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"
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 ROUND_UP(new_l1_size2
, 512));
131 if (new_l1_table
== NULL
) {
134 memset(new_l1_table
, 0, ROUND_UP(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
,
156 new_l1_size2
, false);
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 * @bs: The BlockDriverState
200 * @offset: A guest offset, used to calculate what slice of the L2
202 * @l2_offset: Offset to the L2 table in the image file.
203 * @l2_slice: Location to store the pointer to the L2 slice.
205 * Loads a L2 slice into memory (L2 slices are the parts of L2 tables
206 * that are loaded by the qcow2 cache). If the slice is in the cache,
207 * the cache is used; otherwise the L2 slice is loaded from the image
210 static int l2_load(BlockDriverState
*bs
, uint64_t offset
,
211 uint64_t l2_offset
, uint64_t **l2_slice
)
213 BDRVQcow2State
*s
= bs
->opaque
;
214 int start_of_slice
= sizeof(uint64_t) *
215 (offset_to_l2_index(s
, offset
) - offset_to_l2_slice_index(s
, offset
));
217 return qcow2_cache_get(bs
, s
->l2_table_cache
, l2_offset
+ start_of_slice
,
222 * Writes one sector of the L1 table to the disk (can't update single entries
223 * and we really don't want bdrv_pread to perform a read-modify-write)
225 #define L1_ENTRIES_PER_SECTOR (512 / 8)
226 int qcow2_write_l1_entry(BlockDriverState
*bs
, int l1_index
)
228 BDRVQcow2State
*s
= bs
->opaque
;
229 uint64_t buf
[L1_ENTRIES_PER_SECTOR
] = { 0 };
233 l1_start_index
= l1_index
& ~(L1_ENTRIES_PER_SECTOR
- 1);
234 for (i
= 0; i
< L1_ENTRIES_PER_SECTOR
&& l1_start_index
+ i
< s
->l1_size
;
237 buf
[i
] = cpu_to_be64(s
->l1_table
[l1_start_index
+ i
]);
240 ret
= qcow2_pre_write_overlap_check(bs
, QCOW2_OL_ACTIVE_L1
,
241 s
->l1_table_offset
+ 8 * l1_start_index
, sizeof(buf
), false);
246 BLKDBG_EVENT(bs
->file
, BLKDBG_L1_UPDATE
);
247 ret
= bdrv_pwrite_sync(bs
->file
,
248 s
->l1_table_offset
+ 8 * l1_start_index
,
260 * Allocate a new l2 entry in the file. If l1_index points to an already
261 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
262 * table) copy the contents of the old L2 table into the newly allocated one.
263 * Otherwise the new table is initialized with zeros.
267 static int l2_allocate(BlockDriverState
*bs
, int l1_index
)
269 BDRVQcow2State
*s
= bs
->opaque
;
270 uint64_t old_l2_offset
;
271 uint64_t *l2_slice
= NULL
;
272 unsigned slice
, slice_size2
, n_slices
;
276 old_l2_offset
= s
->l1_table
[l1_index
];
278 trace_qcow2_l2_allocate(bs
, l1_index
);
280 /* allocate a new l2 entry */
282 l2_offset
= qcow2_alloc_clusters(bs
, s
->l2_size
* sizeof(uint64_t));
288 /* The offset must fit in the offset field of the L1 table entry */
289 assert((l2_offset
& L1E_OFFSET_MASK
) == l2_offset
);
291 /* If we're allocating the table at offset 0 then something is wrong */
292 if (l2_offset
== 0) {
293 qcow2_signal_corruption(bs
, true, -1, -1, "Preventing invalid "
294 "allocation of L2 table at offset 0");
299 ret
= qcow2_cache_flush(bs
, s
->refcount_block_cache
);
304 /* allocate a new entry in the l2 cache */
306 slice_size2
= s
->l2_slice_size
* sizeof(uint64_t);
307 n_slices
= s
->cluster_size
/ slice_size2
;
309 trace_qcow2_l2_allocate_get_empty(bs
, l1_index
);
310 for (slice
= 0; slice
< n_slices
; slice
++) {
311 ret
= qcow2_cache_get_empty(bs
, s
->l2_table_cache
,
312 l2_offset
+ slice
* slice_size2
,
313 (void **) &l2_slice
);
318 if ((old_l2_offset
& L1E_OFFSET_MASK
) == 0) {
319 /* if there was no old l2 table, clear the new slice */
320 memset(l2_slice
, 0, slice_size2
);
323 uint64_t old_l2_slice_offset
=
324 (old_l2_offset
& L1E_OFFSET_MASK
) + slice
* slice_size2
;
326 /* if there was an old l2 table, read a slice from the disk */
327 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_ALLOC_COW_READ
);
328 ret
= qcow2_cache_get(bs
, s
->l2_table_cache
, old_l2_slice_offset
,
329 (void **) &old_slice
);
334 memcpy(l2_slice
, old_slice
, slice_size2
);
336 qcow2_cache_put(s
->l2_table_cache
, (void **) &old_slice
);
339 /* write the l2 slice to the file */
340 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_ALLOC_WRITE
);
342 trace_qcow2_l2_allocate_write_l2(bs
, l1_index
);
343 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
344 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
347 ret
= qcow2_cache_flush(bs
, s
->l2_table_cache
);
352 /* update the L1 entry */
353 trace_qcow2_l2_allocate_write_l1(bs
, l1_index
);
354 s
->l1_table
[l1_index
] = l2_offset
| QCOW_OFLAG_COPIED
;
355 ret
= qcow2_write_l1_entry(bs
, l1_index
);
360 trace_qcow2_l2_allocate_done(bs
, l1_index
, 0);
364 trace_qcow2_l2_allocate_done(bs
, l1_index
, ret
);
365 if (l2_slice
!= NULL
) {
366 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
368 s
->l1_table
[l1_index
] = old_l2_offset
;
370 qcow2_free_clusters(bs
, l2_offset
, s
->l2_size
* sizeof(uint64_t),
371 QCOW2_DISCARD_ALWAYS
);
377 * Checks how many clusters in a given L2 slice are contiguous in the image
378 * file. As soon as one of the flags in the bitmask stop_flags changes compared
379 * to the first cluster, the search is stopped and the cluster is not counted
380 * as contiguous. (This allows it, for example, to stop at the first compressed
381 * cluster which may require a different handling)
383 static int count_contiguous_clusters(BlockDriverState
*bs
, int nb_clusters
,
384 int cluster_size
, uint64_t *l2_slice
, uint64_t stop_flags
)
387 QCow2ClusterType first_cluster_type
;
388 uint64_t mask
= stop_flags
| L2E_OFFSET_MASK
| QCOW_OFLAG_COMPRESSED
;
389 uint64_t first_entry
= be64_to_cpu(l2_slice
[0]);
390 uint64_t offset
= first_entry
& mask
;
392 first_cluster_type
= qcow2_get_cluster_type(bs
, first_entry
);
393 if (first_cluster_type
== QCOW2_CLUSTER_UNALLOCATED
) {
397 /* must be allocated */
398 assert(first_cluster_type
== QCOW2_CLUSTER_NORMAL
||
399 first_cluster_type
== QCOW2_CLUSTER_ZERO_ALLOC
);
401 for (i
= 0; i
< nb_clusters
; i
++) {
402 uint64_t l2_entry
= be64_to_cpu(l2_slice
[i
]) & mask
;
403 if (offset
+ (uint64_t) i
* cluster_size
!= l2_entry
) {
412 * Checks how many consecutive unallocated clusters in a given L2
413 * slice have the same cluster type.
415 static int count_contiguous_clusters_unallocated(BlockDriverState
*bs
,
418 QCow2ClusterType wanted_type
)
422 assert(wanted_type
== QCOW2_CLUSTER_ZERO_PLAIN
||
423 wanted_type
== QCOW2_CLUSTER_UNALLOCATED
);
424 for (i
= 0; i
< nb_clusters
; i
++) {
425 uint64_t entry
= be64_to_cpu(l2_slice
[i
]);
426 QCow2ClusterType type
= qcow2_get_cluster_type(bs
, entry
);
428 if (type
!= wanted_type
) {
436 static int coroutine_fn
do_perform_cow_read(BlockDriverState
*bs
,
437 uint64_t src_cluster_offset
,
438 unsigned offset_in_cluster
,
443 if (qiov
->size
== 0) {
447 BLKDBG_EVENT(bs
->file
, BLKDBG_COW_READ
);
453 /* Call .bdrv_co_readv() directly instead of using the public block-layer
454 * interface. This avoids double I/O throttling and request tracking,
455 * which can lead to deadlock when block layer copy-on-read is enabled.
457 ret
= bs
->drv
->bdrv_co_preadv(bs
, src_cluster_offset
+ offset_in_cluster
,
458 qiov
->size
, qiov
, 0);
466 static bool coroutine_fn
do_perform_cow_encrypt(BlockDriverState
*bs
,
467 uint64_t src_cluster_offset
,
468 uint64_t cluster_offset
,
469 unsigned offset_in_cluster
,
473 if (bytes
&& bs
->encrypted
) {
474 BDRVQcow2State
*s
= bs
->opaque
;
475 int64_t offset
= (s
->crypt_physical_offset
?
476 (cluster_offset
+ offset_in_cluster
) :
477 (src_cluster_offset
+ offset_in_cluster
));
478 assert((offset_in_cluster
& ~BDRV_SECTOR_MASK
) == 0);
479 assert((bytes
& ~BDRV_SECTOR_MASK
) == 0);
481 if (qcrypto_block_encrypt(s
->crypto
, offset
, buffer
, bytes
, NULL
) < 0) {
488 static int coroutine_fn
do_perform_cow_write(BlockDriverState
*bs
,
489 uint64_t cluster_offset
,
490 unsigned offset_in_cluster
,
493 BDRVQcow2State
*s
= bs
->opaque
;
496 if (qiov
->size
== 0) {
500 ret
= qcow2_pre_write_overlap_check(bs
, 0,
501 cluster_offset
+ offset_in_cluster
, qiov
->size
, true);
506 BLKDBG_EVENT(bs
->file
, BLKDBG_COW_WRITE
);
507 ret
= bdrv_co_pwritev(s
->data_file
, cluster_offset
+ offset_in_cluster
,
508 qiov
->size
, qiov
, 0);
520 * For a given offset of the virtual disk, find the cluster type and offset in
521 * the qcow2 file. The offset is stored in *cluster_offset.
523 * On entry, *bytes is the maximum number of contiguous bytes starting at
524 * offset that we are interested in.
526 * On exit, *bytes is the number of bytes starting at offset that have the same
527 * cluster type and (if applicable) are stored contiguously in the image file.
528 * Compressed clusters are always returned one by one.
530 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
533 int qcow2_get_cluster_offset(BlockDriverState
*bs
, uint64_t offset
,
534 unsigned int *bytes
, uint64_t *cluster_offset
)
536 BDRVQcow2State
*s
= bs
->opaque
;
537 unsigned int l2_index
;
538 uint64_t l1_index
, l2_offset
, *l2_slice
;
540 unsigned int offset_in_cluster
;
541 uint64_t bytes_available
, bytes_needed
, nb_clusters
;
542 QCow2ClusterType type
;
545 offset_in_cluster
= offset_into_cluster(s
, offset
);
546 bytes_needed
= (uint64_t) *bytes
+ offset_in_cluster
;
548 /* compute how many bytes there are between the start of the cluster
549 * containing offset and the end of the l2 slice that contains
550 * the entry pointing to it */
552 ((uint64_t) (s
->l2_slice_size
- offset_to_l2_slice_index(s
, offset
)))
555 if (bytes_needed
> bytes_available
) {
556 bytes_needed
= bytes_available
;
561 /* seek to the l2 offset in the l1 table */
563 l1_index
= offset_to_l1_index(s
, offset
);
564 if (l1_index
>= s
->l1_size
) {
565 type
= QCOW2_CLUSTER_UNALLOCATED
;
569 l2_offset
= s
->l1_table
[l1_index
] & L1E_OFFSET_MASK
;
571 type
= QCOW2_CLUSTER_UNALLOCATED
;
575 if (offset_into_cluster(s
, l2_offset
)) {
576 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#" PRIx64
577 " unaligned (L1 index: %#" PRIx64
")",
578 l2_offset
, l1_index
);
582 /* load the l2 slice in memory */
584 ret
= l2_load(bs
, offset
, l2_offset
, &l2_slice
);
589 /* find the cluster offset for the given disk offset */
591 l2_index
= offset_to_l2_slice_index(s
, offset
);
592 *cluster_offset
= be64_to_cpu(l2_slice
[l2_index
]);
594 nb_clusters
= size_to_clusters(s
, bytes_needed
);
595 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
596 * integers; the minimum cluster size is 512, so this assertion is always
598 assert(nb_clusters
<= INT_MAX
);
600 type
= qcow2_get_cluster_type(bs
, *cluster_offset
);
601 if (s
->qcow_version
< 3 && (type
== QCOW2_CLUSTER_ZERO_PLAIN
||
602 type
== QCOW2_CLUSTER_ZERO_ALLOC
)) {
603 qcow2_signal_corruption(bs
, true, -1, -1, "Zero cluster entry found"
604 " in pre-v3 image (L2 offset: %#" PRIx64
605 ", L2 index: %#x)", l2_offset
, l2_index
);
610 case QCOW2_CLUSTER_COMPRESSED
:
611 if (has_data_file(bs
)) {
612 qcow2_signal_corruption(bs
, true, -1, -1, "Compressed cluster "
613 "entry found in image with external data "
614 "file (L2 offset: %#" PRIx64
", L2 index: "
615 "%#x)", l2_offset
, l2_index
);
619 /* Compressed clusters can only be processed one by one */
621 *cluster_offset
&= L2E_COMPRESSED_OFFSET_SIZE_MASK
;
623 case QCOW2_CLUSTER_ZERO_PLAIN
:
624 case QCOW2_CLUSTER_UNALLOCATED
:
625 /* how many empty clusters ? */
626 c
= count_contiguous_clusters_unallocated(bs
, nb_clusters
,
627 &l2_slice
[l2_index
], type
);
630 case QCOW2_CLUSTER_ZERO_ALLOC
:
631 case QCOW2_CLUSTER_NORMAL
:
632 /* how many allocated clusters ? */
633 c
= count_contiguous_clusters(bs
, nb_clusters
, s
->cluster_size
,
634 &l2_slice
[l2_index
], QCOW_OFLAG_ZERO
);
635 *cluster_offset
&= L2E_OFFSET_MASK
;
636 if (offset_into_cluster(s
, *cluster_offset
)) {
637 qcow2_signal_corruption(bs
, true, -1, -1,
638 "Cluster allocation offset %#"
639 PRIx64
" unaligned (L2 offset: %#" PRIx64
640 ", L2 index: %#x)", *cluster_offset
,
641 l2_offset
, l2_index
);
645 if (has_data_file(bs
) && *cluster_offset
!= offset
- offset_in_cluster
)
647 qcow2_signal_corruption(bs
, true, -1, -1,
648 "External data file host cluster offset %#"
649 PRIx64
" does not match guest cluster "
651 ", L2 index: %#x)", *cluster_offset
,
652 offset
- offset_in_cluster
, l2_index
);
661 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
663 bytes_available
= (int64_t)c
* s
->cluster_size
;
666 if (bytes_available
> bytes_needed
) {
667 bytes_available
= bytes_needed
;
670 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
671 * subtracting offset_in_cluster will therefore definitely yield something
672 * not exceeding UINT_MAX */
673 assert(bytes_available
- offset_in_cluster
<= UINT_MAX
);
674 *bytes
= bytes_available
- offset_in_cluster
;
679 qcow2_cache_put(s
->l2_table_cache
, (void **)&l2_slice
);
686 * for a given disk offset, load (and allocate if needed)
687 * the appropriate slice of its l2 table.
689 * the cluster index in the l2 slice is given to the caller.
691 * Returns 0 on success, -errno in failure case
693 static int get_cluster_table(BlockDriverState
*bs
, uint64_t offset
,
694 uint64_t **new_l2_slice
,
697 BDRVQcow2State
*s
= bs
->opaque
;
698 unsigned int l2_index
;
699 uint64_t l1_index
, l2_offset
;
700 uint64_t *l2_slice
= NULL
;
703 /* seek to the l2 offset in the l1 table */
705 l1_index
= offset_to_l1_index(s
, offset
);
706 if (l1_index
>= s
->l1_size
) {
707 ret
= qcow2_grow_l1_table(bs
, l1_index
+ 1, false);
713 assert(l1_index
< s
->l1_size
);
714 l2_offset
= s
->l1_table
[l1_index
] & L1E_OFFSET_MASK
;
715 if (offset_into_cluster(s
, l2_offset
)) {
716 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#" PRIx64
717 " unaligned (L1 index: %#" PRIx64
")",
718 l2_offset
, l1_index
);
722 if (!(s
->l1_table
[l1_index
] & QCOW_OFLAG_COPIED
)) {
723 /* First allocate a new L2 table (and do COW if needed) */
724 ret
= l2_allocate(bs
, l1_index
);
729 /* Then decrease the refcount of the old table */
731 qcow2_free_clusters(bs
, l2_offset
, s
->l2_size
* sizeof(uint64_t),
732 QCOW2_DISCARD_OTHER
);
735 /* Get the offset of the newly-allocated l2 table */
736 l2_offset
= s
->l1_table
[l1_index
] & L1E_OFFSET_MASK
;
737 assert(offset_into_cluster(s
, l2_offset
) == 0);
740 /* load the l2 slice in memory */
741 ret
= l2_load(bs
, offset
, l2_offset
, &l2_slice
);
746 /* find the cluster offset for the given disk offset */
748 l2_index
= offset_to_l2_slice_index(s
, offset
);
750 *new_l2_slice
= l2_slice
;
751 *new_l2_index
= l2_index
;
757 * alloc_compressed_cluster_offset
759 * For a given offset on the virtual disk, allocate a new compressed cluster
760 * and put the host offset of the cluster into *host_offset. If a cluster is
761 * already allocated at the offset, return an error.
763 * Return 0 on success and -errno in error cases
765 int qcow2_alloc_compressed_cluster_offset(BlockDriverState
*bs
,
768 uint64_t *host_offset
)
770 BDRVQcow2State
*s
= bs
->opaque
;
773 int64_t cluster_offset
;
776 if (has_data_file(bs
)) {
780 ret
= get_cluster_table(bs
, offset
, &l2_slice
, &l2_index
);
785 /* Compression can't overwrite anything. Fail if the cluster was already
787 cluster_offset
= be64_to_cpu(l2_slice
[l2_index
]);
788 if (cluster_offset
& L2E_OFFSET_MASK
) {
789 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
793 cluster_offset
= qcow2_alloc_bytes(bs
, compressed_size
);
794 if (cluster_offset
< 0) {
795 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
796 return cluster_offset
;
800 (cluster_offset
+ compressed_size
- 1) / QCOW2_COMPRESSED_SECTOR_SIZE
-
801 (cluster_offset
/ QCOW2_COMPRESSED_SECTOR_SIZE
);
803 cluster_offset
|= QCOW_OFLAG_COMPRESSED
|
804 ((uint64_t)nb_csectors
<< s
->csize_shift
);
806 /* update L2 table */
808 /* compressed clusters never have the copied flag */
810 BLKDBG_EVENT(bs
->file
, BLKDBG_L2_UPDATE_COMPRESSED
);
811 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
812 l2_slice
[l2_index
] = cpu_to_be64(cluster_offset
);
813 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
815 *host_offset
= cluster_offset
& s
->cluster_offset_mask
;
819 static int perform_cow(BlockDriverState
*bs
, QCowL2Meta
*m
)
821 BDRVQcow2State
*s
= bs
->opaque
;
822 Qcow2COWRegion
*start
= &m
->cow_start
;
823 Qcow2COWRegion
*end
= &m
->cow_end
;
824 unsigned buffer_size
;
825 unsigned data_bytes
= end
->offset
- (start
->offset
+ start
->nb_bytes
);
827 uint8_t *start_buffer
, *end_buffer
;
831 assert(start
->nb_bytes
<= UINT_MAX
- end
->nb_bytes
);
832 assert(start
->nb_bytes
+ end
->nb_bytes
<= UINT_MAX
- data_bytes
);
833 assert(start
->offset
+ start
->nb_bytes
<= end
->offset
);
834 assert(!m
->data_qiov
|| m
->data_qiov
->size
== data_bytes
);
836 if (start
->nb_bytes
== 0 && end
->nb_bytes
== 0) {
840 /* If we have to read both the start and end COW regions and the
841 * middle region is not too large then perform just one read
843 merge_reads
= start
->nb_bytes
&& end
->nb_bytes
&& data_bytes
<= 16384;
845 buffer_size
= start
->nb_bytes
+ data_bytes
+ end
->nb_bytes
;
847 /* If we have to do two reads, add some padding in the middle
848 * if necessary to make sure that the end region is optimally
850 size_t align
= bdrv_opt_mem_align(bs
);
851 assert(align
> 0 && align
<= UINT_MAX
);
852 assert(QEMU_ALIGN_UP(start
->nb_bytes
, align
) <=
853 UINT_MAX
- end
->nb_bytes
);
854 buffer_size
= QEMU_ALIGN_UP(start
->nb_bytes
, align
) + end
->nb_bytes
;
857 /* Reserve a buffer large enough to store all the data that we're
859 start_buffer
= qemu_try_blockalign(bs
, buffer_size
);
860 if (start_buffer
== NULL
) {
863 /* The part of the buffer where the end region is located */
864 end_buffer
= start_buffer
+ buffer_size
- end
->nb_bytes
;
866 qemu_iovec_init(&qiov
, 2 + (m
->data_qiov
? m
->data_qiov
->niov
: 0));
868 qemu_co_mutex_unlock(&s
->lock
);
869 /* First we read the existing data from both COW regions. We
870 * either read the whole region in one go, or the start and end
871 * regions separately. */
873 qemu_iovec_add(&qiov
, start_buffer
, buffer_size
);
874 ret
= do_perform_cow_read(bs
, m
->offset
, start
->offset
, &qiov
);
876 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
877 ret
= do_perform_cow_read(bs
, m
->offset
, start
->offset
, &qiov
);
882 qemu_iovec_reset(&qiov
);
883 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
884 ret
= do_perform_cow_read(bs
, m
->offset
, end
->offset
, &qiov
);
890 /* Encrypt the data if necessary before writing it */
892 if (!do_perform_cow_encrypt(bs
, m
->offset
, m
->alloc_offset
,
893 start
->offset
, start_buffer
,
895 !do_perform_cow_encrypt(bs
, m
->offset
, m
->alloc_offset
,
896 end
->offset
, end_buffer
, end
->nb_bytes
)) {
902 /* And now we can write everything. If we have the guest data we
903 * can write everything in one single operation */
905 qemu_iovec_reset(&qiov
);
906 if (start
->nb_bytes
) {
907 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
909 qemu_iovec_concat(&qiov
, m
->data_qiov
, 0, data_bytes
);
911 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
913 /* NOTE: we have a write_aio blkdebug event here followed by
914 * a cow_write one in do_perform_cow_write(), but there's only
915 * one single I/O operation */
916 BLKDBG_EVENT(bs
->file
, BLKDBG_WRITE_AIO
);
917 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, start
->offset
, &qiov
);
919 /* If there's no guest data then write both COW regions separately */
920 qemu_iovec_reset(&qiov
);
921 qemu_iovec_add(&qiov
, start_buffer
, start
->nb_bytes
);
922 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, start
->offset
, &qiov
);
927 qemu_iovec_reset(&qiov
);
928 qemu_iovec_add(&qiov
, end_buffer
, end
->nb_bytes
);
929 ret
= do_perform_cow_write(bs
, m
->alloc_offset
, end
->offset
, &qiov
);
933 qemu_co_mutex_lock(&s
->lock
);
936 * Before we update the L2 table to actually point to the new cluster, we
937 * need to be sure that the refcounts have been increased and COW was
941 qcow2_cache_depends_on_flush(s
->l2_table_cache
);
944 qemu_vfree(start_buffer
);
945 qemu_iovec_destroy(&qiov
);
949 int qcow2_alloc_cluster_link_l2(BlockDriverState
*bs
, QCowL2Meta
*m
)
951 BDRVQcow2State
*s
= bs
->opaque
;
952 int i
, j
= 0, l2_index
, ret
;
953 uint64_t *old_cluster
, *l2_slice
;
954 uint64_t cluster_offset
= m
->alloc_offset
;
956 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m
->nb_clusters
);
957 assert(m
->nb_clusters
> 0);
959 old_cluster
= g_try_new(uint64_t, m
->nb_clusters
);
960 if (old_cluster
== NULL
) {
965 /* copy content of unmodified sectors */
966 ret
= perform_cow(bs
, m
);
971 /* Update L2 table. */
972 if (s
->use_lazy_refcounts
) {
973 qcow2_mark_dirty(bs
);
975 if (qcow2_need_accurate_refcounts(s
)) {
976 qcow2_cache_set_dependency(bs
, s
->l2_table_cache
,
977 s
->refcount_block_cache
);
980 ret
= get_cluster_table(bs
, m
->offset
, &l2_slice
, &l2_index
);
984 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
986 assert(l2_index
+ m
->nb_clusters
<= s
->l2_slice_size
);
987 for (i
= 0; i
< m
->nb_clusters
; i
++) {
988 /* if two concurrent writes happen to the same unallocated cluster
989 * each write allocates separate cluster and writes data concurrently.
990 * The first one to complete updates l2 table with pointer to its
991 * cluster the second one has to do RMW (which is done above by
992 * perform_cow()), update l2 table with its cluster pointer and free
993 * old cluster. This is what this loop does */
994 if (l2_slice
[l2_index
+ i
] != 0) {
995 old_cluster
[j
++] = l2_slice
[l2_index
+ i
];
998 l2_slice
[l2_index
+ i
] = cpu_to_be64((cluster_offset
+
999 (i
<< s
->cluster_bits
)) | QCOW_OFLAG_COPIED
);
1003 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1006 * If this was a COW, we need to decrease the refcount of the old cluster.
1008 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1009 * clusters), the next write will reuse them anyway.
1011 if (!m
->keep_old_clusters
&& j
!= 0) {
1012 for (i
= 0; i
< j
; i
++) {
1013 qcow2_free_any_clusters(bs
, be64_to_cpu(old_cluster
[i
]), 1,
1014 QCOW2_DISCARD_NEVER
);
1020 g_free(old_cluster
);
1025 * Frees the allocated clusters because the request failed and they won't
1026 * actually be linked.
1028 void qcow2_alloc_cluster_abort(BlockDriverState
*bs
, QCowL2Meta
*m
)
1030 BDRVQcow2State
*s
= bs
->opaque
;
1031 qcow2_free_clusters(bs
, m
->alloc_offset
, m
->nb_clusters
<< s
->cluster_bits
,
1032 QCOW2_DISCARD_NEVER
);
1036 * Returns the number of contiguous clusters that can be used for an allocating
1037 * write, but require COW to be performed (this includes yet unallocated space,
1038 * which must copy from the backing file)
1040 static int count_cow_clusters(BlockDriverState
*bs
, int nb_clusters
,
1041 uint64_t *l2_slice
, int l2_index
)
1045 for (i
= 0; i
< nb_clusters
; i
++) {
1046 uint64_t l2_entry
= be64_to_cpu(l2_slice
[l2_index
+ i
]);
1047 QCow2ClusterType cluster_type
= qcow2_get_cluster_type(bs
, l2_entry
);
1049 switch(cluster_type
) {
1050 case QCOW2_CLUSTER_NORMAL
:
1051 if (l2_entry
& QCOW_OFLAG_COPIED
) {
1055 case QCOW2_CLUSTER_UNALLOCATED
:
1056 case QCOW2_CLUSTER_COMPRESSED
:
1057 case QCOW2_CLUSTER_ZERO_PLAIN
:
1058 case QCOW2_CLUSTER_ZERO_ALLOC
:
1066 assert(i
<= nb_clusters
);
1071 * Check if there already is an AIO write request in flight which allocates
1072 * the same cluster. In this case we need to wait until the previous
1073 * request has completed and updated the L2 table accordingly.
1076 * 0 if there was no dependency. *cur_bytes indicates the number of
1077 * bytes from guest_offset that can be read before the next
1078 * dependency must be processed (or the request is complete)
1080 * -EAGAIN if we had to wait for another request, previously gathered
1081 * information on cluster allocation may be invalid now. The caller
1082 * must start over anyway, so consider *cur_bytes undefined.
1084 static int handle_dependencies(BlockDriverState
*bs
, uint64_t guest_offset
,
1085 uint64_t *cur_bytes
, QCowL2Meta
**m
)
1087 BDRVQcow2State
*s
= bs
->opaque
;
1088 QCowL2Meta
*old_alloc
;
1089 uint64_t bytes
= *cur_bytes
;
1091 QLIST_FOREACH(old_alloc
, &s
->cluster_allocs
, next_in_flight
) {
1093 uint64_t start
= guest_offset
;
1094 uint64_t end
= start
+ bytes
;
1095 uint64_t old_start
= l2meta_cow_start(old_alloc
);
1096 uint64_t old_end
= l2meta_cow_end(old_alloc
);
1098 if (end
<= old_start
|| start
>= old_end
) {
1099 /* No intersection */
1101 if (start
< old_start
) {
1102 /* Stop at the start of a running allocation */
1103 bytes
= old_start
- start
;
1108 /* Stop if already an l2meta exists. After yielding, it wouldn't
1109 * be valid any more, so we'd have to clean up the old L2Metas
1110 * and deal with requests depending on them before starting to
1111 * gather new ones. Not worth the trouble. */
1112 if (bytes
== 0 && *m
) {
1118 /* Wait for the dependency to complete. We need to recheck
1119 * the free/allocated clusters when we continue. */
1120 qemu_co_queue_wait(&old_alloc
->dependent_requests
, &s
->lock
);
1126 /* Make sure that existing clusters and new allocations are only used up to
1127 * the next dependency if we shortened the request above */
1134 * Checks how many already allocated clusters that don't require a copy on
1135 * write there are at the given guest_offset (up to *bytes). If *host_offset is
1136 * not INV_OFFSET, only physically contiguous clusters beginning at this host
1137 * offset are counted.
1139 * Note that guest_offset may not be cluster aligned. In this case, the
1140 * returned *host_offset points to exact byte referenced by guest_offset and
1141 * therefore isn't cluster aligned as well.
1144 * 0: if no allocated clusters are available at the given offset.
1145 * *bytes is normally unchanged. It is set to 0 if the cluster
1146 * is allocated and doesn't need COW, but doesn't have the right
1149 * 1: if allocated clusters that don't require a COW are available at
1150 * the requested offset. *bytes may have decreased and describes
1151 * the length of the area that can be written to.
1153 * -errno: in error cases
1155 static int handle_copied(BlockDriverState
*bs
, uint64_t guest_offset
,
1156 uint64_t *host_offset
, uint64_t *bytes
, QCowL2Meta
**m
)
1158 BDRVQcow2State
*s
= bs
->opaque
;
1160 uint64_t cluster_offset
;
1162 uint64_t nb_clusters
;
1163 unsigned int keep_clusters
;
1166 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset
, *host_offset
,
1169 assert(*host_offset
== INV_OFFSET
|| offset_into_cluster(s
, guest_offset
)
1170 == offset_into_cluster(s
, *host_offset
));
1173 * Calculate the number of clusters to look for. We stop at L2 slice
1174 * boundaries to keep things simple.
1177 size_to_clusters(s
, offset_into_cluster(s
, guest_offset
) + *bytes
);
1179 l2_index
= offset_to_l2_slice_index(s
, guest_offset
);
1180 nb_clusters
= MIN(nb_clusters
, s
->l2_slice_size
- l2_index
);
1181 assert(nb_clusters
<= INT_MAX
);
1183 /* Find L2 entry for the first involved cluster */
1184 ret
= get_cluster_table(bs
, guest_offset
, &l2_slice
, &l2_index
);
1189 cluster_offset
= be64_to_cpu(l2_slice
[l2_index
]);
1191 /* Check how many clusters are already allocated and don't need COW */
1192 if (qcow2_get_cluster_type(bs
, cluster_offset
) == QCOW2_CLUSTER_NORMAL
1193 && (cluster_offset
& QCOW_OFLAG_COPIED
))
1195 /* If a specific host_offset is required, check it */
1196 bool offset_matches
=
1197 (cluster_offset
& L2E_OFFSET_MASK
) == *host_offset
;
1199 if (offset_into_cluster(s
, cluster_offset
& L2E_OFFSET_MASK
)) {
1200 qcow2_signal_corruption(bs
, true, -1, -1, "Data cluster offset "
1201 "%#llx unaligned (guest offset: %#" PRIx64
1202 ")", cluster_offset
& L2E_OFFSET_MASK
,
1208 if (*host_offset
!= INV_OFFSET
&& !offset_matches
) {
1214 /* We keep all QCOW_OFLAG_COPIED clusters */
1216 count_contiguous_clusters(bs
, nb_clusters
, s
->cluster_size
,
1217 &l2_slice
[l2_index
],
1218 QCOW_OFLAG_COPIED
| QCOW_OFLAG_ZERO
);
1219 assert(keep_clusters
<= nb_clusters
);
1221 *bytes
= MIN(*bytes
,
1222 keep_clusters
* s
->cluster_size
1223 - offset_into_cluster(s
, guest_offset
));
1232 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1234 /* Only return a host offset if we actually made progress. Otherwise we
1235 * would make requirements for handle_alloc() that it can't fulfill */
1237 *host_offset
= (cluster_offset
& L2E_OFFSET_MASK
)
1238 + offset_into_cluster(s
, guest_offset
);
1245 * Allocates new clusters for the given guest_offset.
1247 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1248 * contain the number of clusters that have been allocated and are contiguous
1249 * in the image file.
1251 * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1252 * at which the new clusters must start. *nb_clusters can be 0 on return in
1253 * this case if the cluster at host_offset is already in use. If *host_offset
1254 * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1256 * *host_offset is updated to contain the offset into the image file at which
1257 * the first allocated cluster starts.
1259 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1260 * function has been waiting for another request and the allocation must be
1261 * restarted, but the whole request should not be failed.
1263 static int do_alloc_cluster_offset(BlockDriverState
*bs
, uint64_t guest_offset
,
1264 uint64_t *host_offset
, uint64_t *nb_clusters
)
1266 BDRVQcow2State
*s
= bs
->opaque
;
1268 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset
,
1269 *host_offset
, *nb_clusters
);
1271 if (has_data_file(bs
)) {
1272 assert(*host_offset
== INV_OFFSET
||
1273 *host_offset
== start_of_cluster(s
, guest_offset
));
1274 *host_offset
= start_of_cluster(s
, guest_offset
);
1278 /* Allocate new clusters */
1279 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1280 if (*host_offset
== INV_OFFSET
) {
1281 int64_t cluster_offset
=
1282 qcow2_alloc_clusters(bs
, *nb_clusters
* s
->cluster_size
);
1283 if (cluster_offset
< 0) {
1284 return cluster_offset
;
1286 *host_offset
= cluster_offset
;
1289 int64_t ret
= qcow2_alloc_clusters_at(bs
, *host_offset
, *nb_clusters
);
1299 * Allocates new clusters for an area that either is yet unallocated or needs a
1300 * copy on write. If *host_offset is not INV_OFFSET, clusters are only
1301 * allocated if the new allocation can match the specified host offset.
1303 * Note that guest_offset may not be cluster aligned. In this case, the
1304 * returned *host_offset points to exact byte referenced by guest_offset and
1305 * therefore isn't cluster aligned as well.
1308 * 0: if no clusters could be allocated. *bytes is set to 0,
1309 * *host_offset is left unchanged.
1311 * 1: if new clusters were allocated. *bytes may be decreased if the
1312 * new allocation doesn't cover all of the requested area.
1313 * *host_offset is updated to contain the host offset of the first
1314 * newly allocated cluster.
1316 * -errno: in error cases
1318 static int handle_alloc(BlockDriverState
*bs
, uint64_t guest_offset
,
1319 uint64_t *host_offset
, uint64_t *bytes
, QCowL2Meta
**m
)
1321 BDRVQcow2State
*s
= bs
->opaque
;
1325 uint64_t nb_clusters
;
1327 bool keep_old_clusters
= false;
1329 uint64_t alloc_cluster_offset
= INV_OFFSET
;
1331 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset
, *host_offset
,
1336 * Calculate the number of clusters to look for. We stop at L2 slice
1337 * boundaries to keep things simple.
1340 size_to_clusters(s
, offset_into_cluster(s
, guest_offset
) + *bytes
);
1342 l2_index
= offset_to_l2_slice_index(s
, guest_offset
);
1343 nb_clusters
= MIN(nb_clusters
, s
->l2_slice_size
- l2_index
);
1344 assert(nb_clusters
<= INT_MAX
);
1346 /* Find L2 entry for the first involved cluster */
1347 ret
= get_cluster_table(bs
, guest_offset
, &l2_slice
, &l2_index
);
1352 entry
= be64_to_cpu(l2_slice
[l2_index
]);
1354 /* For the moment, overwrite compressed clusters one by one */
1355 if (entry
& QCOW_OFLAG_COMPRESSED
) {
1358 nb_clusters
= count_cow_clusters(bs
, nb_clusters
, l2_slice
, l2_index
);
1361 /* This function is only called when there were no non-COW clusters, so if
1362 * we can't find any unallocated or COW clusters either, something is
1363 * wrong with our code. */
1364 assert(nb_clusters
> 0);
1366 if (qcow2_get_cluster_type(bs
, entry
) == QCOW2_CLUSTER_ZERO_ALLOC
&&
1367 (entry
& QCOW_OFLAG_COPIED
) &&
1368 (*host_offset
== INV_OFFSET
||
1369 start_of_cluster(s
, *host_offset
) == (entry
& L2E_OFFSET_MASK
)))
1371 int preallocated_nb_clusters
;
1373 if (offset_into_cluster(s
, entry
& L2E_OFFSET_MASK
)) {
1374 qcow2_signal_corruption(bs
, true, -1, -1, "Preallocated zero "
1375 "cluster offset %#llx unaligned (guest "
1376 "offset: %#" PRIx64
")",
1377 entry
& L2E_OFFSET_MASK
, guest_offset
);
1382 /* Try to reuse preallocated zero clusters; contiguous normal clusters
1383 * would be fine, too, but count_cow_clusters() above has limited
1384 * nb_clusters already to a range of COW clusters */
1385 preallocated_nb_clusters
=
1386 count_contiguous_clusters(bs
, nb_clusters
, s
->cluster_size
,
1387 &l2_slice
[l2_index
], QCOW_OFLAG_COPIED
);
1388 assert(preallocated_nb_clusters
> 0);
1390 nb_clusters
= preallocated_nb_clusters
;
1391 alloc_cluster_offset
= entry
& L2E_OFFSET_MASK
;
1393 /* We want to reuse these clusters, so qcow2_alloc_cluster_link_l2()
1394 * should not free them. */
1395 keep_old_clusters
= true;
1398 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1400 if (alloc_cluster_offset
== INV_OFFSET
) {
1401 /* Allocate, if necessary at a given offset in the image file */
1402 alloc_cluster_offset
= *host_offset
== INV_OFFSET
? INV_OFFSET
:
1403 start_of_cluster(s
, *host_offset
);
1404 ret
= do_alloc_cluster_offset(bs
, guest_offset
, &alloc_cluster_offset
,
1410 /* Can't extend contiguous allocation */
1411 if (nb_clusters
== 0) {
1416 assert(alloc_cluster_offset
!= INV_OFFSET
);
1420 * Save info needed for meta data update.
1422 * requested_bytes: Number of bytes from the start of the first
1423 * newly allocated cluster to the end of the (possibly shortened
1424 * before) write request.
1426 * avail_bytes: Number of bytes from the start of the first
1427 * newly allocated to the end of the last newly allocated cluster.
1429 * nb_bytes: The number of bytes from the start of the first
1430 * newly allocated cluster to the end of the area that the write
1431 * request actually writes to (excluding COW at the end)
1433 uint64_t requested_bytes
= *bytes
+ offset_into_cluster(s
, guest_offset
);
1434 int avail_bytes
= MIN(INT_MAX
, nb_clusters
<< s
->cluster_bits
);
1435 int nb_bytes
= MIN(requested_bytes
, avail_bytes
);
1436 QCowL2Meta
*old_m
= *m
;
1438 *m
= g_malloc0(sizeof(**m
));
1440 **m
= (QCowL2Meta
) {
1443 .alloc_offset
= alloc_cluster_offset
,
1444 .offset
= start_of_cluster(s
, guest_offset
),
1445 .nb_clusters
= nb_clusters
,
1447 .keep_old_clusters
= keep_old_clusters
,
1451 .nb_bytes
= offset_into_cluster(s
, guest_offset
),
1455 .nb_bytes
= avail_bytes
- nb_bytes
,
1458 qemu_co_queue_init(&(*m
)->dependent_requests
);
1459 QLIST_INSERT_HEAD(&s
->cluster_allocs
, *m
, next_in_flight
);
1461 *host_offset
= alloc_cluster_offset
+ offset_into_cluster(s
, guest_offset
);
1462 *bytes
= MIN(*bytes
, nb_bytes
- offset_into_cluster(s
, guest_offset
));
1463 assert(*bytes
!= 0);
1468 if (*m
&& (*m
)->nb_clusters
> 0) {
1469 QLIST_REMOVE(*m
, next_in_flight
);
1475 * alloc_cluster_offset
1477 * For a given offset on the virtual disk, find the cluster offset in qcow2
1478 * file. If the offset is not found, allocate a new cluster.
1480 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1481 * other fields in m are meaningless.
1483 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1484 * contiguous clusters that have been allocated. In this case, the other
1485 * fields of m are valid and contain information about the first allocated
1488 * If the request conflicts with another write request in flight, the coroutine
1489 * is queued and will be reentered when the dependency has completed.
1491 * Return 0 on success and -errno in error cases
1493 int qcow2_alloc_cluster_offset(BlockDriverState
*bs
, uint64_t offset
,
1494 unsigned int *bytes
, uint64_t *host_offset
,
1497 BDRVQcow2State
*s
= bs
->opaque
;
1498 uint64_t start
, remaining
;
1499 uint64_t cluster_offset
;
1503 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset
, *bytes
);
1508 cluster_offset
= INV_OFFSET
;
1509 *host_offset
= INV_OFFSET
;
1515 if (*host_offset
== INV_OFFSET
&& cluster_offset
!= INV_OFFSET
) {
1516 *host_offset
= start_of_cluster(s
, cluster_offset
);
1519 assert(remaining
>= cur_bytes
);
1522 remaining
-= cur_bytes
;
1524 if (cluster_offset
!= INV_OFFSET
) {
1525 cluster_offset
+= cur_bytes
;
1528 if (remaining
== 0) {
1532 cur_bytes
= remaining
;
1535 * Now start gathering as many contiguous clusters as possible:
1537 * 1. Check for overlaps with in-flight allocations
1539 * a) Overlap not in the first cluster -> shorten this request and
1540 * let the caller handle the rest in its next loop iteration.
1542 * b) Real overlaps of two requests. Yield and restart the search
1543 * for contiguous clusters (the situation could have changed
1544 * while we were sleeping)
1546 * c) TODO: Request starts in the same cluster as the in-flight
1547 * allocation ends. Shorten the COW of the in-fight allocation,
1548 * set cluster_offset to write to the same cluster and set up
1549 * the right synchronisation between the in-flight request and
1552 ret
= handle_dependencies(bs
, start
, &cur_bytes
, m
);
1553 if (ret
== -EAGAIN
) {
1554 /* Currently handle_dependencies() doesn't yield if we already had
1555 * an allocation. If it did, we would have to clean up the L2Meta
1556 * structs before starting over. */
1559 } else if (ret
< 0) {
1561 } else if (cur_bytes
== 0) {
1564 /* handle_dependencies() may have decreased cur_bytes (shortened
1565 * the allocations below) so that the next dependency is processed
1566 * correctly during the next loop iteration. */
1570 * 2. Count contiguous COPIED clusters.
1572 ret
= handle_copied(bs
, start
, &cluster_offset
, &cur_bytes
, m
);
1577 } else if (cur_bytes
== 0) {
1582 * 3. If the request still hasn't completed, allocate new clusters,
1583 * considering any cluster_offset of steps 1c or 2.
1585 ret
= handle_alloc(bs
, start
, &cluster_offset
, &cur_bytes
, m
);
1591 assert(cur_bytes
== 0);
1596 *bytes
-= remaining
;
1598 assert(*host_offset
!= INV_OFFSET
);
1604 * This discards as many clusters of nb_clusters as possible at once (i.e.
1605 * all clusters in the same L2 slice) and returns the number of discarded
1608 static int discard_in_l2_slice(BlockDriverState
*bs
, uint64_t offset
,
1609 uint64_t nb_clusters
,
1610 enum qcow2_discard_type type
, bool full_discard
)
1612 BDRVQcow2State
*s
= bs
->opaque
;
1618 ret
= get_cluster_table(bs
, offset
, &l2_slice
, &l2_index
);
1623 /* Limit nb_clusters to one L2 slice */
1624 nb_clusters
= MIN(nb_clusters
, s
->l2_slice_size
- l2_index
);
1625 assert(nb_clusters
<= INT_MAX
);
1627 for (i
= 0; i
< nb_clusters
; i
++) {
1628 uint64_t old_l2_entry
;
1630 old_l2_entry
= be64_to_cpu(l2_slice
[l2_index
+ i
]);
1633 * If full_discard is false, make sure that a discarded area reads back
1634 * as zeroes for v3 images (we cannot do it for v2 without actually
1635 * writing a zero-filled buffer). We can skip the operation if the
1636 * cluster is already marked as zero, or if it's unallocated and we
1637 * don't have a backing file.
1639 * TODO We might want to use bdrv_block_status(bs) here, but we're
1640 * holding s->lock, so that doesn't work today.
1642 * If full_discard is true, the sector should not read back as zeroes,
1643 * but rather fall through to the backing file.
1645 switch (qcow2_get_cluster_type(bs
, old_l2_entry
)) {
1646 case QCOW2_CLUSTER_UNALLOCATED
:
1647 if (full_discard
|| !bs
->backing
) {
1652 case QCOW2_CLUSTER_ZERO_PLAIN
:
1653 if (!full_discard
) {
1658 case QCOW2_CLUSTER_ZERO_ALLOC
:
1659 case QCOW2_CLUSTER_NORMAL
:
1660 case QCOW2_CLUSTER_COMPRESSED
:
1667 /* First remove L2 entries */
1668 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
1669 if (!full_discard
&& s
->qcow_version
>= 3) {
1670 l2_slice
[l2_index
+ i
] = cpu_to_be64(QCOW_OFLAG_ZERO
);
1672 l2_slice
[l2_index
+ i
] = cpu_to_be64(0);
1675 /* Then decrease the refcount */
1676 qcow2_free_any_clusters(bs
, old_l2_entry
, 1, type
);
1679 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1684 int qcow2_cluster_discard(BlockDriverState
*bs
, uint64_t offset
,
1685 uint64_t bytes
, enum qcow2_discard_type type
,
1688 BDRVQcow2State
*s
= bs
->opaque
;
1689 uint64_t end_offset
= offset
+ bytes
;
1690 uint64_t nb_clusters
;
1694 /* Caller must pass aligned values, except at image end */
1695 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
1696 assert(QEMU_IS_ALIGNED(end_offset
, s
->cluster_size
) ||
1697 end_offset
== bs
->total_sectors
<< BDRV_SECTOR_BITS
);
1699 nb_clusters
= size_to_clusters(s
, bytes
);
1701 s
->cache_discards
= true;
1703 /* Each L2 slice is handled by its own loop iteration */
1704 while (nb_clusters
> 0) {
1705 cleared
= discard_in_l2_slice(bs
, offset
, nb_clusters
, type
,
1712 nb_clusters
-= cleared
;
1713 offset
+= (cleared
* s
->cluster_size
);
1718 s
->cache_discards
= false;
1719 qcow2_process_discards(bs
, ret
);
1725 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1726 * all clusters in the same L2 slice) and returns the number of zeroed
1729 static int zero_in_l2_slice(BlockDriverState
*bs
, uint64_t offset
,
1730 uint64_t nb_clusters
, int flags
)
1732 BDRVQcow2State
*s
= bs
->opaque
;
1737 bool unmap
= !!(flags
& BDRV_REQ_MAY_UNMAP
);
1739 ret
= get_cluster_table(bs
, offset
, &l2_slice
, &l2_index
);
1744 /* Limit nb_clusters to one L2 slice */
1745 nb_clusters
= MIN(nb_clusters
, s
->l2_slice_size
- l2_index
);
1746 assert(nb_clusters
<= INT_MAX
);
1748 for (i
= 0; i
< nb_clusters
; i
++) {
1749 uint64_t old_offset
;
1750 QCow2ClusterType cluster_type
;
1752 old_offset
= be64_to_cpu(l2_slice
[l2_index
+ i
]);
1755 * Minimize L2 changes if the cluster already reads back as
1756 * zeroes with correct allocation.
1758 cluster_type
= qcow2_get_cluster_type(bs
, old_offset
);
1759 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
||
1760 (cluster_type
== QCOW2_CLUSTER_ZERO_ALLOC
&& !unmap
)) {
1764 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
1765 if (cluster_type
== QCOW2_CLUSTER_COMPRESSED
|| unmap
) {
1766 l2_slice
[l2_index
+ i
] = cpu_to_be64(QCOW_OFLAG_ZERO
);
1767 qcow2_free_any_clusters(bs
, old_offset
, 1, QCOW2_DISCARD_REQUEST
);
1769 l2_slice
[l2_index
+ i
] |= cpu_to_be64(QCOW_OFLAG_ZERO
);
1773 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1778 int qcow2_cluster_zeroize(BlockDriverState
*bs
, uint64_t offset
,
1779 uint64_t bytes
, int flags
)
1781 BDRVQcow2State
*s
= bs
->opaque
;
1782 uint64_t end_offset
= offset
+ bytes
;
1783 uint64_t nb_clusters
;
1787 /* If we have to stay in sync with an external data file, zero out
1788 * s->data_file first. */
1789 if (data_file_is_raw(bs
)) {
1790 assert(has_data_file(bs
));
1791 ret
= bdrv_co_pwrite_zeroes(s
->data_file
, offset
, bytes
, flags
);
1797 /* Caller must pass aligned values, except at image end */
1798 assert(QEMU_IS_ALIGNED(offset
, s
->cluster_size
));
1799 assert(QEMU_IS_ALIGNED(end_offset
, s
->cluster_size
) ||
1800 end_offset
== bs
->total_sectors
<< BDRV_SECTOR_BITS
);
1802 /* The zero flag is only supported by version 3 and newer */
1803 if (s
->qcow_version
< 3) {
1807 /* Each L2 slice is handled by its own loop iteration */
1808 nb_clusters
= size_to_clusters(s
, bytes
);
1810 s
->cache_discards
= true;
1812 while (nb_clusters
> 0) {
1813 cleared
= zero_in_l2_slice(bs
, offset
, nb_clusters
, flags
);
1819 nb_clusters
-= cleared
;
1820 offset
+= (cleared
* s
->cluster_size
);
1825 s
->cache_discards
= false;
1826 qcow2_process_discards(bs
, ret
);
1832 * Expands all zero clusters in a specific L1 table (or deallocates them, for
1833 * non-backed non-pre-allocated zero clusters).
1835 * l1_entries and *visited_l1_entries are used to keep track of progress for
1836 * status_cb(). l1_entries contains the total number of L1 entries and
1837 * *visited_l1_entries counts all visited L1 entries.
1839 static int expand_zero_clusters_in_l1(BlockDriverState
*bs
, uint64_t *l1_table
,
1840 int l1_size
, int64_t *visited_l1_entries
,
1842 BlockDriverAmendStatusCB
*status_cb
,
1845 BDRVQcow2State
*s
= bs
->opaque
;
1846 bool is_active_l1
= (l1_table
== s
->l1_table
);
1847 uint64_t *l2_slice
= NULL
;
1848 unsigned slice
, slice_size2
, n_slices
;
1852 slice_size2
= s
->l2_slice_size
* sizeof(uint64_t);
1853 n_slices
= s
->cluster_size
/ slice_size2
;
1855 if (!is_active_l1
) {
1856 /* inactive L2 tables require a buffer to be stored in when loading
1858 l2_slice
= qemu_try_blockalign(bs
->file
->bs
, slice_size2
);
1859 if (l2_slice
== NULL
) {
1864 for (i
= 0; i
< l1_size
; i
++) {
1865 uint64_t l2_offset
= l1_table
[i
] & L1E_OFFSET_MASK
;
1866 uint64_t l2_refcount
;
1870 (*visited_l1_entries
)++;
1872 status_cb(bs
, *visited_l1_entries
, l1_entries
, cb_opaque
);
1877 if (offset_into_cluster(s
, l2_offset
)) {
1878 qcow2_signal_corruption(bs
, true, -1, -1, "L2 table offset %#"
1879 PRIx64
" unaligned (L1 index: %#x)",
1885 ret
= qcow2_get_refcount(bs
, l2_offset
>> s
->cluster_bits
,
1891 for (slice
= 0; slice
< n_slices
; slice
++) {
1892 uint64_t slice_offset
= l2_offset
+ slice
* slice_size2
;
1893 bool l2_dirty
= false;
1895 /* get active L2 tables from cache */
1896 ret
= qcow2_cache_get(bs
, s
->l2_table_cache
, slice_offset
,
1897 (void **)&l2_slice
);
1899 /* load inactive L2 tables from disk */
1900 ret
= bdrv_pread(bs
->file
, slice_offset
, l2_slice
, slice_size2
);
1906 for (j
= 0; j
< s
->l2_slice_size
; j
++) {
1907 uint64_t l2_entry
= be64_to_cpu(l2_slice
[j
]);
1908 int64_t offset
= l2_entry
& L2E_OFFSET_MASK
;
1909 QCow2ClusterType cluster_type
=
1910 qcow2_get_cluster_type(bs
, l2_entry
);
1912 if (cluster_type
!= QCOW2_CLUSTER_ZERO_PLAIN
&&
1913 cluster_type
!= QCOW2_CLUSTER_ZERO_ALLOC
) {
1917 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1919 /* not backed; therefore we can simply deallocate the
1926 offset
= qcow2_alloc_clusters(bs
, s
->cluster_size
);
1932 if (l2_refcount
> 1) {
1933 /* For shared L2 tables, set the refcount accordingly
1934 * (it is already 1 and needs to be l2_refcount) */
1935 ret
= qcow2_update_cluster_refcount(
1936 bs
, offset
>> s
->cluster_bits
,
1937 refcount_diff(1, l2_refcount
), false,
1938 QCOW2_DISCARD_OTHER
);
1940 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1941 QCOW2_DISCARD_OTHER
);
1947 if (offset_into_cluster(s
, offset
)) {
1948 int l2_index
= slice
* s
->l2_slice_size
+ j
;
1949 qcow2_signal_corruption(
1951 "Cluster allocation offset "
1952 "%#" PRIx64
" unaligned (L2 offset: %#"
1953 PRIx64
", L2 index: %#x)", offset
,
1954 l2_offset
, l2_index
);
1955 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1956 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1957 QCOW2_DISCARD_ALWAYS
);
1963 ret
= qcow2_pre_write_overlap_check(bs
, 0, offset
,
1964 s
->cluster_size
, true);
1966 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1967 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1968 QCOW2_DISCARD_ALWAYS
);
1973 ret
= bdrv_pwrite_zeroes(s
->data_file
, offset
,
1974 s
->cluster_size
, 0);
1976 if (cluster_type
== QCOW2_CLUSTER_ZERO_PLAIN
) {
1977 qcow2_free_clusters(bs
, offset
, s
->cluster_size
,
1978 QCOW2_DISCARD_ALWAYS
);
1983 if (l2_refcount
== 1) {
1984 l2_slice
[j
] = cpu_to_be64(offset
| QCOW_OFLAG_COPIED
);
1986 l2_slice
[j
] = cpu_to_be64(offset
);
1993 qcow2_cache_entry_mark_dirty(s
->l2_table_cache
, l2_slice
);
1994 qcow2_cache_depends_on_flush(s
->l2_table_cache
);
1996 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
1999 ret
= qcow2_pre_write_overlap_check(
2000 bs
, QCOW2_OL_INACTIVE_L2
| QCOW2_OL_ACTIVE_L2
,
2001 slice_offset
, slice_size2
, false);
2006 ret
= bdrv_pwrite(bs
->file
, slice_offset
,
2007 l2_slice
, slice_size2
);
2015 (*visited_l1_entries
)++;
2017 status_cb(bs
, *visited_l1_entries
, l1_entries
, cb_opaque
);
2025 if (!is_active_l1
) {
2026 qemu_vfree(l2_slice
);
2028 qcow2_cache_put(s
->l2_table_cache
, (void **) &l2_slice
);
2035 * For backed images, expands all zero clusters on the image. For non-backed
2036 * images, deallocates all non-pre-allocated zero clusters (and claims the
2037 * allocation for pre-allocated ones). This is important for downgrading to a
2038 * qcow2 version which doesn't yet support metadata zero clusters.
2040 int qcow2_expand_zero_clusters(BlockDriverState
*bs
,
2041 BlockDriverAmendStatusCB
*status_cb
,
2044 BDRVQcow2State
*s
= bs
->opaque
;
2045 uint64_t *l1_table
= NULL
;
2046 int64_t l1_entries
= 0, visited_l1_entries
= 0;
2051 l1_entries
= s
->l1_size
;
2052 for (i
= 0; i
< s
->nb_snapshots
; i
++) {
2053 l1_entries
+= s
->snapshots
[i
].l1_size
;
2057 ret
= expand_zero_clusters_in_l1(bs
, s
->l1_table
, s
->l1_size
,
2058 &visited_l1_entries
, l1_entries
,
2059 status_cb
, cb_opaque
);
2064 /* Inactive L1 tables may point to active L2 tables - therefore it is
2065 * necessary to flush the L2 table cache before trying to access the L2
2066 * tables pointed to by inactive L1 entries (else we might try to expand
2067 * zero clusters that have already been expanded); furthermore, it is also
2068 * necessary to empty the L2 table cache, since it may contain tables which
2069 * are now going to be modified directly on disk, bypassing the cache.
2070 * qcow2_cache_empty() does both for us. */
2071 ret
= qcow2_cache_empty(bs
, s
->l2_table_cache
);
2076 for (i
= 0; i
< s
->nb_snapshots
; i
++) {
2078 uint64_t *new_l1_table
;
2079 Error
*local_err
= NULL
;
2081 ret
= qcow2_validate_table(bs
, s
->snapshots
[i
].l1_table_offset
,
2082 s
->snapshots
[i
].l1_size
, sizeof(uint64_t),
2083 QCOW_MAX_L1_SIZE
, "Snapshot L1 table",
2086 error_report_err(local_err
);
2090 l1_size2
= s
->snapshots
[i
].l1_size
* sizeof(uint64_t);
2091 new_l1_table
= g_try_realloc(l1_table
, l1_size2
);
2093 if (!new_l1_table
) {
2098 l1_table
= new_l1_table
;
2100 ret
= bdrv_pread(bs
->file
, s
->snapshots
[i
].l1_table_offset
,
2101 l1_table
, l1_size2
);
2106 for (j
= 0; j
< s
->snapshots
[i
].l1_size
; j
++) {
2107 be64_to_cpus(&l1_table
[j
]);
2110 ret
= expand_zero_clusters_in_l1(bs
, l1_table
, s
->snapshots
[i
].l1_size
,
2111 &visited_l1_entries
, l1_entries
,
2112 status_cb
, cb_opaque
);