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
27
#include "qemu-common.h"
28
#include "block/block_int.h"
29
#include "block/qcow2.h"
32
int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
35
BDRVQcowState *s = bs->opaque;
36
int new_l1_size2, ret, i;
37
uint64_t *new_l1_table;
38
int64_t new_l1_table_offset, new_l1_size;
41
if (min_size <= s->l1_size)
45
new_l1_size = min_size;
47
/* Bump size up to reduce the number of times we have to grow */
48
new_l1_size = s->l1_size;
49
if (new_l1_size == 0) {
52
while (min_size > new_l1_size) {
53
new_l1_size = (new_l1_size * 3 + 1) / 2;
57
if (new_l1_size > INT_MAX) {
62
fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
63
s->l1_size, new_l1_size);
66
new_l1_size2 = sizeof(uint64_t) * new_l1_size;
67
new_l1_table = g_malloc0(align_offset(new_l1_size2, 512));
68
memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
70
/* write new table (align to cluster) */
71
BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
72
new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
73
if (new_l1_table_offset < 0) {
75
return new_l1_table_offset;
78
ret = qcow2_cache_flush(bs, s->refcount_block_cache);
83
BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
84
for(i = 0; i < s->l1_size; i++)
85
new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
86
ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_table, new_l1_size2);
89
for(i = 0; i < s->l1_size; i++)
90
new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
93
BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
94
cpu_to_be32w((uint32_t*)data, new_l1_size);
95
cpu_to_be64wu((uint64_t*)(data + 4), new_l1_table_offset);
96
ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), data,sizeof(data));
101
qcow2_free_clusters(bs, s->l1_table_offset, s->l1_size * sizeof(uint64_t),
102
QCOW2_DISCARD_OTHER);
103
s->l1_table_offset = new_l1_table_offset;
104
s->l1_table = new_l1_table;
105
s->l1_size = new_l1_size;
108
g_free(new_l1_table);
109
qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
110
QCOW2_DISCARD_OTHER);
117
* Loads a L2 table into memory. If the table is in the cache, the cache
118
* is used; otherwise the L2 table is loaded from the image file.
120
* Returns a pointer to the L2 table on success, or NULL if the read from
121
* the image file failed.
124
static int l2_load(BlockDriverState *bs, uint64_t l2_offset,
127
BDRVQcowState *s = bs->opaque;
130
ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, (void**) l2_table);
136
* Writes one sector of the L1 table to the disk (can't update single entries
137
* and we really don't want bdrv_pread to perform a read-modify-write)
139
#define L1_ENTRIES_PER_SECTOR (512 / 8)
140
static int write_l1_entry(BlockDriverState *bs, int l1_index)
142
BDRVQcowState *s = bs->opaque;
143
uint64_t buf[L1_ENTRIES_PER_SECTOR];
147
l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
148
for (i = 0; i < L1_ENTRIES_PER_SECTOR; i++) {
149
buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
152
BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
153
ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset + 8 * l1_start_index,
165
* Allocate a new l2 entry in the file. If l1_index points to an already
166
* used entry in the L2 table (i.e. we are doing a copy on write for the L2
167
* table) copy the contents of the old L2 table into the newly allocated one.
168
* Otherwise the new table is initialized with zeros.
172
static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table)
174
BDRVQcowState *s = bs->opaque;
175
uint64_t old_l2_offset;
180
old_l2_offset = s->l1_table[l1_index];
182
trace_qcow2_l2_allocate(bs, l1_index);
184
/* allocate a new l2 entry */
186
l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
191
ret = qcow2_cache_flush(bs, s->refcount_block_cache);
196
/* allocate a new entry in the l2 cache */
198
trace_qcow2_l2_allocate_get_empty(bs, l1_index);
199
ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table);
206
if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
207
/* if there was no old l2 table, clear the new table */
208
memset(l2_table, 0, s->l2_size * sizeof(uint64_t));
212
/* if there was an old l2 table, read it from the disk */
213
BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
214
ret = qcow2_cache_get(bs, s->l2_table_cache,
215
old_l2_offset & L1E_OFFSET_MASK,
216
(void**) &old_table);
221
memcpy(l2_table, old_table, s->cluster_size);
223
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &old_table);
229
/* write the l2 table to the file */
230
BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
232
trace_qcow2_l2_allocate_write_l2(bs, l1_index);
233
qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
234
ret = qcow2_cache_flush(bs, s->l2_table_cache);
239
/* update the L1 entry */
240
trace_qcow2_l2_allocate_write_l1(bs, l1_index);
241
s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
242
ret = write_l1_entry(bs, l1_index);
248
trace_qcow2_l2_allocate_done(bs, l1_index, 0);
252
trace_qcow2_l2_allocate_done(bs, l1_index, ret);
253
qcow2_cache_put(bs, s->l2_table_cache, (void**) table);
254
s->l1_table[l1_index] = old_l2_offset;
259
* Checks how many clusters in a given L2 table are contiguous in the image
260
* file. As soon as one of the flags in the bitmask stop_flags changes compared
261
* to the first cluster, the search is stopped and the cluster is not counted
262
* as contiguous. (This allows it, for example, to stop at the first compressed
263
* cluster which may require a different handling)
265
static int count_contiguous_clusters(uint64_t nb_clusters, int cluster_size,
266
uint64_t *l2_table, uint64_t start, uint64_t stop_flags)
269
uint64_t mask = stop_flags | L2E_OFFSET_MASK;
270
uint64_t offset = be64_to_cpu(l2_table[0]) & mask;
275
for (i = start; i < start + nb_clusters; i++) {
276
uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask;
277
if (offset + (uint64_t) i * cluster_size != l2_entry) {
285
static int count_contiguous_free_clusters(uint64_t nb_clusters, uint64_t *l2_table)
289
for (i = 0; i < nb_clusters; i++) {
290
int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i]));
292
if (type != QCOW2_CLUSTER_UNALLOCATED) {
300
/* The crypt function is compatible with the linux cryptoloop
301
algorithm for < 4 GB images. NOTE: out_buf == in_buf is
303
void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
304
uint8_t *out_buf, const uint8_t *in_buf,
305
int nb_sectors, int enc,
314
for(i = 0; i < nb_sectors; i++) {
315
ivec.ll[0] = cpu_to_le64(sector_num);
317
AES_cbc_encrypt(in_buf, out_buf, 512, key,
325
static int coroutine_fn copy_sectors(BlockDriverState *bs,
327
uint64_t cluster_offset,
328
int n_start, int n_end)
330
BDRVQcowState *s = bs->opaque;
336
* If this is the last cluster and it is only partially used, we must only
337
* copy until the end of the image, or bdrv_check_request will fail for the
338
* bdrv_read/write calls below.
340
if (start_sect + n_end > bs->total_sectors) {
341
n_end = bs->total_sectors - start_sect;
349
iov.iov_len = n * BDRV_SECTOR_SIZE;
350
iov.iov_base = qemu_blockalign(bs, iov.iov_len);
352
qemu_iovec_init_external(&qiov, &iov, 1);
354
BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
356
/* Call .bdrv_co_readv() directly instead of using the public block-layer
357
* interface. This avoids double I/O throttling and request tracking,
358
* which can lead to deadlock when block layer copy-on-read is enabled.
360
ret = bs->drv->bdrv_co_readv(bs, start_sect + n_start, n, &qiov);
365
if (s->crypt_method) {
366
qcow2_encrypt_sectors(s, start_sect + n_start,
367
iov.iov_base, iov.iov_base, n, 1,
368
&s->aes_encrypt_key);
371
BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
372
ret = bdrv_co_writev(bs->file, (cluster_offset >> 9) + n_start, n, &qiov);
379
qemu_vfree(iov.iov_base);
387
* For a given offset of the disk image, find the cluster offset in
388
* qcow2 file. The offset is stored in *cluster_offset.
390
* on entry, *num is the number of contiguous sectors we'd like to
391
* access following offset.
393
* on exit, *num is the number of contiguous sectors we can read.
395
* Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
398
int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
399
int *num, uint64_t *cluster_offset)
401
BDRVQcowState *s = bs->opaque;
402
unsigned int l2_index;
403
uint64_t l1_index, l2_offset, *l2_table;
405
unsigned int index_in_cluster, nb_clusters;
406
uint64_t nb_available, nb_needed;
409
index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1);
410
nb_needed = *num + index_in_cluster;
412
l1_bits = s->l2_bits + s->cluster_bits;
414
/* compute how many bytes there are between the offset and
415
* the end of the l1 entry
418
nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1));
420
/* compute the number of available sectors */
422
nb_available = (nb_available >> 9) + index_in_cluster;
424
if (nb_needed > nb_available) {
425
nb_needed = nb_available;
430
/* seek the the l2 offset in the l1 table */
432
l1_index = offset >> l1_bits;
433
if (l1_index >= s->l1_size) {
434
ret = QCOW2_CLUSTER_UNALLOCATED;
438
l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
440
ret = QCOW2_CLUSTER_UNALLOCATED;
444
/* load the l2 table in memory */
446
ret = l2_load(bs, l2_offset, &l2_table);
451
/* find the cluster offset for the given disk offset */
453
l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
454
*cluster_offset = be64_to_cpu(l2_table[l2_index]);
455
nb_clusters = size_to_clusters(s, nb_needed << 9);
457
ret = qcow2_get_cluster_type(*cluster_offset);
459
case QCOW2_CLUSTER_COMPRESSED:
460
/* Compressed clusters can only be processed one by one */
462
*cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
464
case QCOW2_CLUSTER_ZERO:
465
if (s->qcow_version < 3) {
468
c = count_contiguous_clusters(nb_clusters, s->cluster_size,
469
&l2_table[l2_index], 0,
470
QCOW_OFLAG_COMPRESSED | QCOW_OFLAG_ZERO);
473
case QCOW2_CLUSTER_UNALLOCATED:
474
/* how many empty clusters ? */
475
c = count_contiguous_free_clusters(nb_clusters, &l2_table[l2_index]);
478
case QCOW2_CLUSTER_NORMAL:
479
/* how many allocated clusters ? */
480
c = count_contiguous_clusters(nb_clusters, s->cluster_size,
481
&l2_table[l2_index], 0,
482
QCOW_OFLAG_COMPRESSED | QCOW_OFLAG_ZERO);
483
*cluster_offset &= L2E_OFFSET_MASK;
489
qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
491
nb_available = (c * s->cluster_sectors);
494
if (nb_available > nb_needed)
495
nb_available = nb_needed;
497
*num = nb_available - index_in_cluster;
505
* for a given disk offset, load (and allocate if needed)
508
* the l2 table offset in the qcow2 file and the cluster index
509
* in the l2 table are given to the caller.
511
* Returns 0 on success, -errno in failure case
513
static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
514
uint64_t **new_l2_table,
517
BDRVQcowState *s = bs->opaque;
518
unsigned int l2_index;
519
uint64_t l1_index, l2_offset;
520
uint64_t *l2_table = NULL;
523
/* seek the the l2 offset in the l1 table */
525
l1_index = offset >> (s->l2_bits + s->cluster_bits);
526
if (l1_index >= s->l1_size) {
527
ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
533
assert(l1_index < s->l1_size);
534
l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
536
/* seek the l2 table of the given l2 offset */
538
if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) {
539
/* load the l2 table in memory */
540
ret = l2_load(bs, l2_offset, &l2_table);
545
/* First allocate a new L2 table (and do COW if needed) */
546
ret = l2_allocate(bs, l1_index, &l2_table);
551
/* Then decrease the refcount of the old table */
553
qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
554
QCOW2_DISCARD_OTHER);
558
/* find the cluster offset for the given disk offset */
560
l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
562
*new_l2_table = l2_table;
563
*new_l2_index = l2_index;
569
* alloc_compressed_cluster_offset
571
* For a given offset of the disk image, return cluster offset in
574
* If the offset is not found, allocate a new compressed cluster.
576
* Return the cluster offset if successful,
577
* Return 0, otherwise.
581
uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
585
BDRVQcowState *s = bs->opaque;
588
int64_t cluster_offset;
591
ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
596
/* Compression can't overwrite anything. Fail if the cluster was already
598
cluster_offset = be64_to_cpu(l2_table[l2_index]);
599
if (cluster_offset & L2E_OFFSET_MASK) {
600
qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
604
cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
605
if (cluster_offset < 0) {
606
qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
610
nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) -
611
(cluster_offset >> 9);
613
cluster_offset |= QCOW_OFLAG_COMPRESSED |
614
((uint64_t)nb_csectors << s->csize_shift);
616
/* update L2 table */
618
/* compressed clusters never have the copied flag */
620
BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
621
qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
622
l2_table[l2_index] = cpu_to_be64(cluster_offset);
623
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
628
return cluster_offset;
631
static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r)
633
BDRVQcowState *s = bs->opaque;
636
if (r->nb_sectors == 0) {
640
qemu_co_mutex_unlock(&s->lock);
641
ret = copy_sectors(bs, m->offset / BDRV_SECTOR_SIZE, m->alloc_offset,
642
r->offset / BDRV_SECTOR_SIZE,
643
r->offset / BDRV_SECTOR_SIZE + r->nb_sectors);
644
qemu_co_mutex_lock(&s->lock);
651
* Before we update the L2 table to actually point to the new cluster, we
652
* need to be sure that the refcounts have been increased and COW was
655
qcow2_cache_depends_on_flush(s->l2_table_cache);
660
int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
662
BDRVQcowState *s = bs->opaque;
663
int i, j = 0, l2_index, ret;
664
uint64_t *old_cluster, *l2_table;
665
uint64_t cluster_offset = m->alloc_offset;
667
trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
668
assert(m->nb_clusters > 0);
670
old_cluster = g_malloc(m->nb_clusters * sizeof(uint64_t));
672
/* copy content of unmodified sectors */
673
ret = perform_cow(bs, m, &m->cow_start);
678
ret = perform_cow(bs, m, &m->cow_end);
683
/* Update L2 table. */
684
if (s->use_lazy_refcounts) {
685
qcow2_mark_dirty(bs);
687
if (qcow2_need_accurate_refcounts(s)) {
688
qcow2_cache_set_dependency(bs, s->l2_table_cache,
689
s->refcount_block_cache);
692
ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index);
696
qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
698
for (i = 0; i < m->nb_clusters; i++) {
699
/* if two concurrent writes happen to the same unallocated cluster
700
* each write allocates separate cluster and writes data concurrently.
701
* The first one to complete updates l2 table with pointer to its
702
* cluster the second one has to do RMW (which is done above by
703
* copy_sectors()), update l2 table with its cluster pointer and free
704
* old cluster. This is what this loop does */
705
if(l2_table[l2_index + i] != 0)
706
old_cluster[j++] = l2_table[l2_index + i];
708
l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
709
(i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
713
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
719
* If this was a COW, we need to decrease the refcount of the old cluster.
720
* Also flush bs->file to get the right order for L2 and refcount update.
722
* Don't discard clusters that reach a refcount of 0 (e.g. compressed
723
* clusters), the next write will reuse them anyway.
726
for (i = 0; i < j; i++) {
727
qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
728
QCOW2_DISCARD_NEVER);
739
* Returns the number of contiguous clusters that can be used for an allocating
740
* write, but require COW to be performed (this includes yet unallocated space,
741
* which must copy from the backing file)
743
static int count_cow_clusters(BDRVQcowState *s, int nb_clusters,
744
uint64_t *l2_table, int l2_index)
748
for (i = 0; i < nb_clusters; i++) {
749
uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]);
750
int cluster_type = qcow2_get_cluster_type(l2_entry);
752
switch(cluster_type) {
753
case QCOW2_CLUSTER_NORMAL:
754
if (l2_entry & QCOW_OFLAG_COPIED) {
758
case QCOW2_CLUSTER_UNALLOCATED:
759
case QCOW2_CLUSTER_COMPRESSED:
760
case QCOW2_CLUSTER_ZERO:
768
assert(i <= nb_clusters);
773
* Check if there already is an AIO write request in flight which allocates
774
* the same cluster. In this case we need to wait until the previous
775
* request has completed and updated the L2 table accordingly.
778
* 0 if there was no dependency. *cur_bytes indicates the number of
779
* bytes from guest_offset that can be read before the next
780
* dependency must be processed (or the request is complete)
782
* -EAGAIN if we had to wait for another request, previously gathered
783
* information on cluster allocation may be invalid now. The caller
784
* must start over anyway, so consider *cur_bytes undefined.
786
static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
787
uint64_t *cur_bytes, QCowL2Meta **m)
789
BDRVQcowState *s = bs->opaque;
790
QCowL2Meta *old_alloc;
791
uint64_t bytes = *cur_bytes;
793
QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
795
uint64_t start = guest_offset;
796
uint64_t end = start + bytes;
797
uint64_t old_start = l2meta_cow_start(old_alloc);
798
uint64_t old_end = l2meta_cow_end(old_alloc);
800
if (end <= old_start || start >= old_end) {
801
/* No intersection */
803
if (start < old_start) {
804
/* Stop at the start of a running allocation */
805
bytes = old_start - start;
810
/* Stop if already an l2meta exists. After yielding, it wouldn't
811
* be valid any more, so we'd have to clean up the old L2Metas
812
* and deal with requests depending on them before starting to
813
* gather new ones. Not worth the trouble. */
814
if (bytes == 0 && *m) {
820
/* Wait for the dependency to complete. We need to recheck
821
* the free/allocated clusters when we continue. */
822
qemu_co_mutex_unlock(&s->lock);
823
qemu_co_queue_wait(&old_alloc->dependent_requests);
824
qemu_co_mutex_lock(&s->lock);
830
/* Make sure that existing clusters and new allocations are only used up to
831
* the next dependency if we shortened the request above */
838
* Checks how many already allocated clusters that don't require a copy on
839
* write there are at the given guest_offset (up to *bytes). If
840
* *host_offset is not zero, only physically contiguous clusters beginning at
841
* this host offset are counted.
843
* Note that guest_offset may not be cluster aligned. In this case, the
844
* returned *host_offset points to exact byte referenced by guest_offset and
845
* therefore isn't cluster aligned as well.
848
* 0: if no allocated clusters are available at the given offset.
849
* *bytes is normally unchanged. It is set to 0 if the cluster
850
* is allocated and doesn't need COW, but doesn't have the right
853
* 1: if allocated clusters that don't require a COW are available at
854
* the requested offset. *bytes may have decreased and describes
855
* the length of the area that can be written to.
857
* -errno: in error cases
859
static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
860
uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
862
BDRVQcowState *s = bs->opaque;
864
uint64_t cluster_offset;
866
unsigned int nb_clusters;
867
unsigned int keep_clusters;
870
trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
873
assert(*host_offset == 0 || offset_into_cluster(s, guest_offset)
874
== offset_into_cluster(s, *host_offset));
877
* Calculate the number of clusters to look for. We stop at L2 table
878
* boundaries to keep things simple.
881
size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
883
l2_index = offset_to_l2_index(s, guest_offset);
884
nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
886
/* Find L2 entry for the first involved cluster */
887
ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
892
cluster_offset = be64_to_cpu(l2_table[l2_index]);
894
/* Check how many clusters are already allocated and don't need COW */
895
if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL
896
&& (cluster_offset & QCOW_OFLAG_COPIED))
898
/* If a specific host_offset is required, check it */
899
bool offset_matches =
900
(cluster_offset & L2E_OFFSET_MASK) == *host_offset;
902
if (*host_offset != 0 && !offset_matches) {
908
/* We keep all QCOW_OFLAG_COPIED clusters */
910
count_contiguous_clusters(nb_clusters, s->cluster_size,
911
&l2_table[l2_index], 0,
912
QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
913
assert(keep_clusters <= nb_clusters);
916
keep_clusters * s->cluster_size
917
- offset_into_cluster(s, guest_offset));
926
pret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
931
/* Only return a host offset if we actually made progress. Otherwise we
932
* would make requirements for handle_alloc() that it can't fulfill */
934
*host_offset = (cluster_offset & L2E_OFFSET_MASK)
935
+ offset_into_cluster(s, guest_offset);
942
* Allocates new clusters for the given guest_offset.
944
* At most *nb_clusters are allocated, and on return *nb_clusters is updated to
945
* contain the number of clusters that have been allocated and are contiguous
948
* If *host_offset is non-zero, it specifies the offset in the image file at
949
* which the new clusters must start. *nb_clusters can be 0 on return in this
950
* case if the cluster at host_offset is already in use. If *host_offset is
951
* zero, the clusters can be allocated anywhere in the image file.
953
* *host_offset is updated to contain the offset into the image file at which
954
* the first allocated cluster starts.
956
* Return 0 on success and -errno in error cases. -EAGAIN means that the
957
* function has been waiting for another request and the allocation must be
958
* restarted, but the whole request should not be failed.
960
static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
961
uint64_t *host_offset, unsigned int *nb_clusters)
963
BDRVQcowState *s = bs->opaque;
965
trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
966
*host_offset, *nb_clusters);
968
/* Allocate new clusters */
969
trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
970
if (*host_offset == 0) {
971
int64_t cluster_offset =
972
qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
973
if (cluster_offset < 0) {
974
return cluster_offset;
976
*host_offset = cluster_offset;
979
int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
989
* Allocates new clusters for an area that either is yet unallocated or needs a
990
* copy on write. If *host_offset is non-zero, clusters are only allocated if
991
* the new allocation can match the specified host offset.
993
* Note that guest_offset may not be cluster aligned. In this case, the
994
* returned *host_offset points to exact byte referenced by guest_offset and
995
* therefore isn't cluster aligned as well.
998
* 0: if no clusters could be allocated. *bytes is set to 0,
999
* *host_offset is left unchanged.
1001
* 1: if new clusters were allocated. *bytes may be decreased if the
1002
* new allocation doesn't cover all of the requested area.
1003
* *host_offset is updated to contain the host offset of the first
1004
* newly allocated cluster.
1006
* -errno: in error cases
1008
static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1009
uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1011
BDRVQcowState *s = bs->opaque;
1015
unsigned int nb_clusters;
1018
uint64_t alloc_cluster_offset;
1020
trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1025
* Calculate the number of clusters to look for. We stop at L2 table
1026
* boundaries to keep things simple.
1029
size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1031
l2_index = offset_to_l2_index(s, guest_offset);
1032
nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1034
/* Find L2 entry for the first involved cluster */
1035
ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1040
entry = be64_to_cpu(l2_table[l2_index]);
1042
/* For the moment, overwrite compressed clusters one by one */
1043
if (entry & QCOW_OFLAG_COMPRESSED) {
1046
nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index);
1049
/* This function is only called when there were no non-COW clusters, so if
1050
* we can't find any unallocated or COW clusters either, something is
1051
* wrong with our code. */
1052
assert(nb_clusters > 0);
1054
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1059
/* Allocate, if necessary at a given offset in the image file */
1060
alloc_cluster_offset = start_of_cluster(s, *host_offset);
1061
ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1067
/* Can't extend contiguous allocation */
1068
if (nb_clusters == 0) {
1074
* Save info needed for meta data update.
1076
* requested_sectors: Number of sectors from the start of the first
1077
* newly allocated cluster to the end of the (possibly shortened
1078
* before) write request.
1080
* avail_sectors: Number of sectors from the start of the first
1081
* newly allocated to the end of the last newly allocated cluster.
1083
* nb_sectors: The number of sectors from the start of the first
1084
* newly allocated cluster to the end of the area that the write
1085
* request actually writes to (excluding COW at the end)
1087
int requested_sectors =
1088
(*bytes + offset_into_cluster(s, guest_offset))
1089
>> BDRV_SECTOR_BITS;
1090
int avail_sectors = nb_clusters
1091
<< (s->cluster_bits - BDRV_SECTOR_BITS);
1092
int alloc_n_start = offset_into_cluster(s, guest_offset)
1093
>> BDRV_SECTOR_BITS;
1094
int nb_sectors = MIN(requested_sectors, avail_sectors);
1095
QCowL2Meta *old_m = *m;
1097
*m = g_malloc0(sizeof(**m));
1099
**m = (QCowL2Meta) {
1102
.alloc_offset = alloc_cluster_offset,
1103
.offset = start_of_cluster(s, guest_offset),
1104
.nb_clusters = nb_clusters,
1105
.nb_available = nb_sectors,
1109
.nb_sectors = alloc_n_start,
1112
.offset = nb_sectors * BDRV_SECTOR_SIZE,
1113
.nb_sectors = avail_sectors - nb_sectors,
1116
qemu_co_queue_init(&(*m)->dependent_requests);
1117
QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1119
*host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1120
*bytes = MIN(*bytes, (nb_sectors * BDRV_SECTOR_SIZE)
1121
- offset_into_cluster(s, guest_offset));
1122
assert(*bytes != 0);
1127
if (*m && (*m)->nb_clusters > 0) {
1128
QLIST_REMOVE(*m, next_in_flight);
1134
* alloc_cluster_offset
1136
* For a given offset on the virtual disk, find the cluster offset in qcow2
1137
* file. If the offset is not found, allocate a new cluster.
1139
* If the cluster was already allocated, m->nb_clusters is set to 0 and
1140
* other fields in m are meaningless.
1142
* If the cluster is newly allocated, m->nb_clusters is set to the number of
1143
* contiguous clusters that have been allocated. In this case, the other
1144
* fields of m are valid and contain information about the first allocated
1147
* If the request conflicts with another write request in flight, the coroutine
1148
* is queued and will be reentered when the dependency has completed.
1150
* Return 0 on success and -errno in error cases
1152
int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1153
int n_start, int n_end, int *num, uint64_t *host_offset, QCowL2Meta **m)
1155
BDRVQcowState *s = bs->opaque;
1156
uint64_t start, remaining;
1157
uint64_t cluster_offset;
1161
trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset,
1164
assert(n_start * BDRV_SECTOR_SIZE == offset_into_cluster(s, offset));
1165
offset = start_of_cluster(s, offset);
1168
start = offset + (n_start << BDRV_SECTOR_BITS);
1169
remaining = (n_end - n_start) << BDRV_SECTOR_BITS;
1177
if (!*host_offset) {
1178
*host_offset = start_of_cluster(s, cluster_offset);
1181
assert(remaining >= cur_bytes);
1184
remaining -= cur_bytes;
1185
cluster_offset += cur_bytes;
1187
if (remaining == 0) {
1191
cur_bytes = remaining;
1194
* Now start gathering as many contiguous clusters as possible:
1196
* 1. Check for overlaps with in-flight allocations
1198
* a) Overlap not in the first cluster -> shorten this request and
1199
* let the caller handle the rest in its next loop iteration.
1201
* b) Real overlaps of two requests. Yield and restart the search
1202
* for contiguous clusters (the situation could have changed
1203
* while we were sleeping)
1205
* c) TODO: Request starts in the same cluster as the in-flight
1206
* allocation ends. Shorten the COW of the in-fight allocation,
1207
* set cluster_offset to write to the same cluster and set up
1208
* the right synchronisation between the in-flight request and
1211
ret = handle_dependencies(bs, start, &cur_bytes, m);
1212
if (ret == -EAGAIN) {
1213
/* Currently handle_dependencies() doesn't yield if we already had
1214
* an allocation. If it did, we would have to clean up the L2Meta
1215
* structs before starting over. */
1218
} else if (ret < 0) {
1220
} else if (cur_bytes == 0) {
1223
/* handle_dependencies() may have decreased cur_bytes (shortened
1224
* the allocations below) so that the next dependency is processed
1225
* correctly during the next loop iteration. */
1229
* 2. Count contiguous COPIED clusters.
1231
ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1236
} else if (cur_bytes == 0) {
1241
* 3. If the request still hasn't completed, allocate new clusters,
1242
* considering any cluster_offset of steps 1c or 2.
1244
ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1250
assert(cur_bytes == 0);
1255
*num = (n_end - n_start) - (remaining >> BDRV_SECTOR_BITS);
1257
assert(*host_offset != 0);
1262
static int decompress_buffer(uint8_t *out_buf, int out_buf_size,
1263
const uint8_t *buf, int buf_size)
1265
z_stream strm1, *strm = &strm1;
1268
memset(strm, 0, sizeof(*strm));
1270
strm->next_in = (uint8_t *)buf;
1271
strm->avail_in = buf_size;
1272
strm->next_out = out_buf;
1273
strm->avail_out = out_buf_size;
1275
ret = inflateInit2(strm, -12);
1278
ret = inflate(strm, Z_FINISH);
1279
out_len = strm->next_out - out_buf;
1280
if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) ||
1281
out_len != out_buf_size) {
1289
int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
1291
BDRVQcowState *s = bs->opaque;
1292
int ret, csize, nb_csectors, sector_offset;
1295
coffset = cluster_offset & s->cluster_offset_mask;
1296
if (s->cluster_cache_offset != coffset) {
1297
nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1;
1298
sector_offset = coffset & 511;
1299
csize = nb_csectors * 512 - sector_offset;
1300
BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);
1301
ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, nb_csectors);
1305
if (decompress_buffer(s->cluster_cache, s->cluster_size,
1306
s->cluster_data + sector_offset, csize) < 0) {
1309
s->cluster_cache_offset = coffset;
1315
* This discards as many clusters of nb_clusters as possible at once (i.e.
1316
* all clusters in the same L2 table) and returns the number of discarded
1319
static int discard_single_l2(BlockDriverState *bs, uint64_t offset,
1320
unsigned int nb_clusters)
1322
BDRVQcowState *s = bs->opaque;
1328
ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1333
/* Limit nb_clusters to one L2 table */
1334
nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1336
for (i = 0; i < nb_clusters; i++) {
1337
uint64_t old_offset;
1339
old_offset = be64_to_cpu(l2_table[l2_index + i]);
1340
if ((old_offset & L2E_OFFSET_MASK) == 0) {
1344
/* First remove L2 entries */
1345
qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1346
l2_table[l2_index + i] = cpu_to_be64(0);
1348
/* Then decrease the refcount */
1349
qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1352
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1360
int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
1363
BDRVQcowState *s = bs->opaque;
1364
uint64_t end_offset;
1365
unsigned int nb_clusters;
1368
end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
1370
/* Round start up and end down */
1371
offset = align_offset(offset, s->cluster_size);
1372
end_offset &= ~(s->cluster_size - 1);
1374
if (offset > end_offset) {
1378
nb_clusters = size_to_clusters(s, end_offset - offset);
1380
s->cache_discards = true;
1382
/* Each L2 table is handled by its own loop iteration */
1383
while (nb_clusters > 0) {
1384
ret = discard_single_l2(bs, offset, nb_clusters);
1390
offset += (ret * s->cluster_size);
1395
s->cache_discards = false;
1396
qcow2_process_discards(bs, ret);
1402
* This zeroes as many clusters of nb_clusters as possible at once (i.e.
1403
* all clusters in the same L2 table) and returns the number of zeroed
1406
static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
1407
unsigned int nb_clusters)
1409
BDRVQcowState *s = bs->opaque;
1415
ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1420
/* Limit nb_clusters to one L2 table */
1421
nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1423
for (i = 0; i < nb_clusters; i++) {
1424
uint64_t old_offset;
1426
old_offset = be64_to_cpu(l2_table[l2_index + i]);
1428
/* Update L2 entries */
1429
qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1430
if (old_offset & QCOW_OFLAG_COMPRESSED) {
1431
l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1432
qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1434
l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1438
ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1446
int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors)
1448
BDRVQcowState *s = bs->opaque;
1449
unsigned int nb_clusters;
1452
/* The zero flag is only supported by version 3 and newer */
1453
if (s->qcow_version < 3) {
1457
/* Each L2 table is handled by its own loop iteration */
1458
nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS);
1460
s->cache_discards = true;
1462
while (nb_clusters > 0) {
1463
ret = zero_single_l2(bs, offset, nb_clusters);
1469
offset += (ret * s->cluster_size);
1474
s->cache_discards = false;
1475
qcow2_process_discards(bs, ret);