 static int load_refcount_block(BlockDriverState *bs,
230 int64_t refcount_block_offset,
231 void **refcount_block)
233 BDRVQcow2State *s = bs->opaque;
235 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD);
236 return qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
237 refcount_block);
241 * Retrieves the refcount of the cluster given by its index and stores it in
242 * *refcount. Returns 0 on success and -errno on failure.
244 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index,
245 uint64_t *refcount)
247 BDRVQcow2State *s = bs->opaque;
248 uint64_t refcount_table_index, block_index;
249 int64_t refcount_block_offset;
250 int ret;
251 void *refcount_block;
253 refcount_table_index = cluster_index >> s->refcount_block_bits;
254 if (refcount_table_index >= s->refcount_table_size) {
255 *refcount = 0;
256 return 0;
258 refcount_block_offset =
259 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
260 if (!refcount_block_offset) {
261 *refcount = 0;
262 return 0;
265 if (offset_into_cluster(s, refcount_block_offset)) {
266 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64
267 " unaligned (reftable index: %#" PRIx64 ")",
268 refcount_block_offset, refcount_table_index);
269 return -EIO;
272 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset,
273 &refcount_block);
274 if (ret < 0) {
275 return ret;
278 block_index = cluster_index & (s->refcount_block_size - 1);
279 *refcount = s->get_refcount(refcount_block, block_index);
281 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
283 return 0;
286 /* Checks if two offsets are described by the same refcount block */
287 static int in_same_refcount_block(BDRVQcow2State *s, uint64_t offset_a,
288 uint64_t offset_b)
290 uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits);
291 uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits);
293 return (block_a == block_b);
297 * Loads a refcount block. If it doesn't exist yet, it is allocated first
298 * (including growing the refcount table if needed).
300 * Returns 0 on success or -errno in error case
302 static int alloc_refcount_block(BlockDriverState *bs,
303 int64_t cluster_index, void **refcount_block)
305 BDRVQcow2State *s = bs->opaque;
306 unsigned int refcount_table_index;
307 int64_t ret;
309 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC);
311 /* Find the refcount block for the given cluster */
312 refcount_table_index = cluster_index >> s->refcount_block_bits;
314 if (refcount_table_index < s->refcount_table_size) {
316 uint64_t refcount_block_offset =
317 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK;
319 /* If it's already there, we're done */
320 if (refcount_block_offset) {
321 if (offset_into_cluster(s, refcount_block_offset)) {
322 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
323 PRIx64 " unaligned (reftable index: "
324 "%#x)", refcount_block_offset,
325 refcount_table_index);
326 return -EIO;
329 return load_refcount_block(bs, refcount_block_offset,
330 refcount_block);
335 * If we came here, we need to allocate something. Something is at least
336 * a cluster for the new refcount block. It may also include a new refcount
337 * table if the old refcount table is too small.
339 * Note that allocating clusters here needs some special care:
341 * - We can't use the normal qcow2_alloc_clusters(), it would try to
342 * increase the refcount and very likely we would end up with an endless
343 * recursion. Instead we must place the refcount blocks in a way that
344 * they can describe them themselves.
346 * - We need to consider that at this point we are inside update_refcounts
347 * and potentially doing an initial refcount increase. This means that
348 * some clusters have already been allocated by the caller, but their
349 * refcount isn't accurate yet. If we allocate clusters for metadata, we
350 * need to return -EAGAIN to signal the caller that it needs to restart
351 * the search for free clusters.
353 * - alloc_clusters_noref and qcow2_free_clusters may load a different
354 * refcount block into the cache
357 *refcount_block = NULL;
359 /* We write to the refcount table, so we might depend on L2 tables */
360 ret = qcow2_cache_flush(bs, s->l2_table_cache);
361 if (ret < 0) {
362 return ret;
365 /* Allocate the refcount block itself and mark it as used */
366 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size, INT64_MAX);
367 if (new_block < 0) {
368 return new_block;
371 /* The offset must fit in the offset field of the refcount table entry */
372 assert((new_block & REFT_OFFSET_MASK) == new_block);
374 /* If we're allocating the block at offset 0 then something is wrong */
375 if (new_block == 0) {
376 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
377 "allocation of refcount block at offset 0");
378 return -EIO;
381 #ifdef DEBUG_ALLOC2
382 fprintf(stderr, "qcow2: Allocate refcount block %d for %" PRIx64
383 " at %" PRIx64 "\n",
384 refcount_table_index, cluster_index << s->cluster_bits, new_block);
385 #endif
387 if (in_same_refcount_block(s, new_block, cluster_index << s->cluster_bits)) {
388 /* Zero the new refcount block before updating it */
389 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
390 refcount_block);
391 if (ret < 0) {
392 goto fail;
395 memset(*refcount_block, 0, s->cluster_size);
397 /* The block describes itself, need to update the cache */
398 int block_index = (new_block >> s->cluster_bits) &
399 (s->refcount_block_size - 1);
400 s->set_refcount(*refcount_block, block_index, 1);
401 } else {
402 /* Described somewhere else. This can recurse at most twice before we
403 * arrive at a block that describes itself. */
404 ret = update_refcount(bs, new_block, s->cluster_size, 1, false,
405 QCOW2_DISCARD_NEVER);
406 if (ret < 0) {
407 goto fail;
410 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
411 if (ret < 0) {
412 goto fail;
415 /* Initialize the new refcount block only after updating its refcount,
416 * update_refcount uses the refcount cache itself */
417 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, new_block,
418 refcount_block);
419 if (ret < 0) {
420 goto fail;
423 memset(*refcount_block, 0, s->cluster_size);
426 /* Now the new refcount block needs to be written to disk */
427 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE);
428 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, *refcount_block);
429 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
430 if (ret < 0) {
431 goto fail;
434 /* If the refcount table is big enough, just hook the block up there */
435 if (refcount_table_index < s->refcount_table_size) {
436 uint64_t data64 = cpu_to_be64(new_block);
437 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_HOOKUP);
438 ret = bdrv_pwrite_sync(bs->file,
439 s->refcount_table_offset + refcount_table_index * sizeof(uint64_t),
440 &data64, sizeof(data64));
441 if (ret < 0) {
442 goto fail;
445 s->refcount_table[refcount_table_index] = new_block;
446 /* If there's a hole in s->refcount_table then it can happen
447 * that refcount_table_index < s->max_refcount_table_index */
448 s->max_refcount_table_index =
449 MAX(s->max_refcount_table_index, refcount_table_index);
451 /* The new refcount block may be where the caller intended to put its
452 * data, so let it restart the search. */
453 return -EAGAIN;
456 qcow2_cache_put(s->refcount_block_cache, refcount_block);
459 * If we come here, we need to grow the refcount table. Again, a new
460 * refcount table needs some space and we can't simply allocate to avoid
461 * endless recursion.
463 * Therefore let's grab new refcount blocks at the end of the image, which
464 * will describe themselves and the new refcount table. This way we can
465 * reference them only in the new table and do the switch to the new
466 * refcount table at once without producing an inconsistent state in
467 * between.
469 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW);
471 /* Calculate the number of refcount blocks needed so far; this will be the
472 * basis for calculating the index of the first cluster used for the
473 * self-describing refcount structures which we are about to create.
475 * Because we reached this point, there cannot be any refcount entries for
476 * cluster_index or higher indices yet. However, because new_block has been
477 * allocated to describe that cluster (and it will assume this role later
478 * on), we cannot use that index; also, new_block may actually have a higher
479 * cluster index than cluster_index, so it needs to be taken into account
480 * here (and 1 needs to be added to its value because that cluster is used).
482 uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1,
483 (new_block >> s->cluster_bits) + 1),
484 s->refcount_block_size);
486 /* Create the new refcount table and blocks */
487 uint64_t meta_offset = (blocks_used * s->refcount_block_size) *
488 s->cluster_size;
490 ret = qcow2_refcount_area(bs, meta_offset, 0, false,
491 refcount_table_index, new_block);
492 if (ret < 0) {
493 return ret;
496 ret = load_refcount_block(bs, new_block, refcount_block);
497 if (ret < 0) {
498 return ret;
501 /* If we were trying to do the initial refcount update for some cluster
502 * allocation, we might have used the same clusters to store newly
503 * allocated metadata. Make the caller search some new space. */
504 return -EAGAIN;
506 fail:
507 if (*refcount_block != NULL) {
508 qcow2_cache_put(s->refcount_block_cache, refcount_block);
510 return ret;
514 * Starting at @start_offset, this function creates new self-covering refcount
515 * structures: A new refcount table and refcount blocks which cover all of
516 * themselves, and a number of @additional_clusters beyond their end.
517 * @start_offset must be at the end of the image file, that is, there must be
518 * only empty space beyond it.
519 * If @exact_size is false, the refcount table will have 50 % more entries than
520 * necessary so it will not need to grow again soon.
521 * If @new_refblock_offset is not zero, it contains the offset of a refcount
522 * block that should be entered into the new refcount table at index
523 * @new_refblock_index.
525 * Returns: The offset after the new refcount structures (i.e. where the
526 * @additional_clusters may be placed) on success, -errno on error.
528 int64_t qcow2_refcount_area(BlockDriverState *bs, uint64_t start_offset,
529 uint64_t additional_clusters, bool exact_size,
530 int new_refblock_index,
531 uint64_t new_refblock_offset)
533 BDRVQcow2State *s = bs->opaque;
534 uint64_t total_refblock_count_u64, additional_refblock_count;
535 int total_refblock_count, table_size, area_reftable_index, table_clusters;
536 int i;
537 uint64_t table_offset, block_offset, end_offset;
538 int ret;
539 uint64_t *new_table;
541 assert(!(start_offset % s->cluster_size));
543 qcow2_refcount_metadata_size(start_offset / s->cluster_size +
544 additional_clusters,
545 s->cluster_size, s->refcount_order,
546 !exact_size, &total_refblock_count_u64);
547 if (total_refblock_count_u64 > QCOW_MAX_REFTABLE_SIZE) {
548 return -EFBIG;
550 total_refblock_count = total_refblock_count_u64;
552 /* Index in the refcount table of the first refcount block to cover the area
553 * of refcount structures we are about to create; we know that
554 * @total_refblock_count can cover @start_offset, so this will definitely
555 * fit into an int. */
556 area_reftable_index = (start_offset / s->cluster_size) /
557 s->refcount_block_size;
559 if (exact_size) {
560 table_size = total_refblock_count;
561 } else {
562 table_size = total_refblock_count +
563 DIV_ROUND_UP(total_refblock_count, 2);
565 /* The qcow2 file can only store the reftable size in number of clusters */
566 table_size = ROUND_UP(table_size, s->cluster_size / sizeof(uint64_t));
567 table_clusters = (table_size * sizeof(uint64_t)) / s->cluster_size;
569 if (table_size > QCOW_MAX_REFTABLE_SIZE) {
570 return -EFBIG;
573 new_table = g_try_new0(uint64_t, table_size);
575 assert(table_size > 0);
576 if (new_table == NULL) {
577 ret = -ENOMEM;
578 goto fail;
581 /* Fill the new refcount table */
582 if (table_size > s->max_refcount_table_index) {
583 /* We're actually growing the reftable */
584 memcpy(new_table, s->refcount_table,
585 (s->max_refcount_table_index + 1) * sizeof(uint64_t));
586 } else {
587 /* Improbable case: We're shrinking the reftable. However, the caller
588 * has assured us that there is only empty space beyond @start_offset,
589 * so we can simply drop all of the refblocks that won't fit into the
590 * new reftable. */
591 memcpy(new_table, s->refcount_table, table_size * sizeof(uint64_t));
594 if (new_refblock_offset) {
595 assert(new_refblock_index < total_refblock_count);
596 new_table[new_refblock_index] = new_refblock_offset;
599 /* Count how many new refblocks we have to create */
600 additional_refblock_count = 0;
601 for (i = area_reftable_index; i < total_refblock_count; i++) {
602 if (!new_table[i]) {
603 additional_refblock_count++;
607 table_offset = start_offset + additional_refblock_count * s->cluster_size;
608 end_offset = table_offset + table_clusters * s->cluster_size;
610 /* Fill the refcount blocks, and create new ones, if necessary */
611 block_offset = start_offset;
612 for (i = area_reftable_index; i < total_refblock_count; i++) {
613 void *refblock_data;
614 uint64_t first_offset_covered;
616 /* Reuse an existing refblock if possible, create a new one otherwise */
617 if (new_table[i]) {
618 ret = qcow2_cache_get(bs, s->refcount_block_cache, new_table[i],
619 &refblock_data);
620 if (ret < 0) {
621 goto fail;
623 } else {
624 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache,
625 block_offset, &refblock_data);
626 if (ret < 0) {
627 goto fail;
629 memset(refblock_data, 0, s->cluster_size);
630 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
631 refblock_data);
633 new_table[i] = block_offset;
634 block_offset += s->cluster_size;
637 /* First host offset covered by this refblock */
638 first_offset_covered = (uint64_t)i * s->refcount_block_size *
639 s->cluster_size;
640 if (first_offset_covered < end_offset) {
641 int j, end_index;
643 /* Set the refcount of all of the new refcount structures to 1 */
645 if (first_offset_covered < start_offset) {
646 assert(i == area_reftable_index);
647 j = (start_offset - first_offset_covered) / s->cluster_size;
648 assert(j < s->refcount_block_size);
649 } else {
650 j = 0;
653 end_index = MIN((end_offset - first_offset_covered) /
654 s->cluster_size,
655 s->refcount_block_size);
657 for (; j < end_index; j++) {
658 /* The caller guaranteed us this space would be empty */
659 assert(s->get_refcount(refblock_data, j) == 0);
660 s->set_refcount(refblock_data, j, 1);
663 qcow2_cache_entry_mark_dirty(s->refcount_block_cache,
664 refblock_data);
667 qcow2_cache_put(s->refcount_block_cache, &refblock_data);
670 assert(block_offset == table_offset);
672 /* Write refcount blocks to disk */
673 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS);
674 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
675 if (ret < 0) {
676 goto fail;
679 /* Write refcount table to disk */
680 for (i = 0; i < total_refblock_count; i++) {
681 cpu_to_be64s(&new_table[i]);
684 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE);
685 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table,
686 table_size * sizeof(uint64_t));
687 if (ret < 0) {
688 goto fail;
691 for (i = 0; i < total_refblock_count; i++) {
692 be64_to_cpus(&new_table[i]);
695 /* Hook up the new refcount table in the qcow2 header */
696 struct QEMU_PACKED {
697 uint64_t d64;
698 uint32_t d32;
699 } data;
700 data.d64 = cpu_to_be64(table_offset);
701 data.d32 = cpu_to_be32(table_clusters);
702 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE);
703 ret = bdrv_pwrite_sync(bs->file,
704 offsetof(QCowHeader, refcount_table_offset),
705 &data, sizeof(data));
706 if (ret < 0) {
707 goto fail;
710 /* And switch it in memory */
711 uint64_t old_table_offset = s->refcount_table_offset;
712 uint64_t old_table_size = s->refcount_table_size;
714 g_free(s->refcount_table);
715 s->refcount_table = new_table;
716 s->refcount_table_size = table_size;
717 s->refcount_table_offset = table_offset;
718 update_max_refcount_table_index(s);
720 /* Free old table. */
721 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t),
722 QCOW2_DISCARD_OTHER);
724 return end_offset;
726 fail:
727 g_free(new_table);
728 return ret;
731 void qcow2_process_discards(BlockDriverState *bs, int ret)
733 BDRVQcow2State *s = bs->opaque;
734 Qcow2DiscardRegion *d, *next;
736 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) {
737 QTAILQ_REMOVE(&s->discards, d, next);
739 /* Discard is optional, ignore the return value */
740 if (ret >= 0) {
741 bdrv_pdiscard(bs->file, d->offset, d->bytes);
744 g_free(d);
748 static void update_refcount_discard(BlockDriverState *bs,
749 uint64_t offset, uint64_t length)
751 BDRVQcow2State *s = bs->opaque;
752 Qcow2DiscardRegion *d, *p, *next;
754 QTAILQ_FOREACH(d, &s->discards, next) {
755 uint64_t new_start = MIN(offset, d->offset);
756 uint64_t new_end = MAX(offset + length, d->offset + d->bytes);
758 if (new_end - new_start <= length + d->bytes) {
759 /* There can't be any overlap, areas ending up here have no
760 * references any more and therefore shouldn't get freed another
761 * time. */
762 assert(d->bytes + length == new_end - new_start);
763 d->offset = new_start;
764 d->bytes = new_end - new_start;
765 goto found;
769 d = g_malloc(sizeof(*d));
770 *d = (Qcow2DiscardRegion) {
771 .bs = bs,
772 .offset = offset,
773 .bytes = length,
775 QTAILQ_INSERT_TAIL(&s->discards, d, next);
777 found:
778 /* Merge discard requests if they are adjacent now */
779 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) {
780 if (p == d
781 || p->offset > d->offset + d->bytes
782 || d->offset > p->offset + p->bytes)
784 continue;
787 /* Still no overlap possible */
788 assert(p->offset == d->offset + d->bytes
789 || d->offset == p->offset + p->bytes);
791 QTAILQ_REMOVE(&s->discards, p, next);
792 d->offset = MIN(d->offset, p->offset);
793 d->bytes += p->bytes;
794 g_free(p);
798 /* XXX: cache several refcount block clusters ? */
799 /* @addend is the absolute value of the addend; if @decrease is set, @addend
800 * will be subtracted from the current refcount, otherwise it will be added */
801 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs,
802 int64_t offset,
803 int64_t length,
804 uint64_t addend,
805 bool decrease,
806 enum qcow2_discard_type type)
808 BDRVQcow2State *s = bs->opaque;
809 int64_t start, last, cluster_offset;
810 void *refcount_block = NULL;
811 int64_t old_table_index = -1;
812 int ret;
814 #ifdef DEBUG_ALLOC2
815 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64
816 " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "",
817 addend);
818 #endif
819 if (length < 0) {
820 return -EINVAL;
821 } else if (length == 0) {
822 return 0;
825 if (decrease) {
826 qcow2_cache_set_dependency(bs, s->refcount_block_cache,
827 s->l2_table_cache);
830 start = start_of_cluster(s, offset);
831 last = start_of_cluster(s, offset + length - 1);
832 for(cluster_offset = start; cluster_offset <= last;
833 cluster_offset += s->cluster_size)
835 int block_index;
836 uint64_t refcount;
837 int64_t cluster_index = cluster_offset >> s->cluster_bits;
838 int64_t table_index = cluster_index >> s->refcount_block_bits;
840 /* Load the refcount block and allocate it if needed */
841 if (table_index != old_table_index) {
842 if (refcount_block) {
843 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
845 ret = alloc_refcount_block(bs, cluster_index, &refcount_block);
846 /* If the caller needs to restart the search for free clusters,
847 * try the same ones first to see if they're still free. */
848 if (ret == -EAGAIN) {
849 if (s->free_cluster_index > (start >> s->cluster_bits)) {
850 s->free_cluster_index = (start >> s->cluster_bits);
853 if (ret < 0) {
854 goto fail;
857 old_table_index = table_index;
859 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refcount_block);
861 /* we can update the count and save it */
862 block_index = cluster_index & (s->refcount_block_size - 1);
864 refcount = s->get_refcount(refcount_block, block_index);
865 if (decrease ? (refcount - addend > refcount)
866 : (refcount + addend < refcount ||
867 refcount + addend > s->refcount_max))
869 ret = -EINVAL;
870 goto fail;
872 if (decrease) {
873 refcount -= addend;
874 } else {
875 refcount += addend;
877 if (refcount == 0 && cluster_index < s->free_cluster_index) {
878 s->free_cluster_index = cluster_index;
880 s->set_refcount(refcount_block, block_index, refcount);
882 if (refcount == 0) {
883 void *table;
885 table = qcow2_cache_is_table_offset(s->refcount_block_cache,
886 offset);
887 if (table != NULL) {
888 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
889 qcow2_cache_discard(s->refcount_block_cache, table);
892 table = qcow2_cache_is_table_offset(s->l2_table_cache, offset);
893 if (table != NULL) {
894 qcow2_cache_discard(s->l2_table_cache, table);
897 if (s->discard_passthrough[type]) {
898 update_refcount_discard(bs, cluster_offset, s->cluster_size);
903 ret = 0;
904 fail:
905 if (!s->cache_discards) {
906 qcow2_process_discards(bs, ret);
909 /* Write last changed block to disk */
910 if (refcount_block) {
911 qcow2_cache_put(s->refcount_block_cache, &refcount_block);
915 * Try do undo any updates if an error is returned (This may succeed in
916 * some cases like ENOSPC for allocating a new refcount block)
918 if (ret < 0) {
919 int dummy;
920 dummy = update_refcount(bs, offset, cluster_offset - offset, addend,
921 !decrease, QCOW2_DISCARD_NEVER);
922 (void)dummy;
925 return ret;
929 * Increases or decreases the refcount of a given cluster.
931 * @addend is the absolute value of the addend; if @decrease is set, @addend
932 * will be subtracted from the current refcount, otherwise it will be added.
934 * On success 0 is returned; on failure -errno is returned.
936 int qcow2_update_cluster_refcount(BlockDriverState *bs,
937 int64_t cluster_index,
938 uint64_t addend, bool decrease,
939 enum qcow2_discard_type type)
941 BDRVQcow2State *s = bs->opaque;
942 int ret;
944 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend,
945 decrease, type);
946 if (ret < 0) {
947 return ret;
950 return 0;
955 /*********************************************************/
956 /* cluster allocation functions */
960 /* return < 0 if error */
961 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size,
962 uint64_t max)
964 BDRVQcow2State *s = bs->opaque;
965 uint64_t i, nb_clusters, refcount;
966 int ret;
968 /* We can't allocate clusters if they may still be queued for discard. */
969 if (s->cache_discards) {
970 qcow2_process_discards(bs, 0);
973 nb_clusters = size_to_clusters(s, size);
974 retry:
975 for(i = 0; i < nb_clusters; i++) {
976 uint64_t next_cluster_index = s->free_cluster_index++;
977 ret = qcow2_get_refcount(bs, next_cluster_index, &refcount);
979 if (ret < 0) {
980 return ret;
981 } else if (refcount != 0) {
982 goto retry;
986 /* Make sure that all offsets in the "allocated" range are representable
987 * in the requested max */
988 if (s->free_cluster_index > 0 &&
989 s->free_cluster_index - 1 > (max >> s->cluster_bits))
991 return -EFBIG;
994 #ifdef DEBUG_ALLOC2
995 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n",
996 size,
997 (s->free_cluster_index - nb_clusters) << s->cluster_bits);
998 #endif
999 return (s->free_cluster_index - nb_clusters) << s->cluster_bits;
1002 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size)
1004 int64_t offset;
1005 int ret;
1007 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC);
1008 do {
1009 offset = alloc_clusters_noref(bs, size, QCOW_MAX_CLUSTER_OFFSET);
1010 if (offset < 0) {
1011 return offset;
1014 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1015 } while (ret == -EAGAIN);
1017 if (ret < 0) {
1018 return ret;
1021 return offset;
1024 int64_t qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset,
1025 int64_t nb_clusters)
1027 BDRVQcow2State *s = bs->opaque;
1028 uint64_t cluster_index, refcount;
1029 uint64_t i;
1030 int ret;
1032 assert(nb_clusters >= 0);
1033 if (nb_clusters == 0) {
1034 return 0;
1037 do {
1038 /* Check how many clusters there are free */
1039 cluster_index = offset >> s->cluster_bits;
1040 for(i = 0; i < nb_clusters; i++) {
1041 ret = qcow2_get_refcount(bs, cluster_index++, &refcount);
1042 if (ret < 0) {
1043 return ret;
1044 } else if (refcount != 0) {
1045 break;
1049 /* And then allocate them */
1050 ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false,
1051 QCOW2_DISCARD_NEVER);
1052 } while (ret == -EAGAIN);
1054 if (ret < 0) {
1055 return ret;
1058 return i;
1061 /* only used to allocate compressed sectors. We try to allocate
1062 contiguous sectors. size must be <= cluster_size */
1063 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size)
1065 BDRVQcow2State *s = bs->opaque;
1066 int64_t offset;
1067 size_t free_in_cluster;
1068 int ret;
1070 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES);
1071 assert(size > 0 && size <= s->cluster_size);
1072 assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset));
1074 offset = s->free_byte_offset;
1076 if (offset) {
1077 uint64_t refcount;
1078 ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount);
1079 if (ret < 0) {
1080 return ret;
1083 if (refcount == s->refcount_max) {
1084 offset = 0;
1088 free_in_cluster = s->cluster_size - offset_into_cluster(s, offset);
1089 do {
1090 if (!offset || free_in_cluster < size) {
1091 int64_t new_cluster;
1093 new_cluster = alloc_clusters_noref(bs, s->cluster_size,
1094 MIN(s->cluster_offset_mask,
1095 QCOW_MAX_CLUSTER_OFFSET));
1096 if (new_cluster < 0) {
1097 return new_cluster;
1100 if (new_cluster == 0) {
1101 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
1102 "allocation of compressed cluster "
1103 "at offset 0");
1104 return -EIO;
1107 if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) {
1108 offset = new_cluster;
1109 free_in_cluster = s->cluster_size;
1110 } else {
1111 free_in_cluster += s->cluster_size;
1115 assert(offset);
1116 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER);
1117 if (ret < 0) {
1118 offset = 0;
1120 } while (ret == -EAGAIN);
1121 if (ret < 0) {
1122 return ret;
1125 /* The cluster refcount was incremented; refcount blocks must be flushed
1126 * before the caller's L2 table updates. */
1127 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache);
1129 s->free_byte_offset = offset + size;
1130 if (!offset_into_cluster(s, s->free_byte_offset)) {
1131 s->free_byte_offset = 0;
1134 return offset;
1137 void qcow2_free_clusters(BlockDriverState *bs,
1138 int64_t offset, int64_t size,
1139 enum qcow2_discard_type type)
1141 int ret;
1143 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE);
1144 ret = update_refcount(bs, offset, size, 1, true, type);
1145 if (ret < 0) {
1146 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret));
1147 /* TODO Remember the clusters to free them later and avoid leaking */
1152 * Free a cluster using its L2 entry (handles clusters of all types, e.g.
1153 * normal cluster, compressed cluster, etc.)
1155 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry,
1156 int nb_clusters, enum qcow2_discard_type type)
1158 BDRVQcow2State *s = bs->opaque;
1160 switch (qcow2_get_cluster_type(l2_entry)) {
1161 case QCOW2_CLUSTER_COMPRESSED:
1163 int nb_csectors;
1164 nb_csectors = ((l2_entry >> s->csize_shift) &
1165 s->csize_mask) + 1;
1166 qcow2_free_clusters(bs,
1167 (l2_entry & s->cluster_offset_mask) & ~511,
1168 nb_csectors * 512, type);
1170 break;
1171 case QCOW2_CLUSTER_NORMAL:
1172 case QCOW2_CLUSTER_ZERO_ALLOC:
1173 if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) {
1174 qcow2_signal_corruption(bs, false, -1, -1,
1175 "Cannot free unaligned cluster %#llx",
1176 l2_entry & L2E_OFFSET_MASK);
1177 } else {
1178 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK,
1179 nb_clusters << s->cluster_bits, type);
1181 break;
1182 case QCOW2_CLUSTER_ZERO_PLAIN:
1183 case QCOW2_CLUSTER_UNALLOCATED:
1184 break;
1185 default:
1186 abort();
1190 int coroutine_fn qcow2_write_caches(BlockDriverState *bs)
1192 BDRVQcow2State *s = bs->opaque;
1193 int ret;
1195 ret = qcow2_cache_write(bs, s->l2_table_cache);
1196 if (ret < 0) {
1197 return ret;
1200 if (qcow2_need_accurate_refcounts(s)) {
1201 ret = qcow2_cache_write(bs, s->refcount_block_cache);
1202 if (ret < 0) {
1203 return ret;
1207 return 0;
1210 int coroutine_fn qcow2_flush_caches(BlockDriverState *bs)
1212 int ret = qcow2_write_caches(bs);
1213 if (ret < 0) {
1214 return ret;
1217 return bdrv_flush(bs->file->bs);
1220 /*********************************************************/
1221 /* snapshots and image creation */
1225 /* update the refcounts of snapshots and the copied flag */
1226 int qcow2_update_snapshot_refcount(BlockDriverState *bs,
1227 int64_t l1_table_offset, int l1_size, int addend)
1229 BDRVQcow2State *s = bs->opaque;
1230 uint64_t *l1_table, *l2_slice, l2_offset, entry, l1_size2, refcount;
1231 bool l1_allocated = false;
1232 int64_t old_entry, old_l2_offset;
1233 unsigned slice, slice_size2, n_slices;
1234 int i, j, l1_modified = 0, nb_csectors;
1235 int ret;
1237 assert(addend >= -1 && addend <= 1);
1239 l2_slice = NULL;
1240 l1_table = NULL;
1241 l1_size2 = l1_size * sizeof(uint64_t);
1242 slice_size2 = s->l2_slice_size * sizeof(uint64_t);
1243 n_slices = s->cluster_size / slice_size2;
1245 s->cache_discards = true;
1247 /* WARNING: qcow2_snapshot_goto relies on this function not using the
1248 * l1_table_offset when it is the current s->l1_table_offset! Be careful
1249 * when changing this! */
1250 if (l1_table_offset != s->l1_table_offset) {
1251 l1_table = g_try_malloc0(ROUND_UP(l1_size2, 512));
1252 if (l1_size2 && l1_table == NULL) {
1253 ret = -ENOMEM;
1254 goto fail;
1256 l1_allocated = true;
1258 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1259 if (ret < 0) {
1260 goto fail;
1263 for (i = 0; i < l1_size; i++) {
1264 be64_to_cpus(&l1_table[i]);
1266 } else {
1267 assert(l1_size == s->l1_size);
1268 l1_table = s->l1_table;
1269 l1_allocated = false;
1272 for (i = 0; i < l1_size; i++) {
1273 l2_offset = l1_table[i];
1274 if (l2_offset) {
1275 old_l2_offset = l2_offset;
1276 l2_offset &= L1E_OFFSET_MASK;
1278 if (offset_into_cluster(s, l2_offset)) {
1279 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1280 PRIx64 " unaligned (L1 index: %#x)",
1281 l2_offset, i);
1282 ret = -EIO;
1283 goto fail;
1286 for (slice = 0; slice < n_slices; slice++) {
1287 ret = qcow2_cache_get(bs, s->l2_table_cache,
1288 l2_offset + slice * slice_size2,
1289 (void **) &l2_slice);
1290 if (ret < 0) {
1291 goto fail;
1294 for (j = 0; j < s->l2_slice_size; j++) {
1295 uint64_t cluster_index;
1296 uint64_t offset;
1298 entry = be64_to_cpu(l2_slice[j]);
1299 old_entry = entry;
1300 entry &= ~QCOW_OFLAG_COPIED;
1301 offset = entry & L2E_OFFSET_MASK;
1303 switch (qcow2_get_cluster_type(entry)) {
1304 case QCOW2_CLUSTER_COMPRESSED:
1305 nb_csectors = ((entry >> s->csize_shift) &
1306 s->csize_mask) + 1;
1307 if (addend != 0) {
1308 ret = update_refcount(
1309 bs, (entry & s->cluster_offset_mask) & ~511,
1310 nb_csectors * 512, abs(addend), addend < 0,
1311 QCOW2_DISCARD_SNAPSHOT);
1312 if (ret < 0) {
1313 goto fail;
1316 /* compressed clusters are never modified */
1317 refcount = 2;
1318 break;
1320 case QCOW2_CLUSTER_NORMAL:
1321 case QCOW2_CLUSTER_ZERO_ALLOC:
1322 if (offset_into_cluster(s, offset)) {
1323 /* Here l2_index means table (not slice) index */
1324 int l2_index = slice * s->l2_slice_size + j;
1325 qcow2_signal_corruption(
1326 bs, true, -1, -1, "Cluster "
1327 "allocation offset %#" PRIx64
1328 " unaligned (L2 offset: %#"
1329 PRIx64 ", L2 index: %#x)",
1330 offset, l2_offset, l2_index);
1331 ret = -EIO;
1332 goto fail;
1335 cluster_index = offset >> s->cluster_bits;
1336 assert(cluster_index);
1337 if (addend != 0) {
1338 ret = qcow2_update_cluster_refcount(
1339 bs, cluster_index, abs(addend), addend < 0,
1340 QCOW2_DISCARD_SNAPSHOT);
1341 if (ret < 0) {
1342 goto fail;
1346 ret = qcow2_get_refcount(bs, cluster_index, &refcount);
1347 if (ret < 0) {
1348 goto fail;
1350 break;
1352 case QCOW2_CLUSTER_ZERO_PLAIN:
1353 case QCOW2_CLUSTER_UNALLOCATED:
1354 refcount = 0;
1355 break;
1357 default:
1358 abort();
1361 if (refcount == 1) {
1362 entry |= QCOW_OFLAG_COPIED;
1364 if (entry != old_entry) {
1365 if (addend > 0) {
1366 qcow2_cache_set_dependency(bs, s->l2_table_cache,
1367 s->refcount_block_cache);
1369 l2_slice[j] = cpu_to_be64(entry);
1370 qcow2_cache_entry_mark_dirty(s->l2_table_cache,
1371 l2_slice);
1375 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1378 if (addend != 0) {
1379 ret = qcow2_update_cluster_refcount(bs, l2_offset >>
1380 s->cluster_bits,
1381 abs(addend), addend < 0,
1382 QCOW2_DISCARD_SNAPSHOT);
1383 if (ret < 0) {
1384 goto fail;
1387 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1388 &refcount);
1389 if (ret < 0) {
1390 goto fail;
1391 } else if (refcount == 1) {
1392 l2_offset |= QCOW_OFLAG_COPIED;
1394 if (l2_offset != old_l2_offset) {
1395 l1_table[i] = l2_offset;
1396 l1_modified = 1;
1401 ret = bdrv_flush(bs);
1402 fail:
1403 if (l2_slice) {
1404 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1407 s->cache_discards = false;
1408 qcow2_process_discards(bs, ret);
1410 /* Update L1 only if it isn't deleted anyway (addend = -1) */
1411 if (ret == 0 && addend >= 0 && l1_modified) {
1412 for (i = 0; i < l1_size; i++) {
1413 cpu_to_be64s(&l1_table[i]);
1416 ret = bdrv_pwrite_sync(bs->file, l1_table_offset,
1417 l1_table, l1_size2);
1419 for (i = 0; i < l1_size; i++) {
1420 be64_to_cpus(&l1_table[i]);
1423 if (l1_allocated)
1424 g_free(l1_table);
1425 return ret;
1431 /*********************************************************/
1432 /* refcount checking functions */
1435 static uint64_t refcount_array_byte_size(BDRVQcow2State *s, uint64_t entries)
1437 /* This assertion holds because there is no way we can address more than
1438 * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because
1439 * offsets have to be representable in bytes); due to every cluster
1440 * corresponding to one refcount entry, we are well below that limit */
1441 assert(entries < (UINT64_C(1) << (64 - 9)));
1443 /* Thanks to the assertion this will not overflow, because
1444 * s->refcount_order < 7.
1445 * (note: x << s->refcount_order == x * s->refcount_bits) */
1446 return DIV_ROUND_UP(entries << s->refcount_order, 8);
1450 * Reallocates *array so that it can hold new_size entries. *size must contain
1451 * the current number of entries in *array. If the reallocation fails, *array
1452 * and *size will not be modified and -errno will be returned. If the
1453 * reallocation is successful, *array will be set to the new buffer, *size
1454 * will be set to new_size and 0 will be returned. The size of the reallocated
1455 * refcount array buffer will be aligned to a cluster boundary, and the newly
1456 * allocated area will be zeroed.
1458 static int realloc_refcount_array(BDRVQcow2State *s, void **array,
1459 int64_t *size, int64_t new_size)
1461 int64_t old_byte_size, new_byte_size;
1462 void *new_ptr;
1464 /* Round to clusters so the array can be directly written to disk */
1465 old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size))
1466 * s->cluster_size;
1467 new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size))
1468 * s->cluster_size;
1470 if (new_byte_size == old_byte_size) {
1471 *size = new_size;
1472 return 0;
1475 assert(new_byte_size > 0);
1477 if (new_byte_size > SIZE_MAX) {
1478 return -ENOMEM;
1481 new_ptr = g_try_realloc(*array, new_byte_size);
1482 if (!new_ptr) {
1483 return -ENOMEM;
1486 if (new_byte_size > old_byte_size) {
1487 memset((char *)new_ptr + old_byte_size, 0,
1488 new_byte_size - old_byte_size);
1491 *array = new_ptr;
1492 *size = new_size;
1494 return 0;
1498 * Increases the refcount for a range of clusters in a given refcount table.
1499 * This is used to construct a temporary refcount table out of L1 and L2 tables
1500 * which can be compared to the refcount table saved in the image.
1502 * Modifies the number of errors in res.
1504 int qcow2_inc_refcounts_imrt(BlockDriverState *bs, BdrvCheckResult *res,
1505 void **refcount_table,
1506 int64_t *refcount_table_size,
1507 int64_t offset, int64_t size)
1509 BDRVQcow2State *s = bs->opaque;
1510 uint64_t start, last, cluster_offset, k, refcount;
1511 int ret;
1513 if (size <= 0) {
1514 return 0;
1517 start = start_of_cluster(s, offset);
1518 last = start_of_cluster(s, offset + size - 1);
1519 for(cluster_offset = start; cluster_offset <= last;
1520 cluster_offset += s->cluster_size) {
1521 k = cluster_offset >> s->cluster_bits;
1522 if (k >= *refcount_table_size) {
1523 ret = realloc_refcount_array(s, refcount_table,
1524 refcount_table_size, k + 1);
1525 if (ret < 0) {
1526 res->check_errors++;
1527 return ret;
1531 refcount = s->get_refcount(*refcount_table, k);
1532 if (refcount == s->refcount_max) {
1533 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64
1534 "\n", cluster_offset);
1535 fprintf(stderr, "Use qemu-img amend to increase the refcount entry "
1536 "width or qemu-img convert to create a clean copy if the "
1537 "image cannot be opened for writing\n");
1538 res->corruptions++;
1539 continue;
1541 s->set_refcount(*refcount_table, k, refcount + 1);
1544 return 0;
1547 /* Flags for check_refcounts_l1() and check_refcounts_l2() */
1548 enum {
1549 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */
1553 * Increases the refcount in the given refcount table for the all clusters
1554 * referenced in the L2 table. While doing so, performs some checks on L2
1555 * entries.
1557 * Returns the number of errors found by the checks or -errno if an internal
1558 * error occurred.
1560 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res,
1561 void **refcount_table,
1562 int64_t *refcount_table_size, int64_t l2_offset,
1563 int flags, BdrvCheckMode fix)
1565 BDRVQcow2State *s = bs->opaque;
1566 uint64_t *l2_table, l2_entry;
1567 uint64_t next_contiguous_offset = 0;
1568 int i, l2_size, nb_csectors, ret;
1570 /* Read L2 table from disk */
1571 l2_size = s->l2_size * sizeof(uint64_t);
1572 l2_table = g_malloc(l2_size);
1574 ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size);
1575 if (ret < 0) {
1576 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n");
1577 res->check_errors++;
1578 goto fail;
1581 /* Do the actual checks */
1582 for(i = 0; i < s->l2_size; i++) {
1583 l2_entry = be64_to_cpu(l2_table[i]);
1585 switch (qcow2_get_cluster_type(l2_entry)) {
1586 case QCOW2_CLUSTER_COMPRESSED:
1587 /* Compressed clusters don't have QCOW_OFLAG_COPIED */
1588 if (l2_entry & QCOW_OFLAG_COPIED) {
1589 fprintf(stderr, "ERROR: coffset=0x%" PRIx64 ": "
1590 "copied flag must never be set for compressed "
1591 "clusters\n", l2_entry & s->cluster_offset_mask);
1592 l2_entry &= ~QCOW_OFLAG_COPIED;
1593 res->corruptions++;
1596 /* Mark cluster as used */
1597 nb_csectors = ((l2_entry >> s->csize_shift) &
1598 s->csize_mask) + 1;
1599 l2_entry &= s->cluster_offset_mask;
1600 ret = qcow2_inc_refcounts_imrt(bs, res,
1601 refcount_table, refcount_table_size,
1602 l2_entry & ~511, nb_csectors * 512);
1603 if (ret < 0) {
1604 goto fail;
1607 if (flags & CHECK_FRAG_INFO) {
1608 res->bfi.allocated_clusters++;
1609 res->bfi.compressed_clusters++;
1611 /* Compressed clusters are fragmented by nature. Since they
1612 * take up sub-sector space but we only have sector granularity
1613 * I/O we need to re-read the same sectors even for adjacent
1614 * compressed clusters.
1616 res->bfi.fragmented_clusters++;
1618 break;
1620 case QCOW2_CLUSTER_ZERO_ALLOC:
1621 case QCOW2_CLUSTER_NORMAL:
1623 uint64_t offset = l2_entry & L2E_OFFSET_MASK;
1625 if (flags & CHECK_FRAG_INFO) {
1626 res->bfi.allocated_clusters++;
1627 if (next_contiguous_offset &&
1628 offset != next_contiguous_offset) {
1629 res->bfi.fragmented_clusters++;
1631 next_contiguous_offset = offset + s->cluster_size;
1634 /* Correct offsets are cluster aligned */
1635 if (offset_into_cluster(s, offset)) {
1636 if (qcow2_get_cluster_type(l2_entry) ==
1637 QCOW2_CLUSTER_ZERO_ALLOC)
1639 fprintf(stderr, "%s offset=%" PRIx64 ": Preallocated zero "
1640 "cluster is not properly aligned; L2 entry "
1641 "corrupted.\n",
1642 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR",
1643 offset);
1644 if (fix & BDRV_FIX_ERRORS) {
1645 uint64_t l2e_offset =
1646 l2_offset + (uint64_t)i * sizeof(uint64_t);
1648 l2_entry = QCOW_OFLAG_ZERO;
1649 l2_table[i] = cpu_to_be64(l2_entry);
1650 ret = qcow2_pre_write_overlap_check(bs,
1651 QCOW2_OL_ACTIVE_L2 | QCOW2_OL_INACTIVE_L2,
1652 l2e_offset, sizeof(uint64_t));
1653 if (ret < 0) {
1654 fprintf(stderr, "ERROR: Overlap check failed\n");
1655 res->check_errors++;
1656 /* Something is seriously wrong, so abort checking
1657 * this L2 table */
1658 goto fail;
1661 ret = bdrv_pwrite_sync(bs->file, l2e_offset,
1662 &l2_table[i], sizeof(uint64_t));
1663 if (ret < 0) {
1664 fprintf(stderr, "ERROR: Failed to overwrite L2 "
1665 "table entry: %s\n", strerror(-ret));
1666 res->check_errors++;
1667 /* Do not abort, continue checking the rest of this
1668 * L2 table's entries */
1669 } else {
1670 res->corruptions_fixed++;
1671 /* Skip marking the cluster as used
1672 * (it is unused now) */
1673 continue;
1675 } else {
1676 res->corruptions++;
1678 } else {
1679 fprintf(stderr, "ERROR offset=%" PRIx64 ": Data cluster is "
1680 "not properly aligned; L2 entry corrupted.\n", offset);
1681 res->corruptions++;
1685 /* Mark cluster as used */
1686 ret = qcow2_inc_refcounts_imrt(bs, res,
1687 refcount_table, refcount_table_size,
1688 offset, s->cluster_size);
1689 if (ret < 0) {
1690 goto fail;
1692 break;
1695 case QCOW2_CLUSTER_ZERO_PLAIN:
1696 case QCOW2_CLUSTER_UNALLOCATED:
1697 break;
1699 default:
1700 abort();
1704 g_free(l2_table);
1705 return 0;
1707 fail:
1708 g_free(l2_table);
1709 return ret;
1713 * Increases the refcount for the L1 table, its L2 tables and all referenced
1714 * clusters in the given refcount table. While doing so, performs some checks
1715 * on L1 and L2 entries.
1717 * Returns the number of errors found by the checks or -errno if an internal
1718 * error occurred.
1720 static int check_refcounts_l1(BlockDriverState *bs,
1721 BdrvCheckResult *res,
1722 void **refcount_table,
1723 int64_t *refcount_table_size,
1724 int64_t l1_table_offset, int l1_size,
1725 int flags, BdrvCheckMode fix)
1727 BDRVQcow2State *s = bs->opaque;
1728 uint64_t *l1_table = NULL, l2_offset, l1_size2;
1729 int i, ret;
1731 l1_size2 = l1_size * sizeof(uint64_t);
1733 /* Mark L1 table as used */
1734 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, refcount_table_size,
1735 l1_table_offset, l1_size2);
1736 if (ret < 0) {
1737 goto fail;
1740 /* Read L1 table entries from disk */
1741 if (l1_size2 > 0) {
1742 l1_table = g_try_malloc(l1_size2);
1743 if (l1_table == NULL) {
1744 ret = -ENOMEM;
1745 res->check_errors++;
1746 goto fail;
1748 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2);
1749 if (ret < 0) {
1750 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n");
1751 res->check_errors++;
1752 goto fail;
1754 for(i = 0;i < l1_size; i++)
1755 be64_to_cpus(&l1_table[i]);
1758 /* Do the actual checks */
1759 for(i = 0; i < l1_size; i++) {
1760 l2_offset = l1_table[i];
1761 if (l2_offset) {
1762 /* Mark L2 table as used */
1763 l2_offset &= L1E_OFFSET_MASK;
1764 ret = qcow2_inc_refcounts_imrt(bs, res,
1765 refcount_table, refcount_table_size,
1766 l2_offset, s->cluster_size);
1767 if (ret < 0) {
1768 goto fail;
1771 /* L2 tables are cluster aligned */
1772 if (offset_into_cluster(s, l2_offset)) {
1773 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not "
1774 "cluster aligned; L1 entry corrupted\n", l2_offset);
1775 res->corruptions++;
1778 /* Process and check L2 entries */
1779 ret = check_refcounts_l2(bs, res, refcount_table,
1780 refcount_table_size, l2_offset, flags,
1781 fix);
1782 if (ret < 0) {
1783 goto fail;
1787 g_free(l1_table);
1788 return 0;
1790 fail:
1791 g_free(l1_table);
1792 return ret;
1796 * Checks the OFLAG_COPIED flag for all L1 and L2 entries.
1798 * This function does not print an error message nor does it increment
1799 * check_errors if qcow2_get_refcount fails (this is because such an error will
1800 * have been already detected and sufficiently signaled by the calling function
1801 * (qcow2_check_refcounts) by the time this function is called).
1803 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res,
1804 BdrvCheckMode fix)
1806 BDRVQcow2State *s = bs->opaque;
1807 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size);
1808 int ret;
1809 uint64_t refcount;
1810 int i, j;
1811 bool repair;
1813 if (fix & BDRV_FIX_ERRORS) {
1814 /* Always repair */
1815 repair = true;
1816 } else if (fix & BDRV_FIX_LEAKS) {
1817 /* Repair only if that seems safe: This function is always
1818 * called after the refcounts have been fixed, so the refcount
1819 * is accurate if that repair was successful */
1820 repair = !res->check_errors && !res->corruptions && !res->leaks;
1821 } else {
1822 repair = false;
1825 for (i = 0; i < s->l1_size; i++) {
1826 uint64_t l1_entry = s->l1_table[i];
1827 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK;
1828 bool l2_dirty = false;
1830 if (!l2_offset) {
1831 continue;
1834 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1835 &refcount);
1836 if (ret < 0) {
1837 /* don't print message nor increment check_errors */
1838 continue;
1840 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) {
1841 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d "
1842 "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1843 repair ? "Repairing" : "ERROR", i, l1_entry, refcount);
1844 if (repair) {
1845 s->l1_table[i] = refcount == 1
1846 ? l1_entry | QCOW_OFLAG_COPIED
1847 : l1_entry & ~QCOW_OFLAG_COPIED;
1848 ret = qcow2_write_l1_entry(bs, i);
1849 if (ret < 0) {
1850 res->check_errors++;
1851 goto fail;
1853 res->corruptions_fixed++;
1854 } else {
1855 res->corruptions++;
1859 ret = bdrv_pread(bs->file, l2_offset, l2_table,
1860 s->l2_size * sizeof(uint64_t));
1861 if (ret < 0) {
1862 fprintf(stderr, "ERROR: Could not read L2 table: %s\n",
1863 strerror(-ret));
1864 res->check_errors++;
1865 goto fail;
1868 for (j = 0; j < s->l2_size; j++) {
1869 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1870 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK;
1871 QCow2ClusterType cluster_type = qcow2_get_cluster_type(l2_entry);
1873 if (cluster_type == QCOW2_CLUSTER_NORMAL ||
1874 cluster_type == QCOW2_CLUSTER_ZERO_ALLOC) {
1875 ret = qcow2_get_refcount(bs,
1876 data_offset >> s->cluster_bits,
1877 &refcount);
1878 if (ret < 0) {
1879 /* don't print message nor increment check_errors */
1880 continue;
1882 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) {
1883 fprintf(stderr, "%s OFLAG_COPIED data cluster: "
1884 "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n",
1885 repair ? "Repairing" : "ERROR", l2_entry, refcount);
1886 if (repair) {
1887 l2_table[j] = cpu_to_be64(refcount == 1
1888 ? l2_entry | QCOW_OFLAG_COPIED
1889 : l2_entry & ~QCOW_OFLAG_COPIED);
1890 l2_dirty = true;
1891 res->corruptions_fixed++;
1892 } else {
1893 res->corruptions++;
1899 if (l2_dirty) {
1900 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2,
1901 l2_offset, s->cluster_size);
1902 if (ret < 0) {
1903 fprintf(stderr, "ERROR: Could not write L2 table; metadata "
1904 "overlap check failed: %s\n", strerror(-ret));
1905 res->check_errors++;
1906 goto fail;
1909 ret = bdrv_pwrite(bs->file, l2_offset, l2_table,
1910 s->cluster_size);
1911 if (ret < 0) {
1912 fprintf(stderr, "ERROR: Could not write L2 table: %s\n",
1913 strerror(-ret));
1914 res->check_errors++;
1915 goto fail;
1920 ret = 0;
1922 fail:
1923 qemu_vfree(l2_table);
1924 return ret;
1928 * Checks consistency of refblocks and accounts for each refblock in
1929 * *refcount_table.
1931 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res,
1932 BdrvCheckMode fix, bool *rebuild,
1933 void **refcount_table, int64_t *nb_clusters)
1935 BDRVQcow2State *s = bs->opaque;
1936 int64_t i, size;
1937 int ret;
1939 for(i = 0; i < s->refcount_table_size; i++) {
1940 uint64_t offset, cluster;
1941 offset = s->refcount_table[i];
1942 cluster = offset >> s->cluster_bits;
1944 /* Refcount blocks are cluster aligned */
1945 if (offset_into_cluster(s, offset)) {
1946 fprintf(stderr, "ERROR refcount block %" PRId64 " is not "
1947 "cluster aligned; refcount table entry corrupted\n", i);
1948 res->corruptions++;
1949 *rebuild = true;
1950 continue;
1953 if (cluster >= *nb_clusters) {
1954 fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n",
1955 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i);
1957 if (fix & BDRV_FIX_ERRORS) {
1958 int64_t new_nb_clusters;
1959 Error *local_err = NULL;
1961 if (offset > INT64_MAX - s->cluster_size) {
1962 ret = -EINVAL;
1963 goto resize_fail;
1966 ret = bdrv_truncate(bs->file, offset + s->cluster_size,
1967 PREALLOC_MODE_OFF, &local_err);
1968 if (ret < 0) {
1969 error_report_err(local_err);
1970 goto resize_fail;
1972 size = bdrv_getlength(bs->file->bs);
1973 if (size < 0) {
1974 ret = size;
1975 goto resize_fail;
1978 new_nb_clusters = size_to_clusters(s, size);
1979 assert(new_nb_clusters >= *nb_clusters);
1981 ret = realloc_refcount_array(s, refcount_table,
1982 nb_clusters, new_nb_clusters);
1983 if (ret < 0) {
1984 res->check_errors++;
1985 return ret;
1988 if (cluster >= *nb_clusters) {
1989 ret = -EINVAL;
1990 goto resize_fail;
1993 res->corruptions_fixed++;
1994 ret = qcow2_inc_refcounts_imrt(bs, res,
1995 refcount_table, nb_clusters,
1996 offset, s->cluster_size);
1997 if (ret < 0) {
1998 return ret;
2000 /* No need to check whether the refcount is now greater than 1:
2001 * This area was just allocated and zeroed, so it can only be
2002 * exactly 1 after qcow2_inc_refcounts_imrt() */
2003 continue;
2005 resize_fail:
2006 res->corruptions++;
2007 *rebuild = true;
2008 fprintf(stderr, "ERROR could not resize image: %s\n",
2009 strerror(-ret));
2010 } else {
2011 res->corruptions++;
2013 continue;
2016 if (offset != 0) {
2017 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2018 offset, s->cluster_size);
2019 if (ret < 0) {
2020 return ret;
2022 if (s->get_refcount(*refcount_table, cluster) != 1) {
2023 fprintf(stderr, "ERROR refcount block %" PRId64
2024 " refcount=%" PRIu64 "\n", i,
2025 s->get_refcount(*refcount_table, cluster));
2026 res->corruptions++;
2027 *rebuild = true;
2032 return 0;
2036 * Calculates an in-memory refcount table.
2038 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2039 BdrvCheckMode fix, bool *rebuild,
2040 void **refcount_table, int64_t *nb_clusters)
2042 BDRVQcow2State *s = bs->opaque;
2043 int64_t i;
2044 QCowSnapshot *sn;
2045 int ret;
2047 if (!*refcount_table) {
2048 int64_t old_size = 0;
2049 ret = realloc_refcount_array(s, refcount_table,
2050 &old_size, *nb_clusters);
2051 if (ret < 0) {
2052 res->check_errors++;
2053 return ret;
2057 /* header */
2058 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2059 0, s->cluster_size);
2060 if (ret < 0) {
2061 return ret;
2064 /* current L1 table */
2065 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2066 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO,
2067 fix);
2068 if (ret < 0) {
2069 return ret;
2072 /* snapshots */
2073 for (i = 0; i < s->nb_snapshots; i++) {
2074 sn = s->snapshots + i;
2075 if (offset_into_cluster(s, sn->l1_table_offset)) {
2076 fprintf(stderr, "ERROR snapshot %s (%s) l1_offset=%#" PRIx64 ": "
2077 "L1 table is not cluster aligned; snapshot table entry "
2078 "corrupted\n", sn->id_str, sn->name, sn->l1_table_offset);
2079 res->corruptions++;
2080 continue;
2082 if (sn->l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) {
2083 fprintf(stderr, "ERROR snapshot %s (%s) l1_size=%#" PRIx32 ": "
2084 "L1 table is too large; snapshot table entry corrupted\n",
2085 sn->id_str, sn->name, sn->l1_size);
2086 res->corruptions++;
2087 continue;
2089 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters,
2090 sn->l1_table_offset, sn->l1_size, 0, fix);
2091 if (ret < 0) {
2092 return ret;
2095 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2096 s->snapshots_offset, s->snapshots_size);
2097 if (ret < 0) {
2098 return ret;
2101 /* refcount data */
2102 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2103 s->refcount_table_offset,
2104 s->refcount_table_size * sizeof(uint64_t));
2105 if (ret < 0) {
2106 return ret;
2109 /* encryption */
2110 if (s->crypto_header.length) {
2111 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters,
2112 s->crypto_header.offset,
2113 s->crypto_header.length);
2114 if (ret < 0) {
2115 return ret;
2119 /* bitmaps */
2120 ret = qcow2_check_bitmaps_refcounts(bs, res, refcount_table, nb_clusters);
2121 if (ret < 0) {
2122 return ret;
2125 return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters);
2129 * Compares the actual reference count for each cluster in the image against the
2130 * refcount as reported by the refcount structures on-disk.
2132 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2133 BdrvCheckMode fix, bool *rebuild,
2134 int64_t *highest_cluster,
2135 void *refcount_table, int64_t nb_clusters)
2137 BDRVQcow2State *s = bs->opaque;
2138 int64_t i;
2139 uint64_t refcount1, refcount2;
2140 int ret;
2142 for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) {
2143 ret = qcow2_get_refcount(bs, i, &refcount1);
2144 if (ret < 0) {
2145 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
2146 i, strerror(-ret));
2147 res->check_errors++;
2148 continue;
2151 refcount2 = s->get_refcount(refcount_table, i);
2153 if (refcount1 > 0 || refcount2 > 0) {
2154 *highest_cluster = i;
2157 if (refcount1 != refcount2) {
2158 /* Check if we're allowed to fix the mismatch */
2159 int *num_fixed = NULL;
2160 if (refcount1 == 0) {
2161 *rebuild = true;
2162 } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) {
2163 num_fixed = &res->leaks_fixed;
2164 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) {
2165 num_fixed = &res->corruptions_fixed;
2168 fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64
2169 " reference=%" PRIu64 "\n",
2170 num_fixed != NULL ? "Repairing" :
2171 refcount1 < refcount2 ? "ERROR" :
2172 "Leaked",
2173 i, refcount1, refcount2);
2175 if (num_fixed) {
2176 ret = update_refcount(bs, i << s->cluster_bits, 1,
2177 refcount_diff(refcount1, refcount2),
2178 refcount1 > refcount2,
2179 QCOW2_DISCARD_ALWAYS);
2180 if (ret >= 0) {
2181 (*num_fixed)++;
2182 continue;
2186 /* And if we couldn't, print an error */
2187 if (refcount1 < refcount2) {
2188 res->corruptions++;
2189 } else {
2190 res->leaks++;
2197 * Allocates clusters using an in-memory refcount table (IMRT) in contrast to
2198 * the on-disk refcount structures.
2200 * On input, *first_free_cluster tells where to start looking, and need not
2201 * actually be a free cluster; the returned offset will not be before that
2202 * cluster. On output, *first_free_cluster points to the first gap found, even
2203 * if that gap was too small to be used as the returned offset.
2205 * Note that *first_free_cluster is a cluster index whereas the return value is
2206 * an offset.
2208 static int64_t alloc_clusters_imrt(BlockDriverState *bs,
2209 int cluster_count,
2210 void **refcount_table,
2211 int64_t *imrt_nb_clusters,
2212 int64_t *first_free_cluster)
2214 BDRVQcow2State *s = bs->opaque;
2215 int64_t cluster = *first_free_cluster, i;
2216 bool first_gap = true;
2217 int contiguous_free_clusters;
2218 int ret;
2220 /* Starting at *first_free_cluster, find a range of at least cluster_count
2221 * continuously free clusters */
2222 for (contiguous_free_clusters = 0;
2223 cluster < *imrt_nb_clusters &&
2224 contiguous_free_clusters < cluster_count;
2225 cluster++)
2227 if (!s->get_refcount(*refcount_table, cluster)) {
2228 contiguous_free_clusters++;
2229 if (first_gap) {
2230 /* If this is the first free cluster found, update
2231 * *first_free_cluster accordingly */
2232 *first_free_cluster = cluster;
2233 first_gap = false;
2235 } else if (contiguous_free_clusters) {
2236 contiguous_free_clusters = 0;
2240 /* If contiguous_free_clusters is greater than zero, it contains the number
2241 * of continuously free clusters until the current cluster; the first free
2242 * cluster in the current "gap" is therefore
2243 * cluster - contiguous_free_clusters */
2245 /* If no such range could be found, grow the in-memory refcount table
2246 * accordingly to append free clusters at the end of the image */
2247 if (contiguous_free_clusters < cluster_count) {
2248 /* contiguous_free_clusters clusters are already empty at the image end;
2249 * we need cluster_count clusters; therefore, we have to allocate
2250 * cluster_count - contiguous_free_clusters new clusters at the end of
2251 * the image (which is the current value of cluster; note that cluster
2252 * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond
2253 * the image end) */
2254 ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters,
2255 cluster + cluster_count
2256 - contiguous_free_clusters);
2257 if (ret < 0) {
2258 return ret;
2262 /* Go back to the first free cluster */
2263 cluster -= contiguous_free_clusters;
2264 for (i = 0; i < cluster_count; i++) {
2265 s->set_refcount(*refcount_table, cluster + i, 1);
2268 return cluster << s->cluster_bits;
2272 * Creates a new refcount structure based solely on the in-memory information
2273 * given through *refcount_table. All necessary allocations will be reflected
2274 * in that array.
2276 * On success, the old refcount structure is leaked (it will be covered by the
2277 * new refcount structure).
2279 static int rebuild_refcount_structure(BlockDriverState *bs,
2280 BdrvCheckResult *res,
2281 void **refcount_table,
2282 int64_t *nb_clusters)
2284 BDRVQcow2State *s = bs->opaque;
2285 int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0;
2286 int64_t refblock_offset, refblock_start, refblock_index;
2287 uint32_t reftable_size = 0;
2288 uint64_t *on_disk_reftable = NULL;
2289 void *on_disk_refblock;
2290 int ret = 0;
2291 struct {
2292 uint64_t reftable_offset;
2293 uint32_t reftable_clusters;
2294 } QEMU_PACKED reftable_offset_and_clusters;
2296 qcow2_cache_empty(bs, s->refcount_block_cache);
2298 write_refblocks:
2299 for (; cluster < *nb_clusters; cluster++) {
2300 if (!s->get_refcount(*refcount_table, cluster)) {
2301 continue;
2304 refblock_index = cluster >> s->refcount_block_bits;
2305 refblock_start = refblock_index << s->refcount_block_bits;
2307 /* Don't allocate a cluster in a refblock already written to disk */
2308 if (first_free_cluster < refblock_start) {
2309 first_free_cluster = refblock_start;
2311 refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table,
2312 nb_clusters, &first_free_cluster);
2313 if (refblock_offset < 0) {
2314 fprintf(stderr, "ERROR allocating refblock: %s\n",
2315 strerror(-refblock_offset));
2316 res->check_errors++;
2317 ret = refblock_offset;
2318 goto fail;
2321 if (reftable_size <= refblock_index) {
2322 uint32_t old_reftable_size = reftable_size;
2323 uint64_t *new_on_disk_reftable;
2325 reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t),
2326 s->cluster_size) / sizeof(uint64_t);
2327 new_on_disk_reftable = g_try_realloc(on_disk_reftable,
2328 reftable_size *
2329 sizeof(uint64_t));
2330 if (!new_on_disk_reftable) {
2331 res->check_errors++;
2332 ret = -ENOMEM;
2333 goto fail;
2335 on_disk_reftable = new_on_disk_reftable;
2337 memset(on_disk_reftable + old_reftable_size, 0,
2338 (reftable_size - old_reftable_size) * sizeof(uint64_t));
2340 /* The offset we have for the reftable is now no longer valid;
2341 * this will leak that range, but we can easily fix that by running
2342 * a leak-fixing check after this rebuild operation */
2343 reftable_offset = -1;
2344 } else {
2345 assert(on_disk_reftable);
2347 on_disk_reftable[refblock_index] = refblock_offset;
2349 /* If this is apparently the last refblock (for now), try to squeeze the
2350 * reftable in */
2351 if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits &&
2352 reftable_offset < 0)
2354 uint64_t reftable_clusters = size_to_clusters(s, reftable_size *
2355 sizeof(uint64_t));
2356 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2357 refcount_table, nb_clusters,
2358 &first_free_cluster);
2359 if (reftable_offset < 0) {
2360 fprintf(stderr, "ERROR allocating reftable: %s\n",
2361 strerror(-reftable_offset));
2362 res->check_errors++;
2363 ret = reftable_offset;
2364 goto fail;
2368 ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset,
2369 s->cluster_size);
2370 if (ret < 0) {
2371 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2372 goto fail;
2375 /* The size of *refcount_table is always cluster-aligned, therefore the
2376 * write operation will not overflow */
2377 on_disk_refblock = (void *)((char *) *refcount_table +
2378 refblock_index * s->cluster_size);
2380 ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE,
2381 on_disk_refblock, s->cluster_sectors);
2382 if (ret < 0) {
2383 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret));
2384 goto fail;
2387 /* Go to the end of this refblock */
2388 cluster = refblock_start + s->refcount_block_size - 1;
2391 if (reftable_offset < 0) {
2392 uint64_t post_refblock_start, reftable_clusters;
2394 post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size);
2395 reftable_clusters = size_to_clusters(s,
2396 reftable_size * sizeof(uint64_t));
2397 /* Not pretty but simple */
2398 if (first_free_cluster < post_refblock_start) {
2399 first_free_cluster = post_refblock_start;
2401 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters,
2402 refcount_table, nb_clusters,
2403 &first_free_cluster);
2404 if (reftable_offset < 0) {
2405 fprintf(stderr, "ERROR allocating reftable: %s\n",
2406 strerror(-reftable_offset));
2407 res->check_errors++;
2408 ret = reftable_offset;
2409 goto fail;
2412 goto write_refblocks;
2415 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2416 cpu_to_be64s(&on_disk_reftable[refblock_index]);
2419 ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset,
2420 reftable_size * sizeof(uint64_t));
2421 if (ret < 0) {
2422 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2423 goto fail;
2426 assert(reftable_size < INT_MAX / sizeof(uint64_t));
2427 ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable,
2428 reftable_size * sizeof(uint64_t));
2429 if (ret < 0) {
2430 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret));
2431 goto fail;
2434 /* Enter new reftable into the image header */
2435 reftable_offset_and_clusters.reftable_offset = cpu_to_be64(reftable_offset);
2436 reftable_offset_and_clusters.reftable_clusters =
2437 cpu_to_be32(size_to_clusters(s, reftable_size * sizeof(uint64_t)));
2438 ret = bdrv_pwrite_sync(bs->file,
2439 offsetof(QCowHeader, refcount_table_offset),
2440 &reftable_offset_and_clusters,
2441 sizeof(reftable_offset_and_clusters));
2442 if (ret < 0) {
2443 fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret));
2444 goto fail;
2447 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) {
2448 be64_to_cpus(&on_disk_reftable[refblock_index]);
2450 s->refcount_table = on_disk_reftable;
2451 s->refcount_table_offset = reftable_offset;
2452 s->refcount_table_size = reftable_size;
2453 update_max_refcount_table_index(s);
2455 return 0;
2457 fail:
2458 g_free(on_disk_reftable);
2459 return ret;
2463 * Checks an image for refcount consistency.
2465 * Returns 0 if no errors are found, the number of errors in case the image is
2466 * detected as corrupted, and -errno when an internal error occurred.
2468 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res,
2469 BdrvCheckMode fix)
2471 BDRVQcow2State *s = bs->opaque;
2472 BdrvCheckResult pre_compare_res;
2473 int64_t size, highest_cluster, nb_clusters;
2474 void *refcount_table = NULL;
2475 bool rebuild = false;
2476 int ret;
2478 size = bdrv_getlength(bs->file->bs);
2479 if (size < 0) {
2480 res->check_errors++;
2481 return size;
2484 nb_clusters = size_to_clusters(s, size);
2485 if (nb_clusters > INT_MAX) {
2486 res->check_errors++;
2487 return -EFBIG;
2490 res->bfi.total_clusters =
2491 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE);
2493 ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table,
2494 &nb_clusters);
2495 if (ret < 0) {
2496 goto fail;
2499 /* In case we don't need to rebuild the refcount structure (but want to fix
2500 * something), this function is immediately called again, in which case the
2501 * result should be ignored */
2502 pre_compare_res = *res;
2503 compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table,
2504 nb_clusters);
2506 if (rebuild && (fix & BDRV_FIX_ERRORS)) {
2507 BdrvCheckResult old_res = *res;
2508 int fresh_leaks = 0;
2510 fprintf(stderr, "Rebuilding refcount structure\n");
2511 ret = rebuild_refcount_structure(bs, res, &refcount_table,
2512 &nb_clusters);
2513 if (ret < 0) {
2514 goto fail;
2517 res->corruptions = 0;
2518 res->leaks = 0;
2520 /* Because the old reftable has been exchanged for a new one the
2521 * references have to be recalculated */
2522 rebuild = false;
2523 memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters));
2524 ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table,
2525 &nb_clusters);
2526 if (ret < 0) {
2527 goto fail;
2530 if (fix & BDRV_FIX_LEAKS) {
2531 /* The old refcount structures are now leaked, fix it; the result
2532 * can be ignored, aside from leaks which were introduced by
2533 * rebuild_refcount_structure() that could not be fixed */
2534 BdrvCheckResult saved_res = *res;
2535 *res = (BdrvCheckResult){ 0 };
2537 compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild,
2538 &highest_cluster, refcount_table, nb_clusters);
2539 if (rebuild) {
2540 fprintf(stderr, "ERROR rebuilt refcount structure is still "
2541 "broken\n");
2544 /* Any leaks accounted for here were introduced by
2545 * rebuild_refcount_structure() because that function has created a
2546 * new refcount structure from scratch */
2547 fresh_leaks = res->leaks;
2548 *res = saved_res;
2551 if (res->corruptions < old_res.corruptions) {
2552 res->corruptions_fixed += old_res.corruptions - res->corruptions;
2554 if (res->leaks < old_res.leaks) {
2555 res->leaks_fixed += old_res.leaks - res->leaks;
2557 res->leaks += fresh_leaks;
2558 } else if (fix) {
2559 if (rebuild) {
2560 fprintf(stderr, "ERROR need to rebuild refcount structures\n");
2561 res->check_errors++;
2562 ret = -EIO;
2563 goto fail;
2566 if (res->leaks || res->corruptions) {
2567 *res = pre_compare_res;
2568 compare_refcounts(bs, res, fix, &rebuild, &highest_cluster,
2569 refcount_table, nb_clusters);
2573 /* check OFLAG_COPIED */
2574 ret = check_oflag_copied(bs, res, fix);
2575 if (ret < 0) {
2576 goto fail;
2579 res->image_end_offset = (highest_cluster + 1) * s->cluster_size;
2580 ret = 0;
2582 fail:
2583 g_free(refcount_table);
2585 return ret;
2588 #define overlaps_with(ofs, sz) \
2589 ranges_overlap(offset, size, ofs, sz)
2592 * Checks if the given offset into the image file is actually free to use by
2593 * looking for overlaps with important metadata sections (L1/L2 tables etc.),
2594 * i.e. a sanity check without relying on the refcount tables.
2596 * The ign parameter specifies what checks not to perform (being a bitmask of
2597 * QCow2MetadataOverlap values), i.e., what sections to ignore.
2599 * Returns:
2600 * - 0 if writing to this offset will not affect the mentioned metadata
2601 * - a positive QCow2MetadataOverlap value indicating one overlapping section
2602 * - a negative value (-errno) indicating an error while performing a check,
2603 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2
2605 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset,
2606 int64_t size)
2608 BDRVQcow2State *s = bs->opaque;
2609 int chk = s->overlap_check & ~ign;
2610 int i, j;
2612 if (!size) {
2613 return 0;
2616 if (chk & QCOW2_OL_MAIN_HEADER) {
2617 if (offset < s->cluster_size) {
2618 return QCOW2_OL_MAIN_HEADER;
2622 /* align range to test to cluster boundaries */
2623 size = ROUND_UP(offset_into_cluster(s, offset) + size, s->cluster_size);
2624 offset = start_of_cluster(s, offset);
2626 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) {
2627 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) {
2628 return QCOW2_OL_ACTIVE_L1;
2632 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) {
2633 if (overlaps_with(s->refcount_table_offset,
2634 s->refcount_table_size * sizeof(uint64_t))) {
2635 return QCOW2_OL_REFCOUNT_TABLE;
2639 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) {
2640 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) {
2641 return QCOW2_OL_SNAPSHOT_TABLE;
2645 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) {
2646 for (i = 0; i < s->nb_snapshots; i++) {
2647 if (s->snapshots[i].l1_size &&
2648 overlaps_with(s->snapshots[i].l1_table_offset,
2649 s->snapshots[i].l1_size * sizeof(uint64_t))) {
2650 return QCOW2_OL_INACTIVE_L1;
2655 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) {
2656 for (i = 0; i < s->l1_size; i++) {
2657 if ((s->l1_table[i] & L1E_OFFSET_MASK) &&
2658 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK,
2659 s->cluster_size)) {
2660 return QCOW2_OL_ACTIVE_L2;
2665 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) {
2666 unsigned last_entry = s->max_refcount_table_index;
2667 assert(last_entry < s->refcount_table_size);
2668 assert(last_entry + 1 == s->refcount_table_size ||
2669 (s->refcount_table[last_entry + 1] & REFT_OFFSET_MASK) == 0);
2670 for (i = 0; i <= last_entry; i++) {
2671 if ((s->refcount_table[i] & REFT_OFFSET_MASK) &&
2672 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK,
2673 s->cluster_size)) {
2674 return QCOW2_OL_REFCOUNT_BLOCK;
2679 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) {
2680 for (i = 0; i < s->nb_snapshots; i++) {
2681 uint64_t l1_ofs = s->snapshots[i].l1_table_offset;
2682 uint32_t l1_sz = s->snapshots[i].l1_size;
2683 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t);
2684 uint64_t *l1;
2685 int ret;
2687 ret = qcow2_validate_table(bs, l1_ofs, l1_sz, sizeof(uint64_t),
2688 QCOW_MAX_L1_SIZE, "", NULL);
2689 if (ret < 0) {
2690 return ret;
2693 l1 = g_try_malloc(l1_sz2);
2695 if (l1_sz2 && l1 == NULL) {
2696 return -ENOMEM;
2699 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2);
2700 if (ret < 0) {
2701 g_free(l1);
2702 return ret;
2705 for (j = 0; j < l1_sz; j++) {
2706 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK;
2707 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) {
2708 g_free(l1);
2709 return QCOW2_OL_INACTIVE_L2;
2713 g_free(l1);
2717 if ((chk & QCOW2_OL_BITMAP_DIRECTORY) &&
2718 (s->autoclear_features & QCOW2_AUTOCLEAR_BITMAPS))
2720 if (overlaps_with(s->bitmap_directory_offset,
2721 s->bitmap_directory_size))
2723 return QCOW2_OL_BITMAP_DIRECTORY;
2727 return 0;
2730 static const char *metadata_ol_names[] = {
2731 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header",
2732 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table",
2733 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table",
2734 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table",
2735 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block",
2736 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table",
2737 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table",
2738 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table",
2739 [QCOW2_OL_BITMAP_DIRECTORY_BITNR] = "bitmap directory",
2741 QEMU_BUILD_BUG_ON(QCOW2_OL_MAX_BITNR != ARRAY_SIZE(metadata_ol_names));
2744 * First performs a check for metadata overlaps (through
2745 * qcow2_check_metadata_overlap); if that fails with a negative value (error
2746 * while performing a check), that value is returned. If an impending overlap
2747 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt
2748 * and -EIO returned.
2750 * Returns 0 if there were neither overlaps nor errors while checking for
2751 * overlaps; or a negative value (-errno) on error.
2753 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset,
2754 int64_t size)
2756 int ret = qcow2_check_metadata_overlap(bs, ign, offset, size);
2758 if (ret < 0) {
2759 return ret;
2760 } else if (ret > 0) {
2761 int metadata_ol_bitnr = ctz32(ret);
2762 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR);
2764 qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid "
2765 "write on metadata (overlaps with %s)",
2766 metadata_ol_names[metadata_ol_bitnr]);
2767 return -EIO;
2770 return 0;
2773 /* A pointer to a function of this type is given to walk_over_reftable(). That
2774 * function will create refblocks and pass them to a RefblockFinishOp once they
2775 * are completed (@refblock). @refblock_empty is set if the refblock is
2776 * completely empty.
2778 * Along with the refblock, a corresponding reftable entry is passed, in the
2779 * reftable @reftable (which may be reallocated) at @reftable_index.
2781 * @allocated should be set to true if a new cluster has been allocated.
2783 typedef int (RefblockFinishOp)(BlockDriverState *bs, uint64_t **reftable,
2784 uint64_t reftable_index, uint64_t *reftable_size,
2785 void *refblock, bool refblock_empty,
2786 bool *allocated, Error **errp);
2789 * This "operation" for walk_over_reftable() allocates the refblock on disk (if
2790 * it is not empty) and inserts its offset into the new reftable. The size of
2791 * this new reftable is increased as required.
2793 static int alloc_refblock(BlockDriverState *bs, uint64_t **reftable,
2794 uint64_t reftable_index, uint64_t *reftable_size,
2795 void *refblock, bool refblock_empty, bool *allocated,
2796 Error **errp)
2798 BDRVQcow2State *s = bs->opaque;
2799 int64_t offset;
2801 if (!refblock_empty && reftable_index >= *reftable_size) {
2802 uint64_t *new_reftable;
2803 uint64_t new_reftable_size;
2805 new_reftable_size = ROUND_UP(reftable_index + 1,
2806 s->cluster_size / sizeof(uint64_t));
2807 if (new_reftable_size > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) {
2808 error_setg(errp,
2809 "This operation would make the refcount table grow "
2810 "beyond the maximum size supported by QEMU, aborting");
2811 return -ENOTSUP;
2814 new_reftable = g_try_realloc(*reftable, new_reftable_size *
2815 sizeof(uint64_t));
2816 if (!new_reftable) {
2817 error_setg(errp, "Failed to increase reftable buffer size");
2818 return -ENOMEM;
2821 memset(new_reftable + *reftable_size, 0,
2822 (new_reftable_size - *reftable_size) * sizeof(uint64_t));
2824 *reftable = new_reftable;
2825 *reftable_size = new_reftable_size;
2828 if (!refblock_empty && !(*reftable)[reftable_index]) {
2829 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2830 if (offset < 0) {
2831 error_setg_errno(errp, -offset, "Failed to allocate refblock");
2832 return offset;
2834 (*reftable)[reftable_index] = offset;
2835 *allocated = true;
2838 return 0;
2842 * This "operation" for walk_over_reftable() writes the refblock to disk at the
2843 * offset specified by the new reftable's entry. It does not modify the new
2844 * reftable or change any refcounts.
2846 static int flush_refblock(BlockDriverState *bs, uint64_t **reftable,
2847 uint64_t reftable_index, uint64_t *reftable_size,
2848 void *refblock, bool refblock_empty, bool *allocated,
2849 Error **errp)
2851 BDRVQcow2State *s = bs->opaque;
2852 int64_t offset;
2853 int ret;
2855 if (reftable_index < *reftable_size && (*reftable)[reftable_index]) {
2856 offset = (*reftable)[reftable_index];
2858 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size);
2859 if (ret < 0) {
2860 error_setg_errno(errp, -ret, "Overlap check failed");
2861 return ret;
2864 ret = bdrv_pwrite(bs->file, offset, refblock, s->cluster_size);
2865 if (ret < 0) {
2866 error_setg_errno(errp, -ret, "Failed to write refblock");
2867 return ret;
2869 } else {
2870 assert(refblock_empty);
2873 return 0;
2877 * This function walks over the existing reftable and every referenced refblock;
2878 * if @new_set_refcount is non-NULL, it is called for every refcount entry to
2879 * create an equal new entry in the passed @new_refblock. Once that
2880 * @new_refblock is completely filled, @operation will be called.
2882 * @status_cb and @cb_opaque are used for the amend operation's status callback.
2883 * @index is the index of the walk_over_reftable() calls and @total is the total
2884 * number of walk_over_reftable() calls per amend operation. Both are used for
2885 * calculating the parameters for the status callback.
2887 * @allocated is set to true if a new cluster has been allocated.
2889 static int walk_over_reftable(BlockDriverState *bs, uint64_t **new_reftable,
2890 uint64_t *new_reftable_index,
2891 uint64_t *new_reftable_size,
2892 void *new_refblock, int new_refblock_size,
2893 int new_refcount_bits,
2894 RefblockFinishOp *operation, bool *allocated,
2895 Qcow2SetRefcountFunc *new_set_refcount,
2896 BlockDriverAmendStatusCB *status_cb,
2897 void *cb_opaque, int index, int total,
2898 Error **errp)
2900 BDRVQcow2State *s = bs->opaque;
2901 uint64_t reftable_index;
2902 bool new_refblock_empty = true;
2903 int refblock_index;
2904 int new_refblock_index = 0;
2905 int ret;
2907 for (reftable_index = 0; reftable_index < s->refcount_table_size;
2908 reftable_index++)
2910 uint64_t refblock_offset = s->refcount_table[reftable_index]
2911 & REFT_OFFSET_MASK;
2913 status_cb(bs, (uint64_t)index * s->refcount_table_size + reftable_index,
2914 (uint64_t)total * s->refcount_table_size, cb_opaque);
2916 if (refblock_offset) {
2917 void *refblock;
2919 if (offset_into_cluster(s, refblock_offset)) {
2920 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#"
2921 PRIx64 " unaligned (reftable index: %#"
2922 PRIx64 ")", refblock_offset,
2923 reftable_index);
2924 error_setg(errp,
2925 "Image is corrupt (unaligned refblock offset)");
2926 return -EIO;
2929 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offset,
2930 &refblock);
2931 if (ret < 0) {
2932 error_setg_errno(errp, -ret, "Failed to retrieve refblock");
2933 return ret;
2936 for (refblock_index = 0; refblock_index < s->refcount_block_size;
2937 refblock_index++)
2939 uint64_t refcount;
2941 if (new_refblock_index >= new_refblock_size) {
2942 /* new_refblock is now complete */
2943 ret = operation(bs, new_reftable, *new_reftable_index,
2944 new_reftable_size, new_refblock,
2945 new_refblock_empty, allocated, errp);
2946 if (ret < 0) {
2947 qcow2_cache_put(s->refcount_block_cache, &refblock);
2948 return ret;
2951 (*new_reftable_index)++;
2952 new_refblock_index = 0;
2953 new_refblock_empty = true;
2956 refcount = s->get_refcount(refblock, refblock_index);
2957 if (new_refcount_bits < 64 && refcount >> new_refcount_bits) {
2958 uint64_t offset;
2960 qcow2_cache_put(s->refcount_block_cache, &refblock);
2962 offset = ((reftable_index << s->refcount_block_bits)
2963 + refblock_index) << s->cluster_bits;
2965 error_setg(errp, "Cannot decrease refcount entry width to "
2966 "%i bits: Cluster at offset %#" PRIx64 " has a "
2967 "refcount of %" PRIu64, new_refcount_bits,
2968 offset, refcount);
2969 return -EINVAL;
2972 if (new_set_refcount) {
2973 new_set_refcount(new_refblock, new_refblock_index++,
2974 refcount);
2975 } else {
2976 new_refblock_index++;
2978 new_refblock_empty = new_refblock_empty && refcount == 0;
2981 qcow2_cache_put(s->refcount_block_cache, &refblock);
2982 } else {
2983 /* No refblock means every refcount is 0 */
2984 for (refblock_index = 0; refblock_index < s->refcount_block_size;
2985 refblock_index++)
2987 if (new_refblock_index >= new_refblock_size) {
2988 /* new_refblock is now complete */
2989 ret = operation(bs, new_reftable, *new_reftable_index,
2990 new_reftable_size, new_refblock,
2991 new_refblock_empty, allocated, errp);
2992 if (ret < 0) {
2993 return ret;
2996 (*new_reftable_index)++;
2997 new_refblock_index = 0;
2998 new_refblock_empty = true;
3001 if (new_set_refcount) {
3002 new_set_refcount(new_refblock, new_refblock_index++, 0);
3003 } else {
3004 new_refblock_index++;
3010 if (new_refblock_index > 0) {
3011 /* Complete the potentially existing partially filled final refblock */
3012 if (new_set_refcount) {
3013 for (; new_refblock_index < new_refblock_size;
3014 new_refblock_index++)
3016 new_set_refcount(new_refblock, new_refblock_index, 0);
3020 ret = operation(bs, new_reftable, *new_reftable_index,
3021 new_reftable_size, new_refblock, new_refblock_empty,
3022 allocated, errp);
3023 if (ret < 0) {
3024 return ret;
3027 (*new_reftable_index)++;
3030 status_cb(bs, (uint64_t)(index + 1) * s->refcount_table_size,
3031 (uint64_t)total * s->refcount_table_size, cb_opaque);
3033 return 0;
3036 int qcow2_change_refcount_order(BlockDriverState *bs, int refcount_order,
3037 BlockDriverAmendStatusCB *status_cb,
3038 void *cb_opaque, Error **errp)
3040 BDRVQcow2State *s = bs->opaque;
3041 Qcow2GetRefcountFunc *new_get_refcount;
3042 Qcow2SetRefcountFunc *new_set_refcount;
3043 void *new_refblock = qemu_blockalign(bs->file->bs, s->cluster_size);
3044 uint64_t *new_reftable = NULL, new_reftable_size = 0;
3045 uint64_t *old_reftable, old_reftable_size, old_reftable_offset;
3046 uint64_t new_reftable_index = 0;
3047 uint64_t i;
3048 int64_t new_reftable_offset = 0, allocated_reftable_size = 0;
3049 int new_refblock_size, new_refcount_bits = 1 << refcount_order;
3050 int old_refcount_order;
3051 int walk_index = 0;
3052 int ret;
3053 bool new_allocation;
3055 assert(s->qcow_version >= 3);
3056 assert(refcount_order >= 0 && refcount_order <= 6);
3058 /* see qcow2_open() */
3059 new_refblock_size = 1 << (s->cluster_bits - (refcount_order - 3));
3061 new_get_refcount = get_refcount_funcs[refcount_order];
3062 new_set_refcount = set_refcount_funcs[refcount_order];
3065 do {
3066 int total_walks;
3068 new_allocation = false;
3070 /* At least we have to do this walk and the one which writes the
3071 * refblocks; also, at least we have to do this loop here at least
3072 * twice (normally), first to do the allocations, and second to
3073 * determine that everything is correctly allocated, this then makes
3074 * three walks in total */
3075 total_walks = MAX(walk_index + 2, 3);
3077 /* First, allocate the structures so they are present in the refcount
3078 * structures */
3079 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3080 &new_reftable_size, NULL, new_refblock_size,
3081 new_refcount_bits, &alloc_refblock,
3082 &new_allocation, NULL, status_cb, cb_opaque,
3083 walk_index++, total_walks, errp);
3084 if (ret < 0) {
3085 goto done;
3088 new_reftable_index = 0;
3090 if (new_allocation) {
3091 if (new_reftable_offset) {
3092 qcow2_free_clusters(bs, new_reftable_offset,
3093 allocated_reftable_size * sizeof(uint64_t),
3094 QCOW2_DISCARD_NEVER);
3097 new_reftable_offset = qcow2_alloc_clusters(bs, new_reftable_size *
3098 sizeof(uint64_t));
3099 if (new_reftable_offset < 0) {
3100 error_setg_errno(errp, -new_reftable_offset,
3101 "Failed to allocate the new reftable");
3102 ret = new_reftable_offset;
3103 goto done;
3105 allocated_reftable_size = new_reftable_size;
3107 } while (new_allocation);
3109 /* Second, write the new refblocks */
3110 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index,
3111 &new_reftable_size, new_refblock,
3112 new_refblock_size, new_refcount_bits,
3113 &flush_refblock, &new_allocation, new_set_refcount,
3114 status_cb, cb_opaque, walk_index, walk_index + 1,
3115 errp);
3116 if (ret < 0) {
3117 goto done;
3119 assert(!new_allocation);
3122 /* Write the new reftable */
3123 ret = qcow2_pre_write_overlap_check(bs, 0, new_reftable_offset,
3124 new_reftable_size * sizeof(uint64_t));
3125 if (ret < 0) {
3126 error_setg_errno(errp, -ret, "Overlap check failed");
3127 goto done;
3130 for (i = 0; i < new_reftable_size; i++) {
3131 cpu_to_be64s(&new_reftable[i]);
3134 ret = bdrv_pwrite(bs->file, new_reftable_offset, new_reftable,
3135 new_reftable_size * sizeof(uint64_t));
3137 for (i = 0; i < new_reftable_size; i++) {
3138 be64_to_cpus(&new_reftable[i]);
3141 if (ret < 0) {
3142 error_setg_errno(errp, -ret, "Failed to write the new reftable");
3143 goto done;
3147 /* Empty the refcount cache */
3148 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
3149 if (ret < 0) {
3150 error_setg_errno(errp, -ret, "Failed to flush the refblock cache");
3151 goto done;
3154 /* Update the image header to point to the new reftable; this only updates
3155 * the fields which are relevant to qcow2_update_header(); other fields
3156 * such as s->refcount_table or s->refcount_bits stay stale for now
3157 * (because we have to restore everything if qcow2_update_header() fails) */
3158 old_refcount_order = s->refcount_order;
3159 old_reftable_size = s->refcount_table_size;
3160 old_reftable_offset = s->refcount_table_offset;
3162 s->refcount_order = refcount_order;
3163 s->refcount_table_size = new_reftable_size;
3164 s->refcount_table_offset = new_reftable_offset;
3166 ret = qcow2_update_header(bs);
3167 if (ret < 0) {
3168 s->refcount_order = old_refcount_order;
3169 s->refcount_table_size = old_reftable_size;
3170 s->refcount_table_offset = old_reftable_offset;
3171 error_setg_errno(errp, -ret, "Failed to update the qcow2 header");
3172 goto done;
3175 /* Now update the rest of the in-memory information */
3176 old_reftable = s->refcount_table;
3177 s->refcount_table = new_reftable;
3178 update_max_refcount_table_index(s);
3180 s->refcount_bits = 1 << refcount_order;
3181 s->refcount_max = UINT64_C(1) << (s->refcount_bits - 1);
3182 s->refcount_max += s->refcount_max - 1;
3184 s->refcount_block_bits = s->cluster_bits - (refcount_order - 3);
3185 s->refcount_block_size = 1 << s->refcount_block_bits;
3187 s->get_refcount = new_get_refcount;
3188 s->set_refcount = new_set_refcount;
3190 /* For cleaning up all old refblocks and the old reftable below the "done"
3191 * label */
3192 new_reftable = old_reftable;
3193 new_reftable_size = old_reftable_size;
3194 new_reftable_offset = old_reftable_offset;
3196 done:
3197 if (new_reftable) {
3198 /* On success, new_reftable actually points to the old reftable (and
3199 * new_reftable_size is the old reftable's size); but that is just
3200 * fine */
3201 for (i = 0; i < new_reftable_size; i++) {
3202 uint64_t offset = new_reftable[i] & REFT_OFFSET_MASK;
3203 if (offset) {
3204 qcow2_free_clusters(bs, offset, s->cluster_size,
3205 QCOW2_DISCARD_OTHER);
3208 g_free(new_reftable);
3210 if (new_reftable_offset > 0) {
3211 qcow2_free_clusters(bs, new_reftable_offset,
3212 new_reftable_size * sizeof(uint64_t),
3213 QCOW2_DISCARD_OTHER);
3217 qemu_vfree(new_refblock);
3218 return ret;
3221 static int64_t get_refblock_offset(BlockDriverState *bs, uint64_t offset)
3223 BDRVQcow2State *s = bs->opaque;
3224 uint32_t index = offset_to_reftable_index(s, offset);
3225 int64_t covering_refblock_offset = 0;
3227 if (index < s->refcount_table_size) {
3228 covering_refblock_offset = s->refcount_table[index] & REFT_OFFSET_MASK;
3230 if (!covering_refblock_offset) {
3231 qcow2_signal_corruption(bs, true, -1, -1, "Refblock at %#" PRIx64 " is "
3232 "not covered by the refcount structures",
3233 offset);
3234 return -EIO;
3237 return covering_refblock_offset;
3240 static int qcow2_discard_refcount_block(BlockDriverState *bs,
3241 uint64_t discard_block_offs)
3243 BDRVQcow2State *s = bs->opaque;
3244 int64_t refblock_offs;
3245 uint64_t cluster_index = discard_block_offs >> s->cluster_bits;
3246 uint32_t block_index = cluster_index & (s->refcount_block_size - 1);
3247 void *refblock;
3248 int ret;
3250 refblock_offs = get_refblock_offset(bs, discard_block_offs);
3251 if (refblock_offs < 0) {
3252 return refblock_offs;
3255 assert(discard_block_offs != 0);
3257 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3258 &refblock);
3259 if (ret < 0) {
3260 return ret;
3263 if (s->get_refcount(refblock, block_index) != 1) {
3264 qcow2_signal_corruption(bs, true, -1, -1, "Invalid refcount:"
3265 " refblock offset %#" PRIx64
3266 ", reftable index %u"
3267 ", block offset %#" PRIx64
3268 ", refcount %#" PRIx64,
3269 refblock_offs,
3270 offset_to_reftable_index(s, discard_block_offs),
3271 discard_block_offs,
3272 s->get_refcount(refblock, block_index));
3273 qcow2_cache_put(s->refcount_block_cache, &refblock);
3274 return -EINVAL;
3276 s->set_refcount(refblock, block_index, 0);
3278 qcow2_cache_entry_mark_dirty(s->refcount_block_cache, refblock);
3280 qcow2_cache_put(s->refcount_block_cache, &refblock);
3282 if (cluster_index < s->free_cluster_index) {
3283 s->free_cluster_index = cluster_index;
3286 refblock = qcow2_cache_is_table_offset(s->refcount_block_cache,
3287 discard_block_offs);
3288 if (refblock) {
3289 /* discard refblock from the cache if refblock is cached */
3290 qcow2_cache_discard(s->refcount_block_cache, refblock);
3292 update_refcount_discard(bs, discard_block_offs, s->cluster_size);
3294 return 0;
3297 int qcow2_shrink_reftable(BlockDriverState *bs)
3299 BDRVQcow2State *s = bs->opaque;
3300 uint64_t *reftable_tmp =
3301 g_malloc(s->refcount_table_size * sizeof(uint64_t));
3302 int i, ret;
3304 for (i = 0; i < s->refcount_table_size; i++) {
3305 int64_t refblock_offs = s->refcount_table[i] & REFT_OFFSET_MASK;
3306 void *refblock;
3307 bool unused_block;
3309 if (refblock_offs == 0) {
3310 reftable_tmp[i] = 0;
3311 continue;
3313 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs,
3314 &refblock);
3315 if (ret < 0) {
3316 goto out;
3319 /* the refblock has own reference */
3320 if (i == offset_to_reftable_index(s, refblock_offs)) {
3321 uint64_t block_index = (refblock_offs >> s->cluster_bits) &
3322 (s->refcount_block_size - 1);
3323 uint64_t refcount = s->get_refcount(refblock, block_index);
3325 s->set_refcount(refblock, block_index, 0);
3327 unused_block = buffer_is_zero(refblock, s->cluster_size);
3329 s->set_refcount(refblock, block_index, refcount);
3330 } else {
3331 unused_block = buffer_is_zero(refblock, s->cluster_size);
3333 qcow2_cache_put(s->refcount_block_cache, &refblock);
3335 reftable_tmp[i] = unused_block ? 0 : cpu_to_be64(s->refcount_table[i]);
3338 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset, reftable_tmp,
3339 s->refcount_table_size * sizeof(uint64_t));
3341 * If the write in the reftable failed the image may contain a partially
3342 * overwritten reftable. In this case it would be better to clear the
3343 * reftable in memory to avoid possible image corruption.
3345 for (i = 0; i < s->refcount_table_size; i++) {
3346 if (s->refcount_table[i] && !reftable_tmp[i]) {
3347 if (ret == 0) {
3348 ret = qcow2_discard_refcount_block(bs, s->refcount_table[i] &
3349 REFT_OFFSET_MASK);
3351 s->refcount_table[i] = 0;
3355 if (!s->cache_discards) {
3356 qcow2_process_discards(bs, ret);
3359 out:
3360 g_free(reftable_tmp);
3361 return ret;
3364 int64_t qcow2_get_last_cluster(BlockDriverState *bs, int64_t size)
3366 BDRVQcow2State *s = bs->opaque;
3367 int64_t i;
3369 for (i = size_to_clusters(s, size) - 1; i >= 0; i--) {
3370 uint64_t refcount;
3371 int ret = qcow2_get_refcount(bs, i, &refcount);
3372 if (ret < 0) {
3373 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n",
3374 i, strerror(-ret));
3375 return ret;
3377 if (refcount > 0) {
3378 return i;
3381 qcow2_signal_corruption(bs, true, -1, -1,
3382 "There are no references in the refcount table.");
3383 return -EIO;