throttle: add throttle_max_is_missing_limit() test
[qemu/ar7.git] / block / qcow2-cluster.c
blobb43f186eb898dc7e7482da3e4ebc3637477d8720
1 /*
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
22 * THE SOFTWARE.
25 #include <zlib.h>
27 #include "qemu-common.h"
28 #include "block/block_int.h"
29 #include "block/qcow2.h"
30 #include "trace.h"
32 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
33 bool exact_size)
35 BDRVQcowState *s = bs->opaque;
36 int new_l1_size2, ret, i;
37 uint64_t *new_l1_table;
38 int64_t old_l1_table_offset, old_l1_size;
39 int64_t new_l1_table_offset, new_l1_size;
40 uint8_t data[12];
42 if (min_size <= s->l1_size)
43 return 0;
45 /* Do a sanity check on min_size before trying to calculate new_l1_size
46 * (this prevents overflows during the while loop for the calculation of
47 * new_l1_size) */
48 if (min_size > INT_MAX / sizeof(uint64_t)) {
49 return -EFBIG;
52 if (exact_size) {
53 new_l1_size = min_size;
54 } else {
55 /* Bump size up to reduce the number of times we have to grow */
56 new_l1_size = s->l1_size;
57 if (new_l1_size == 0) {
58 new_l1_size = 1;
60 while (min_size > new_l1_size) {
61 new_l1_size = (new_l1_size * 3 + 1) / 2;
65 if (new_l1_size > INT_MAX / sizeof(uint64_t)) {
66 return -EFBIG;
69 #ifdef DEBUG_ALLOC2
70 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
71 s->l1_size, new_l1_size);
72 #endif
74 new_l1_size2 = sizeof(uint64_t) * new_l1_size;
75 new_l1_table = qemu_try_blockalign(bs->file,
76 align_offset(new_l1_size2, 512));
77 if (new_l1_table == NULL) {
78 return -ENOMEM;
80 memset(new_l1_table, 0, align_offset(new_l1_size2, 512));
82 memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
84 /* write new table (align to cluster) */
85 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
86 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
87 if (new_l1_table_offset < 0) {
88 qemu_vfree(new_l1_table);
89 return new_l1_table_offset;
92 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
93 if (ret < 0) {
94 goto fail;
97 /* the L1 position has not yet been updated, so these clusters must
98 * indeed be completely free */
99 ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
100 new_l1_size2);
101 if (ret < 0) {
102 goto fail;
105 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
106 for(i = 0; i < s->l1_size; i++)
107 new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
108 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_table, new_l1_size2);
109 if (ret < 0)
110 goto fail;
111 for(i = 0; i < s->l1_size; i++)
112 new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
114 /* set new table */
115 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
116 cpu_to_be32w((uint32_t*)data, new_l1_size);
117 stq_be_p(data + 4, new_l1_table_offset);
118 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), data,sizeof(data));
119 if (ret < 0) {
120 goto fail;
122 qemu_vfree(s->l1_table);
123 old_l1_table_offset = s->l1_table_offset;
124 s->l1_table_offset = new_l1_table_offset;
125 s->l1_table = new_l1_table;
126 old_l1_size = s->l1_size;
127 s->l1_size = new_l1_size;
128 qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
129 QCOW2_DISCARD_OTHER);
130 return 0;
131 fail:
132 qemu_vfree(new_l1_table);
133 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
134 QCOW2_DISCARD_OTHER);
135 return ret;
139 * l2_load
141 * Loads a L2 table into memory. If the table is in the cache, the cache
142 * is used; otherwise the L2 table is loaded from the image file.
144 * Returns a pointer to the L2 table on success, or NULL if the read from
145 * the image file failed.
148 static int l2_load(BlockDriverState *bs, uint64_t l2_offset,
149 uint64_t **l2_table)
151 BDRVQcowState *s = bs->opaque;
152 int ret;
154 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, (void**) l2_table);
156 return ret;
160 * Writes one sector of the L1 table to the disk (can't update single entries
161 * and we really don't want bdrv_pread to perform a read-modify-write)
163 #define L1_ENTRIES_PER_SECTOR (512 / 8)
164 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
166 BDRVQcowState *s = bs->opaque;
167 uint64_t buf[L1_ENTRIES_PER_SECTOR] = { 0 };
168 int l1_start_index;
169 int i, ret;
171 l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
172 for (i = 0; i < L1_ENTRIES_PER_SECTOR && l1_start_index + i < s->l1_size;
173 i++)
175 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
178 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
179 s->l1_table_offset + 8 * l1_start_index, sizeof(buf));
180 if (ret < 0) {
181 return ret;
184 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
185 ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset + 8 * l1_start_index,
186 buf, sizeof(buf));
187 if (ret < 0) {
188 return ret;
191 return 0;
195 * l2_allocate
197 * Allocate a new l2 entry in the file. If l1_index points to an already
198 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
199 * table) copy the contents of the old L2 table into the newly allocated one.
200 * Otherwise the new table is initialized with zeros.
204 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table)
206 BDRVQcowState *s = bs->opaque;
207 uint64_t old_l2_offset;
208 uint64_t *l2_table = NULL;
209 int64_t l2_offset;
210 int ret;
212 old_l2_offset = s->l1_table[l1_index];
214 trace_qcow2_l2_allocate(bs, l1_index);
216 /* allocate a new l2 entry */
218 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
219 if (l2_offset < 0) {
220 ret = l2_offset;
221 goto fail;
224 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
225 if (ret < 0) {
226 goto fail;
229 /* allocate a new entry in the l2 cache */
231 trace_qcow2_l2_allocate_get_empty(bs, l1_index);
232 ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table);
233 if (ret < 0) {
234 goto fail;
237 l2_table = *table;
239 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
240 /* if there was no old l2 table, clear the new table */
241 memset(l2_table, 0, s->l2_size * sizeof(uint64_t));
242 } else {
243 uint64_t* old_table;
245 /* if there was an old l2 table, read it from the disk */
246 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
247 ret = qcow2_cache_get(bs, s->l2_table_cache,
248 old_l2_offset & L1E_OFFSET_MASK,
249 (void**) &old_table);
250 if (ret < 0) {
251 goto fail;
254 memcpy(l2_table, old_table, s->cluster_size);
256 qcow2_cache_put(bs, s->l2_table_cache, (void **) &old_table);
259 /* write the l2 table to the file */
260 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
262 trace_qcow2_l2_allocate_write_l2(bs, l1_index);
263 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
264 ret = qcow2_cache_flush(bs, s->l2_table_cache);
265 if (ret < 0) {
266 goto fail;
269 /* update the L1 entry */
270 trace_qcow2_l2_allocate_write_l1(bs, l1_index);
271 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
272 ret = qcow2_write_l1_entry(bs, l1_index);
273 if (ret < 0) {
274 goto fail;
277 *table = l2_table;
278 trace_qcow2_l2_allocate_done(bs, l1_index, 0);
279 return 0;
281 fail:
282 trace_qcow2_l2_allocate_done(bs, l1_index, ret);
283 if (l2_table != NULL) {
284 qcow2_cache_put(bs, s->l2_table_cache, (void**) table);
286 s->l1_table[l1_index] = old_l2_offset;
287 if (l2_offset > 0) {
288 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
289 QCOW2_DISCARD_ALWAYS);
291 return ret;
295 * Checks how many clusters in a given L2 table are contiguous in the image
296 * file. As soon as one of the flags in the bitmask stop_flags changes compared
297 * to the first cluster, the search is stopped and the cluster is not counted
298 * as contiguous. (This allows it, for example, to stop at the first compressed
299 * cluster which may require a different handling)
301 static int count_contiguous_clusters(uint64_t nb_clusters, int cluster_size,
302 uint64_t *l2_table, uint64_t stop_flags)
304 int i;
305 uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
306 uint64_t first_entry = be64_to_cpu(l2_table[0]);
307 uint64_t offset = first_entry & mask;
309 if (!offset)
310 return 0;
312 assert(qcow2_get_cluster_type(first_entry) != QCOW2_CLUSTER_COMPRESSED);
314 for (i = 0; i < nb_clusters; i++) {
315 uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask;
316 if (offset + (uint64_t) i * cluster_size != l2_entry) {
317 break;
321 return i;
324 static int count_contiguous_free_clusters(uint64_t nb_clusters, uint64_t *l2_table)
326 int i;
328 for (i = 0; i < nb_clusters; i++) {
329 int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i]));
331 if (type != QCOW2_CLUSTER_UNALLOCATED) {
332 break;
336 return i;
339 /* The crypt function is compatible with the linux cryptoloop
340 algorithm for < 4 GB images. NOTE: out_buf == in_buf is
341 supported */
342 int qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
343 uint8_t *out_buf, const uint8_t *in_buf,
344 int nb_sectors, bool enc,
345 Error **errp)
347 union {
348 uint64_t ll[2];
349 uint8_t b[16];
350 } ivec;
351 int i;
352 int ret;
354 for(i = 0; i < nb_sectors; i++) {
355 ivec.ll[0] = cpu_to_le64(sector_num);
356 ivec.ll[1] = 0;
357 if (qcrypto_cipher_setiv(s->cipher,
358 ivec.b, G_N_ELEMENTS(ivec.b),
359 errp) < 0) {
360 return -1;
362 if (enc) {
363 ret = qcrypto_cipher_encrypt(s->cipher,
364 in_buf,
365 out_buf,
366 512,
367 errp);
368 } else {
369 ret = qcrypto_cipher_decrypt(s->cipher,
370 in_buf,
371 out_buf,
372 512,
373 errp);
375 if (ret < 0) {
376 return -1;
378 sector_num++;
379 in_buf += 512;
380 out_buf += 512;
382 return 0;
385 static int coroutine_fn copy_sectors(BlockDriverState *bs,
386 uint64_t start_sect,
387 uint64_t cluster_offset,
388 int n_start, int n_end)
390 BDRVQcowState *s = bs->opaque;
391 QEMUIOVector qiov;
392 struct iovec iov;
393 int n, ret;
395 n = n_end - n_start;
396 if (n <= 0) {
397 return 0;
400 iov.iov_len = n * BDRV_SECTOR_SIZE;
401 iov.iov_base = qemu_try_blockalign(bs, iov.iov_len);
402 if (iov.iov_base == NULL) {
403 return -ENOMEM;
406 qemu_iovec_init_external(&qiov, &iov, 1);
408 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
410 if (!bs->drv) {
411 ret = -ENOMEDIUM;
412 goto out;
415 /* Call .bdrv_co_readv() directly instead of using the public block-layer
416 * interface. This avoids double I/O throttling and request tracking,
417 * which can lead to deadlock when block layer copy-on-read is enabled.
419 ret = bs->drv->bdrv_co_readv(bs, start_sect + n_start, n, &qiov);
420 if (ret < 0) {
421 goto out;
424 if (bs->encrypted) {
425 Error *err = NULL;
426 assert(s->cipher);
427 if (qcow2_encrypt_sectors(s, start_sect + n_start,
428 iov.iov_base, iov.iov_base, n,
429 true, &err) < 0) {
430 ret = -EIO;
431 error_free(err);
432 goto out;
436 ret = qcow2_pre_write_overlap_check(bs, 0,
437 cluster_offset + n_start * BDRV_SECTOR_SIZE, n * BDRV_SECTOR_SIZE);
438 if (ret < 0) {
439 goto out;
442 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
443 ret = bdrv_co_writev(bs->file, (cluster_offset >> 9) + n_start, n, &qiov);
444 if (ret < 0) {
445 goto out;
448 ret = 0;
449 out:
450 qemu_vfree(iov.iov_base);
451 return ret;
456 * get_cluster_offset
458 * For a given offset of the disk image, find the cluster offset in
459 * qcow2 file. The offset is stored in *cluster_offset.
461 * on entry, *num is the number of contiguous sectors we'd like to
462 * access following offset.
464 * on exit, *num is the number of contiguous sectors we can read.
466 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
467 * cases.
469 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
470 int *num, uint64_t *cluster_offset)
472 BDRVQcowState *s = bs->opaque;
473 unsigned int l2_index;
474 uint64_t l1_index, l2_offset, *l2_table;
475 int l1_bits, c;
476 unsigned int index_in_cluster, nb_clusters;
477 uint64_t nb_available, nb_needed;
478 int ret;
480 index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1);
481 nb_needed = *num + index_in_cluster;
483 l1_bits = s->l2_bits + s->cluster_bits;
485 /* compute how many bytes there are between the offset and
486 * the end of the l1 entry
489 nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1));
491 /* compute the number of available sectors */
493 nb_available = (nb_available >> 9) + index_in_cluster;
495 if (nb_needed > nb_available) {
496 nb_needed = nb_available;
499 *cluster_offset = 0;
501 /* seek the the l2 offset in the l1 table */
503 l1_index = offset >> l1_bits;
504 if (l1_index >= s->l1_size) {
505 ret = QCOW2_CLUSTER_UNALLOCATED;
506 goto out;
509 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
510 if (!l2_offset) {
511 ret = QCOW2_CLUSTER_UNALLOCATED;
512 goto out;
515 if (offset_into_cluster(s, l2_offset)) {
516 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
517 " unaligned (L1 index: %#" PRIx64 ")",
518 l2_offset, l1_index);
519 return -EIO;
522 /* load the l2 table in memory */
524 ret = l2_load(bs, l2_offset, &l2_table);
525 if (ret < 0) {
526 return ret;
529 /* find the cluster offset for the given disk offset */
531 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
532 *cluster_offset = be64_to_cpu(l2_table[l2_index]);
533 nb_clusters = size_to_clusters(s, nb_needed << 9);
535 ret = qcow2_get_cluster_type(*cluster_offset);
536 switch (ret) {
537 case QCOW2_CLUSTER_COMPRESSED:
538 /* Compressed clusters can only be processed one by one */
539 c = 1;
540 *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
541 break;
542 case QCOW2_CLUSTER_ZERO:
543 if (s->qcow_version < 3) {
544 qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
545 " in pre-v3 image (L2 offset: %#" PRIx64
546 ", L2 index: %#x)", l2_offset, l2_index);
547 ret = -EIO;
548 goto fail;
550 c = count_contiguous_clusters(nb_clusters, s->cluster_size,
551 &l2_table[l2_index], QCOW_OFLAG_ZERO);
552 *cluster_offset = 0;
553 break;
554 case QCOW2_CLUSTER_UNALLOCATED:
555 /* how many empty clusters ? */
556 c = count_contiguous_free_clusters(nb_clusters, &l2_table[l2_index]);
557 *cluster_offset = 0;
558 break;
559 case QCOW2_CLUSTER_NORMAL:
560 /* how many allocated clusters ? */
561 c = count_contiguous_clusters(nb_clusters, s->cluster_size,
562 &l2_table[l2_index], QCOW_OFLAG_ZERO);
563 *cluster_offset &= L2E_OFFSET_MASK;
564 if (offset_into_cluster(s, *cluster_offset)) {
565 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset %#"
566 PRIx64 " unaligned (L2 offset: %#" PRIx64
567 ", L2 index: %#x)", *cluster_offset,
568 l2_offset, l2_index);
569 ret = -EIO;
570 goto fail;
572 break;
573 default:
574 abort();
577 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
579 nb_available = (c * s->cluster_sectors);
581 out:
582 if (nb_available > nb_needed)
583 nb_available = nb_needed;
585 *num = nb_available - index_in_cluster;
587 return ret;
589 fail:
590 qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
591 return ret;
595 * get_cluster_table
597 * for a given disk offset, load (and allocate if needed)
598 * the l2 table.
600 * the l2 table offset in the qcow2 file and the cluster index
601 * in the l2 table are given to the caller.
603 * Returns 0 on success, -errno in failure case
605 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
606 uint64_t **new_l2_table,
607 int *new_l2_index)
609 BDRVQcowState *s = bs->opaque;
610 unsigned int l2_index;
611 uint64_t l1_index, l2_offset;
612 uint64_t *l2_table = NULL;
613 int ret;
615 /* seek the the l2 offset in the l1 table */
617 l1_index = offset >> (s->l2_bits + s->cluster_bits);
618 if (l1_index >= s->l1_size) {
619 ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
620 if (ret < 0) {
621 return ret;
625 assert(l1_index < s->l1_size);
626 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
627 if (offset_into_cluster(s, l2_offset)) {
628 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
629 " unaligned (L1 index: %#" PRIx64 ")",
630 l2_offset, l1_index);
631 return -EIO;
634 /* seek the l2 table of the given l2 offset */
636 if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) {
637 /* load the l2 table in memory */
638 ret = l2_load(bs, l2_offset, &l2_table);
639 if (ret < 0) {
640 return ret;
642 } else {
643 /* First allocate a new L2 table (and do COW if needed) */
644 ret = l2_allocate(bs, l1_index, &l2_table);
645 if (ret < 0) {
646 return ret;
649 /* Then decrease the refcount of the old table */
650 if (l2_offset) {
651 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
652 QCOW2_DISCARD_OTHER);
656 /* find the cluster offset for the given disk offset */
658 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
660 *new_l2_table = l2_table;
661 *new_l2_index = l2_index;
663 return 0;
667 * alloc_compressed_cluster_offset
669 * For a given offset of the disk image, return cluster offset in
670 * qcow2 file.
672 * If the offset is not found, allocate a new compressed cluster.
674 * Return the cluster offset if successful,
675 * Return 0, otherwise.
679 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
680 uint64_t offset,
681 int compressed_size)
683 BDRVQcowState *s = bs->opaque;
684 int l2_index, ret;
685 uint64_t *l2_table;
686 int64_t cluster_offset;
687 int nb_csectors;
689 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
690 if (ret < 0) {
691 return 0;
694 /* Compression can't overwrite anything. Fail if the cluster was already
695 * allocated. */
696 cluster_offset = be64_to_cpu(l2_table[l2_index]);
697 if (cluster_offset & L2E_OFFSET_MASK) {
698 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
699 return 0;
702 cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
703 if (cluster_offset < 0) {
704 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
705 return 0;
708 nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) -
709 (cluster_offset >> 9);
711 cluster_offset |= QCOW_OFLAG_COMPRESSED |
712 ((uint64_t)nb_csectors << s->csize_shift);
714 /* update L2 table */
716 /* compressed clusters never have the copied flag */
718 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
719 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
720 l2_table[l2_index] = cpu_to_be64(cluster_offset);
721 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
723 return cluster_offset;
726 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r)
728 BDRVQcowState *s = bs->opaque;
729 int ret;
731 if (r->nb_sectors == 0) {
732 return 0;
735 qemu_co_mutex_unlock(&s->lock);
736 ret = copy_sectors(bs, m->offset / BDRV_SECTOR_SIZE, m->alloc_offset,
737 r->offset / BDRV_SECTOR_SIZE,
738 r->offset / BDRV_SECTOR_SIZE + r->nb_sectors);
739 qemu_co_mutex_lock(&s->lock);
741 if (ret < 0) {
742 return ret;
746 * Before we update the L2 table to actually point to the new cluster, we
747 * need to be sure that the refcounts have been increased and COW was
748 * handled.
750 qcow2_cache_depends_on_flush(s->l2_table_cache);
752 return 0;
755 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
757 BDRVQcowState *s = bs->opaque;
758 int i, j = 0, l2_index, ret;
759 uint64_t *old_cluster, *l2_table;
760 uint64_t cluster_offset = m->alloc_offset;
762 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
763 assert(m->nb_clusters > 0);
765 old_cluster = g_try_new(uint64_t, m->nb_clusters);
766 if (old_cluster == NULL) {
767 ret = -ENOMEM;
768 goto err;
771 /* copy content of unmodified sectors */
772 ret = perform_cow(bs, m, &m->cow_start);
773 if (ret < 0) {
774 goto err;
777 ret = perform_cow(bs, m, &m->cow_end);
778 if (ret < 0) {
779 goto err;
782 /* Update L2 table. */
783 if (s->use_lazy_refcounts) {
784 qcow2_mark_dirty(bs);
786 if (qcow2_need_accurate_refcounts(s)) {
787 qcow2_cache_set_dependency(bs, s->l2_table_cache,
788 s->refcount_block_cache);
791 ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index);
792 if (ret < 0) {
793 goto err;
795 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
797 assert(l2_index + m->nb_clusters <= s->l2_size);
798 for (i = 0; i < m->nb_clusters; i++) {
799 /* if two concurrent writes happen to the same unallocated cluster
800 * each write allocates separate cluster and writes data concurrently.
801 * The first one to complete updates l2 table with pointer to its
802 * cluster the second one has to do RMW (which is done above by
803 * copy_sectors()), update l2 table with its cluster pointer and free
804 * old cluster. This is what this loop does */
805 if(l2_table[l2_index + i] != 0)
806 old_cluster[j++] = l2_table[l2_index + i];
808 l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
809 (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
813 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
816 * If this was a COW, we need to decrease the refcount of the old cluster.
817 * Also flush bs->file to get the right order for L2 and refcount update.
819 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
820 * clusters), the next write will reuse them anyway.
822 if (j != 0) {
823 for (i = 0; i < j; i++) {
824 qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
825 QCOW2_DISCARD_NEVER);
829 ret = 0;
830 err:
831 g_free(old_cluster);
832 return ret;
836 * Returns the number of contiguous clusters that can be used for an allocating
837 * write, but require COW to be performed (this includes yet unallocated space,
838 * which must copy from the backing file)
840 static int count_cow_clusters(BDRVQcowState *s, int nb_clusters,
841 uint64_t *l2_table, int l2_index)
843 int i;
845 for (i = 0; i < nb_clusters; i++) {
846 uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]);
847 int cluster_type = qcow2_get_cluster_type(l2_entry);
849 switch(cluster_type) {
850 case QCOW2_CLUSTER_NORMAL:
851 if (l2_entry & QCOW_OFLAG_COPIED) {
852 goto out;
854 break;
855 case QCOW2_CLUSTER_UNALLOCATED:
856 case QCOW2_CLUSTER_COMPRESSED:
857 case QCOW2_CLUSTER_ZERO:
858 break;
859 default:
860 abort();
864 out:
865 assert(i <= nb_clusters);
866 return i;
870 * Check if there already is an AIO write request in flight which allocates
871 * the same cluster. In this case we need to wait until the previous
872 * request has completed and updated the L2 table accordingly.
874 * Returns:
875 * 0 if there was no dependency. *cur_bytes indicates the number of
876 * bytes from guest_offset that can be read before the next
877 * dependency must be processed (or the request is complete)
879 * -EAGAIN if we had to wait for another request, previously gathered
880 * information on cluster allocation may be invalid now. The caller
881 * must start over anyway, so consider *cur_bytes undefined.
883 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
884 uint64_t *cur_bytes, QCowL2Meta **m)
886 BDRVQcowState *s = bs->opaque;
887 QCowL2Meta *old_alloc;
888 uint64_t bytes = *cur_bytes;
890 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
892 uint64_t start = guest_offset;
893 uint64_t end = start + bytes;
894 uint64_t old_start = l2meta_cow_start(old_alloc);
895 uint64_t old_end = l2meta_cow_end(old_alloc);
897 if (end <= old_start || start >= old_end) {
898 /* No intersection */
899 } else {
900 if (start < old_start) {
901 /* Stop at the start of a running allocation */
902 bytes = old_start - start;
903 } else {
904 bytes = 0;
907 /* Stop if already an l2meta exists. After yielding, it wouldn't
908 * be valid any more, so we'd have to clean up the old L2Metas
909 * and deal with requests depending on them before starting to
910 * gather new ones. Not worth the trouble. */
911 if (bytes == 0 && *m) {
912 *cur_bytes = 0;
913 return 0;
916 if (bytes == 0) {
917 /* Wait for the dependency to complete. We need to recheck
918 * the free/allocated clusters when we continue. */
919 qemu_co_mutex_unlock(&s->lock);
920 qemu_co_queue_wait(&old_alloc->dependent_requests);
921 qemu_co_mutex_lock(&s->lock);
922 return -EAGAIN;
927 /* Make sure that existing clusters and new allocations are only used up to
928 * the next dependency if we shortened the request above */
929 *cur_bytes = bytes;
931 return 0;
935 * Checks how many already allocated clusters that don't require a copy on
936 * write there are at the given guest_offset (up to *bytes). If
937 * *host_offset is not zero, only physically contiguous clusters beginning at
938 * this host offset are counted.
940 * Note that guest_offset may not be cluster aligned. In this case, the
941 * returned *host_offset points to exact byte referenced by guest_offset and
942 * therefore isn't cluster aligned as well.
944 * Returns:
945 * 0: if no allocated clusters are available at the given offset.
946 * *bytes is normally unchanged. It is set to 0 if the cluster
947 * is allocated and doesn't need COW, but doesn't have the right
948 * physical offset.
950 * 1: if allocated clusters that don't require a COW are available at
951 * the requested offset. *bytes may have decreased and describes
952 * the length of the area that can be written to.
954 * -errno: in error cases
956 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
957 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
959 BDRVQcowState *s = bs->opaque;
960 int l2_index;
961 uint64_t cluster_offset;
962 uint64_t *l2_table;
963 unsigned int nb_clusters;
964 unsigned int keep_clusters;
965 int ret;
967 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
968 *bytes);
970 assert(*host_offset == 0 || offset_into_cluster(s, guest_offset)
971 == offset_into_cluster(s, *host_offset));
974 * Calculate the number of clusters to look for. We stop at L2 table
975 * boundaries to keep things simple.
977 nb_clusters =
978 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
980 l2_index = offset_to_l2_index(s, guest_offset);
981 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
983 /* Find L2 entry for the first involved cluster */
984 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
985 if (ret < 0) {
986 return ret;
989 cluster_offset = be64_to_cpu(l2_table[l2_index]);
991 /* Check how many clusters are already allocated and don't need COW */
992 if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL
993 && (cluster_offset & QCOW_OFLAG_COPIED))
995 /* If a specific host_offset is required, check it */
996 bool offset_matches =
997 (cluster_offset & L2E_OFFSET_MASK) == *host_offset;
999 if (offset_into_cluster(s, cluster_offset & L2E_OFFSET_MASK)) {
1000 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset "
1001 "%#llx unaligned (guest offset: %#" PRIx64
1002 ")", cluster_offset & L2E_OFFSET_MASK,
1003 guest_offset);
1004 ret = -EIO;
1005 goto out;
1008 if (*host_offset != 0 && !offset_matches) {
1009 *bytes = 0;
1010 ret = 0;
1011 goto out;
1014 /* We keep all QCOW_OFLAG_COPIED clusters */
1015 keep_clusters =
1016 count_contiguous_clusters(nb_clusters, s->cluster_size,
1017 &l2_table[l2_index],
1018 QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
1019 assert(keep_clusters <= nb_clusters);
1021 *bytes = MIN(*bytes,
1022 keep_clusters * s->cluster_size
1023 - offset_into_cluster(s, guest_offset));
1025 ret = 1;
1026 } else {
1027 ret = 0;
1030 /* Cleanup */
1031 out:
1032 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1034 /* Only return a host offset if we actually made progress. Otherwise we
1035 * would make requirements for handle_alloc() that it can't fulfill */
1036 if (ret > 0) {
1037 *host_offset = (cluster_offset & L2E_OFFSET_MASK)
1038 + offset_into_cluster(s, guest_offset);
1041 return ret;
1045 * Allocates new clusters for the given guest_offset.
1047 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1048 * contain the number of clusters that have been allocated and are contiguous
1049 * in the image file.
1051 * If *host_offset is non-zero, it specifies the offset in the image file at
1052 * which the new clusters must start. *nb_clusters can be 0 on return in this
1053 * case if the cluster at host_offset is already in use. If *host_offset is
1054 * zero, the clusters can be allocated anywhere in the image file.
1056 * *host_offset is updated to contain the offset into the image file at which
1057 * the first allocated cluster starts.
1059 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1060 * function has been waiting for another request and the allocation must be
1061 * restarted, but the whole request should not be failed.
1063 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1064 uint64_t *host_offset, unsigned int *nb_clusters)
1066 BDRVQcowState *s = bs->opaque;
1068 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1069 *host_offset, *nb_clusters);
1071 /* Allocate new clusters */
1072 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1073 if (*host_offset == 0) {
1074 int64_t cluster_offset =
1075 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1076 if (cluster_offset < 0) {
1077 return cluster_offset;
1079 *host_offset = cluster_offset;
1080 return 0;
1081 } else {
1082 int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1083 if (ret < 0) {
1084 return ret;
1086 *nb_clusters = ret;
1087 return 0;
1092 * Allocates new clusters for an area that either is yet unallocated or needs a
1093 * copy on write. If *host_offset is non-zero, clusters are only allocated if
1094 * the new allocation can match the specified host offset.
1096 * Note that guest_offset may not be cluster aligned. In this case, the
1097 * returned *host_offset points to exact byte referenced by guest_offset and
1098 * therefore isn't cluster aligned as well.
1100 * Returns:
1101 * 0: if no clusters could be allocated. *bytes is set to 0,
1102 * *host_offset is left unchanged.
1104 * 1: if new clusters were allocated. *bytes may be decreased if the
1105 * new allocation doesn't cover all of the requested area.
1106 * *host_offset is updated to contain the host offset of the first
1107 * newly allocated cluster.
1109 * -errno: in error cases
1111 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1112 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1114 BDRVQcowState *s = bs->opaque;
1115 int l2_index;
1116 uint64_t *l2_table;
1117 uint64_t entry;
1118 unsigned int nb_clusters;
1119 int ret;
1121 uint64_t alloc_cluster_offset;
1123 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1124 *bytes);
1125 assert(*bytes > 0);
1128 * Calculate the number of clusters to look for. We stop at L2 table
1129 * boundaries to keep things simple.
1131 nb_clusters =
1132 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1134 l2_index = offset_to_l2_index(s, guest_offset);
1135 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1137 /* Find L2 entry for the first involved cluster */
1138 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1139 if (ret < 0) {
1140 return ret;
1143 entry = be64_to_cpu(l2_table[l2_index]);
1145 /* For the moment, overwrite compressed clusters one by one */
1146 if (entry & QCOW_OFLAG_COMPRESSED) {
1147 nb_clusters = 1;
1148 } else {
1149 nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index);
1152 /* This function is only called when there were no non-COW clusters, so if
1153 * we can't find any unallocated or COW clusters either, something is
1154 * wrong with our code. */
1155 assert(nb_clusters > 0);
1157 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1159 /* Allocate, if necessary at a given offset in the image file */
1160 alloc_cluster_offset = start_of_cluster(s, *host_offset);
1161 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1162 &nb_clusters);
1163 if (ret < 0) {
1164 goto fail;
1167 /* Can't extend contiguous allocation */
1168 if (nb_clusters == 0) {
1169 *bytes = 0;
1170 return 0;
1173 /* !*host_offset would overwrite the image header and is reserved for "no
1174 * host offset preferred". If 0 was a valid host offset, it'd trigger the
1175 * following overlap check; do that now to avoid having an invalid value in
1176 * *host_offset. */
1177 if (!alloc_cluster_offset) {
1178 ret = qcow2_pre_write_overlap_check(bs, 0, alloc_cluster_offset,
1179 nb_clusters * s->cluster_size);
1180 assert(ret < 0);
1181 goto fail;
1185 * Save info needed for meta data update.
1187 * requested_sectors: Number of sectors from the start of the first
1188 * newly allocated cluster to the end of the (possibly shortened
1189 * before) write request.
1191 * avail_sectors: Number of sectors from the start of the first
1192 * newly allocated to the end of the last newly allocated cluster.
1194 * nb_sectors: The number of sectors from the start of the first
1195 * newly allocated cluster to the end of the area that the write
1196 * request actually writes to (excluding COW at the end)
1198 int requested_sectors =
1199 (*bytes + offset_into_cluster(s, guest_offset))
1200 >> BDRV_SECTOR_BITS;
1201 int avail_sectors = nb_clusters
1202 << (s->cluster_bits - BDRV_SECTOR_BITS);
1203 int alloc_n_start = offset_into_cluster(s, guest_offset)
1204 >> BDRV_SECTOR_BITS;
1205 int nb_sectors = MIN(requested_sectors, avail_sectors);
1206 QCowL2Meta *old_m = *m;
1208 *m = g_malloc0(sizeof(**m));
1210 **m = (QCowL2Meta) {
1211 .next = old_m,
1213 .alloc_offset = alloc_cluster_offset,
1214 .offset = start_of_cluster(s, guest_offset),
1215 .nb_clusters = nb_clusters,
1216 .nb_available = nb_sectors,
1218 .cow_start = {
1219 .offset = 0,
1220 .nb_sectors = alloc_n_start,
1222 .cow_end = {
1223 .offset = nb_sectors * BDRV_SECTOR_SIZE,
1224 .nb_sectors = avail_sectors - nb_sectors,
1227 qemu_co_queue_init(&(*m)->dependent_requests);
1228 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1230 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1231 *bytes = MIN(*bytes, (nb_sectors * BDRV_SECTOR_SIZE)
1232 - offset_into_cluster(s, guest_offset));
1233 assert(*bytes != 0);
1235 return 1;
1237 fail:
1238 if (*m && (*m)->nb_clusters > 0) {
1239 QLIST_REMOVE(*m, next_in_flight);
1241 return ret;
1245 * alloc_cluster_offset
1247 * For a given offset on the virtual disk, find the cluster offset in qcow2
1248 * file. If the offset is not found, allocate a new cluster.
1250 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1251 * other fields in m are meaningless.
1253 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1254 * contiguous clusters that have been allocated. In this case, the other
1255 * fields of m are valid and contain information about the first allocated
1256 * cluster.
1258 * If the request conflicts with another write request in flight, the coroutine
1259 * is queued and will be reentered when the dependency has completed.
1261 * Return 0 on success and -errno in error cases
1263 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1264 int *num, uint64_t *host_offset, QCowL2Meta **m)
1266 BDRVQcowState *s = bs->opaque;
1267 uint64_t start, remaining;
1268 uint64_t cluster_offset;
1269 uint64_t cur_bytes;
1270 int ret;
1272 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *num);
1274 assert((offset & ~BDRV_SECTOR_MASK) == 0);
1276 again:
1277 start = offset;
1278 remaining = (uint64_t)*num << BDRV_SECTOR_BITS;
1279 cluster_offset = 0;
1280 *host_offset = 0;
1281 cur_bytes = 0;
1282 *m = NULL;
1284 while (true) {
1286 if (!*host_offset) {
1287 *host_offset = start_of_cluster(s, cluster_offset);
1290 assert(remaining >= cur_bytes);
1292 start += cur_bytes;
1293 remaining -= cur_bytes;
1294 cluster_offset += cur_bytes;
1296 if (remaining == 0) {
1297 break;
1300 cur_bytes = remaining;
1303 * Now start gathering as many contiguous clusters as possible:
1305 * 1. Check for overlaps with in-flight allocations
1307 * a) Overlap not in the first cluster -> shorten this request and
1308 * let the caller handle the rest in its next loop iteration.
1310 * b) Real overlaps of two requests. Yield and restart the search
1311 * for contiguous clusters (the situation could have changed
1312 * while we were sleeping)
1314 * c) TODO: Request starts in the same cluster as the in-flight
1315 * allocation ends. Shorten the COW of the in-fight allocation,
1316 * set cluster_offset to write to the same cluster and set up
1317 * the right synchronisation between the in-flight request and
1318 * the new one.
1320 ret = handle_dependencies(bs, start, &cur_bytes, m);
1321 if (ret == -EAGAIN) {
1322 /* Currently handle_dependencies() doesn't yield if we already had
1323 * an allocation. If it did, we would have to clean up the L2Meta
1324 * structs before starting over. */
1325 assert(*m == NULL);
1326 goto again;
1327 } else if (ret < 0) {
1328 return ret;
1329 } else if (cur_bytes == 0) {
1330 break;
1331 } else {
1332 /* handle_dependencies() may have decreased cur_bytes (shortened
1333 * the allocations below) so that the next dependency is processed
1334 * correctly during the next loop iteration. */
1338 * 2. Count contiguous COPIED clusters.
1340 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1341 if (ret < 0) {
1342 return ret;
1343 } else if (ret) {
1344 continue;
1345 } else if (cur_bytes == 0) {
1346 break;
1350 * 3. If the request still hasn't completed, allocate new clusters,
1351 * considering any cluster_offset of steps 1c or 2.
1353 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1354 if (ret < 0) {
1355 return ret;
1356 } else if (ret) {
1357 continue;
1358 } else {
1359 assert(cur_bytes == 0);
1360 break;
1364 *num -= remaining >> BDRV_SECTOR_BITS;
1365 assert(*num > 0);
1366 assert(*host_offset != 0);
1368 return 0;
1371 static int decompress_buffer(uint8_t *out_buf, int out_buf_size,
1372 const uint8_t *buf, int buf_size)
1374 z_stream strm1, *strm = &strm1;
1375 int ret, out_len;
1377 memset(strm, 0, sizeof(*strm));
1379 strm->next_in = (uint8_t *)buf;
1380 strm->avail_in = buf_size;
1381 strm->next_out = out_buf;
1382 strm->avail_out = out_buf_size;
1384 ret = inflateInit2(strm, -12);
1385 if (ret != Z_OK)
1386 return -1;
1387 ret = inflate(strm, Z_FINISH);
1388 out_len = strm->next_out - out_buf;
1389 if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) ||
1390 out_len != out_buf_size) {
1391 inflateEnd(strm);
1392 return -1;
1394 inflateEnd(strm);
1395 return 0;
1398 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
1400 BDRVQcowState *s = bs->opaque;
1401 int ret, csize, nb_csectors, sector_offset;
1402 uint64_t coffset;
1404 coffset = cluster_offset & s->cluster_offset_mask;
1405 if (s->cluster_cache_offset != coffset) {
1406 nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1;
1407 sector_offset = coffset & 511;
1408 csize = nb_csectors * 512 - sector_offset;
1409 BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);
1410 ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, nb_csectors);
1411 if (ret < 0) {
1412 return ret;
1414 if (decompress_buffer(s->cluster_cache, s->cluster_size,
1415 s->cluster_data + sector_offset, csize) < 0) {
1416 return -EIO;
1418 s->cluster_cache_offset = coffset;
1420 return 0;
1424 * This discards as many clusters of nb_clusters as possible at once (i.e.
1425 * all clusters in the same L2 table) and returns the number of discarded
1426 * clusters.
1428 static int discard_single_l2(BlockDriverState *bs, uint64_t offset,
1429 unsigned int nb_clusters, enum qcow2_discard_type type, bool full_discard)
1431 BDRVQcowState *s = bs->opaque;
1432 uint64_t *l2_table;
1433 int l2_index;
1434 int ret;
1435 int i;
1437 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1438 if (ret < 0) {
1439 return ret;
1442 /* Limit nb_clusters to one L2 table */
1443 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1445 for (i = 0; i < nb_clusters; i++) {
1446 uint64_t old_l2_entry;
1448 old_l2_entry = be64_to_cpu(l2_table[l2_index + i]);
1451 * If full_discard is false, make sure that a discarded area reads back
1452 * as zeroes for v3 images (we cannot do it for v2 without actually
1453 * writing a zero-filled buffer). We can skip the operation if the
1454 * cluster is already marked as zero, or if it's unallocated and we
1455 * don't have a backing file.
1457 * TODO We might want to use bdrv_get_block_status(bs) here, but we're
1458 * holding s->lock, so that doesn't work today.
1460 * If full_discard is true, the sector should not read back as zeroes,
1461 * but rather fall through to the backing file.
1463 switch (qcow2_get_cluster_type(old_l2_entry)) {
1464 case QCOW2_CLUSTER_UNALLOCATED:
1465 if (full_discard || !bs->backing_hd) {
1466 continue;
1468 break;
1470 case QCOW2_CLUSTER_ZERO:
1471 if (!full_discard) {
1472 continue;
1474 break;
1476 case QCOW2_CLUSTER_NORMAL:
1477 case QCOW2_CLUSTER_COMPRESSED:
1478 break;
1480 default:
1481 abort();
1484 /* First remove L2 entries */
1485 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1486 if (!full_discard && s->qcow_version >= 3) {
1487 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1488 } else {
1489 l2_table[l2_index + i] = cpu_to_be64(0);
1492 /* Then decrease the refcount */
1493 qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1496 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1498 return nb_clusters;
1501 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
1502 int nb_sectors, enum qcow2_discard_type type, bool full_discard)
1504 BDRVQcowState *s = bs->opaque;
1505 uint64_t end_offset;
1506 unsigned int nb_clusters;
1507 int ret;
1509 end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
1511 /* Round start up and end down */
1512 offset = align_offset(offset, s->cluster_size);
1513 end_offset = start_of_cluster(s, end_offset);
1515 if (offset > end_offset) {
1516 return 0;
1519 nb_clusters = size_to_clusters(s, end_offset - offset);
1521 s->cache_discards = true;
1523 /* Each L2 table is handled by its own loop iteration */
1524 while (nb_clusters > 0) {
1525 ret = discard_single_l2(bs, offset, nb_clusters, type, full_discard);
1526 if (ret < 0) {
1527 goto fail;
1530 nb_clusters -= ret;
1531 offset += (ret * s->cluster_size);
1534 ret = 0;
1535 fail:
1536 s->cache_discards = false;
1537 qcow2_process_discards(bs, ret);
1539 return ret;
1543 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1544 * all clusters in the same L2 table) and returns the number of zeroed
1545 * clusters.
1547 static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
1548 unsigned int nb_clusters)
1550 BDRVQcowState *s = bs->opaque;
1551 uint64_t *l2_table;
1552 int l2_index;
1553 int ret;
1554 int i;
1556 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1557 if (ret < 0) {
1558 return ret;
1561 /* Limit nb_clusters to one L2 table */
1562 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1564 for (i = 0; i < nb_clusters; i++) {
1565 uint64_t old_offset;
1567 old_offset = be64_to_cpu(l2_table[l2_index + i]);
1569 /* Update L2 entries */
1570 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1571 if (old_offset & QCOW_OFLAG_COMPRESSED) {
1572 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1573 qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1574 } else {
1575 l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1579 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1581 return nb_clusters;
1584 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors)
1586 BDRVQcowState *s = bs->opaque;
1587 unsigned int nb_clusters;
1588 int ret;
1590 /* The zero flag is only supported by version 3 and newer */
1591 if (s->qcow_version < 3) {
1592 return -ENOTSUP;
1595 /* Each L2 table is handled by its own loop iteration */
1596 nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS);
1598 s->cache_discards = true;
1600 while (nb_clusters > 0) {
1601 ret = zero_single_l2(bs, offset, nb_clusters);
1602 if (ret < 0) {
1603 goto fail;
1606 nb_clusters -= ret;
1607 offset += (ret * s->cluster_size);
1610 ret = 0;
1611 fail:
1612 s->cache_discards = false;
1613 qcow2_process_discards(bs, ret);
1615 return ret;
1619 * Expands all zero clusters in a specific L1 table (or deallocates them, for
1620 * non-backed non-pre-allocated zero clusters).
1622 * l1_entries and *visited_l1_entries are used to keep track of progress for
1623 * status_cb(). l1_entries contains the total number of L1 entries and
1624 * *visited_l1_entries counts all visited L1 entries.
1626 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1627 int l1_size, int64_t *visited_l1_entries,
1628 int64_t l1_entries,
1629 BlockDriverAmendStatusCB *status_cb)
1631 BDRVQcowState *s = bs->opaque;
1632 bool is_active_l1 = (l1_table == s->l1_table);
1633 uint64_t *l2_table = NULL;
1634 int ret;
1635 int i, j;
1637 if (!is_active_l1) {
1638 /* inactive L2 tables require a buffer to be stored in when loading
1639 * them from disk */
1640 l2_table = qemu_try_blockalign(bs->file, s->cluster_size);
1641 if (l2_table == NULL) {
1642 return -ENOMEM;
1646 for (i = 0; i < l1_size; i++) {
1647 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1648 bool l2_dirty = false;
1649 uint64_t l2_refcount;
1651 if (!l2_offset) {
1652 /* unallocated */
1653 (*visited_l1_entries)++;
1654 if (status_cb) {
1655 status_cb(bs, *visited_l1_entries, l1_entries);
1657 continue;
1660 if (offset_into_cluster(s, l2_offset)) {
1661 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1662 PRIx64 " unaligned (L1 index: %#x)",
1663 l2_offset, i);
1664 ret = -EIO;
1665 goto fail;
1668 if (is_active_l1) {
1669 /* get active L2 tables from cache */
1670 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
1671 (void **)&l2_table);
1672 } else {
1673 /* load inactive L2 tables from disk */
1674 ret = bdrv_read(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1675 (void *)l2_table, s->cluster_sectors);
1677 if (ret < 0) {
1678 goto fail;
1681 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1682 &l2_refcount);
1683 if (ret < 0) {
1684 goto fail;
1687 for (j = 0; j < s->l2_size; j++) {
1688 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1689 int64_t offset = l2_entry & L2E_OFFSET_MASK;
1690 int cluster_type = qcow2_get_cluster_type(l2_entry);
1691 bool preallocated = offset != 0;
1693 if (cluster_type != QCOW2_CLUSTER_ZERO) {
1694 continue;
1697 if (!preallocated) {
1698 if (!bs->backing_hd) {
1699 /* not backed; therefore we can simply deallocate the
1700 * cluster */
1701 l2_table[j] = 0;
1702 l2_dirty = true;
1703 continue;
1706 offset = qcow2_alloc_clusters(bs, s->cluster_size);
1707 if (offset < 0) {
1708 ret = offset;
1709 goto fail;
1712 if (l2_refcount > 1) {
1713 /* For shared L2 tables, set the refcount accordingly (it is
1714 * already 1 and needs to be l2_refcount) */
1715 ret = qcow2_update_cluster_refcount(bs,
1716 offset >> s->cluster_bits,
1717 refcount_diff(1, l2_refcount), false,
1718 QCOW2_DISCARD_OTHER);
1719 if (ret < 0) {
1720 qcow2_free_clusters(bs, offset, s->cluster_size,
1721 QCOW2_DISCARD_OTHER);
1722 goto fail;
1727 if (offset_into_cluster(s, offset)) {
1728 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset "
1729 "%#" PRIx64 " unaligned (L2 offset: %#"
1730 PRIx64 ", L2 index: %#x)", offset,
1731 l2_offset, j);
1732 if (!preallocated) {
1733 qcow2_free_clusters(bs, offset, s->cluster_size,
1734 QCOW2_DISCARD_ALWAYS);
1736 ret = -EIO;
1737 goto fail;
1740 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size);
1741 if (ret < 0) {
1742 if (!preallocated) {
1743 qcow2_free_clusters(bs, offset, s->cluster_size,
1744 QCOW2_DISCARD_ALWAYS);
1746 goto fail;
1749 ret = bdrv_write_zeroes(bs->file, offset / BDRV_SECTOR_SIZE,
1750 s->cluster_sectors, 0);
1751 if (ret < 0) {
1752 if (!preallocated) {
1753 qcow2_free_clusters(bs, offset, s->cluster_size,
1754 QCOW2_DISCARD_ALWAYS);
1756 goto fail;
1759 if (l2_refcount == 1) {
1760 l2_table[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
1761 } else {
1762 l2_table[j] = cpu_to_be64(offset);
1764 l2_dirty = true;
1767 if (is_active_l1) {
1768 if (l2_dirty) {
1769 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
1770 qcow2_cache_depends_on_flush(s->l2_table_cache);
1772 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1773 } else {
1774 if (l2_dirty) {
1775 ret = qcow2_pre_write_overlap_check(bs,
1776 QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2, l2_offset,
1777 s->cluster_size);
1778 if (ret < 0) {
1779 goto fail;
1782 ret = bdrv_write(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1783 (void *)l2_table, s->cluster_sectors);
1784 if (ret < 0) {
1785 goto fail;
1790 (*visited_l1_entries)++;
1791 if (status_cb) {
1792 status_cb(bs, *visited_l1_entries, l1_entries);
1796 ret = 0;
1798 fail:
1799 if (l2_table) {
1800 if (!is_active_l1) {
1801 qemu_vfree(l2_table);
1802 } else {
1803 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table);
1806 return ret;
1810 * For backed images, expands all zero clusters on the image. For non-backed
1811 * images, deallocates all non-pre-allocated zero clusters (and claims the
1812 * allocation for pre-allocated ones). This is important for downgrading to a
1813 * qcow2 version which doesn't yet support metadata zero clusters.
1815 int qcow2_expand_zero_clusters(BlockDriverState *bs,
1816 BlockDriverAmendStatusCB *status_cb)
1818 BDRVQcowState *s = bs->opaque;
1819 uint64_t *l1_table = NULL;
1820 int64_t l1_entries = 0, visited_l1_entries = 0;
1821 int ret;
1822 int i, j;
1824 if (status_cb) {
1825 l1_entries = s->l1_size;
1826 for (i = 0; i < s->nb_snapshots; i++) {
1827 l1_entries += s->snapshots[i].l1_size;
1831 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
1832 &visited_l1_entries, l1_entries,
1833 status_cb);
1834 if (ret < 0) {
1835 goto fail;
1838 /* Inactive L1 tables may point to active L2 tables - therefore it is
1839 * necessary to flush the L2 table cache before trying to access the L2
1840 * tables pointed to by inactive L1 entries (else we might try to expand
1841 * zero clusters that have already been expanded); furthermore, it is also
1842 * necessary to empty the L2 table cache, since it may contain tables which
1843 * are now going to be modified directly on disk, bypassing the cache.
1844 * qcow2_cache_empty() does both for us. */
1845 ret = qcow2_cache_empty(bs, s->l2_table_cache);
1846 if (ret < 0) {
1847 goto fail;
1850 for (i = 0; i < s->nb_snapshots; i++) {
1851 int l1_sectors = (s->snapshots[i].l1_size * sizeof(uint64_t) +
1852 BDRV_SECTOR_SIZE - 1) / BDRV_SECTOR_SIZE;
1854 l1_table = g_realloc(l1_table, l1_sectors * BDRV_SECTOR_SIZE);
1856 ret = bdrv_read(bs->file, s->snapshots[i].l1_table_offset /
1857 BDRV_SECTOR_SIZE, (void *)l1_table, l1_sectors);
1858 if (ret < 0) {
1859 goto fail;
1862 for (j = 0; j < s->snapshots[i].l1_size; j++) {
1863 be64_to_cpus(&l1_table[j]);
1866 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
1867 &visited_l1_entries, l1_entries,
1868 status_cb);
1869 if (ret < 0) {
1870 goto fail;
1874 ret = 0;
1876 fail:
1877 g_free(l1_table);
1878 return ret;