xref: /qemu/block/qcow2-cluster.c (revision 372b69f5)
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 "block/block-io.h"
29 #include "qapi/error.h"
30 #include "qcow2.h"
31 #include "qemu/bswap.h"
32 #include "qemu/memalign.h"
33 #include "trace.h"
34 
35 int coroutine_fn qcow2_shrink_l1_table(BlockDriverState *bs,
36                                        uint64_t exact_size)
37 {
38     BDRVQcow2State *s = bs->opaque;
39     int new_l1_size, i, ret;
40 
41     if (exact_size >= s->l1_size) {
42         return 0;
43     }
44 
45     new_l1_size = exact_size;
46 
47 #ifdef DEBUG_ALLOC2
48     fprintf(stderr, "shrink l1_table from %d to %d\n", s->l1_size, new_l1_size);
49 #endif
50 
51     BLKDBG_CO_EVENT(bs->file, BLKDBG_L1_SHRINK_WRITE_TABLE);
52     ret = bdrv_co_pwrite_zeroes(bs->file,
53                                 s->l1_table_offset + new_l1_size * L1E_SIZE,
54                                 (s->l1_size - new_l1_size) * L1E_SIZE, 0);
55     if (ret < 0) {
56         goto fail;
57     }
58 
59     ret = bdrv_co_flush(bs->file->bs);
60     if (ret < 0) {
61         goto fail;
62     }
63 
64     BLKDBG_CO_EVENT(bs->file, BLKDBG_L1_SHRINK_FREE_L2_CLUSTERS);
65     for (i = s->l1_size - 1; i > new_l1_size - 1; i--) {
66         if ((s->l1_table[i] & L1E_OFFSET_MASK) == 0) {
67             continue;
68         }
69         qcow2_free_clusters(bs, s->l1_table[i] & L1E_OFFSET_MASK,
70                             s->cluster_size, QCOW2_DISCARD_ALWAYS);
71         s->l1_table[i] = 0;
72     }
73     return 0;
74 
75 fail:
76     /*
77      * If the write in the l1_table failed the image may contain a partially
78      * overwritten l1_table. In this case it would be better to clear the
79      * l1_table in memory to avoid possible image corruption.
80      */
81     memset(s->l1_table + new_l1_size, 0,
82            (s->l1_size - new_l1_size) * L1E_SIZE);
83     return ret;
84 }
85 
86 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
87                         bool exact_size)
88 {
89     BDRVQcow2State *s = bs->opaque;
90     int new_l1_size2, ret, i;
91     uint64_t *new_l1_table;
92     int64_t old_l1_table_offset, old_l1_size;
93     int64_t new_l1_table_offset, new_l1_size;
94     uint8_t data[12];
95 
96     if (min_size <= s->l1_size)
97         return 0;
98 
99     /* Do a sanity check on min_size before trying to calculate new_l1_size
100      * (this prevents overflows during the while loop for the calculation of
101      * new_l1_size) */
102     if (min_size > INT_MAX / L1E_SIZE) {
103         return -EFBIG;
104     }
105 
106     if (exact_size) {
107         new_l1_size = min_size;
108     } else {
109         /* Bump size up to reduce the number of times we have to grow */
110         new_l1_size = s->l1_size;
111         if (new_l1_size == 0) {
112             new_l1_size = 1;
113         }
114         while (min_size > new_l1_size) {
115             new_l1_size = DIV_ROUND_UP(new_l1_size * 3, 2);
116         }
117     }
118 
119     QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX);
120     if (new_l1_size > QCOW_MAX_L1_SIZE / L1E_SIZE) {
121         return -EFBIG;
122     }
123 
124 #ifdef DEBUG_ALLOC2
125     fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
126             s->l1_size, new_l1_size);
127 #endif
128 
129     new_l1_size2 = L1E_SIZE * new_l1_size;
130     new_l1_table = qemu_try_blockalign(bs->file->bs, new_l1_size2);
131     if (new_l1_table == NULL) {
132         return -ENOMEM;
133     }
134     memset(new_l1_table, 0, new_l1_size2);
135 
136     if (s->l1_size) {
137         memcpy(new_l1_table, s->l1_table, s->l1_size * L1E_SIZE);
138     }
139 
140     /* write new table (align to cluster) */
141     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
142     new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
143     if (new_l1_table_offset < 0) {
144         qemu_vfree(new_l1_table);
145         return new_l1_table_offset;
146     }
147 
148     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
149     if (ret < 0) {
150         goto fail;
151     }
152 
153     /* the L1 position has not yet been updated, so these clusters must
154      * indeed be completely free */
155     ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
156                                         new_l1_size2, false);
157     if (ret < 0) {
158         goto fail;
159     }
160 
161     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
162     for(i = 0; i < s->l1_size; i++)
163         new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
164     ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_size2,
165                            new_l1_table, 0);
166     if (ret < 0)
167         goto fail;
168     for(i = 0; i < s->l1_size; i++)
169         new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
170 
171     /* set new table */
172     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
173     stl_be_p(data, new_l1_size);
174     stq_be_p(data + 4, new_l1_table_offset);
175     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size),
176                            sizeof(data), data, 0);
177     if (ret < 0) {
178         goto fail;
179     }
180     qemu_vfree(s->l1_table);
181     old_l1_table_offset = s->l1_table_offset;
182     s->l1_table_offset = new_l1_table_offset;
183     s->l1_table = new_l1_table;
184     old_l1_size = s->l1_size;
185     s->l1_size = new_l1_size;
186     qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * L1E_SIZE,
187                         QCOW2_DISCARD_OTHER);
188     return 0;
189  fail:
190     qemu_vfree(new_l1_table);
191     qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
192                         QCOW2_DISCARD_OTHER);
193     return ret;
194 }
195 
196 /*
197  * l2_load
198  *
199  * @bs: The BlockDriverState
200  * @offset: A guest offset, used to calculate what slice of the L2
201  *          table to load.
202  * @l2_offset: Offset to the L2 table in the image file.
203  * @l2_slice: Location to store the pointer to the L2 slice.
204  *
205  * Loads a L2 slice into memory (L2 slices are the parts of L2 tables
206  * that are loaded by the qcow2 cache). If the slice is in the cache,
207  * the cache is used; otherwise the L2 slice is loaded from the image
208  * file.
209  */
210 static int GRAPH_RDLOCK
211 l2_load(BlockDriverState *bs, uint64_t offset,
212         uint64_t l2_offset, uint64_t **l2_slice)
213 {
214     BDRVQcow2State *s = bs->opaque;
215     int start_of_slice = l2_entry_size(s) *
216         (offset_to_l2_index(s, offset) - offset_to_l2_slice_index(s, offset));
217 
218     return qcow2_cache_get(bs, s->l2_table_cache, l2_offset + start_of_slice,
219                            (void **)l2_slice);
220 }
221 
222 /*
223  * Writes an L1 entry to disk (note that depending on the alignment
224  * requirements this function may write more that just one entry in
225  * order to prevent bdrv_pwrite from performing a read-modify-write)
226  */
227 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
228 {
229     BDRVQcow2State *s = bs->opaque;
230     int l1_start_index;
231     int i, ret;
232     int bufsize = MAX(L1E_SIZE,
233                       MIN(bs->file->bs->bl.request_alignment, s->cluster_size));
234     int nentries = bufsize / L1E_SIZE;
235     g_autofree uint64_t *buf = g_try_new0(uint64_t, nentries);
236 
237     if (buf == NULL) {
238         return -ENOMEM;
239     }
240 
241     l1_start_index = QEMU_ALIGN_DOWN(l1_index, nentries);
242     for (i = 0; i < MIN(nentries, s->l1_size - l1_start_index); i++) {
243         buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
244     }
245 
246     ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
247             s->l1_table_offset + L1E_SIZE * l1_start_index, bufsize, false);
248     if (ret < 0) {
249         return ret;
250     }
251 
252     BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
253     ret = bdrv_pwrite_sync(bs->file,
254                            s->l1_table_offset + L1E_SIZE * l1_start_index,
255                            bufsize, buf, 0);
256     if (ret < 0) {
257         return ret;
258     }
259 
260     return 0;
261 }
262 
263 /*
264  * l2_allocate
265  *
266  * Allocate a new l2 entry in the file. If l1_index points to an already
267  * used entry in the L2 table (i.e. we are doing a copy on write for the L2
268  * table) copy the contents of the old L2 table into the newly allocated one.
269  * Otherwise the new table is initialized with zeros.
270  *
271  */
272 
273 static int GRAPH_RDLOCK l2_allocate(BlockDriverState *bs, int l1_index)
274 {
275     BDRVQcow2State *s = bs->opaque;
276     uint64_t old_l2_offset;
277     uint64_t *l2_slice = NULL;
278     unsigned slice, slice_size2, n_slices;
279     int64_t l2_offset;
280     int ret;
281 
282     old_l2_offset = s->l1_table[l1_index];
283 
284     trace_qcow2_l2_allocate(bs, l1_index);
285 
286     /* allocate a new l2 entry */
287 
288     l2_offset = qcow2_alloc_clusters(bs, s->l2_size * l2_entry_size(s));
289     if (l2_offset < 0) {
290         ret = l2_offset;
291         goto fail;
292     }
293 
294     /* The offset must fit in the offset field of the L1 table entry */
295     assert((l2_offset & L1E_OFFSET_MASK) == l2_offset);
296 
297     /* If we're allocating the table at offset 0 then something is wrong */
298     if (l2_offset == 0) {
299         qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
300                                 "allocation of L2 table at offset 0");
301         ret = -EIO;
302         goto fail;
303     }
304 
305     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
306     if (ret < 0) {
307         goto fail;
308     }
309 
310     /* allocate a new entry in the l2 cache */
311 
312     slice_size2 = s->l2_slice_size * l2_entry_size(s);
313     n_slices = s->cluster_size / slice_size2;
314 
315     trace_qcow2_l2_allocate_get_empty(bs, l1_index);
316     for (slice = 0; slice < n_slices; slice++) {
317         ret = qcow2_cache_get_empty(bs, s->l2_table_cache,
318                                     l2_offset + slice * slice_size2,
319                                     (void **) &l2_slice);
320         if (ret < 0) {
321             goto fail;
322         }
323 
324         if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
325             /* if there was no old l2 table, clear the new slice */
326             memset(l2_slice, 0, slice_size2);
327         } else {
328             uint64_t *old_slice;
329             uint64_t old_l2_slice_offset =
330                 (old_l2_offset & L1E_OFFSET_MASK) + slice * slice_size2;
331 
332             /* if there was an old l2 table, read a slice from the disk */
333             BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
334             ret = qcow2_cache_get(bs, s->l2_table_cache, old_l2_slice_offset,
335                                   (void **) &old_slice);
336             if (ret < 0) {
337                 goto fail;
338             }
339 
340             memcpy(l2_slice, old_slice, slice_size2);
341 
342             qcow2_cache_put(s->l2_table_cache, (void **) &old_slice);
343         }
344 
345         /* write the l2 slice to the file */
346         BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
347 
348         trace_qcow2_l2_allocate_write_l2(bs, l1_index);
349         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
350         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
351     }
352 
353     ret = qcow2_cache_flush(bs, s->l2_table_cache);
354     if (ret < 0) {
355         goto fail;
356     }
357 
358     /* update the L1 entry */
359     trace_qcow2_l2_allocate_write_l1(bs, l1_index);
360     s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
361     ret = qcow2_write_l1_entry(bs, l1_index);
362     if (ret < 0) {
363         goto fail;
364     }
365 
366     trace_qcow2_l2_allocate_done(bs, l1_index, 0);
367     return 0;
368 
369 fail:
370     trace_qcow2_l2_allocate_done(bs, l1_index, ret);
371     if (l2_slice != NULL) {
372         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
373     }
374     s->l1_table[l1_index] = old_l2_offset;
375     if (l2_offset > 0) {
376         qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
377                             QCOW2_DISCARD_ALWAYS);
378     }
379     return ret;
380 }
381 
382 /*
383  * For a given L2 entry, count the number of contiguous subclusters of
384  * the same type starting from @sc_from. Compressed clusters are
385  * treated as if they were divided into subclusters of size
386  * s->subcluster_size.
387  *
388  * Return the number of contiguous subclusters and set @type to the
389  * subcluster type.
390  *
391  * If the L2 entry is invalid return -errno and set @type to
392  * QCOW2_SUBCLUSTER_INVALID.
393  */
394 static int qcow2_get_subcluster_range_type(BlockDriverState *bs,
395                                            uint64_t l2_entry,
396                                            uint64_t l2_bitmap,
397                                            unsigned sc_from,
398                                            QCow2SubclusterType *type)
399 {
400     BDRVQcow2State *s = bs->opaque;
401     uint32_t val;
402 
403     *type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_from);
404 
405     if (*type == QCOW2_SUBCLUSTER_INVALID) {
406         return -EINVAL;
407     } else if (!has_subclusters(s) || *type == QCOW2_SUBCLUSTER_COMPRESSED) {
408         return s->subclusters_per_cluster - sc_from;
409     }
410 
411     switch (*type) {
412     case QCOW2_SUBCLUSTER_NORMAL:
413         val = l2_bitmap | QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
414         return cto32(val) - sc_from;
415 
416     case QCOW2_SUBCLUSTER_ZERO_PLAIN:
417     case QCOW2_SUBCLUSTER_ZERO_ALLOC:
418         val = (l2_bitmap | QCOW_OFLAG_SUB_ZERO_RANGE(0, sc_from)) >> 32;
419         return cto32(val) - sc_from;
420 
421     case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
422     case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
423         val = ((l2_bitmap >> 32) | l2_bitmap)
424             & ~QCOW_OFLAG_SUB_ALLOC_RANGE(0, sc_from);
425         return ctz32(val) - sc_from;
426 
427     default:
428         g_assert_not_reached();
429     }
430 }
431 
432 /*
433  * Return the number of contiguous subclusters of the exact same type
434  * in a given L2 slice, starting from cluster @l2_index, subcluster
435  * @sc_index. Allocated subclusters are required to be contiguous in
436  * the image file.
437  * At most @nb_clusters are checked (note that this means clusters,
438  * not subclusters).
439  * Compressed clusters are always processed one by one but for the
440  * purpose of this count they are treated as if they were divided into
441  * subclusters of size s->subcluster_size.
442  * On failure return -errno and update @l2_index to point to the
443  * invalid entry.
444  */
445 static int count_contiguous_subclusters(BlockDriverState *bs, int nb_clusters,
446                                         unsigned sc_index, uint64_t *l2_slice,
447                                         unsigned *l2_index)
448 {
449     BDRVQcow2State *s = bs->opaque;
450     int i, count = 0;
451     bool check_offset = false;
452     uint64_t expected_offset = 0;
453     QCow2SubclusterType expected_type = QCOW2_SUBCLUSTER_NORMAL, type;
454 
455     assert(*l2_index + nb_clusters <= s->l2_slice_size);
456 
457     for (i = 0; i < nb_clusters; i++) {
458         unsigned first_sc = (i == 0) ? sc_index : 0;
459         uint64_t l2_entry = get_l2_entry(s, l2_slice, *l2_index + i);
460         uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, *l2_index + i);
461         int ret = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
462                                                   first_sc, &type);
463         if (ret < 0) {
464             *l2_index += i; /* Point to the invalid entry */
465             return -EIO;
466         }
467         if (i == 0) {
468             if (type == QCOW2_SUBCLUSTER_COMPRESSED) {
469                 /* Compressed clusters are always processed one by one */
470                 return ret;
471             }
472             expected_type = type;
473             expected_offset = l2_entry & L2E_OFFSET_MASK;
474             check_offset = (type == QCOW2_SUBCLUSTER_NORMAL ||
475                             type == QCOW2_SUBCLUSTER_ZERO_ALLOC ||
476                             type == QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC);
477         } else if (type != expected_type) {
478             break;
479         } else if (check_offset) {
480             expected_offset += s->cluster_size;
481             if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
482                 break;
483             }
484         }
485         count += ret;
486         /* Stop if there are type changes before the end of the cluster */
487         if (first_sc + ret < s->subclusters_per_cluster) {
488             break;
489         }
490     }
491 
492     return count;
493 }
494 
495 static int coroutine_fn GRAPH_RDLOCK
496 do_perform_cow_read(BlockDriverState *bs, uint64_t src_cluster_offset,
497                     unsigned offset_in_cluster, QEMUIOVector *qiov)
498 {
499     int ret;
500 
501     if (qiov->size == 0) {
502         return 0;
503     }
504 
505     BLKDBG_CO_EVENT(bs->file, BLKDBG_COW_READ);
506 
507     if (!bs->drv) {
508         return -ENOMEDIUM;
509     }
510 
511     /*
512      * We never deal with requests that don't satisfy
513      * bdrv_check_qiov_request(), and aligning requests to clusters never
514      * breaks this condition. So, do some assertions before calling
515      * bs->drv->bdrv_co_preadv_part() which has int64_t arguments.
516      */
517     assert(src_cluster_offset <= INT64_MAX);
518     assert(src_cluster_offset + offset_in_cluster <= INT64_MAX);
519     /* Cast qiov->size to uint64_t to silence a compiler warning on -m32 */
520     assert((uint64_t)qiov->size <= INT64_MAX);
521     bdrv_check_qiov_request(src_cluster_offset + offset_in_cluster, qiov->size,
522                             qiov, 0, &error_abort);
523     /*
524      * Call .bdrv_co_readv() directly instead of using the public block-layer
525      * interface.  This avoids double I/O throttling and request tracking,
526      * which can lead to deadlock when block layer copy-on-read is enabled.
527      */
528     ret = bs->drv->bdrv_co_preadv_part(bs,
529                                        src_cluster_offset + offset_in_cluster,
530                                        qiov->size, qiov, 0, 0);
531     if (ret < 0) {
532         return ret;
533     }
534 
535     return 0;
536 }
537 
538 static int coroutine_fn GRAPH_RDLOCK
539 do_perform_cow_write(BlockDriverState *bs, uint64_t cluster_offset,
540                      unsigned offset_in_cluster, QEMUIOVector *qiov)
541 {
542     BDRVQcow2State *s = bs->opaque;
543     int ret;
544 
545     if (qiov->size == 0) {
546         return 0;
547     }
548 
549     ret = qcow2_pre_write_overlap_check(bs, 0,
550             cluster_offset + offset_in_cluster, qiov->size, true);
551     if (ret < 0) {
552         return ret;
553     }
554 
555     BLKDBG_CO_EVENT(bs->file, BLKDBG_COW_WRITE);
556     ret = bdrv_co_pwritev(s->data_file, cluster_offset + offset_in_cluster,
557                           qiov->size, qiov, 0);
558     if (ret < 0) {
559         return ret;
560     }
561 
562     return 0;
563 }
564 
565 
566 /*
567  * get_host_offset
568  *
569  * For a given offset of the virtual disk find the equivalent host
570  * offset in the qcow2 file and store it in *host_offset. Neither
571  * offset needs to be aligned to a cluster boundary.
572  *
573  * If the cluster is unallocated then *host_offset will be 0.
574  * If the cluster is compressed then *host_offset will contain the l2 entry.
575  *
576  * On entry, *bytes is the maximum number of contiguous bytes starting at
577  * offset that we are interested in.
578  *
579  * On exit, *bytes is the number of bytes starting at offset that have the same
580  * subcluster type and (if applicable) are stored contiguously in the image
581  * file. The subcluster type is stored in *subcluster_type.
582  * Compressed clusters are always processed one by one.
583  *
584  * Returns 0 on success, -errno in error cases.
585  */
586 int qcow2_get_host_offset(BlockDriverState *bs, uint64_t offset,
587                           unsigned int *bytes, uint64_t *host_offset,
588                           QCow2SubclusterType *subcluster_type)
589 {
590     BDRVQcow2State *s = bs->opaque;
591     unsigned int l2_index, sc_index;
592     uint64_t l1_index, l2_offset, *l2_slice, l2_entry, l2_bitmap;
593     int sc;
594     unsigned int offset_in_cluster;
595     uint64_t bytes_available, bytes_needed, nb_clusters;
596     QCow2SubclusterType type;
597     int ret;
598 
599     offset_in_cluster = offset_into_cluster(s, offset);
600     bytes_needed = (uint64_t) *bytes + offset_in_cluster;
601 
602     /* compute how many bytes there are between the start of the cluster
603      * containing offset and the end of the l2 slice that contains
604      * the entry pointing to it */
605     bytes_available =
606         ((uint64_t) (s->l2_slice_size - offset_to_l2_slice_index(s, offset)))
607         << s->cluster_bits;
608 
609     if (bytes_needed > bytes_available) {
610         bytes_needed = bytes_available;
611     }
612 
613     *host_offset = 0;
614 
615     /* seek to the l2 offset in the l1 table */
616 
617     l1_index = offset_to_l1_index(s, offset);
618     if (l1_index >= s->l1_size) {
619         type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
620         goto out;
621     }
622 
623     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
624     if (!l2_offset) {
625         type = QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN;
626         goto out;
627     }
628 
629     if (offset_into_cluster(s, l2_offset)) {
630         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
631                                 " unaligned (L1 index: %#" PRIx64 ")",
632                                 l2_offset, l1_index);
633         return -EIO;
634     }
635 
636     /* load the l2 slice in memory */
637 
638     ret = l2_load(bs, offset, l2_offset, &l2_slice);
639     if (ret < 0) {
640         return ret;
641     }
642 
643     /* find the cluster offset for the given disk offset */
644 
645     l2_index = offset_to_l2_slice_index(s, offset);
646     sc_index = offset_to_sc_index(s, offset);
647     l2_entry = get_l2_entry(s, l2_slice, l2_index);
648     l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
649 
650     nb_clusters = size_to_clusters(s, bytes_needed);
651     /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
652      * integers; the minimum cluster size is 512, so this assertion is always
653      * true */
654     assert(nb_clusters <= INT_MAX);
655 
656     type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
657     if (s->qcow_version < 3 && (type == QCOW2_SUBCLUSTER_ZERO_PLAIN ||
658                                 type == QCOW2_SUBCLUSTER_ZERO_ALLOC)) {
659         qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
660                                 " in pre-v3 image (L2 offset: %#" PRIx64
661                                 ", L2 index: %#x)", l2_offset, l2_index);
662         ret = -EIO;
663         goto fail;
664     }
665     switch (type) {
666     case QCOW2_SUBCLUSTER_INVALID:
667         break; /* This is handled by count_contiguous_subclusters() below */
668     case QCOW2_SUBCLUSTER_COMPRESSED:
669         if (has_data_file(bs)) {
670             qcow2_signal_corruption(bs, true, -1, -1, "Compressed cluster "
671                                     "entry found in image with external data "
672                                     "file (L2 offset: %#" PRIx64 ", L2 index: "
673                                     "%#x)", l2_offset, l2_index);
674             ret = -EIO;
675             goto fail;
676         }
677         *host_offset = l2_entry;
678         break;
679     case QCOW2_SUBCLUSTER_ZERO_PLAIN:
680     case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
681         break;
682     case QCOW2_SUBCLUSTER_ZERO_ALLOC:
683     case QCOW2_SUBCLUSTER_NORMAL:
684     case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC: {
685         uint64_t host_cluster_offset = l2_entry & L2E_OFFSET_MASK;
686         *host_offset = host_cluster_offset + offset_in_cluster;
687         if (offset_into_cluster(s, host_cluster_offset)) {
688             qcow2_signal_corruption(bs, true, -1, -1,
689                                     "Cluster allocation offset %#"
690                                     PRIx64 " unaligned (L2 offset: %#" PRIx64
691                                     ", L2 index: %#x)", host_cluster_offset,
692                                     l2_offset, l2_index);
693             ret = -EIO;
694             goto fail;
695         }
696         if (has_data_file(bs) && *host_offset != offset) {
697             qcow2_signal_corruption(bs, true, -1, -1,
698                                     "External data file host cluster offset %#"
699                                     PRIx64 " does not match guest cluster "
700                                     "offset: %#" PRIx64
701                                     ", L2 index: %#x)", host_cluster_offset,
702                                     offset - offset_in_cluster, l2_index);
703             ret = -EIO;
704             goto fail;
705         }
706         break;
707     }
708     default:
709         abort();
710     }
711 
712     sc = count_contiguous_subclusters(bs, nb_clusters, sc_index,
713                                       l2_slice, &l2_index);
714     if (sc < 0) {
715         qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster entry found "
716                                 " (L2 offset: %#" PRIx64 ", L2 index: %#x)",
717                                 l2_offset, l2_index);
718         ret = -EIO;
719         goto fail;
720     }
721     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
722 
723     bytes_available = ((int64_t)sc + sc_index) << s->subcluster_bits;
724 
725 out:
726     if (bytes_available > bytes_needed) {
727         bytes_available = bytes_needed;
728     }
729 
730     /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
731      * subtracting offset_in_cluster will therefore definitely yield something
732      * not exceeding UINT_MAX */
733     assert(bytes_available - offset_in_cluster <= UINT_MAX);
734     *bytes = bytes_available - offset_in_cluster;
735 
736     *subcluster_type = type;
737 
738     return 0;
739 
740 fail:
741     qcow2_cache_put(s->l2_table_cache, (void **)&l2_slice);
742     return ret;
743 }
744 
745 /*
746  * get_cluster_table
747  *
748  * for a given disk offset, load (and allocate if needed)
749  * the appropriate slice of its l2 table.
750  *
751  * the cluster index in the l2 slice is given to the caller.
752  *
753  * Returns 0 on success, -errno in failure case
754  */
755 static int GRAPH_RDLOCK
756 get_cluster_table(BlockDriverState *bs, uint64_t offset,
757                   uint64_t **new_l2_slice, int *new_l2_index)
758 {
759     BDRVQcow2State *s = bs->opaque;
760     unsigned int l2_index;
761     uint64_t l1_index, l2_offset;
762     uint64_t *l2_slice = NULL;
763     int ret;
764 
765     /* seek to the l2 offset in the l1 table */
766 
767     l1_index = offset_to_l1_index(s, offset);
768     if (l1_index >= s->l1_size) {
769         ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
770         if (ret < 0) {
771             return ret;
772         }
773     }
774 
775     assert(l1_index < s->l1_size);
776     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
777     if (offset_into_cluster(s, l2_offset)) {
778         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
779                                 " unaligned (L1 index: %#" PRIx64 ")",
780                                 l2_offset, l1_index);
781         return -EIO;
782     }
783 
784     if (!(s->l1_table[l1_index] & QCOW_OFLAG_COPIED)) {
785         /* First allocate a new L2 table (and do COW if needed) */
786         ret = l2_allocate(bs, l1_index);
787         if (ret < 0) {
788             return ret;
789         }
790 
791         /* Then decrease the refcount of the old table */
792         if (l2_offset) {
793             qcow2_free_clusters(bs, l2_offset, s->l2_size * l2_entry_size(s),
794                                 QCOW2_DISCARD_OTHER);
795         }
796 
797         /* Get the offset of the newly-allocated l2 table */
798         l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
799         assert(offset_into_cluster(s, l2_offset) == 0);
800     }
801 
802     /* load the l2 slice in memory */
803     ret = l2_load(bs, offset, l2_offset, &l2_slice);
804     if (ret < 0) {
805         return ret;
806     }
807 
808     /* find the cluster offset for the given disk offset */
809 
810     l2_index = offset_to_l2_slice_index(s, offset);
811 
812     *new_l2_slice = l2_slice;
813     *new_l2_index = l2_index;
814 
815     return 0;
816 }
817 
818 /*
819  * alloc_compressed_cluster_offset
820  *
821  * For a given offset on the virtual disk, allocate a new compressed cluster
822  * and put the host offset of the cluster into *host_offset. If a cluster is
823  * already allocated at the offset, return an error.
824  *
825  * Return 0 on success and -errno in error cases
826  */
827 int coroutine_fn GRAPH_RDLOCK
828 qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs, uint64_t offset,
829                                       int compressed_size, uint64_t *host_offset)
830 {
831     BDRVQcow2State *s = bs->opaque;
832     int l2_index, ret;
833     uint64_t *l2_slice;
834     int64_t cluster_offset;
835     int nb_csectors;
836 
837     if (has_data_file(bs)) {
838         return 0;
839     }
840 
841     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
842     if (ret < 0) {
843         return ret;
844     }
845 
846     /* Compression can't overwrite anything. Fail if the cluster was already
847      * allocated. */
848     cluster_offset = get_l2_entry(s, l2_slice, l2_index);
849     if (cluster_offset & L2E_OFFSET_MASK) {
850         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
851         return -EIO;
852     }
853 
854     cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
855     if (cluster_offset < 0) {
856         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
857         return cluster_offset;
858     }
859 
860     nb_csectors =
861         (cluster_offset + compressed_size - 1) / QCOW2_COMPRESSED_SECTOR_SIZE -
862         (cluster_offset / QCOW2_COMPRESSED_SECTOR_SIZE);
863 
864     /* The offset and size must fit in their fields of the L2 table entry */
865     assert((cluster_offset & s->cluster_offset_mask) == cluster_offset);
866     assert((nb_csectors & s->csize_mask) == nb_csectors);
867 
868     cluster_offset |= QCOW_OFLAG_COMPRESSED |
869                       ((uint64_t)nb_csectors << s->csize_shift);
870 
871     /* update L2 table */
872 
873     /* compressed clusters never have the copied flag */
874 
875     BLKDBG_CO_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
876     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
877     set_l2_entry(s, l2_slice, l2_index, cluster_offset);
878     if (has_subclusters(s)) {
879         set_l2_bitmap(s, l2_slice, l2_index, 0);
880     }
881     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
882 
883     *host_offset = cluster_offset & s->cluster_offset_mask;
884     return 0;
885 }
886 
887 static int coroutine_fn GRAPH_RDLOCK
888 perform_cow(BlockDriverState *bs, QCowL2Meta *m)
889 {
890     BDRVQcow2State *s = bs->opaque;
891     Qcow2COWRegion *start = &m->cow_start;
892     Qcow2COWRegion *end = &m->cow_end;
893     unsigned buffer_size;
894     unsigned data_bytes = end->offset - (start->offset + start->nb_bytes);
895     bool merge_reads;
896     uint8_t *start_buffer, *end_buffer;
897     QEMUIOVector qiov;
898     int ret;
899 
900     assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
901     assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
902     assert(start->offset + start->nb_bytes <= end->offset);
903 
904     if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
905         return 0;
906     }
907 
908     /* If we have to read both the start and end COW regions and the
909      * middle region is not too large then perform just one read
910      * operation */
911     merge_reads = start->nb_bytes && end->nb_bytes && data_bytes <= 16384;
912     if (merge_reads) {
913         buffer_size = start->nb_bytes + data_bytes + end->nb_bytes;
914     } else {
915         /* If we have to do two reads, add some padding in the middle
916          * if necessary to make sure that the end region is optimally
917          * aligned. */
918         size_t align = bdrv_opt_mem_align(bs);
919         assert(align > 0 && align <= UINT_MAX);
920         assert(QEMU_ALIGN_UP(start->nb_bytes, align) <=
921                UINT_MAX - end->nb_bytes);
922         buffer_size = QEMU_ALIGN_UP(start->nb_bytes, align) + end->nb_bytes;
923     }
924 
925     /* Reserve a buffer large enough to store all the data that we're
926      * going to read */
927     start_buffer = qemu_try_blockalign(bs, buffer_size);
928     if (start_buffer == NULL) {
929         return -ENOMEM;
930     }
931     /* The part of the buffer where the end region is located */
932     end_buffer = start_buffer + buffer_size - end->nb_bytes;
933 
934     qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
935                                 qemu_iovec_subvec_niov(m->data_qiov,
936                                                        m->data_qiov_offset,
937                                                        data_bytes)
938                                 : 0));
939 
940     qemu_co_mutex_unlock(&s->lock);
941     /* First we read the existing data from both COW regions. We
942      * either read the whole region in one go, or the start and end
943      * regions separately. */
944     if (merge_reads) {
945         qemu_iovec_add(&qiov, start_buffer, buffer_size);
946         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
947     } else {
948         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
949         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
950         if (ret < 0) {
951             goto fail;
952         }
953 
954         qemu_iovec_reset(&qiov);
955         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
956         ret = do_perform_cow_read(bs, m->offset, end->offset, &qiov);
957     }
958     if (ret < 0) {
959         goto fail;
960     }
961 
962     /* Encrypt the data if necessary before writing it */
963     if (bs->encrypted) {
964         ret = qcow2_co_encrypt(bs,
965                                m->alloc_offset + start->offset,
966                                m->offset + start->offset,
967                                start_buffer, start->nb_bytes);
968         if (ret < 0) {
969             goto fail;
970         }
971 
972         ret = qcow2_co_encrypt(bs,
973                                m->alloc_offset + end->offset,
974                                m->offset + end->offset,
975                                end_buffer, end->nb_bytes);
976         if (ret < 0) {
977             goto fail;
978         }
979     }
980 
981     /* And now we can write everything. If we have the guest data we
982      * can write everything in one single operation */
983     if (m->data_qiov) {
984         qemu_iovec_reset(&qiov);
985         if (start->nb_bytes) {
986             qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
987         }
988         qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
989         if (end->nb_bytes) {
990             qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
991         }
992         /* NOTE: we have a write_aio blkdebug event here followed by
993          * a cow_write one in do_perform_cow_write(), but there's only
994          * one single I/O operation */
995         BLKDBG_CO_EVENT(bs->file, BLKDBG_WRITE_AIO);
996         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
997     } else {
998         /* If there's no guest data then write both COW regions separately */
999         qemu_iovec_reset(&qiov);
1000         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
1001         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
1002         if (ret < 0) {
1003             goto fail;
1004         }
1005 
1006         qemu_iovec_reset(&qiov);
1007         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
1008         ret = do_perform_cow_write(bs, m->alloc_offset, end->offset, &qiov);
1009     }
1010 
1011 fail:
1012     qemu_co_mutex_lock(&s->lock);
1013 
1014     /*
1015      * Before we update the L2 table to actually point to the new cluster, we
1016      * need to be sure that the refcounts have been increased and COW was
1017      * handled.
1018      */
1019     if (ret == 0) {
1020         qcow2_cache_depends_on_flush(s->l2_table_cache);
1021     }
1022 
1023     qemu_vfree(start_buffer);
1024     qemu_iovec_destroy(&qiov);
1025     return ret;
1026 }
1027 
1028 int coroutine_fn qcow2_alloc_cluster_link_l2(BlockDriverState *bs,
1029                                              QCowL2Meta *m)
1030 {
1031     BDRVQcow2State *s = bs->opaque;
1032     int i, j = 0, l2_index, ret;
1033     uint64_t *old_cluster, *l2_slice;
1034     uint64_t cluster_offset = m->alloc_offset;
1035 
1036     trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
1037     assert(m->nb_clusters > 0);
1038 
1039     old_cluster = g_try_new(uint64_t, m->nb_clusters);
1040     if (old_cluster == NULL) {
1041         ret = -ENOMEM;
1042         goto err;
1043     }
1044 
1045     /* copy content of unmodified sectors */
1046     ret = perform_cow(bs, m);
1047     if (ret < 0) {
1048         goto err;
1049     }
1050 
1051     /* Update L2 table. */
1052     if (s->use_lazy_refcounts) {
1053         qcow2_mark_dirty(bs);
1054     }
1055     if (qcow2_need_accurate_refcounts(s)) {
1056         qcow2_cache_set_dependency(bs, s->l2_table_cache,
1057                                    s->refcount_block_cache);
1058     }
1059 
1060     ret = get_cluster_table(bs, m->offset, &l2_slice, &l2_index);
1061     if (ret < 0) {
1062         goto err;
1063     }
1064     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1065 
1066     assert(l2_index + m->nb_clusters <= s->l2_slice_size);
1067     assert(m->cow_end.offset + m->cow_end.nb_bytes <=
1068            m->nb_clusters << s->cluster_bits);
1069     for (i = 0; i < m->nb_clusters; i++) {
1070         uint64_t offset = cluster_offset + ((uint64_t)i << s->cluster_bits);
1071         /* if two concurrent writes happen to the same unallocated cluster
1072          * each write allocates separate cluster and writes data concurrently.
1073          * The first one to complete updates l2 table with pointer to its
1074          * cluster the second one has to do RMW (which is done above by
1075          * perform_cow()), update l2 table with its cluster pointer and free
1076          * old cluster. This is what this loop does */
1077         if (get_l2_entry(s, l2_slice, l2_index + i) != 0) {
1078             old_cluster[j++] = get_l2_entry(s, l2_slice, l2_index + i);
1079         }
1080 
1081         /* The offset must fit in the offset field of the L2 table entry */
1082         assert((offset & L2E_OFFSET_MASK) == offset);
1083 
1084         set_l2_entry(s, l2_slice, l2_index + i, offset | QCOW_OFLAG_COPIED);
1085 
1086         /* Update bitmap with the subclusters that were just written */
1087         if (has_subclusters(s) && !m->prealloc) {
1088             uint64_t l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1089             unsigned written_from = m->cow_start.offset;
1090             unsigned written_to = m->cow_end.offset + m->cow_end.nb_bytes;
1091             int first_sc, last_sc;
1092             /* Narrow written_from and written_to down to the current cluster */
1093             written_from = MAX(written_from, i << s->cluster_bits);
1094             written_to   = MIN(written_to, (i + 1) << s->cluster_bits);
1095             assert(written_from < written_to);
1096             first_sc = offset_to_sc_index(s, written_from);
1097             last_sc  = offset_to_sc_index(s, written_to - 1);
1098             l2_bitmap |= QCOW_OFLAG_SUB_ALLOC_RANGE(first_sc, last_sc + 1);
1099             l2_bitmap &= ~QCOW_OFLAG_SUB_ZERO_RANGE(first_sc, last_sc + 1);
1100             set_l2_bitmap(s, l2_slice, l2_index + i, l2_bitmap);
1101         }
1102      }
1103 
1104 
1105     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1106 
1107     /*
1108      * If this was a COW, we need to decrease the refcount of the old cluster.
1109      *
1110      * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1111      * clusters), the next write will reuse them anyway.
1112      */
1113     if (!m->keep_old_clusters && j != 0) {
1114         for (i = 0; i < j; i++) {
1115             qcow2_free_any_cluster(bs, old_cluster[i], QCOW2_DISCARD_NEVER);
1116         }
1117     }
1118 
1119     ret = 0;
1120 err:
1121     g_free(old_cluster);
1122     return ret;
1123  }
1124 
1125 /**
1126  * Frees the allocated clusters because the request failed and they won't
1127  * actually be linked.
1128  */
1129 void coroutine_fn qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1130 {
1131     BDRVQcow2State *s = bs->opaque;
1132     if (!has_data_file(bs) && !m->keep_old_clusters) {
1133         qcow2_free_clusters(bs, m->alloc_offset,
1134                             m->nb_clusters << s->cluster_bits,
1135                             QCOW2_DISCARD_NEVER);
1136     }
1137 }
1138 
1139 /*
1140  * For a given write request, create a new QCowL2Meta structure, add
1141  * it to @m and the BDRVQcow2State.cluster_allocs list. If the write
1142  * request does not need copy-on-write or changes to the L2 metadata
1143  * then this function does nothing.
1144  *
1145  * @host_cluster_offset points to the beginning of the first cluster.
1146  *
1147  * @guest_offset and @bytes indicate the offset and length of the
1148  * request.
1149  *
1150  * @l2_slice contains the L2 entries of all clusters involved in this
1151  * write request.
1152  *
1153  * If @keep_old is true it means that the clusters were already
1154  * allocated and will be overwritten. If false then the clusters are
1155  * new and we have to decrease the reference count of the old ones.
1156  *
1157  * Returns 0 on success, -errno on failure.
1158  */
1159 static int coroutine_fn GRAPH_RDLOCK
1160 calculate_l2_meta(BlockDriverState *bs, uint64_t host_cluster_offset,
1161                   uint64_t guest_offset, unsigned bytes, uint64_t *l2_slice,
1162                   QCowL2Meta **m, bool keep_old)
1163 {
1164     BDRVQcow2State *s = bs->opaque;
1165     int sc_index, l2_index = offset_to_l2_slice_index(s, guest_offset);
1166     uint64_t l2_entry, l2_bitmap;
1167     unsigned cow_start_from, cow_end_to;
1168     unsigned cow_start_to = offset_into_cluster(s, guest_offset);
1169     unsigned cow_end_from = cow_start_to + bytes;
1170     unsigned nb_clusters = size_to_clusters(s, cow_end_from);
1171     QCowL2Meta *old_m = *m;
1172     QCow2SubclusterType type;
1173     int i;
1174     bool skip_cow = keep_old;
1175 
1176     assert(nb_clusters <= s->l2_slice_size - l2_index);
1177 
1178     /* Check the type of all affected subclusters */
1179     for (i = 0; i < nb_clusters; i++) {
1180         l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1181         l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1182         if (skip_cow) {
1183             unsigned write_from = MAX(cow_start_to, i << s->cluster_bits);
1184             unsigned write_to = MIN(cow_end_from, (i + 1) << s->cluster_bits);
1185             int first_sc = offset_to_sc_index(s, write_from);
1186             int last_sc = offset_to_sc_index(s, write_to - 1);
1187             int cnt = qcow2_get_subcluster_range_type(bs, l2_entry, l2_bitmap,
1188                                                       first_sc, &type);
1189             /* Is any of the subclusters of type != QCOW2_SUBCLUSTER_NORMAL ? */
1190             if (type != QCOW2_SUBCLUSTER_NORMAL || first_sc + cnt <= last_sc) {
1191                 skip_cow = false;
1192             }
1193         } else {
1194             /* If we can't skip the cow we can still look for invalid entries */
1195             type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, 0);
1196         }
1197         if (type == QCOW2_SUBCLUSTER_INVALID) {
1198             int l1_index = offset_to_l1_index(s, guest_offset);
1199             uint64_t l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
1200             qcow2_signal_corruption(bs, true, -1, -1, "Invalid cluster "
1201                                     "entry found (L2 offset: %#" PRIx64
1202                                     ", L2 index: %#x)",
1203                                     l2_offset, l2_index + i);
1204             return -EIO;
1205         }
1206     }
1207 
1208     if (skip_cow) {
1209         return 0;
1210     }
1211 
1212     /* Get the L2 entry of the first cluster */
1213     l2_entry = get_l2_entry(s, l2_slice, l2_index);
1214     l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1215     sc_index = offset_to_sc_index(s, guest_offset);
1216     type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1217 
1218     if (!keep_old) {
1219         switch (type) {
1220         case QCOW2_SUBCLUSTER_COMPRESSED:
1221             cow_start_from = 0;
1222             break;
1223         case QCOW2_SUBCLUSTER_NORMAL:
1224         case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1225         case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1226             if (has_subclusters(s)) {
1227                 /* Skip all leading zero and unallocated subclusters */
1228                 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1229                 cow_start_from =
1230                     MIN(sc_index, ctz32(alloc_bitmap)) << s->subcluster_bits;
1231             } else {
1232                 cow_start_from = 0;
1233             }
1234             break;
1235         case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1236         case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1237             cow_start_from = sc_index << s->subcluster_bits;
1238             break;
1239         default:
1240             g_assert_not_reached();
1241         }
1242     } else {
1243         switch (type) {
1244         case QCOW2_SUBCLUSTER_NORMAL:
1245             cow_start_from = cow_start_to;
1246             break;
1247         case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1248         case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1249             cow_start_from = sc_index << s->subcluster_bits;
1250             break;
1251         default:
1252             g_assert_not_reached();
1253         }
1254     }
1255 
1256     /* Get the L2 entry of the last cluster */
1257     l2_index += nb_clusters - 1;
1258     l2_entry = get_l2_entry(s, l2_slice, l2_index);
1259     l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
1260     sc_index = offset_to_sc_index(s, guest_offset + bytes - 1);
1261     type = qcow2_get_subcluster_type(bs, l2_entry, l2_bitmap, sc_index);
1262 
1263     if (!keep_old) {
1264         switch (type) {
1265         case QCOW2_SUBCLUSTER_COMPRESSED:
1266             cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1267             break;
1268         case QCOW2_SUBCLUSTER_NORMAL:
1269         case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1270         case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1271             cow_end_to = ROUND_UP(cow_end_from, s->cluster_size);
1272             if (has_subclusters(s)) {
1273                 /* Skip all trailing zero and unallocated subclusters */
1274                 uint32_t alloc_bitmap = l2_bitmap & QCOW_L2_BITMAP_ALL_ALLOC;
1275                 cow_end_to -=
1276                     MIN(s->subclusters_per_cluster - sc_index - 1,
1277                         clz32(alloc_bitmap)) << s->subcluster_bits;
1278             }
1279             break;
1280         case QCOW2_SUBCLUSTER_ZERO_PLAIN:
1281         case QCOW2_SUBCLUSTER_UNALLOCATED_PLAIN:
1282             cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1283             break;
1284         default:
1285             g_assert_not_reached();
1286         }
1287     } else {
1288         switch (type) {
1289         case QCOW2_SUBCLUSTER_NORMAL:
1290             cow_end_to = cow_end_from;
1291             break;
1292         case QCOW2_SUBCLUSTER_ZERO_ALLOC:
1293         case QCOW2_SUBCLUSTER_UNALLOCATED_ALLOC:
1294             cow_end_to = ROUND_UP(cow_end_from, s->subcluster_size);
1295             break;
1296         default:
1297             g_assert_not_reached();
1298         }
1299     }
1300 
1301     *m = g_malloc0(sizeof(**m));
1302     **m = (QCowL2Meta) {
1303         .next           = old_m,
1304 
1305         .alloc_offset   = host_cluster_offset,
1306         .offset         = start_of_cluster(s, guest_offset),
1307         .nb_clusters    = nb_clusters,
1308 
1309         .keep_old_clusters = keep_old,
1310 
1311         .cow_start = {
1312             .offset     = cow_start_from,
1313             .nb_bytes   = cow_start_to - cow_start_from,
1314         },
1315         .cow_end = {
1316             .offset     = cow_end_from,
1317             .nb_bytes   = cow_end_to - cow_end_from,
1318         },
1319     };
1320 
1321     qemu_co_queue_init(&(*m)->dependent_requests);
1322     QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1323 
1324     return 0;
1325 }
1326 
1327 /*
1328  * Returns true if writing to the cluster pointed to by @l2_entry
1329  * requires a new allocation (that is, if the cluster is unallocated
1330  * or has refcount > 1 and therefore cannot be written in-place).
1331  */
1332 static bool cluster_needs_new_alloc(BlockDriverState *bs, uint64_t l2_entry)
1333 {
1334     switch (qcow2_get_cluster_type(bs, l2_entry)) {
1335     case QCOW2_CLUSTER_NORMAL:
1336     case QCOW2_CLUSTER_ZERO_ALLOC:
1337         if (l2_entry & QCOW_OFLAG_COPIED) {
1338             return false;
1339         }
1340         /* fallthrough */
1341     case QCOW2_CLUSTER_UNALLOCATED:
1342     case QCOW2_CLUSTER_COMPRESSED:
1343     case QCOW2_CLUSTER_ZERO_PLAIN:
1344         return true;
1345     default:
1346         abort();
1347     }
1348 }
1349 
1350 /*
1351  * Returns the number of contiguous clusters that can be written to
1352  * using one single write request, starting from @l2_index.
1353  * At most @nb_clusters are checked.
1354  *
1355  * If @new_alloc is true this counts clusters that are either
1356  * unallocated, or allocated but with refcount > 1 (so they need to be
1357  * newly allocated and COWed).
1358  *
1359  * If @new_alloc is false this counts clusters that are already
1360  * allocated and can be overwritten in-place (this includes clusters
1361  * of type QCOW2_CLUSTER_ZERO_ALLOC).
1362  */
1363 static int count_single_write_clusters(BlockDriverState *bs, int nb_clusters,
1364                                        uint64_t *l2_slice, int l2_index,
1365                                        bool new_alloc)
1366 {
1367     BDRVQcow2State *s = bs->opaque;
1368     uint64_t l2_entry = get_l2_entry(s, l2_slice, l2_index);
1369     uint64_t expected_offset = l2_entry & L2E_OFFSET_MASK;
1370     int i;
1371 
1372     for (i = 0; i < nb_clusters; i++) {
1373         l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1374         if (cluster_needs_new_alloc(bs, l2_entry) != new_alloc) {
1375             break;
1376         }
1377         if (!new_alloc) {
1378             if (expected_offset != (l2_entry & L2E_OFFSET_MASK)) {
1379                 break;
1380             }
1381             expected_offset += s->cluster_size;
1382         }
1383     }
1384 
1385     assert(i <= nb_clusters);
1386     return i;
1387 }
1388 
1389 /*
1390  * Check if there already is an AIO write request in flight which allocates
1391  * the same cluster. In this case we need to wait until the previous
1392  * request has completed and updated the L2 table accordingly.
1393  *
1394  * Returns:
1395  *   0       if there was no dependency. *cur_bytes indicates the number of
1396  *           bytes from guest_offset that can be read before the next
1397  *           dependency must be processed (or the request is complete)
1398  *
1399  *   -EAGAIN if we had to wait for another request, previously gathered
1400  *           information on cluster allocation may be invalid now. The caller
1401  *           must start over anyway, so consider *cur_bytes undefined.
1402  */
1403 static int coroutine_fn handle_dependencies(BlockDriverState *bs,
1404                                             uint64_t guest_offset,
1405                                             uint64_t *cur_bytes, QCowL2Meta **m)
1406 {
1407     BDRVQcow2State *s = bs->opaque;
1408     QCowL2Meta *old_alloc;
1409     uint64_t bytes = *cur_bytes;
1410 
1411     QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
1412 
1413         uint64_t start = guest_offset;
1414         uint64_t end = start + bytes;
1415         uint64_t old_start = start_of_cluster(s, l2meta_cow_start(old_alloc));
1416         uint64_t old_end = ROUND_UP(l2meta_cow_end(old_alloc), s->cluster_size);
1417 
1418         if (end <= old_start || start >= old_end) {
1419             /* No intersection */
1420             continue;
1421         }
1422 
1423         if (old_alloc->keep_old_clusters &&
1424             (end <= l2meta_cow_start(old_alloc) ||
1425              start >= l2meta_cow_end(old_alloc)))
1426         {
1427             /*
1428              * Clusters intersect but COW areas don't. And cluster itself is
1429              * already allocated. So, there is no actual conflict.
1430              */
1431             continue;
1432         }
1433 
1434         /* Conflict */
1435 
1436         if (start < old_start) {
1437             /* Stop at the start of a running allocation */
1438             bytes = old_start - start;
1439         } else {
1440             bytes = 0;
1441         }
1442 
1443         /*
1444          * Stop if an l2meta already exists. After yielding, it wouldn't
1445          * be valid any more, so we'd have to clean up the old L2Metas
1446          * and deal with requests depending on them before starting to
1447          * gather new ones. Not worth the trouble.
1448          */
1449         if (bytes == 0 && *m) {
1450             *cur_bytes = 0;
1451             return 0;
1452         }
1453 
1454         if (bytes == 0) {
1455             /*
1456              * Wait for the dependency to complete. We need to recheck
1457              * the free/allocated clusters when we continue.
1458              */
1459             qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
1460             return -EAGAIN;
1461         }
1462     }
1463 
1464     /* Make sure that existing clusters and new allocations are only used up to
1465      * the next dependency if we shortened the request above */
1466     *cur_bytes = bytes;
1467 
1468     return 0;
1469 }
1470 
1471 /*
1472  * Checks how many already allocated clusters that don't require a new
1473  * allocation there are at the given guest_offset (up to *bytes).
1474  * If *host_offset is not INV_OFFSET, only physically contiguous clusters
1475  * beginning at this host offset are counted.
1476  *
1477  * Note that guest_offset may not be cluster aligned. In this case, the
1478  * returned *host_offset points to exact byte referenced by guest_offset and
1479  * therefore isn't cluster aligned as well.
1480  *
1481  * Returns:
1482  *   0:     if no allocated clusters are available at the given offset.
1483  *          *bytes is normally unchanged. It is set to 0 if the cluster
1484  *          is allocated and can be overwritten in-place but doesn't have
1485  *          the right physical offset.
1486  *
1487  *   1:     if allocated clusters that can be overwritten in place are
1488  *          available at the requested offset. *bytes may have decreased
1489  *          and describes the length of the area that can be written to.
1490  *
1491  *  -errno: in error cases
1492  */
1493 static int coroutine_fn GRAPH_RDLOCK
1494 handle_copied(BlockDriverState *bs, uint64_t guest_offset,
1495               uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1496 {
1497     BDRVQcow2State *s = bs->opaque;
1498     int l2_index;
1499     uint64_t l2_entry, cluster_offset;
1500     uint64_t *l2_slice;
1501     uint64_t nb_clusters;
1502     unsigned int keep_clusters;
1503     int ret;
1504 
1505     trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
1506                               *bytes);
1507 
1508     assert(*host_offset == INV_OFFSET || offset_into_cluster(s, guest_offset)
1509                                       == offset_into_cluster(s, *host_offset));
1510 
1511     /*
1512      * Calculate the number of clusters to look for. We stop at L2 slice
1513      * boundaries to keep things simple.
1514      */
1515     nb_clusters =
1516         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1517 
1518     l2_index = offset_to_l2_slice_index(s, guest_offset);
1519     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1520     /* Limit total byte count to BDRV_REQUEST_MAX_BYTES */
1521     nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1522 
1523     /* Find L2 entry for the first involved cluster */
1524     ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1525     if (ret < 0) {
1526         return ret;
1527     }
1528 
1529     l2_entry = get_l2_entry(s, l2_slice, l2_index);
1530     cluster_offset = l2_entry & L2E_OFFSET_MASK;
1531 
1532     if (!cluster_needs_new_alloc(bs, l2_entry)) {
1533         if (offset_into_cluster(s, cluster_offset)) {
1534             qcow2_signal_corruption(bs, true, -1, -1, "%s cluster offset "
1535                                     "%#" PRIx64 " unaligned (guest offset: %#"
1536                                     PRIx64 ")", l2_entry & QCOW_OFLAG_ZERO ?
1537                                     "Preallocated zero" : "Data",
1538                                     cluster_offset, guest_offset);
1539             ret = -EIO;
1540             goto out;
1541         }
1542 
1543         /* If a specific host_offset is required, check it */
1544         if (*host_offset != INV_OFFSET && cluster_offset != *host_offset) {
1545             *bytes = 0;
1546             ret = 0;
1547             goto out;
1548         }
1549 
1550         /* We keep all QCOW_OFLAG_COPIED clusters */
1551         keep_clusters = count_single_write_clusters(bs, nb_clusters, l2_slice,
1552                                                     l2_index, false);
1553         assert(keep_clusters <= nb_clusters);
1554 
1555         *bytes = MIN(*bytes,
1556                  keep_clusters * s->cluster_size
1557                  - offset_into_cluster(s, guest_offset));
1558         assert(*bytes != 0);
1559 
1560         ret = calculate_l2_meta(bs, cluster_offset, guest_offset,
1561                                 *bytes, l2_slice, m, true);
1562         if (ret < 0) {
1563             goto out;
1564         }
1565 
1566         ret = 1;
1567     } else {
1568         ret = 0;
1569     }
1570 
1571     /* Cleanup */
1572 out:
1573     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1574 
1575     /* Only return a host offset if we actually made progress. Otherwise we
1576      * would make requirements for handle_alloc() that it can't fulfill */
1577     if (ret > 0) {
1578         *host_offset = cluster_offset + offset_into_cluster(s, guest_offset);
1579     }
1580 
1581     return ret;
1582 }
1583 
1584 /*
1585  * Allocates new clusters for the given guest_offset.
1586  *
1587  * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1588  * contain the number of clusters that have been allocated and are contiguous
1589  * in the image file.
1590  *
1591  * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1592  * at which the new clusters must start. *nb_clusters can be 0 on return in
1593  * this case if the cluster at host_offset is already in use. If *host_offset
1594  * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1595  *
1596  * *host_offset is updated to contain the offset into the image file at which
1597  * the first allocated cluster starts.
1598  *
1599  * Return 0 on success and -errno in error cases. -EAGAIN means that the
1600  * function has been waiting for another request and the allocation must be
1601  * restarted, but the whole request should not be failed.
1602  */
1603 static int coroutine_fn GRAPH_RDLOCK
1604 do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1605                         uint64_t *host_offset, uint64_t *nb_clusters)
1606 {
1607     BDRVQcow2State *s = bs->opaque;
1608 
1609     trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1610                                          *host_offset, *nb_clusters);
1611 
1612     if (has_data_file(bs)) {
1613         assert(*host_offset == INV_OFFSET ||
1614                *host_offset == start_of_cluster(s, guest_offset));
1615         *host_offset = start_of_cluster(s, guest_offset);
1616         return 0;
1617     }
1618 
1619     /* Allocate new clusters */
1620     trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1621     if (*host_offset == INV_OFFSET) {
1622         int64_t cluster_offset =
1623             qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1624         if (cluster_offset < 0) {
1625             return cluster_offset;
1626         }
1627         *host_offset = cluster_offset;
1628         return 0;
1629     } else {
1630         int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1631         if (ret < 0) {
1632             return ret;
1633         }
1634         *nb_clusters = ret;
1635         return 0;
1636     }
1637 }
1638 
1639 /*
1640  * Allocates new clusters for an area that is either still unallocated or
1641  * cannot be overwritten in-place. If *host_offset is not INV_OFFSET,
1642  * clusters are only allocated if the new allocation can match the specified
1643  * host offset.
1644  *
1645  * Note that guest_offset may not be cluster aligned. In this case, the
1646  * returned *host_offset points to exact byte referenced by guest_offset and
1647  * therefore isn't cluster aligned as well.
1648  *
1649  * Returns:
1650  *   0:     if no clusters could be allocated. *bytes is set to 0,
1651  *          *host_offset is left unchanged.
1652  *
1653  *   1:     if new clusters were allocated. *bytes may be decreased if the
1654  *          new allocation doesn't cover all of the requested area.
1655  *          *host_offset is updated to contain the host offset of the first
1656  *          newly allocated cluster.
1657  *
1658  *  -errno: in error cases
1659  */
1660 static int coroutine_fn GRAPH_RDLOCK
1661 handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1662              uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1663 {
1664     BDRVQcow2State *s = bs->opaque;
1665     int l2_index;
1666     uint64_t *l2_slice;
1667     uint64_t nb_clusters;
1668     int ret;
1669 
1670     uint64_t alloc_cluster_offset;
1671 
1672     trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1673                              *bytes);
1674     assert(*bytes > 0);
1675 
1676     /*
1677      * Calculate the number of clusters to look for. We stop at L2 slice
1678      * boundaries to keep things simple.
1679      */
1680     nb_clusters =
1681         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1682 
1683     l2_index = offset_to_l2_slice_index(s, guest_offset);
1684     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1685     /* Limit total allocation byte count to BDRV_REQUEST_MAX_BYTES */
1686     nb_clusters = MIN(nb_clusters, BDRV_REQUEST_MAX_BYTES >> s->cluster_bits);
1687 
1688     /* Find L2 entry for the first involved cluster */
1689     ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1690     if (ret < 0) {
1691         return ret;
1692     }
1693 
1694     nb_clusters = count_single_write_clusters(bs, nb_clusters,
1695                                               l2_slice, l2_index, true);
1696 
1697     /* This function is only called when there were no non-COW clusters, so if
1698      * we can't find any unallocated or COW clusters either, something is
1699      * wrong with our code. */
1700     assert(nb_clusters > 0);
1701 
1702     /* Allocate at a given offset in the image file */
1703     alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1704         start_of_cluster(s, *host_offset);
1705     ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1706                                   &nb_clusters);
1707     if (ret < 0) {
1708         goto out;
1709     }
1710 
1711     /* Can't extend contiguous allocation */
1712     if (nb_clusters == 0) {
1713         *bytes = 0;
1714         ret = 0;
1715         goto out;
1716     }
1717 
1718     assert(alloc_cluster_offset != INV_OFFSET);
1719 
1720     /*
1721      * Save info needed for meta data update.
1722      *
1723      * requested_bytes: Number of bytes from the start of the first
1724      * newly allocated cluster to the end of the (possibly shortened
1725      * before) write request.
1726      *
1727      * avail_bytes: Number of bytes from the start of the first
1728      * newly allocated to the end of the last newly allocated cluster.
1729      *
1730      * nb_bytes: The number of bytes from the start of the first
1731      * newly allocated cluster to the end of the area that the write
1732      * request actually writes to (excluding COW at the end)
1733      */
1734     uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1735     int avail_bytes = nb_clusters << s->cluster_bits;
1736     int nb_bytes = MIN(requested_bytes, avail_bytes);
1737 
1738     *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1739     *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1740     assert(*bytes != 0);
1741 
1742     ret = calculate_l2_meta(bs, alloc_cluster_offset, guest_offset, *bytes,
1743                             l2_slice, m, false);
1744     if (ret < 0) {
1745         goto out;
1746     }
1747 
1748     ret = 1;
1749 
1750 out:
1751     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1752     return ret;
1753 }
1754 
1755 /*
1756  * For a given area on the virtual disk defined by @offset and @bytes,
1757  * find the corresponding area on the qcow2 image, allocating new
1758  * clusters (or subclusters) if necessary. The result can span a
1759  * combination of allocated and previously unallocated clusters.
1760  *
1761  * Note that offset may not be cluster aligned. In this case, the returned
1762  * *host_offset points to exact byte referenced by offset and therefore
1763  * isn't cluster aligned as well.
1764  *
1765  * On return, @host_offset is set to the beginning of the requested
1766  * area. This area is guaranteed to be contiguous on the qcow2 file
1767  * but it can be smaller than initially requested. In this case @bytes
1768  * is updated with the actual size.
1769  *
1770  * If any clusters or subclusters were allocated then @m contains a
1771  * list with the information of all the affected regions. Note that
1772  * this can happen regardless of whether this function succeeds or
1773  * not. The caller is responsible for updating the L2 metadata of the
1774  * allocated clusters (on success) or freeing them (on failure), and
1775  * for clearing the contents of @m afterwards in both cases.
1776  *
1777  * If the request conflicts with another write request in flight, the coroutine
1778  * is queued and will be reentered when the dependency has completed.
1779  *
1780  * Return 0 on success and -errno in error cases
1781  */
1782 int coroutine_fn qcow2_alloc_host_offset(BlockDriverState *bs, uint64_t offset,
1783                                          unsigned int *bytes,
1784                                          uint64_t *host_offset,
1785                                          QCowL2Meta **m)
1786 {
1787     BDRVQcow2State *s = bs->opaque;
1788     uint64_t start, remaining;
1789     uint64_t cluster_offset;
1790     uint64_t cur_bytes;
1791     int ret;
1792 
1793     trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1794 
1795 again:
1796     start = offset;
1797     remaining = *bytes;
1798     cluster_offset = INV_OFFSET;
1799     *host_offset = INV_OFFSET;
1800     cur_bytes = 0;
1801     *m = NULL;
1802 
1803     while (true) {
1804 
1805         if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1806             *host_offset = cluster_offset;
1807         }
1808 
1809         assert(remaining >= cur_bytes);
1810 
1811         start           += cur_bytes;
1812         remaining       -= cur_bytes;
1813 
1814         if (cluster_offset != INV_OFFSET) {
1815             cluster_offset += cur_bytes;
1816         }
1817 
1818         if (remaining == 0) {
1819             break;
1820         }
1821 
1822         cur_bytes = remaining;
1823 
1824         /*
1825          * Now start gathering as many contiguous clusters as possible:
1826          *
1827          * 1. Check for overlaps with in-flight allocations
1828          *
1829          *      a) Overlap not in the first cluster -> shorten this request and
1830          *         let the caller handle the rest in its next loop iteration.
1831          *
1832          *      b) Real overlaps of two requests. Yield and restart the search
1833          *         for contiguous clusters (the situation could have changed
1834          *         while we were sleeping)
1835          *
1836          *      c) TODO: Request starts in the same cluster as the in-flight
1837          *         allocation ends. Shorten the COW of the in-fight allocation,
1838          *         set cluster_offset to write to the same cluster and set up
1839          *         the right synchronisation between the in-flight request and
1840          *         the new one.
1841          */
1842         ret = handle_dependencies(bs, start, &cur_bytes, m);
1843         if (ret == -EAGAIN) {
1844             /* Currently handle_dependencies() doesn't yield if we already had
1845              * an allocation. If it did, we would have to clean up the L2Meta
1846              * structs before starting over. */
1847             assert(*m == NULL);
1848             goto again;
1849         } else if (ret < 0) {
1850             return ret;
1851         } else if (cur_bytes == 0) {
1852             break;
1853         } else {
1854             /* handle_dependencies() may have decreased cur_bytes (shortened
1855              * the allocations below) so that the next dependency is processed
1856              * correctly during the next loop iteration. */
1857         }
1858 
1859         /*
1860          * 2. Count contiguous COPIED clusters.
1861          */
1862         ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1863         if (ret < 0) {
1864             return ret;
1865         } else if (ret) {
1866             continue;
1867         } else if (cur_bytes == 0) {
1868             break;
1869         }
1870 
1871         /*
1872          * 3. If the request still hasn't completed, allocate new clusters,
1873          *    considering any cluster_offset of steps 1c or 2.
1874          */
1875         ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1876         if (ret < 0) {
1877             return ret;
1878         } else if (ret) {
1879             continue;
1880         } else {
1881             assert(cur_bytes == 0);
1882             break;
1883         }
1884     }
1885 
1886     *bytes -= remaining;
1887     assert(*bytes > 0);
1888     assert(*host_offset != INV_OFFSET);
1889     assert(offset_into_cluster(s, *host_offset) ==
1890            offset_into_cluster(s, offset));
1891 
1892     return 0;
1893 }
1894 
1895 /*
1896  * This discards as many clusters of nb_clusters as possible at once (i.e.
1897  * all clusters in the same L2 slice) and returns the number of discarded
1898  * clusters.
1899  */
1900 static int GRAPH_RDLOCK
1901 discard_in_l2_slice(BlockDriverState *bs, uint64_t offset, uint64_t nb_clusters,
1902                     enum qcow2_discard_type type, bool full_discard)
1903 {
1904     BDRVQcow2State *s = bs->opaque;
1905     uint64_t *l2_slice;
1906     int l2_index;
1907     int ret;
1908     int i;
1909 
1910     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1911     if (ret < 0) {
1912         return ret;
1913     }
1914 
1915     /* Limit nb_clusters to one L2 slice */
1916     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1917     assert(nb_clusters <= INT_MAX);
1918 
1919     for (i = 0; i < nb_clusters; i++) {
1920         uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
1921         uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
1922         uint64_t new_l2_entry = old_l2_entry;
1923         uint64_t new_l2_bitmap = old_l2_bitmap;
1924         QCow2ClusterType cluster_type =
1925             qcow2_get_cluster_type(bs, old_l2_entry);
1926         bool keep_reference = (cluster_type != QCOW2_CLUSTER_COMPRESSED) &&
1927                               !full_discard &&
1928                               (s->discard_no_unref &&
1929                                type == QCOW2_DISCARD_REQUEST);
1930 
1931         /*
1932          * If full_discard is true, the cluster should not read back as zeroes,
1933          * but rather fall through to the backing file.
1934          *
1935          * If full_discard is false, make sure that a discarded area reads back
1936          * as zeroes for v3 images (we cannot do it for v2 without actually
1937          * writing a zero-filled buffer). We can skip the operation if the
1938          * cluster is already marked as zero, or if it's unallocated and we
1939          * don't have a backing file.
1940          *
1941          * TODO We might want to use bdrv_block_status(bs) here, but we're
1942          * holding s->lock, so that doesn't work today.
1943          */
1944         if (full_discard) {
1945             new_l2_entry = new_l2_bitmap = 0;
1946         } else if (bs->backing || qcow2_cluster_is_allocated(cluster_type)) {
1947             if (has_subclusters(s)) {
1948                 if (keep_reference) {
1949                     new_l2_entry = old_l2_entry;
1950                 } else {
1951                     new_l2_entry = 0;
1952                 }
1953                 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
1954             } else {
1955                 if (s->qcow_version >= 3) {
1956                     if (keep_reference) {
1957                         new_l2_entry |= QCOW_OFLAG_ZERO;
1958                     } else {
1959                         new_l2_entry = QCOW_OFLAG_ZERO;
1960                     }
1961                 } else {
1962                     new_l2_entry = 0;
1963                 }
1964             }
1965         }
1966 
1967         if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
1968             continue;
1969         }
1970 
1971         /* First remove L2 entries */
1972         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1973         set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
1974         if (has_subclusters(s)) {
1975             set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
1976         }
1977         if (!keep_reference) {
1978             /* Then decrease the refcount */
1979             qcow2_free_any_cluster(bs, old_l2_entry, type);
1980         } else if (s->discard_passthrough[type] &&
1981                    (cluster_type == QCOW2_CLUSTER_NORMAL ||
1982                     cluster_type == QCOW2_CLUSTER_ZERO_ALLOC)) {
1983             /* If we keep the reference, pass on the discard still */
1984             bdrv_pdiscard(s->data_file, old_l2_entry & L2E_OFFSET_MASK,
1985                           s->cluster_size);
1986         }
1987     }
1988 
1989     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1990 
1991     return nb_clusters;
1992 }
1993 
1994 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1995                           uint64_t bytes, enum qcow2_discard_type type,
1996                           bool full_discard)
1997 {
1998     BDRVQcow2State *s = bs->opaque;
1999     uint64_t end_offset = offset + bytes;
2000     uint64_t nb_clusters;
2001     int64_t cleared;
2002     int ret;
2003 
2004     /* Caller must pass aligned values, except at image end */
2005     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
2006     assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
2007            end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
2008 
2009     nb_clusters = size_to_clusters(s, bytes);
2010 
2011     s->cache_discards = true;
2012 
2013     /* Each L2 slice is handled by its own loop iteration */
2014     while (nb_clusters > 0) {
2015         cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
2016                                       full_discard);
2017         if (cleared < 0) {
2018             ret = cleared;
2019             goto fail;
2020         }
2021 
2022         nb_clusters -= cleared;
2023         offset += (cleared * s->cluster_size);
2024     }
2025 
2026     ret = 0;
2027 fail:
2028     s->cache_discards = false;
2029     qcow2_process_discards(bs, ret);
2030 
2031     return ret;
2032 }
2033 
2034 /*
2035  * This zeroes as many clusters of nb_clusters as possible at once (i.e.
2036  * all clusters in the same L2 slice) and returns the number of zeroed
2037  * clusters.
2038  */
2039 static int coroutine_fn GRAPH_RDLOCK
2040 zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
2041                  uint64_t nb_clusters, int flags)
2042 {
2043     BDRVQcow2State *s = bs->opaque;
2044     uint64_t *l2_slice;
2045     int l2_index;
2046     int ret;
2047     int i;
2048 
2049     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2050     if (ret < 0) {
2051         return ret;
2052     }
2053 
2054     /* Limit nb_clusters to one L2 slice */
2055     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
2056     assert(nb_clusters <= INT_MAX);
2057 
2058     for (i = 0; i < nb_clusters; i++) {
2059         uint64_t old_l2_entry = get_l2_entry(s, l2_slice, l2_index + i);
2060         uint64_t old_l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index + i);
2061         QCow2ClusterType type = qcow2_get_cluster_type(bs, old_l2_entry);
2062         bool unmap = (type == QCOW2_CLUSTER_COMPRESSED) ||
2063             ((flags & BDRV_REQ_MAY_UNMAP) && qcow2_cluster_is_allocated(type));
2064         bool keep_reference =
2065             (s->discard_no_unref && type != QCOW2_CLUSTER_COMPRESSED);
2066         uint64_t new_l2_entry = old_l2_entry;
2067         uint64_t new_l2_bitmap = old_l2_bitmap;
2068 
2069         if (unmap && !keep_reference) {
2070             new_l2_entry = 0;
2071         }
2072 
2073         if (has_subclusters(s)) {
2074             new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
2075         } else {
2076             new_l2_entry |= QCOW_OFLAG_ZERO;
2077         }
2078 
2079         if (old_l2_entry == new_l2_entry && old_l2_bitmap == new_l2_bitmap) {
2080             continue;
2081         }
2082 
2083         /* First update L2 entries */
2084         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2085         set_l2_entry(s, l2_slice, l2_index + i, new_l2_entry);
2086         if (has_subclusters(s)) {
2087             set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
2088         }
2089 
2090         if (unmap) {
2091             if (!keep_reference) {
2092                 /* Then decrease the refcount */
2093                 qcow2_free_any_cluster(bs, old_l2_entry, QCOW2_DISCARD_REQUEST);
2094             } else if (s->discard_passthrough[QCOW2_DISCARD_REQUEST] &&
2095                        (type == QCOW2_CLUSTER_NORMAL ||
2096                         type == QCOW2_CLUSTER_ZERO_ALLOC)) {
2097                 /* If we keep the reference, pass on the discard still */
2098                 bdrv_pdiscard(s->data_file, old_l2_entry & L2E_OFFSET_MASK,
2099                             s->cluster_size);
2100             }
2101         }
2102     }
2103 
2104     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2105 
2106     return nb_clusters;
2107 }
2108 
2109 static int coroutine_fn GRAPH_RDLOCK
2110 zero_l2_subclusters(BlockDriverState *bs, uint64_t offset,
2111                     unsigned nb_subclusters)
2112 {
2113     BDRVQcow2State *s = bs->opaque;
2114     uint64_t *l2_slice;
2115     uint64_t old_l2_bitmap, l2_bitmap;
2116     int l2_index, ret, sc = offset_to_sc_index(s, offset);
2117 
2118     /* For full clusters use zero_in_l2_slice() instead */
2119     assert(nb_subclusters > 0 && nb_subclusters < s->subclusters_per_cluster);
2120     assert(sc + nb_subclusters <= s->subclusters_per_cluster);
2121     assert(offset_into_subcluster(s, offset) == 0);
2122 
2123     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
2124     if (ret < 0) {
2125         return ret;
2126     }
2127 
2128     switch (qcow2_get_cluster_type(bs, get_l2_entry(s, l2_slice, l2_index))) {
2129     case QCOW2_CLUSTER_COMPRESSED:
2130         ret = -ENOTSUP; /* We cannot partially zeroize compressed clusters */
2131         goto out;
2132     case QCOW2_CLUSTER_NORMAL:
2133     case QCOW2_CLUSTER_UNALLOCATED:
2134         break;
2135     default:
2136         g_assert_not_reached();
2137     }
2138 
2139     old_l2_bitmap = l2_bitmap = get_l2_bitmap(s, l2_slice, l2_index);
2140 
2141     l2_bitmap |=  QCOW_OFLAG_SUB_ZERO_RANGE(sc, sc + nb_subclusters);
2142     l2_bitmap &= ~QCOW_OFLAG_SUB_ALLOC_RANGE(sc, sc + nb_subclusters);
2143 
2144     if (old_l2_bitmap != l2_bitmap) {
2145         set_l2_bitmap(s, l2_slice, l2_index, l2_bitmap);
2146         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2147     }
2148 
2149     ret = 0;
2150 out:
2151     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2152 
2153     return ret;
2154 }
2155 
2156 int coroutine_fn qcow2_subcluster_zeroize(BlockDriverState *bs, uint64_t offset,
2157                                           uint64_t bytes, int flags)
2158 {
2159     BDRVQcow2State *s = bs->opaque;
2160     uint64_t end_offset = offset + bytes;
2161     uint64_t nb_clusters;
2162     unsigned head, tail;
2163     int64_t cleared;
2164     int ret;
2165 
2166     /* If we have to stay in sync with an external data file, zero out
2167      * s->data_file first. */
2168     if (data_file_is_raw(bs)) {
2169         assert(has_data_file(bs));
2170         ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
2171         if (ret < 0) {
2172             return ret;
2173         }
2174     }
2175 
2176     /* Caller must pass aligned values, except at image end */
2177     assert(offset_into_subcluster(s, offset) == 0);
2178     assert(offset_into_subcluster(s, end_offset) == 0 ||
2179            end_offset >= bs->total_sectors << BDRV_SECTOR_BITS);
2180 
2181     /*
2182      * The zero flag is only supported by version 3 and newer. However, if we
2183      * have no backing file, we can resort to discard in version 2.
2184      */
2185     if (s->qcow_version < 3) {
2186         if (!bs->backing) {
2187             return qcow2_cluster_discard(bs, offset, bytes,
2188                                          QCOW2_DISCARD_REQUEST, false);
2189         }
2190         return -ENOTSUP;
2191     }
2192 
2193     head = MIN(end_offset, ROUND_UP(offset, s->cluster_size)) - offset;
2194     offset += head;
2195 
2196     tail = (end_offset >= bs->total_sectors << BDRV_SECTOR_BITS) ? 0 :
2197         end_offset - MAX(offset, start_of_cluster(s, end_offset));
2198     end_offset -= tail;
2199 
2200     s->cache_discards = true;
2201 
2202     if (head) {
2203         ret = zero_l2_subclusters(bs, offset - head,
2204                                   size_to_subclusters(s, head));
2205         if (ret < 0) {
2206             goto fail;
2207         }
2208     }
2209 
2210     /* Each L2 slice is handled by its own loop iteration */
2211     nb_clusters = size_to_clusters(s, end_offset - offset);
2212 
2213     while (nb_clusters > 0) {
2214         cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
2215         if (cleared < 0) {
2216             ret = cleared;
2217             goto fail;
2218         }
2219 
2220         nb_clusters -= cleared;
2221         offset += (cleared * s->cluster_size);
2222     }
2223 
2224     if (tail) {
2225         ret = zero_l2_subclusters(bs, end_offset, size_to_subclusters(s, tail));
2226         if (ret < 0) {
2227             goto fail;
2228         }
2229     }
2230 
2231     ret = 0;
2232 fail:
2233     s->cache_discards = false;
2234     qcow2_process_discards(bs, ret);
2235 
2236     return ret;
2237 }
2238 
2239 /*
2240  * Expands all zero clusters in a specific L1 table (or deallocates them, for
2241  * non-backed non-pre-allocated zero clusters).
2242  *
2243  * l1_entries and *visited_l1_entries are used to keep track of progress for
2244  * status_cb(). l1_entries contains the total number of L1 entries and
2245  * *visited_l1_entries counts all visited L1 entries.
2246  */
2247 static int GRAPH_RDLOCK
2248 expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
2249                            int l1_size, int64_t *visited_l1_entries,
2250                            int64_t l1_entries,
2251                            BlockDriverAmendStatusCB *status_cb,
2252                            void *cb_opaque)
2253 {
2254     BDRVQcow2State *s = bs->opaque;
2255     bool is_active_l1 = (l1_table == s->l1_table);
2256     uint64_t *l2_slice = NULL;
2257     unsigned slice, slice_size2, n_slices;
2258     int ret;
2259     int i, j;
2260 
2261     /* qcow2_downgrade() is not allowed in images with subclusters */
2262     assert(!has_subclusters(s));
2263 
2264     slice_size2 = s->l2_slice_size * l2_entry_size(s);
2265     n_slices = s->cluster_size / slice_size2;
2266 
2267     if (!is_active_l1) {
2268         /* inactive L2 tables require a buffer to be stored in when loading
2269          * them from disk */
2270         l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
2271         if (l2_slice == NULL) {
2272             return -ENOMEM;
2273         }
2274     }
2275 
2276     for (i = 0; i < l1_size; i++) {
2277         uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
2278         uint64_t l2_refcount;
2279 
2280         if (!l2_offset) {
2281             /* unallocated */
2282             (*visited_l1_entries)++;
2283             if (status_cb) {
2284                 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2285             }
2286             continue;
2287         }
2288 
2289         if (offset_into_cluster(s, l2_offset)) {
2290             qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
2291                                     PRIx64 " unaligned (L1 index: %#x)",
2292                                     l2_offset, i);
2293             ret = -EIO;
2294             goto fail;
2295         }
2296 
2297         ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
2298                                  &l2_refcount);
2299         if (ret < 0) {
2300             goto fail;
2301         }
2302 
2303         for (slice = 0; slice < n_slices; slice++) {
2304             uint64_t slice_offset = l2_offset + slice * slice_size2;
2305             bool l2_dirty = false;
2306             if (is_active_l1) {
2307                 /* get active L2 tables from cache */
2308                 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
2309                                       (void **)&l2_slice);
2310             } else {
2311                 /* load inactive L2 tables from disk */
2312                 ret = bdrv_pread(bs->file, slice_offset, slice_size2,
2313                                  l2_slice, 0);
2314             }
2315             if (ret < 0) {
2316                 goto fail;
2317             }
2318 
2319             for (j = 0; j < s->l2_slice_size; j++) {
2320                 uint64_t l2_entry = get_l2_entry(s, l2_slice, j);
2321                 int64_t offset = l2_entry & L2E_OFFSET_MASK;
2322                 QCow2ClusterType cluster_type =
2323                     qcow2_get_cluster_type(bs, l2_entry);
2324 
2325                 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
2326                     cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
2327                     continue;
2328                 }
2329 
2330                 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2331                     if (!bs->backing) {
2332                         /*
2333                          * not backed; therefore we can simply deallocate the
2334                          * cluster. No need to call set_l2_bitmap(), this
2335                          * function doesn't support images with subclusters.
2336                          */
2337                         set_l2_entry(s, l2_slice, j, 0);
2338                         l2_dirty = true;
2339                         continue;
2340                     }
2341 
2342                     offset = qcow2_alloc_clusters(bs, s->cluster_size);
2343                     if (offset < 0) {
2344                         ret = offset;
2345                         goto fail;
2346                     }
2347 
2348                     /* The offset must fit in the offset field */
2349                     assert((offset & L2E_OFFSET_MASK) == offset);
2350 
2351                     if (l2_refcount > 1) {
2352                         /* For shared L2 tables, set the refcount accordingly
2353                          * (it is already 1 and needs to be l2_refcount) */
2354                         ret = qcow2_update_cluster_refcount(
2355                             bs, offset >> s->cluster_bits,
2356                             refcount_diff(1, l2_refcount), false,
2357                             QCOW2_DISCARD_OTHER);
2358                         if (ret < 0) {
2359                             qcow2_free_clusters(bs, offset, s->cluster_size,
2360                                                 QCOW2_DISCARD_OTHER);
2361                             goto fail;
2362                         }
2363                     }
2364                 }
2365 
2366                 if (offset_into_cluster(s, offset)) {
2367                     int l2_index = slice * s->l2_slice_size + j;
2368                     qcow2_signal_corruption(
2369                         bs, true, -1, -1,
2370                         "Cluster allocation offset "
2371                         "%#" PRIx64 " unaligned (L2 offset: %#"
2372                         PRIx64 ", L2 index: %#x)", offset,
2373                         l2_offset, l2_index);
2374                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2375                         qcow2_free_clusters(bs, offset, s->cluster_size,
2376                                             QCOW2_DISCARD_ALWAYS);
2377                     }
2378                     ret = -EIO;
2379                     goto fail;
2380                 }
2381 
2382                 ret = qcow2_pre_write_overlap_check(bs, 0, offset,
2383                                                     s->cluster_size, true);
2384                 if (ret < 0) {
2385                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2386                         qcow2_free_clusters(bs, offset, s->cluster_size,
2387                                             QCOW2_DISCARD_ALWAYS);
2388                     }
2389                     goto fail;
2390                 }
2391 
2392                 ret = bdrv_pwrite_zeroes(s->data_file, offset,
2393                                          s->cluster_size, 0);
2394                 if (ret < 0) {
2395                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
2396                         qcow2_free_clusters(bs, offset, s->cluster_size,
2397                                             QCOW2_DISCARD_ALWAYS);
2398                     }
2399                     goto fail;
2400                 }
2401 
2402                 if (l2_refcount == 1) {
2403                     set_l2_entry(s, l2_slice, j, offset | QCOW_OFLAG_COPIED);
2404                 } else {
2405                     set_l2_entry(s, l2_slice, j, offset);
2406                 }
2407                 /*
2408                  * No need to call set_l2_bitmap() after set_l2_entry() because
2409                  * this function doesn't support images with subclusters.
2410                  */
2411                 l2_dirty = true;
2412             }
2413 
2414             if (is_active_l1) {
2415                 if (l2_dirty) {
2416                     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
2417                     qcow2_cache_depends_on_flush(s->l2_table_cache);
2418                 }
2419                 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2420             } else {
2421                 if (l2_dirty) {
2422                     ret = qcow2_pre_write_overlap_check(
2423                         bs, QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2,
2424                         slice_offset, slice_size2, false);
2425                     if (ret < 0) {
2426                         goto fail;
2427                     }
2428 
2429                     ret = bdrv_pwrite(bs->file, slice_offset, slice_size2,
2430                                       l2_slice, 0);
2431                     if (ret < 0) {
2432                         goto fail;
2433                     }
2434                 }
2435             }
2436         }
2437 
2438         (*visited_l1_entries)++;
2439         if (status_cb) {
2440             status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2441         }
2442     }
2443 
2444     ret = 0;
2445 
2446 fail:
2447     if (l2_slice) {
2448         if (!is_active_l1) {
2449             qemu_vfree(l2_slice);
2450         } else {
2451             qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2452         }
2453     }
2454     return ret;
2455 }
2456 
2457 /*
2458  * For backed images, expands all zero clusters on the image. For non-backed
2459  * images, deallocates all non-pre-allocated zero clusters (and claims the
2460  * allocation for pre-allocated ones). This is important for downgrading to a
2461  * qcow2 version which doesn't yet support metadata zero clusters.
2462  */
2463 int qcow2_expand_zero_clusters(BlockDriverState *bs,
2464                                BlockDriverAmendStatusCB *status_cb,
2465                                void *cb_opaque)
2466 {
2467     BDRVQcow2State *s = bs->opaque;
2468     uint64_t *l1_table = NULL;
2469     int64_t l1_entries = 0, visited_l1_entries = 0;
2470     int ret;
2471     int i, j;
2472 
2473     if (status_cb) {
2474         l1_entries = s->l1_size;
2475         for (i = 0; i < s->nb_snapshots; i++) {
2476             l1_entries += s->snapshots[i].l1_size;
2477         }
2478     }
2479 
2480     ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
2481                                      &visited_l1_entries, l1_entries,
2482                                      status_cb, cb_opaque);
2483     if (ret < 0) {
2484         goto fail;
2485     }
2486 
2487     /* Inactive L1 tables may point to active L2 tables - therefore it is
2488      * necessary to flush the L2 table cache before trying to access the L2
2489      * tables pointed to by inactive L1 entries (else we might try to expand
2490      * zero clusters that have already been expanded); furthermore, it is also
2491      * necessary to empty the L2 table cache, since it may contain tables which
2492      * are now going to be modified directly on disk, bypassing the cache.
2493      * qcow2_cache_empty() does both for us. */
2494     ret = qcow2_cache_empty(bs, s->l2_table_cache);
2495     if (ret < 0) {
2496         goto fail;
2497     }
2498 
2499     for (i = 0; i < s->nb_snapshots; i++) {
2500         int l1_size2;
2501         uint64_t *new_l1_table;
2502         Error *local_err = NULL;
2503 
2504         ret = qcow2_validate_table(bs, s->snapshots[i].l1_table_offset,
2505                                    s->snapshots[i].l1_size, L1E_SIZE,
2506                                    QCOW_MAX_L1_SIZE, "Snapshot L1 table",
2507                                    &local_err);
2508         if (ret < 0) {
2509             error_report_err(local_err);
2510             goto fail;
2511         }
2512 
2513         l1_size2 = s->snapshots[i].l1_size * L1E_SIZE;
2514         new_l1_table = g_try_realloc(l1_table, l1_size2);
2515 
2516         if (!new_l1_table) {
2517             ret = -ENOMEM;
2518             goto fail;
2519         }
2520 
2521         l1_table = new_l1_table;
2522 
2523         ret = bdrv_pread(bs->file, s->snapshots[i].l1_table_offset, l1_size2,
2524                          l1_table, 0);
2525         if (ret < 0) {
2526             goto fail;
2527         }
2528 
2529         for (j = 0; j < s->snapshots[i].l1_size; j++) {
2530             be64_to_cpus(&l1_table[j]);
2531         }
2532 
2533         ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
2534                                          &visited_l1_entries, l1_entries,
2535                                          status_cb, cb_opaque);
2536         if (ret < 0) {
2537             goto fail;
2538         }
2539     }
2540 
2541     ret = 0;
2542 
2543 fail:
2544     g_free(l1_table);
2545     return ret;
2546 }
2547 
2548 void qcow2_parse_compressed_l2_entry(BlockDriverState *bs, uint64_t l2_entry,
2549                                      uint64_t *coffset, int *csize)
2550 {
2551     BDRVQcow2State *s = bs->opaque;
2552     int nb_csectors;
2553 
2554     assert(qcow2_get_cluster_type(bs, l2_entry) == QCOW2_CLUSTER_COMPRESSED);
2555 
2556     *coffset = l2_entry & s->cluster_offset_mask;
2557 
2558     nb_csectors = ((l2_entry >> s->csize_shift) & s->csize_mask) + 1;
2559     *csize = nb_csectors * QCOW2_COMPRESSED_SECTOR_SIZE -
2560         (*coffset & (QCOW2_COMPRESSED_SECTOR_SIZE - 1));
2561 }
2562