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
25 #include "qemu/osdep.h"
28 #include "qapi/error.h"
29 #include "qemu-common.h"
30 #include "block/block_int.h"
31 #include "block/qcow2.h"
32 #include "qemu/bswap.h"
35 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
38 BDRVQcow2State *s = bs->opaque;
39 int new_l1_size2, ret, i;
40 uint64_t *new_l1_table;
41 int64_t old_l1_table_offset, old_l1_size;
42 int64_t new_l1_table_offset, new_l1_size;
45 if (min_size <= s->l1_size)
48 /* Do a sanity check on min_size before trying to calculate new_l1_size
49 * (this prevents overflows during the while loop for the calculation of
51 if (min_size > INT_MAX / sizeof(uint64_t)) {
56 new_l1_size = min_size;
58 /* Bump size up to reduce the number of times we have to grow */
59 new_l1_size = s->l1_size;
60 if (new_l1_size == 0) {
63 while (min_size > new_l1_size) {
64 new_l1_size = (new_l1_size * 3 + 1) / 2;
68 QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX);
69 if (new_l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) {
74 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
75 s->l1_size, new_l1_size);
78 new_l1_size2 = sizeof(uint64_t) * new_l1_size;
79 new_l1_table = qemu_try_blockalign(bs->file->bs,
80 align_offset(new_l1_size2, 512));
81 if (new_l1_table == NULL) {
84 memset(new_l1_table, 0, align_offset(new_l1_size2, 512));
86 memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
88 /* write new table (align to cluster) */
89 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
90 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
91 if (new_l1_table_offset < 0) {
92 qemu_vfree(new_l1_table);
93 return new_l1_table_offset;
96 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
101 /* the L1 position has not yet been updated, so these clusters must
102 * indeed be completely free */
103 ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
109 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
110 for(i = 0; i < s->l1_size; i++)
111 new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
112 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset,
113 new_l1_table, new_l1_size2);
116 for(i = 0; i < s->l1_size; i++)
117 new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
120 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
121 stl_be_p(data, new_l1_size);
122 stq_be_p(data + 4, new_l1_table_offset);
123 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size),
128 qemu_vfree(s->l1_table);
129 old_l1_table_offset = s->l1_table_offset;
130 s->l1_table_offset = new_l1_table_offset;
131 s->l1_table = new_l1_table;
132 old_l1_size = s->l1_size;
133 s->l1_size = new_l1_size;
134 qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
135 QCOW2_DISCARD_OTHER);
138 qemu_vfree(new_l1_table);
139 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
140 QCOW2_DISCARD_OTHER);
147 * Loads a L2 table into memory. If the table is in the cache, the cache
148 * is used; otherwise the L2 table is loaded from the image file.
150 * Returns a pointer to the L2 table on success, or NULL if the read from
151 * the image file failed.
154 static int l2_load(BlockDriverState *bs, uint64_t l2_offset,
157 BDRVQcow2State *s = bs->opaque;
159 return qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
164 * Writes one sector of the L1 table to the disk (can't update single entries
165 * and we really don't want bdrv_pread to perform a read-modify-write)
167 #define L1_ENTRIES_PER_SECTOR (512 / 8)
168 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
170 BDRVQcow2State *s = bs->opaque;
171 uint64_t buf[L1_ENTRIES_PER_SECTOR] = { 0 };
175 l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
176 for (i = 0; i < L1_ENTRIES_PER_SECTOR && l1_start_index + i < s->l1_size;
179 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
182 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
183 s->l1_table_offset + 8 * l1_start_index, sizeof(buf));
188 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
189 ret = bdrv_pwrite_sync(bs->file,
190 s->l1_table_offset + 8 * l1_start_index,
202 * Allocate a new l2 entry in the file. If l1_index points to an already
203 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
204 * table) copy the contents of the old L2 table into the newly allocated one.
205 * Otherwise the new table is initialized with zeros.
209 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table)
211 BDRVQcow2State *s = bs->opaque;
212 uint64_t old_l2_offset;
213 uint64_t *l2_table = NULL;
217 old_l2_offset = s->l1_table[l1_index];
219 trace_qcow2_l2_allocate(bs, l1_index);
221 /* allocate a new l2 entry */
223 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
229 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
234 /* allocate a new entry in the l2 cache */
236 trace_qcow2_l2_allocate_get_empty(bs, l1_index);
237 ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table);
244 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
245 /* if there was no old l2 table, clear the new table */
246 memset(l2_table, 0, s->l2_size * sizeof(uint64_t));
250 /* if there was an old l2 table, read it from the disk */
251 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
252 ret = qcow2_cache_get(bs, s->l2_table_cache,
253 old_l2_offset & L1E_OFFSET_MASK,
254 (void**) &old_table);
259 memcpy(l2_table, old_table, s->cluster_size);
261 qcow2_cache_put(bs, s->l2_table_cache, (void **) &old_table);
264 /* write the l2 table to the file */
265 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
267 trace_qcow2_l2_allocate_write_l2(bs, l1_index);
268 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
269 ret = qcow2_cache_flush(bs, s->l2_table_cache);
274 /* update the L1 entry */
275 trace_qcow2_l2_allocate_write_l1(bs, l1_index);
276 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
277 ret = qcow2_write_l1_entry(bs, l1_index);
283 trace_qcow2_l2_allocate_done(bs, l1_index, 0);
287 trace_qcow2_l2_allocate_done(bs, l1_index, ret);
288 if (l2_table != NULL) {
289 qcow2_cache_put(bs, s->l2_table_cache, (void**) table);
291 s->l1_table[l1_index] = old_l2_offset;
293 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
294 QCOW2_DISCARD_ALWAYS);
300 * Checks how many clusters in a given L2 table are contiguous in the image
301 * file. As soon as one of the flags in the bitmask stop_flags changes compared
302 * to the first cluster, the search is stopped and the cluster is not counted
303 * as contiguous. (This allows it, for example, to stop at the first compressed
304 * cluster which may require a different handling)
306 static int count_contiguous_clusters(int nb_clusters, int cluster_size,
307 uint64_t *l2_table, uint64_t stop_flags)
310 uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
311 uint64_t first_entry = be64_to_cpu(l2_table[0]);
312 uint64_t offset = first_entry & mask;
317 assert(qcow2_get_cluster_type(first_entry) == QCOW2_CLUSTER_NORMAL);
319 for (i = 0; i < nb_clusters; i++) {
320 uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask;
321 if (offset + (uint64_t) i * cluster_size != l2_entry) {
329 static int count_contiguous_clusters_by_type(int nb_clusters,
335 for (i = 0; i < nb_clusters; i++) {
336 int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i]));
338 if (type != wanted_type) {
346 /* The crypt function is compatible with the linux cryptoloop
347 algorithm for < 4 GB images. NOTE: out_buf == in_buf is
349 int qcow2_encrypt_sectors(BDRVQcow2State *s, int64_t sector_num,
350 uint8_t *out_buf, const uint8_t *in_buf,
351 int nb_sectors, bool enc,
361 for(i = 0; i < nb_sectors; i++) {
362 ivec.ll[0] = cpu_to_le64(sector_num);
364 if (qcrypto_cipher_setiv(s->cipher,
365 ivec.b, G_N_ELEMENTS(ivec.b),
370 ret = qcrypto_cipher_encrypt(s->cipher,
376 ret = qcrypto_cipher_decrypt(s->cipher,
392 static int coroutine_fn do_perform_cow(BlockDriverState *bs,
393 uint64_t src_cluster_offset,
394 uint64_t cluster_offset,
395 int offset_in_cluster,
398 BDRVQcow2State *s = bs->opaque;
404 iov.iov_base = qemu_try_blockalign(bs, iov.iov_len);
405 if (iov.iov_base == NULL) {
409 qemu_iovec_init_external(&qiov, &iov, 1);
411 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
418 /* Call .bdrv_co_readv() directly instead of using the public block-layer
419 * interface. This avoids double I/O throttling and request tracking,
420 * which can lead to deadlock when block layer copy-on-read is enabled.
422 ret = bs->drv->bdrv_co_preadv(bs, src_cluster_offset + offset_in_cluster,
430 int64_t sector = (cluster_offset + offset_in_cluster)
433 assert((offset_in_cluster & ~BDRV_SECTOR_MASK) == 0);
434 assert((bytes & ~BDRV_SECTOR_MASK) == 0);
435 if (qcow2_encrypt_sectors(s, sector, iov.iov_base, iov.iov_base,
436 bytes >> BDRV_SECTOR_BITS, true, &err) < 0) {
443 ret = qcow2_pre_write_overlap_check(bs, 0,
444 cluster_offset + offset_in_cluster, bytes);
449 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
450 ret = bdrv_co_pwritev(bs->file, cluster_offset + offset_in_cluster,
458 qemu_vfree(iov.iov_base);
466 * For a given offset of the virtual disk, find the cluster type and offset in
467 * the qcow2 file. The offset is stored in *cluster_offset.
469 * On entry, *bytes is the maximum number of contiguous bytes starting at
470 * offset that we are interested in.
472 * On exit, *bytes is the number of bytes starting at offset that have the same
473 * cluster type and (if applicable) are stored contiguously in the image file.
474 * Compressed clusters are always returned one by one.
476 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
479 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
480 unsigned int *bytes, uint64_t *cluster_offset)
482 BDRVQcow2State *s = bs->opaque;
483 unsigned int l2_index;
484 uint64_t l1_index, l2_offset, *l2_table;
486 unsigned int offset_in_cluster;
487 uint64_t bytes_available, bytes_needed, nb_clusters;
490 offset_in_cluster = offset_into_cluster(s, offset);
491 bytes_needed = (uint64_t) *bytes + offset_in_cluster;
493 l1_bits = s->l2_bits + s->cluster_bits;
495 /* compute how many bytes there are between the start of the cluster
496 * containing offset and the end of the l1 entry */
497 bytes_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1))
500 if (bytes_needed > bytes_available) {
501 bytes_needed = bytes_available;
506 /* seek to the l2 offset in the l1 table */
508 l1_index = offset >> l1_bits;
509 if (l1_index >= s->l1_size) {
510 ret = QCOW2_CLUSTER_UNALLOCATED;
514 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
516 ret = QCOW2_CLUSTER_UNALLOCATED;
520 if (offset_into_cluster(s, l2_offset)) {
521 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
522 " unaligned (L1 index: %#" PRIx64 ")",
523 l2_offset, l1_index);
527 /* load the l2 table in memory */
529 ret = l2_load(bs, l2_offset, &l2_table);
534 /* find the cluster offset for the given disk offset */
536 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
537 *cluster_offset = be64_to_cpu(l2_table[l2_index]);
539 nb_clusters = size_to_clusters(s, bytes_needed);
540 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
541 * integers; the minimum cluster size is 512, so this assertion is always
543 assert(nb_clusters <= INT_MAX);
545 ret = qcow2_get_cluster_type(*cluster_offset);
547 case QCOW2_CLUSTER_COMPRESSED:
548 /* Compressed clusters can only be processed one by one */
550 *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
552 case QCOW2_CLUSTER_ZERO:
553 if (s->qcow_version < 3) {
554 qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
555 " in pre-v3 image (L2 offset: %#" PRIx64
556 ", L2 index: %#x)", l2_offset, l2_index);
560 c = count_contiguous_clusters_by_type(nb_clusters, &l2_table[l2_index],
564 case QCOW2_CLUSTER_UNALLOCATED:
565 /* how many empty clusters ? */
566 c = count_contiguous_clusters_by_type(nb_clusters, &l2_table[l2_index],
567 QCOW2_CLUSTER_UNALLOCATED);
570 case QCOW2_CLUSTER_NORMAL:
571 /* how many allocated clusters ? */
572 c = count_contiguous_clusters(nb_clusters, s->cluster_size,
573 &l2_table[l2_index], QCOW_OFLAG_ZERO);
574 *cluster_offset &= L2E_OFFSET_MASK;
575 if (offset_into_cluster(s, *cluster_offset)) {
576 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset %#"
577 PRIx64 " unaligned (L2 offset: %#" PRIx64
578 ", L2 index: %#x)", *cluster_offset,
579 l2_offset, l2_index);
588 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
590 bytes_available = (int64_t)c * s->cluster_size;
593 if (bytes_available > bytes_needed) {
594 bytes_available = bytes_needed;
597 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
598 * subtracting offset_in_cluster will therefore definitely yield something
599 * not exceeding UINT_MAX */
600 assert(bytes_available - offset_in_cluster <= UINT_MAX);
601 *bytes = bytes_available - offset_in_cluster;
606 qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
613 * for a given disk offset, load (and allocate if needed)
616 * the l2 table offset in the qcow2 file and the cluster index
617 * in the l2 table are given to the caller.
619 * Returns 0 on success, -errno in failure case
621 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
622 uint64_t **new_l2_table,
625 BDRVQcow2State *s = bs->opaque;
626 unsigned int l2_index;
627 uint64_t l1_index, l2_offset;
628 uint64_t *l2_table = NULL;
631 /* seek to the l2 offset in the l1 table */
633 l1_index = offset >> (s->l2_bits + s->cluster_bits);
634 if (l1_index >= s->l1_size) {
635 ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
641 assert(l1_index < s->l1_size);
642 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
643 if (offset_into_cluster(s, l2_offset)) {
644 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
645 " unaligned (L1 index: %#" PRIx64 ")",
646 l2_offset, l1_index);
650 /* seek the l2 table of the given l2 offset */
652 if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) {
653 /* load the l2 table in memory */
654 ret = l2_load(bs, l2_offset, &l2_table);
659 /* First allocate a new L2 table (and do COW if needed) */
660 ret = l2_allocate(bs, l1_index, &l2_table);
665 /* Then decrease the refcount of the old table */
667 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
668 QCOW2_DISCARD_OTHER);
672 /* find the cluster offset for the given disk offset */
674 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
676 *new_l2_table = l2_table;
677 *new_l2_index = l2_index;
683 * alloc_compressed_cluster_offset
685 * For a given offset of the disk image, return cluster offset in
688 * If the offset is not found, allocate a new compressed cluster.
690 * Return the cluster offset if successful,
691 * Return 0, otherwise.
695 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
699 BDRVQcow2State *s = bs->opaque;
702 int64_t cluster_offset;
705 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
710 /* Compression can't overwrite anything. Fail if the cluster was already
712 cluster_offset = be64_to_cpu(l2_table[l2_index]);
713 if (cluster_offset & L2E_OFFSET_MASK) {
714 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
718 cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
719 if (cluster_offset < 0) {
720 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
724 nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) -
725 (cluster_offset >> 9);
727 cluster_offset |= QCOW_OFLAG_COMPRESSED |
728 ((uint64_t)nb_csectors << s->csize_shift);
730 /* update L2 table */
732 /* compressed clusters never have the copied flag */
734 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
735 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
736 l2_table[l2_index] = cpu_to_be64(cluster_offset);
737 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
739 return cluster_offset;
742 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r)
744 BDRVQcow2State *s = bs->opaque;
747 if (r->nb_bytes == 0) {
751 qemu_co_mutex_unlock(&s->lock);
752 ret = do_perform_cow(bs, m->offset, m->alloc_offset, r->offset, r->nb_bytes);
753 qemu_co_mutex_lock(&s->lock);
760 * Before we update the L2 table to actually point to the new cluster, we
761 * need to be sure that the refcounts have been increased and COW was
764 qcow2_cache_depends_on_flush(s->l2_table_cache);
769 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
771 BDRVQcow2State *s = bs->opaque;
772 int i, j = 0, l2_index, ret;
773 uint64_t *old_cluster, *l2_table;
774 uint64_t cluster_offset = m->alloc_offset;
776 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
777 assert(m->nb_clusters > 0);
779 old_cluster = g_try_new(uint64_t, m->nb_clusters);
780 if (old_cluster == NULL) {
785 /* copy content of unmodified sectors */
786 ret = perform_cow(bs, m, &m->cow_start);
791 ret = perform_cow(bs, m, &m->cow_end);
796 /* Update L2 table. */
797 if (s->use_lazy_refcounts) {
798 qcow2_mark_dirty(bs);
800 if (qcow2_need_accurate_refcounts(s)) {
801 qcow2_cache_set_dependency(bs, s->l2_table_cache,
802 s->refcount_block_cache);
805 ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index);
809 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
811 assert(l2_index + m->nb_clusters <= s->l2_size);
812 for (i = 0; i < m->nb_clusters; i++) {
813 /* if two concurrent writes happen to the same unallocated cluster
814 * each write allocates separate cluster and writes data concurrently.
815 * The first one to complete updates l2 table with pointer to its
816 * cluster the second one has to do RMW (which is done above by
817 * perform_cow()), update l2 table with its cluster pointer and free
818 * old cluster. This is what this loop does */
819 if (l2_table[l2_index + i] != 0) {
820 old_cluster[j++] = l2_table[l2_index + i];
823 l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
824 (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
828 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
831 * If this was a COW, we need to decrease the refcount of the old cluster.
833 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
834 * clusters), the next write will reuse them anyway.
837 for (i = 0; i < j; i++) {
838 qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
839 QCOW2_DISCARD_NEVER);
850 * Returns the number of contiguous clusters that can be used for an allocating
851 * write, but require COW to be performed (this includes yet unallocated space,
852 * which must copy from the backing file)
854 static int count_cow_clusters(BDRVQcow2State *s, int nb_clusters,
855 uint64_t *l2_table, int l2_index)
859 for (i = 0; i < nb_clusters; i++) {
860 uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]);
861 int cluster_type = qcow2_get_cluster_type(l2_entry);
863 switch(cluster_type) {
864 case QCOW2_CLUSTER_NORMAL:
865 if (l2_entry & QCOW_OFLAG_COPIED) {
869 case QCOW2_CLUSTER_UNALLOCATED:
870 case QCOW2_CLUSTER_COMPRESSED:
871 case QCOW2_CLUSTER_ZERO:
879 assert(i <= nb_clusters);
884 * Check if there already is an AIO write request in flight which allocates
885 * the same cluster. In this case we need to wait until the previous
886 * request has completed and updated the L2 table accordingly.
889 * 0 if there was no dependency. *cur_bytes indicates the number of
890 * bytes from guest_offset that can be read before the next
891 * dependency must be processed (or the request is complete)
893 * -EAGAIN if we had to wait for another request, previously gathered
894 * information on cluster allocation may be invalid now. The caller
895 * must start over anyway, so consider *cur_bytes undefined.
897 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
898 uint64_t *cur_bytes, QCowL2Meta **m)
900 BDRVQcow2State *s = bs->opaque;
901 QCowL2Meta *old_alloc;
902 uint64_t bytes = *cur_bytes;
904 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
906 uint64_t start = guest_offset;
907 uint64_t end = start + bytes;
908 uint64_t old_start = l2meta_cow_start(old_alloc);
909 uint64_t old_end = l2meta_cow_end(old_alloc);
911 if (end <= old_start || start >= old_end) {
912 /* No intersection */
914 if (start < old_start) {
915 /* Stop at the start of a running allocation */
916 bytes = old_start - start;
921 /* Stop if already an l2meta exists. After yielding, it wouldn't
922 * be valid any more, so we'd have to clean up the old L2Metas
923 * and deal with requests depending on them before starting to
924 * gather new ones. Not worth the trouble. */
925 if (bytes == 0 && *m) {
931 /* Wait for the dependency to complete. We need to recheck
932 * the free/allocated clusters when we continue. */
933 qemu_co_mutex_unlock(&s->lock);
934 qemu_co_queue_wait(&old_alloc->dependent_requests);
935 qemu_co_mutex_lock(&s->lock);
941 /* Make sure that existing clusters and new allocations are only used up to
942 * the next dependency if we shortened the request above */
949 * Checks how many already allocated clusters that don't require a copy on
950 * write there are at the given guest_offset (up to *bytes). If
951 * *host_offset is not zero, only physically contiguous clusters beginning at
952 * this host offset are counted.
954 * Note that guest_offset may not be cluster aligned. In this case, the
955 * returned *host_offset points to exact byte referenced by guest_offset and
956 * therefore isn't cluster aligned as well.
959 * 0: if no allocated clusters are available at the given offset.
960 * *bytes is normally unchanged. It is set to 0 if the cluster
961 * is allocated and doesn't need COW, but doesn't have the right
964 * 1: if allocated clusters that don't require a COW are available at
965 * the requested offset. *bytes may have decreased and describes
966 * the length of the area that can be written to.
968 * -errno: in error cases
970 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
971 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
973 BDRVQcow2State *s = bs->opaque;
975 uint64_t cluster_offset;
977 uint64_t nb_clusters;
978 unsigned int keep_clusters;
981 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
984 assert(*host_offset == 0 || offset_into_cluster(s, guest_offset)
985 == offset_into_cluster(s, *host_offset));
988 * Calculate the number of clusters to look for. We stop at L2 table
989 * boundaries to keep things simple.
992 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
994 l2_index = offset_to_l2_index(s, guest_offset);
995 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
996 assert(nb_clusters <= INT_MAX);
998 /* Find L2 entry for the first involved cluster */
999 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1004 cluster_offset = be64_to_cpu(l2_table[l2_index]);
1006 /* Check how many clusters are already allocated and don't need COW */
1007 if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL
1008 && (cluster_offset & QCOW_OFLAG_COPIED))
1010 /* If a specific host_offset is required, check it */
1011 bool offset_matches =
1012 (cluster_offset & L2E_OFFSET_MASK) == *host_offset;
1014 if (offset_into_cluster(s, cluster_offset & L2E_OFFSET_MASK)) {
1015 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset "
1016 "%#llx unaligned (guest offset: %#" PRIx64
1017 ")", cluster_offset & L2E_OFFSET_MASK,
1023 if (*host_offset != 0 && !offset_matches) {
1029 /* We keep all QCOW_OFLAG_COPIED clusters */
1031 count_contiguous_clusters(nb_clusters, s->cluster_size,
1032 &l2_table[l2_index],
1033 QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
1034 assert(keep_clusters <= nb_clusters);
1036 *bytes = MIN(*bytes,
1037 keep_clusters * s->cluster_size
1038 - offset_into_cluster(s, guest_offset));
1047 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1049 /* Only return a host offset if we actually made progress. Otherwise we
1050 * would make requirements for handle_alloc() that it can't fulfill */
1052 *host_offset = (cluster_offset & L2E_OFFSET_MASK)
1053 + offset_into_cluster(s, guest_offset);
1060 * Allocates new clusters for the given guest_offset.
1062 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1063 * contain the number of clusters that have been allocated and are contiguous
1064 * in the image file.
1066 * If *host_offset is non-zero, it specifies the offset in the image file at
1067 * which the new clusters must start. *nb_clusters can be 0 on return in this
1068 * case if the cluster at host_offset is already in use. If *host_offset is
1069 * zero, the clusters can be allocated anywhere in the image file.
1071 * *host_offset is updated to contain the offset into the image file at which
1072 * the first allocated cluster starts.
1074 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1075 * function has been waiting for another request and the allocation must be
1076 * restarted, but the whole request should not be failed.
1078 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1079 uint64_t *host_offset, uint64_t *nb_clusters)
1081 BDRVQcow2State *s = bs->opaque;
1083 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1084 *host_offset, *nb_clusters);
1086 /* Allocate new clusters */
1087 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1088 if (*host_offset == 0) {
1089 int64_t cluster_offset =
1090 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1091 if (cluster_offset < 0) {
1092 return cluster_offset;
1094 *host_offset = cluster_offset;
1097 int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1107 * Allocates new clusters for an area that either is yet unallocated or needs a
1108 * copy on write. If *host_offset is non-zero, clusters are only allocated if
1109 * the new allocation can match the specified host offset.
1111 * Note that guest_offset may not be cluster aligned. In this case, the
1112 * returned *host_offset points to exact byte referenced by guest_offset and
1113 * therefore isn't cluster aligned as well.
1116 * 0: if no clusters could be allocated. *bytes is set to 0,
1117 * *host_offset is left unchanged.
1119 * 1: if new clusters were allocated. *bytes may be decreased if the
1120 * new allocation doesn't cover all of the requested area.
1121 * *host_offset is updated to contain the host offset of the first
1122 * newly allocated cluster.
1124 * -errno: in error cases
1126 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1127 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1129 BDRVQcow2State *s = bs->opaque;
1133 uint64_t nb_clusters;
1136 uint64_t alloc_cluster_offset;
1138 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1143 * Calculate the number of clusters to look for. We stop at L2 table
1144 * boundaries to keep things simple.
1147 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1149 l2_index = offset_to_l2_index(s, guest_offset);
1150 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1151 assert(nb_clusters <= INT_MAX);
1153 /* Find L2 entry for the first involved cluster */
1154 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1159 entry = be64_to_cpu(l2_table[l2_index]);
1161 /* For the moment, overwrite compressed clusters one by one */
1162 if (entry & QCOW_OFLAG_COMPRESSED) {
1165 nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index);
1168 /* This function is only called when there were no non-COW clusters, so if
1169 * we can't find any unallocated or COW clusters either, something is
1170 * wrong with our code. */
1171 assert(nb_clusters > 0);
1173 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1175 /* Allocate, if necessary at a given offset in the image file */
1176 alloc_cluster_offset = start_of_cluster(s, *host_offset);
1177 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1183 /* Can't extend contiguous allocation */
1184 if (nb_clusters == 0) {
1189 /* !*host_offset would overwrite the image header and is reserved for "no
1190 * host offset preferred". If 0 was a valid host offset, it'd trigger the
1191 * following overlap check; do that now to avoid having an invalid value in
1193 if (!alloc_cluster_offset) {
1194 ret = qcow2_pre_write_overlap_check(bs, 0, alloc_cluster_offset,
1195 nb_clusters * s->cluster_size);
1201 * Save info needed for meta data update.
1203 * requested_bytes: Number of bytes from the start of the first
1204 * newly allocated cluster to the end of the (possibly shortened
1205 * before) write request.
1207 * avail_bytes: Number of bytes from the start of the first
1208 * newly allocated to the end of the last newly allocated cluster.
1210 * nb_bytes: The number of bytes from the start of the first
1211 * newly allocated cluster to the end of the area that the write
1212 * request actually writes to (excluding COW at the end)
1214 uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1215 int avail_bytes = MIN(INT_MAX, nb_clusters << s->cluster_bits);
1216 int nb_bytes = MIN(requested_bytes, avail_bytes);
1217 QCowL2Meta *old_m = *m;
1219 *m = g_malloc0(sizeof(**m));
1221 **m = (QCowL2Meta) {
1224 .alloc_offset = alloc_cluster_offset,
1225 .offset = start_of_cluster(s, guest_offset),
1226 .nb_clusters = nb_clusters,
1230 .nb_bytes = offset_into_cluster(s, guest_offset),
1234 .nb_bytes = avail_bytes - nb_bytes,
1237 qemu_co_queue_init(&(*m)->dependent_requests);
1238 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1240 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1241 *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1242 assert(*bytes != 0);
1247 if (*m && (*m)->nb_clusters > 0) {
1248 QLIST_REMOVE(*m, next_in_flight);
1254 * alloc_cluster_offset
1256 * For a given offset on the virtual disk, find the cluster offset in qcow2
1257 * file. If the offset is not found, allocate a new cluster.
1259 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1260 * other fields in m are meaningless.
1262 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1263 * contiguous clusters that have been allocated. In this case, the other
1264 * fields of m are valid and contain information about the first allocated
1267 * If the request conflicts with another write request in flight, the coroutine
1268 * is queued and will be reentered when the dependency has completed.
1270 * Return 0 on success and -errno in error cases
1272 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1273 unsigned int *bytes, uint64_t *host_offset,
1276 BDRVQcow2State *s = bs->opaque;
1277 uint64_t start, remaining;
1278 uint64_t cluster_offset;
1282 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1294 if (!*host_offset) {
1295 *host_offset = start_of_cluster(s, cluster_offset);
1298 assert(remaining >= cur_bytes);
1301 remaining -= cur_bytes;
1302 cluster_offset += cur_bytes;
1304 if (remaining == 0) {
1308 cur_bytes = remaining;
1311 * Now start gathering as many contiguous clusters as possible:
1313 * 1. Check for overlaps with in-flight allocations
1315 * a) Overlap not in the first cluster -> shorten this request and
1316 * let the caller handle the rest in its next loop iteration.
1318 * b) Real overlaps of two requests. Yield and restart the search
1319 * for contiguous clusters (the situation could have changed
1320 * while we were sleeping)
1322 * c) TODO: Request starts in the same cluster as the in-flight
1323 * allocation ends. Shorten the COW of the in-fight allocation,
1324 * set cluster_offset to write to the same cluster and set up
1325 * the right synchronisation between the in-flight request and
1328 ret = handle_dependencies(bs, start, &cur_bytes, m);
1329 if (ret == -EAGAIN) {
1330 /* Currently handle_dependencies() doesn't yield if we already had
1331 * an allocation. If it did, we would have to clean up the L2Meta
1332 * structs before starting over. */
1335 } else if (ret < 0) {
1337 } else if (cur_bytes == 0) {
1340 /* handle_dependencies() may have decreased cur_bytes (shortened
1341 * the allocations below) so that the next dependency is processed
1342 * correctly during the next loop iteration. */
1346 * 2. Count contiguous COPIED clusters.
1348 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1353 } else if (cur_bytes == 0) {
1358 * 3. If the request still hasn't completed, allocate new clusters,
1359 * considering any cluster_offset of steps 1c or 2.
1361 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1367 assert(cur_bytes == 0);
1372 *bytes -= remaining;
1374 assert(*host_offset != 0);
1379 static int decompress_buffer(uint8_t *out_buf, int out_buf_size,
1380 const uint8_t *buf, int buf_size)
1382 z_stream strm1, *strm = &strm1;
1385 memset(strm, 0, sizeof(*strm));
1387 strm->next_in = (uint8_t *)buf;
1388 strm->avail_in = buf_size;
1389 strm->next_out = out_buf;
1390 strm->avail_out = out_buf_size;
1392 ret = inflateInit2(strm, -12);
1395 ret = inflate(strm, Z_FINISH);
1396 out_len = strm->next_out - out_buf;
1397 if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) ||
1398 out_len != out_buf_size) {
1406 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
1408 BDRVQcow2State *s = bs->opaque;
1409 int ret, csize, nb_csectors, sector_offset;
1412 coffset = cluster_offset & s->cluster_offset_mask;
1413 if (s->cluster_cache_offset != coffset) {
1414 nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1;
1415 sector_offset = coffset & 511;
1416 csize = nb_csectors * 512 - sector_offset;
1417 BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);
1418 ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data,
1423 if (decompress_buffer(s->cluster_cache, s->cluster_size,
1424 s->cluster_data + sector_offset, csize) < 0) {
1427 s->cluster_cache_offset = coffset;
1433 * This discards as many clusters of nb_clusters as possible at once (i.e.
1434 * all clusters in the same L2 table) and returns the number of discarded
1437 static int discard_single_l2(BlockDriverState *bs, uint64_t offset,
1438 uint64_t nb_clusters, enum qcow2_discard_type type,
1441 BDRVQcow2State *s = bs->opaque;
1447 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1452 /* Limit nb_clusters to one L2 table */
1453 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1454 assert(nb_clusters <= INT_MAX);
1456 for (i = 0; i < nb_clusters; i++) {
1457 uint64_t old_l2_entry;
1459 old_l2_entry = be64_to_cpu(l2_table[l2_index + i]);
1462 * If full_discard is false, make sure that a discarded area reads back
1463 * as zeroes for v3 images (we cannot do it for v2 without actually
1464 * writing a zero-filled buffer). We can skip the operation if the
1465 * cluster is already marked as zero, or if it's unallocated and we
1466 * don't have a backing file.
1468 * TODO We might want to use bdrv_get_block_status(bs) here, but we're
1469 * holding s->lock, so that doesn't work today.
1471 * If full_discard is true, the sector should not read back as zeroes,
1472 * but rather fall through to the backing file.
1474 switch (qcow2_get_cluster_type(old_l2_entry)) {
1475 case QCOW2_CLUSTER_UNALLOCATED:
1476 if (full_discard || !bs->backing) {
1481 case QCOW2_CLUSTER_ZERO:
1482 if (!full_discard) {
1487 case QCOW2_CLUSTER_NORMAL:
1488 case QCOW2_CLUSTER_COMPRESSED:
1495 /* First remove L2 entries */
1496 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1497 if (!full_discard && s->qcow_version >= 3) {
1498 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1500 l2_table[l2_index + i] = cpu_to_be64(0);
1503 /* Then decrease the refcount */
1504 qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1507 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1512 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
1513 int nb_sectors, enum qcow2_discard_type type, bool full_discard)
1515 BDRVQcow2State *s = bs->opaque;
1516 uint64_t end_offset;
1517 uint64_t nb_clusters;
1520 end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
1522 /* Round start up and end down */
1523 offset = align_offset(offset, s->cluster_size);
1524 end_offset = start_of_cluster(s, end_offset);
1526 if (offset > end_offset) {
1530 nb_clusters = size_to_clusters(s, end_offset - offset);
1532 s->cache_discards = true;
1534 /* Each L2 table is handled by its own loop iteration */
1535 while (nb_clusters > 0) {
1536 ret = discard_single_l2(bs, offset, nb_clusters, type, full_discard);
1542 offset += (ret * s->cluster_size);
1547 s->cache_discards = false;
1548 qcow2_process_discards(bs, ret);
1554 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1555 * all clusters in the same L2 table) and returns the number of zeroed
1558 static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
1559 uint64_t nb_clusters)
1561 BDRVQcow2State *s = bs->opaque;
1567 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1572 /* Limit nb_clusters to one L2 table */
1573 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1574 assert(nb_clusters <= INT_MAX);
1576 for (i = 0; i < nb_clusters; i++) {
1577 uint64_t old_offset;
1579 old_offset = be64_to_cpu(l2_table[l2_index + i]);
1581 /* Update L2 entries */
1582 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1583 if (old_offset & QCOW_OFLAG_COMPRESSED) {
1584 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1585 qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1587 l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1591 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1596 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors)
1598 BDRVQcow2State *s = bs->opaque;
1599 uint64_t nb_clusters;
1602 /* The zero flag is only supported by version 3 and newer */
1603 if (s->qcow_version < 3) {
1607 /* Each L2 table is handled by its own loop iteration */
1608 nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS);
1610 s->cache_discards = true;
1612 while (nb_clusters > 0) {
1613 ret = zero_single_l2(bs, offset, nb_clusters);
1619 offset += (ret * s->cluster_size);
1624 s->cache_discards = false;
1625 qcow2_process_discards(bs, ret);
1631 * Expands all zero clusters in a specific L1 table (or deallocates them, for
1632 * non-backed non-pre-allocated zero clusters).
1634 * l1_entries and *visited_l1_entries are used to keep track of progress for
1635 * status_cb(). l1_entries contains the total number of L1 entries and
1636 * *visited_l1_entries counts all visited L1 entries.
1638 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1639 int l1_size, int64_t *visited_l1_entries,
1641 BlockDriverAmendStatusCB *status_cb,
1644 BDRVQcow2State *s = bs->opaque;
1645 bool is_active_l1 = (l1_table == s->l1_table);
1646 uint64_t *l2_table = NULL;
1650 if (!is_active_l1) {
1651 /* inactive L2 tables require a buffer to be stored in when loading
1653 l2_table = qemu_try_blockalign(bs->file->bs, s->cluster_size);
1654 if (l2_table == NULL) {
1659 for (i = 0; i < l1_size; i++) {
1660 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1661 bool l2_dirty = false;
1662 uint64_t l2_refcount;
1666 (*visited_l1_entries)++;
1668 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
1673 if (offset_into_cluster(s, l2_offset)) {
1674 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1675 PRIx64 " unaligned (L1 index: %#x)",
1682 /* get active L2 tables from cache */
1683 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
1684 (void **)&l2_table);
1686 /* load inactive L2 tables from disk */
1687 ret = bdrv_read(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1688 (void *)l2_table, s->cluster_sectors);
1694 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1700 for (j = 0; j < s->l2_size; j++) {
1701 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1702 int64_t offset = l2_entry & L2E_OFFSET_MASK;
1703 int cluster_type = qcow2_get_cluster_type(l2_entry);
1704 bool preallocated = offset != 0;
1706 if (cluster_type != QCOW2_CLUSTER_ZERO) {
1710 if (!preallocated) {
1712 /* not backed; therefore we can simply deallocate the
1719 offset = qcow2_alloc_clusters(bs, s->cluster_size);
1725 if (l2_refcount > 1) {
1726 /* For shared L2 tables, set the refcount accordingly (it is
1727 * already 1 and needs to be l2_refcount) */
1728 ret = qcow2_update_cluster_refcount(bs,
1729 offset >> s->cluster_bits,
1730 refcount_diff(1, l2_refcount), false,
1731 QCOW2_DISCARD_OTHER);
1733 qcow2_free_clusters(bs, offset, s->cluster_size,
1734 QCOW2_DISCARD_OTHER);
1740 if (offset_into_cluster(s, offset)) {
1741 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset "
1742 "%#" PRIx64 " unaligned (L2 offset: %#"
1743 PRIx64 ", L2 index: %#x)", offset,
1745 if (!preallocated) {
1746 qcow2_free_clusters(bs, offset, s->cluster_size,
1747 QCOW2_DISCARD_ALWAYS);
1753 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size);
1755 if (!preallocated) {
1756 qcow2_free_clusters(bs, offset, s->cluster_size,
1757 QCOW2_DISCARD_ALWAYS);
1762 ret = bdrv_pwrite_zeroes(bs->file, offset, s->cluster_size, 0);
1764 if (!preallocated) {
1765 qcow2_free_clusters(bs, offset, s->cluster_size,
1766 QCOW2_DISCARD_ALWAYS);
1771 if (l2_refcount == 1) {
1772 l2_table[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
1774 l2_table[j] = cpu_to_be64(offset);
1781 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1782 qcow2_cache_depends_on_flush(s->l2_table_cache);
1784 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1787 ret = qcow2_pre_write_overlap_check(bs,
1788 QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2, l2_offset,
1794 ret = bdrv_write(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1795 (void *)l2_table, s->cluster_sectors);
1802 (*visited_l1_entries)++;
1804 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
1812 if (!is_active_l1) {
1813 qemu_vfree(l2_table);
1815 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1822 * For backed images, expands all zero clusters on the image. For non-backed
1823 * images, deallocates all non-pre-allocated zero clusters (and claims the
1824 * allocation for pre-allocated ones). This is important for downgrading to a
1825 * qcow2 version which doesn't yet support metadata zero clusters.
1827 int qcow2_expand_zero_clusters(BlockDriverState *bs,
1828 BlockDriverAmendStatusCB *status_cb,
1831 BDRVQcow2State *s = bs->opaque;
1832 uint64_t *l1_table = NULL;
1833 int64_t l1_entries = 0, visited_l1_entries = 0;
1838 l1_entries = s->l1_size;
1839 for (i = 0; i < s->nb_snapshots; i++) {
1840 l1_entries += s->snapshots[i].l1_size;
1844 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
1845 &visited_l1_entries, l1_entries,
1846 status_cb, cb_opaque);
1851 /* Inactive L1 tables may point to active L2 tables - therefore it is
1852 * necessary to flush the L2 table cache before trying to access the L2
1853 * tables pointed to by inactive L1 entries (else we might try to expand
1854 * zero clusters that have already been expanded); furthermore, it is also
1855 * necessary to empty the L2 table cache, since it may contain tables which
1856 * are now going to be modified directly on disk, bypassing the cache.
1857 * qcow2_cache_empty() does both for us. */
1858 ret = qcow2_cache_empty(bs, s->l2_table_cache);
1863 for (i = 0; i < s->nb_snapshots; i++) {
1864 int l1_sectors = DIV_ROUND_UP(s->snapshots[i].l1_size *
1865 sizeof(uint64_t), BDRV_SECTOR_SIZE);
1867 l1_table = g_realloc(l1_table, l1_sectors * BDRV_SECTOR_SIZE);
1869 ret = bdrv_read(bs->file,
1870 s->snapshots[i].l1_table_offset / BDRV_SECTOR_SIZE,
1871 (void *)l1_table, l1_sectors);
1876 for (j = 0; j < s->snapshots[i].l1_size; j++) {
1877 be64_to_cpus(&l1_table[j]);
1880 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
1881 &visited_l1_entries, l1_entries,
1882 status_cb, cb_opaque);