qcow2: Use macros for the L1, refcount and bitmap table entry sizes
[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 "qemu/osdep.h"
26 #include <zlib.h>
27
28 #include "qapi/error.h"
29 #include "qcow2.h"
30 #include "qemu/bswap.h"
31 #include "trace.h"
32
33 int qcow2_shrink_l1_table(BlockDriverState *bs, uint64_t exact_size)
34 {
35 BDRVQcow2State *s = bs->opaque;
36 int new_l1_size, i, ret;
37
38 if (exact_size >= s->l1_size) {
39 return 0;
40 }
41
42 new_l1_size = exact_size;
43
44 #ifdef DEBUG_ALLOC2
45 fprintf(stderr, "shrink l1_table from %d to %d\n", s->l1_size, new_l1_size);
46 #endif
47
48 BLKDBG_EVENT(bs->file, BLKDBG_L1_SHRINK_WRITE_TABLE);
49 ret = bdrv_pwrite_zeroes(bs->file, s->l1_table_offset +
50 new_l1_size * L1E_SIZE,
51 (s->l1_size - new_l1_size) * L1E_SIZE, 0);
52 if (ret < 0) {
53 goto fail;
54 }
55
56 ret = bdrv_flush(bs->file->bs);
57 if (ret < 0) {
58 goto fail;
59 }
60
61 BLKDBG_EVENT(bs->file, BLKDBG_L1_SHRINK_FREE_L2_CLUSTERS);
62 for (i = s->l1_size - 1; i > new_l1_size - 1; i--) {
63 if ((s->l1_table[i] & L1E_OFFSET_MASK) == 0) {
64 continue;
65 }
66 qcow2_free_clusters(bs, s->l1_table[i] & L1E_OFFSET_MASK,
67 s->cluster_size, QCOW2_DISCARD_ALWAYS);
68 s->l1_table[i] = 0;
69 }
70 return 0;
71
72 fail:
73 /*
74 * If the write in the l1_table failed the image may contain a partially
75 * overwritten l1_table. In this case it would be better to clear the
76 * l1_table in memory to avoid possible image corruption.
77 */
78 memset(s->l1_table + new_l1_size, 0,
79 (s->l1_size - new_l1_size) * L1E_SIZE);
80 return ret;
81 }
82
83 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
84 bool exact_size)
85 {
86 BDRVQcow2State *s = bs->opaque;
87 int new_l1_size2, ret, i;
88 uint64_t *new_l1_table;
89 int64_t old_l1_table_offset, old_l1_size;
90 int64_t new_l1_table_offset, new_l1_size;
91 uint8_t data[12];
92
93 if (min_size <= s->l1_size)
94 return 0;
95
96 /* Do a sanity check on min_size before trying to calculate new_l1_size
97 * (this prevents overflows during the while loop for the calculation of
98 * new_l1_size) */
99 if (min_size > INT_MAX / L1E_SIZE) {
100 return -EFBIG;
101 }
102
103 if (exact_size) {
104 new_l1_size = min_size;
105 } else {
106 /* Bump size up to reduce the number of times we have to grow */
107 new_l1_size = s->l1_size;
108 if (new_l1_size == 0) {
109 new_l1_size = 1;
110 }
111 while (min_size > new_l1_size) {
112 new_l1_size = DIV_ROUND_UP(new_l1_size * 3, 2);
113 }
114 }
115
116 QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX);
117 if (new_l1_size > QCOW_MAX_L1_SIZE / L1E_SIZE) {
118 return -EFBIG;
119 }
120
121 #ifdef DEBUG_ALLOC2
122 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
123 s->l1_size, new_l1_size);
124 #endif
125
126 new_l1_size2 = L1E_SIZE * new_l1_size;
127 new_l1_table = qemu_try_blockalign(bs->file->bs, new_l1_size2);
128 if (new_l1_table == NULL) {
129 return -ENOMEM;
130 }
131 memset(new_l1_table, 0, new_l1_size2);
132
133 if (s->l1_size) {
134 memcpy(new_l1_table, s->l1_table, s->l1_size * L1E_SIZE);
135 }
136
137 /* write new table (align to cluster) */
138 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
139 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
140 if (new_l1_table_offset < 0) {
141 qemu_vfree(new_l1_table);
142 return new_l1_table_offset;
143 }
144
145 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
146 if (ret < 0) {
147 goto fail;
148 }
149
150 /* the L1 position has not yet been updated, so these clusters must
151 * indeed be completely free */
152 ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
153 new_l1_size2, false);
154 if (ret < 0) {
155 goto fail;
156 }
157
158 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
159 for(i = 0; i < s->l1_size; i++)
160 new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
161 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset,
162 new_l1_table, new_l1_size2);
163 if (ret < 0)
164 goto fail;
165 for(i = 0; i < s->l1_size; i++)
166 new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
167
168 /* set new table */
169 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
170 stl_be_p(data, new_l1_size);
171 stq_be_p(data + 4, new_l1_table_offset);
172 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size),
173 data, sizeof(data));
174 if (ret < 0) {
175 goto fail;
176 }
177 qemu_vfree(s->l1_table);
178 old_l1_table_offset = s->l1_table_offset;
179 s->l1_table_offset = new_l1_table_offset;
180 s->l1_table = new_l1_table;
181 old_l1_size = s->l1_size;
182 s->l1_size = new_l1_size;
183 qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * L1E_SIZE,
184 QCOW2_DISCARD_OTHER);
185 return 0;
186 fail:
187 qemu_vfree(new_l1_table);
188 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
189 QCOW2_DISCARD_OTHER);
190 return ret;
191 }
192
193 /*
194 * l2_load
195 *
196 * @bs: The BlockDriverState
197 * @offset: A guest offset, used to calculate what slice of the L2
198 * table to load.
199 * @l2_offset: Offset to the L2 table in the image file.
200 * @l2_slice: Location to store the pointer to the L2 slice.
201 *
202 * Loads a L2 slice into memory (L2 slices are the parts of L2 tables
203 * that are loaded by the qcow2 cache). If the slice is in the cache,
204 * the cache is used; otherwise the L2 slice is loaded from the image
205 * file.
206 */
207 static int l2_load(BlockDriverState *bs, uint64_t offset,
208 uint64_t l2_offset, uint64_t **l2_slice)
209 {
210 BDRVQcow2State *s = bs->opaque;
211 int start_of_slice = l2_entry_size(s) *
212 (offset_to_l2_index(s, offset) - offset_to_l2_slice_index(s, offset));
213
214 return qcow2_cache_get(bs, s->l2_table_cache, l2_offset + start_of_slice,
215 (void **)l2_slice);
216 }
217
218 /*
219 * Writes an L1 entry to disk (note that depending on the alignment
220 * requirements this function may write more that just one entry in
221 * order to prevent bdrv_pwrite from performing a read-modify-write)
222 */
223 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
224 {
225 BDRVQcow2State *s = bs->opaque;
226 int l1_start_index;
227 int i, ret;
228 int bufsize = MAX(L1E_SIZE,
229 MIN(bs->file->bs->bl.request_alignment, s->cluster_size));
230 int nentries = bufsize / L1E_SIZE;
231 g_autofree uint64_t *buf = g_try_new0(uint64_t, nentries);
232
233 if (buf == NULL) {
234 return -ENOMEM;
235 }
236
237 l1_start_index = QEMU_ALIGN_DOWN(l1_index, nentries);
238 for (i = 0; i < MIN(nentries, s->l1_size - l1_start_index); i++) {
239 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
240 }
241
242 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
243 s->l1_table_offset + 8 * l1_start_index, bufsize, false);
244 if (ret < 0) {
245 return ret;
246 }
247
248 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
249 ret = bdrv_pwrite_sync(bs->file,
250 s->l1_table_offset + 8 * l1_start_index,
251 buf, bufsize);
252 if (ret < 0) {
253 return ret;
254 }
255
256 return 0;
257 }
258
259 /*
260 * l2_allocate
261 *
262 * Allocate a new l2 entry in the file. If l1_index points to an already
263 * used entry in the L2 table (i.e. we are doing a copy on write for the L2
264 * table) copy the contents of the old L2 table into the newly allocated one.
265 * Otherwise the new table is initialized with zeros.
266 *
267 */
268
269 static int l2_allocate(BlockDriverState *bs, int l1_index)
270 {
271 BDRVQcow2State *s = bs->opaque;
272 uint64_t old_l2_offset;
273 uint64_t *l2_slice = NULL;
274 unsigned slice, slice_size2, n_slices;
275 int64_t l2_offset;
276 int ret;
277
278 old_l2_offset = s->l1_table[l1_index];
279
280 trace_qcow2_l2_allocate(bs, l1_index);
281
282 /* allocate a new l2 entry */
283
284 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * l2_entry_size(s));
285 if (l2_offset < 0) {
286 ret = l2_offset;
287 goto fail;
288 }
289
290 /* The offset must fit in the offset field of the L1 table entry */
291 assert((l2_offset & L1E_OFFSET_MASK) == l2_offset);
292
293 /* If we're allocating the table at offset 0 then something is wrong */
294 if (l2_offset == 0) {
295 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
296 "allocation of L2 table at offset 0");
297 ret = -EIO;
298 goto fail;
299 }
300
301 ret = qcow2_cache_flush(bs, s->refcount_block_cache);
302 if (ret < 0) {
303 goto fail;
304 }
305
306 /* allocate a new entry in the l2 cache */
307
308 slice_size2 = s->l2_slice_size * l2_entry_size(s);
309 n_slices = s->cluster_size / slice_size2;
310
311 trace_qcow2_l2_allocate_get_empty(bs, l1_index);
312 for (slice = 0; slice < n_slices; slice++) {
313 ret = qcow2_cache_get_empty(bs, s->l2_table_cache,
314 l2_offset + slice * slice_size2,
315 (void **) &l2_slice);
316 if (ret < 0) {
317 goto fail;
318 }
319
320 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
321 /* if there was no old l2 table, clear the new slice */
322 memset(l2_slice, 0, slice_size2);
323 } else {
324 uint64_t *old_slice;
325 uint64_t old_l2_slice_offset =
326 (old_l2_offset & L1E_OFFSET_MASK) + slice * slice_size2;
327
328 /* if there was an old l2 table, read a slice from the disk */
329 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
330 ret = qcow2_cache_get(bs, s->l2_table_cache, old_l2_slice_offset,
331 (void **) &old_slice);
332 if (ret < 0) {
333 goto fail;
334 }
335
336 memcpy(l2_slice, old_slice, slice_size2);
337
338 qcow2_cache_put(s->l2_table_cache, (void **) &old_slice);
339 }
340
341 /* write the l2 slice to the file */
342 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
343
344 trace_qcow2_l2_allocate_write_l2(bs, l1_index);
345 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
346 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
347 }
348
349 ret = qcow2_cache_flush(bs, s->l2_table_cache);
350 if (ret < 0) {
351 goto fail;
352 }
353
354 /* update the L1 entry */
355 trace_qcow2_l2_allocate_write_l1(bs, l1_index);
356 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
357 ret = qcow2_write_l1_entry(bs, l1_index);
358 if (ret < 0) {
359 goto fail;
360 }
361
362 trace_qcow2_l2_allocate_done(bs, l1_index, 0);
363 return 0;
364
365 fail:
366 trace_qcow2_l2_allocate_done(bs, l1_index, ret);
367 if (l2_slice != NULL) {
368 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
369 }
370 s->l1_table[l1_index] = old_l2_offset;
371 if (l2_offset > 0) {
372 qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
373 QCOW2_DISCARD_ALWAYS);
374 }
375 return ret;
376 }
377
378 /*
379 * For a given L2 entry, count the number of contiguous subclusters of
380 * the same type starting from @sc_from. Compressed clusters are
381 * treated as if they were divided into subclusters of size
382 * s->subcluster_size.
383 *
384 * Return the number of contiguous subclusters and set @type to the
385 * subcluster type.
386 *
387 * If the L2 entry is invalid return -errno and set @type to
388 * QCOW2_SUBCLUSTER_INVALID.
389 */
390 static int qcow2_get_subcluster_range_type(BlockDriverState *bs,
391 uint64_t l2_entry,
392 uint64_t l2_bitmap,
393 unsigned sc_from,
394 QCow2SubclusterType *type)
395 {
396 BDRVQcow2State *s = bs->opaque;
397 uint32_t val;
398
399 *type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_from);
400
401 if (*type == QCOW2_SUBCLUSTER_INVALID) {
402 return -EINVAL;
403 } else if (!has_subclusters(s) || *type == QCOW2_SUBCLUSTER_COMPRESSED) {
404 return s->subclusters_per_cluster - sc_from;
405 }
406
407 switch (*type) {
408 case QCOW2_SUBCLUSTER_NORMAL:
409 val = l2_bitmap | QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
410 return cto32(val) - sc_from;
411
412 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
413 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
414 val = (l2_bitmap | QCOW_OFLAG_SUB_ZERO_RANGE(0, sc_from)) >> 32;
415 return cto32(val) - sc_from;
416
417 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
418 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
419 val = ((l2_bitmap >> 32) | l2_bitmap)
420 & ~QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
421 return ctz32(val) - sc_from;
422
423 default:
424 g_assert_not_reached();
425 }
426 }
427
428 /*
429 * Return the number of contiguous subclusters of the exact same type
430 * in a given L2 slice, starting from cluster @l2_index, subcluster
431 * @sc_index. Allocated subclusters are required to be contiguous in
432 * the image file.
433 * At most @nb_clusters are checked (note that this means clusters,
434 * not subclusters).
435 * Compressed clusters are always processed one by one but for the
436 * purpose of this count they are treated as if they were divided into
437 * subclusters of size s->subcluster_size.
438 * On failure return -errno and update @l2_index to point to the
439 * invalid entry.
440 */
441 static int count_contiguous_subclusters(BlockDriverState *bs, int nb_clusters,
442 unsigned sc_index, uint64_t *l2_slice,
443 unsigned *l2_index)
444 {
445 BDRVQcow2State *s = bs->opaque;
446 int i, count = 0;
447 bool check_offset = false;
448 uint64_t expected_offset = 0;
449 QCow2SubclusterType expected_type = QCOW2_SUBCLUSTER_NORMAL, type;
450
451 assert(*l2_index + nb_clusters <= s->l2_slice_size);
452
453 for (i = 0; i < nb_clusters; i++) {
454 unsigned first_sc = (i == 0) ? sc_index : 0;
455 uint64_t l2_entry = get_l2_entry(s, l2_slice, *l2_index + i);
456 uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, *l2_index + i);
457 int ret = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
458 first_sc, &type);
459 if (ret < 0) {
460 *l2_index += i; /* Point to the invalid entry */
461 return -EIO;
462 }
463 if (i == 0) {
464 if (type == QCOW2_SUBCLUSTER_COMPRESSED) {
465 /* Compressed clusters are always processed one by one */
466 return ret;
467 }
468 expected_type = type;
469 expected_offset = l2_entry & L2E_OFFSET_MASK;
470 check_offset = (type == QCOW2_SUBCLUSTER_NORMAL ||
471 type == QCOW2_SUBCLUSTER_ZERO_ALLOC ||
472 type == QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC);
473 } else if (type != expected_type) {
474 break;
475 } else if (check_offset) {
476 expected_offset += s->cluster_size;
477 if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
478 break;
479 }
480 }
481 count += ret;
482 /* Stop if there are type changes before the end of the cluster */
483 if (first_sc + ret < s->subclusters_per_cluster) {
484 break;
485 }
486 }
487
488 return count;
489 }
490
491 static int coroutine_fn do_perform_cow_read(BlockDriverState *bs,
492 uint64_t src_cluster_offset,
493 unsigned offset_in_cluster,
494 QEMUIOVector *qiov)
495 {
496 int ret;
497
498 if (qiov->size == 0) {
499 return 0;
500 }
501
502 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
503
504 if (!bs->drv) {
505 return -ENOMEDIUM;
506 }
507
508 /* Call .bdrv_co_readv() directly instead of using the public block-layer
509 * interface. This avoids double I/O throttling and request tracking,
510 * which can lead to deadlock when block layer copy-on-read is enabled.
511 */
512 ret = bs->drv->bdrv_co_preadv_part(bs,
513 src_cluster_offset + offset_in_cluster,
514 qiov->size, qiov, 0, 0);
515 if (ret < 0) {
516 return ret;
517 }
518
519 return 0;
520 }
521
522 static int coroutine_fn do_perform_cow_write(BlockDriverState *bs,
523 uint64_t cluster_offset,
524 unsigned offset_in_cluster,
525 QEMUIOVector *qiov)
526 {
527 BDRVQcow2State *s = bs->opaque;
528 int ret;
529
530 if (qiov->size == 0) {
531 return 0;
532 }
533
534 ret = qcow2_pre_write_overlap_check(bs, 0,
535 cluster_offset + offset_in_cluster, qiov->size, true);
536 if (ret < 0) {
537 return ret;
538 }
539
540 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
541 ret = bdrv_co_pwritev(s->data_file, cluster_offset + offset_in_cluster,
542 qiov->size, qiov, 0);
543 if (ret < 0) {
544 return ret;
545 }
546
547 return 0;
548 }
549
550
551 /*
552 * get_host_offset
553 *
554 * For a given offset of the virtual disk find the equivalent host
555 * offset in the qcow2 file and store it in *host_offset. Neither
556 * offset needs to be aligned to a cluster boundary.
557 *
558 * If the cluster is unallocated then *host_offset will be 0.
559 * If the cluster is compressed then *host_offset will contain the
560 * complete compressed cluster descriptor.
561 *
562 * On entry, *bytes is the maximum number of contiguous bytes starting at
563 * offset that we are interested in.
564 *
565 * On exit, *bytes is the number of bytes starting at offset that have the same
566 * subcluster type and (if applicable) are stored contiguously in the image
567 * file. The subcluster type is stored in *subcluster_type.
568 * Compressed clusters are always processed one by one.
569 *
570 * Returns 0 on success, -errno in error cases.
571 */
572 int qcow2_get_host_offset(BlockDriverState *bs, uint64_t offset,
573 unsigned int *bytes, uint64_t *host_offset,
574 QCow2SubclusterType *subcluster_type)
575 {
576 BDRVQcow2State *s = bs->opaque;
577 unsigned int l2_index, sc_index;
578 uint64_t l1_index, l2_offset, *l2_slice, l2_entry, l2_bitmap;
579 int sc;
580 unsigned int offset_in_cluster;
581 uint64_t bytes_available, bytes_needed, nb_clusters;
582 QCow2SubclusterType type;
583 int ret;
584
585 offset_in_cluster = offset_into_cluster(s, offset);
586 bytes_needed = (uint64_t) *bytes + offset_in_cluster;
587
588 /* compute how many bytes there are between the start of the cluster
589 * containing offset and the end of the l2 slice that contains
590 * the entry pointing to it */
591 bytes_available =
592 ((uint64_t) (s->l2_slice_size - offset_to_l2_slice_index(s, offset)))
593 << s->cluster_bits;
594
595 if (bytes_needed > bytes_available) {
596 bytes_needed = bytes_available;
597 }
598
599 *host_offset = 0;
600
601 /* seek to the l2 offset in the l1 table */
602
603 l1_index = offset_to_l1_index(s, offset);
604 if (l1_index >= s->l1_size) {
605 type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
606 goto out;
607 }
608
609 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
610 if (!l2_offset) {
611 type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
612 goto out;
613 }
614
615 if (offset_into_cluster(s, l2_offset)) {
616 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
617 " unaligned (L1 index: %#" PRIx64 ")",
618 l2_offset, l1_index);
619 return -EIO;
620 }
621
622 /* load the l2 slice in memory */
623
624 ret = l2_load(bs, offset, l2_offset, &l2_slice);
625 if (ret < 0) {
626 return ret;
627 }
628
629 /* find the cluster offset for the given disk offset */
630
631 l2_index = offset_to_l2_slice_index(s, offset);
632 sc_index = offset_to_sc_index(s, offset);
633 l2_entry = get_l2_entry(s, l2_slice, l2_index);
634 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
635
636 nb_clusters = size_to_clusters(s, bytes_needed);
637 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
638 * integers; the minimum cluster size is 512, so this assertion is always
639 * true */
640 assert(nb_clusters <= INT_MAX);
641
642 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
643 if (s->qcow_version < 3 && (type == QCOW2_SUBCLUSTER_ZERO_PLAIN ||
644 type == QCOW2_SUBCLUSTER_ZERO_ALLOC)) {
645 qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
646 " in pre-v3 image (L2 offset: %#" PRIx64
647 ", L2 index: %#x)", l2_offset, l2_index);
648 ret = -EIO;
649 goto fail;
650 }
651 switch (type) {
652 case QCOW2_SUBCLUSTER_INVALID:
653 break; /* This is handled by count_contiguous_subclusters() below */
654 case QCOW2_SUBCLUSTER_COMPRESSED:
655 if (has_data_file(bs)) {
656 qcow2_signal_corruption(bs, true, -1, -1, "Compressed cluster "
657 "entry found in image with external data "
658 "file (L2 offset: %#" PRIx64 ", L2 index: "
659 "%#x)", l2_offset, l2_index);
660 ret = -EIO;
661 goto fail;
662 }
663 *host_offset = l2_entry & L2E_COMPRESSED_OFFSET_SIZE_MASK;
664 break;
665 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
666 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
667 break;
668 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
669 case QCOW2_SUBCLUSTER_NORMAL:
670 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC: {
671 uint64_t host_cluster_offset = l2_entry & L2E_OFFSET_MASK;
672 *host_offset = host_cluster_offset + offset_in_cluster;
673 if (offset_into_cluster(s, host_cluster_offset)) {
674 qcow2_signal_corruption(bs, true, -1, -1,
675 "Cluster allocation offset %#"
676 PRIx64 " unaligned (L2 offset: %#" PRIx64
677 ", L2 index: %#x)", host_cluster_offset,
678 l2_offset, l2_index);
679 ret = -EIO;
680 goto fail;
681 }
682 if (has_data_file(bs) && *host_offset != offset) {
683 qcow2_signal_corruption(bs, true, -1, -1,
684 "External data file host cluster offset %#"
685 PRIx64 " does not match guest cluster "
686 "offset: %#" PRIx64
687 ", L2 index: %#x)", host_cluster_offset,
688 offset - offset_in_cluster, l2_index);
689 ret = -EIO;
690 goto fail;
691 }
692 break;
693 }
694 default:
695 abort();
696 }
697
698 sc = count_contiguous_subclusters(bs, nb_clusters, sc_index,
699 l2_slice, &l2_index);
700 if (sc < 0) {
701 qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster entry found "
702 " (L2 offset: %#" PRIx64 ", L2 index: %#x)",
703 l2_offset, l2_index);
704 ret = -EIO;
705 goto fail;
706 }
707 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
708
709 bytes_available = ((int64_t)sc + sc_index) << s->subcluster_bits;
710
711 out:
712 if (bytes_available > bytes_needed) {
713 bytes_available = bytes_needed;
714 }
715
716 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
717 * subtracting offset_in_cluster will therefore definitely yield something
718 * not exceeding UINT_MAX */
719 assert(bytes_available - offset_in_cluster <= UINT_MAX);
720 *bytes = bytes_available - offset_in_cluster;
721
722 *subcluster_type = type;
723
724 return 0;
725
726 fail:
727 qcow2_cache_put(s->l2_table_cache, (void **)&l2_slice);
728 return ret;
729 }
730
731 /*
732 * get_cluster_table
733 *
734 * for a given disk offset, load (and allocate if needed)
735 * the appropriate slice of its l2 table.
736 *
737 * the cluster index in the l2 slice is given to the caller.
738 *
739 * Returns 0 on success, -errno in failure case
740 */
741 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
742 uint64_t **new_l2_slice,
743 int *new_l2_index)
744 {
745 BDRVQcow2State *s = bs->opaque;
746 unsigned int l2_index;
747 uint64_t l1_index, l2_offset;
748 uint64_t *l2_slice = NULL;
749 int ret;
750
751 /* seek to the l2 offset in the l1 table */
752
753 l1_index = offset_to_l1_index(s, offset);
754 if (l1_index >= s->l1_size) {
755 ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
756 if (ret < 0) {
757 return ret;
758 }
759 }
760
761 assert(l1_index < s->l1_size);
762 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
763 if (offset_into_cluster(s, l2_offset)) {
764 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
765 " unaligned (L1 index: %#" PRIx64 ")",
766 l2_offset, l1_index);
767 return -EIO;
768 }
769
770 if (!(s->l1_table[l1_index] & QCOW_OFLAG_COPIED)) {
771 /* First allocate a new L2 table (and do COW if needed) */
772 ret = l2_allocate(bs, l1_index);
773 if (ret < 0) {
774 return ret;
775 }
776
777 /* Then decrease the refcount of the old table */
778 if (l2_offset) {
779 qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
780 QCOW2_DISCARD_OTHER);
781 }
782
783 /* Get the offset of the newly-allocated l2 table */
784 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
785 assert(offset_into_cluster(s, l2_offset) == 0);
786 }
787
788 /* load the l2 slice in memory */
789 ret = l2_load(bs, offset, l2_offset, &l2_slice);
790 if (ret < 0) {
791 return ret;
792 }
793
794 /* find the cluster offset for the given disk offset */
795
796 l2_index = offset_to_l2_slice_index(s, offset);
797
798 *new_l2_slice = l2_slice;
799 *new_l2_index = l2_index;
800
801 return 0;
802 }
803
804 /*
805 * alloc_compressed_cluster_offset
806 *
807 * For a given offset on the virtual disk, allocate a new compressed cluster
808 * and put the host offset of the cluster into *host_offset. If a cluster is
809 * already allocated at the offset, return an error.
810 *
811 * Return 0 on success and -errno in error cases
812 */
813 int qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
814 uint64_t offset,
815 int compressed_size,
816 uint64_t *host_offset)
817 {
818 BDRVQcow2State *s = bs->opaque;
819 int l2_index, ret;
820 uint64_t *l2_slice;
821 int64_t cluster_offset;
822 int nb_csectors;
823
824 if (has_data_file(bs)) {
825 return 0;
826 }
827
828 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
829 if (ret < 0) {
830 return ret;
831 }
832
833 /* Compression can't overwrite anything. Fail if the cluster was already
834 * allocated. */
835 cluster_offset = get_l2_entry(s, l2_slice, l2_index);
836 if (cluster_offset & L2E_OFFSET_MASK) {
837 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
838 return -EIO;
839 }
840
841 cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
842 if (cluster_offset < 0) {
843 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
844 return cluster_offset;
845 }
846
847 nb_csectors =
848 (cluster_offset + compressed_size - 1) / QCOW2_COMPRESSED_SECTOR_SIZE -
849 (cluster_offset / QCOW2_COMPRESSED_SECTOR_SIZE);
850
851 /* The offset and size must fit in their fields of the L2 table entry */
852 assert((cluster_offset & s->cluster_offset_mask) == cluster_offset);
853 assert((nb_csectors & s->csize_mask) == nb_csectors);
854
855 cluster_offset |= QCOW_OFLAG_COMPRESSED |
856 ((uint64_t)nb_csectors << s->csize_shift);
857
858 /* update L2 table */
859
860 /* compressed clusters never have the copied flag */
861
862 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
863 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
864 set_l2_entry(s, l2_slice, l2_index, cluster_offset);
865 if (has_subclusters(s)) {
866 set_l2_bitmap(s, l2_slice, l2_index, 0);
867 }
868 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
869
870 *host_offset = cluster_offset & s->cluster_offset_mask;
871 return 0;
872 }
873
874 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
875 {
876 BDRVQcow2State *s = bs->opaque;
877 Qcow2COWRegion *start = &m->cow_start;
878 Qcow2COWRegion *end = &m->cow_end;
879 unsigned buffer_size;
880 unsigned data_bytes = end->offset - (start->offset + start->nb_bytes);
881 bool merge_reads;
882 uint8_t *start_buffer, *end_buffer;
883 QEMUIOVector qiov;
884 int ret;
885
886 assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
887 assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
888 assert(start->offset + start->nb_bytes <= end->offset);
889
890 if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
891 return 0;
892 }
893
894 /* If we have to read both the start and end COW regions and the
895 * middle region is not too large then perform just one read
896 * operation */
897 merge_reads = start->nb_bytes && end->nb_bytes && data_bytes <= 16384;
898 if (merge_reads) {
899 buffer_size = start->nb_bytes + data_bytes + end->nb_bytes;
900 } else {
901 /* If we have to do two reads, add some padding in the middle
902 * if necessary to make sure that the end region is optimally
903 * aligned. */
904 size_t align = bdrv_opt_mem_align(bs);
905 assert(align > 0 && align <= UINT_MAX);
906 assert(QEMU_ALIGN_UP(start->nb_bytes, align) <=
907 UINT_MAX - end->nb_bytes);
908 buffer_size = QEMU_ALIGN_UP(start->nb_bytes, align) + end->nb_bytes;
909 }
910
911 /* Reserve a buffer large enough to store all the data that we're
912 * going to read */
913 start_buffer = qemu_try_blockalign(bs, buffer_size);
914 if (start_buffer == NULL) {
915 return -ENOMEM;
916 }
917 /* The part of the buffer where the end region is located */
918 end_buffer = start_buffer + buffer_size - end->nb_bytes;
919
920 qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
921 qemu_iovec_subvec_niov(m->data_qiov,
922 m->data_qiov_offset,
923 data_bytes)
924 : 0));
925
926 qemu_co_mutex_unlock(&s->lock);
927 /* First we read the existing data from both COW regions. We
928 * either read the whole region in one go, or the start and end
929 * regions separately. */
930 if (merge_reads) {
931 qemu_iovec_add(&qiov, start_buffer, buffer_size);
932 ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
933 } else {
934 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
935 ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
936 if (ret < 0) {
937 goto fail;
938 }
939
940 qemu_iovec_reset(&qiov);
941 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
942 ret = do_perform_cow_read(bs, m->offset, end->offset, &qiov);
943 }
944 if (ret < 0) {
945 goto fail;
946 }
947
948 /* Encrypt the data if necessary before writing it */
949 if (bs->encrypted) {
950 ret = qcow2_co_encrypt(bs,
951 m->alloc_offset + start->offset,
952 m->offset + start->offset,
953 start_buffer, start->nb_bytes);
954 if (ret < 0) {
955 goto fail;
956 }
957
958 ret = qcow2_co_encrypt(bs,
959 m->alloc_offset + end->offset,
960 m->offset + end->offset,
961 end_buffer, end->nb_bytes);
962 if (ret < 0) {
963 goto fail;
964 }
965 }
966
967 /* And now we can write everything. If we have the guest data we
968 * can write everything in one single operation */
969 if (m->data_qiov) {
970 qemu_iovec_reset(&qiov);
971 if (start->nb_bytes) {
972 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
973 }
974 qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
975 if (end->nb_bytes) {
976 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
977 }
978 /* NOTE: we have a write_aio blkdebug event here followed by
979 * a cow_write one in do_perform_cow_write(), but there's only
980 * one single I/O operation */
981 BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
982 ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
983 } else {
984 /* If there's no guest data then write both COW regions separately */
985 qemu_iovec_reset(&qiov);
986 qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
987 ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
988 if (ret < 0) {
989 goto fail;
990 }
991
992 qemu_iovec_reset(&qiov);
993 qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
994 ret = do_perform_cow_write(bs, m->alloc_offset, end->offset, &qiov);
995 }
996
997 fail:
998 qemu_co_mutex_lock(&s->lock);
999
1000 /*
1001 * Before we update the L2 table to actually point to the new cluster, we
1002 * need to be sure that the refcounts have been increased and COW was
1003 * handled.
1004 */
1005 if (ret == 0) {
1006 qcow2_cache_depends_on_flush(s->l2_table_cache);
1007 }
1008
1009 qemu_vfree(start_buffer);
1010 qemu_iovec_destroy(&qiov);
1011 return ret;
1012 }
1013
1014 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
1015 {
1016 BDRVQcow2State *s = bs->opaque;
1017 int i, j = 0, l2_index, ret;
1018 uint64_t *old_cluster, *l2_slice;
1019 uint64_t cluster_offset = m->alloc_offset;
1020
1021 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
1022 assert(m->nb_clusters > 0);
1023
1024 old_cluster = g_try_new(uint64_t, m->nb_clusters);
1025 if (old_cluster == NULL) {
1026 ret = -ENOMEM;
1027 goto err;
1028 }
1029
1030 /* copy content of unmodified sectors */
1031 ret = perform_cow(bs, m);
1032 if (ret < 0) {
1033 goto err;
1034 }
1035
1036 /* Update L2 table. */
1037 if (s->use_lazy_refcounts) {
1038 qcow2_mark_dirty(bs);
1039 }
1040 if (qcow2_need_accurate_refcounts(s)) {
1041 qcow2_cache_set_dependency(bs, s->l2_table_cache,
1042 s->refcount_block_cache);
1043 }
1044
1045 ret = get_cluster_table(bs, m->offset, &l2_slice, &l2_index);
1046 if (ret < 0) {
1047 goto err;
1048 }
1049 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1050
1051 assert(l2_index + m->nb_clusters <= s->l2_slice_size);
1052 for (i = 0; i < m->nb_clusters; i++) {
1053 uint64_t offset = cluster_offset + ((uint64_t)i << s->cluster_bits);
1054 /* if two concurrent writes happen to the same unallocated cluster
1055 * each write allocates separate cluster and writes data concurrently.
1056 * The first one to complete updates l2 table with pointer to its
1057 * cluster the second one has to do RMW (which is done above by
1058 * perform_cow()), update l2 table with its cluster pointer and free
1059 * old cluster. This is what this loop does */
1060 if (get_l2_entry(s, l2_slice, l2_index + i) != 0) {
1061 old_cluster[j++] = get_l2_entry(s, l2_slice, l2_index + i);
1062 }
1063
1064 /* The offset must fit in the offset field of the L2 table entry */
1065 assert((offset & L2E_OFFSET_MASK) == offset);
1066
1067 set_l2_entry(s, l2_slice, l2_index + i, offset | QCOW_OFLAG_COPIED);
1068
1069 /* Update bitmap with the subclusters that were just written */
1070 if (has_subclusters(s) && !m->prealloc) {
1071 uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1072 unsigned written_from = m->cow_start.offset;
1073 unsigned written_to = m->cow_end.offset + m->cow_end.nb_bytes ?:
1074 m->nb_clusters << s->cluster_bits;
1075 int first_sc, last_sc;
1076 /* Narrow written_from and written_to down to the current cluster */
1077 written_from = MAX(written_from, i << s->cluster_bits);
1078 written_to = MIN(written_to, (i + 1) << s->cluster_bits);
1079 assert(written_from < written_to);
1080 first_sc = offset_to_sc_index(s, written_from);
1081 last_sc = offset_to_sc_index(s, written_to - 1);
1082 l2_bitmap |= QCOW_OFLAG_SUB_ALLOC_RANGE(first_sc, last_sc + 1);
1083 l2_bitmap &= ~QCOW_OFLAG_SUB_ZERO_RANGE(first_sc, last_sc + 1);
1084 set_l2_bitmap(s, l2_slice, l2_index + i, l2_bitmap);
1085 }
1086 }
1087
1088
1089 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1090
1091 /*
1092 * If this was a COW, we need to decrease the refcount of the old cluster.
1093 *
1094 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1095 * clusters), the next write will reuse them anyway.
1096 */
1097 if (!m->keep_old_clusters && j != 0) {
1098 for (i = 0; i < j; i++) {
1099 qcow2_free_any_clusters(bs, old_cluster[i], 1, QCOW2_DISCARD_NEVER);
1100 }
1101 }
1102
1103 ret = 0;
1104 err:
1105 g_free(old_cluster);
1106 return ret;
1107 }
1108
1109 /**
1110 * Frees the allocated clusters because the request failed and they won't
1111 * actually be linked.
1112 */
1113 void qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1114 {
1115 BDRVQcow2State *s = bs->opaque;
1116 if (!has_data_file(bs) && !m->keep_old_clusters) {
1117 qcow2_free_clusters(bs, m->alloc_offset,
1118 m->nb_clusters << s->cluster_bits,
1119 QCOW2_DISCARD_NEVER);
1120 }
1121 }
1122
1123 /*
1124 * For a given write request, create a new QCowL2Meta structure, add
1125 * it to @m and the BDRVQcow2State.cluster_allocs list. If the write
1126 * request does not need copy-on-write or changes to the L2 metadata
1127 * then this function does nothing.
1128 *
1129 * @host_cluster_offset points to the beginning of the first cluster.
1130 *
1131 * @guest_offset and @bytes indicate the offset and length of the
1132 * request.
1133 *
1134 * @l2_slice contains the L2 entries of all clusters involved in this
1135 * write request.
1136 *
1137 * If @keep_old is true it means that the clusters were already
1138 * allocated and will be overwritten. If false then the clusters are
1139 * new and we have to decrease the reference count of the old ones.
1140 *
1141 * Returns 0 on success, -errno on failure.
1142 */
1143 static int calculate_l2_meta(BlockDriverState *bs, uint64_t host_cluster_offset,
1144 uint64_t guest_offset, unsigned bytes,
1145 uint64_t *l2_slice, QCowL2Meta **m, bool keep_old)
1146 {
1147 BDRVQcow2State *s = bs->opaque;
1148 int sc_index, l2_index = offset_to_l2_slice_index(s, guest_offset);
1149 uint64_t l2_entry, l2_bitmap;
1150 unsigned cow_start_from, cow_end_to;
1151 unsigned cow_start_to = offset_into_cluster(s, guest_offset);
1152 unsigned cow_end_from = cow_start_to + bytes;
1153 unsigned nb_clusters = size_to_clusters(s, cow_end_from);
1154 QCowL2Meta *old_m = *m;
1155 QCow2SubclusterType type;
1156 int i;
1157 bool skip_cow = keep_old;
1158
1159 assert(nb_clusters <= s->l2_slice_size - l2_index);
1160
1161 /* Check the type of all affected subclusters */
1162 for (i = 0; i < nb_clusters; i++) {
1163 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1164 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1165 if (skip_cow) {
1166 unsigned write_from = MAX(cow_start_to, i << s->cluster_bits);
1167 unsigned write_to = MIN(cow_end_from, (i + 1) << s->cluster_bits);
1168 int first_sc = offset_to_sc_index(s, write_from);
1169 int last_sc = offset_to_sc_index(s, write_to - 1);
1170 int cnt = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
1171 first_sc, &type);
1172 /* Is any of the subclusters of type != QCOW2_SUBCLUSTER_NORMAL ? */
1173 if (type != QCOW2_SUBCLUSTER_NORMAL || first_sc + cnt <= last_sc) {
1174 skip_cow = false;
1175 }
1176 } else {
1177 /* If we can't skip the cow we can still look for invalid entries */
1178 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, 0);
1179 }
1180 if (type == QCOW2_SUBCLUSTER_INVALID) {
1181 int l1_index = offset_to_l1_index(s, guest_offset);
1182 uint64_t l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
1183 qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster "
1184 "entry found (L2 offset: %#" PRIx64
1185 ", L2 index: %#x)",
1186 l2_offset, l2_index + i);
1187 return -EIO;
1188 }
1189 }
1190
1191 if (skip_cow) {
1192 return 0;
1193 }
1194
1195 /* Get the L2 entry of the first cluster */
1196 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1197 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1198 sc_index = offset_to_sc_index(s, guest_offset);
1199 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1200
1201 if (!keep_old) {
1202 switch (type) {
1203 case QCOW2_SUBCLUSTER_COMPRESSED:
1204 cow_start_from = 0;
1205 break;
1206 case QCOW2_SUBCLUSTER_NORMAL:
1207 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1208 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1209 if (has_subclusters(s)) {
1210 /* Skip all leading zero and unallocated subclusters */
1211 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1212 cow_start_from =
1213 MIN(sc_index, ctz32(alloc_bitmap)) << s->subcluster_bits;
1214 } else {
1215 cow_start_from = 0;
1216 }
1217 break;
1218 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1219 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1220 cow_start_from = sc_index << s->subcluster_bits;
1221 break;
1222 default:
1223 g_assert_not_reached();
1224 }
1225 } else {
1226 switch (type) {
1227 case QCOW2_SUBCLUSTER_NORMAL:
1228 cow_start_from = cow_start_to;
1229 break;
1230 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1231 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1232 cow_start_from = sc_index << s->subcluster_bits;
1233 break;
1234 default:
1235 g_assert_not_reached();
1236 }
1237 }
1238
1239 /* Get the L2 entry of the last cluster */
1240 l2_index += nb_clusters - 1;
1241 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1242 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1243 sc_index = offset_to_sc_index(s, guest_offset + bytes - 1);
1244 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1245
1246 if (!keep_old) {
1247 switch (type) {
1248 case QCOW2_SUBCLUSTER_COMPRESSED:
1249 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1250 break;
1251 case QCOW2_SUBCLUSTER_NORMAL:
1252 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1253 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1254 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1255 if (has_subclusters(s)) {
1256 /* Skip all trailing zero and unallocated subclusters */
1257 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1258 cow_end_to -=
1259 MIN(s->subclusters_per_cluster - sc_index - 1,
1260 clz32(alloc_bitmap)) << s->subcluster_bits;
1261 }
1262 break;
1263 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1264 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1265 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1266 break;
1267 default:
1268 g_assert_not_reached();
1269 }
1270 } else {
1271 switch (type) {
1272 case QCOW2_SUBCLUSTER_NORMAL:
1273 cow_end_to = cow_end_from;
1274 break;
1275 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1276 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1277 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1278 break;
1279 default:
1280 g_assert_not_reached();
1281 }
1282 }
1283
1284 *m = g_malloc0(sizeof(**m));
1285 **m = (QCowL2Meta) {
1286 .next = old_m,
1287
1288 .alloc_offset = host_cluster_offset,
1289 .offset = start_of_cluster(s, guest_offset),
1290 .nb_clusters = nb_clusters,
1291
1292 .keep_old_clusters = keep_old,
1293
1294 .cow_start = {
1295 .offset = cow_start_from,
1296 .nb_bytes = cow_start_to - cow_start_from,
1297 },
1298 .cow_end = {
1299 .offset = cow_end_from,
1300 .nb_bytes = cow_end_to - cow_end_from,
1301 },
1302 };
1303
1304 qemu_co_queue_init(&(*m)->dependent_requests);
1305 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1306
1307 return 0;
1308 }
1309
1310 /*
1311 * Returns true if writing to the cluster pointed to by @l2_entry
1312 * requires a new allocation (that is, if the cluster is unallocated
1313 * or has refcount > 1 and therefore cannot be written in-place).
1314 */
1315 static bool cluster_needs_new_alloc(BlockDriverState *bs, uint64_t l2_entry)
1316 {
1317 switch (qcow2_get_cluster_type(bs, l2_entry)) {
1318 case QCOW2_CLUSTER_NORMAL:
1319 case QCOW2_CLUSTER_ZERO_ALLOC:
1320 if (l2_entry & QCOW_OFLAG_COPIED) {
1321 return false;
1322 }
1323 /* fallthrough */
1324 case QCOW2_CLUSTER_UNALLOCATED:
1325 case QCOW2_CLUSTER_COMPRESSED:
1326 case QCOW2_CLUSTER_ZERO_PLAIN:
1327 return true;
1328 default:
1329 abort();
1330 }
1331 }
1332
1333 /*
1334 * Returns the number of contiguous clusters that can be written to
1335 * using one single write request, starting from @l2_index.
1336 * At most @nb_clusters are checked.
1337 *
1338 * If @new_alloc is true this counts clusters that are either
1339 * unallocated, or allocated but with refcount > 1 (so they need to be
1340 * newly allocated and COWed).
1341 *
1342 * If @new_alloc is false this counts clusters that are already
1343 * allocated and can be overwritten in-place (this includes clusters
1344 * of type QCOW2_CLUSTER_ZERO_ALLOC).
1345 */
1346 static int count_single_write_clusters(BlockDriverState *bs, int nb_clusters,
1347 uint64_t *l2_slice, int l2_index,
1348 bool new_alloc)
1349 {
1350 BDRVQcow2State *s = bs->opaque;
1351 uint64_t l2_entry = get_l2_entry(s, l2_slice, l2_index);
1352 uint64_t expected_offset = l2_entry & L2E_OFFSET_MASK;
1353 int i;
1354
1355 for (i = 0; i < nb_clusters; i++) {
1356 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1357 if (cluster_needs_new_alloc(bs, l2_entry) != new_alloc) {
1358 break;
1359 }
1360 if (!new_alloc) {
1361 if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
1362 break;
1363 }
1364 expected_offset += s->cluster_size;
1365 }
1366 }
1367
1368 assert(i <= nb_clusters);
1369 return i;
1370 }
1371
1372 /*
1373 * Check if there already is an AIO write request in flight which allocates
1374 * the same cluster. In this case we need to wait until the previous
1375 * request has completed and updated the L2 table accordingly.
1376 *
1377 * Returns:
1378 * 0 if there was no dependency. *cur_bytes indicates the number of
1379 * bytes from guest_offset that can be read before the next
1380 * dependency must be processed (or the request is complete)
1381 *
1382 * -EAGAIN if we had to wait for another request, previously gathered
1383 * information on cluster allocation may be invalid now. The caller
1384 * must start over anyway, so consider *cur_bytes undefined.
1385 */
1386 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
1387 uint64_t *cur_bytes, QCowL2Meta **m)
1388 {
1389 BDRVQcow2State *s = bs->opaque;
1390 QCowL2Meta *old_alloc;
1391 uint64_t bytes = *cur_bytes;
1392
1393 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
1394
1395 uint64_t start = guest_offset;
1396 uint64_t end = start + bytes;
1397 uint64_t old_start = start_of_cluster(s, l2meta_cow_start(old_alloc));
1398 uint64_t old_end = ROUND_UP(l2meta_cow_end(old_alloc), s->cluster_size);
1399
1400 if (end <= old_start || start >= old_end) {
1401 /* No intersection */
1402 } else {
1403 if (start < old_start) {
1404 /* Stop at the start of a running allocation */
1405 bytes = old_start - start;
1406 } else {
1407 bytes = 0;
1408 }
1409
1410 /* Stop if already an l2meta exists. After yielding, it wouldn't
1411 * be valid any more, so we'd have to clean up the old L2Metas
1412 * and deal with requests depending on them before starting to
1413 * gather new ones. Not worth the trouble. */
1414 if (bytes == 0 && *m) {
1415 *cur_bytes = 0;
1416 return 0;
1417 }
1418
1419 if (bytes == 0) {
1420 /* Wait for the dependency to complete. We need to recheck
1421 * the free/allocated clusters when we continue. */
1422 qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
1423 return -EAGAIN;
1424 }
1425 }
1426 }
1427
1428 /* Make sure that existing clusters and new allocations are only used up to
1429 * the next dependency if we shortened the request above */
1430 *cur_bytes = bytes;
1431
1432 return 0;
1433 }
1434
1435 /*
1436 * Checks how many already allocated clusters that don't require a new
1437 * allocation there are at the given guest_offset (up to *bytes).
1438 * If *host_offset is not INV_OFFSET, only physically contiguous clusters
1439 * beginning at this host offset are counted.
1440 *
1441 * Note that guest_offset may not be cluster aligned. In this case, the
1442 * returned *host_offset points to exact byte referenced by guest_offset and
1443 * therefore isn't cluster aligned as well.
1444 *
1445 * Returns:
1446 * 0: if no allocated clusters are available at the given offset.
1447 * *bytes is normally unchanged. It is set to 0 if the cluster
1448 * is allocated and can be overwritten in-place but doesn't have
1449 * the right physical offset.
1450 *
1451 * 1: if allocated clusters that can be overwritten in place are
1452 * available at the requested offset. *bytes may have decreased
1453 * and describes the length of the area that can be written to.
1454 *
1455 * -errno: in error cases
1456 */
1457 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
1458 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1459 {
1460 BDRVQcow2State *s = bs->opaque;
1461 int l2_index;
1462 uint64_t l2_entry, cluster_offset;
1463 uint64_t *l2_slice;
1464 uint64_t nb_clusters;
1465 unsigned int keep_clusters;
1466 int ret;
1467
1468 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
1469 *bytes);
1470
1471 assert(*host_offset == INV_OFFSET || offset_into_cluster(s, guest_offset)
1472 == offset_into_cluster(s, *host_offset));
1473
1474 /*
1475 * Calculate the number of clusters to look for. We stop at L2 slice
1476 * boundaries to keep things simple.
1477 */
1478 nb_clusters =
1479 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1480
1481 l2_index = offset_to_l2_slice_index(s, guest_offset);
1482 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1483 /* Limit total byte count to BDRV_REQUEST_MAX_BYTES */
1484 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1485
1486 /* Find L2 entry for the first involved cluster */
1487 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1488 if (ret < 0) {
1489 return ret;
1490 }
1491
1492 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1493 cluster_offset = l2_entry & L2E_OFFSET_MASK;
1494
1495 if (!cluster_needs_new_alloc(bs, l2_entry)) {
1496 if (offset_into_cluster(s, cluster_offset)) {
1497 qcow2_signal_corruption(bs, true, -1, -1, "%s cluster offset "
1498 "%#" PRIx64 " unaligned (guest offset: %#"
1499 PRIx64 ")", l2_entry & QCOW_OFLAG_ZERO ?
1500 "Preallocated zero" : "Data",
1501 cluster_offset, guest_offset);
1502 ret = -EIO;
1503 goto out;
1504 }
1505
1506 /* If a specific host_offset is required, check it */
1507 if (*host_offset != INV_OFFSET && cluster_offset != *host_offset) {
1508 *bytes = 0;
1509 ret = 0;
1510 goto out;
1511 }
1512
1513 /* We keep all QCOW_OFLAG_COPIED clusters */
1514 keep_clusters = count_single_write_clusters(bs, nb_clusters, l2_slice,
1515 l2_index, false);
1516 assert(keep_clusters <= nb_clusters);
1517
1518 *bytes = MIN(*bytes,
1519 keep_clusters * s->cluster_size
1520 - offset_into_cluster(s, guest_offset));
1521 assert(*bytes != 0);
1522
1523 ret = calculate_l2_meta(bs, cluster_offset, guest_offset,
1524 *bytes, l2_slice, m, true);
1525 if (ret < 0) {
1526 goto out;
1527 }
1528
1529 ret = 1;
1530 } else {
1531 ret = 0;
1532 }
1533
1534 /* Cleanup */
1535 out:
1536 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1537
1538 /* Only return a host offset if we actually made progress. Otherwise we
1539 * would make requirements for handle_alloc() that it can't fulfill */
1540 if (ret > 0) {
1541 *host_offset = cluster_offset + offset_into_cluster(s, guest_offset);
1542 }
1543
1544 return ret;
1545 }
1546
1547 /*
1548 * Allocates new clusters for the given guest_offset.
1549 *
1550 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1551 * contain the number of clusters that have been allocated and are contiguous
1552 * in the image file.
1553 *
1554 * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1555 * at which the new clusters must start. *nb_clusters can be 0 on return in
1556 * this case if the cluster at host_offset is already in use. If *host_offset
1557 * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1558 *
1559 * *host_offset is updated to contain the offset into the image file at which
1560 * the first allocated cluster starts.
1561 *
1562 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1563 * function has been waiting for another request and the allocation must be
1564 * restarted, but the whole request should not be failed.
1565 */
1566 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1567 uint64_t *host_offset, uint64_t *nb_clusters)
1568 {
1569 BDRVQcow2State *s = bs->opaque;
1570
1571 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1572 *host_offset, *nb_clusters);
1573
1574 if (has_data_file(bs)) {
1575 assert(*host_offset == INV_OFFSET ||
1576 *host_offset == start_of_cluster(s, guest_offset));
1577 *host_offset = start_of_cluster(s, guest_offset);
1578 return 0;
1579 }
1580
1581 /* Allocate new clusters */
1582 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1583 if (*host_offset == INV_OFFSET) {
1584 int64_t cluster_offset =
1585 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1586 if (cluster_offset < 0) {
1587 return cluster_offset;
1588 }
1589 *host_offset = cluster_offset;
1590 return 0;
1591 } else {
1592 int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1593 if (ret < 0) {
1594 return ret;
1595 }
1596 *nb_clusters = ret;
1597 return 0;
1598 }
1599 }
1600
1601 /*
1602 * Allocates new clusters for an area that is either still unallocated or
1603 * cannot be overwritten in-place. If *host_offset is not INV_OFFSET,
1604 * clusters are only allocated if the new allocation can match the specified
1605 * host offset.
1606 *
1607 * Note that guest_offset may not be cluster aligned. In this case, the
1608 * returned *host_offset points to exact byte referenced by guest_offset and
1609 * therefore isn't cluster aligned as well.
1610 *
1611 * Returns:
1612 * 0: if no clusters could be allocated. *bytes is set to 0,
1613 * *host_offset is left unchanged.
1614 *
1615 * 1: if new clusters were allocated. *bytes may be decreased if the
1616 * new allocation doesn't cover all of the requested area.
1617 * *host_offset is updated to contain the host offset of the first
1618 * newly allocated cluster.
1619 *
1620 * -errno: in error cases
1621 */
1622 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1623 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1624 {
1625 BDRVQcow2State *s = bs->opaque;
1626 int l2_index;
1627 uint64_t *l2_slice;
1628 uint64_t nb_clusters;
1629 int ret;
1630
1631 uint64_t alloc_cluster_offset;
1632
1633 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1634 *bytes);
1635 assert(*bytes > 0);
1636
1637 /*
1638 * Calculate the number of clusters to look for. We stop at L2 slice
1639 * boundaries to keep things simple.
1640 */
1641 nb_clusters =
1642 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1643
1644 l2_index = offset_to_l2_slice_index(s, guest_offset);
1645 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1646 /* Limit total allocation byte count to BDRV_REQUEST_MAX_BYTES */
1647 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1648
1649 /* Find L2 entry for the first involved cluster */
1650 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1651 if (ret < 0) {
1652 return ret;
1653 }
1654
1655 nb_clusters = count_single_write_clusters(bs, nb_clusters,
1656 l2_slice, l2_index, true);
1657
1658 /* This function is only called when there were no non-COW clusters, so if
1659 * we can't find any unallocated or COW clusters either, something is
1660 * wrong with our code. */
1661 assert(nb_clusters > 0);
1662
1663 /* Allocate at a given offset in the image file */
1664 alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1665 start_of_cluster(s, *host_offset);
1666 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1667 &nb_clusters);
1668 if (ret < 0) {
1669 goto out;
1670 }
1671
1672 /* Can't extend contiguous allocation */
1673 if (nb_clusters == 0) {
1674 *bytes = 0;
1675 ret = 0;
1676 goto out;
1677 }
1678
1679 assert(alloc_cluster_offset != INV_OFFSET);
1680
1681 /*
1682 * Save info needed for meta data update.
1683 *
1684 * requested_bytes: Number of bytes from the start of the first
1685 * newly allocated cluster to the end of the (possibly shortened
1686 * before) write request.
1687 *
1688 * avail_bytes: Number of bytes from the start of the first
1689 * newly allocated to the end of the last newly allocated cluster.
1690 *
1691 * nb_bytes: The number of bytes from the start of the first
1692 * newly allocated cluster to the end of the area that the write
1693 * request actually writes to (excluding COW at the end)
1694 */
1695 uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1696 int avail_bytes = nb_clusters << s->cluster_bits;
1697 int nb_bytes = MIN(requested_bytes, avail_bytes);
1698
1699 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1700 *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1701 assert(*bytes != 0);
1702
1703 ret = calculate_l2_meta(bs, alloc_cluster_offset, guest_offset, *bytes,
1704 l2_slice, m, false);
1705 if (ret < 0) {
1706 goto out;
1707 }
1708
1709 ret = 1;
1710
1711 out:
1712 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1713 if (ret < 0 && *m && (*m)->nb_clusters > 0) {
1714 QLIST_REMOVE(*m, next_in_flight);
1715 }
1716 return ret;
1717 }
1718
1719 /*
1720 * alloc_cluster_offset
1721 *
1722 * For a given offset on the virtual disk, find the cluster offset in qcow2
1723 * file. If the offset is not found, allocate a new cluster.
1724 *
1725 * If the cluster was already allocated, m->nb_clusters is set to 0 and
1726 * other fields in m are meaningless.
1727 *
1728 * If the cluster is newly allocated, m->nb_clusters is set to the number of
1729 * contiguous clusters that have been allocated. In this case, the other
1730 * fields of m are valid and contain information about the first allocated
1731 * cluster.
1732 *
1733 * If the request conflicts with another write request in flight, the coroutine
1734 * is queued and will be reentered when the dependency has completed.
1735 *
1736 * Return 0 on success and -errno in error cases
1737 */
1738 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1739 unsigned int *bytes, uint64_t *host_offset,
1740 QCowL2Meta **m)
1741 {
1742 BDRVQcow2State *s = bs->opaque;
1743 uint64_t start, remaining;
1744 uint64_t cluster_offset;
1745 uint64_t cur_bytes;
1746 int ret;
1747
1748 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1749
1750 again:
1751 start = offset;
1752 remaining = *bytes;
1753 cluster_offset = INV_OFFSET;
1754 *host_offset = INV_OFFSET;
1755 cur_bytes = 0;
1756 *m = NULL;
1757
1758 while (true) {
1759
1760 if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1761 *host_offset = start_of_cluster(s, cluster_offset);
1762 }
1763
1764 assert(remaining >= cur_bytes);
1765
1766 start += cur_bytes;
1767 remaining -= cur_bytes;
1768
1769 if (cluster_offset != INV_OFFSET) {
1770 cluster_offset += cur_bytes;
1771 }
1772
1773 if (remaining == 0) {
1774 break;
1775 }
1776
1777 cur_bytes = remaining;
1778
1779 /*
1780 * Now start gathering as many contiguous clusters as possible:
1781 *
1782 * 1. Check for overlaps with in-flight allocations
1783 *
1784 * a) Overlap not in the first cluster -> shorten this request and
1785 * let the caller handle the rest in its next loop iteration.
1786 *
1787 * b) Real overlaps of two requests. Yield and restart the search
1788 * for contiguous clusters (the situation could have changed
1789 * while we were sleeping)
1790 *
1791 * c) TODO: Request starts in the same cluster as the in-flight
1792 * allocation ends. Shorten the COW of the in-fight allocation,
1793 * set cluster_offset to write to the same cluster and set up
1794 * the right synchronisation between the in-flight request and
1795 * the new one.
1796 */
1797 ret = handle_dependencies(bs, start, &cur_bytes, m);
1798 if (ret == -EAGAIN) {
1799 /* Currently handle_dependencies() doesn't yield if we already had
1800 * an allocation. If it did, we would have to clean up the L2Meta
1801 * structs before starting over. */
1802 assert(*m == NULL);
1803 goto again;
1804 } else if (ret < 0) {
1805 return ret;
1806 } else if (cur_bytes == 0) {
1807 break;
1808 } else {
1809 /* handle_dependencies() may have decreased cur_bytes (shortened
1810 * the allocations below) so that the next dependency is processed
1811 * correctly during the next loop iteration. */
1812 }
1813
1814 /*
1815 * 2. Count contiguous COPIED clusters.
1816 */
1817 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1818 if (ret < 0) {
1819 return ret;
1820 } else if (ret) {
1821 continue;
1822 } else if (cur_bytes == 0) {
1823 break;
1824 }
1825
1826 /*
1827 * 3. If the request still hasn't completed, allocate new clusters,
1828 * considering any cluster_offset of steps 1c or 2.
1829 */
1830 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1831 if (ret < 0) {
1832 return ret;
1833 } else if (ret) {
1834 continue;
1835 } else {
1836 assert(cur_bytes == 0);
1837 break;
1838 }
1839 }
1840
1841 *bytes -= remaining;
1842 assert(*bytes > 0);
1843 assert(*host_offset != INV_OFFSET);
1844
1845 return 0;
1846 }
1847
1848 /*
1849 * This discards as many clusters of nb_clusters as possible at once (i.e.
1850 * all clusters in the same L2 slice) and returns the number of discarded
1851 * clusters.
1852 */
1853 static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1854 uint64_t nb_clusters,
1855 enum qcow2_discard_type type, bool full_discard)
1856 {
1857 BDRVQcow2State *s = bs->opaque;
1858 uint64_t *l2_slice;
1859 int l2_index;
1860 int ret;
1861 int i;
1862
1863 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1864 if (ret < 0) {
1865 return ret;
1866 }
1867
1868 /* Limit nb_clusters to one L2 slice */
1869 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1870 assert(nb_clusters <= INT_MAX);
1871
1872 for (i = 0; i < nb_clusters; i++) {
1873 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1874 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1875 uint64_t new_l2_entry = old_l2_entry;
1876 uint64_t new_l2_bitmap = old_l2_bitmap;
1877 QCow2ClusterType cluster_type =
1878 qcow2_get_cluster_type(bs, old_l2_entry);
1879
1880 /*
1881 * If full_discard is true, the cluster should not read back as zeroes,
1882 * but rather fall through to the backing file.
1883 *
1884 * If full_discard is false, make sure that a discarded area reads back
1885 * as zeroes for v3 images (we cannot do it for v2 without actually
1886 * writing a zero-filled buffer). We can skip the operation if the
1887 * cluster is already marked as zero, or if it's unallocated and we
1888 * don't have a backing file.
1889 *
1890 * TODO We might want to use bdrv_block_status(bs) here, but we're
1891 * holding s->lock, so that doesn't work today.
1892 */
1893 if (full_discard) {
1894 new_l2_entry = new_l2_bitmap = 0;
1895 } else if (bs->backing || qcow2_cluster_is_allocated(cluster_type)) {
1896 if (has_subclusters(s)) {
1897 new_l2_entry = 0;
1898 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
1899 } else {
1900 new_l2_entry = s->qcow_version >= 3 ? QCOW_OFLAG_ZERO : 0;
1901 }
1902 }
1903
1904 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
1905 continue;
1906 }
1907
1908 /* First remove L2 entries */
1909 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1910 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
1911 if (has_subclusters(s)) {
1912 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
1913 }
1914 /* Then decrease the refcount */
1915 qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1916 }
1917
1918 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1919
1920 return nb_clusters;
1921 }
1922
1923 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1924 uint64_t bytes, enum qcow2_discard_type type,
1925 bool full_discard)
1926 {
1927 BDRVQcow2State *s = bs->opaque;
1928 uint64_t end_offset = offset + bytes;
1929 uint64_t nb_clusters;
1930 int64_t cleared;
1931 int ret;
1932
1933 /* Caller must pass aligned values, except at image end */
1934 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1935 assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1936 end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
1937
1938 nb_clusters = size_to_clusters(s, bytes);
1939
1940 s->cache_discards = true;
1941
1942 /* Each L2 slice is handled by its own loop iteration */
1943 while (nb_clusters > 0) {
1944 cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
1945 full_discard);
1946 if (cleared < 0) {
1947 ret = cleared;
1948 goto fail;
1949 }
1950
1951 nb_clusters -= cleared;
1952 offset += (cleared * s->cluster_size);
1953 }
1954
1955 ret = 0;
1956 fail:
1957 s->cache_discards = false;
1958 qcow2_process_discards(bs, ret);
1959
1960 return ret;
1961 }
1962
1963 /*
1964 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1965 * all clusters in the same L2 slice) and returns the number of zeroed
1966 * clusters.
1967 */
1968 static int zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1969 uint64_t nb_clusters, int flags)
1970 {
1971 BDRVQcow2State *s = bs->opaque;
1972 uint64_t *l2_slice;
1973 int l2_index;
1974 int ret;
1975 int i;
1976
1977 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1978 if (ret < 0) {
1979 return ret;
1980 }
1981
1982 /* Limit nb_clusters to one L2 slice */
1983 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1984 assert(nb_clusters <= INT_MAX);
1985
1986 for (i = 0; i < nb_clusters; i++) {
1987 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1988 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1989 QCow2ClusterType type = qcow2_get_cluster_type(bs, old_l2_entry);
1990 bool unmap = (type == QCOW2_CLUSTER_COMPRESSED) ||
1991 ((flags & BDRV_REQ_MAY_UNMAP) && qcow2_cluster_is_allocated(type));
1992 uint64_t new_l2_entry = unmap ? 0 : old_l2_entry;
1993 uint64_t new_l2_bitmap = old_l2_bitmap;
1994
1995 if (has_subclusters(s)) {
1996 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
1997 } else {
1998 new_l2_entry |= QCOW_OFLAG_ZERO;
1999 }
2000
2001 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
2002 continue;
2003 }
2004
2005 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2006 if (unmap) {
2007 qcow2_free_any_clusters(bs, old_l2_entry, 1, QCOW2_DISCARD_REQUEST);
2008 }
2009 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
2010 if (has_subclusters(s)) {
2011 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
2012 }
2013 }
2014
2015 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2016
2017 return nb_clusters;
2018 }
2019
2020 static int zero_l2_subclusters(BlockDriverState *bs, uint64_t offset,
2021 unsigned nb_subclusters)
2022 {
2023 BDRVQcow2State *s = bs->opaque;
2024 uint64_t *l2_slice;
2025 uint64_t old_l2_bitmap, l2_bitmap;
2026 int l2_index, ret, sc = offset_to_sc_index(s, offset);
2027
2028 /* For full clusters use zero_in_l2_slice() instead */
2029 assert(nb_subclusters > 0 && nb_subclusters < s->subclusters_per_cluster);
2030 assert(sc + nb_subclusters <= s->subclusters_per_cluster);
2031 assert(offset_into_subcluster(s, offset) == 0);
2032
2033 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2034 if (ret < 0) {
2035 return ret;
2036 }
2037
2038 switch (qcow2_get_cluster_type(bs, get_l2_entry(s, l2_slice, l2_index))) {
2039 case QCOW2_CLUSTER_COMPRESSED:
2040 ret = -ENOTSUP; /* We cannot partially zeroize compressed clusters */
2041 goto out;
2042 case QCOW2_CLUSTER_NORMAL:
2043 case QCOW2_CLUSTER_UNALLOCATED:
2044 break;
2045 default:
2046 g_assert_not_reached();
2047 }
2048
2049 old_l2_bitmap = l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
2050
2051 l2_bitmap |= QCOW_OFLAG_SUB_ZERO_RANGE(sc, sc + nb_subclusters);
2052 l2_bitmap &= ~QCOW_OFLAG_SUB_ALLOC_RANGE(sc, sc + nb_subclusters);
2053
2054 if (old_l2_bitmap != l2_bitmap) {
2055 set_l2_bitmap(s, l2_slice, l2_index, l2_bitmap);
2056 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2057 }
2058
2059 ret = 0;
2060 out:
2061 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2062
2063 return ret;
2064 }
2065
2066 int qcow2_subcluster_zeroize(BlockDriverState *bs, uint64_t offset,
2067 uint64_t bytes, int flags)
2068 {
2069 BDRVQcow2State *s = bs->opaque;
2070 uint64_t end_offset = offset + bytes;
2071 uint64_t nb_clusters;
2072 unsigned head, tail;
2073 int64_t cleared;
2074 int ret;
2075
2076 /* If we have to stay in sync with an external data file, zero out
2077 * s->data_file first. */
2078 if (data_file_is_raw(bs)) {
2079 assert(has_data_file(bs));
2080 ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
2081 if (ret < 0) {
2082 return ret;
2083 }
2084 }
2085
2086 /* Caller must pass aligned values, except at image end */
2087 assert(offset_into_subcluster(s, offset) == 0);
2088 assert(offset_into_subcluster(s, end_offset) == 0 ||
2089 end_offset >= bs->total_sectors << BDRV_SECTOR_BITS);
2090
2091 /*
2092 * The zero flag is only supported by version 3 and newer. However, if we
2093 * have no backing file, we can resort to discard in version 2.
2094 */
2095 if (s->qcow_version < 3) {
2096 if (!bs->backing) {
2097 return qcow2_cluster_discard(bs, offset, bytes,
2098 QCOW2_DISCARD_REQUEST, false);
2099 }
2100 return -ENOTSUP;
2101 }
2102
2103 head = MIN(end_offset, ROUND_UP(offset, s->cluster_size)) - offset;
2104 offset += head;
2105
2106 tail = (end_offset >= bs->total_sectors << BDRV_SECTOR_BITS) ? 0 :
2107 end_offset - MAX(offset, start_of_cluster(s, end_offset));
2108 end_offset -= tail;
2109
2110 s->cache_discards = true;
2111
2112 if (head) {
2113 ret = zero_l2_subclusters(bs, offset - head,
2114 size_to_subclusters(s, head));
2115 if (ret < 0) {
2116 goto fail;
2117 }
2118 }
2119
2120 /* Each L2 slice is handled by its own loop iteration */
2121 nb_clusters = size_to_clusters(s, end_offset - offset);
2122
2123 while (nb_clusters > 0) {
2124 cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
2125 if (cleared < 0) {
2126 ret = cleared;
2127 goto fail;
2128 }
2129
2130 nb_clusters -= cleared;
2131 offset += (cleared * s->cluster_size);
2132 }
2133
2134 if (tail) {
2135 ret = zero_l2_subclusters(bs, end_offset, size_to_subclusters(s, tail));
2136 if (ret < 0) {
2137 goto fail;
2138 }
2139 }
2140
2141 ret = 0;
2142 fail:
2143 s->cache_discards = false;
2144 qcow2_process_discards(bs, ret);
2145
2146 return ret;
2147 }
2148
2149 /*
2150 * Expands all zero clusters in a specific L1 table (or deallocates them, for
2151 * non-backed non-pre-allocated zero clusters).
2152 *
2153 * l1_entries and *visited_l1_entries are used to keep track of progress for
2154 * status_cb(). l1_entries contains the total number of L1 entries and
2155 * *visited_l1_entries counts all visited L1 entries.
2156 */
2157 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
2158 int l1_size, int64_t *visited_l1_entries,
2159 int64_t l1_entries,
2160 BlockDriverAmendStatusCB *status_cb,
2161 void *cb_opaque)
2162 {
2163 BDRVQcow2State *s = bs->opaque;
2164 bool is_active_l1 = (l1_table == s->l1_table);
2165 uint64_t *l2_slice = NULL;
2166 unsigned slice, slice_size2, n_slices;
2167 int ret;
2168 int i, j;
2169
2170 /* qcow2_downgrade() is not allowed in images with subclusters */
2171 assert(!has_subclusters(s));
2172
2173 slice_size2 = s->l2_slice_size * l2_entry_size(s);
2174 n_slices = s->cluster_size / slice_size2;
2175
2176 if (!is_active_l1) {
2177 /* inactive L2 tables require a buffer to be stored in when loading
2178 * them from disk */
2179 l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
2180 if (l2_slice == NULL) {
2181 return -ENOMEM;
2182 }
2183 }
2184
2185 for (i = 0; i < l1_size; i++) {
2186 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
2187 uint64_t l2_refcount;
2188
2189 if (!l2_offset) {
2190 /* unallocated */
2191 (*visited_l1_entries)++;
2192 if (status_cb) {
2193 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2194 }
2195 continue;
2196 }
2197
2198 if (offset_into_cluster(s, l2_offset)) {
2199 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
2200 PRIx64 " unaligned (L1 index: %#x)",
2201 l2_offset, i);
2202 ret = -EIO;
2203 goto fail;
2204 }
2205
2206 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
2207 &l2_refcount);
2208 if (ret < 0) {
2209 goto fail;
2210 }
2211
2212 for (slice = 0; slice < n_slices; slice++) {
2213 uint64_t slice_offset = l2_offset + slice * slice_size2;
2214 bool l2_dirty = false;
2215 if (is_active_l1) {
2216 /* get active L2 tables from cache */
2217 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
2218 (void **)&l2_slice);
2219 } else {
2220 /* load inactive L2 tables from disk */
2221 ret = bdrv_pread(bs->file, slice_offset, l2_slice, slice_size2);
2222 }
2223 if (ret < 0) {
2224 goto fail;
2225 }
2226
2227 for (j = 0; j < s->l2_slice_size; j++) {
2228 uint64_t l2_entry = get_l2_entry(s, l2_slice, j);
2229 int64_t offset = l2_entry & L2E_OFFSET_MASK;
2230 QCow2ClusterType cluster_type =
2231 qcow2_get_cluster_type(bs, l2_entry);
2232
2233 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
2234 cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
2235 continue;
2236 }
2237
2238 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2239 if (!bs->backing) {
2240 /*
2241 * not backed; therefore we can simply deallocate the
2242 * cluster. No need to call set_l2_bitmap(), this
2243 * function doesn't support images with subclusters.
2244 */
2245 set_l2_entry(s, l2_slice, j, 0);
2246 l2_dirty = true;
2247 continue;
2248 }
2249
2250 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2251 if (offset < 0) {
2252 ret = offset;
2253 goto fail;
2254 }
2255
2256 /* The offset must fit in the offset field */
2257 assert((offset & L2E_OFFSET_MASK) == offset);
2258
2259 if (l2_refcount > 1) {
2260 /* For shared L2 tables, set the refcount accordingly
2261 * (it is already 1 and needs to be l2_refcount) */
2262 ret = qcow2_update_cluster_refcount(
2263 bs, offset >> s->cluster_bits,
2264 refcount_diff(1, l2_refcount), false,
2265 QCOW2_DISCARD_OTHER);
2266 if (ret < 0) {
2267 qcow2_free_clusters(bs, offset, s->cluster_size,
2268 QCOW2_DISCARD_OTHER);
2269 goto fail;
2270 }
2271 }
2272 }
2273
2274 if (offset_into_cluster(s, offset)) {
2275 int l2_index = slice * s->l2_slice_size + j;
2276 qcow2_signal_corruption(
2277 bs, true, -1, -1,
2278 "Cluster allocation offset "
2279 "%#" PRIx64 " unaligned (L2 offset: %#"
2280 PRIx64 ", L2 index: %#x)", offset,
2281 l2_offset, l2_index);
2282 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2283 qcow2_free_clusters(bs, offset, s->cluster_size,
2284 QCOW2_DISCARD_ALWAYS);
2285 }
2286 ret = -EIO;
2287 goto fail;
2288 }
2289
2290 ret = qcow2_pre_write_overlap_check(bs, 0, offset,
2291 s->cluster_size, true);
2292 if (ret < 0) {
2293 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2294 qcow2_free_clusters(bs, offset, s->cluster_size,
2295 QCOW2_DISCARD_ALWAYS);
2296 }
2297 goto fail;
2298 }
2299
2300 ret = bdrv_pwrite_zeroes(s->data_file, offset,
2301 s->cluster_size, 0);
2302 if (ret < 0) {
2303 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2304 qcow2_free_clusters(bs, offset, s->cluster_size,
2305 QCOW2_DISCARD_ALWAYS);
2306 }
2307 goto fail;
2308 }
2309
2310 if (l2_refcount == 1) {
2311 set_l2_entry(s, l2_slice, j, offset | QCOW_OFLAG_COPIED);
2312 } else {
2313 set_l2_entry(s, l2_slice, j, offset);
2314 }
2315 /*
2316 * No need to call set_l2_bitmap() after set_l2_entry() because
2317 * this function doesn't support images with subclusters.
2318 */
2319 l2_dirty = true;
2320 }
2321
2322 if (is_active_l1) {
2323 if (l2_dirty) {
2324 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2325 qcow2_cache_depends_on_flush(s->l2_table_cache);
2326 }
2327 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2328 } else {
2329 if (l2_dirty) {
2330 ret = qcow2_pre_write_overlap_check(
2331 bs, QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2,
2332 slice_offset, slice_size2, false);
2333 if (ret < 0) {
2334 goto fail;
2335 }
2336
2337 ret = bdrv_pwrite(bs->file, slice_offset,
2338 l2_slice, slice_size2);
2339 if (ret < 0) {
2340 goto fail;
2341 }
2342 }
2343 }
2344 }
2345
2346 (*visited_l1_entries)++;
2347 if (status_cb) {
2348 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2349 }
2350 }
2351
2352 ret = 0;
2353
2354 fail:
2355 if (l2_slice) {
2356 if (!is_active_l1) {
2357 qemu_vfree(l2_slice);
2358 } else {
2359 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2360 }
2361 }
2362 return ret;
2363 }
2364
2365 /*
2366 * For backed images, expands all zero clusters on the image. For non-backed
2367 * images, deallocates all non-pre-allocated zero clusters (and claims the
2368 * allocation for pre-allocated ones). This is important for downgrading to a
2369 * qcow2 version which doesn't yet support metadata zero clusters.
2370 */
2371 int qcow2_expand_zero_clusters(BlockDriverState *bs,
2372 BlockDriverAmendStatusCB *status_cb,
2373 void *cb_opaque)
2374 {
2375 BDRVQcow2State *s = bs->opaque;
2376 uint64_t *l1_table = NULL;
2377 int64_t l1_entries = 0, visited_l1_entries = 0;
2378 int ret;
2379 int i, j;
2380
2381 if (status_cb) {
2382 l1_entries = s->l1_size;
2383 for (i = 0; i < s->nb_snapshots; i++) {
2384 l1_entries += s->snapshots[i].l1_size;
2385 }
2386 }
2387
2388 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
2389 &visited_l1_entries, l1_entries,
2390 status_cb, cb_opaque);
2391 if (ret < 0) {
2392 goto fail;
2393 }
2394
2395 /* Inactive L1 tables may point to active L2 tables - therefore it is
2396 * necessary to flush the L2 table cache before trying to access the L2
2397 * tables pointed to by inactive L1 entries (else we might try to expand
2398 * zero clusters that have already been expanded); furthermore, it is also
2399 * necessary to empty the L2 table cache, since it may contain tables which
2400 * are now going to be modified directly on disk, bypassing the cache.
2401 * qcow2_cache_empty() does both for us. */
2402 ret = qcow2_cache_empty(bs, s->l2_table_cache);
2403 if (ret < 0) {
2404 goto fail;
2405 }
2406
2407 for (i = 0; i < s->nb_snapshots; i++) {
2408 int l1_size2;
2409 uint64_t *new_l1_table;
2410 Error *local_err = NULL;
2411
2412 ret = qcow2_validate_table(bs, s->snapshots[i].l1_table_offset,
2413 s->snapshots[i].l1_size, L1E_SIZE,
2414 QCOW_MAX_L1_SIZE, "Snapshot L1 table",
2415 &local_err);
2416 if (ret < 0) {
2417 error_report_err(local_err);
2418 goto fail;
2419 }
2420
2421 l1_size2 = s->snapshots[i].l1_size * L1E_SIZE;
2422 new_l1_table = g_try_realloc(l1_table, l1_size2);
2423
2424 if (!new_l1_table) {
2425 ret = -ENOMEM;
2426 goto fail;
2427 }
2428
2429 l1_table = new_l1_table;
2430
2431 ret = bdrv_pread(bs->file, s->snapshots[i].l1_table_offset,
2432 l1_table, l1_size2);
2433 if (ret < 0) {
2434 goto fail;
2435 }
2436
2437 for (j = 0; j < s->snapshots[i].l1_size; j++) {
2438 be64_to_cpus(&l1_table[j]);
2439 }
2440
2441 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
2442 &visited_l1_entries, l1_entries,
2443 status_cb, cb_opaque);
2444 if (ret < 0) {
2445 goto fail;
2446 }
2447 }
2448
2449 ret = 0;
2450
2451 fail:
2452 g_free(l1_table);
2453 return ret;
2454 }