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_shrink_l1_table(BlockDriverState
*bs
, uint64_t exact_size
)
37 BDRVQcow2State
*s
= bs
->opaque
;
38 int new_l1_size
, i
, ret
;
40 if (exact_size
>= s
->l1_size
) {
44 new_l1_size
= exact_size
;
47 fprintf(stderr
, "shrink l1_table from %d to %d\n", s
->l1_size
, new_l1_size
);
50 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_SHRINK_WRITE_TABLE
);
51 ret
= bdrv_pwrite_zeroes(bs
->file
, s
->l1_table_offset
+
52 new_l1_size
* sizeof(uint64_t),
53 (s
->l1_size
- new_l1_size
) * sizeof(uint64_t), 0);
58 ret
= bdrv_flush(bs
->file
->bs
);
63 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_SHRINK_FREE_L2_CLUSTERS
);
64 for (i
= s
->l1_size
- 1; i
> new_l1_size
- 1; i
--) {
65 if ((s
->l1_table
[i
] & L1E_OFFSET_MASK
) == 0) {
68 qcow2_free_clusters(bs
, s
->l1_table
[i
] & L1E_OFFSET_MASK
,
69 s
->cluster_size
, QCOW2_DISCARD_ALWAYS
);
76 * If the write in the l1_table failed the image may contain a partially
77 * overwritten l1_table. In this case it would be better to clear the
78 * l1_table in memory to avoid possible image corruption.
80 memset(s
->l1_table
+ new_l1_size
, 0,
81 (s
->l1_size
- new_l1_size
) * sizeof(uint64_t));
85 int qcow2_grow_l1_table(BlockDriverState
*bs
, uint64_t min_size
,
88 BDRVQcow2State
*s
= bs
->opaque
;
89 int new_l1_size2
, ret
, i
;
90 uint64_t *new_l1_table
;
91 int64_t old_l1_table_offset
, old_l1_size
;
92 int64_t new_l1_table_offset
, new_l1_size
;
95 if (min_size
<= s
->l1_size
)
98 /* Do a sanity check on min_size before trying to calculate new_l1_size
99 * (this prevents overflows during the while loop for the calculation of
101 if (min_size
> INT_MAX
/ sizeof(uint64_t)) {
106 new_l1_size
= min_size
;
108 /* Bump size up to reduce the number of times we have to grow */
109 new_l1_size
= s
->l1_size
;
110 if (new_l1_size
== 0) {
113 while (min_size
> new_l1_size
) {
114 new_l1_size
= DIV_ROUND_UP(new_l1_size
* 3, 2);
118 QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE
> INT_MAX
);
119 if (new_l1_size
> QCOW_MAX_L1_SIZE
/ sizeof(uint64_t)) {
124 fprintf(stderr
, "grow l1_table from %d to %" PRId64
"\n",
125 s
->l1_size
, new_l1_size
);
128 new_l1_size2
= sizeof(uint64_t) * new_l1_size
;
129 new_l1_table
= qemu_try_blockalign(bs
->file
->bs
,
130 align_offset(new_l1_size2
, 512));
131 if (new_l1_table
== NULL
) {
134 memset(new_l1_table
, 0, align_offset(new_l1_size2
, 512));
137 memcpy(new_l1_table
, s
->l1_table
, s
->l1_size
* sizeof(uint64_t));
140 /* write new table (align to cluster) */
141 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_GROW_ALLOC_TABLE
);
142 new_l1_table_offset
= qcow2_alloc_clusters(bs
, new_l1_size2
);
143 if (new_l1_table_offset
< 0) {
144 qemu_vfree(new_l1_table
);
145 return new_l1_table_offset
;
148 ret
= qcow2_cache_flush(bs
, s
->refcount_block_cache
);
153 /* the L1 position has not yet been updated, so these clusters must
154 * indeed be completely free */
155 ret
= qcow2_pre_write_overlap_check(bs
, 0, new_l1_table_offset
,
161 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_GROW_WRITE_TABLE
);
162 for(i
= 0; i
< s
->l1_size
; i
++)
163 new_l1_table
[i
] = cpu_to_be64(new_l1_table
[i
]);
164 ret
= bdrv_pwrite_sync(bs
->file
, new_l1_table_offset
,
165 new_l1_table
, new_l1_size2
);
168 for(i
= 0; i
< s
->l1_size
; i
++)
169 new_l1_table
[i
] = be64_to_cpu(new_l1_table
[i
]);
172 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_GROW_ACTIVATE_TABLE
);
173 stl_be_p(data
, new_l1_size
);
174 stq_be_p(data
+ 4, new_l1_table_offset
);
175 ret
= bdrv_pwrite_sync(bs
->file
, offsetof(QCowHeader
, l1_size
),
180 qemu_vfree(s
->l1_table
);
181 old_l1_table_offset
= s
->l1_table_offset
;
182 s
->l1_table_offset
= new_l1_table_offset
;
183 s
->l1_table
= new_l1_table
;
184 old_l1_size
= s
->l1_size
;
185 s
->l1_size
= new_l1_size
;
186 qcow2_free_clusters(bs
, old_l1_table_offset
, old_l1_size
* sizeof(uint64_t),
187 QCOW2_DISCARD_OTHER
);
190 qemu_vfree(new_l1_table
);
191 qcow2_free_clusters(bs
, new_l1_table_offset
, new_l1_size2
,
192 QCOW2_DISCARD_OTHER
);
199 * Loads a L2 table into memory. If the table is in the cache, the cache
200 * is used; otherwise the L2 table is loaded from the image file.
202 * Returns a pointer to the L2 table on success, or NULL if the read from
203 * the image file failed.
206 static int l2_load(BlockDriverState
*bs
, uint64_t l2_offset
,
209 BDRVQcow2State
*s
= bs
->opaque
;
211 return qcow2_cache_get(bs
, s
->l2_table_cache
, l2_offset
,
216 * Writes one sector of the L1 table to the disk (can't update single entries
217 * and we really don't want bdrv_pread to perform a read-modify-write)
219 #define L1_ENTRIES_PER_SECTOR (512 / 8)
220 int qcow2_write_l1_entry(BlockDriverState
*bs
, int l1_index
)
222 BDRVQcow2State
*s
= bs
->opaque
;
223 uint64_t buf
[L1_ENTRIES_PER_SECTOR
] = { 0 };
227 l1_start_index
= l1_index
& ~(L1_ENTRIES_PER_SECTOR
- 1);
228 for (i
= 0; i
< L1_ENTRIES_PER_SECTOR
&& l1_start_index
+ i
< s
->l1_size
;
231 buf
[i
] = cpu_to_be64(s
->l1_table
[l1_start_index
+ i
]);
234 ret
= qcow2_pre_write_overlap_check(bs
, QCOW2_OL_ACTIVE_L1
,
235 s
->l1_table_offset
+ 8 * l1_start_index
, sizeof(buf
));
240 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_UPDATE
);
241 ret
= bdrv_pwrite_sync(bs
->file
,
242 s
->l1_table_offset
+ 8 * l1_start_index
,
254 * Allocate a new l2 entry in the file. If l1_index points to an already
255 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
256 * table) copy the contents of the old L2 table into the newly allocated one.
257 * Otherwise the new table is initialized with zeros.
261 static int l2_allocate(BlockDriverState
*bs
, int l1_index
, uint64_t **table
)
263 BDRVQcow2State
*s
= bs
->opaque
;
264 uint64_t old_l2_offset
;
265 uint64_t *l2_table
= NULL
;
269 old_l2_offset
= s
->l1_table
[l1_index
];
271 trace_qcow2_l2_allocate(bs
, l1_index
);
273 /* allocate a new l2 entry */
275 l2_offset
= qcow2_alloc_clusters(bs
, s
->l2_size
* sizeof(uint64_t));
281 ret
= qcow2_cache_flush(bs
, s
->refcount_block_cache
);
286 /* allocate a new entry in the l2 cache */
288 trace_qcow2_l2_allocate_get_empty(bs
, l1_index
);
289 ret
= qcow2_cache_get_empty(bs
, s
->l2_table_cache
, l2_offset
, (void**) table
);
296 if ((old_l2_offset
& L1E_OFFSET_MASK
) == 0) {
297 /* if there was no old l2 table, clear the new table */
298 memset(l2_table
, 0, s
->l2_size
* sizeof(uint64_t));
302 /* if there was an old l2 table, read it from the disk */
303 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_ALLOC_COW_READ
);
304 ret
= qcow2_cache_get(bs
, s
->l2_table_cache
,
305 old_l2_offset
& L1E_OFFSET_MASK
,
306 (void**) &old_table
);
311 memcpy(l2_table
, old_table
, s
->cluster_size
);
313 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &old_table
);
316 /* write the l2 table to the file */
317 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_ALLOC_WRITE
);
319 trace_qcow2_l2_allocate_write_l2(bs
, l1_index
);
320 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
321 ret
= qcow2_cache_flush(bs
, s
->l2_table_cache
);
326 /* update the L1 entry */
327 trace_qcow2_l2_allocate_write_l1(bs
, l1_index
);
328 s
->l1_table
[l1_index
] = l2_offset
| QCOW_OFLAG_COPIED
;
329 ret
= qcow2_write_l1_entry(bs
, l1_index
);
335 trace_qcow2_l2_allocate_done(bs
, l1_index
, 0);
339 trace_qcow2_l2_allocate_done(bs
, l1_index
, ret
);
340 if (l2_table
!= NULL
) {
341 qcow2_cache_put(bs
, s
->l2_table_cache
, (void**) table
);
343 s
->l1_table
[l1_index
] = old_l2_offset
;
345 qcow2_free_clusters(bs
, l2_offset
, s
->l2_size
* sizeof(uint64_t),
346 QCOW2_DISCARD_ALWAYS
);
352 * Checks how many clusters in a given L2 table are contiguous in the image
353 * file. As soon as one of the flags in the bitmask stop_flags changes compared
354 * to the first cluster, the search is stopped and the cluster is not counted
355 * as contiguous. (This allows it, for example, to stop at the first compressed
356 * cluster which may require a different handling)
358 static int count_contiguous_clusters(int nb_clusters
, int cluster_size
,
359 uint64_t *l2_table
, uint64_t stop_flags
)
362 QCow2ClusterType first_cluster_type
;
363 uint64_t mask
= stop_flags
| L2E_OFFSET_MASK
| QCOW_OFLAG_COMPRESSED
;
364 uint64_t first_entry
= be64_to_cpu(l2_table
[0]);
365 uint64_t offset
= first_entry
& mask
;
371 /* must be allocated */
372 first_cluster_type
= qcow2_get_cluster_type(first_entry
);
373 assert(first_cluster_type
== QCOW2_CLUSTER_NORMAL
||
374 first_cluster_type
== QCOW2_CLUSTER_ZERO_ALLOC
);
376 for (i
= 0; i
< nb_clusters
; i
++) {
377 uint64_t l2_entry
= be64_to_cpu(l2_table
[i
]) & mask
;
378 if (offset
+ (uint64_t) i
* cluster_size
!= l2_entry
) {
387 * Checks how many consecutive unallocated clusters in a given L2
388 * table have the same cluster type.
390 static int count_contiguous_clusters_unallocated(int nb_clusters
,
392 QCow2ClusterType wanted_type
)
396 assert(wanted_type
== QCOW2_CLUSTER_ZERO_PLAIN
||
397 wanted_type
== QCOW2_CLUSTER_UNALLOCATED
);
398 for (i
= 0; i
< nb_clusters
; i
++) {
399 uint64_t entry
= be64_to_cpu(l2_table
[i
]);
400 QCow2ClusterType type
= qcow2_get_cluster_type(entry
);
402 if (type
!= wanted_type
) {
410 static int coroutine_fn
do_perform_cow_read(BlockDriverState
*bs
,
411 uint64_t src_cluster_offset
,
412 unsigned offset_in_cluster
,
417 if (qiov
->size
== 0) {
421 BLKDBG_EVENT(bs
->file
, BLKDBG_COW_READ
);
427 /* Call .bdrv_co_readv() directly instead of using the public block-layer
428 * interface. This avoids double I/O throttling and request tracking,
429 * which can lead to deadlock when block layer copy-on-read is enabled.
431 ret
= bs
->drv
->bdrv_co_preadv(bs
, src_cluster_offset
+ offset_in_cluster
,
432 qiov
->size
, qiov
, 0);
440 static bool coroutine_fn
do_perform_cow_encrypt(BlockDriverState
*bs
,
441 uint64_t src_cluster_offset
,
442 uint64_t cluster_offset
,
443 unsigned offset_in_cluster
,
447 if (bytes
&& bs
->encrypted
) {
448 BDRVQcow2State
*s
= bs
->opaque
;
449 int64_t sector
= (s
->crypt_physical_offset
?
450 (cluster_offset
+ offset_in_cluster
) :
451 (src_cluster_offset
+ offset_in_cluster
))
453 assert((offset_in_cluster
& ~BDRV_SECTOR_MASK
) == 0);
454 assert((bytes
& ~BDRV_SECTOR_MASK
) == 0);
456 if (qcrypto_block_encrypt(s
->crypto
, sector
, buffer
,
464 static int coroutine_fn
do_perform_cow_write(BlockDriverState
*bs
,
465 uint64_t cluster_offset
,
466 unsigned offset_in_cluster
,
471 if (qiov
->size
== 0) {
475 ret
= qcow2_pre_write_overlap_check(bs
, 0,
476 cluster_offset
+ offset_in_cluster
, qiov
->size
);
481 BLKDBG_EVENT(bs
->file
, BLKDBG_COW_WRITE
);
482 ret
= bdrv_co_pwritev(bs
->file
, cluster_offset
+ offset_in_cluster
,
483 qiov
->size
, qiov
, 0);
495 * For a given offset of the virtual disk, find the cluster type and offset in
496 * the qcow2 file. The offset is stored in *cluster_offset.
498 * On entry, *bytes is the maximum number of contiguous bytes starting at
499 * offset that we are interested in.
501 * On exit, *bytes is the number of bytes starting at offset that have the same
502 * cluster type and (if applicable) are stored contiguously in the image file.
503 * Compressed clusters are always returned one by one.
505 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
508 int qcow2_get_cluster_offset(BlockDriverState
*bs
, uint64_t offset
,
509 unsigned int *bytes
, uint64_t *cluster_offset
)
511 BDRVQcow2State
*s
= bs
->opaque
;
512 unsigned int l2_index
;
513 uint64_t l1_index
, l2_offset
, *l2_table
;
515 unsigned int offset_in_cluster
;
516 uint64_t bytes_available
, bytes_needed
, nb_clusters
;
517 QCow2ClusterType type
;
520 offset_in_cluster
= offset_into_cluster(s
, offset
);
521 bytes_needed
= (uint64_t) *bytes
+ offset_in_cluster
;
523 l1_bits
= s
->l2_bits
+ s
->cluster_bits
;
525 /* compute how many bytes there are between the start of the cluster
526 * containing offset and the end of the l1 entry */
527 bytes_available
= (1ULL << l1_bits
) - (offset
& ((1ULL << l1_bits
) - 1))
530 if (bytes_needed
> bytes_available
) {
531 bytes_needed
= bytes_available
;
536 /* seek to the l2 offset in the l1 table */
538 l1_index
= offset
>> l1_bits
;
539 if (l1_index
>= s
->l1_size
) {
540 type
= QCOW2_CLUSTER_UNALLOCATED
;
544 l2_offset
= s
->l1_table
[l1_index
] & L1E_OFFSET_MASK
;
546 type
= QCOW2_CLUSTER_UNALLOCATED
;
550 if (offset_into_cluster(s
, l2_offset
)) {
551 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#" PRIx64
552 " unaligned (L1 index: %#" PRIx64
")",
553 l2_offset
, l1_index
);
557 /* load the l2 table in memory */
559 ret
= l2_load(bs
, l2_offset
, &l2_table
);
564 /* find the cluster offset for the given disk offset */
566 l2_index
= offset_to_l2_index(s
, offset
);
567 *cluster_offset
= be64_to_cpu(l2_table
[l2_index
]);
569 nb_clusters
= size_to_clusters(s
, bytes_needed
);
570 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
571 * integers; the minimum cluster size is 512, so this assertion is always
573 assert(nb_clusters
<= INT_MAX
);
575 type
= qcow2_get_cluster_type(*cluster_offset
);
576 if (s
->qcow_version
< 3 && (type
== QCOW2_CLUSTER_ZERO_PLAIN
||
577 type
== QCOW2_CLUSTER_ZERO_ALLOC
)) {
578 qcow2_signal_corruption(bs
, true, -1, -1, "Zero cluster entry found"
579 " in pre-v3 image (L2 offset: %#" PRIx64
580 ", L2 index: %#x)", l2_offset
, l2_index
);
585 case QCOW2_CLUSTER_COMPRESSED
:
586 /* Compressed clusters can only be processed one by one */
588 *cluster_offset
&= L2E_COMPRESSED_OFFSET_SIZE_MASK
;
590 case QCOW2_CLUSTER_ZERO_PLAIN
:
591 case QCOW2_CLUSTER_UNALLOCATED
:
592 /* how many empty clusters ? */
593 c
= count_contiguous_clusters_unallocated(nb_clusters
,
594 &l2_table
[l2_index
], type
);
597 case QCOW2_CLUSTER_ZERO_ALLOC
:
598 case QCOW2_CLUSTER_NORMAL
:
599 /* how many allocated clusters ? */
600 c
= count_contiguous_clusters(nb_clusters
, s
->cluster_size
,
601 &l2_table
[l2_index
], QCOW_OFLAG_ZERO
);
602 *cluster_offset
&= L2E_OFFSET_MASK
;
603 if (offset_into_cluster(s
, *cluster_offset
)) {
604 qcow2_signal_corruption(bs
, true, -1, -1,
605 "Cluster allocation offset %#"
606 PRIx64
" unaligned (L2 offset: %#" PRIx64
607 ", L2 index: %#x)", *cluster_offset
,
608 l2_offset
, l2_index
);
617 qcow2_cache_put(bs
, s
->l2_table_cache
, (void**) &l2_table
);
619 bytes_available
= (int64_t)c
* s
->cluster_size
;
622 if (bytes_available
> bytes_needed
) {
623 bytes_available
= bytes_needed
;
626 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
627 * subtracting offset_in_cluster will therefore definitely yield something
628 * not exceeding UINT_MAX */
629 assert(bytes_available
- offset_in_cluster
<= UINT_MAX
);
630 *bytes
= bytes_available
- offset_in_cluster
;
635 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **)&l2_table
);
642 * for a given disk offset, load (and allocate if needed)
645 * the l2 table offset in the qcow2 file and the cluster index
646 * in the l2 table are given to the caller.
648 * Returns 0 on success, -errno in failure case
650 static int get_cluster_table(BlockDriverState
*bs
, uint64_t offset
,
651 uint64_t **new_l2_table
,
654 BDRVQcow2State
*s
= bs
->opaque
;
655 unsigned int l2_index
;
656 uint64_t l1_index
, l2_offset
;
657 uint64_t *l2_table
= NULL
;
660 /* seek to the l2 offset in the l1 table */
662 l1_index
= offset
>> (s
->l2_bits
+ s
->cluster_bits
);
663 if (l1_index
>= s
->l1_size
) {
664 ret
= qcow2_grow_l1_table(bs
, l1_index
+ 1, false);
670 assert(l1_index
< s
->l1_size
);
671 l2_offset
= s
->l1_table
[l1_index
] & L1E_OFFSET_MASK
;
672 if (offset_into_cluster(s
, l2_offset
)) {
673 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#" PRIx64
674 " unaligned (L1 index: %#" PRIx64
")",
675 l2_offset
, l1_index
);
679 /* seek the l2 table of the given l2 offset */
681 if (s
->l1_table
[l1_index
] & QCOW_OFLAG_COPIED
) {
682 /* load the l2 table in memory */
683 ret
= l2_load(bs
, l2_offset
, &l2_table
);
688 /* First allocate a new L2 table (and do COW if needed) */
689 ret
= l2_allocate(bs
, l1_index
, &l2_table
);
694 /* Then decrease the refcount of the old table */
696 qcow2_free_clusters(bs
, l2_offset
, s
->l2_size
* sizeof(uint64_t),
697 QCOW2_DISCARD_OTHER
);
701 /* find the cluster offset for the given disk offset */
703 l2_index
= offset_to_l2_index(s
, offset
);
705 *new_l2_table
= l2_table
;
706 *new_l2_index
= l2_index
;
712 * alloc_compressed_cluster_offset
714 * For a given offset of the disk image, return cluster offset in
717 * If the offset is not found, allocate a new compressed cluster.
719 * Return the cluster offset if successful,
720 * Return 0, otherwise.
724 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState
*bs
,
728 BDRVQcow2State
*s
= bs
->opaque
;
731 int64_t cluster_offset
;
734 ret
= get_cluster_table(bs
, offset
, &l2_table
, &l2_index
);
739 /* Compression can't overwrite anything. Fail if the cluster was already
741 cluster_offset
= be64_to_cpu(l2_table
[l2_index
]);
742 if (cluster_offset
& L2E_OFFSET_MASK
) {
743 qcow2_cache_put(bs
, s
->l2_table_cache
, (void**) &l2_table
);
747 cluster_offset
= qcow2_alloc_bytes(bs
, compressed_size
);
748 if (cluster_offset
< 0) {
749 qcow2_cache_put(bs
, s
->l2_table_cache
, (void**) &l2_table
);
753 nb_csectors
= ((cluster_offset
+ compressed_size
- 1) >> 9) -
754 (cluster_offset
>> 9);
756 cluster_offset
|= QCOW_OFLAG_COMPRESSED
|
757 ((uint64_t)nb_csectors
<< s
->csize_shift
);
759 /* update L2 table */
761 /* compressed clusters never have the copied flag */
763 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_UPDATE_COMPRESSED
);
764 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
765 l2_table
[l2_index
] = cpu_to_be64(cluster_offset
);
766 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
768 return cluster_offset
;
771 static int perform_cow(BlockDriverState
*bs
, QCowL2Meta
*m
)
773 BDRVQcow2State
*s
= bs
->opaque
;
774 Qcow2COWRegion
*start
= &m
->cow_start
;
775 Qcow2COWRegion
*end
= &m
->cow_end
;
776 unsigned buffer_size
;
777 unsigned data_bytes
= end
->offset
- (start
->offset
+ start
->nb_bytes
);
779 uint8_t *start_buffer
, *end_buffer
;
783 assert(start
->nb_bytes
<= UINT_MAX
- end
->nb_bytes
);
784 assert(start
->nb_bytes
+ end
->nb_bytes
<= UINT_MAX
- data_bytes
);
785 assert(start
->offset
+ start
->nb_bytes
<= end
->offset
);
786 assert(!m
->data_qiov
|| m
->data_qiov
->size
== data_bytes
);
788 if (start
->nb_bytes
== 0 && end
->nb_bytes
== 0) {
792 /* If we have to read both the start and end COW regions and the
793 * middle region is not too large then perform just one read
795 merge_reads
= start
->nb_bytes
&& end
->nb_bytes
&& data_bytes
<= 16384;
797 buffer_size
= start
->nb_bytes
+ data_bytes
+ end
->nb_bytes
;
799 /* If we have to do two reads, add some padding in the middle
800 * if necessary to make sure that the end region is optimally
802 size_t align
= bdrv_opt_mem_align(bs
);
803 assert(align
> 0 && align
<= UINT_MAX
);
804 assert(QEMU_ALIGN_UP(start
->nb_bytes
, align
) <=
805 UINT_MAX
- end
->nb_bytes
);
806 buffer_size
= QEMU_ALIGN_UP(start
->nb_bytes
, align
) + end
->nb_bytes
;
809 /* Reserve a buffer large enough to store all the data that we're
811 start_buffer
= qemu_try_blockalign(bs
, buffer_size
);
812 if (start_buffer
== NULL
) {
815 /* The part of the buffer where the end region is located */
816 end_buffer
= start_buffer
+ buffer_size
- end
->nb_bytes
;
818 qemu_iovec_init(&qiov
, 2 + (m
->data_qiov
? m
->data_qiov
->niov
: 0));
820 qemu_co_mutex_unlock(&s
->lock
);
821 /* First we read the existing data from both COW regions. We
822 * either read the whole region in one go, or the start and end
823 * regions separately. */
825 qemu_iovec_add(&qiov
, start_buffer
, buffer_size
);
826 ret
= do_perform_cow_read(bs
, m
->offset
, start
->offset
, &qiov
);
828 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
829 ret
= do_perform_cow_read(bs
, m
->offset
, start
->offset
, &qiov
);
834 qemu_iovec_reset(&qiov
);
835 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
836 ret
= do_perform_cow_read(bs
, m
->offset
, end
->offset
, &qiov
);
842 /* Encrypt the data if necessary before writing it */
844 if (!do_perform_cow_encrypt(bs
, m
->offset
, m
->alloc_offset
,
845 start
->offset
, start_buffer
,
847 !do_perform_cow_encrypt(bs
, m
->offset
, m
->alloc_offset
,
848 end
->offset
, end_buffer
, end
->nb_bytes
)) {
854 /* And now we can write everything. If we have the guest data we
855 * can write everything in one single operation */
857 qemu_iovec_reset(&qiov
);
858 if (start
->nb_bytes
) {
859 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
861 qemu_iovec_concat(&qiov
, m
->data_qiov
, 0, data_bytes
);
863 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
865 /* NOTE: we have a write_aio blkdebug event here followed by
866 * a cow_write one in do_perform_cow_write(), but there's only
867 * one single I/O operation */
868 BLKDBG_EVENT(bs
->file
, BLKDBG_WRITE_AIO
);
869 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, start
->offset
, &qiov
);
871 /* If there's no guest data then write both COW regions separately */
872 qemu_iovec_reset(&qiov
);
873 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
874 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, start
->offset
, &qiov
);
879 qemu_iovec_reset(&qiov
);
880 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
881 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, end
->offset
, &qiov
);
885 qemu_co_mutex_lock(&s
->lock
);
888 * Before we update the L2 table to actually point to the new cluster, we
889 * need to be sure that the refcounts have been increased and COW was
893 qcow2_cache_depends_on_flush(s
->l2_table_cache
);
896 qemu_vfree(start_buffer
);
897 qemu_iovec_destroy(&qiov
);
901 int qcow2_alloc_cluster_link_l2(BlockDriverState
*bs
, QCowL2Meta
*m
)
903 BDRVQcow2State
*s
= bs
->opaque
;
904 int i
, j
= 0, l2_index
, ret
;
905 uint64_t *old_cluster
, *l2_table
;
906 uint64_t cluster_offset
= m
->alloc_offset
;
908 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m
->nb_clusters
);
909 assert(m
->nb_clusters
> 0);
911 old_cluster
= g_try_new(uint64_t, m
->nb_clusters
);
912 if (old_cluster
== NULL
) {
917 /* copy content of unmodified sectors */
918 ret
= perform_cow(bs
, m
);
923 /* Update L2 table. */
924 if (s
->use_lazy_refcounts
) {
925 qcow2_mark_dirty(bs
);
927 if (qcow2_need_accurate_refcounts(s
)) {
928 qcow2_cache_set_dependency(bs
, s
->l2_table_cache
,
929 s
->refcount_block_cache
);
932 ret
= get_cluster_table(bs
, m
->offset
, &l2_table
, &l2_index
);
936 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
938 assert(l2_index
+ m
->nb_clusters
<= s
->l2_size
);
939 for (i
= 0; i
< m
->nb_clusters
; i
++) {
940 /* if two concurrent writes happen to the same unallocated cluster
941 * each write allocates separate cluster and writes data concurrently.
942 * The first one to complete updates l2 table with pointer to its
943 * cluster the second one has to do RMW (which is done above by
944 * perform_cow()), update l2 table with its cluster pointer and free
945 * old cluster. This is what this loop does */
946 if (l2_table
[l2_index
+ i
] != 0) {
947 old_cluster
[j
++] = l2_table
[l2_index
+ i
];
950 l2_table
[l2_index
+ i
] = cpu_to_be64((cluster_offset
+
951 (i
<< s
->cluster_bits
)) | QCOW_OFLAG_COPIED
);
955 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
958 * If this was a COW, we need to decrease the refcount of the old cluster.
960 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
961 * clusters), the next write will reuse them anyway.
963 if (!m
->keep_old_clusters
&& j
!= 0) {
964 for (i
= 0; i
< j
; i
++) {
965 qcow2_free_any_clusters(bs
, be64_to_cpu(old_cluster
[i
]), 1,
966 QCOW2_DISCARD_NEVER
);
977 * Returns the number of contiguous clusters that can be used for an allocating
978 * write, but require COW to be performed (this includes yet unallocated space,
979 * which must copy from the backing file)
981 static int count_cow_clusters(BDRVQcow2State
*s
, int nb_clusters
,
982 uint64_t *l2_table
, int l2_index
)
986 for (i
= 0; i
< nb_clusters
; i
++) {
987 uint64_t l2_entry
= be64_to_cpu(l2_table
[l2_index
+ i
]);
988 QCow2ClusterType cluster_type
= qcow2_get_cluster_type(l2_entry
);
990 switch(cluster_type
) {
991 case QCOW2_CLUSTER_NORMAL
:
992 if (l2_entry
& QCOW_OFLAG_COPIED
) {
996 case QCOW2_CLUSTER_UNALLOCATED
:
997 case QCOW2_CLUSTER_COMPRESSED
:
998 case QCOW2_CLUSTER_ZERO_PLAIN
:
999 case QCOW2_CLUSTER_ZERO_ALLOC
:
1007 assert(i
<= nb_clusters
);
1012 * Check if there already is an AIO write request in flight which allocates
1013 * the same cluster. In this case we need to wait until the previous
1014 * request has completed and updated the L2 table accordingly.
1017 * 0 if there was no dependency. *cur_bytes indicates the number of
1018 * bytes from guest_offset that can be read before the next
1019 * dependency must be processed (or the request is complete)
1021 * -EAGAIN if we had to wait for another request, previously gathered
1022 * information on cluster allocation may be invalid now. The caller
1023 * must start over anyway, so consider *cur_bytes undefined.
1025 static int handle_dependencies(BlockDriverState
*bs
, uint64_t guest_offset
,
1026 uint64_t *cur_bytes
, QCowL2Meta
**m
)
1028 BDRVQcow2State
*s
= bs
->opaque
;
1029 QCowL2Meta
*old_alloc
;
1030 uint64_t bytes
= *cur_bytes
;
1032 QLIST_FOREACH(old_alloc
, &s
->cluster_allocs
, next_in_flight
) {
1034 uint64_t start
= guest_offset
;
1035 uint64_t end
= start
+ bytes
;
1036 uint64_t old_start
= l2meta_cow_start(old_alloc
);
1037 uint64_t old_end
= l2meta_cow_end(old_alloc
);
1039 if (end
<= old_start
|| start
>= old_end
) {
1040 /* No intersection */
1042 if (start
< old_start
) {
1043 /* Stop at the start of a running allocation */
1044 bytes
= old_start
- start
;
1049 /* Stop if already an l2meta exists. After yielding, it wouldn't
1050 * be valid any more, so we'd have to clean up the old L2Metas
1051 * and deal with requests depending on them before starting to
1052 * gather new ones. Not worth the trouble. */
1053 if (bytes
== 0 && *m
) {
1059 /* Wait for the dependency to complete. We need to recheck
1060 * the free/allocated clusters when we continue. */
1061 qemu_co_queue_wait(&old_alloc
->dependent_requests
, &s
->lock
);
1067 /* Make sure that existing clusters and new allocations are only used up to
1068 * the next dependency if we shortened the request above */
1075 * Checks how many already allocated clusters that don't require a copy on
1076 * write there are at the given guest_offset (up to *bytes). If
1077 * *host_offset is not zero, only physically contiguous clusters beginning at
1078 * this host offset are counted.
1080 * Note that guest_offset may not be cluster aligned. In this case, the
1081 * returned *host_offset points to exact byte referenced by guest_offset and
1082 * therefore isn't cluster aligned as well.
1085 * 0: if no allocated clusters are available at the given offset.
1086 * *bytes is normally unchanged. It is set to 0 if the cluster
1087 * is allocated and doesn't need COW, but doesn't have the right
1090 * 1: if allocated clusters that don't require a COW are available at
1091 * the requested offset. *bytes may have decreased and describes
1092 * the length of the area that can be written to.
1094 * -errno: in error cases
1096 static int handle_copied(BlockDriverState
*bs
, uint64_t guest_offset
,
1097 uint64_t *host_offset
, uint64_t *bytes
, QCowL2Meta
**m
)
1099 BDRVQcow2State
*s
= bs
->opaque
;
1101 uint64_t cluster_offset
;
1103 uint64_t nb_clusters
;
1104 unsigned int keep_clusters
;
1107 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset
, *host_offset
,
1110 assert(*host_offset
== 0 || offset_into_cluster(s
, guest_offset
)
1111 == offset_into_cluster(s
, *host_offset
));
1114 * Calculate the number of clusters to look for. We stop at L2 table
1115 * boundaries to keep things simple.
1118 size_to_clusters(s
, offset_into_cluster(s
, guest_offset
) + *bytes
);
1120 l2_index
= offset_to_l2_index(s
, guest_offset
);
1121 nb_clusters
= MIN(nb_clusters
, s
->l2_size
- l2_index
);
1122 assert(nb_clusters
<= INT_MAX
);
1124 /* Find L2 entry for the first involved cluster */
1125 ret
= get_cluster_table(bs
, guest_offset
, &l2_table
, &l2_index
);
1130 cluster_offset
= be64_to_cpu(l2_table
[l2_index
]);
1132 /* Check how many clusters are already allocated and don't need COW */
1133 if (qcow2_get_cluster_type(cluster_offset
) == QCOW2_CLUSTER_NORMAL
1134 && (cluster_offset
& QCOW_OFLAG_COPIED
))
1136 /* If a specific host_offset is required, check it */
1137 bool offset_matches
=
1138 (cluster_offset
& L2E_OFFSET_MASK
) == *host_offset
;
1140 if (offset_into_cluster(s
, cluster_offset
& L2E_OFFSET_MASK
)) {
1141 qcow2_signal_corruption(bs
, true, -1, -1, "Data cluster offset "
1142 "%#llx unaligned (guest offset: %#" PRIx64
1143 ")", cluster_offset
& L2E_OFFSET_MASK
,
1149 if (*host_offset
!= 0 && !offset_matches
) {
1155 /* We keep all QCOW_OFLAG_COPIED clusters */
1157 count_contiguous_clusters(nb_clusters
, s
->cluster_size
,
1158 &l2_table
[l2_index
],
1159 QCOW_OFLAG_COPIED
| QCOW_OFLAG_ZERO
);
1160 assert(keep_clusters
<= nb_clusters
);
1162 *bytes
= MIN(*bytes
,
1163 keep_clusters
* s
->cluster_size
1164 - offset_into_cluster(s
, guest_offset
));
1173 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
1175 /* Only return a host offset if we actually made progress. Otherwise we
1176 * would make requirements for handle_alloc() that it can't fulfill */
1178 *host_offset
= (cluster_offset
& L2E_OFFSET_MASK
)
1179 + offset_into_cluster(s
, guest_offset
);
1186 * Allocates new clusters for the given guest_offset.
1188 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1189 * contain the number of clusters that have been allocated and are contiguous
1190 * in the image file.
1192 * If *host_offset is non-zero, it specifies the offset in the image file at
1193 * which the new clusters must start. *nb_clusters can be 0 on return in this
1194 * case if the cluster at host_offset is already in use. If *host_offset is
1195 * zero, the clusters can be allocated anywhere in the image file.
1197 * *host_offset is updated to contain the offset into the image file at which
1198 * the first allocated cluster starts.
1200 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1201 * function has been waiting for another request and the allocation must be
1202 * restarted, but the whole request should not be failed.
1204 static int do_alloc_cluster_offset(BlockDriverState
*bs
, uint64_t guest_offset
,
1205 uint64_t *host_offset
, uint64_t *nb_clusters
)
1207 BDRVQcow2State
*s
= bs
->opaque
;
1209 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset
,
1210 *host_offset
, *nb_clusters
);
1212 /* Allocate new clusters */
1213 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1214 if (*host_offset
== 0) {
1215 int64_t cluster_offset
=
1216 qcow2_alloc_clusters(bs
, *nb_clusters
* s
->cluster_size
);
1217 if (cluster_offset
< 0) {
1218 return cluster_offset
;
1220 *host_offset
= cluster_offset
;
1223 int64_t ret
= qcow2_alloc_clusters_at(bs
, *host_offset
, *nb_clusters
);
1233 * Allocates new clusters for an area that either is yet unallocated or needs a
1234 * copy on write. If *host_offset is non-zero, clusters are only allocated if
1235 * the new allocation can match the specified host offset.
1237 * Note that guest_offset may not be cluster aligned. In this case, the
1238 * returned *host_offset points to exact byte referenced by guest_offset and
1239 * therefore isn't cluster aligned as well.
1242 * 0: if no clusters could be allocated. *bytes is set to 0,
1243 * *host_offset is left unchanged.
1245 * 1: if new clusters were allocated. *bytes may be decreased if the
1246 * new allocation doesn't cover all of the requested area.
1247 * *host_offset is updated to contain the host offset of the first
1248 * newly allocated cluster.
1250 * -errno: in error cases
1252 static int handle_alloc(BlockDriverState
*bs
, uint64_t guest_offset
,
1253 uint64_t *host_offset
, uint64_t *bytes
, QCowL2Meta
**m
)
1255 BDRVQcow2State
*s
= bs
->opaque
;
1259 uint64_t nb_clusters
;
1261 bool keep_old_clusters
= false;
1263 uint64_t alloc_cluster_offset
= 0;
1265 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset
, *host_offset
,
1270 * Calculate the number of clusters to look for. We stop at L2 table
1271 * boundaries to keep things simple.
1274 size_to_clusters(s
, offset_into_cluster(s
, guest_offset
) + *bytes
);
1276 l2_index
= offset_to_l2_index(s
, guest_offset
);
1277 nb_clusters
= MIN(nb_clusters
, s
->l2_size
- l2_index
);
1278 assert(nb_clusters
<= INT_MAX
);
1280 /* Find L2 entry for the first involved cluster */
1281 ret
= get_cluster_table(bs
, guest_offset
, &l2_table
, &l2_index
);
1286 entry
= be64_to_cpu(l2_table
[l2_index
]);
1288 /* For the moment, overwrite compressed clusters one by one */
1289 if (entry
& QCOW_OFLAG_COMPRESSED
) {
1292 nb_clusters
= count_cow_clusters(s
, nb_clusters
, l2_table
, l2_index
);
1295 /* This function is only called when there were no non-COW clusters, so if
1296 * we can't find any unallocated or COW clusters either, something is
1297 * wrong with our code. */
1298 assert(nb_clusters
> 0);
1300 if (qcow2_get_cluster_type(entry
) == QCOW2_CLUSTER_ZERO_ALLOC
&&
1301 (entry
& QCOW_OFLAG_COPIED
) &&
1303 start_of_cluster(s
, *host_offset
) == (entry
& L2E_OFFSET_MASK
)))
1305 /* Try to reuse preallocated zero clusters; contiguous normal clusters
1306 * would be fine, too, but count_cow_clusters() above has limited
1307 * nb_clusters already to a range of COW clusters */
1308 int preallocated_nb_clusters
=
1309 count_contiguous_clusters(nb_clusters
, s
->cluster_size
,
1310 &l2_table
[l2_index
], QCOW_OFLAG_COPIED
);
1311 assert(preallocated_nb_clusters
> 0);
1313 nb_clusters
= preallocated_nb_clusters
;
1314 alloc_cluster_offset
= entry
& L2E_OFFSET_MASK
;
1316 /* We want to reuse these clusters, so qcow2_alloc_cluster_link_l2()
1317 * should not free them. */
1318 keep_old_clusters
= true;
1321 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
1323 if (!alloc_cluster_offset
) {
1324 /* Allocate, if necessary at a given offset in the image file */
1325 alloc_cluster_offset
= start_of_cluster(s
, *host_offset
);
1326 ret
= do_alloc_cluster_offset(bs
, guest_offset
, &alloc_cluster_offset
,
1332 /* Can't extend contiguous allocation */
1333 if (nb_clusters
== 0) {
1338 /* !*host_offset would overwrite the image header and is reserved for
1339 * "no host offset preferred". If 0 was a valid host offset, it'd
1340 * trigger the following overlap check; do that now to avoid having an
1341 * invalid value in *host_offset. */
1342 if (!alloc_cluster_offset
) {
1343 ret
= qcow2_pre_write_overlap_check(bs
, 0, alloc_cluster_offset
,
1344 nb_clusters
* s
->cluster_size
);
1351 * Save info needed for meta data update.
1353 * requested_bytes: Number of bytes from the start of the first
1354 * newly allocated cluster to the end of the (possibly shortened
1355 * before) write request.
1357 * avail_bytes: Number of bytes from the start of the first
1358 * newly allocated to the end of the last newly allocated cluster.
1360 * nb_bytes: The number of bytes from the start of the first
1361 * newly allocated cluster to the end of the area that the write
1362 * request actually writes to (excluding COW at the end)
1364 uint64_t requested_bytes
= *bytes
+ offset_into_cluster(s
, guest_offset
);
1365 int avail_bytes
= MIN(INT_MAX
, nb_clusters
<< s
->cluster_bits
);
1366 int nb_bytes
= MIN(requested_bytes
, avail_bytes
);
1367 QCowL2Meta
*old_m
= *m
;
1369 *m
= g_malloc0(sizeof(**m
));
1371 **m
= (QCowL2Meta
) {
1374 .alloc_offset
= alloc_cluster_offset
,
1375 .offset
= start_of_cluster(s
, guest_offset
),
1376 .nb_clusters
= nb_clusters
,
1378 .keep_old_clusters
= keep_old_clusters
,
1382 .nb_bytes
= offset_into_cluster(s
, guest_offset
),
1386 .nb_bytes
= avail_bytes
- nb_bytes
,
1389 qemu_co_queue_init(&(*m
)->dependent_requests
);
1390 QLIST_INSERT_HEAD(&s
->cluster_allocs
, *m
, next_in_flight
);
1392 *host_offset
= alloc_cluster_offset
+ offset_into_cluster(s
, guest_offset
);
1393 *bytes
= MIN(*bytes
, nb_bytes
- offset_into_cluster(s
, guest_offset
));
1394 assert(*bytes
!= 0);
1399 if (*m
&& (*m
)->nb_clusters
> 0) {
1400 QLIST_REMOVE(*m
, next_in_flight
);
1406 * alloc_cluster_offset
1408 * For a given offset on the virtual disk, find the cluster offset in qcow2
1409 * file. If the offset is not found, allocate a new cluster.
1411 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1412 * other fields in m are meaningless.
1414 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1415 * contiguous clusters that have been allocated. In this case, the other
1416 * fields of m are valid and contain information about the first allocated
1419 * If the request conflicts with another write request in flight, the coroutine
1420 * is queued and will be reentered when the dependency has completed.
1422 * Return 0 on success and -errno in error cases
1424 int qcow2_alloc_cluster_offset(BlockDriverState
*bs
, uint64_t offset
,
1425 unsigned int *bytes
, uint64_t *host_offset
,
1428 BDRVQcow2State
*s
= bs
->opaque
;
1429 uint64_t start
, remaining
;
1430 uint64_t cluster_offset
;
1434 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset
, *bytes
);
1446 if (!*host_offset
) {
1447 *host_offset
= start_of_cluster(s
, cluster_offset
);
1450 assert(remaining
>= cur_bytes
);
1453 remaining
-= cur_bytes
;
1454 cluster_offset
+= cur_bytes
;
1456 if (remaining
== 0) {
1460 cur_bytes
= remaining
;
1463 * Now start gathering as many contiguous clusters as possible:
1465 * 1. Check for overlaps with in-flight allocations
1467 * a) Overlap not in the first cluster -> shorten this request and
1468 * let the caller handle the rest in its next loop iteration.
1470 * b) Real overlaps of two requests. Yield and restart the search
1471 * for contiguous clusters (the situation could have changed
1472 * while we were sleeping)
1474 * c) TODO: Request starts in the same cluster as the in-flight
1475 * allocation ends. Shorten the COW of the in-fight allocation,
1476 * set cluster_offset to write to the same cluster and set up
1477 * the right synchronisation between the in-flight request and
1480 ret
= handle_dependencies(bs
, start
, &cur_bytes
, m
);
1481 if (ret
== -EAGAIN
) {
1482 /* Currently handle_dependencies() doesn't yield if we already had
1483 * an allocation. If it did, we would have to clean up the L2Meta
1484 * structs before starting over. */
1487 } else if (ret
< 0) {
1489 } else if (cur_bytes
== 0) {
1492 /* handle_dependencies() may have decreased cur_bytes (shortened
1493 * the allocations below) so that the next dependency is processed
1494 * correctly during the next loop iteration. */
1498 * 2. Count contiguous COPIED clusters.
1500 ret
= handle_copied(bs
, start
, &cluster_offset
, &cur_bytes
, m
);
1505 } else if (cur_bytes
== 0) {
1510 * 3. If the request still hasn't completed, allocate new clusters,
1511 * considering any cluster_offset of steps 1c or 2.
1513 ret
= handle_alloc(bs
, start
, &cluster_offset
, &cur_bytes
, m
);
1519 assert(cur_bytes
== 0);
1524 *bytes
-= remaining
;
1526 assert(*host_offset
!= 0);
1531 static int decompress_buffer(uint8_t *out_buf
, int out_buf_size
,
1532 const uint8_t *buf
, int buf_size
)
1534 z_stream strm1
, *strm
= &strm1
;
1537 memset(strm
, 0, sizeof(*strm
));
1539 strm
->next_in
= (uint8_t *)buf
;
1540 strm
->avail_in
= buf_size
;
1541 strm
->next_out
= out_buf
;
1542 strm
->avail_out
= out_buf_size
;
1544 ret
= inflateInit2(strm
, -12);
1547 ret
= inflate(strm
, Z_FINISH
);
1548 out_len
= strm
->next_out
- out_buf
;
1549 if ((ret
!= Z_STREAM_END
&& ret
!= Z_BUF_ERROR
) ||
1550 out_len
!= out_buf_size
) {
1558 int qcow2_decompress_cluster(BlockDriverState
*bs
, uint64_t cluster_offset
)
1560 BDRVQcow2State
*s
= bs
->opaque
;
1561 int ret
, csize
, nb_csectors
, sector_offset
;
1564 coffset
= cluster_offset
& s
->cluster_offset_mask
;
1565 if (s
->cluster_cache_offset
!= coffset
) {
1566 nb_csectors
= ((cluster_offset
>> s
->csize_shift
) & s
->csize_mask
) + 1;
1567 sector_offset
= coffset
& 511;
1568 csize
= nb_csectors
* 512 - sector_offset
;
1570 /* Allocate buffers on first decompress operation, most images are
1571 * uncompressed and the memory overhead can be avoided. The buffers
1572 * are freed in .bdrv_close().
1574 if (!s
->cluster_data
) {
1575 /* one more sector for decompressed data alignment */
1576 s
->cluster_data
= qemu_try_blockalign(bs
->file
->bs
,
1577 QCOW_MAX_CRYPT_CLUSTERS
* s
->cluster_size
+ 512);
1578 if (!s
->cluster_data
) {
1582 if (!s
->cluster_cache
) {
1583 s
->cluster_cache
= g_malloc(s
->cluster_size
);
1586 BLKDBG_EVENT(bs
->file
, BLKDBG_READ_COMPRESSED
);
1587 ret
= bdrv_read(bs
->file
, coffset
>> 9, s
->cluster_data
,
1592 if (decompress_buffer(s
->cluster_cache
, s
->cluster_size
,
1593 s
->cluster_data
+ sector_offset
, csize
) < 0) {
1596 s
->cluster_cache_offset
= coffset
;
1602 * This discards as many clusters of nb_clusters as possible at once (i.e.
1603 * all clusters in the same L2 table) and returns the number of discarded
1606 static int discard_single_l2(BlockDriverState
*bs
, uint64_t offset
,
1607 uint64_t nb_clusters
, enum qcow2_discard_type type
,
1610 BDRVQcow2State
*s
= bs
->opaque
;
1616 ret
= get_cluster_table(bs
, offset
, &l2_table
, &l2_index
);
1621 /* Limit nb_clusters to one L2 table */
1622 nb_clusters
= MIN(nb_clusters
, s
->l2_size
- l2_index
);
1623 assert(nb_clusters
<= INT_MAX
);
1625 for (i
= 0; i
< nb_clusters
; i
++) {
1626 uint64_t old_l2_entry
;
1628 old_l2_entry
= be64_to_cpu(l2_table
[l2_index
+ i
]);
1631 * If full_discard is false, make sure that a discarded area reads back
1632 * as zeroes for v3 images (we cannot do it for v2 without actually
1633 * writing a zero-filled buffer). We can skip the operation if the
1634 * cluster is already marked as zero, or if it's unallocated and we
1635 * don't have a backing file.
1637 * TODO We might want to use bdrv_get_block_status(bs) here, but we're
1638 * holding s->lock, so that doesn't work today.
1640 * If full_discard is true, the sector should not read back as zeroes,
1641 * but rather fall through to the backing file.
1643 switch (qcow2_get_cluster_type(old_l2_entry
)) {
1644 case QCOW2_CLUSTER_UNALLOCATED
:
1645 if (full_discard
|| !bs
->backing
) {
1650 case QCOW2_CLUSTER_ZERO_PLAIN
:
1651 if (!full_discard
) {
1656 case QCOW2_CLUSTER_ZERO_ALLOC
:
1657 case QCOW2_CLUSTER_NORMAL
:
1658 case QCOW2_CLUSTER_COMPRESSED
:
1665 /* First remove L2 entries */
1666 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
1667 if (!full_discard
&& s
->qcow_version
>= 3) {
1668 l2_table
[l2_index
+ i
] = cpu_to_be64(QCOW_OFLAG_ZERO
);
1670 l2_table
[l2_index
+ i
] = cpu_to_be64(0);
1673 /* Then decrease the refcount */
1674 qcow2_free_any_clusters(bs
, old_l2_entry
, 1, type
);
1677 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
1682 int qcow2_cluster_discard(BlockDriverState
*bs
, uint64_t offset
,
1683 uint64_t bytes
, enum qcow2_discard_type type
,
1686 BDRVQcow2State
*s
= bs
->opaque
;
1687 uint64_t end_offset
= offset
+ bytes
;
1688 uint64_t nb_clusters
;
1692 /* Caller must pass aligned values, except at image end */
1693 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
1694 assert(QEMU_IS_ALIGNED(end_offset
, s
->cluster_size
) ||
1695 end_offset
== bs
->total_sectors
<< BDRV_SECTOR_BITS
);
1697 nb_clusters
= size_to_clusters(s
, bytes
);
1699 s
->cache_discards
= true;
1701 /* Each L2 table is handled by its own loop iteration */
1702 while (nb_clusters
> 0) {
1703 cleared
= discard_single_l2(bs
, offset
, nb_clusters
, type
,
1710 nb_clusters
-= cleared
;
1711 offset
+= (cleared
* s
->cluster_size
);
1716 s
->cache_discards
= false;
1717 qcow2_process_discards(bs
, ret
);
1723 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1724 * all clusters in the same L2 table) and returns the number of zeroed
1727 static int zero_single_l2(BlockDriverState
*bs
, uint64_t offset
,
1728 uint64_t nb_clusters
, int flags
)
1730 BDRVQcow2State
*s
= bs
->opaque
;
1735 bool unmap
= !!(flags
& BDRV_REQ_MAY_UNMAP
);
1737 ret
= get_cluster_table(bs
, offset
, &l2_table
, &l2_index
);
1742 /* Limit nb_clusters to one L2 table */
1743 nb_clusters
= MIN(nb_clusters
, s
->l2_size
- l2_index
);
1744 assert(nb_clusters
<= INT_MAX
);
1746 for (i
= 0; i
< nb_clusters
; i
++) {
1747 uint64_t old_offset
;
1748 QCow2ClusterType cluster_type
;
1750 old_offset
= be64_to_cpu(l2_table
[l2_index
+ i
]);
1753 * Minimize L2 changes if the cluster already reads back as
1754 * zeroes with correct allocation.
1756 cluster_type
= qcow2_get_cluster_type(old_offset
);
1757 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
||
1758 (cluster_type
== QCOW2_CLUSTER_ZERO_ALLOC
&& !unmap
)) {
1762 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
1763 if (cluster_type
== QCOW2_CLUSTER_COMPRESSED
|| unmap
) {
1764 l2_table
[l2_index
+ i
] = cpu_to_be64(QCOW_OFLAG_ZERO
);
1765 qcow2_free_any_clusters(bs
, old_offset
, 1, QCOW2_DISCARD_REQUEST
);
1767 l2_table
[l2_index
+ i
] |= cpu_to_be64(QCOW_OFLAG_ZERO
);
1771 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
1776 int qcow2_cluster_zeroize(BlockDriverState
*bs
, uint64_t offset
,
1777 uint64_t bytes
, int flags
)
1779 BDRVQcow2State
*s
= bs
->opaque
;
1780 uint64_t end_offset
= offset
+ bytes
;
1781 uint64_t nb_clusters
;
1785 /* Caller must pass aligned values, except at image end */
1786 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
1787 assert(QEMU_IS_ALIGNED(end_offset
, s
->cluster_size
) ||
1788 end_offset
== bs
->total_sectors
<< BDRV_SECTOR_BITS
);
1790 /* The zero flag is only supported by version 3 and newer */
1791 if (s
->qcow_version
< 3) {
1795 /* Each L2 table is handled by its own loop iteration */
1796 nb_clusters
= size_to_clusters(s
, bytes
);
1798 s
->cache_discards
= true;
1800 while (nb_clusters
> 0) {
1801 cleared
= zero_single_l2(bs
, offset
, nb_clusters
, flags
);
1807 nb_clusters
-= cleared
;
1808 offset
+= (cleared
* s
->cluster_size
);
1813 s
->cache_discards
= false;
1814 qcow2_process_discards(bs
, ret
);
1820 * Expands all zero clusters in a specific L1 table (or deallocates them, for
1821 * non-backed non-pre-allocated zero clusters).
1823 * l1_entries and *visited_l1_entries are used to keep track of progress for
1824 * status_cb(). l1_entries contains the total number of L1 entries and
1825 * *visited_l1_entries counts all visited L1 entries.
1827 static int expand_zero_clusters_in_l1(BlockDriverState
*bs
, uint64_t *l1_table
,
1828 int l1_size
, int64_t *visited_l1_entries
,
1830 BlockDriverAmendStatusCB
*status_cb
,
1833 BDRVQcow2State
*s
= bs
->opaque
;
1834 bool is_active_l1
= (l1_table
== s
->l1_table
);
1835 uint64_t *l2_table
= NULL
;
1839 if (!is_active_l1
) {
1840 /* inactive L2 tables require a buffer to be stored in when loading
1842 l2_table
= qemu_try_blockalign(bs
->file
->bs
, s
->cluster_size
);
1843 if (l2_table
== NULL
) {
1848 for (i
= 0; i
< l1_size
; i
++) {
1849 uint64_t l2_offset
= l1_table
[i
] & L1E_OFFSET_MASK
;
1850 bool l2_dirty
= false;
1851 uint64_t l2_refcount
;
1855 (*visited_l1_entries
)++;
1857 status_cb(bs
, *visited_l1_entries
, l1_entries
, cb_opaque
);
1862 if (offset_into_cluster(s
, l2_offset
)) {
1863 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#"
1864 PRIx64
" unaligned (L1 index: %#x)",
1871 /* get active L2 tables from cache */
1872 ret
= qcow2_cache_get(bs
, s
->l2_table_cache
, l2_offset
,
1873 (void **)&l2_table
);
1875 /* load inactive L2 tables from disk */
1876 ret
= bdrv_read(bs
->file
, l2_offset
/ BDRV_SECTOR_SIZE
,
1877 (void *)l2_table
, s
->cluster_sectors
);
1883 ret
= qcow2_get_refcount(bs
, l2_offset
>> s
->cluster_bits
,
1889 for (j
= 0; j
< s
->l2_size
; j
++) {
1890 uint64_t l2_entry
= be64_to_cpu(l2_table
[j
]);
1891 int64_t offset
= l2_entry
& L2E_OFFSET_MASK
;
1892 QCow2ClusterType cluster_type
= qcow2_get_cluster_type(l2_entry
);
1894 if (cluster_type
!= QCOW2_CLUSTER_ZERO_PLAIN
&&
1895 cluster_type
!= QCOW2_CLUSTER_ZERO_ALLOC
) {
1899 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1901 /* not backed; therefore we can simply deallocate the
1908 offset
= qcow2_alloc_clusters(bs
, s
->cluster_size
);
1914 if (l2_refcount
> 1) {
1915 /* For shared L2 tables, set the refcount accordingly (it is
1916 * already 1 and needs to be l2_refcount) */
1917 ret
= qcow2_update_cluster_refcount(bs
,
1918 offset
>> s
->cluster_bits
,
1919 refcount_diff(1, l2_refcount
), false,
1920 QCOW2_DISCARD_OTHER
);
1922 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1923 QCOW2_DISCARD_OTHER
);
1929 if (offset_into_cluster(s
, offset
)) {
1930 qcow2_signal_corruption(bs
, true, -1, -1,
1931 "Cluster allocation offset "
1932 "%#" PRIx64
" unaligned (L2 offset: %#"
1933 PRIx64
", L2 index: %#x)", offset
,
1935 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1936 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1937 QCOW2_DISCARD_ALWAYS
);
1943 ret
= qcow2_pre_write_overlap_check(bs
, 0, offset
, s
->cluster_size
);
1945 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1946 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1947 QCOW2_DISCARD_ALWAYS
);
1952 ret
= bdrv_pwrite_zeroes(bs
->file
, offset
, s
->cluster_size
, 0);
1954 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1955 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1956 QCOW2_DISCARD_ALWAYS
);
1961 if (l2_refcount
== 1) {
1962 l2_table
[j
] = cpu_to_be64(offset
| QCOW_OFLAG_COPIED
);
1964 l2_table
[j
] = cpu_to_be64(offset
);
1971 qcow2_cache_entry_mark_dirty(bs
, s
->l2_table_cache
, l2_table
);
1972 qcow2_cache_depends_on_flush(s
->l2_table_cache
);
1974 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
1977 ret
= qcow2_pre_write_overlap_check(bs
,
1978 QCOW2_OL_INACTIVE_L2
| QCOW2_OL_ACTIVE_L2
, l2_offset
,
1984 ret
= bdrv_write(bs
->file
, l2_offset
/ BDRV_SECTOR_SIZE
,
1985 (void *)l2_table
, s
->cluster_sectors
);
1992 (*visited_l1_entries
)++;
1994 status_cb(bs
, *visited_l1_entries
, l1_entries
, cb_opaque
);
2002 if (!is_active_l1
) {
2003 qemu_vfree(l2_table
);
2005 qcow2_cache_put(bs
, s
->l2_table_cache
, (void **) &l2_table
);
2012 * For backed images, expands all zero clusters on the image. For non-backed
2013 * images, deallocates all non-pre-allocated zero clusters (and claims the
2014 * allocation for pre-allocated ones). This is important for downgrading to a
2015 * qcow2 version which doesn't yet support metadata zero clusters.
2017 int qcow2_expand_zero_clusters(BlockDriverState
*bs
,
2018 BlockDriverAmendStatusCB
*status_cb
,
2021 BDRVQcow2State
*s
= bs
->opaque
;
2022 uint64_t *l1_table
= NULL
;
2023 int64_t l1_entries
= 0, visited_l1_entries
= 0;
2028 l1_entries
= s
->l1_size
;
2029 for (i
= 0; i
< s
->nb_snapshots
; i
++) {
2030 l1_entries
+= s
->snapshots
[i
].l1_size
;
2034 ret
= expand_zero_clusters_in_l1(bs
, s
->l1_table
, s
->l1_size
,
2035 &visited_l1_entries
, l1_entries
,
2036 status_cb
, cb_opaque
);
2041 /* Inactive L1 tables may point to active L2 tables - therefore it is
2042 * necessary to flush the L2 table cache before trying to access the L2
2043 * tables pointed to by inactive L1 entries (else we might try to expand
2044 * zero clusters that have already been expanded); furthermore, it is also
2045 * necessary to empty the L2 table cache, since it may contain tables which
2046 * are now going to be modified directly on disk, bypassing the cache.
2047 * qcow2_cache_empty() does both for us. */
2048 ret
= qcow2_cache_empty(bs
, s
->l2_table_cache
);
2053 for (i
= 0; i
< s
->nb_snapshots
; i
++) {
2054 int l1_sectors
= DIV_ROUND_UP(s
->snapshots
[i
].l1_size
*
2055 sizeof(uint64_t), BDRV_SECTOR_SIZE
);
2057 l1_table
= g_realloc(l1_table
, l1_sectors
* BDRV_SECTOR_SIZE
);
2059 ret
= bdrv_read(bs
->file
,
2060 s
->snapshots
[i
].l1_table_offset
/ BDRV_SECTOR_SIZE
,
2061 (void *)l1_table
, l1_sectors
);
2066 for (j
= 0; j
< s
->snapshots
[i
].l1_size
; j
++) {
2067 be64_to_cpus(&l1_table
[j
]);
2070 ret
= expand_zero_clusters_in_l1(bs
, l1_table
, s
->snapshots
[i
].l1_size
,
2071 &visited_l1_entries
, l1_entries
,
2072 status_cb
, cb_opaque
);