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 "block/block-io.h"
29 #include "qapi/error.h"
31 #include "qemu/bswap.h"
32 #include "qemu/memalign.h"
35 int coroutine_fn qcow2_shrink_l1_table(BlockDriverState *bs,
38 BDRVQcow2State *s = bs->opaque;
39 int new_l1_size, i, ret;
41 if (exact_size >= s->l1_size) {
45 new_l1_size = exact_size;
48 fprintf(stderr, "shrink l1_table from %d to %d\n", s->l1_size, new_l1_size);
51 BLKDBG_CO_EVENT(bs->file, BLKDBG_L1_SHRINK_WRITE_TABLE);
52 ret = bdrv_co_pwrite_zeroes(bs->file,
53 s->l1_table_offset + new_l1_size * L1E_SIZE,
54 (s->l1_size - new_l1_size) * L1E_SIZE, 0);
59 ret = bdrv_co_flush(bs->file->bs);
64 BLKDBG_CO_EVENT(bs->file, BLKDBG_L1_SHRINK_FREE_L2_CLUSTERS);
65 for (i = s->l1_size - 1; i > new_l1_size - 1; i--) {
66 if ((s->l1_table[i] & L1E_OFFSET_MASK) == 0) {
69 qcow2_free_clusters(bs, s->l1_table[i] & L1E_OFFSET_MASK,
70 s->cluster_size, QCOW2_DISCARD_ALWAYS);
77 * If the write in the l1_table failed the image may contain a partially
78 * overwritten l1_table. In this case it would be better to clear the
79 * l1_table in memory to avoid possible image corruption.
81 memset(s->l1_table + new_l1_size, 0,
82 (s->l1_size - new_l1_size) * L1E_SIZE);
86 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
89 BDRVQcow2State *s = bs->opaque;
90 int new_l1_size2, ret, i;
91 uint64_t *new_l1_table;
92 int64_t old_l1_table_offset, old_l1_size;
93 int64_t new_l1_table_offset, new_l1_size;
96 if (min_size <= s->l1_size)
99 /* Do a sanity check on min_size before trying to calculate new_l1_size
100 * (this prevents overflows during the while loop for the calculation of
102 if (min_size > INT_MAX / L1E_SIZE) {
107 new_l1_size = min_size;
109 /* Bump size up to reduce the number of times we have to grow */
110 new_l1_size = s->l1_size;
111 if (new_l1_size == 0) {
114 while (min_size > new_l1_size) {
115 new_l1_size = DIV_ROUND_UP(new_l1_size * 3, 2);
119 QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX);
120 if (new_l1_size > QCOW_MAX_L1_SIZE / L1E_SIZE) {
125 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
126 s->l1_size, new_l1_size);
129 new_l1_size2 = L1E_SIZE * new_l1_size;
130 new_l1_table = qemu_try_blockalign(bs->file->bs, new_l1_size2);
131 if (new_l1_table == NULL) {
134 memset(new_l1_table, 0, new_l1_size2);
137 memcpy(new_l1_table, s->l1_table, s->l1_size * L1E_SIZE);
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, 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),
176 sizeof(data), data, 0);
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 * L1E_SIZE,
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 GRAPH_RDLOCK
211 l2_load(BlockDriverState *bs, uint64_t offset,
212 uint64_t l2_offset, uint64_t **l2_slice)
214 BDRVQcow2State *s = bs->opaque;
215 int start_of_slice = l2_entry_size(s) *
216 (offset_to_l2_index(s, offset) - offset_to_l2_slice_index(s, offset));
218 return qcow2_cache_get(bs, s->l2_table_cache, l2_offset + start_of_slice,
223 * Writes an L1 entry to disk (note that depending on the alignment
224 * requirements this function may write more that just one entry in
225 * order to prevent bdrv_pwrite from performing a read-modify-write)
227 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
229 BDRVQcow2State *s = bs->opaque;
232 int bufsize = MAX(L1E_SIZE,
233 MIN(bs->file->bs->bl.request_alignment, s->cluster_size));
234 int nentries = bufsize / L1E_SIZE;
235 g_autofree uint64_t *buf = g_try_new0(uint64_t, nentries);
241 l1_start_index = QEMU_ALIGN_DOWN(l1_index, nentries);
242 for (i = 0; i < MIN(nentries, s->l1_size - l1_start_index); i++) {
243 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
246 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
247 s->l1_table_offset + L1E_SIZE * l1_start_index, bufsize, false);
252 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
253 ret = bdrv_pwrite_sync(bs->file,
254 s->l1_table_offset + L1E_SIZE * l1_start_index,
266 * Allocate a new l2 entry in the file. If l1_index points to an already
267 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
268 * table) copy the contents of the old L2 table into the newly allocated one.
269 * Otherwise the new table is initialized with zeros.
273 static int GRAPH_RDLOCK l2_allocate(BlockDriverState *bs, int l1_index)
275 BDRVQcow2State *s = bs->opaque;
276 uint64_t old_l2_offset;
277 uint64_t *l2_slice = NULL;
278 unsigned slice, slice_size2, n_slices;
282 old_l2_offset = s->l1_table[l1_index];
284 trace_qcow2_l2_allocate(bs, l1_index);
286 /* allocate a new l2 entry */
288 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * l2_entry_size(s));
294 /* The offset must fit in the offset field of the L1 table entry */
295 assert((l2_offset & L1E_OFFSET_MASK) == l2_offset);
297 /* If we're allocating the table at offset 0 then something is wrong */
298 if (l2_offset == 0) {
299 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
300 "allocation of L2 table at offset 0");
305 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
310 /* allocate a new entry in the l2 cache */
312 slice_size2 = s->l2_slice_size * l2_entry_size(s);
313 n_slices = s->cluster_size / slice_size2;
315 trace_qcow2_l2_allocate_get_empty(bs, l1_index);
316 for (slice = 0; slice < n_slices; slice++) {
317 ret = qcow2_cache_get_empty(bs, s->l2_table_cache,
318 l2_offset + slice * slice_size2,
319 (void **) &l2_slice);
324 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
325 /* if there was no old l2 table, clear the new slice */
326 memset(l2_slice, 0, slice_size2);
329 uint64_t old_l2_slice_offset =
330 (old_l2_offset & L1E_OFFSET_MASK) + slice * slice_size2;
332 /* if there was an old l2 table, read a slice from the disk */
333 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
334 ret = qcow2_cache_get(bs, s->l2_table_cache, old_l2_slice_offset,
335 (void **) &old_slice);
340 memcpy(l2_slice, old_slice, slice_size2);
342 qcow2_cache_put(s->l2_table_cache, (void **) &old_slice);
345 /* write the l2 slice to the file */
346 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
348 trace_qcow2_l2_allocate_write_l2(bs, l1_index);
349 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
350 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
353 ret = qcow2_cache_flush(bs, s->l2_table_cache);
358 /* update the L1 entry */
359 trace_qcow2_l2_allocate_write_l1(bs, l1_index);
360 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
361 ret = qcow2_write_l1_entry(bs, l1_index);
366 trace_qcow2_l2_allocate_done(bs, l1_index, 0);
370 trace_qcow2_l2_allocate_done(bs, l1_index, ret);
371 if (l2_slice != NULL) {
372 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
374 s->l1_table[l1_index] = old_l2_offset;
376 qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
377 QCOW2_DISCARD_ALWAYS);
383 * For a given L2 entry, count the number of contiguous subclusters of
384 * the same type starting from @sc_from. Compressed clusters are
385 * treated as if they were divided into subclusters of size
386 * s->subcluster_size.
388 * Return the number of contiguous subclusters and set @type to the
391 * If the L2 entry is invalid return -errno and set @type to
392 * QCOW2_SUBCLUSTER_INVALID.
394 static int qcow2_get_subcluster_range_type(BlockDriverState *bs,
398 QCow2SubclusterType *type)
400 BDRVQcow2State *s = bs->opaque;
403 *type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_from);
405 if (*type == QCOW2_SUBCLUSTER_INVALID) {
407 } else if (!has_subclusters(s) || *type == QCOW2_SUBCLUSTER_COMPRESSED) {
408 return s->subclusters_per_cluster - sc_from;
412 case QCOW2_SUBCLUSTER_NORMAL:
413 val = l2_bitmap | QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
414 return cto32(val) - sc_from;
416 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
417 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
418 val = (l2_bitmap | QCOW_OFLAG_SUB_ZERO_RANGE(0, sc_from)) >> 32;
419 return cto32(val) - sc_from;
421 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
422 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
423 val = ((l2_bitmap >> 32) | l2_bitmap)
424 & ~QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
425 return ctz32(val) - sc_from;
428 g_assert_not_reached();
433 * Return the number of contiguous subclusters of the exact same type
434 * in a given L2 slice, starting from cluster @l2_index, subcluster
435 * @sc_index. Allocated subclusters are required to be contiguous in
437 * At most @nb_clusters are checked (note that this means clusters,
439 * Compressed clusters are always processed one by one but for the
440 * purpose of this count they are treated as if they were divided into
441 * subclusters of size s->subcluster_size.
442 * On failure return -errno and update @l2_index to point to the
445 static int count_contiguous_subclusters(BlockDriverState *bs, int nb_clusters,
446 unsigned sc_index, uint64_t *l2_slice,
449 BDRVQcow2State *s = bs->opaque;
451 bool check_offset = false;
452 uint64_t expected_offset = 0;
453 QCow2SubclusterType expected_type = QCOW2_SUBCLUSTER_NORMAL, type;
455 assert(*l2_index + nb_clusters <= s->l2_slice_size);
457 for (i = 0; i < nb_clusters; i++) {
458 unsigned first_sc = (i == 0) ? sc_index : 0;
459 uint64_t l2_entry = get_l2_entry(s, l2_slice, *l2_index + i);
460 uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, *l2_index + i);
461 int ret = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
464 *l2_index += i; /* Point to the invalid entry */
468 if (type == QCOW2_SUBCLUSTER_COMPRESSED) {
469 /* Compressed clusters are always processed one by one */
472 expected_type = type;
473 expected_offset = l2_entry & L2E_OFFSET_MASK;
474 check_offset = (type == QCOW2_SUBCLUSTER_NORMAL ||
475 type == QCOW2_SUBCLUSTER_ZERO_ALLOC ||
476 type == QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC);
477 } else if (type != expected_type) {
479 } else if (check_offset) {
480 expected_offset += s->cluster_size;
481 if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
486 /* Stop if there are type changes before the end of the cluster */
487 if (first_sc + ret < s->subclusters_per_cluster) {
495 static int coroutine_fn GRAPH_RDLOCK
496 do_perform_cow_read(BlockDriverState *bs, uint64_t src_cluster_offset,
497 unsigned offset_in_cluster, QEMUIOVector *qiov)
501 if (qiov->size == 0) {
505 BLKDBG_CO_EVENT(bs->file, BLKDBG_COW_READ);
512 * We never deal with requests that don't satisfy
513 * bdrv_check_qiov_request(), and aligning requests to clusters never
514 * breaks this condition. So, do some assertions before calling
515 * bs->drv->bdrv_co_preadv_part() which has int64_t arguments.
517 assert(src_cluster_offset <= INT64_MAX);
518 assert(src_cluster_offset + offset_in_cluster <= INT64_MAX);
519 /* Cast qiov->size to uint64_t to silence a compiler warning on -m32 */
520 assert((uint64_t)qiov->size <= INT64_MAX);
521 bdrv_check_qiov_request(src_cluster_offset + offset_in_cluster, qiov->size,
522 qiov, 0, &error_abort);
524 * Call .bdrv_co_readv() directly instead of using the public block-layer
525 * interface. This avoids double I/O throttling and request tracking,
526 * which can lead to deadlock when block layer copy-on-read is enabled.
528 ret = bs->drv->bdrv_co_preadv_part(bs,
529 src_cluster_offset + offset_in_cluster,
530 qiov->size, qiov, 0, 0);
538 static int coroutine_fn GRAPH_RDLOCK
539 do_perform_cow_write(BlockDriverState *bs, uint64_t cluster_offset,
540 unsigned offset_in_cluster, QEMUIOVector *qiov)
542 BDRVQcow2State *s = bs->opaque;
545 if (qiov->size == 0) {
549 ret = qcow2_pre_write_overlap_check(bs, 0,
550 cluster_offset + offset_in_cluster, qiov->size, true);
555 BLKDBG_CO_EVENT(bs->file, BLKDBG_COW_WRITE);
556 ret = bdrv_co_pwritev(s->data_file, cluster_offset + offset_in_cluster,
557 qiov->size, qiov, 0);
569 * For a given offset of the virtual disk find the equivalent host
570 * offset in the qcow2 file and store it in *host_offset. Neither
571 * offset needs to be aligned to a cluster boundary.
573 * If the cluster is unallocated then *host_offset will be 0.
574 * If the cluster is compressed then *host_offset will contain the l2 entry.
576 * On entry, *bytes is the maximum number of contiguous bytes starting at
577 * offset that we are interested in.
579 * On exit, *bytes is the number of bytes starting at offset that have the same
580 * subcluster type and (if applicable) are stored contiguously in the image
581 * file. The subcluster type is stored in *subcluster_type.
582 * Compressed clusters are always processed one by one.
584 * Returns 0 on success, -errno in error cases.
586 int qcow2_get_host_offset(BlockDriverState *bs, uint64_t offset,
587 unsigned int *bytes, uint64_t *host_offset,
588 QCow2SubclusterType *subcluster_type)
590 BDRVQcow2State *s = bs->opaque;
591 unsigned int l2_index, sc_index;
592 uint64_t l1_index, l2_offset, *l2_slice, l2_entry, l2_bitmap;
594 unsigned int offset_in_cluster;
595 uint64_t bytes_available, bytes_needed, nb_clusters;
596 QCow2SubclusterType type;
599 offset_in_cluster = offset_into_cluster(s, offset);
600 bytes_needed = (uint64_t) *bytes + offset_in_cluster;
602 /* compute how many bytes there are between the start of the cluster
603 * containing offset and the end of the l2 slice that contains
604 * the entry pointing to it */
606 ((uint64_t) (s->l2_slice_size - offset_to_l2_slice_index(s, offset)))
609 if (bytes_needed > bytes_available) {
610 bytes_needed = bytes_available;
615 /* seek to the l2 offset in the l1 table */
617 l1_index = offset_to_l1_index(s, offset);
618 if (l1_index >= s->l1_size) {
619 type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
623 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
625 type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
629 if (offset_into_cluster(s, l2_offset)) {
630 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
631 " unaligned (L1 index: %#" PRIx64 ")",
632 l2_offset, l1_index);
636 /* load the l2 slice in memory */
638 ret = l2_load(bs, offset, l2_offset, &l2_slice);
643 /* find the cluster offset for the given disk offset */
645 l2_index = offset_to_l2_slice_index(s, offset);
646 sc_index = offset_to_sc_index(s, offset);
647 l2_entry = get_l2_entry(s, l2_slice, l2_index);
648 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
650 nb_clusters = size_to_clusters(s, bytes_needed);
651 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
652 * integers; the minimum cluster size is 512, so this assertion is always
654 assert(nb_clusters <= INT_MAX);
656 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
657 if (s->qcow_version < 3 && (type == QCOW2_SUBCLUSTER_ZERO_PLAIN ||
658 type == QCOW2_SUBCLUSTER_ZERO_ALLOC)) {
659 qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
660 " in pre-v3 image (L2 offset: %#" PRIx64
661 ", L2 index: %#x)", l2_offset, l2_index);
666 case QCOW2_SUBCLUSTER_INVALID:
667 break; /* This is handled by count_contiguous_subclusters() below */
668 case QCOW2_SUBCLUSTER_COMPRESSED:
669 if (has_data_file(bs)) {
670 qcow2_signal_corruption(bs, true, -1, -1, "Compressed cluster "
671 "entry found in image with external data "
672 "file (L2 offset: %#" PRIx64 ", L2 index: "
673 "%#x)", l2_offset, l2_index);
677 *host_offset = l2_entry;
679 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
680 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
682 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
683 case QCOW2_SUBCLUSTER_NORMAL:
684 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC: {
685 uint64_t host_cluster_offset = l2_entry & L2E_OFFSET_MASK;
686 *host_offset = host_cluster_offset + offset_in_cluster;
687 if (offset_into_cluster(s, host_cluster_offset)) {
688 qcow2_signal_corruption(bs, true, -1, -1,
689 "Cluster allocation offset %#"
690 PRIx64 " unaligned (L2 offset: %#" PRIx64
691 ", L2 index: %#x)", host_cluster_offset,
692 l2_offset, l2_index);
696 if (has_data_file(bs) && *host_offset != offset) {
697 qcow2_signal_corruption(bs, true, -1, -1,
698 "External data file host cluster offset %#"
699 PRIx64 " does not match guest cluster "
701 ", L2 index: %#x)", host_cluster_offset,
702 offset - offset_in_cluster, l2_index);
712 sc = count_contiguous_subclusters(bs, nb_clusters, sc_index,
713 l2_slice, &l2_index);
715 qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster entry found "
716 " (L2 offset: %#" PRIx64 ", L2 index: %#x)",
717 l2_offset, l2_index);
721 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
723 bytes_available = ((int64_t)sc + sc_index) << s->subcluster_bits;
726 if (bytes_available > bytes_needed) {
727 bytes_available = bytes_needed;
730 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
731 * subtracting offset_in_cluster will therefore definitely yield something
732 * not exceeding UINT_MAX */
733 assert(bytes_available - offset_in_cluster <= UINT_MAX);
734 *bytes = bytes_available - offset_in_cluster;
736 *subcluster_type = type;
741 qcow2_cache_put(s->l2_table_cache, (void **)&l2_slice);
748 * for a given disk offset, load (and allocate if needed)
749 * the appropriate slice of its l2 table.
751 * the cluster index in the l2 slice is given to the caller.
753 * Returns 0 on success, -errno in failure case
755 static int GRAPH_RDLOCK
756 get_cluster_table(BlockDriverState *bs, uint64_t offset,
757 uint64_t **new_l2_slice, int *new_l2_index)
759 BDRVQcow2State *s = bs->opaque;
760 unsigned int l2_index;
761 uint64_t l1_index, l2_offset;
762 uint64_t *l2_slice = NULL;
765 /* seek to the l2 offset in the l1 table */
767 l1_index = offset_to_l1_index(s, offset);
768 if (l1_index >= s->l1_size) {
769 ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
775 assert(l1_index < s->l1_size);
776 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
777 if (offset_into_cluster(s, l2_offset)) {
778 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
779 " unaligned (L1 index: %#" PRIx64 ")",
780 l2_offset, l1_index);
784 if (!(s->l1_table[l1_index] & QCOW_OFLAG_COPIED)) {
785 /* First allocate a new L2 table (and do COW if needed) */
786 ret = l2_allocate(bs, l1_index);
791 /* Then decrease the refcount of the old table */
793 qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
794 QCOW2_DISCARD_OTHER);
797 /* Get the offset of the newly-allocated l2 table */
798 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
799 assert(offset_into_cluster(s, l2_offset) == 0);
802 /* load the l2 slice in memory */
803 ret = l2_load(bs, offset, l2_offset, &l2_slice);
808 /* find the cluster offset for the given disk offset */
810 l2_index = offset_to_l2_slice_index(s, offset);
812 *new_l2_slice = l2_slice;
813 *new_l2_index = l2_index;
819 * alloc_compressed_cluster_offset
821 * For a given offset on the virtual disk, allocate a new compressed cluster
822 * and put the host offset of the cluster into *host_offset. If a cluster is
823 * already allocated at the offset, return an error.
825 * Return 0 on success and -errno in error cases
827 int coroutine_fn GRAPH_RDLOCK
828 qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs, uint64_t offset,
829 int compressed_size, uint64_t *host_offset)
831 BDRVQcow2State *s = bs->opaque;
834 int64_t cluster_offset;
837 if (has_data_file(bs)) {
841 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
846 /* Compression can't overwrite anything. Fail if the cluster was already
848 cluster_offset = get_l2_entry(s, l2_slice, l2_index);
849 if (cluster_offset & L2E_OFFSET_MASK) {
850 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
854 cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
855 if (cluster_offset < 0) {
856 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
857 return cluster_offset;
861 (cluster_offset + compressed_size - 1) / QCOW2_COMPRESSED_SECTOR_SIZE -
862 (cluster_offset / QCOW2_COMPRESSED_SECTOR_SIZE);
864 /* The offset and size must fit in their fields of the L2 table entry */
865 assert((cluster_offset & s->cluster_offset_mask) == cluster_offset);
866 assert((nb_csectors & s->csize_mask) == nb_csectors);
868 cluster_offset |= QCOW_OFLAG_COMPRESSED |
869 ((uint64_t)nb_csectors << s->csize_shift);
871 /* update L2 table */
873 /* compressed clusters never have the copied flag */
875 BLKDBG_CO_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
876 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
877 set_l2_entry(s, l2_slice, l2_index, cluster_offset);
878 if (has_subclusters(s)) {
879 set_l2_bitmap(s, l2_slice, l2_index, 0);
881 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
883 *host_offset = cluster_offset & s->cluster_offset_mask;
887 static int coroutine_fn GRAPH_RDLOCK
888 perform_cow(BlockDriverState *bs, QCowL2Meta *m)
890 BDRVQcow2State *s = bs->opaque;
891 Qcow2COWRegion *start = &m->cow_start;
892 Qcow2COWRegion *end = &m->cow_end;
893 unsigned buffer_size;
894 unsigned data_bytes = end->offset - (start->offset + start->nb_bytes);
896 uint8_t *start_buffer, *end_buffer;
900 assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
901 assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
902 assert(start->offset + start->nb_bytes <= end->offset);
904 if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
908 /* If we have to read both the start and end COW regions and the
909 * middle region is not too large then perform just one read
911 merge_reads = start->nb_bytes && end->nb_bytes && data_bytes <= 16384;
913 buffer_size = start->nb_bytes + data_bytes + end->nb_bytes;
915 /* If we have to do two reads, add some padding in the middle
916 * if necessary to make sure that the end region is optimally
918 size_t align = bdrv_opt_mem_align(bs);
919 assert(align > 0 && align <= UINT_MAX);
920 assert(QEMU_ALIGN_UP(start->nb_bytes, align) <=
921 UINT_MAX - end->nb_bytes);
922 buffer_size = QEMU_ALIGN_UP(start->nb_bytes, align) + end->nb_bytes;
925 /* Reserve a buffer large enough to store all the data that we're
927 start_buffer = qemu_try_blockalign(bs, buffer_size);
928 if (start_buffer == NULL) {
931 /* The part of the buffer where the end region is located */
932 end_buffer = start_buffer + buffer_size - end->nb_bytes;
934 qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
935 qemu_iovec_subvec_niov(m->data_qiov,
940 qemu_co_mutex_unlock(&s->lock);
941 /* First we read the existing data from both COW regions. We
942 * either read the whole region in one go, or the start and end
943 * regions separately. */
945 qemu_iovec_add(&qiov, start_buffer, buffer_size);
946 ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
948 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
949 ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
954 qemu_iovec_reset(&qiov);
955 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
956 ret = do_perform_cow_read(bs, m->offset, end->offset, &qiov);
962 /* Encrypt the data if necessary before writing it */
964 ret = qcow2_co_encrypt(bs,
965 m->alloc_offset + start->offset,
966 m->offset + start->offset,
967 start_buffer, start->nb_bytes);
972 ret = qcow2_co_encrypt(bs,
973 m->alloc_offset + end->offset,
974 m->offset + end->offset,
975 end_buffer, end->nb_bytes);
981 /* And now we can write everything. If we have the guest data we
982 * can write everything in one single operation */
984 qemu_iovec_reset(&qiov);
985 if (start->nb_bytes) {
986 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
988 qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
990 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
992 /* NOTE: we have a write_aio blkdebug event here followed by
993 * a cow_write one in do_perform_cow_write(), but there's only
994 * one single I/O operation */
995 BLKDBG_CO_EVENT(bs->file, BLKDBG_WRITE_AIO);
996 ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
998 /* If there's no guest data then write both COW regions separately */
999 qemu_iovec_reset(&qiov);
1000 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
1001 ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
1006 qemu_iovec_reset(&qiov);
1007 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
1008 ret = do_perform_cow_write(bs, m->alloc_offset, end->offset, &qiov);
1012 qemu_co_mutex_lock(&s->lock);
1015 * Before we update the L2 table to actually point to the new cluster, we
1016 * need to be sure that the refcounts have been increased and COW was
1020 qcow2_cache_depends_on_flush(s->l2_table_cache);
1023 qemu_vfree(start_buffer);
1024 qemu_iovec_destroy(&qiov);
1028 int coroutine_fn qcow2_alloc_cluster_link_l2(BlockDriverState *bs,
1031 BDRVQcow2State *s = bs->opaque;
1032 int i, j = 0, l2_index, ret;
1033 uint64_t *old_cluster, *l2_slice;
1034 uint64_t cluster_offset = m->alloc_offset;
1036 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
1037 assert(m->nb_clusters > 0);
1039 old_cluster = g_try_new(uint64_t, m->nb_clusters);
1040 if (old_cluster == NULL) {
1045 /* copy content of unmodified sectors */
1046 ret = perform_cow(bs, m);
1051 /* Update L2 table. */
1052 if (s->use_lazy_refcounts) {
1053 qcow2_mark_dirty(bs);
1055 if (qcow2_need_accurate_refcounts(s)) {
1056 qcow2_cache_set_dependency(bs, s->l2_table_cache,
1057 s->refcount_block_cache);
1060 ret = get_cluster_table(bs, m->offset, &l2_slice, &l2_index);
1064 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1066 assert(l2_index + m->nb_clusters <= s->l2_slice_size);
1067 assert(m->cow_end.offset + m->cow_end.nb_bytes <=
1068 m->nb_clusters << s->cluster_bits);
1069 for (i = 0; i < m->nb_clusters; i++) {
1070 uint64_t offset = cluster_offset + ((uint64_t)i << s->cluster_bits);
1071 /* if two concurrent writes happen to the same unallocated cluster
1072 * each write allocates separate cluster and writes data concurrently.
1073 * The first one to complete updates l2 table with pointer to its
1074 * cluster the second one has to do RMW (which is done above by
1075 * perform_cow()), update l2 table with its cluster pointer and free
1076 * old cluster. This is what this loop does */
1077 if (get_l2_entry(s, l2_slice, l2_index + i) != 0) {
1078 old_cluster[j++] = get_l2_entry(s, l2_slice, l2_index + i);
1081 /* The offset must fit in the offset field of the L2 table entry */
1082 assert((offset & L2E_OFFSET_MASK) == offset);
1084 set_l2_entry(s, l2_slice, l2_index + i, offset | QCOW_OFLAG_COPIED);
1086 /* Update bitmap with the subclusters that were just written */
1087 if (has_subclusters(s) && !m->prealloc) {
1088 uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1089 unsigned written_from = m->cow_start.offset;
1090 unsigned written_to = m->cow_end.offset + m->cow_end.nb_bytes;
1091 int first_sc, last_sc;
1092 /* Narrow written_from and written_to down to the current cluster */
1093 written_from = MAX(written_from, i << s->cluster_bits);
1094 written_to = MIN(written_to, (i + 1) << s->cluster_bits);
1095 assert(written_from < written_to);
1096 first_sc = offset_to_sc_index(s, written_from);
1097 last_sc = offset_to_sc_index(s, written_to - 1);
1098 l2_bitmap |= QCOW_OFLAG_SUB_ALLOC_RANGE(first_sc, last_sc + 1);
1099 l2_bitmap &= ~QCOW_OFLAG_SUB_ZERO_RANGE(first_sc, last_sc + 1);
1100 set_l2_bitmap(s, l2_slice, l2_index + i, l2_bitmap);
1105 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1108 * If this was a COW, we need to decrease the refcount of the old cluster.
1110 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1111 * clusters), the next write will reuse them anyway.
1113 if (!m->keep_old_clusters && j != 0) {
1114 for (i = 0; i < j; i++) {
1115 qcow2_free_any_cluster(bs, old_cluster[i], QCOW2_DISCARD_NEVER);
1121 g_free(old_cluster);
1126 * Frees the allocated clusters because the request failed and they won't
1127 * actually be linked.
1129 void coroutine_fn qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1131 BDRVQcow2State *s = bs->opaque;
1132 if (!has_data_file(bs) && !m->keep_old_clusters) {
1133 qcow2_free_clusters(bs, m->alloc_offset,
1134 m->nb_clusters << s->cluster_bits,
1135 QCOW2_DISCARD_NEVER);
1140 * For a given write request, create a new QCowL2Meta structure, add
1141 * it to @m and the BDRVQcow2State.cluster_allocs list. If the write
1142 * request does not need copy-on-write or changes to the L2 metadata
1143 * then this function does nothing.
1145 * @host_cluster_offset points to the beginning of the first cluster.
1147 * @guest_offset and @bytes indicate the offset and length of the
1150 * @l2_slice contains the L2 entries of all clusters involved in this
1153 * If @keep_old is true it means that the clusters were already
1154 * allocated and will be overwritten. If false then the clusters are
1155 * new and we have to decrease the reference count of the old ones.
1157 * Returns 0 on success, -errno on failure.
1159 static int coroutine_fn GRAPH_RDLOCK
1160 calculate_l2_meta(BlockDriverState *bs, uint64_t host_cluster_offset,
1161 uint64_t guest_offset, unsigned bytes, uint64_t *l2_slice,
1162 QCowL2Meta **m, bool keep_old)
1164 BDRVQcow2State *s = bs->opaque;
1165 int sc_index, l2_index = offset_to_l2_slice_index(s, guest_offset);
1166 uint64_t l2_entry, l2_bitmap;
1167 unsigned cow_start_from, cow_end_to;
1168 unsigned cow_start_to = offset_into_cluster(s, guest_offset);
1169 unsigned cow_end_from = cow_start_to + bytes;
1170 unsigned nb_clusters = size_to_clusters(s, cow_end_from);
1171 QCowL2Meta *old_m = *m;
1172 QCow2SubclusterType type;
1174 bool skip_cow = keep_old;
1176 assert(nb_clusters <= s->l2_slice_size - l2_index);
1178 /* Check the type of all affected subclusters */
1179 for (i = 0; i < nb_clusters; i++) {
1180 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1181 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1183 unsigned write_from = MAX(cow_start_to, i << s->cluster_bits);
1184 unsigned write_to = MIN(cow_end_from, (i + 1) << s->cluster_bits);
1185 int first_sc = offset_to_sc_index(s, write_from);
1186 int last_sc = offset_to_sc_index(s, write_to - 1);
1187 int cnt = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
1189 /* Is any of the subclusters of type != QCOW2_SUBCLUSTER_NORMAL ? */
1190 if (type != QCOW2_SUBCLUSTER_NORMAL || first_sc + cnt <= last_sc) {
1194 /* If we can't skip the cow we can still look for invalid entries */
1195 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, 0);
1197 if (type == QCOW2_SUBCLUSTER_INVALID) {
1198 int l1_index = offset_to_l1_index(s, guest_offset);
1199 uint64_t l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
1200 qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster "
1201 "entry found (L2 offset: %#" PRIx64
1203 l2_offset, l2_index + i);
1212 /* Get the L2 entry of the first cluster */
1213 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1214 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1215 sc_index = offset_to_sc_index(s, guest_offset);
1216 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1220 case QCOW2_SUBCLUSTER_COMPRESSED:
1223 case QCOW2_SUBCLUSTER_NORMAL:
1224 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1225 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1226 if (has_subclusters(s)) {
1227 /* Skip all leading zero and unallocated subclusters */
1228 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1230 MIN(sc_index, ctz32(alloc_bitmap)) << s->subcluster_bits;
1235 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1236 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1237 cow_start_from = sc_index << s->subcluster_bits;
1240 g_assert_not_reached();
1244 case QCOW2_SUBCLUSTER_NORMAL:
1245 cow_start_from = cow_start_to;
1247 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1248 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1249 cow_start_from = sc_index << s->subcluster_bits;
1252 g_assert_not_reached();
1256 /* Get the L2 entry of the last cluster */
1257 l2_index += nb_clusters - 1;
1258 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1259 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1260 sc_index = offset_to_sc_index(s, guest_offset + bytes - 1);
1261 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1265 case QCOW2_SUBCLUSTER_COMPRESSED:
1266 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1268 case QCOW2_SUBCLUSTER_NORMAL:
1269 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1270 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1271 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1272 if (has_subclusters(s)) {
1273 /* Skip all trailing zero and unallocated subclusters */
1274 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1276 MIN(s->subclusters_per_cluster - sc_index - 1,
1277 clz32(alloc_bitmap)) << s->subcluster_bits;
1280 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1281 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1282 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1285 g_assert_not_reached();
1289 case QCOW2_SUBCLUSTER_NORMAL:
1290 cow_end_to = cow_end_from;
1292 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1293 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1294 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1297 g_assert_not_reached();
1301 *m = g_malloc0(sizeof(**m));
1302 **m = (QCowL2Meta) {
1305 .alloc_offset = host_cluster_offset,
1306 .offset = start_of_cluster(s, guest_offset),
1307 .nb_clusters = nb_clusters,
1309 .keep_old_clusters = keep_old,
1312 .offset = cow_start_from,
1313 .nb_bytes = cow_start_to - cow_start_from,
1316 .offset = cow_end_from,
1317 .nb_bytes = cow_end_to - cow_end_from,
1321 qemu_co_queue_init(&(*m)->dependent_requests);
1322 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1328 * Returns true if writing to the cluster pointed to by @l2_entry
1329 * requires a new allocation (that is, if the cluster is unallocated
1330 * or has refcount > 1 and therefore cannot be written in-place).
1332 static bool cluster_needs_new_alloc(BlockDriverState *bs, uint64_t l2_entry)
1334 switch (qcow2_get_cluster_type(bs, l2_entry)) {
1335 case QCOW2_CLUSTER_NORMAL:
1336 case QCOW2_CLUSTER_ZERO_ALLOC:
1337 if (l2_entry & QCOW_OFLAG_COPIED) {
1341 case QCOW2_CLUSTER_UNALLOCATED:
1342 case QCOW2_CLUSTER_COMPRESSED:
1343 case QCOW2_CLUSTER_ZERO_PLAIN:
1351 * Returns the number of contiguous clusters that can be written to
1352 * using one single write request, starting from @l2_index.
1353 * At most @nb_clusters are checked.
1355 * If @new_alloc is true this counts clusters that are either
1356 * unallocated, or allocated but with refcount > 1 (so they need to be
1357 * newly allocated and COWed).
1359 * If @new_alloc is false this counts clusters that are already
1360 * allocated and can be overwritten in-place (this includes clusters
1361 * of type QCOW2_CLUSTER_ZERO_ALLOC).
1363 static int count_single_write_clusters(BlockDriverState *bs, int nb_clusters,
1364 uint64_t *l2_slice, int l2_index,
1367 BDRVQcow2State *s = bs->opaque;
1368 uint64_t l2_entry = get_l2_entry(s, l2_slice, l2_index);
1369 uint64_t expected_offset = l2_entry & L2E_OFFSET_MASK;
1372 for (i = 0; i < nb_clusters; i++) {
1373 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1374 if (cluster_needs_new_alloc(bs, l2_entry) != new_alloc) {
1378 if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
1381 expected_offset += s->cluster_size;
1385 assert(i <= nb_clusters);
1390 * Check if there already is an AIO write request in flight which allocates
1391 * the same cluster. In this case we need to wait until the previous
1392 * request has completed and updated the L2 table accordingly.
1395 * 0 if there was no dependency. *cur_bytes indicates the number of
1396 * bytes from guest_offset that can be read before the next
1397 * dependency must be processed (or the request is complete)
1399 * -EAGAIN if we had to wait for another request, previously gathered
1400 * information on cluster allocation may be invalid now. The caller
1401 * must start over anyway, so consider *cur_bytes undefined.
1403 static int coroutine_fn handle_dependencies(BlockDriverState *bs,
1404 uint64_t guest_offset,
1405 uint64_t *cur_bytes, QCowL2Meta **m)
1407 BDRVQcow2State *s = bs->opaque;
1408 QCowL2Meta *old_alloc;
1409 uint64_t bytes = *cur_bytes;
1411 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
1413 uint64_t start = guest_offset;
1414 uint64_t end = start + bytes;
1415 uint64_t old_start = start_of_cluster(s, l2meta_cow_start(old_alloc));
1416 uint64_t old_end = ROUND_UP(l2meta_cow_end(old_alloc), s->cluster_size);
1418 if (end <= old_start || start >= old_end) {
1419 /* No intersection */
1423 if (old_alloc->keep_old_clusters &&
1424 (end <= l2meta_cow_start(old_alloc) ||
1425 start >= l2meta_cow_end(old_alloc)))
1428 * Clusters intersect but COW areas don't. And cluster itself is
1429 * already allocated. So, there is no actual conflict.
1436 if (start < old_start) {
1437 /* Stop at the start of a running allocation */
1438 bytes = old_start - start;
1444 * Stop if an l2meta already exists. After yielding, it wouldn't
1445 * be valid any more, so we'd have to clean up the old L2Metas
1446 * and deal with requests depending on them before starting to
1447 * gather new ones. Not worth the trouble.
1449 if (bytes == 0 && *m) {
1456 * Wait for the dependency to complete. We need to recheck
1457 * the free/allocated clusters when we continue.
1459 qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
1464 /* Make sure that existing clusters and new allocations are only used up to
1465 * the next dependency if we shortened the request above */
1472 * Checks how many already allocated clusters that don't require a new
1473 * allocation there are at the given guest_offset (up to *bytes).
1474 * If *host_offset is not INV_OFFSET, only physically contiguous clusters
1475 * beginning at this host offset are counted.
1477 * Note that guest_offset may not be cluster aligned. In this case, the
1478 * returned *host_offset points to exact byte referenced by guest_offset and
1479 * therefore isn't cluster aligned as well.
1482 * 0: if no allocated clusters are available at the given offset.
1483 * *bytes is normally unchanged. It is set to 0 if the cluster
1484 * is allocated and can be overwritten in-place but doesn't have
1485 * the right physical offset.
1487 * 1: if allocated clusters that can be overwritten in place are
1488 * available at the requested offset. *bytes may have decreased
1489 * and describes the length of the area that can be written to.
1491 * -errno: in error cases
1493 static int coroutine_fn GRAPH_RDLOCK
1494 handle_copied(BlockDriverState *bs, uint64_t guest_offset,
1495 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1497 BDRVQcow2State *s = bs->opaque;
1499 uint64_t l2_entry, cluster_offset;
1501 uint64_t nb_clusters;
1502 unsigned int keep_clusters;
1505 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
1508 assert(*host_offset == INV_OFFSET || offset_into_cluster(s, guest_offset)
1509 == offset_into_cluster(s, *host_offset));
1512 * Calculate the number of clusters to look for. We stop at L2 slice
1513 * boundaries to keep things simple.
1516 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1518 l2_index = offset_to_l2_slice_index(s, guest_offset);
1519 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1520 /* Limit total byte count to BDRV_REQUEST_MAX_BYTES */
1521 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1523 /* Find L2 entry for the first involved cluster */
1524 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1529 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1530 cluster_offset = l2_entry & L2E_OFFSET_MASK;
1532 if (!cluster_needs_new_alloc(bs, l2_entry)) {
1533 if (offset_into_cluster(s, cluster_offset)) {
1534 qcow2_signal_corruption(bs, true, -1, -1, "%s cluster offset "
1535 "%#" PRIx64 " unaligned (guest offset: %#"
1536 PRIx64 ")", l2_entry & QCOW_OFLAG_ZERO ?
1537 "Preallocated zero" : "Data",
1538 cluster_offset, guest_offset);
1543 /* If a specific host_offset is required, check it */
1544 if (*host_offset != INV_OFFSET && cluster_offset != *host_offset) {
1550 /* We keep all QCOW_OFLAG_COPIED clusters */
1551 keep_clusters = count_single_write_clusters(bs, nb_clusters, l2_slice,
1553 assert(keep_clusters <= nb_clusters);
1555 *bytes = MIN(*bytes,
1556 keep_clusters * s->cluster_size
1557 - offset_into_cluster(s, guest_offset));
1558 assert(*bytes != 0);
1560 ret = calculate_l2_meta(bs, cluster_offset, guest_offset,
1561 *bytes, l2_slice, m, true);
1573 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1575 /* Only return a host offset if we actually made progress. Otherwise we
1576 * would make requirements for handle_alloc() that it can't fulfill */
1578 *host_offset = cluster_offset + offset_into_cluster(s, guest_offset);
1585 * Allocates new clusters for the given guest_offset.
1587 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1588 * contain the number of clusters that have been allocated and are contiguous
1589 * in the image file.
1591 * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1592 * at which the new clusters must start. *nb_clusters can be 0 on return in
1593 * this case if the cluster at host_offset is already in use. If *host_offset
1594 * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1596 * *host_offset is updated to contain the offset into the image file at which
1597 * the first allocated cluster starts.
1599 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1600 * function has been waiting for another request and the allocation must be
1601 * restarted, but the whole request should not be failed.
1603 static int coroutine_fn GRAPH_RDLOCK
1604 do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1605 uint64_t *host_offset, uint64_t *nb_clusters)
1607 BDRVQcow2State *s = bs->opaque;
1609 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1610 *host_offset, *nb_clusters);
1612 if (has_data_file(bs)) {
1613 assert(*host_offset == INV_OFFSET ||
1614 *host_offset == start_of_cluster(s, guest_offset));
1615 *host_offset = start_of_cluster(s, guest_offset);
1619 /* Allocate new clusters */
1620 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1621 if (*host_offset == INV_OFFSET) {
1622 int64_t cluster_offset =
1623 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1624 if (cluster_offset < 0) {
1625 return cluster_offset;
1627 *host_offset = cluster_offset;
1630 int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1640 * Allocates new clusters for an area that is either still unallocated or
1641 * cannot be overwritten in-place. If *host_offset is not INV_OFFSET,
1642 * clusters are only allocated if the new allocation can match the specified
1645 * Note that guest_offset may not be cluster aligned. In this case, the
1646 * returned *host_offset points to exact byte referenced by guest_offset and
1647 * therefore isn't cluster aligned as well.
1650 * 0: if no clusters could be allocated. *bytes is set to 0,
1651 * *host_offset is left unchanged.
1653 * 1: if new clusters were allocated. *bytes may be decreased if the
1654 * new allocation doesn't cover all of the requested area.
1655 * *host_offset is updated to contain the host offset of the first
1656 * newly allocated cluster.
1658 * -errno: in error cases
1660 static int coroutine_fn GRAPH_RDLOCK
1661 handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1662 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1664 BDRVQcow2State *s = bs->opaque;
1667 uint64_t nb_clusters;
1670 uint64_t alloc_cluster_offset;
1672 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1677 * Calculate the number of clusters to look for. We stop at L2 slice
1678 * boundaries to keep things simple.
1681 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1683 l2_index = offset_to_l2_slice_index(s, guest_offset);
1684 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1685 /* Limit total allocation byte count to BDRV_REQUEST_MAX_BYTES */
1686 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1688 /* Find L2 entry for the first involved cluster */
1689 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1694 nb_clusters = count_single_write_clusters(bs, nb_clusters,
1695 l2_slice, l2_index, true);
1697 /* This function is only called when there were no non-COW clusters, so if
1698 * we can't find any unallocated or COW clusters either, something is
1699 * wrong with our code. */
1700 assert(nb_clusters > 0);
1702 /* Allocate at a given offset in the image file */
1703 alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1704 start_of_cluster(s, *host_offset);
1705 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1711 /* Can't extend contiguous allocation */
1712 if (nb_clusters == 0) {
1718 assert(alloc_cluster_offset != INV_OFFSET);
1721 * Save info needed for meta data update.
1723 * requested_bytes: Number of bytes from the start of the first
1724 * newly allocated cluster to the end of the (possibly shortened
1725 * before) write request.
1727 * avail_bytes: Number of bytes from the start of the first
1728 * newly allocated to the end of the last newly allocated cluster.
1730 * nb_bytes: The number of bytes from the start of the first
1731 * newly allocated cluster to the end of the area that the write
1732 * request actually writes to (excluding COW at the end)
1734 uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1735 int avail_bytes = nb_clusters << s->cluster_bits;
1736 int nb_bytes = MIN(requested_bytes, avail_bytes);
1738 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1739 *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1740 assert(*bytes != 0);
1742 ret = calculate_l2_meta(bs, alloc_cluster_offset, guest_offset, *bytes,
1743 l2_slice, m, false);
1751 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1756 * For a given area on the virtual disk defined by @offset and @bytes,
1757 * find the corresponding area on the qcow2 image, allocating new
1758 * clusters (or subclusters) if necessary. The result can span a
1759 * combination of allocated and previously unallocated clusters.
1761 * Note that offset may not be cluster aligned. In this case, the returned
1762 * *host_offset points to exact byte referenced by offset and therefore
1763 * isn't cluster aligned as well.
1765 * On return, @host_offset is set to the beginning of the requested
1766 * area. This area is guaranteed to be contiguous on the qcow2 file
1767 * but it can be smaller than initially requested. In this case @bytes
1768 * is updated with the actual size.
1770 * If any clusters or subclusters were allocated then @m contains a
1771 * list with the information of all the affected regions. Note that
1772 * this can happen regardless of whether this function succeeds or
1773 * not. The caller is responsible for updating the L2 metadata of the
1774 * allocated clusters (on success) or freeing them (on failure), and
1775 * for clearing the contents of @m afterwards in both cases.
1777 * If the request conflicts with another write request in flight, the coroutine
1778 * is queued and will be reentered when the dependency has completed.
1780 * Return 0 on success and -errno in error cases
1782 int coroutine_fn qcow2_alloc_host_offset(BlockDriverState *bs, uint64_t offset,
1783 unsigned int *bytes,
1784 uint64_t *host_offset,
1787 BDRVQcow2State *s = bs->opaque;
1788 uint64_t start, remaining;
1789 uint64_t cluster_offset;
1793 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1798 cluster_offset = INV_OFFSET;
1799 *host_offset = INV_OFFSET;
1805 if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1806 *host_offset = cluster_offset;
1809 assert(remaining >= cur_bytes);
1812 remaining -= cur_bytes;
1814 if (cluster_offset != INV_OFFSET) {
1815 cluster_offset += cur_bytes;
1818 if (remaining == 0) {
1822 cur_bytes = remaining;
1825 * Now start gathering as many contiguous clusters as possible:
1827 * 1. Check for overlaps with in-flight allocations
1829 * a) Overlap not in the first cluster -> shorten this request and
1830 * let the caller handle the rest in its next loop iteration.
1832 * b) Real overlaps of two requests. Yield and restart the search
1833 * for contiguous clusters (the situation could have changed
1834 * while we were sleeping)
1836 * c) TODO: Request starts in the same cluster as the in-flight
1837 * allocation ends. Shorten the COW of the in-fight allocation,
1838 * set cluster_offset to write to the same cluster and set up
1839 * the right synchronisation between the in-flight request and
1842 ret = handle_dependencies(bs, start, &cur_bytes, m);
1843 if (ret == -EAGAIN) {
1844 /* Currently handle_dependencies() doesn't yield if we already had
1845 * an allocation. If it did, we would have to clean up the L2Meta
1846 * structs before starting over. */
1849 } else if (ret < 0) {
1851 } else if (cur_bytes == 0) {
1854 /* handle_dependencies() may have decreased cur_bytes (shortened
1855 * the allocations below) so that the next dependency is processed
1856 * correctly during the next loop iteration. */
1860 * 2. Count contiguous COPIED clusters.
1862 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1867 } else if (cur_bytes == 0) {
1872 * 3. If the request still hasn't completed, allocate new clusters,
1873 * considering any cluster_offset of steps 1c or 2.
1875 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1881 assert(cur_bytes == 0);
1886 *bytes -= remaining;
1888 assert(*host_offset != INV_OFFSET);
1889 assert(offset_into_cluster(s, *host_offset) ==
1890 offset_into_cluster(s, offset));
1896 * This discards as many clusters of nb_clusters as possible at once (i.e.
1897 * all clusters in the same L2 slice) and returns the number of discarded
1900 static int GRAPH_RDLOCK
1901 discard_in_l2_slice(BlockDriverState *bs, uint64_t offset, uint64_t nb_clusters,
1902 enum qcow2_discard_type type, bool full_discard)
1904 BDRVQcow2State *s = bs->opaque;
1910 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1915 /* Limit nb_clusters to one L2 slice */
1916 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1917 assert(nb_clusters <= INT_MAX);
1919 for (i = 0; i < nb_clusters; i++) {
1920 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1921 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1922 uint64_t new_l2_entry = old_l2_entry;
1923 uint64_t new_l2_bitmap = old_l2_bitmap;
1924 QCow2ClusterType cluster_type =
1925 qcow2_get_cluster_type(bs, old_l2_entry);
1926 bool keep_reference = (cluster_type != QCOW2_CLUSTER_COMPRESSED) &&
1928 (s->discard_no_unref &&
1929 type == QCOW2_DISCARD_REQUEST);
1932 * If full_discard is true, the cluster should not read back as zeroes,
1933 * but rather fall through to the backing file.
1935 * If full_discard is false, make sure that a discarded area reads back
1936 * as zeroes for v3 images (we cannot do it for v2 without actually
1937 * writing a zero-filled buffer). We can skip the operation if the
1938 * cluster is already marked as zero, or if it's unallocated and we
1939 * don't have a backing file.
1941 * TODO We might want to use bdrv_block_status(bs) here, but we're
1942 * holding s->lock, so that doesn't work today.
1945 new_l2_entry = new_l2_bitmap = 0;
1946 } else if (bs->backing || qcow2_cluster_is_allocated(cluster_type)) {
1947 if (has_subclusters(s)) {
1948 if (keep_reference) {
1949 new_l2_entry = old_l2_entry;
1953 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
1955 if (s->qcow_version >= 3) {
1956 if (keep_reference) {
1957 new_l2_entry |= QCOW_OFLAG_ZERO;
1959 new_l2_entry = QCOW_OFLAG_ZERO;
1967 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
1971 /* First remove L2 entries */
1972 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1973 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
1974 if (has_subclusters(s)) {
1975 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
1977 if (!keep_reference) {
1978 /* Then decrease the refcount */
1979 qcow2_free_any_cluster(bs, old_l2_entry, type);
1980 } else if (s->discard_passthrough[type] &&
1981 (cluster_type == QCOW2_CLUSTER_NORMAL ||
1982 cluster_type == QCOW2_CLUSTER_ZERO_ALLOC)) {
1983 /* If we keep the reference, pass on the discard still */
1984 bdrv_pdiscard(s->data_file, old_l2_entry & L2E_OFFSET_MASK,
1989 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1994 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1995 uint64_t bytes, enum qcow2_discard_type type,
1998 BDRVQcow2State *s = bs->opaque;
1999 uint64_t end_offset = offset + bytes;
2000 uint64_t nb_clusters;
2004 /* Caller must pass aligned values, except at image end */
2005 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
2006 assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
2007 end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
2009 nb_clusters = size_to_clusters(s, bytes);
2011 s->cache_discards = true;
2013 /* Each L2 slice is handled by its own loop iteration */
2014 while (nb_clusters > 0) {
2015 cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
2022 nb_clusters -= cleared;
2023 offset += (cleared * s->cluster_size);
2028 s->cache_discards = false;
2029 qcow2_process_discards(bs, ret);
2035 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
2036 * all clusters in the same L2 slice) and returns the number of zeroed
2039 static int coroutine_fn GRAPH_RDLOCK
2040 zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
2041 uint64_t nb_clusters, int flags)
2043 BDRVQcow2State *s = bs->opaque;
2049 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2054 /* Limit nb_clusters to one L2 slice */
2055 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
2056 assert(nb_clusters <= INT_MAX);
2058 for (i = 0; i < nb_clusters; i++) {
2059 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
2060 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
2061 QCow2ClusterType type = qcow2_get_cluster_type(bs, old_l2_entry);
2062 bool unmap = (type == QCOW2_CLUSTER_COMPRESSED) ||
2063 ((flags & BDRV_REQ_MAY_UNMAP) && qcow2_cluster_is_allocated(type));
2064 bool keep_reference =
2065 (s->discard_no_unref && type != QCOW2_CLUSTER_COMPRESSED);
2066 uint64_t new_l2_entry = old_l2_entry;
2067 uint64_t new_l2_bitmap = old_l2_bitmap;
2069 if (unmap && !keep_reference) {
2073 if (has_subclusters(s)) {
2074 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
2076 new_l2_entry |= QCOW_OFLAG_ZERO;
2079 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
2083 /* First update L2 entries */
2084 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2085 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
2086 if (has_subclusters(s)) {
2087 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
2091 if (!keep_reference) {
2092 /* Then decrease the refcount */
2093 qcow2_free_any_cluster(bs, old_l2_entry, QCOW2_DISCARD_REQUEST);
2094 } else if (s->discard_passthrough[QCOW2_DISCARD_REQUEST] &&
2095 (type == QCOW2_CLUSTER_NORMAL ||
2096 type == QCOW2_CLUSTER_ZERO_ALLOC)) {
2097 /* If we keep the reference, pass on the discard still */
2098 bdrv_pdiscard(s->data_file, old_l2_entry & L2E_OFFSET_MASK,
2104 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2109 static int coroutine_fn GRAPH_RDLOCK
2110 zero_l2_subclusters(BlockDriverState *bs, uint64_t offset,
2111 unsigned nb_subclusters)
2113 BDRVQcow2State *s = bs->opaque;
2115 uint64_t old_l2_bitmap, l2_bitmap;
2116 int l2_index, ret, sc = offset_to_sc_index(s, offset);
2118 /* For full clusters use zero_in_l2_slice() instead */
2119 assert(nb_subclusters > 0 && nb_subclusters < s->subclusters_per_cluster);
2120 assert(sc + nb_subclusters <= s->subclusters_per_cluster);
2121 assert(offset_into_subcluster(s, offset) == 0);
2123 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2128 switch (qcow2_get_cluster_type(bs, get_l2_entry(s, l2_slice, l2_index))) {
2129 case QCOW2_CLUSTER_COMPRESSED:
2130 ret = -ENOTSUP; /* We cannot partially zeroize compressed clusters */
2132 case QCOW2_CLUSTER_NORMAL:
2133 case QCOW2_CLUSTER_UNALLOCATED:
2136 g_assert_not_reached();
2139 old_l2_bitmap = l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
2141 l2_bitmap |= QCOW_OFLAG_SUB_ZERO_RANGE(sc, sc + nb_subclusters);
2142 l2_bitmap &= ~QCOW_OFLAG_SUB_ALLOC_RANGE(sc, sc + nb_subclusters);
2144 if (old_l2_bitmap != l2_bitmap) {
2145 set_l2_bitmap(s, l2_slice, l2_index, l2_bitmap);
2146 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2151 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2156 int coroutine_fn qcow2_subcluster_zeroize(BlockDriverState *bs, uint64_t offset,
2157 uint64_t bytes, int flags)
2159 BDRVQcow2State *s = bs->opaque;
2160 uint64_t end_offset = offset + bytes;
2161 uint64_t nb_clusters;
2162 unsigned head, tail;
2166 /* If we have to stay in sync with an external data file, zero out
2167 * s->data_file first. */
2168 if (data_file_is_raw(bs)) {
2169 assert(has_data_file(bs));
2170 ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
2176 /* Caller must pass aligned values, except at image end */
2177 assert(offset_into_subcluster(s, offset) == 0);
2178 assert(offset_into_subcluster(s, end_offset) == 0 ||
2179 end_offset >= bs->total_sectors << BDRV_SECTOR_BITS);
2182 * The zero flag is only supported by version 3 and newer. However, if we
2183 * have no backing file, we can resort to discard in version 2.
2185 if (s->qcow_version < 3) {
2187 return qcow2_cluster_discard(bs, offset, bytes,
2188 QCOW2_DISCARD_REQUEST, false);
2193 head = MIN(end_offset, ROUND_UP(offset, s->cluster_size)) - offset;
2196 tail = (end_offset >= bs->total_sectors << BDRV_SECTOR_BITS) ? 0 :
2197 end_offset - MAX(offset, start_of_cluster(s, end_offset));
2200 s->cache_discards = true;
2203 ret = zero_l2_subclusters(bs, offset - head,
2204 size_to_subclusters(s, head));
2210 /* Each L2 slice is handled by its own loop iteration */
2211 nb_clusters = size_to_clusters(s, end_offset - offset);
2213 while (nb_clusters > 0) {
2214 cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
2220 nb_clusters -= cleared;
2221 offset += (cleared * s->cluster_size);
2225 ret = zero_l2_subclusters(bs, end_offset, size_to_subclusters(s, tail));
2233 s->cache_discards = false;
2234 qcow2_process_discards(bs, ret);
2240 * Expands all zero clusters in a specific L1 table (or deallocates them, for
2241 * non-backed non-pre-allocated zero clusters).
2243 * l1_entries and *visited_l1_entries are used to keep track of progress for
2244 * status_cb(). l1_entries contains the total number of L1 entries and
2245 * *visited_l1_entries counts all visited L1 entries.
2247 static int GRAPH_RDLOCK
2248 expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
2249 int l1_size, int64_t *visited_l1_entries,
2251 BlockDriverAmendStatusCB *status_cb,
2254 BDRVQcow2State *s = bs->opaque;
2255 bool is_active_l1 = (l1_table == s->l1_table);
2256 uint64_t *l2_slice = NULL;
2257 unsigned slice, slice_size2, n_slices;
2261 /* qcow2_downgrade() is not allowed in images with subclusters */
2262 assert(!has_subclusters(s));
2264 slice_size2 = s->l2_slice_size * l2_entry_size(s);
2265 n_slices = s->cluster_size / slice_size2;
2267 if (!is_active_l1) {
2268 /* inactive L2 tables require a buffer to be stored in when loading
2270 l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
2271 if (l2_slice == NULL) {
2276 for (i = 0; i < l1_size; i++) {
2277 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
2278 uint64_t l2_refcount;
2282 (*visited_l1_entries)++;
2284 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2289 if (offset_into_cluster(s, l2_offset)) {
2290 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
2291 PRIx64 " unaligned (L1 index: %#x)",
2297 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
2303 for (slice = 0; slice < n_slices; slice++) {
2304 uint64_t slice_offset = l2_offset + slice * slice_size2;
2305 bool l2_dirty = false;
2307 /* get active L2 tables from cache */
2308 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
2309 (void **)&l2_slice);
2311 /* load inactive L2 tables from disk */
2312 ret = bdrv_pread(bs->file, slice_offset, slice_size2,
2319 for (j = 0; j < s->l2_slice_size; j++) {
2320 uint64_t l2_entry = get_l2_entry(s, l2_slice, j);
2321 int64_t offset = l2_entry & L2E_OFFSET_MASK;
2322 QCow2ClusterType cluster_type =
2323 qcow2_get_cluster_type(bs, l2_entry);
2325 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
2326 cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
2330 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2333 * not backed; therefore we can simply deallocate the
2334 * cluster. No need to call set_l2_bitmap(), this
2335 * function doesn't support images with subclusters.
2337 set_l2_entry(s, l2_slice, j, 0);
2342 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2348 /* The offset must fit in the offset field */
2349 assert((offset & L2E_OFFSET_MASK) == offset);
2351 if (l2_refcount > 1) {
2352 /* For shared L2 tables, set the refcount accordingly
2353 * (it is already 1 and needs to be l2_refcount) */
2354 ret = qcow2_update_cluster_refcount(
2355 bs, offset >> s->cluster_bits,
2356 refcount_diff(1, l2_refcount), false,
2357 QCOW2_DISCARD_OTHER);
2359 qcow2_free_clusters(bs, offset, s->cluster_size,
2360 QCOW2_DISCARD_OTHER);
2366 if (offset_into_cluster(s, offset)) {
2367 int l2_index = slice * s->l2_slice_size + j;
2368 qcow2_signal_corruption(
2370 "Cluster allocation offset "
2371 "%#" PRIx64 " unaligned (L2 offset: %#"
2372 PRIx64 ", L2 index: %#x)", offset,
2373 l2_offset, l2_index);
2374 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2375 qcow2_free_clusters(bs, offset, s->cluster_size,
2376 QCOW2_DISCARD_ALWAYS);
2382 ret = qcow2_pre_write_overlap_check(bs, 0, offset,
2383 s->cluster_size, true);
2385 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2386 qcow2_free_clusters(bs, offset, s->cluster_size,
2387 QCOW2_DISCARD_ALWAYS);
2392 ret = bdrv_pwrite_zeroes(s->data_file, offset,
2393 s->cluster_size, 0);
2395 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2396 qcow2_free_clusters(bs, offset, s->cluster_size,
2397 QCOW2_DISCARD_ALWAYS);
2402 if (l2_refcount == 1) {
2403 set_l2_entry(s, l2_slice, j, offset | QCOW_OFLAG_COPIED);
2405 set_l2_entry(s, l2_slice, j, offset);
2408 * No need to call set_l2_bitmap() after set_l2_entry() because
2409 * this function doesn't support images with subclusters.
2416 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2417 qcow2_cache_depends_on_flush(s->l2_table_cache);
2419 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2422 ret = qcow2_pre_write_overlap_check(
2423 bs, QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2,
2424 slice_offset, slice_size2, false);
2429 ret = bdrv_pwrite(bs->file, slice_offset, slice_size2,
2438 (*visited_l1_entries)++;
2440 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2448 if (!is_active_l1) {
2449 qemu_vfree(l2_slice);
2451 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2458 * For backed images, expands all zero clusters on the image. For non-backed
2459 * images, deallocates all non-pre-allocated zero clusters (and claims the
2460 * allocation for pre-allocated ones). This is important for downgrading to a
2461 * qcow2 version which doesn't yet support metadata zero clusters.
2463 int qcow2_expand_zero_clusters(BlockDriverState *bs,
2464 BlockDriverAmendStatusCB *status_cb,
2467 BDRVQcow2State *s = bs->opaque;
2468 uint64_t *l1_table = NULL;
2469 int64_t l1_entries = 0, visited_l1_entries = 0;
2474 l1_entries = s->l1_size;
2475 for (i = 0; i < s->nb_snapshots; i++) {
2476 l1_entries += s->snapshots[i].l1_size;
2480 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
2481 &visited_l1_entries, l1_entries,
2482 status_cb, cb_opaque);
2487 /* Inactive L1 tables may point to active L2 tables - therefore it is
2488 * necessary to flush the L2 table cache before trying to access the L2
2489 * tables pointed to by inactive L1 entries (else we might try to expand
2490 * zero clusters that have already been expanded); furthermore, it is also
2491 * necessary to empty the L2 table cache, since it may contain tables which
2492 * are now going to be modified directly on disk, bypassing the cache.
2493 * qcow2_cache_empty() does both for us. */
2494 ret = qcow2_cache_empty(bs, s->l2_table_cache);
2499 for (i = 0; i < s->nb_snapshots; i++) {
2501 uint64_t *new_l1_table;
2502 Error *local_err = NULL;
2504 ret = qcow2_validate_table(bs, s->snapshots[i].l1_table_offset,
2505 s->snapshots[i].l1_size, L1E_SIZE,
2506 QCOW_MAX_L1_SIZE, "Snapshot L1 table",
2509 error_report_err(local_err);
2513 l1_size2 = s->snapshots[i].l1_size * L1E_SIZE;
2514 new_l1_table = g_try_realloc(l1_table, l1_size2);
2516 if (!new_l1_table) {
2521 l1_table = new_l1_table;
2523 ret = bdrv_pread(bs->file, s->snapshots[i].l1_table_offset, l1_size2,
2529 for (j = 0; j < s->snapshots[i].l1_size; j++) {
2530 be64_to_cpus(&l1_table[j]);
2533 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
2534 &visited_l1_entries, l1_entries,
2535 status_cb, cb_opaque);
2548 void qcow2_parse_compressed_l2_entry(BlockDriverState *bs, uint64_t l2_entry,
2549 uint64_t *coffset, int *csize)
2551 BDRVQcow2State *s = bs->opaque;
2554 assert(qcow2_get_cluster_type(bs, l2_entry) == QCOW2_CLUSTER_COMPRESSED);
2556 *coffset = l2_entry & s->cluster_offset_mask;
2558 nb_csectors = ((l2_entry >> s->csize_shift) & s->csize_mask) + 1;
2559 *csize = nb_csectors * QCOW2_COMPRESSED_SECTOR_SIZE -
2560 (*coffset & (QCOW2_COMPRESSED_SECTOR_SIZE - 1));