qcow2: Fix corruption on write_zeroes with MAY_UNMAP
[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 + L1E_SIZE * 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 + L1E_SIZE * 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 assert(m->cow_end.offset + m->cow_end.nb_bytes <=
1053 m->nb_clusters << s->cluster_bits);
1054 for (i = 0; i < m->nb_clusters; i++) {
1055 uint64_t offset = cluster_offset + ((uint64_t)i << s->cluster_bits);
1056 /* if two concurrent writes happen to the same unallocated cluster
1057 * each write allocates separate cluster and writes data concurrently.
1058 * The first one to complete updates l2 table with pointer to its
1059 * cluster the second one has to do RMW (which is done above by
1060 * perform_cow()), update l2 table with its cluster pointer and free
1061 * old cluster. This is what this loop does */
1062 if (get_l2_entry(s, l2_slice, l2_index + i) != 0) {
1063 old_cluster[j++] = get_l2_entry(s, l2_slice, l2_index + i);
1064 }
1065
1066 /* The offset must fit in the offset field of the L2 table entry */
1067 assert((offset & L2E_OFFSET_MASK) == offset);
1068
1069 set_l2_entry(s, l2_slice, l2_index + i, offset | QCOW_OFLAG_COPIED);
1070
1071 /* Update bitmap with the subclusters that were just written */
1072 if (has_subclusters(s) && !m->prealloc) {
1073 uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1074 unsigned written_from = m->cow_start.offset;
1075 unsigned written_to = m->cow_end.offset + m->cow_end.nb_bytes;
1076 int first_sc, last_sc;
1077 /* Narrow written_from and written_to down to the current cluster */
1078 written_from = MAX(written_from, i << s->cluster_bits);
1079 written_to = MIN(written_to, (i + 1) << s->cluster_bits);
1080 assert(written_from < written_to);
1081 first_sc = offset_to_sc_index(s, written_from);
1082 last_sc = offset_to_sc_index(s, written_to - 1);
1083 l2_bitmap |= QCOW_OFLAG_SUB_ALLOC_RANGE(first_sc, last_sc + 1);
1084 l2_bitmap &= ~QCOW_OFLAG_SUB_ZERO_RANGE(first_sc, last_sc + 1);
1085 set_l2_bitmap(s, l2_slice, l2_index + i, l2_bitmap);
1086 }
1087 }
1088
1089
1090 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1091
1092 /*
1093 * If this was a COW, we need to decrease the refcount of the old cluster.
1094 *
1095 * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1096 * clusters), the next write will reuse them anyway.
1097 */
1098 if (!m->keep_old_clusters && j != 0) {
1099 for (i = 0; i < j; i++) {
1100 qcow2_free_any_cluster(bs, old_cluster[i], QCOW2_DISCARD_NEVER);
1101 }
1102 }
1103
1104 ret = 0;
1105 err:
1106 g_free(old_cluster);
1107 return ret;
1108 }
1109
1110 /**
1111 * Frees the allocated clusters because the request failed and they won't
1112 * actually be linked.
1113 */
1114 void qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1115 {
1116 BDRVQcow2State *s = bs->opaque;
1117 if (!has_data_file(bs) && !m->keep_old_clusters) {
1118 qcow2_free_clusters(bs, m->alloc_offset,
1119 m->nb_clusters << s->cluster_bits,
1120 QCOW2_DISCARD_NEVER);
1121 }
1122 }
1123
1124 /*
1125 * For a given write request, create a new QCowL2Meta structure, add
1126 * it to @m and the BDRVQcow2State.cluster_allocs list. If the write
1127 * request does not need copy-on-write or changes to the L2 metadata
1128 * then this function does nothing.
1129 *
1130 * @host_cluster_offset points to the beginning of the first cluster.
1131 *
1132 * @guest_offset and @bytes indicate the offset and length of the
1133 * request.
1134 *
1135 * @l2_slice contains the L2 entries of all clusters involved in this
1136 * write request.
1137 *
1138 * If @keep_old is true it means that the clusters were already
1139 * allocated and will be overwritten. If false then the clusters are
1140 * new and we have to decrease the reference count of the old ones.
1141 *
1142 * Returns 0 on success, -errno on failure.
1143 */
1144 static int calculate_l2_meta(BlockDriverState *bs, uint64_t host_cluster_offset,
1145 uint64_t guest_offset, unsigned bytes,
1146 uint64_t *l2_slice, QCowL2Meta **m, bool keep_old)
1147 {
1148 BDRVQcow2State *s = bs->opaque;
1149 int sc_index, l2_index = offset_to_l2_slice_index(s, guest_offset);
1150 uint64_t l2_entry, l2_bitmap;
1151 unsigned cow_start_from, cow_end_to;
1152 unsigned cow_start_to = offset_into_cluster(s, guest_offset);
1153 unsigned cow_end_from = cow_start_to + bytes;
1154 unsigned nb_clusters = size_to_clusters(s, cow_end_from);
1155 QCowL2Meta *old_m = *m;
1156 QCow2SubclusterType type;
1157 int i;
1158 bool skip_cow = keep_old;
1159
1160 assert(nb_clusters <= s->l2_slice_size - l2_index);
1161
1162 /* Check the type of all affected subclusters */
1163 for (i = 0; i < nb_clusters; i++) {
1164 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1165 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1166 if (skip_cow) {
1167 unsigned write_from = MAX(cow_start_to, i << s->cluster_bits);
1168 unsigned write_to = MIN(cow_end_from, (i + 1) << s->cluster_bits);
1169 int first_sc = offset_to_sc_index(s, write_from);
1170 int last_sc = offset_to_sc_index(s, write_to - 1);
1171 int cnt = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
1172 first_sc, &type);
1173 /* Is any of the subclusters of type != QCOW2_SUBCLUSTER_NORMAL ? */
1174 if (type != QCOW2_SUBCLUSTER_NORMAL || first_sc + cnt <= last_sc) {
1175 skip_cow = false;
1176 }
1177 } else {
1178 /* If we can't skip the cow we can still look for invalid entries */
1179 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, 0);
1180 }
1181 if (type == QCOW2_SUBCLUSTER_INVALID) {
1182 int l1_index = offset_to_l1_index(s, guest_offset);
1183 uint64_t l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
1184 qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster "
1185 "entry found (L2 offset: %#" PRIx64
1186 ", L2 index: %#x)",
1187 l2_offset, l2_index + i);
1188 return -EIO;
1189 }
1190 }
1191
1192 if (skip_cow) {
1193 return 0;
1194 }
1195
1196 /* Get the L2 entry of the first cluster */
1197 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1198 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1199 sc_index = offset_to_sc_index(s, guest_offset);
1200 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1201
1202 if (!keep_old) {
1203 switch (type) {
1204 case QCOW2_SUBCLUSTER_COMPRESSED:
1205 cow_start_from = 0;
1206 break;
1207 case QCOW2_SUBCLUSTER_NORMAL:
1208 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1209 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1210 if (has_subclusters(s)) {
1211 /* Skip all leading zero and unallocated subclusters */
1212 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1213 cow_start_from =
1214 MIN(sc_index, ctz32(alloc_bitmap)) << s->subcluster_bits;
1215 } else {
1216 cow_start_from = 0;
1217 }
1218 break;
1219 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1220 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1221 cow_start_from = sc_index << s->subcluster_bits;
1222 break;
1223 default:
1224 g_assert_not_reached();
1225 }
1226 } else {
1227 switch (type) {
1228 case QCOW2_SUBCLUSTER_NORMAL:
1229 cow_start_from = cow_start_to;
1230 break;
1231 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1232 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1233 cow_start_from = sc_index << s->subcluster_bits;
1234 break;
1235 default:
1236 g_assert_not_reached();
1237 }
1238 }
1239
1240 /* Get the L2 entry of the last cluster */
1241 l2_index += nb_clusters - 1;
1242 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1243 l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1244 sc_index = offset_to_sc_index(s, guest_offset + bytes - 1);
1245 type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1246
1247 if (!keep_old) {
1248 switch (type) {
1249 case QCOW2_SUBCLUSTER_COMPRESSED:
1250 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1251 break;
1252 case QCOW2_SUBCLUSTER_NORMAL:
1253 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1254 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1255 cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1256 if (has_subclusters(s)) {
1257 /* Skip all trailing zero and unallocated subclusters */
1258 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1259 cow_end_to -=
1260 MIN(s->subclusters_per_cluster - sc_index - 1,
1261 clz32(alloc_bitmap)) << s->subcluster_bits;
1262 }
1263 break;
1264 case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1265 case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1266 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1267 break;
1268 default:
1269 g_assert_not_reached();
1270 }
1271 } else {
1272 switch (type) {
1273 case QCOW2_SUBCLUSTER_NORMAL:
1274 cow_end_to = cow_end_from;
1275 break;
1276 case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1277 case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1278 cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1279 break;
1280 default:
1281 g_assert_not_reached();
1282 }
1283 }
1284
1285 *m = g_malloc0(sizeof(**m));
1286 **m = (QCowL2Meta) {
1287 .next = old_m,
1288
1289 .alloc_offset = host_cluster_offset,
1290 .offset = start_of_cluster(s, guest_offset),
1291 .nb_clusters = nb_clusters,
1292
1293 .keep_old_clusters = keep_old,
1294
1295 .cow_start = {
1296 .offset = cow_start_from,
1297 .nb_bytes = cow_start_to - cow_start_from,
1298 },
1299 .cow_end = {
1300 .offset = cow_end_from,
1301 .nb_bytes = cow_end_to - cow_end_from,
1302 },
1303 };
1304
1305 qemu_co_queue_init(&(*m)->dependent_requests);
1306 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1307
1308 return 0;
1309 }
1310
1311 /*
1312 * Returns true if writing to the cluster pointed to by @l2_entry
1313 * requires a new allocation (that is, if the cluster is unallocated
1314 * or has refcount > 1 and therefore cannot be written in-place).
1315 */
1316 static bool cluster_needs_new_alloc(BlockDriverState *bs, uint64_t l2_entry)
1317 {
1318 switch (qcow2_get_cluster_type(bs, l2_entry)) {
1319 case QCOW2_CLUSTER_NORMAL:
1320 case QCOW2_CLUSTER_ZERO_ALLOC:
1321 if (l2_entry & QCOW_OFLAG_COPIED) {
1322 return false;
1323 }
1324 /* fallthrough */
1325 case QCOW2_CLUSTER_UNALLOCATED:
1326 case QCOW2_CLUSTER_COMPRESSED:
1327 case QCOW2_CLUSTER_ZERO_PLAIN:
1328 return true;
1329 default:
1330 abort();
1331 }
1332 }
1333
1334 /*
1335 * Returns the number of contiguous clusters that can be written to
1336 * using one single write request, starting from @l2_index.
1337 * At most @nb_clusters are checked.
1338 *
1339 * If @new_alloc is true this counts clusters that are either
1340 * unallocated, or allocated but with refcount > 1 (so they need to be
1341 * newly allocated and COWed).
1342 *
1343 * If @new_alloc is false this counts clusters that are already
1344 * allocated and can be overwritten in-place (this includes clusters
1345 * of type QCOW2_CLUSTER_ZERO_ALLOC).
1346 */
1347 static int count_single_write_clusters(BlockDriverState *bs, int nb_clusters,
1348 uint64_t *l2_slice, int l2_index,
1349 bool new_alloc)
1350 {
1351 BDRVQcow2State *s = bs->opaque;
1352 uint64_t l2_entry = get_l2_entry(s, l2_slice, l2_index);
1353 uint64_t expected_offset = l2_entry & L2E_OFFSET_MASK;
1354 int i;
1355
1356 for (i = 0; i < nb_clusters; i++) {
1357 l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1358 if (cluster_needs_new_alloc(bs, l2_entry) != new_alloc) {
1359 break;
1360 }
1361 if (!new_alloc) {
1362 if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
1363 break;
1364 }
1365 expected_offset += s->cluster_size;
1366 }
1367 }
1368
1369 assert(i <= nb_clusters);
1370 return i;
1371 }
1372
1373 /*
1374 * Check if there already is an AIO write request in flight which allocates
1375 * the same cluster. In this case we need to wait until the previous
1376 * request has completed and updated the L2 table accordingly.
1377 *
1378 * Returns:
1379 * 0 if there was no dependency. *cur_bytes indicates the number of
1380 * bytes from guest_offset that can be read before the next
1381 * dependency must be processed (or the request is complete)
1382 *
1383 * -EAGAIN if we had to wait for another request, previously gathered
1384 * information on cluster allocation may be invalid now. The caller
1385 * must start over anyway, so consider *cur_bytes undefined.
1386 */
1387 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
1388 uint64_t *cur_bytes, QCowL2Meta **m)
1389 {
1390 BDRVQcow2State *s = bs->opaque;
1391 QCowL2Meta *old_alloc;
1392 uint64_t bytes = *cur_bytes;
1393
1394 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
1395
1396 uint64_t start = guest_offset;
1397 uint64_t end = start + bytes;
1398 uint64_t old_start = start_of_cluster(s, l2meta_cow_start(old_alloc));
1399 uint64_t old_end = ROUND_UP(l2meta_cow_end(old_alloc), s->cluster_size);
1400
1401 if (end <= old_start || start >= old_end) {
1402 /* No intersection */
1403 } else {
1404 if (start < old_start) {
1405 /* Stop at the start of a running allocation */
1406 bytes = old_start - start;
1407 } else {
1408 bytes = 0;
1409 }
1410
1411 /* Stop if already an l2meta exists. After yielding, it wouldn't
1412 * be valid any more, so we'd have to clean up the old L2Metas
1413 * and deal with requests depending on them before starting to
1414 * gather new ones. Not worth the trouble. */
1415 if (bytes == 0 && *m) {
1416 *cur_bytes = 0;
1417 return 0;
1418 }
1419
1420 if (bytes == 0) {
1421 /* Wait for the dependency to complete. We need to recheck
1422 * the free/allocated clusters when we continue. */
1423 qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
1424 return -EAGAIN;
1425 }
1426 }
1427 }
1428
1429 /* Make sure that existing clusters and new allocations are only used up to
1430 * the next dependency if we shortened the request above */
1431 *cur_bytes = bytes;
1432
1433 return 0;
1434 }
1435
1436 /*
1437 * Checks how many already allocated clusters that don't require a new
1438 * allocation there are at the given guest_offset (up to *bytes).
1439 * If *host_offset is not INV_OFFSET, only physically contiguous clusters
1440 * beginning at this host offset are counted.
1441 *
1442 * Note that guest_offset may not be cluster aligned. In this case, the
1443 * returned *host_offset points to exact byte referenced by guest_offset and
1444 * therefore isn't cluster aligned as well.
1445 *
1446 * Returns:
1447 * 0: if no allocated clusters are available at the given offset.
1448 * *bytes is normally unchanged. It is set to 0 if the cluster
1449 * is allocated and can be overwritten in-place but doesn't have
1450 * the right physical offset.
1451 *
1452 * 1: if allocated clusters that can be overwritten in place are
1453 * available at the requested offset. *bytes may have decreased
1454 * and describes the length of the area that can be written to.
1455 *
1456 * -errno: in error cases
1457 */
1458 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
1459 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1460 {
1461 BDRVQcow2State *s = bs->opaque;
1462 int l2_index;
1463 uint64_t l2_entry, cluster_offset;
1464 uint64_t *l2_slice;
1465 uint64_t nb_clusters;
1466 unsigned int keep_clusters;
1467 int ret;
1468
1469 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
1470 *bytes);
1471
1472 assert(*host_offset == INV_OFFSET || offset_into_cluster(s, guest_offset)
1473 == offset_into_cluster(s, *host_offset));
1474
1475 /*
1476 * Calculate the number of clusters to look for. We stop at L2 slice
1477 * boundaries to keep things simple.
1478 */
1479 nb_clusters =
1480 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1481
1482 l2_index = offset_to_l2_slice_index(s, guest_offset);
1483 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1484 /* Limit total byte count to BDRV_REQUEST_MAX_BYTES */
1485 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1486
1487 /* Find L2 entry for the first involved cluster */
1488 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1489 if (ret < 0) {
1490 return ret;
1491 }
1492
1493 l2_entry = get_l2_entry(s, l2_slice, l2_index);
1494 cluster_offset = l2_entry & L2E_OFFSET_MASK;
1495
1496 if (!cluster_needs_new_alloc(bs, l2_entry)) {
1497 if (offset_into_cluster(s, cluster_offset)) {
1498 qcow2_signal_corruption(bs, true, -1, -1, "%s cluster offset "
1499 "%#" PRIx64 " unaligned (guest offset: %#"
1500 PRIx64 ")", l2_entry & QCOW_OFLAG_ZERO ?
1501 "Preallocated zero" : "Data",
1502 cluster_offset, guest_offset);
1503 ret = -EIO;
1504 goto out;
1505 }
1506
1507 /* If a specific host_offset is required, check it */
1508 if (*host_offset != INV_OFFSET && cluster_offset != *host_offset) {
1509 *bytes = 0;
1510 ret = 0;
1511 goto out;
1512 }
1513
1514 /* We keep all QCOW_OFLAG_COPIED clusters */
1515 keep_clusters = count_single_write_clusters(bs, nb_clusters, l2_slice,
1516 l2_index, false);
1517 assert(keep_clusters <= nb_clusters);
1518
1519 *bytes = MIN(*bytes,
1520 keep_clusters * s->cluster_size
1521 - offset_into_cluster(s, guest_offset));
1522 assert(*bytes != 0);
1523
1524 ret = calculate_l2_meta(bs, cluster_offset, guest_offset,
1525 *bytes, l2_slice, m, true);
1526 if (ret < 0) {
1527 goto out;
1528 }
1529
1530 ret = 1;
1531 } else {
1532 ret = 0;
1533 }
1534
1535 /* Cleanup */
1536 out:
1537 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1538
1539 /* Only return a host offset if we actually made progress. Otherwise we
1540 * would make requirements for handle_alloc() that it can't fulfill */
1541 if (ret > 0) {
1542 *host_offset = cluster_offset + offset_into_cluster(s, guest_offset);
1543 }
1544
1545 return ret;
1546 }
1547
1548 /*
1549 * Allocates new clusters for the given guest_offset.
1550 *
1551 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1552 * contain the number of clusters that have been allocated and are contiguous
1553 * in the image file.
1554 *
1555 * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1556 * at which the new clusters must start. *nb_clusters can be 0 on return in
1557 * this case if the cluster at host_offset is already in use. If *host_offset
1558 * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1559 *
1560 * *host_offset is updated to contain the offset into the image file at which
1561 * the first allocated cluster starts.
1562 *
1563 * Return 0 on success and -errno in error cases. -EAGAIN means that the
1564 * function has been waiting for another request and the allocation must be
1565 * restarted, but the whole request should not be failed.
1566 */
1567 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1568 uint64_t *host_offset, uint64_t *nb_clusters)
1569 {
1570 BDRVQcow2State *s = bs->opaque;
1571
1572 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1573 *host_offset, *nb_clusters);
1574
1575 if (has_data_file(bs)) {
1576 assert(*host_offset == INV_OFFSET ||
1577 *host_offset == start_of_cluster(s, guest_offset));
1578 *host_offset = start_of_cluster(s, guest_offset);
1579 return 0;
1580 }
1581
1582 /* Allocate new clusters */
1583 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1584 if (*host_offset == INV_OFFSET) {
1585 int64_t cluster_offset =
1586 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1587 if (cluster_offset < 0) {
1588 return cluster_offset;
1589 }
1590 *host_offset = cluster_offset;
1591 return 0;
1592 } else {
1593 int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1594 if (ret < 0) {
1595 return ret;
1596 }
1597 *nb_clusters = ret;
1598 return 0;
1599 }
1600 }
1601
1602 /*
1603 * Allocates new clusters for an area that is either still unallocated or
1604 * cannot be overwritten in-place. If *host_offset is not INV_OFFSET,
1605 * clusters are only allocated if the new allocation can match the specified
1606 * host offset.
1607 *
1608 * Note that guest_offset may not be cluster aligned. In this case, the
1609 * returned *host_offset points to exact byte referenced by guest_offset and
1610 * therefore isn't cluster aligned as well.
1611 *
1612 * Returns:
1613 * 0: if no clusters could be allocated. *bytes is set to 0,
1614 * *host_offset is left unchanged.
1615 *
1616 * 1: if new clusters were allocated. *bytes may be decreased if the
1617 * new allocation doesn't cover all of the requested area.
1618 * *host_offset is updated to contain the host offset of the first
1619 * newly allocated cluster.
1620 *
1621 * -errno: in error cases
1622 */
1623 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1624 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1625 {
1626 BDRVQcow2State *s = bs->opaque;
1627 int l2_index;
1628 uint64_t *l2_slice;
1629 uint64_t nb_clusters;
1630 int ret;
1631
1632 uint64_t alloc_cluster_offset;
1633
1634 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1635 *bytes);
1636 assert(*bytes > 0);
1637
1638 /*
1639 * Calculate the number of clusters to look for. We stop at L2 slice
1640 * boundaries to keep things simple.
1641 */
1642 nb_clusters =
1643 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1644
1645 l2_index = offset_to_l2_slice_index(s, guest_offset);
1646 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1647 /* Limit total allocation byte count to BDRV_REQUEST_MAX_BYTES */
1648 nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1649
1650 /* Find L2 entry for the first involved cluster */
1651 ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1652 if (ret < 0) {
1653 return ret;
1654 }
1655
1656 nb_clusters = count_single_write_clusters(bs, nb_clusters,
1657 l2_slice, l2_index, true);
1658
1659 /* This function is only called when there were no non-COW clusters, so if
1660 * we can't find any unallocated or COW clusters either, something is
1661 * wrong with our code. */
1662 assert(nb_clusters > 0);
1663
1664 /* Allocate at a given offset in the image file */
1665 alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1666 start_of_cluster(s, *host_offset);
1667 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1668 &nb_clusters);
1669 if (ret < 0) {
1670 goto out;
1671 }
1672
1673 /* Can't extend contiguous allocation */
1674 if (nb_clusters == 0) {
1675 *bytes = 0;
1676 ret = 0;
1677 goto out;
1678 }
1679
1680 assert(alloc_cluster_offset != INV_OFFSET);
1681
1682 /*
1683 * Save info needed for meta data update.
1684 *
1685 * requested_bytes: Number of bytes from the start of the first
1686 * newly allocated cluster to the end of the (possibly shortened
1687 * before) write request.
1688 *
1689 * avail_bytes: Number of bytes from the start of the first
1690 * newly allocated to the end of the last newly allocated cluster.
1691 *
1692 * nb_bytes: The number of bytes from the start of the first
1693 * newly allocated cluster to the end of the area that the write
1694 * request actually writes to (excluding COW at the end)
1695 */
1696 uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1697 int avail_bytes = nb_clusters << s->cluster_bits;
1698 int nb_bytes = MIN(requested_bytes, avail_bytes);
1699
1700 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1701 *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1702 assert(*bytes != 0);
1703
1704 ret = calculate_l2_meta(bs, alloc_cluster_offset, guest_offset, *bytes,
1705 l2_slice, m, false);
1706 if (ret < 0) {
1707 goto out;
1708 }
1709
1710 ret = 1;
1711
1712 out:
1713 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1714 return ret;
1715 }
1716
1717 /*
1718 * For a given area on the virtual disk defined by @offset and @bytes,
1719 * find the corresponding area on the qcow2 image, allocating new
1720 * clusters (or subclusters) if necessary. The result can span a
1721 * combination of allocated and previously unallocated clusters.
1722 *
1723 * Note that offset may not be cluster aligned. In this case, the returned
1724 * *host_offset points to exact byte referenced by offset and therefore
1725 * isn't cluster aligned as well.
1726 *
1727 * On return, @host_offset is set to the beginning of the requested
1728 * area. This area is guaranteed to be contiguous on the qcow2 file
1729 * but it can be smaller than initially requested. In this case @bytes
1730 * is updated with the actual size.
1731 *
1732 * If any clusters or subclusters were allocated then @m contains a
1733 * list with the information of all the affected regions. Note that
1734 * this can happen regardless of whether this function succeeds or
1735 * not. The caller is responsible for updating the L2 metadata of the
1736 * allocated clusters (on success) or freeing them (on failure), and
1737 * for clearing the contents of @m afterwards in both cases.
1738 *
1739 * If the request conflicts with another write request in flight, the coroutine
1740 * is queued and will be reentered when the dependency has completed.
1741 *
1742 * Return 0 on success and -errno in error cases
1743 */
1744 int qcow2_alloc_host_offset(BlockDriverState *bs, uint64_t offset,
1745 unsigned int *bytes, uint64_t *host_offset,
1746 QCowL2Meta **m)
1747 {
1748 BDRVQcow2State *s = bs->opaque;
1749 uint64_t start, remaining;
1750 uint64_t cluster_offset;
1751 uint64_t cur_bytes;
1752 int ret;
1753
1754 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1755
1756 again:
1757 start = offset;
1758 remaining = *bytes;
1759 cluster_offset = INV_OFFSET;
1760 *host_offset = INV_OFFSET;
1761 cur_bytes = 0;
1762 *m = NULL;
1763
1764 while (true) {
1765
1766 if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1767 *host_offset = cluster_offset;
1768 }
1769
1770 assert(remaining >= cur_bytes);
1771
1772 start += cur_bytes;
1773 remaining -= cur_bytes;
1774
1775 if (cluster_offset != INV_OFFSET) {
1776 cluster_offset += cur_bytes;
1777 }
1778
1779 if (remaining == 0) {
1780 break;
1781 }
1782
1783 cur_bytes = remaining;
1784
1785 /*
1786 * Now start gathering as many contiguous clusters as possible:
1787 *
1788 * 1. Check for overlaps with in-flight allocations
1789 *
1790 * a) Overlap not in the first cluster -> shorten this request and
1791 * let the caller handle the rest in its next loop iteration.
1792 *
1793 * b) Real overlaps of two requests. Yield and restart the search
1794 * for contiguous clusters (the situation could have changed
1795 * while we were sleeping)
1796 *
1797 * c) TODO: Request starts in the same cluster as the in-flight
1798 * allocation ends. Shorten the COW of the in-fight allocation,
1799 * set cluster_offset to write to the same cluster and set up
1800 * the right synchronisation between the in-flight request and
1801 * the new one.
1802 */
1803 ret = handle_dependencies(bs, start, &cur_bytes, m);
1804 if (ret == -EAGAIN) {
1805 /* Currently handle_dependencies() doesn't yield if we already had
1806 * an allocation. If it did, we would have to clean up the L2Meta
1807 * structs before starting over. */
1808 assert(*m == NULL);
1809 goto again;
1810 } else if (ret < 0) {
1811 return ret;
1812 } else if (cur_bytes == 0) {
1813 break;
1814 } else {
1815 /* handle_dependencies() may have decreased cur_bytes (shortened
1816 * the allocations below) so that the next dependency is processed
1817 * correctly during the next loop iteration. */
1818 }
1819
1820 /*
1821 * 2. Count contiguous COPIED clusters.
1822 */
1823 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1824 if (ret < 0) {
1825 return ret;
1826 } else if (ret) {
1827 continue;
1828 } else if (cur_bytes == 0) {
1829 break;
1830 }
1831
1832 /*
1833 * 3. If the request still hasn't completed, allocate new clusters,
1834 * considering any cluster_offset of steps 1c or 2.
1835 */
1836 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1837 if (ret < 0) {
1838 return ret;
1839 } else if (ret) {
1840 continue;
1841 } else {
1842 assert(cur_bytes == 0);
1843 break;
1844 }
1845 }
1846
1847 *bytes -= remaining;
1848 assert(*bytes > 0);
1849 assert(*host_offset != INV_OFFSET);
1850 assert(offset_into_cluster(s, *host_offset) ==
1851 offset_into_cluster(s, offset));
1852
1853 return 0;
1854 }
1855
1856 /*
1857 * This discards as many clusters of nb_clusters as possible at once (i.e.
1858 * all clusters in the same L2 slice) and returns the number of discarded
1859 * clusters.
1860 */
1861 static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1862 uint64_t nb_clusters,
1863 enum qcow2_discard_type type, bool full_discard)
1864 {
1865 BDRVQcow2State *s = bs->opaque;
1866 uint64_t *l2_slice;
1867 int l2_index;
1868 int ret;
1869 int i;
1870
1871 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1872 if (ret < 0) {
1873 return ret;
1874 }
1875
1876 /* Limit nb_clusters to one L2 slice */
1877 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1878 assert(nb_clusters <= INT_MAX);
1879
1880 for (i = 0; i < nb_clusters; i++) {
1881 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1882 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1883 uint64_t new_l2_entry = old_l2_entry;
1884 uint64_t new_l2_bitmap = old_l2_bitmap;
1885 QCow2ClusterType cluster_type =
1886 qcow2_get_cluster_type(bs, old_l2_entry);
1887
1888 /*
1889 * If full_discard is true, the cluster should not read back as zeroes,
1890 * but rather fall through to the backing file.
1891 *
1892 * If full_discard is false, make sure that a discarded area reads back
1893 * as zeroes for v3 images (we cannot do it for v2 without actually
1894 * writing a zero-filled buffer). We can skip the operation if the
1895 * cluster is already marked as zero, or if it's unallocated and we
1896 * don't have a backing file.
1897 *
1898 * TODO We might want to use bdrv_block_status(bs) here, but we're
1899 * holding s->lock, so that doesn't work today.
1900 */
1901 if (full_discard) {
1902 new_l2_entry = new_l2_bitmap = 0;
1903 } else if (bs->backing || qcow2_cluster_is_allocated(cluster_type)) {
1904 if (has_subclusters(s)) {
1905 new_l2_entry = 0;
1906 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
1907 } else {
1908 new_l2_entry = s->qcow_version >= 3 ? QCOW_OFLAG_ZERO : 0;
1909 }
1910 }
1911
1912 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
1913 continue;
1914 }
1915
1916 /* First remove L2 entries */
1917 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1918 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
1919 if (has_subclusters(s)) {
1920 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
1921 }
1922 /* Then decrease the refcount */
1923 qcow2_free_any_cluster(bs, old_l2_entry, type);
1924 }
1925
1926 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1927
1928 return nb_clusters;
1929 }
1930
1931 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1932 uint64_t bytes, enum qcow2_discard_type type,
1933 bool full_discard)
1934 {
1935 BDRVQcow2State *s = bs->opaque;
1936 uint64_t end_offset = offset + bytes;
1937 uint64_t nb_clusters;
1938 int64_t cleared;
1939 int ret;
1940
1941 /* Caller must pass aligned values, except at image end */
1942 assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1943 assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1944 end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
1945
1946 nb_clusters = size_to_clusters(s, bytes);
1947
1948 s->cache_discards = true;
1949
1950 /* Each L2 slice is handled by its own loop iteration */
1951 while (nb_clusters > 0) {
1952 cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
1953 full_discard);
1954 if (cleared < 0) {
1955 ret = cleared;
1956 goto fail;
1957 }
1958
1959 nb_clusters -= cleared;
1960 offset += (cleared * s->cluster_size);
1961 }
1962
1963 ret = 0;
1964 fail:
1965 s->cache_discards = false;
1966 qcow2_process_discards(bs, ret);
1967
1968 return ret;
1969 }
1970
1971 /*
1972 * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1973 * all clusters in the same L2 slice) and returns the number of zeroed
1974 * clusters.
1975 */
1976 static int zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1977 uint64_t nb_clusters, int flags)
1978 {
1979 BDRVQcow2State *s = bs->opaque;
1980 uint64_t *l2_slice;
1981 int l2_index;
1982 int ret;
1983 int i;
1984
1985 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1986 if (ret < 0) {
1987 return ret;
1988 }
1989
1990 /* Limit nb_clusters to one L2 slice */
1991 nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1992 assert(nb_clusters <= INT_MAX);
1993
1994 for (i = 0; i < nb_clusters; i++) {
1995 uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1996 uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1997 QCow2ClusterType type = qcow2_get_cluster_type(bs, old_l2_entry);
1998 bool unmap = (type == QCOW2_CLUSTER_COMPRESSED) ||
1999 ((flags & BDRV_REQ_MAY_UNMAP) && qcow2_cluster_is_allocated(type));
2000 uint64_t new_l2_entry = unmap ? 0 : old_l2_entry;
2001 uint64_t new_l2_bitmap = old_l2_bitmap;
2002
2003 if (has_subclusters(s)) {
2004 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
2005 } else {
2006 new_l2_entry |= QCOW_OFLAG_ZERO;
2007 }
2008
2009 if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
2010 continue;
2011 }
2012
2013 /* First update L2 entries */
2014 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2015 set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
2016 if (has_subclusters(s)) {
2017 set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
2018 }
2019
2020 /* Then decrease the refcount */
2021 if (unmap) {
2022 qcow2_free_any_cluster(bs, old_l2_entry, QCOW2_DISCARD_REQUEST);
2023 }
2024 }
2025
2026 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2027
2028 return nb_clusters;
2029 }
2030
2031 static int zero_l2_subclusters(BlockDriverState *bs, uint64_t offset,
2032 unsigned nb_subclusters)
2033 {
2034 BDRVQcow2State *s = bs->opaque;
2035 uint64_t *l2_slice;
2036 uint64_t old_l2_bitmap, l2_bitmap;
2037 int l2_index, ret, sc = offset_to_sc_index(s, offset);
2038
2039 /* For full clusters use zero_in_l2_slice() instead */
2040 assert(nb_subclusters > 0 && nb_subclusters < s->subclusters_per_cluster);
2041 assert(sc + nb_subclusters <= s->subclusters_per_cluster);
2042 assert(offset_into_subcluster(s, offset) == 0);
2043
2044 ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2045 if (ret < 0) {
2046 return ret;
2047 }
2048
2049 switch (qcow2_get_cluster_type(bs, get_l2_entry(s, l2_slice, l2_index))) {
2050 case QCOW2_CLUSTER_COMPRESSED:
2051 ret = -ENOTSUP; /* We cannot partially zeroize compressed clusters */
2052 goto out;
2053 case QCOW2_CLUSTER_NORMAL:
2054 case QCOW2_CLUSTER_UNALLOCATED:
2055 break;
2056 default:
2057 g_assert_not_reached();
2058 }
2059
2060 old_l2_bitmap = l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
2061
2062 l2_bitmap |= QCOW_OFLAG_SUB_ZERO_RANGE(sc, sc + nb_subclusters);
2063 l2_bitmap &= ~QCOW_OFLAG_SUB_ALLOC_RANGE(sc, sc + nb_subclusters);
2064
2065 if (old_l2_bitmap != l2_bitmap) {
2066 set_l2_bitmap(s, l2_slice, l2_index, l2_bitmap);
2067 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2068 }
2069
2070 ret = 0;
2071 out:
2072 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2073
2074 return ret;
2075 }
2076
2077 int qcow2_subcluster_zeroize(BlockDriverState *bs, uint64_t offset,
2078 uint64_t bytes, int flags)
2079 {
2080 BDRVQcow2State *s = bs->opaque;
2081 uint64_t end_offset = offset + bytes;
2082 uint64_t nb_clusters;
2083 unsigned head, tail;
2084 int64_t cleared;
2085 int ret;
2086
2087 /* If we have to stay in sync with an external data file, zero out
2088 * s->data_file first. */
2089 if (data_file_is_raw(bs)) {
2090 assert(has_data_file(bs));
2091 ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
2092 if (ret < 0) {
2093 return ret;
2094 }
2095 }
2096
2097 /* Caller must pass aligned values, except at image end */
2098 assert(offset_into_subcluster(s, offset) == 0);
2099 assert(offset_into_subcluster(s, end_offset) == 0 ||
2100 end_offset >= bs->total_sectors << BDRV_SECTOR_BITS);
2101
2102 /*
2103 * The zero flag is only supported by version 3 and newer. However, if we
2104 * have no backing file, we can resort to discard in version 2.
2105 */
2106 if (s->qcow_version < 3) {
2107 if (!bs->backing) {
2108 return qcow2_cluster_discard(bs, offset, bytes,
2109 QCOW2_DISCARD_REQUEST, false);
2110 }
2111 return -ENOTSUP;
2112 }
2113
2114 head = MIN(end_offset, ROUND_UP(offset, s->cluster_size)) - offset;
2115 offset += head;
2116
2117 tail = (end_offset >= bs->total_sectors << BDRV_SECTOR_BITS) ? 0 :
2118 end_offset - MAX(offset, start_of_cluster(s, end_offset));
2119 end_offset -= tail;
2120
2121 s->cache_discards = true;
2122
2123 if (head) {
2124 ret = zero_l2_subclusters(bs, offset - head,
2125 size_to_subclusters(s, head));
2126 if (ret < 0) {
2127 goto fail;
2128 }
2129 }
2130
2131 /* Each L2 slice is handled by its own loop iteration */
2132 nb_clusters = size_to_clusters(s, end_offset - offset);
2133
2134 while (nb_clusters > 0) {
2135 cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
2136 if (cleared < 0) {
2137 ret = cleared;
2138 goto fail;
2139 }
2140
2141 nb_clusters -= cleared;
2142 offset += (cleared * s->cluster_size);
2143 }
2144
2145 if (tail) {
2146 ret = zero_l2_subclusters(bs, end_offset, size_to_subclusters(s, tail));
2147 if (ret < 0) {
2148 goto fail;
2149 }
2150 }
2151
2152 ret = 0;
2153 fail:
2154 s->cache_discards = false;
2155 qcow2_process_discards(bs, ret);
2156
2157 return ret;
2158 }
2159
2160 /*
2161 * Expands all zero clusters in a specific L1 table (or deallocates them, for
2162 * non-backed non-pre-allocated zero clusters).
2163 *
2164 * l1_entries and *visited_l1_entries are used to keep track of progress for
2165 * status_cb(). l1_entries contains the total number of L1 entries and
2166 * *visited_l1_entries counts all visited L1 entries.
2167 */
2168 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
2169 int l1_size, int64_t *visited_l1_entries,
2170 int64_t l1_entries,
2171 BlockDriverAmendStatusCB *status_cb,
2172 void *cb_opaque)
2173 {
2174 BDRVQcow2State *s = bs->opaque;
2175 bool is_active_l1 = (l1_table == s->l1_table);
2176 uint64_t *l2_slice = NULL;
2177 unsigned slice, slice_size2, n_slices;
2178 int ret;
2179 int i, j;
2180
2181 /* qcow2_downgrade() is not allowed in images with subclusters */
2182 assert(!has_subclusters(s));
2183
2184 slice_size2 = s->l2_slice_size * l2_entry_size(s);
2185 n_slices = s->cluster_size / slice_size2;
2186
2187 if (!is_active_l1) {
2188 /* inactive L2 tables require a buffer to be stored in when loading
2189 * them from disk */
2190 l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
2191 if (l2_slice == NULL) {
2192 return -ENOMEM;
2193 }
2194 }
2195
2196 for (i = 0; i < l1_size; i++) {
2197 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
2198 uint64_t l2_refcount;
2199
2200 if (!l2_offset) {
2201 /* unallocated */
2202 (*visited_l1_entries)++;
2203 if (status_cb) {
2204 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2205 }
2206 continue;
2207 }
2208
2209 if (offset_into_cluster(s, l2_offset)) {
2210 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
2211 PRIx64 " unaligned (L1 index: %#x)",
2212 l2_offset, i);
2213 ret = -EIO;
2214 goto fail;
2215 }
2216
2217 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
2218 &l2_refcount);
2219 if (ret < 0) {
2220 goto fail;
2221 }
2222
2223 for (slice = 0; slice < n_slices; slice++) {
2224 uint64_t slice_offset = l2_offset + slice * slice_size2;
2225 bool l2_dirty = false;
2226 if (is_active_l1) {
2227 /* get active L2 tables from cache */
2228 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
2229 (void **)&l2_slice);
2230 } else {
2231 /* load inactive L2 tables from disk */
2232 ret = bdrv_pread(bs->file, slice_offset, l2_slice, slice_size2);
2233 }
2234 if (ret < 0) {
2235 goto fail;
2236 }
2237
2238 for (j = 0; j < s->l2_slice_size; j++) {
2239 uint64_t l2_entry = get_l2_entry(s, l2_slice, j);
2240 int64_t offset = l2_entry & L2E_OFFSET_MASK;
2241 QCow2ClusterType cluster_type =
2242 qcow2_get_cluster_type(bs, l2_entry);
2243
2244 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
2245 cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
2246 continue;
2247 }
2248
2249 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2250 if (!bs->backing) {
2251 /*
2252 * not backed; therefore we can simply deallocate the
2253 * cluster. No need to call set_l2_bitmap(), this
2254 * function doesn't support images with subclusters.
2255 */
2256 set_l2_entry(s, l2_slice, j, 0);
2257 l2_dirty = true;
2258 continue;
2259 }
2260
2261 offset = qcow2_alloc_clusters(bs, s->cluster_size);
2262 if (offset < 0) {
2263 ret = offset;
2264 goto fail;
2265 }
2266
2267 /* The offset must fit in the offset field */
2268 assert((offset & L2E_OFFSET_MASK) == offset);
2269
2270 if (l2_refcount > 1) {
2271 /* For shared L2 tables, set the refcount accordingly
2272 * (it is already 1 and needs to be l2_refcount) */
2273 ret = qcow2_update_cluster_refcount(
2274 bs, offset >> s->cluster_bits,
2275 refcount_diff(1, l2_refcount), false,
2276 QCOW2_DISCARD_OTHER);
2277 if (ret < 0) {
2278 qcow2_free_clusters(bs, offset, s->cluster_size,
2279 QCOW2_DISCARD_OTHER);
2280 goto fail;
2281 }
2282 }
2283 }
2284
2285 if (offset_into_cluster(s, offset)) {
2286 int l2_index = slice * s->l2_slice_size + j;
2287 qcow2_signal_corruption(
2288 bs, true, -1, -1,
2289 "Cluster allocation offset "
2290 "%#" PRIx64 " unaligned (L2 offset: %#"
2291 PRIx64 ", L2 index: %#x)", offset,
2292 l2_offset, l2_index);
2293 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2294 qcow2_free_clusters(bs, offset, s->cluster_size,
2295 QCOW2_DISCARD_ALWAYS);
2296 }
2297 ret = -EIO;
2298 goto fail;
2299 }
2300
2301 ret = qcow2_pre_write_overlap_check(bs, 0, offset,
2302 s->cluster_size, true);
2303 if (ret < 0) {
2304 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2305 qcow2_free_clusters(bs, offset, s->cluster_size,
2306 QCOW2_DISCARD_ALWAYS);
2307 }
2308 goto fail;
2309 }
2310
2311 ret = bdrv_pwrite_zeroes(s->data_file, offset,
2312 s->cluster_size, 0);
2313 if (ret < 0) {
2314 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2315 qcow2_free_clusters(bs, offset, s->cluster_size,
2316 QCOW2_DISCARD_ALWAYS);
2317 }
2318 goto fail;
2319 }
2320
2321 if (l2_refcount == 1) {
2322 set_l2_entry(s, l2_slice, j, offset | QCOW_OFLAG_COPIED);
2323 } else {
2324 set_l2_entry(s, l2_slice, j, offset);
2325 }
2326 /*
2327 * No need to call set_l2_bitmap() after set_l2_entry() because
2328 * this function doesn't support images with subclusters.
2329 */
2330 l2_dirty = true;
2331 }
2332
2333 if (is_active_l1) {
2334 if (l2_dirty) {
2335 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2336 qcow2_cache_depends_on_flush(s->l2_table_cache);
2337 }
2338 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2339 } else {
2340 if (l2_dirty) {
2341 ret = qcow2_pre_write_overlap_check(
2342 bs, QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2,
2343 slice_offset, slice_size2, false);
2344 if (ret < 0) {
2345 goto fail;
2346 }
2347
2348 ret = bdrv_pwrite(bs->file, slice_offset,
2349 l2_slice, slice_size2);
2350 if (ret < 0) {
2351 goto fail;
2352 }
2353 }
2354 }
2355 }
2356
2357 (*visited_l1_entries)++;
2358 if (status_cb) {
2359 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2360 }
2361 }
2362
2363 ret = 0;
2364
2365 fail:
2366 if (l2_slice) {
2367 if (!is_active_l1) {
2368 qemu_vfree(l2_slice);
2369 } else {
2370 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2371 }
2372 }
2373 return ret;
2374 }
2375
2376 /*
2377 * For backed images, expands all zero clusters on the image. For non-backed
2378 * images, deallocates all non-pre-allocated zero clusters (and claims the
2379 * allocation for pre-allocated ones). This is important for downgrading to a
2380 * qcow2 version which doesn't yet support metadata zero clusters.
2381 */
2382 int qcow2_expand_zero_clusters(BlockDriverState *bs,
2383 BlockDriverAmendStatusCB *status_cb,
2384 void *cb_opaque)
2385 {
2386 BDRVQcow2State *s = bs->opaque;
2387 uint64_t *l1_table = NULL;
2388 int64_t l1_entries = 0, visited_l1_entries = 0;
2389 int ret;
2390 int i, j;
2391
2392 if (status_cb) {
2393 l1_entries = s->l1_size;
2394 for (i = 0; i < s->nb_snapshots; i++) {
2395 l1_entries += s->snapshots[i].l1_size;
2396 }
2397 }
2398
2399 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
2400 &visited_l1_entries, l1_entries,
2401 status_cb, cb_opaque);
2402 if (ret < 0) {
2403 goto fail;
2404 }
2405
2406 /* Inactive L1 tables may point to active L2 tables - therefore it is
2407 * necessary to flush the L2 table cache before trying to access the L2
2408 * tables pointed to by inactive L1 entries (else we might try to expand
2409 * zero clusters that have already been expanded); furthermore, it is also
2410 * necessary to empty the L2 table cache, since it may contain tables which
2411 * are now going to be modified directly on disk, bypassing the cache.
2412 * qcow2_cache_empty() does both for us. */
2413 ret = qcow2_cache_empty(bs, s->l2_table_cache);
2414 if (ret < 0) {
2415 goto fail;
2416 }
2417
2418 for (i = 0; i < s->nb_snapshots; i++) {
2419 int l1_size2;
2420 uint64_t *new_l1_table;
2421 Error *local_err = NULL;
2422
2423 ret = qcow2_validate_table(bs, s->snapshots[i].l1_table_offset,
2424 s->snapshots[i].l1_size, L1E_SIZE,
2425 QCOW_MAX_L1_SIZE, "Snapshot L1 table",
2426 &local_err);
2427 if (ret < 0) {
2428 error_report_err(local_err);
2429 goto fail;
2430 }
2431
2432 l1_size2 = s->snapshots[i].l1_size * L1E_SIZE;
2433 new_l1_table = g_try_realloc(l1_table, l1_size2);
2434
2435 if (!new_l1_table) {
2436 ret = -ENOMEM;
2437 goto fail;
2438 }
2439
2440 l1_table = new_l1_table;
2441
2442 ret = bdrv_pread(bs->file, s->snapshots[i].l1_table_offset,
2443 l1_table, l1_size2);
2444 if (ret < 0) {
2445 goto fail;
2446 }
2447
2448 for (j = 0; j < s->snapshots[i].l1_size; j++) {
2449 be64_to_cpus(&l1_table[j]);
2450 }
2451
2452 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
2453 &visited_l1_entries, l1_entries,
2454 status_cb, cb_opaque);
2455 if (ret < 0) {
2456 goto fail;
2457 }
2458 }
2459
2460 ret = 0;
2461
2462 fail:
2463 g_free(l1_table);
2464 return ret;
2465 }