Merge remote-tracking branch 'remotes/kevin/tags/for-upstream' into staging
[qemu.git] / block / qcow2-cluster.c
1 /*
2 * Block driver for the QCOW version 2 format
3 *
4 * Copyright (c) 2004-2006 Fabrice Bellard
5 *
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:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
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.
23 */
24
25 #include <zlib.h>
26
27 #include "qemu-common.h"
28 #include "block/block_int.h"
29 #include "block/qcow2.h"
30 #include "trace.h"
31
32 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
33 bool exact_size)
34 {
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];
41
42 if (min_size <= s->l1_size)
43 return 0;
44
45 if (exact_size) {
46 new_l1_size = min_size;
47 } else {
48 /* Bump size up to reduce the number of times we have to grow */
49 new_l1_size = s->l1_size;
50 if (new_l1_size == 0) {
51 new_l1_size = 1;
52 }
53 while (min_size > new_l1_size) {
54 new_l1_size = (new_l1_size * 3 + 1) / 2;
55 }
56 }
57
58 if (new_l1_size > INT_MAX) {
59 return -EFBIG;
60 }
61
62 #ifdef DEBUG_ALLOC2
63 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
64 s->l1_size, new_l1_size);
65 #endif
66
67 new_l1_size2 = sizeof(uint64_t) * new_l1_size;
68 new_l1_table = g_malloc0(align_offset(new_l1_size2, 512));
69 memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
70
71 /* write new table (align to cluster) */
72 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
73 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
74 if (new_l1_table_offset < 0) {
75 g_free(new_l1_table);
76 return new_l1_table_offset;
77 }
78
79 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
80 if (ret < 0) {
81 goto fail;
82 }
83
84 /* the L1 position has not yet been updated, so these clusters must
85 * indeed be completely free */
86 ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
87 new_l1_size2);
88 if (ret < 0) {
89 goto fail;
90 }
91
92 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
93 for(i = 0; i < s->l1_size; i++)
94 new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
95 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_table, new_l1_size2);
96 if (ret < 0)
97 goto fail;
98 for(i = 0; i < s->l1_size; i++)
99 new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
100
101 /* set new table */
102 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
103 cpu_to_be32w((uint32_t*)data, new_l1_size);
104 stq_be_p(data + 4, new_l1_table_offset);
105 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), data,sizeof(data));
106 if (ret < 0) {
107 goto fail;
108 }
109 g_free(s->l1_table);
110 old_l1_table_offset = s->l1_table_offset;
111 s->l1_table_offset = new_l1_table_offset;
112 s->l1_table = new_l1_table;
113 old_l1_size = s->l1_size;
114 s->l1_size = new_l1_size;
115 qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
116 QCOW2_DISCARD_OTHER);
117 return 0;
118 fail:
119 g_free(new_l1_table);
120 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
121 QCOW2_DISCARD_OTHER);
122 return ret;
123 }
124
125 /*
126 * l2_load
127 *
128 * Loads a L2 table into memory. If the table is in the cache, the cache
129 * is used; otherwise the L2 table is loaded from the image file.
130 *
131 * Returns a pointer to the L2 table on success, or NULL if the read from
132 * the image file failed.
133 */
134
135 static int l2_load(BlockDriverState *bs, uint64_t l2_offset,
136 uint64_t **l2_table)
137 {
138 BDRVQcowState *s = bs->opaque;
139 int ret;
140
141 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, (void**) l2_table);
142
143 return ret;
144 }
145
146 /*
147 * Writes one sector of the L1 table to the disk (can't update single entries
148 * and we really don't want bdrv_pread to perform a read-modify-write)
149 */
150 #define L1_ENTRIES_PER_SECTOR (512 / 8)
151 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
152 {
153 BDRVQcowState *s = bs->opaque;
154 uint64_t buf[L1_ENTRIES_PER_SECTOR];
155 int l1_start_index;
156 int i, ret;
157
158 l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
159 for (i = 0; i < L1_ENTRIES_PER_SECTOR; i++) {
160 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
161 }
162
163 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
164 s->l1_table_offset + 8 * l1_start_index, sizeof(buf));
165 if (ret < 0) {
166 return ret;
167 }
168
169 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
170 ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset + 8 * l1_start_index,
171 buf, sizeof(buf));
172 if (ret < 0) {
173 return ret;
174 }
175
176 return 0;
177 }
178
179 /*
180 * l2_allocate
181 *
182 * Allocate a new l2 entry in the file. If l1_index points to an already
183 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
184 * table) copy the contents of the old L2 table into the newly allocated one.
185 * Otherwise the new table is initialized with zeros.
186 *
187 */
188
189 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table)
190 {
191 BDRVQcowState *s = bs->opaque;
192 uint64_t old_l2_offset;
193 uint64_t *l2_table = NULL;
194 int64_t l2_offset;
195 int ret;
196
197 old_l2_offset = s->l1_table[l1_index];
198
199 trace_qcow2_l2_allocate(bs, l1_index);
200
201 /* allocate a new l2 entry */
202
203 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
204 if (l2_offset < 0) {
205 ret = l2_offset;
206 goto fail;
207 }
208
209 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
210 if (ret < 0) {
211 goto fail;
212 }
213
214 /* allocate a new entry in the l2 cache */
215
216 trace_qcow2_l2_allocate_get_empty(bs, l1_index);
217 ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table);
218 if (ret < 0) {
219 goto fail;
220 }
221
222 l2_table = *table;
223
224 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
225 /* if there was no old l2 table, clear the new table */
226 memset(l2_table, 0, s->l2_size * sizeof(uint64_t));
227 } else {
228 uint64_t* old_table;
229
230 /* if there was an old l2 table, read it from the disk */
231 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
232 ret = qcow2_cache_get(bs, s->l2_table_cache,
233 old_l2_offset & L1E_OFFSET_MASK,
234 (void**) &old_table);
235 if (ret < 0) {
236 goto fail;
237 }
238
239 memcpy(l2_table, old_table, s->cluster_size);
240
241 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &old_table);
242 if (ret < 0) {
243 goto fail;
244 }
245 }
246
247 /* write the l2 table to the file */
248 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
249
250 trace_qcow2_l2_allocate_write_l2(bs, l1_index);
251 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
252 ret = qcow2_cache_flush(bs, s->l2_table_cache);
253 if (ret < 0) {
254 goto fail;
255 }
256
257 /* update the L1 entry */
258 trace_qcow2_l2_allocate_write_l1(bs, l1_index);
259 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
260 ret = qcow2_write_l1_entry(bs, l1_index);
261 if (ret < 0) {
262 goto fail;
263 }
264
265 *table = l2_table;
266 trace_qcow2_l2_allocate_done(bs, l1_index, 0);
267 return 0;
268
269 fail:
270 trace_qcow2_l2_allocate_done(bs, l1_index, ret);
271 if (l2_table != NULL) {
272 qcow2_cache_put(bs, s->l2_table_cache, (void**) table);
273 }
274 s->l1_table[l1_index] = old_l2_offset;
275 if (l2_offset > 0) {
276 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
277 QCOW2_DISCARD_ALWAYS);
278 }
279 return ret;
280 }
281
282 /*
283 * Checks how many clusters in a given L2 table are contiguous in the image
284 * file. As soon as one of the flags in the bitmask stop_flags changes compared
285 * to the first cluster, the search is stopped and the cluster is not counted
286 * as contiguous. (This allows it, for example, to stop at the first compressed
287 * cluster which may require a different handling)
288 */
289 static int count_contiguous_clusters(uint64_t nb_clusters, int cluster_size,
290 uint64_t *l2_table, uint64_t stop_flags)
291 {
292 int i;
293 uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
294 uint64_t first_entry = be64_to_cpu(l2_table[0]);
295 uint64_t offset = first_entry & mask;
296
297 if (!offset)
298 return 0;
299
300 assert(qcow2_get_cluster_type(first_entry) != QCOW2_CLUSTER_COMPRESSED);
301
302 for (i = 0; i < nb_clusters; i++) {
303 uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask;
304 if (offset + (uint64_t) i * cluster_size != l2_entry) {
305 break;
306 }
307 }
308
309 return i;
310 }
311
312 static int count_contiguous_free_clusters(uint64_t nb_clusters, uint64_t *l2_table)
313 {
314 int i;
315
316 for (i = 0; i < nb_clusters; i++) {
317 int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i]));
318
319 if (type != QCOW2_CLUSTER_UNALLOCATED) {
320 break;
321 }
322 }
323
324 return i;
325 }
326
327 /* The crypt function is compatible with the linux cryptoloop
328 algorithm for < 4 GB images. NOTE: out_buf == in_buf is
329 supported */
330 void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
331 uint8_t *out_buf, const uint8_t *in_buf,
332 int nb_sectors, int enc,
333 const AES_KEY *key)
334 {
335 union {
336 uint64_t ll[2];
337 uint8_t b[16];
338 } ivec;
339 int i;
340
341 for(i = 0; i < nb_sectors; i++) {
342 ivec.ll[0] = cpu_to_le64(sector_num);
343 ivec.ll[1] = 0;
344 AES_cbc_encrypt(in_buf, out_buf, 512, key,
345 ivec.b, enc);
346 sector_num++;
347 in_buf += 512;
348 out_buf += 512;
349 }
350 }
351
352 static int coroutine_fn copy_sectors(BlockDriverState *bs,
353 uint64_t start_sect,
354 uint64_t cluster_offset,
355 int n_start, int n_end)
356 {
357 BDRVQcowState *s = bs->opaque;
358 QEMUIOVector qiov;
359 struct iovec iov;
360 int n, ret;
361
362 /*
363 * If this is the last cluster and it is only partially used, we must only
364 * copy until the end of the image, or bdrv_check_request will fail for the
365 * bdrv_read/write calls below.
366 */
367 if (start_sect + n_end > bs->total_sectors) {
368 n_end = bs->total_sectors - start_sect;
369 }
370
371 n = n_end - n_start;
372 if (n <= 0) {
373 return 0;
374 }
375
376 iov.iov_len = n * BDRV_SECTOR_SIZE;
377 iov.iov_base = qemu_blockalign(bs, iov.iov_len);
378
379 qemu_iovec_init_external(&qiov, &iov, 1);
380
381 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
382
383 if (!bs->drv) {
384 return -ENOMEDIUM;
385 }
386
387 /* Call .bdrv_co_readv() directly instead of using the public block-layer
388 * interface. This avoids double I/O throttling and request tracking,
389 * which can lead to deadlock when block layer copy-on-read is enabled.
390 */
391 ret = bs->drv->bdrv_co_readv(bs, start_sect + n_start, n, &qiov);
392 if (ret < 0) {
393 goto out;
394 }
395
396 if (s->crypt_method) {
397 qcow2_encrypt_sectors(s, start_sect + n_start,
398 iov.iov_base, iov.iov_base, n, 1,
399 &s->aes_encrypt_key);
400 }
401
402 ret = qcow2_pre_write_overlap_check(bs, 0,
403 cluster_offset + n_start * BDRV_SECTOR_SIZE, n * BDRV_SECTOR_SIZE);
404 if (ret < 0) {
405 goto out;
406 }
407
408 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
409 ret = bdrv_co_writev(bs->file, (cluster_offset >> 9) + n_start, n, &qiov);
410 if (ret < 0) {
411 goto out;
412 }
413
414 ret = 0;
415 out:
416 qemu_vfree(iov.iov_base);
417 return ret;
418 }
419
420
421 /*
422 * get_cluster_offset
423 *
424 * For a given offset of the disk image, find the cluster offset in
425 * qcow2 file. The offset is stored in *cluster_offset.
426 *
427 * on entry, *num is the number of contiguous sectors we'd like to
428 * access following offset.
429 *
430 * on exit, *num is the number of contiguous sectors we can read.
431 *
432 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
433 * cases.
434 */
435 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
436 int *num, uint64_t *cluster_offset)
437 {
438 BDRVQcowState *s = bs->opaque;
439 unsigned int l2_index;
440 uint64_t l1_index, l2_offset, *l2_table;
441 int l1_bits, c;
442 unsigned int index_in_cluster, nb_clusters;
443 uint64_t nb_available, nb_needed;
444 int ret;
445
446 index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1);
447 nb_needed = *num + index_in_cluster;
448
449 l1_bits = s->l2_bits + s->cluster_bits;
450
451 /* compute how many bytes there are between the offset and
452 * the end of the l1 entry
453 */
454
455 nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1));
456
457 /* compute the number of available sectors */
458
459 nb_available = (nb_available >> 9) + index_in_cluster;
460
461 if (nb_needed > nb_available) {
462 nb_needed = nb_available;
463 }
464
465 *cluster_offset = 0;
466
467 /* seek the the l2 offset in the l1 table */
468
469 l1_index = offset >> l1_bits;
470 if (l1_index >= s->l1_size) {
471 ret = QCOW2_CLUSTER_UNALLOCATED;
472 goto out;
473 }
474
475 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
476 if (!l2_offset) {
477 ret = QCOW2_CLUSTER_UNALLOCATED;
478 goto out;
479 }
480
481 /* load the l2 table in memory */
482
483 ret = l2_load(bs, l2_offset, &l2_table);
484 if (ret < 0) {
485 return ret;
486 }
487
488 /* find the cluster offset for the given disk offset */
489
490 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
491 *cluster_offset = be64_to_cpu(l2_table[l2_index]);
492 nb_clusters = size_to_clusters(s, nb_needed << 9);
493
494 ret = qcow2_get_cluster_type(*cluster_offset);
495 switch (ret) {
496 case QCOW2_CLUSTER_COMPRESSED:
497 /* Compressed clusters can only be processed one by one */
498 c = 1;
499 *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
500 break;
501 case QCOW2_CLUSTER_ZERO:
502 if (s->qcow_version < 3) {
503 return -EIO;
504 }
505 c = count_contiguous_clusters(nb_clusters, s->cluster_size,
506 &l2_table[l2_index], QCOW_OFLAG_ZERO);
507 *cluster_offset = 0;
508 break;
509 case QCOW2_CLUSTER_UNALLOCATED:
510 /* how many empty clusters ? */
511 c = count_contiguous_free_clusters(nb_clusters, &l2_table[l2_index]);
512 *cluster_offset = 0;
513 break;
514 case QCOW2_CLUSTER_NORMAL:
515 /* how many allocated clusters ? */
516 c = count_contiguous_clusters(nb_clusters, s->cluster_size,
517 &l2_table[l2_index], QCOW_OFLAG_ZERO);
518 *cluster_offset &= L2E_OFFSET_MASK;
519 break;
520 default:
521 abort();
522 }
523
524 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
525
526 nb_available = (c * s->cluster_sectors);
527
528 out:
529 if (nb_available > nb_needed)
530 nb_available = nb_needed;
531
532 *num = nb_available - index_in_cluster;
533
534 return ret;
535 }
536
537 /*
538 * get_cluster_table
539 *
540 * for a given disk offset, load (and allocate if needed)
541 * the l2 table.
542 *
543 * the l2 table offset in the qcow2 file and the cluster index
544 * in the l2 table are given to the caller.
545 *
546 * Returns 0 on success, -errno in failure case
547 */
548 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
549 uint64_t **new_l2_table,
550 int *new_l2_index)
551 {
552 BDRVQcowState *s = bs->opaque;
553 unsigned int l2_index;
554 uint64_t l1_index, l2_offset;
555 uint64_t *l2_table = NULL;
556 int ret;
557
558 /* seek the the l2 offset in the l1 table */
559
560 l1_index = offset >> (s->l2_bits + s->cluster_bits);
561 if (l1_index >= s->l1_size) {
562 ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
563 if (ret < 0) {
564 return ret;
565 }
566 }
567
568 assert(l1_index < s->l1_size);
569 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
570
571 /* seek the l2 table of the given l2 offset */
572
573 if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) {
574 /* load the l2 table in memory */
575 ret = l2_load(bs, l2_offset, &l2_table);
576 if (ret < 0) {
577 return ret;
578 }
579 } else {
580 /* First allocate a new L2 table (and do COW if needed) */
581 ret = l2_allocate(bs, l1_index, &l2_table);
582 if (ret < 0) {
583 return ret;
584 }
585
586 /* Then decrease the refcount of the old table */
587 if (l2_offset) {
588 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
589 QCOW2_DISCARD_OTHER);
590 }
591 }
592
593 /* find the cluster offset for the given disk offset */
594
595 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
596
597 *new_l2_table = l2_table;
598 *new_l2_index = l2_index;
599
600 return 0;
601 }
602
603 /*
604 * alloc_compressed_cluster_offset
605 *
606 * For a given offset of the disk image, return cluster offset in
607 * qcow2 file.
608 *
609 * If the offset is not found, allocate a new compressed cluster.
610 *
611 * Return the cluster offset if successful,
612 * Return 0, otherwise.
613 *
614 */
615
616 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
617 uint64_t offset,
618 int compressed_size)
619 {
620 BDRVQcowState *s = bs->opaque;
621 int l2_index, ret;
622 uint64_t *l2_table;
623 int64_t cluster_offset;
624 int nb_csectors;
625
626 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
627 if (ret < 0) {
628 return 0;
629 }
630
631 /* Compression can't overwrite anything. Fail if the cluster was already
632 * allocated. */
633 cluster_offset = be64_to_cpu(l2_table[l2_index]);
634 if (cluster_offset & L2E_OFFSET_MASK) {
635 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
636 return 0;
637 }
638
639 cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
640 if (cluster_offset < 0) {
641 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
642 return 0;
643 }
644
645 nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) -
646 (cluster_offset >> 9);
647
648 cluster_offset |= QCOW_OFLAG_COMPRESSED |
649 ((uint64_t)nb_csectors << s->csize_shift);
650
651 /* update L2 table */
652
653 /* compressed clusters never have the copied flag */
654
655 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
656 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
657 l2_table[l2_index] = cpu_to_be64(cluster_offset);
658 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
659 if (ret < 0) {
660 return 0;
661 }
662
663 return cluster_offset;
664 }
665
666 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r)
667 {
668 BDRVQcowState *s = bs->opaque;
669 int ret;
670
671 if (r->nb_sectors == 0) {
672 return 0;
673 }
674
675 qemu_co_mutex_unlock(&s->lock);
676 ret = copy_sectors(bs, m->offset / BDRV_SECTOR_SIZE, m->alloc_offset,
677 r->offset / BDRV_SECTOR_SIZE,
678 r->offset / BDRV_SECTOR_SIZE + r->nb_sectors);
679 qemu_co_mutex_lock(&s->lock);
680
681 if (ret < 0) {
682 return ret;
683 }
684
685 /*
686 * Before we update the L2 table to actually point to the new cluster, we
687 * need to be sure that the refcounts have been increased and COW was
688 * handled.
689 */
690 qcow2_cache_depends_on_flush(s->l2_table_cache);
691
692 return 0;
693 }
694
695 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
696 {
697 BDRVQcowState *s = bs->opaque;
698 int i, j = 0, l2_index, ret;
699 uint64_t *old_cluster, *l2_table;
700 uint64_t cluster_offset = m->alloc_offset;
701
702 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
703 assert(m->nb_clusters > 0);
704
705 old_cluster = g_malloc(m->nb_clusters * sizeof(uint64_t));
706
707 /* copy content of unmodified sectors */
708 ret = perform_cow(bs, m, &m->cow_start);
709 if (ret < 0) {
710 goto err;
711 }
712
713 ret = perform_cow(bs, m, &m->cow_end);
714 if (ret < 0) {
715 goto err;
716 }
717
718 /* Update L2 table. */
719 if (s->use_lazy_refcounts) {
720 qcow2_mark_dirty(bs);
721 }
722 if (qcow2_need_accurate_refcounts(s)) {
723 qcow2_cache_set_dependency(bs, s->l2_table_cache,
724 s->refcount_block_cache);
725 }
726
727 ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index);
728 if (ret < 0) {
729 goto err;
730 }
731 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
732
733 assert(l2_index + m->nb_clusters <= s->l2_size);
734 for (i = 0; i < m->nb_clusters; i++) {
735 /* if two concurrent writes happen to the same unallocated cluster
736 * each write allocates separate cluster and writes data concurrently.
737 * The first one to complete updates l2 table with pointer to its
738 * cluster the second one has to do RMW (which is done above by
739 * copy_sectors()), update l2 table with its cluster pointer and free
740 * old cluster. This is what this loop does */
741 if(l2_table[l2_index + i] != 0)
742 old_cluster[j++] = l2_table[l2_index + i];
743
744 l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
745 (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
746 }
747
748
749 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
750 if (ret < 0) {
751 goto err;
752 }
753
754 /*
755 * If this was a COW, we need to decrease the refcount of the old cluster.
756 * Also flush bs->file to get the right order for L2 and refcount update.
757 *
758 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
759 * clusters), the next write will reuse them anyway.
760 */
761 if (j != 0) {
762 for (i = 0; i < j; i++) {
763 qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
764 QCOW2_DISCARD_NEVER);
765 }
766 }
767
768 ret = 0;
769 err:
770 g_free(old_cluster);
771 return ret;
772 }
773
774 /*
775 * Returns the number of contiguous clusters that can be used for an allocating
776 * write, but require COW to be performed (this includes yet unallocated space,
777 * which must copy from the backing file)
778 */
779 static int count_cow_clusters(BDRVQcowState *s, int nb_clusters,
780 uint64_t *l2_table, int l2_index)
781 {
782 int i;
783
784 for (i = 0; i < nb_clusters; i++) {
785 uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]);
786 int cluster_type = qcow2_get_cluster_type(l2_entry);
787
788 switch(cluster_type) {
789 case QCOW2_CLUSTER_NORMAL:
790 if (l2_entry & QCOW_OFLAG_COPIED) {
791 goto out;
792 }
793 break;
794 case QCOW2_CLUSTER_UNALLOCATED:
795 case QCOW2_CLUSTER_COMPRESSED:
796 case QCOW2_CLUSTER_ZERO:
797 break;
798 default:
799 abort();
800 }
801 }
802
803 out:
804 assert(i <= nb_clusters);
805 return i;
806 }
807
808 /*
809 * Check if there already is an AIO write request in flight which allocates
810 * the same cluster. In this case we need to wait until the previous
811 * request has completed and updated the L2 table accordingly.
812 *
813 * Returns:
814 * 0 if there was no dependency. *cur_bytes indicates the number of
815 * bytes from guest_offset that can be read before the next
816 * dependency must be processed (or the request is complete)
817 *
818 * -EAGAIN if we had to wait for another request, previously gathered
819 * information on cluster allocation may be invalid now. The caller
820 * must start over anyway, so consider *cur_bytes undefined.
821 */
822 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
823 uint64_t *cur_bytes, QCowL2Meta **m)
824 {
825 BDRVQcowState *s = bs->opaque;
826 QCowL2Meta *old_alloc;
827 uint64_t bytes = *cur_bytes;
828
829 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
830
831 uint64_t start = guest_offset;
832 uint64_t end = start + bytes;
833 uint64_t old_start = l2meta_cow_start(old_alloc);
834 uint64_t old_end = l2meta_cow_end(old_alloc);
835
836 if (end <= old_start || start >= old_end) {
837 /* No intersection */
838 } else {
839 if (start < old_start) {
840 /* Stop at the start of a running allocation */
841 bytes = old_start - start;
842 } else {
843 bytes = 0;
844 }
845
846 /* Stop if already an l2meta exists. After yielding, it wouldn't
847 * be valid any more, so we'd have to clean up the old L2Metas
848 * and deal with requests depending on them before starting to
849 * gather new ones. Not worth the trouble. */
850 if (bytes == 0 && *m) {
851 *cur_bytes = 0;
852 return 0;
853 }
854
855 if (bytes == 0) {
856 /* Wait for the dependency to complete. We need to recheck
857 * the free/allocated clusters when we continue. */
858 qemu_co_mutex_unlock(&s->lock);
859 qemu_co_queue_wait(&old_alloc->dependent_requests);
860 qemu_co_mutex_lock(&s->lock);
861 return -EAGAIN;
862 }
863 }
864 }
865
866 /* Make sure that existing clusters and new allocations are only used up to
867 * the next dependency if we shortened the request above */
868 *cur_bytes = bytes;
869
870 return 0;
871 }
872
873 /*
874 * Checks how many already allocated clusters that don't require a copy on
875 * write there are at the given guest_offset (up to *bytes). If
876 * *host_offset is not zero, only physically contiguous clusters beginning at
877 * this host offset are counted.
878 *
879 * Note that guest_offset may not be cluster aligned. In this case, the
880 * returned *host_offset points to exact byte referenced by guest_offset and
881 * therefore isn't cluster aligned as well.
882 *
883 * Returns:
884 * 0: if no allocated clusters are available at the given offset.
885 * *bytes is normally unchanged. It is set to 0 if the cluster
886 * is allocated and doesn't need COW, but doesn't have the right
887 * physical offset.
888 *
889 * 1: if allocated clusters that don't require a COW are available at
890 * the requested offset. *bytes may have decreased and describes
891 * the length of the area that can be written to.
892 *
893 * -errno: in error cases
894 */
895 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
896 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
897 {
898 BDRVQcowState *s = bs->opaque;
899 int l2_index;
900 uint64_t cluster_offset;
901 uint64_t *l2_table;
902 unsigned int nb_clusters;
903 unsigned int keep_clusters;
904 int ret, pret;
905
906 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
907 *bytes);
908
909 assert(*host_offset == 0 || offset_into_cluster(s, guest_offset)
910 == offset_into_cluster(s, *host_offset));
911
912 /*
913 * Calculate the number of clusters to look for. We stop at L2 table
914 * boundaries to keep things simple.
915 */
916 nb_clusters =
917 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
918
919 l2_index = offset_to_l2_index(s, guest_offset);
920 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
921
922 /* Find L2 entry for the first involved cluster */
923 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
924 if (ret < 0) {
925 return ret;
926 }
927
928 cluster_offset = be64_to_cpu(l2_table[l2_index]);
929
930 /* Check how many clusters are already allocated and don't need COW */
931 if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL
932 && (cluster_offset & QCOW_OFLAG_COPIED))
933 {
934 /* If a specific host_offset is required, check it */
935 bool offset_matches =
936 (cluster_offset & L2E_OFFSET_MASK) == *host_offset;
937
938 if (*host_offset != 0 && !offset_matches) {
939 *bytes = 0;
940 ret = 0;
941 goto out;
942 }
943
944 /* We keep all QCOW_OFLAG_COPIED clusters */
945 keep_clusters =
946 count_contiguous_clusters(nb_clusters, s->cluster_size,
947 &l2_table[l2_index],
948 QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
949 assert(keep_clusters <= nb_clusters);
950
951 *bytes = MIN(*bytes,
952 keep_clusters * s->cluster_size
953 - offset_into_cluster(s, guest_offset));
954
955 ret = 1;
956 } else {
957 ret = 0;
958 }
959
960 /* Cleanup */
961 out:
962 pret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
963 if (pret < 0) {
964 return pret;
965 }
966
967 /* Only return a host offset if we actually made progress. Otherwise we
968 * would make requirements for handle_alloc() that it can't fulfill */
969 if (ret) {
970 *host_offset = (cluster_offset & L2E_OFFSET_MASK)
971 + offset_into_cluster(s, guest_offset);
972 }
973
974 return ret;
975 }
976
977 /*
978 * Allocates new clusters for the given guest_offset.
979 *
980 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
981 * contain the number of clusters that have been allocated and are contiguous
982 * in the image file.
983 *
984 * If *host_offset is non-zero, it specifies the offset in the image file at
985 * which the new clusters must start. *nb_clusters can be 0 on return in this
986 * case if the cluster at host_offset is already in use. If *host_offset is
987 * zero, the clusters can be allocated anywhere in the image file.
988 *
989 * *host_offset is updated to contain the offset into the image file at which
990 * the first allocated cluster starts.
991 *
992 * Return 0 on success and -errno in error cases. -EAGAIN means that the
993 * function has been waiting for another request and the allocation must be
994 * restarted, but the whole request should not be failed.
995 */
996 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
997 uint64_t *host_offset, unsigned int *nb_clusters)
998 {
999 BDRVQcowState *s = bs->opaque;
1000
1001 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1002 *host_offset, *nb_clusters);
1003
1004 /* Allocate new clusters */
1005 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1006 if (*host_offset == 0) {
1007 int64_t cluster_offset =
1008 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1009 if (cluster_offset < 0) {
1010 return cluster_offset;
1011 }
1012 *host_offset = cluster_offset;
1013 return 0;
1014 } else {
1015 int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1016 if (ret < 0) {
1017 return ret;
1018 }
1019 *nb_clusters = ret;
1020 return 0;
1021 }
1022 }
1023
1024 /*
1025 * Allocates new clusters for an area that either is yet unallocated or needs a
1026 * copy on write. If *host_offset is non-zero, clusters are only allocated if
1027 * the new allocation can match the specified host offset.
1028 *
1029 * Note that guest_offset may not be cluster aligned. In this case, the
1030 * returned *host_offset points to exact byte referenced by guest_offset and
1031 * therefore isn't cluster aligned as well.
1032 *
1033 * Returns:
1034 * 0: if no clusters could be allocated. *bytes is set to 0,
1035 * *host_offset is left unchanged.
1036 *
1037 * 1: if new clusters were allocated. *bytes may be decreased if the
1038 * new allocation doesn't cover all of the requested area.
1039 * *host_offset is updated to contain the host offset of the first
1040 * newly allocated cluster.
1041 *
1042 * -errno: in error cases
1043 */
1044 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1045 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1046 {
1047 BDRVQcowState *s = bs->opaque;
1048 int l2_index;
1049 uint64_t *l2_table;
1050 uint64_t entry;
1051 unsigned int nb_clusters;
1052 int ret;
1053
1054 uint64_t alloc_cluster_offset;
1055
1056 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1057 *bytes);
1058 assert(*bytes > 0);
1059
1060 /*
1061 * Calculate the number of clusters to look for. We stop at L2 table
1062 * boundaries to keep things simple.
1063 */
1064 nb_clusters =
1065 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1066
1067 l2_index = offset_to_l2_index(s, guest_offset);
1068 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1069
1070 /* Find L2 entry for the first involved cluster */
1071 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1072 if (ret < 0) {
1073 return ret;
1074 }
1075
1076 entry = be64_to_cpu(l2_table[l2_index]);
1077
1078 /* For the moment, overwrite compressed clusters one by one */
1079 if (entry & QCOW_OFLAG_COMPRESSED) {
1080 nb_clusters = 1;
1081 } else {
1082 nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index);
1083 }
1084
1085 /* This function is only called when there were no non-COW clusters, so if
1086 * we can't find any unallocated or COW clusters either, something is
1087 * wrong with our code. */
1088 assert(nb_clusters > 0);
1089
1090 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1091 if (ret < 0) {
1092 return ret;
1093 }
1094
1095 /* Allocate, if necessary at a given offset in the image file */
1096 alloc_cluster_offset = start_of_cluster(s, *host_offset);
1097 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1098 &nb_clusters);
1099 if (ret < 0) {
1100 goto fail;
1101 }
1102
1103 /* Can't extend contiguous allocation */
1104 if (nb_clusters == 0) {
1105 *bytes = 0;
1106 return 0;
1107 }
1108
1109 /*
1110 * Save info needed for meta data update.
1111 *
1112 * requested_sectors: Number of sectors from the start of the first
1113 * newly allocated cluster to the end of the (possibly shortened
1114 * before) write request.
1115 *
1116 * avail_sectors: Number of sectors from the start of the first
1117 * newly allocated to the end of the last newly allocated cluster.
1118 *
1119 * nb_sectors: The number of sectors from the start of the first
1120 * newly allocated cluster to the end of the area that the write
1121 * request actually writes to (excluding COW at the end)
1122 */
1123 int requested_sectors =
1124 (*bytes + offset_into_cluster(s, guest_offset))
1125 >> BDRV_SECTOR_BITS;
1126 int avail_sectors = nb_clusters
1127 << (s->cluster_bits - BDRV_SECTOR_BITS);
1128 int alloc_n_start = offset_into_cluster(s, guest_offset)
1129 >> BDRV_SECTOR_BITS;
1130 int nb_sectors = MIN(requested_sectors, avail_sectors);
1131 QCowL2Meta *old_m = *m;
1132
1133 *m = g_malloc0(sizeof(**m));
1134
1135 **m = (QCowL2Meta) {
1136 .next = old_m,
1137
1138 .alloc_offset = alloc_cluster_offset,
1139 .offset = start_of_cluster(s, guest_offset),
1140 .nb_clusters = nb_clusters,
1141 .nb_available = nb_sectors,
1142
1143 .cow_start = {
1144 .offset = 0,
1145 .nb_sectors = alloc_n_start,
1146 },
1147 .cow_end = {
1148 .offset = nb_sectors * BDRV_SECTOR_SIZE,
1149 .nb_sectors = avail_sectors - nb_sectors,
1150 },
1151 };
1152 qemu_co_queue_init(&(*m)->dependent_requests);
1153 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1154
1155 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1156 *bytes = MIN(*bytes, (nb_sectors * BDRV_SECTOR_SIZE)
1157 - offset_into_cluster(s, guest_offset));
1158 assert(*bytes != 0);
1159
1160 return 1;
1161
1162 fail:
1163 if (*m && (*m)->nb_clusters > 0) {
1164 QLIST_REMOVE(*m, next_in_flight);
1165 }
1166 return ret;
1167 }
1168
1169 /*
1170 * alloc_cluster_offset
1171 *
1172 * For a given offset on the virtual disk, find the cluster offset in qcow2
1173 * file. If the offset is not found, allocate a new cluster.
1174 *
1175 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1176 * other fields in m are meaningless.
1177 *
1178 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1179 * contiguous clusters that have been allocated. In this case, the other
1180 * fields of m are valid and contain information about the first allocated
1181 * cluster.
1182 *
1183 * If the request conflicts with another write request in flight, the coroutine
1184 * is queued and will be reentered when the dependency has completed.
1185 *
1186 * Return 0 on success and -errno in error cases
1187 */
1188 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1189 int *num, uint64_t *host_offset, QCowL2Meta **m)
1190 {
1191 BDRVQcowState *s = bs->opaque;
1192 uint64_t start, remaining;
1193 uint64_t cluster_offset;
1194 uint64_t cur_bytes;
1195 int ret;
1196
1197 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *num);
1198
1199 assert((offset & ~BDRV_SECTOR_MASK) == 0);
1200
1201 again:
1202 start = offset;
1203 remaining = *num << BDRV_SECTOR_BITS;
1204 cluster_offset = 0;
1205 *host_offset = 0;
1206 cur_bytes = 0;
1207 *m = NULL;
1208
1209 while (true) {
1210
1211 if (!*host_offset) {
1212 *host_offset = start_of_cluster(s, cluster_offset);
1213 }
1214
1215 assert(remaining >= cur_bytes);
1216
1217 start += cur_bytes;
1218 remaining -= cur_bytes;
1219 cluster_offset += cur_bytes;
1220
1221 if (remaining == 0) {
1222 break;
1223 }
1224
1225 cur_bytes = remaining;
1226
1227 /*
1228 * Now start gathering as many contiguous clusters as possible:
1229 *
1230 * 1. Check for overlaps with in-flight allocations
1231 *
1232 * a) Overlap not in the first cluster -> shorten this request and
1233 * let the caller handle the rest in its next loop iteration.
1234 *
1235 * b) Real overlaps of two requests. Yield and restart the search
1236 * for contiguous clusters (the situation could have changed
1237 * while we were sleeping)
1238 *
1239 * c) TODO: Request starts in the same cluster as the in-flight
1240 * allocation ends. Shorten the COW of the in-fight allocation,
1241 * set cluster_offset to write to the same cluster and set up
1242 * the right synchronisation between the in-flight request and
1243 * the new one.
1244 */
1245 ret = handle_dependencies(bs, start, &cur_bytes, m);
1246 if (ret == -EAGAIN) {
1247 /* Currently handle_dependencies() doesn't yield if we already had
1248 * an allocation. If it did, we would have to clean up the L2Meta
1249 * structs before starting over. */
1250 assert(*m == NULL);
1251 goto again;
1252 } else if (ret < 0) {
1253 return ret;
1254 } else if (cur_bytes == 0) {
1255 break;
1256 } else {
1257 /* handle_dependencies() may have decreased cur_bytes (shortened
1258 * the allocations below) so that the next dependency is processed
1259 * correctly during the next loop iteration. */
1260 }
1261
1262 /*
1263 * 2. Count contiguous COPIED clusters.
1264 */
1265 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1266 if (ret < 0) {
1267 return ret;
1268 } else if (ret) {
1269 continue;
1270 } else if (cur_bytes == 0) {
1271 break;
1272 }
1273
1274 /*
1275 * 3. If the request still hasn't completed, allocate new clusters,
1276 * considering any cluster_offset of steps 1c or 2.
1277 */
1278 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1279 if (ret < 0) {
1280 return ret;
1281 } else if (ret) {
1282 continue;
1283 } else {
1284 assert(cur_bytes == 0);
1285 break;
1286 }
1287 }
1288
1289 *num -= remaining >> BDRV_SECTOR_BITS;
1290 assert(*num > 0);
1291 assert(*host_offset != 0);
1292
1293 return 0;
1294 }
1295
1296 static int decompress_buffer(uint8_t *out_buf, int out_buf_size,
1297 const uint8_t *buf, int buf_size)
1298 {
1299 z_stream strm1, *strm = &strm1;
1300 int ret, out_len;
1301
1302 memset(strm, 0, sizeof(*strm));
1303
1304 strm->next_in = (uint8_t *)buf;
1305 strm->avail_in = buf_size;
1306 strm->next_out = out_buf;
1307 strm->avail_out = out_buf_size;
1308
1309 ret = inflateInit2(strm, -12);
1310 if (ret != Z_OK)
1311 return -1;
1312 ret = inflate(strm, Z_FINISH);
1313 out_len = strm->next_out - out_buf;
1314 if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) ||
1315 out_len != out_buf_size) {
1316 inflateEnd(strm);
1317 return -1;
1318 }
1319 inflateEnd(strm);
1320 return 0;
1321 }
1322
1323 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
1324 {
1325 BDRVQcowState *s = bs->opaque;
1326 int ret, csize, nb_csectors, sector_offset;
1327 uint64_t coffset;
1328
1329 coffset = cluster_offset & s->cluster_offset_mask;
1330 if (s->cluster_cache_offset != coffset) {
1331 nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1;
1332 sector_offset = coffset & 511;
1333 csize = nb_csectors * 512 - sector_offset;
1334 BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);
1335 ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, nb_csectors);
1336 if (ret < 0) {
1337 return ret;
1338 }
1339 if (decompress_buffer(s->cluster_cache, s->cluster_size,
1340 s->cluster_data + sector_offset, csize) < 0) {
1341 return -EIO;
1342 }
1343 s->cluster_cache_offset = coffset;
1344 }
1345 return 0;
1346 }
1347
1348 /*
1349 * This discards as many clusters of nb_clusters as possible at once (i.e.
1350 * all clusters in the same L2 table) and returns the number of discarded
1351 * clusters.
1352 */
1353 static int discard_single_l2(BlockDriverState *bs, uint64_t offset,
1354 unsigned int nb_clusters, enum qcow2_discard_type type)
1355 {
1356 BDRVQcowState *s = bs->opaque;
1357 uint64_t *l2_table;
1358 int l2_index;
1359 int ret;
1360 int i;
1361
1362 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1363 if (ret < 0) {
1364 return ret;
1365 }
1366
1367 /* Limit nb_clusters to one L2 table */
1368 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1369
1370 for (i = 0; i < nb_clusters; i++) {
1371 uint64_t old_offset;
1372
1373 old_offset = be64_to_cpu(l2_table[l2_index + i]);
1374
1375 /*
1376 * Make sure that a discarded area reads back as zeroes for v3 images
1377 * (we cannot do it for v2 without actually writing a zero-filled
1378 * buffer). We can skip the operation if the cluster is already marked
1379 * as zero, or if it's unallocated and we don't have a backing file.
1380 *
1381 * TODO We might want to use bdrv_get_block_status(bs) here, but we're
1382 * holding s->lock, so that doesn't work today.
1383 */
1384 if (old_offset & QCOW_OFLAG_ZERO) {
1385 continue;
1386 }
1387
1388 if ((old_offset & L2E_OFFSET_MASK) == 0 && !bs->backing_hd) {
1389 continue;
1390 }
1391
1392 /* First remove L2 entries */
1393 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1394 if (s->qcow_version >= 3) {
1395 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1396 } else {
1397 l2_table[l2_index + i] = cpu_to_be64(0);
1398 }
1399
1400 /* Then decrease the refcount */
1401 qcow2_free_any_clusters(bs, old_offset, 1, type);
1402 }
1403
1404 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1405 if (ret < 0) {
1406 return ret;
1407 }
1408
1409 return nb_clusters;
1410 }
1411
1412 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
1413 int nb_sectors, enum qcow2_discard_type type)
1414 {
1415 BDRVQcowState *s = bs->opaque;
1416 uint64_t end_offset;
1417 unsigned int nb_clusters;
1418 int ret;
1419
1420 end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
1421
1422 /* Round start up and end down */
1423 offset = align_offset(offset, s->cluster_size);
1424 end_offset = start_of_cluster(s, end_offset);
1425
1426 if (offset > end_offset) {
1427 return 0;
1428 }
1429
1430 nb_clusters = size_to_clusters(s, end_offset - offset);
1431
1432 s->cache_discards = true;
1433
1434 /* Each L2 table is handled by its own loop iteration */
1435 while (nb_clusters > 0) {
1436 ret = discard_single_l2(bs, offset, nb_clusters, type);
1437 if (ret < 0) {
1438 goto fail;
1439 }
1440
1441 nb_clusters -= ret;
1442 offset += (ret * s->cluster_size);
1443 }
1444
1445 ret = 0;
1446 fail:
1447 s->cache_discards = false;
1448 qcow2_process_discards(bs, ret);
1449
1450 return ret;
1451 }
1452
1453 /*
1454 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1455 * all clusters in the same L2 table) and returns the number of zeroed
1456 * clusters.
1457 */
1458 static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
1459 unsigned int nb_clusters)
1460 {
1461 BDRVQcowState *s = bs->opaque;
1462 uint64_t *l2_table;
1463 int l2_index;
1464 int ret;
1465 int i;
1466
1467 ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1468 if (ret < 0) {
1469 return ret;
1470 }
1471
1472 /* Limit nb_clusters to one L2 table */
1473 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1474
1475 for (i = 0; i < nb_clusters; i++) {
1476 uint64_t old_offset;
1477
1478 old_offset = be64_to_cpu(l2_table[l2_index + i]);
1479
1480 /* Update L2 entries */
1481 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1482 if (old_offset & QCOW_OFLAG_COMPRESSED) {
1483 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1484 qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1485 } else {
1486 l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1487 }
1488 }
1489
1490 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1491 if (ret < 0) {
1492 return ret;
1493 }
1494
1495 return nb_clusters;
1496 }
1497
1498 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors)
1499 {
1500 BDRVQcowState *s = bs->opaque;
1501 unsigned int nb_clusters;
1502 int ret;
1503
1504 /* The zero flag is only supported by version 3 and newer */
1505 if (s->qcow_version < 3) {
1506 return -ENOTSUP;
1507 }
1508
1509 /* Each L2 table is handled by its own loop iteration */
1510 nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS);
1511
1512 s->cache_discards = true;
1513
1514 while (nb_clusters > 0) {
1515 ret = zero_single_l2(bs, offset, nb_clusters);
1516 if (ret < 0) {
1517 goto fail;
1518 }
1519
1520 nb_clusters -= ret;
1521 offset += (ret * s->cluster_size);
1522 }
1523
1524 ret = 0;
1525 fail:
1526 s->cache_discards = false;
1527 qcow2_process_discards(bs, ret);
1528
1529 return ret;
1530 }
1531
1532 /*
1533 * Expands all zero clusters in a specific L1 table (or deallocates them, for
1534 * non-backed non-pre-allocated zero clusters).
1535 *
1536 * expanded_clusters is a bitmap where every bit corresponds to one cluster in
1537 * the image file; a bit gets set if the corresponding cluster has been used for
1538 * zero expansion (i.e., has been filled with zeroes and is referenced from an
1539 * L2 table). nb_clusters contains the total cluster count of the image file,
1540 * i.e., the number of bits in expanded_clusters.
1541 */
1542 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1543 int l1_size, uint8_t **expanded_clusters,
1544 uint64_t *nb_clusters)
1545 {
1546 BDRVQcowState *s = bs->opaque;
1547 bool is_active_l1 = (l1_table == s->l1_table);
1548 uint64_t *l2_table = NULL;
1549 int ret;
1550 int i, j;
1551
1552 if (!is_active_l1) {
1553 /* inactive L2 tables require a buffer to be stored in when loading
1554 * them from disk */
1555 l2_table = qemu_blockalign(bs, s->cluster_size);
1556 }
1557
1558 for (i = 0; i < l1_size; i++) {
1559 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1560 bool l2_dirty = false;
1561
1562 if (!l2_offset) {
1563 /* unallocated */
1564 continue;
1565 }
1566
1567 if (is_active_l1) {
1568 /* get active L2 tables from cache */
1569 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
1570 (void **)&l2_table);
1571 } else {
1572 /* load inactive L2 tables from disk */
1573 ret = bdrv_read(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1574 (void *)l2_table, s->cluster_sectors);
1575 }
1576 if (ret < 0) {
1577 goto fail;
1578 }
1579
1580 for (j = 0; j < s->l2_size; j++) {
1581 uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1582 int64_t offset = l2_entry & L2E_OFFSET_MASK, cluster_index;
1583 int cluster_type = qcow2_get_cluster_type(l2_entry);
1584 bool preallocated = offset != 0;
1585
1586 if (cluster_type == QCOW2_CLUSTER_NORMAL) {
1587 cluster_index = offset >> s->cluster_bits;
1588 assert((cluster_index >= 0) && (cluster_index < *nb_clusters));
1589 if ((*expanded_clusters)[cluster_index / 8] &
1590 (1 << (cluster_index % 8))) {
1591 /* Probably a shared L2 table; this cluster was a zero
1592 * cluster which has been expanded, its refcount
1593 * therefore most likely requires an update. */
1594 ret = qcow2_update_cluster_refcount(bs, cluster_index, 1,
1595 QCOW2_DISCARD_NEVER);
1596 if (ret < 0) {
1597 goto fail;
1598 }
1599 /* Since we just increased the refcount, the COPIED flag may
1600 * no longer be set. */
1601 l2_table[j] = cpu_to_be64(l2_entry & ~QCOW_OFLAG_COPIED);
1602 l2_dirty = true;
1603 }
1604 continue;
1605 }
1606 else if (qcow2_get_cluster_type(l2_entry) != QCOW2_CLUSTER_ZERO) {
1607 continue;
1608 }
1609
1610 if (!preallocated) {
1611 if (!bs->backing_hd) {
1612 /* not backed; therefore we can simply deallocate the
1613 * cluster */
1614 l2_table[j] = 0;
1615 l2_dirty = true;
1616 continue;
1617 }
1618
1619 offset = qcow2_alloc_clusters(bs, s->cluster_size);
1620 if (offset < 0) {
1621 ret = offset;
1622 goto fail;
1623 }
1624 }
1625
1626 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size);
1627 if (ret < 0) {
1628 if (!preallocated) {
1629 qcow2_free_clusters(bs, offset, s->cluster_size,
1630 QCOW2_DISCARD_ALWAYS);
1631 }
1632 goto fail;
1633 }
1634
1635 ret = bdrv_write_zeroes(bs->file, offset / BDRV_SECTOR_SIZE,
1636 s->cluster_sectors, 0);
1637 if (ret < 0) {
1638 if (!preallocated) {
1639 qcow2_free_clusters(bs, offset, s->cluster_size,
1640 QCOW2_DISCARD_ALWAYS);
1641 }
1642 goto fail;
1643 }
1644
1645 l2_table[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
1646 l2_dirty = true;
1647
1648 cluster_index = offset >> s->cluster_bits;
1649
1650 if (cluster_index >= *nb_clusters) {
1651 uint64_t old_bitmap_size = (*nb_clusters + 7) / 8;
1652 uint64_t new_bitmap_size;
1653 /* The offset may lie beyond the old end of the underlying image
1654 * file for growable files only */
1655 assert(bs->file->growable);
1656 *nb_clusters = size_to_clusters(s, bs->file->total_sectors *
1657 BDRV_SECTOR_SIZE);
1658 new_bitmap_size = (*nb_clusters + 7) / 8;
1659 *expanded_clusters = g_realloc(*expanded_clusters,
1660 new_bitmap_size);
1661 /* clear the newly allocated space */
1662 memset(&(*expanded_clusters)[old_bitmap_size], 0,
1663 new_bitmap_size - old_bitmap_size);
1664 }
1665
1666 assert((cluster_index >= 0) && (cluster_index < *nb_clusters));
1667 (*expanded_clusters)[cluster_index / 8] |= 1 << (cluster_index % 8);
1668 }
1669
1670 if (is_active_l1) {
1671 if (l2_dirty) {
1672 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1673 qcow2_cache_depends_on_flush(s->l2_table_cache);
1674 }
1675 ret = qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
1676 if (ret < 0) {
1677 l2_table = NULL;
1678 goto fail;
1679 }
1680 } else {
1681 if (l2_dirty) {
1682 ret = qcow2_pre_write_overlap_check(bs,
1683 QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2, l2_offset,
1684 s->cluster_size);
1685 if (ret < 0) {
1686 goto fail;
1687 }
1688
1689 ret = bdrv_write(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1690 (void *)l2_table, s->cluster_sectors);
1691 if (ret < 0) {
1692 goto fail;
1693 }
1694 }
1695 }
1696 }
1697
1698 ret = 0;
1699
1700 fail:
1701 if (l2_table) {
1702 if (!is_active_l1) {
1703 qemu_vfree(l2_table);
1704 } else {
1705 if (ret < 0) {
1706 qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
1707 } else {
1708 ret = qcow2_cache_put(bs, s->l2_table_cache,
1709 (void **)&l2_table);
1710 }
1711 }
1712 }
1713 return ret;
1714 }
1715
1716 /*
1717 * For backed images, expands all zero clusters on the image. For non-backed
1718 * images, deallocates all non-pre-allocated zero clusters (and claims the
1719 * allocation for pre-allocated ones). This is important for downgrading to a
1720 * qcow2 version which doesn't yet support metadata zero clusters.
1721 */
1722 int qcow2_expand_zero_clusters(BlockDriverState *bs)
1723 {
1724 BDRVQcowState *s = bs->opaque;
1725 uint64_t *l1_table = NULL;
1726 uint64_t nb_clusters;
1727 uint8_t *expanded_clusters;
1728 int ret;
1729 int i, j;
1730
1731 nb_clusters = size_to_clusters(s, bs->file->total_sectors *
1732 BDRV_SECTOR_SIZE);
1733 expanded_clusters = g_malloc0((nb_clusters + 7) / 8);
1734
1735 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
1736 &expanded_clusters, &nb_clusters);
1737 if (ret < 0) {
1738 goto fail;
1739 }
1740
1741 /* Inactive L1 tables may point to active L2 tables - therefore it is
1742 * necessary to flush the L2 table cache before trying to access the L2
1743 * tables pointed to by inactive L1 entries (else we might try to expand
1744 * zero clusters that have already been expanded); furthermore, it is also
1745 * necessary to empty the L2 table cache, since it may contain tables which
1746 * are now going to be modified directly on disk, bypassing the cache.
1747 * qcow2_cache_empty() does both for us. */
1748 ret = qcow2_cache_empty(bs, s->l2_table_cache);
1749 if (ret < 0) {
1750 goto fail;
1751 }
1752
1753 for (i = 0; i < s->nb_snapshots; i++) {
1754 int l1_sectors = (s->snapshots[i].l1_size * sizeof(uint64_t) +
1755 BDRV_SECTOR_SIZE - 1) / BDRV_SECTOR_SIZE;
1756
1757 l1_table = g_realloc(l1_table, l1_sectors * BDRV_SECTOR_SIZE);
1758
1759 ret = bdrv_read(bs->file, s->snapshots[i].l1_table_offset /
1760 BDRV_SECTOR_SIZE, (void *)l1_table, l1_sectors);
1761 if (ret < 0) {
1762 goto fail;
1763 }
1764
1765 for (j = 0; j < s->snapshots[i].l1_size; j++) {
1766 be64_to_cpus(&l1_table[j]);
1767 }
1768
1769 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
1770 &expanded_clusters, &nb_clusters);
1771 if (ret < 0) {
1772 goto fail;
1773 }
1774 }
1775
1776 ret = 0;
1777
1778 fail:
1779 g_free(expanded_clusters);
1780 g_free(l1_table);
1781 return ret;
1782 }