xref: /openbmc/qemu/block/qcow2-cluster.c (revision 0c4e9931)
1 /*
2  * Block driver for the QCOW version 2 format
3  *
4  * Copyright (c) 2004-2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include <zlib.h>
27 
28 #include "qapi/error.h"
29 #include "qcow2.h"
30 #include "qemu/bswap.h"
31 #include "trace.h"
32 
33 int qcow2_shrink_l1_table(BlockDriverState *bs, uint64_t exact_size)
34 {
35     BDRVQcow2State *s = bs->opaque;
36     int new_l1_size, i, ret;
37 
38     if (exact_size >= s->l1_size) {
39         return 0;
40     }
41 
42     new_l1_size = exact_size;
43 
44 #ifdef DEBUG_ALLOC2
45     fprintf(stderr, "shrink l1_table from %d to %d\n", s->l1_size, new_l1_size);
46 #endif
47 
48     BLKDBG_EVENT(bs->file, BLKDBG_L1_SHRINK_WRITE_TABLE);
49     ret = bdrv_pwrite_zeroes(bs->file, s->l1_table_offset +
50                                        new_l1_size * sizeof(uint64_t),
51                              (s->l1_size - new_l1_size) * sizeof(uint64_t), 0);
52     if (ret < 0) {
53         goto fail;
54     }
55 
56     ret = bdrv_flush(bs->file->bs);
57     if (ret < 0) {
58         goto fail;
59     }
60 
61     BLKDBG_EVENT(bs->file, BLKDBG_L1_SHRINK_FREE_L2_CLUSTERS);
62     for (i = s->l1_size - 1; i > new_l1_size - 1; i--) {
63         if ((s->l1_table[i] & L1E_OFFSET_MASK) == 0) {
64             continue;
65         }
66         qcow2_free_clusters(bs, s->l1_table[i] & L1E_OFFSET_MASK,
67                             s->cluster_size, QCOW2_DISCARD_ALWAYS);
68         s->l1_table[i] = 0;
69     }
70     return 0;
71 
72 fail:
73     /*
74      * If the write in the l1_table failed the image may contain a partially
75      * overwritten l1_table. In this case it would be better to clear the
76      * l1_table in memory to avoid possible image corruption.
77      */
78     memset(s->l1_table + new_l1_size, 0,
79            (s->l1_size - new_l1_size) * sizeof(uint64_t));
80     return ret;
81 }
82 
83 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
84                         bool exact_size)
85 {
86     BDRVQcow2State *s = bs->opaque;
87     int new_l1_size2, ret, i;
88     uint64_t *new_l1_table;
89     int64_t old_l1_table_offset, old_l1_size;
90     int64_t new_l1_table_offset, new_l1_size;
91     uint8_t data[12];
92 
93     if (min_size <= s->l1_size)
94         return 0;
95 
96     /* Do a sanity check on min_size before trying to calculate new_l1_size
97      * (this prevents overflows during the while loop for the calculation of
98      * new_l1_size) */
99     if (min_size > INT_MAX / sizeof(uint64_t)) {
100         return -EFBIG;
101     }
102 
103     if (exact_size) {
104         new_l1_size = min_size;
105     } else {
106         /* Bump size up to reduce the number of times we have to grow */
107         new_l1_size = s->l1_size;
108         if (new_l1_size == 0) {
109             new_l1_size = 1;
110         }
111         while (min_size > new_l1_size) {
112             new_l1_size = DIV_ROUND_UP(new_l1_size * 3, 2);
113         }
114     }
115 
116     QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX);
117     if (new_l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) {
118         return -EFBIG;
119     }
120 
121 #ifdef DEBUG_ALLOC2
122     fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
123             s->l1_size, new_l1_size);
124 #endif
125 
126     new_l1_size2 = sizeof(uint64_t) * new_l1_size;
127     new_l1_table = qemu_try_blockalign(bs->file->bs, new_l1_size2);
128     if (new_l1_table == NULL) {
129         return -ENOMEM;
130     }
131     memset(new_l1_table, 0, new_l1_size2);
132 
133     if (s->l1_size) {
134         memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
135     }
136 
137     /* write new table (align to cluster) */
138     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
139     new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
140     if (new_l1_table_offset < 0) {
141         qemu_vfree(new_l1_table);
142         return new_l1_table_offset;
143     }
144 
145     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
146     if (ret < 0) {
147         goto fail;
148     }
149 
150     /* the L1 position has not yet been updated, so these clusters must
151      * indeed be completely free */
152     ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
153                                         new_l1_size2, false);
154     if (ret < 0) {
155         goto fail;
156     }
157 
158     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
159     for(i = 0; i < s->l1_size; i++)
160         new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
161     ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset,
162                            new_l1_table, new_l1_size2);
163     if (ret < 0)
164         goto fail;
165     for(i = 0; i < s->l1_size; i++)
166         new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
167 
168     /* set new table */
169     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
170     stl_be_p(data, new_l1_size);
171     stq_be_p(data + 4, new_l1_table_offset);
172     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size),
173                            data, sizeof(data));
174     if (ret < 0) {
175         goto fail;
176     }
177     qemu_vfree(s->l1_table);
178     old_l1_table_offset = s->l1_table_offset;
179     s->l1_table_offset = new_l1_table_offset;
180     s->l1_table = new_l1_table;
181     old_l1_size = s->l1_size;
182     s->l1_size = new_l1_size;
183     qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
184                         QCOW2_DISCARD_OTHER);
185     return 0;
186  fail:
187     qemu_vfree(new_l1_table);
188     qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
189                         QCOW2_DISCARD_OTHER);
190     return ret;
191 }
192 
193 /*
194  * l2_load
195  *
196  * @bs: The BlockDriverState
197  * @offset: A guest offset, used to calculate what slice of the L2
198  *          table to load.
199  * @l2_offset: Offset to the L2 table in the image file.
200  * @l2_slice: Location to store the pointer to the L2 slice.
201  *
202  * Loads a L2 slice into memory (L2 slices are the parts of L2 tables
203  * that are loaded by the qcow2 cache). If the slice is in the cache,
204  * the cache is used; otherwise the L2 slice is loaded from the image
205  * file.
206  */
207 static int l2_load(BlockDriverState *bs, uint64_t offset,
208                    uint64_t l2_offset, uint64_t **l2_slice)
209 {
210     BDRVQcow2State *s = bs->opaque;
211     int start_of_slice = sizeof(uint64_t) *
212         (offset_to_l2_index(s, offset) - offset_to_l2_slice_index(s, offset));
213 
214     return qcow2_cache_get(bs, s->l2_table_cache, l2_offset + start_of_slice,
215                            (void **)l2_slice);
216 }
217 
218 /*
219  * Writes an L1 entry to disk (note that depending on the alignment
220  * requirements this function may write more that just one entry in
221  * order to prevent bdrv_pwrite from performing a read-modify-write)
222  */
223 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
224 {
225     BDRVQcow2State *s = bs->opaque;
226     int l1_start_index;
227     int i, ret;
228     int bufsize = MAX(sizeof(uint64_t),
229                       MIN(bs->file->bs->bl.request_alignment, s->cluster_size));
230     int nentries = bufsize / sizeof(uint64_t);
231     g_autofree uint64_t *buf = g_try_new0(uint64_t, nentries);
232 
233     if (buf == NULL) {
234         return -ENOMEM;
235     }
236 
237     l1_start_index = QEMU_ALIGN_DOWN(l1_index, nentries);
238     for (i = 0; i < MIN(nentries, s->l1_size - l1_start_index); i++) {
239         buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
240     }
241 
242     ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
243             s->l1_table_offset + 8 * l1_start_index, bufsize, false);
244     if (ret < 0) {
245         return ret;
246     }
247 
248     BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
249     ret = bdrv_pwrite_sync(bs->file,
250                            s->l1_table_offset + 8 * l1_start_index,
251                            buf, bufsize);
252     if (ret < 0) {
253         return ret;
254     }
255 
256     return 0;
257 }
258 
259 /*
260  * l2_allocate
261  *
262  * Allocate a new l2 entry in the file. If l1_index points to an already
263  * used entry in the L2 table (i.e. we are doing a copy on write for the L2
264  * table) copy the contents of the old L2 table into the newly allocated one.
265  * Otherwise the new table is initialized with zeros.
266  *
267  */
268 
269 static int l2_allocate(BlockDriverState *bs, int l1_index)
270 {
271     BDRVQcow2State *s = bs->opaque;
272     uint64_t old_l2_offset;
273     uint64_t *l2_slice = NULL;
274     unsigned slice, slice_size2, n_slices;
275     int64_t l2_offset;
276     int ret;
277 
278     old_l2_offset = s->l1_table[l1_index];
279 
280     trace_qcow2_l2_allocate(bs, l1_index);
281 
282     /* allocate a new l2 entry */
283 
284     l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
285     if (l2_offset < 0) {
286         ret = l2_offset;
287         goto fail;
288     }
289 
290     /* The offset must fit in the offset field of the L1 table entry */
291     assert((l2_offset & L1E_OFFSET_MASK) == l2_offset);
292 
293     /* If we're allocating the table at offset 0 then something is wrong */
294     if (l2_offset == 0) {
295         qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
296                                 "allocation of L2 table at offset 0");
297         ret = -EIO;
298         goto fail;
299     }
300 
301     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
302     if (ret < 0) {
303         goto fail;
304     }
305 
306     /* allocate a new entry in the l2 cache */
307 
308     slice_size2 = s->l2_slice_size * sizeof(uint64_t);
309     n_slices = s->cluster_size / slice_size2;
310 
311     trace_qcow2_l2_allocate_get_empty(bs, l1_index);
312     for (slice = 0; slice < n_slices; slice++) {
313         ret = qcow2_cache_get_empty(bs, s->l2_table_cache,
314                                     l2_offset + slice * slice_size2,
315                                     (void **) &l2_slice);
316         if (ret < 0) {
317             goto fail;
318         }
319 
320         if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
321             /* if there was no old l2 table, clear the new slice */
322             memset(l2_slice, 0, slice_size2);
323         } else {
324             uint64_t *old_slice;
325             uint64_t old_l2_slice_offset =
326                 (old_l2_offset & L1E_OFFSET_MASK) + slice * slice_size2;
327 
328             /* if there was an old l2 table, read a slice from the disk */
329             BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
330             ret = qcow2_cache_get(bs, s->l2_table_cache, old_l2_slice_offset,
331                                   (void **) &old_slice);
332             if (ret < 0) {
333                 goto fail;
334             }
335 
336             memcpy(l2_slice, old_slice, slice_size2);
337 
338             qcow2_cache_put(s->l2_table_cache, (void **) &old_slice);
339         }
340 
341         /* write the l2 slice to the file */
342         BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
343 
344         trace_qcow2_l2_allocate_write_l2(bs, l1_index);
345         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
346         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
347     }
348 
349     ret = qcow2_cache_flush(bs, s->l2_table_cache);
350     if (ret < 0) {
351         goto fail;
352     }
353 
354     /* update the L1 entry */
355     trace_qcow2_l2_allocate_write_l1(bs, l1_index);
356     s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
357     ret = qcow2_write_l1_entry(bs, l1_index);
358     if (ret < 0) {
359         goto fail;
360     }
361 
362     trace_qcow2_l2_allocate_done(bs, l1_index, 0);
363     return 0;
364 
365 fail:
366     trace_qcow2_l2_allocate_done(bs, l1_index, ret);
367     if (l2_slice != NULL) {
368         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
369     }
370     s->l1_table[l1_index] = old_l2_offset;
371     if (l2_offset > 0) {
372         qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
373                             QCOW2_DISCARD_ALWAYS);
374     }
375     return ret;
376 }
377 
378 /*
379  * Checks how many clusters in a given L2 slice are contiguous in the image
380  * file. As soon as one of the flags in the bitmask stop_flags changes compared
381  * to the first cluster, the search is stopped and the cluster is not counted
382  * as contiguous. (This allows it, for example, to stop at the first compressed
383  * cluster which may require a different handling)
384  */
385 static int count_contiguous_clusters(BlockDriverState *bs, int nb_clusters,
386         int cluster_size, uint64_t *l2_slice, uint64_t stop_flags)
387 {
388     int i;
389     QCow2ClusterType first_cluster_type;
390     uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
391     uint64_t first_entry = be64_to_cpu(l2_slice[0]);
392     uint64_t offset = first_entry & mask;
393 
394     first_cluster_type = qcow2_get_cluster_type(bs, first_entry);
395     if (first_cluster_type == QCOW2_CLUSTER_UNALLOCATED) {
396         return 0;
397     }
398 
399     /* must be allocated */
400     assert(first_cluster_type == QCOW2_CLUSTER_NORMAL ||
401            first_cluster_type == QCOW2_CLUSTER_ZERO_ALLOC);
402 
403     for (i = 0; i < nb_clusters; i++) {
404         uint64_t l2_entry = be64_to_cpu(l2_slice[i]) & mask;
405         if (offset + (uint64_t) i * cluster_size != l2_entry) {
406             break;
407         }
408     }
409 
410         return i;
411 }
412 
413 /*
414  * Checks how many consecutive unallocated clusters in a given L2
415  * slice have the same cluster type.
416  */
417 static int count_contiguous_clusters_unallocated(BlockDriverState *bs,
418                                                  int nb_clusters,
419                                                  uint64_t *l2_slice,
420                                                  QCow2ClusterType wanted_type)
421 {
422     int i;
423 
424     assert(wanted_type == QCOW2_CLUSTER_ZERO_PLAIN ||
425            wanted_type == QCOW2_CLUSTER_UNALLOCATED);
426     for (i = 0; i < nb_clusters; i++) {
427         uint64_t entry = be64_to_cpu(l2_slice[i]);
428         QCow2ClusterType type = qcow2_get_cluster_type(bs, entry);
429 
430         if (type != wanted_type) {
431             break;
432         }
433     }
434 
435     return i;
436 }
437 
438 static int coroutine_fn do_perform_cow_read(BlockDriverState *bs,
439                                             uint64_t src_cluster_offset,
440                                             unsigned offset_in_cluster,
441                                             QEMUIOVector *qiov)
442 {
443     int ret;
444 
445     if (qiov->size == 0) {
446         return 0;
447     }
448 
449     BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
450 
451     if (!bs->drv) {
452         return -ENOMEDIUM;
453     }
454 
455     /* Call .bdrv_co_readv() directly instead of using the public block-layer
456      * interface.  This avoids double I/O throttling and request tracking,
457      * which can lead to deadlock when block layer copy-on-read is enabled.
458      */
459     ret = bs->drv->bdrv_co_preadv_part(bs,
460                                        src_cluster_offset + offset_in_cluster,
461                                        qiov->size, qiov, 0, 0);
462     if (ret < 0) {
463         return ret;
464     }
465 
466     return 0;
467 }
468 
469 static int coroutine_fn do_perform_cow_write(BlockDriverState *bs,
470                                              uint64_t cluster_offset,
471                                              unsigned offset_in_cluster,
472                                              QEMUIOVector *qiov)
473 {
474     BDRVQcow2State *s = bs->opaque;
475     int ret;
476 
477     if (qiov->size == 0) {
478         return 0;
479     }
480 
481     ret = qcow2_pre_write_overlap_check(bs, 0,
482             cluster_offset + offset_in_cluster, qiov->size, true);
483     if (ret < 0) {
484         return ret;
485     }
486 
487     BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
488     ret = bdrv_co_pwritev(s->data_file, cluster_offset + offset_in_cluster,
489                           qiov->size, qiov, 0);
490     if (ret < 0) {
491         return ret;
492     }
493 
494     return 0;
495 }
496 
497 
498 /*
499  * get_cluster_offset
500  *
501  * For a given offset of the virtual disk, find the cluster type and offset in
502  * the qcow2 file. The offset is stored in *cluster_offset.
503  *
504  * On entry, *bytes is the maximum number of contiguous bytes starting at
505  * offset that we are interested in.
506  *
507  * On exit, *bytes is the number of bytes starting at offset that have the same
508  * cluster type and (if applicable) are stored contiguously in the image file.
509  * Compressed clusters are always returned one by one.
510  *
511  * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
512  * cases.
513  */
514 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
515                              unsigned int *bytes, uint64_t *cluster_offset)
516 {
517     BDRVQcow2State *s = bs->opaque;
518     unsigned int l2_index;
519     uint64_t l1_index, l2_offset, *l2_slice;
520     int c;
521     unsigned int offset_in_cluster;
522     uint64_t bytes_available, bytes_needed, nb_clusters;
523     QCow2ClusterType type;
524     int ret;
525 
526     offset_in_cluster = offset_into_cluster(s, offset);
527     bytes_needed = (uint64_t) *bytes + offset_in_cluster;
528 
529     /* compute how many bytes there are between the start of the cluster
530      * containing offset and the end of the l2 slice that contains
531      * the entry pointing to it */
532     bytes_available =
533         ((uint64_t) (s->l2_slice_size - offset_to_l2_slice_index(s, offset)))
534         << s->cluster_bits;
535 
536     if (bytes_needed > bytes_available) {
537         bytes_needed = bytes_available;
538     }
539 
540     *cluster_offset = 0;
541 
542     /* seek to the l2 offset in the l1 table */
543 
544     l1_index = offset_to_l1_index(s, offset);
545     if (l1_index >= s->l1_size) {
546         type = QCOW2_CLUSTER_UNALLOCATED;
547         goto out;
548     }
549 
550     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
551     if (!l2_offset) {
552         type = QCOW2_CLUSTER_UNALLOCATED;
553         goto out;
554     }
555 
556     if (offset_into_cluster(s, l2_offset)) {
557         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
558                                 " unaligned (L1 index: %#" PRIx64 ")",
559                                 l2_offset, l1_index);
560         return -EIO;
561     }
562 
563     /* load the l2 slice in memory */
564 
565     ret = l2_load(bs, offset, l2_offset, &l2_slice);
566     if (ret < 0) {
567         return ret;
568     }
569 
570     /* find the cluster offset for the given disk offset */
571 
572     l2_index = offset_to_l2_slice_index(s, offset);
573     *cluster_offset = be64_to_cpu(l2_slice[l2_index]);
574 
575     nb_clusters = size_to_clusters(s, bytes_needed);
576     /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
577      * integers; the minimum cluster size is 512, so this assertion is always
578      * true */
579     assert(nb_clusters <= INT_MAX);
580 
581     type = qcow2_get_cluster_type(bs, *cluster_offset);
582     if (s->qcow_version < 3 && (type == QCOW2_CLUSTER_ZERO_PLAIN ||
583                                 type == QCOW2_CLUSTER_ZERO_ALLOC)) {
584         qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
585                                 " in pre-v3 image (L2 offset: %#" PRIx64
586                                 ", L2 index: %#x)", l2_offset, l2_index);
587         ret = -EIO;
588         goto fail;
589     }
590     switch (type) {
591     case QCOW2_CLUSTER_COMPRESSED:
592         if (has_data_file(bs)) {
593             qcow2_signal_corruption(bs, true, -1, -1, "Compressed cluster "
594                                     "entry found in image with external data "
595                                     "file (L2 offset: %#" PRIx64 ", L2 index: "
596                                     "%#x)", l2_offset, l2_index);
597             ret = -EIO;
598             goto fail;
599         }
600         /* Compressed clusters can only be processed one by one */
601         c = 1;
602         *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
603         break;
604     case QCOW2_CLUSTER_ZERO_PLAIN:
605     case QCOW2_CLUSTER_UNALLOCATED:
606         /* how many empty clusters ? */
607         c = count_contiguous_clusters_unallocated(bs, nb_clusters,
608                                                   &l2_slice[l2_index], type);
609         *cluster_offset = 0;
610         break;
611     case QCOW2_CLUSTER_ZERO_ALLOC:
612     case QCOW2_CLUSTER_NORMAL:
613         /* how many allocated clusters ? */
614         c = count_contiguous_clusters(bs, nb_clusters, s->cluster_size,
615                                       &l2_slice[l2_index], QCOW_OFLAG_ZERO);
616         *cluster_offset &= L2E_OFFSET_MASK;
617         if (offset_into_cluster(s, *cluster_offset)) {
618             qcow2_signal_corruption(bs, true, -1, -1,
619                                     "Cluster allocation offset %#"
620                                     PRIx64 " unaligned (L2 offset: %#" PRIx64
621                                     ", L2 index: %#x)", *cluster_offset,
622                                     l2_offset, l2_index);
623             ret = -EIO;
624             goto fail;
625         }
626         if (has_data_file(bs) && *cluster_offset != offset - offset_in_cluster)
627         {
628             qcow2_signal_corruption(bs, true, -1, -1,
629                                     "External data file host cluster offset %#"
630                                     PRIx64 " does not match guest cluster "
631                                     "offset: %#" PRIx64
632                                     ", L2 index: %#x)", *cluster_offset,
633                                     offset - offset_in_cluster, l2_index);
634             ret = -EIO;
635             goto fail;
636         }
637         break;
638     default:
639         abort();
640     }
641 
642     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
643 
644     bytes_available = (int64_t)c * s->cluster_size;
645 
646 out:
647     if (bytes_available > bytes_needed) {
648         bytes_available = bytes_needed;
649     }
650 
651     /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
652      * subtracting offset_in_cluster will therefore definitely yield something
653      * not exceeding UINT_MAX */
654     assert(bytes_available - offset_in_cluster <= UINT_MAX);
655     *bytes = bytes_available - offset_in_cluster;
656 
657     return type;
658 
659 fail:
660     qcow2_cache_put(s->l2_table_cache, (void **)&l2_slice);
661     return ret;
662 }
663 
664 /*
665  * get_cluster_table
666  *
667  * for a given disk offset, load (and allocate if needed)
668  * the appropriate slice of its l2 table.
669  *
670  * the cluster index in the l2 slice is given to the caller.
671  *
672  * Returns 0 on success, -errno in failure case
673  */
674 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
675                              uint64_t **new_l2_slice,
676                              int *new_l2_index)
677 {
678     BDRVQcow2State *s = bs->opaque;
679     unsigned int l2_index;
680     uint64_t l1_index, l2_offset;
681     uint64_t *l2_slice = NULL;
682     int ret;
683 
684     /* seek to the l2 offset in the l1 table */
685 
686     l1_index = offset_to_l1_index(s, offset);
687     if (l1_index >= s->l1_size) {
688         ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
689         if (ret < 0) {
690             return ret;
691         }
692     }
693 
694     assert(l1_index < s->l1_size);
695     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
696     if (offset_into_cluster(s, l2_offset)) {
697         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
698                                 " unaligned (L1 index: %#" PRIx64 ")",
699                                 l2_offset, l1_index);
700         return -EIO;
701     }
702 
703     if (!(s->l1_table[l1_index] & QCOW_OFLAG_COPIED)) {
704         /* First allocate a new L2 table (and do COW if needed) */
705         ret = l2_allocate(bs, l1_index);
706         if (ret < 0) {
707             return ret;
708         }
709 
710         /* Then decrease the refcount of the old table */
711         if (l2_offset) {
712             qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
713                                 QCOW2_DISCARD_OTHER);
714         }
715 
716         /* Get the offset of the newly-allocated l2 table */
717         l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
718         assert(offset_into_cluster(s, l2_offset) == 0);
719     }
720 
721     /* load the l2 slice in memory */
722     ret = l2_load(bs, offset, l2_offset, &l2_slice);
723     if (ret < 0) {
724         return ret;
725     }
726 
727     /* find the cluster offset for the given disk offset */
728 
729     l2_index = offset_to_l2_slice_index(s, offset);
730 
731     *new_l2_slice = l2_slice;
732     *new_l2_index = l2_index;
733 
734     return 0;
735 }
736 
737 /*
738  * alloc_compressed_cluster_offset
739  *
740  * For a given offset on the virtual disk, allocate a new compressed cluster
741  * and put the host offset of the cluster into *host_offset. If a cluster is
742  * already allocated at the offset, return an error.
743  *
744  * Return 0 on success and -errno in error cases
745  */
746 int qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
747                                           uint64_t offset,
748                                           int compressed_size,
749                                           uint64_t *host_offset)
750 {
751     BDRVQcow2State *s = bs->opaque;
752     int l2_index, ret;
753     uint64_t *l2_slice;
754     int64_t cluster_offset;
755     int nb_csectors;
756 
757     if (has_data_file(bs)) {
758         return 0;
759     }
760 
761     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
762     if (ret < 0) {
763         return ret;
764     }
765 
766     /* Compression can't overwrite anything. Fail if the cluster was already
767      * allocated. */
768     cluster_offset = be64_to_cpu(l2_slice[l2_index]);
769     if (cluster_offset & L2E_OFFSET_MASK) {
770         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
771         return -EIO;
772     }
773 
774     cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
775     if (cluster_offset < 0) {
776         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
777         return cluster_offset;
778     }
779 
780     nb_csectors =
781         (cluster_offset + compressed_size - 1) / QCOW2_COMPRESSED_SECTOR_SIZE -
782         (cluster_offset / QCOW2_COMPRESSED_SECTOR_SIZE);
783 
784     /* The offset and size must fit in their fields of the L2 table entry */
785     assert((cluster_offset & s->cluster_offset_mask) == cluster_offset);
786     assert((nb_csectors & s->csize_mask) == nb_csectors);
787 
788     cluster_offset |= QCOW_OFLAG_COMPRESSED |
789                       ((uint64_t)nb_csectors << s->csize_shift);
790 
791     /* update L2 table */
792 
793     /* compressed clusters never have the copied flag */
794 
795     BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
796     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
797     l2_slice[l2_index] = cpu_to_be64(cluster_offset);
798     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
799 
800     *host_offset = cluster_offset & s->cluster_offset_mask;
801     return 0;
802 }
803 
804 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
805 {
806     BDRVQcow2State *s = bs->opaque;
807     Qcow2COWRegion *start = &m->cow_start;
808     Qcow2COWRegion *end = &m->cow_end;
809     unsigned buffer_size;
810     unsigned data_bytes = end->offset - (start->offset + start->nb_bytes);
811     bool merge_reads;
812     uint8_t *start_buffer, *end_buffer;
813     QEMUIOVector qiov;
814     int ret;
815 
816     assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
817     assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
818     assert(start->offset + start->nb_bytes <= end->offset);
819 
820     if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
821         return 0;
822     }
823 
824     /* If we have to read both the start and end COW regions and the
825      * middle region is not too large then perform just one read
826      * operation */
827     merge_reads = start->nb_bytes && end->nb_bytes && data_bytes <= 16384;
828     if (merge_reads) {
829         buffer_size = start->nb_bytes + data_bytes + end->nb_bytes;
830     } else {
831         /* If we have to do two reads, add some padding in the middle
832          * if necessary to make sure that the end region is optimally
833          * aligned. */
834         size_t align = bdrv_opt_mem_align(bs);
835         assert(align > 0 && align <= UINT_MAX);
836         assert(QEMU_ALIGN_UP(start->nb_bytes, align) <=
837                UINT_MAX - end->nb_bytes);
838         buffer_size = QEMU_ALIGN_UP(start->nb_bytes, align) + end->nb_bytes;
839     }
840 
841     /* Reserve a buffer large enough to store all the data that we're
842      * going to read */
843     start_buffer = qemu_try_blockalign(bs, buffer_size);
844     if (start_buffer == NULL) {
845         return -ENOMEM;
846     }
847     /* The part of the buffer where the end region is located */
848     end_buffer = start_buffer + buffer_size - end->nb_bytes;
849 
850     qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
851                                 qemu_iovec_subvec_niov(m->data_qiov,
852                                                        m->data_qiov_offset,
853                                                        data_bytes)
854                                 : 0));
855 
856     qemu_co_mutex_unlock(&s->lock);
857     /* First we read the existing data from both COW regions. We
858      * either read the whole region in one go, or the start and end
859      * regions separately. */
860     if (merge_reads) {
861         qemu_iovec_add(&qiov, start_buffer, buffer_size);
862         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
863     } else {
864         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
865         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
866         if (ret < 0) {
867             goto fail;
868         }
869 
870         qemu_iovec_reset(&qiov);
871         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
872         ret = do_perform_cow_read(bs, m->offset, end->offset, &qiov);
873     }
874     if (ret < 0) {
875         goto fail;
876     }
877 
878     /* Encrypt the data if necessary before writing it */
879     if (bs->encrypted) {
880         ret = qcow2_co_encrypt(bs,
881                                m->alloc_offset + start->offset,
882                                m->offset + start->offset,
883                                start_buffer, start->nb_bytes);
884         if (ret < 0) {
885             goto fail;
886         }
887 
888         ret = qcow2_co_encrypt(bs,
889                                m->alloc_offset + end->offset,
890                                m->offset + end->offset,
891                                end_buffer, end->nb_bytes);
892         if (ret < 0) {
893             goto fail;
894         }
895     }
896 
897     /* And now we can write everything. If we have the guest data we
898      * can write everything in one single operation */
899     if (m->data_qiov) {
900         qemu_iovec_reset(&qiov);
901         if (start->nb_bytes) {
902             qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
903         }
904         qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
905         if (end->nb_bytes) {
906             qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
907         }
908         /* NOTE: we have a write_aio blkdebug event here followed by
909          * a cow_write one in do_perform_cow_write(), but there's only
910          * one single I/O operation */
911         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
912         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
913     } else {
914         /* If there's no guest data then write both COW regions separately */
915         qemu_iovec_reset(&qiov);
916         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
917         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
918         if (ret < 0) {
919             goto fail;
920         }
921 
922         qemu_iovec_reset(&qiov);
923         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
924         ret = do_perform_cow_write(bs, m->alloc_offset, end->offset, &qiov);
925     }
926 
927 fail:
928     qemu_co_mutex_lock(&s->lock);
929 
930     /*
931      * Before we update the L2 table to actually point to the new cluster, we
932      * need to be sure that the refcounts have been increased and COW was
933      * handled.
934      */
935     if (ret == 0) {
936         qcow2_cache_depends_on_flush(s->l2_table_cache);
937     }
938 
939     qemu_vfree(start_buffer);
940     qemu_iovec_destroy(&qiov);
941     return ret;
942 }
943 
944 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
945 {
946     BDRVQcow2State *s = bs->opaque;
947     int i, j = 0, l2_index, ret;
948     uint64_t *old_cluster, *l2_slice;
949     uint64_t cluster_offset = m->alloc_offset;
950 
951     trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
952     assert(m->nb_clusters > 0);
953 
954     old_cluster = g_try_new(uint64_t, m->nb_clusters);
955     if (old_cluster == NULL) {
956         ret = -ENOMEM;
957         goto err;
958     }
959 
960     /* copy content of unmodified sectors */
961     ret = perform_cow(bs, m);
962     if (ret < 0) {
963         goto err;
964     }
965 
966     /* Update L2 table. */
967     if (s->use_lazy_refcounts) {
968         qcow2_mark_dirty(bs);
969     }
970     if (qcow2_need_accurate_refcounts(s)) {
971         qcow2_cache_set_dependency(bs, s->l2_table_cache,
972                                    s->refcount_block_cache);
973     }
974 
975     ret = get_cluster_table(bs, m->offset, &l2_slice, &l2_index);
976     if (ret < 0) {
977         goto err;
978     }
979     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
980 
981     assert(l2_index + m->nb_clusters <= s->l2_slice_size);
982     for (i = 0; i < m->nb_clusters; i++) {
983         uint64_t offset = cluster_offset + (i << s->cluster_bits);
984         /* if two concurrent writes happen to the same unallocated cluster
985          * each write allocates separate cluster and writes data concurrently.
986          * The first one to complete updates l2 table with pointer to its
987          * cluster the second one has to do RMW (which is done above by
988          * perform_cow()), update l2 table with its cluster pointer and free
989          * old cluster. This is what this loop does */
990         if (l2_slice[l2_index + i] != 0) {
991             old_cluster[j++] = l2_slice[l2_index + i];
992         }
993 
994         /* The offset must fit in the offset field of the L2 table entry */
995         assert((offset & L2E_OFFSET_MASK) == offset);
996 
997         l2_slice[l2_index + i] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
998      }
999 
1000 
1001     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1002 
1003     /*
1004      * If this was a COW, we need to decrease the refcount of the old cluster.
1005      *
1006      * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1007      * clusters), the next write will reuse them anyway.
1008      */
1009     if (!m->keep_old_clusters && j != 0) {
1010         for (i = 0; i < j; i++) {
1011             qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
1012                                     QCOW2_DISCARD_NEVER);
1013         }
1014     }
1015 
1016     ret = 0;
1017 err:
1018     g_free(old_cluster);
1019     return ret;
1020  }
1021 
1022 /**
1023  * Frees the allocated clusters because the request failed and they won't
1024  * actually be linked.
1025  */
1026 void qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1027 {
1028     BDRVQcow2State *s = bs->opaque;
1029     if (!has_data_file(bs) && !m->keep_old_clusters) {
1030         qcow2_free_clusters(bs, m->alloc_offset,
1031                             m->nb_clusters << s->cluster_bits,
1032                             QCOW2_DISCARD_NEVER);
1033     }
1034 }
1035 
1036 /*
1037  * Returns the number of contiguous clusters that can be used for an allocating
1038  * write, but require COW to be performed (this includes yet unallocated space,
1039  * which must copy from the backing file)
1040  */
1041 static int count_cow_clusters(BlockDriverState *bs, int nb_clusters,
1042     uint64_t *l2_slice, int l2_index)
1043 {
1044     int i;
1045 
1046     for (i = 0; i < nb_clusters; i++) {
1047         uint64_t l2_entry = be64_to_cpu(l2_slice[l2_index + i]);
1048         QCow2ClusterType cluster_type = qcow2_get_cluster_type(bs, l2_entry);
1049 
1050         switch(cluster_type) {
1051         case QCOW2_CLUSTER_NORMAL:
1052             if (l2_entry & QCOW_OFLAG_COPIED) {
1053                 goto out;
1054             }
1055             break;
1056         case QCOW2_CLUSTER_UNALLOCATED:
1057         case QCOW2_CLUSTER_COMPRESSED:
1058         case QCOW2_CLUSTER_ZERO_PLAIN:
1059         case QCOW2_CLUSTER_ZERO_ALLOC:
1060             break;
1061         default:
1062             abort();
1063         }
1064     }
1065 
1066 out:
1067     assert(i <= nb_clusters);
1068     return i;
1069 }
1070 
1071 /*
1072  * Check if there already is an AIO write request in flight which allocates
1073  * the same cluster. In this case we need to wait until the previous
1074  * request has completed and updated the L2 table accordingly.
1075  *
1076  * Returns:
1077  *   0       if there was no dependency. *cur_bytes indicates the number of
1078  *           bytes from guest_offset that can be read before the next
1079  *           dependency must be processed (or the request is complete)
1080  *
1081  *   -EAGAIN if we had to wait for another request, previously gathered
1082  *           information on cluster allocation may be invalid now. The caller
1083  *           must start over anyway, so consider *cur_bytes undefined.
1084  */
1085 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
1086     uint64_t *cur_bytes, QCowL2Meta **m)
1087 {
1088     BDRVQcow2State *s = bs->opaque;
1089     QCowL2Meta *old_alloc;
1090     uint64_t bytes = *cur_bytes;
1091 
1092     QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
1093 
1094         uint64_t start = guest_offset;
1095         uint64_t end = start + bytes;
1096         uint64_t old_start = l2meta_cow_start(old_alloc);
1097         uint64_t old_end = l2meta_cow_end(old_alloc);
1098 
1099         if (end <= old_start || start >= old_end) {
1100             /* No intersection */
1101         } else {
1102             if (start < old_start) {
1103                 /* Stop at the start of a running allocation */
1104                 bytes = old_start - start;
1105             } else {
1106                 bytes = 0;
1107             }
1108 
1109             /* Stop if already an l2meta exists. After yielding, it wouldn't
1110              * be valid any more, so we'd have to clean up the old L2Metas
1111              * and deal with requests depending on them before starting to
1112              * gather new ones. Not worth the trouble. */
1113             if (bytes == 0 && *m) {
1114                 *cur_bytes = 0;
1115                 return 0;
1116             }
1117 
1118             if (bytes == 0) {
1119                 /* Wait for the dependency to complete. We need to recheck
1120                  * the free/allocated clusters when we continue. */
1121                 qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
1122                 return -EAGAIN;
1123             }
1124         }
1125     }
1126 
1127     /* Make sure that existing clusters and new allocations are only used up to
1128      * the next dependency if we shortened the request above */
1129     *cur_bytes = bytes;
1130 
1131     return 0;
1132 }
1133 
1134 /*
1135  * Checks how many already allocated clusters that don't require a copy on
1136  * write there are at the given guest_offset (up to *bytes). If *host_offset is
1137  * not INV_OFFSET, only physically contiguous clusters beginning at this host
1138  * offset are counted.
1139  *
1140  * Note that guest_offset may not be cluster aligned. In this case, the
1141  * returned *host_offset points to exact byte referenced by guest_offset and
1142  * therefore isn't cluster aligned as well.
1143  *
1144  * Returns:
1145  *   0:     if no allocated clusters are available at the given offset.
1146  *          *bytes is normally unchanged. It is set to 0 if the cluster
1147  *          is allocated and doesn't need COW, but doesn't have the right
1148  *          physical offset.
1149  *
1150  *   1:     if allocated clusters that don't require a COW are available at
1151  *          the requested offset. *bytes may have decreased and describes
1152  *          the length of the area that can be written to.
1153  *
1154  *  -errno: in error cases
1155  */
1156 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
1157     uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1158 {
1159     BDRVQcow2State *s = bs->opaque;
1160     int l2_index;
1161     uint64_t cluster_offset;
1162     uint64_t *l2_slice;
1163     uint64_t nb_clusters;
1164     unsigned int keep_clusters;
1165     int ret;
1166 
1167     trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
1168                               *bytes);
1169 
1170     assert(*host_offset == INV_OFFSET || offset_into_cluster(s, guest_offset)
1171                                       == offset_into_cluster(s, *host_offset));
1172 
1173     /*
1174      * Calculate the number of clusters to look for. We stop at L2 slice
1175      * boundaries to keep things simple.
1176      */
1177     nb_clusters =
1178         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1179 
1180     l2_index = offset_to_l2_slice_index(s, guest_offset);
1181     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1182     assert(nb_clusters <= INT_MAX);
1183 
1184     /* Find L2 entry for the first involved cluster */
1185     ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1186     if (ret < 0) {
1187         return ret;
1188     }
1189 
1190     cluster_offset = be64_to_cpu(l2_slice[l2_index]);
1191 
1192     /* Check how many clusters are already allocated and don't need COW */
1193     if (qcow2_get_cluster_type(bs, cluster_offset) == QCOW2_CLUSTER_NORMAL
1194         && (cluster_offset & QCOW_OFLAG_COPIED))
1195     {
1196         /* If a specific host_offset is required, check it */
1197         bool offset_matches =
1198             (cluster_offset & L2E_OFFSET_MASK) == *host_offset;
1199 
1200         if (offset_into_cluster(s, cluster_offset & L2E_OFFSET_MASK)) {
1201             qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset "
1202                                     "%#llx unaligned (guest offset: %#" PRIx64
1203                                     ")", cluster_offset & L2E_OFFSET_MASK,
1204                                     guest_offset);
1205             ret = -EIO;
1206             goto out;
1207         }
1208 
1209         if (*host_offset != INV_OFFSET && !offset_matches) {
1210             *bytes = 0;
1211             ret = 0;
1212             goto out;
1213         }
1214 
1215         /* We keep all QCOW_OFLAG_COPIED clusters */
1216         keep_clusters =
1217             count_contiguous_clusters(bs, nb_clusters, s->cluster_size,
1218                                       &l2_slice[l2_index],
1219                                       QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
1220         assert(keep_clusters <= nb_clusters);
1221 
1222         *bytes = MIN(*bytes,
1223                  keep_clusters * s->cluster_size
1224                  - offset_into_cluster(s, guest_offset));
1225 
1226         ret = 1;
1227     } else {
1228         ret = 0;
1229     }
1230 
1231     /* Cleanup */
1232 out:
1233     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1234 
1235     /* Only return a host offset if we actually made progress. Otherwise we
1236      * would make requirements for handle_alloc() that it can't fulfill */
1237     if (ret > 0) {
1238         *host_offset = (cluster_offset & L2E_OFFSET_MASK)
1239                      + offset_into_cluster(s, guest_offset);
1240     }
1241 
1242     return ret;
1243 }
1244 
1245 /*
1246  * Allocates new clusters for the given guest_offset.
1247  *
1248  * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
1249  * contain the number of clusters that have been allocated and are contiguous
1250  * in the image file.
1251  *
1252  * If *host_offset is not INV_OFFSET, it specifies the offset in the image file
1253  * at which the new clusters must start. *nb_clusters can be 0 on return in
1254  * this case if the cluster at host_offset is already in use. If *host_offset
1255  * is INV_OFFSET, the clusters can be allocated anywhere in the image file.
1256  *
1257  * *host_offset is updated to contain the offset into the image file at which
1258  * the first allocated cluster starts.
1259  *
1260  * Return 0 on success and -errno in error cases. -EAGAIN means that the
1261  * function has been waiting for another request and the allocation must be
1262  * restarted, but the whole request should not be failed.
1263  */
1264 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
1265                                    uint64_t *host_offset, uint64_t *nb_clusters)
1266 {
1267     BDRVQcow2State *s = bs->opaque;
1268 
1269     trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1270                                          *host_offset, *nb_clusters);
1271 
1272     if (has_data_file(bs)) {
1273         assert(*host_offset == INV_OFFSET ||
1274                *host_offset == start_of_cluster(s, guest_offset));
1275         *host_offset = start_of_cluster(s, guest_offset);
1276         return 0;
1277     }
1278 
1279     /* Allocate new clusters */
1280     trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1281     if (*host_offset == INV_OFFSET) {
1282         int64_t cluster_offset =
1283             qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1284         if (cluster_offset < 0) {
1285             return cluster_offset;
1286         }
1287         *host_offset = cluster_offset;
1288         return 0;
1289     } else {
1290         int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1291         if (ret < 0) {
1292             return ret;
1293         }
1294         *nb_clusters = ret;
1295         return 0;
1296     }
1297 }
1298 
1299 /*
1300  * Allocates new clusters for an area that either is yet unallocated or needs a
1301  * copy on write. If *host_offset is not INV_OFFSET, clusters are only
1302  * allocated if the new allocation can match the specified host offset.
1303  *
1304  * Note that guest_offset may not be cluster aligned. In this case, the
1305  * returned *host_offset points to exact byte referenced by guest_offset and
1306  * therefore isn't cluster aligned as well.
1307  *
1308  * Returns:
1309  *   0:     if no clusters could be allocated. *bytes is set to 0,
1310  *          *host_offset is left unchanged.
1311  *
1312  *   1:     if new clusters were allocated. *bytes may be decreased if the
1313  *          new allocation doesn't cover all of the requested area.
1314  *          *host_offset is updated to contain the host offset of the first
1315  *          newly allocated cluster.
1316  *
1317  *  -errno: in error cases
1318  */
1319 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1320     uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1321 {
1322     BDRVQcow2State *s = bs->opaque;
1323     int l2_index;
1324     uint64_t *l2_slice;
1325     uint64_t entry;
1326     uint64_t nb_clusters;
1327     int ret;
1328     bool keep_old_clusters = false;
1329 
1330     uint64_t alloc_cluster_offset = INV_OFFSET;
1331 
1332     trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1333                              *bytes);
1334     assert(*bytes > 0);
1335 
1336     /*
1337      * Calculate the number of clusters to look for. We stop at L2 slice
1338      * boundaries to keep things simple.
1339      */
1340     nb_clusters =
1341         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1342 
1343     l2_index = offset_to_l2_slice_index(s, guest_offset);
1344     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1345     assert(nb_clusters <= INT_MAX);
1346 
1347     /* Limit total allocation byte count to INT_MAX */
1348     nb_clusters = MIN(nb_clusters, INT_MAX >> s->cluster_bits);
1349 
1350     /* Find L2 entry for the first involved cluster */
1351     ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1352     if (ret < 0) {
1353         return ret;
1354     }
1355 
1356     entry = be64_to_cpu(l2_slice[l2_index]);
1357     nb_clusters = count_cow_clusters(bs, nb_clusters, l2_slice, l2_index);
1358 
1359     /* This function is only called when there were no non-COW clusters, so if
1360      * we can't find any unallocated or COW clusters either, something is
1361      * wrong with our code. */
1362     assert(nb_clusters > 0);
1363 
1364     if (qcow2_get_cluster_type(bs, entry) == QCOW2_CLUSTER_ZERO_ALLOC &&
1365         (entry & QCOW_OFLAG_COPIED) &&
1366         (*host_offset == INV_OFFSET ||
1367          start_of_cluster(s, *host_offset) == (entry & L2E_OFFSET_MASK)))
1368     {
1369         int preallocated_nb_clusters;
1370 
1371         if (offset_into_cluster(s, entry & L2E_OFFSET_MASK)) {
1372             qcow2_signal_corruption(bs, true, -1, -1, "Preallocated zero "
1373                                     "cluster offset %#llx unaligned (guest "
1374                                     "offset: %#" PRIx64 ")",
1375                                     entry & L2E_OFFSET_MASK, guest_offset);
1376             ret = -EIO;
1377             goto fail;
1378         }
1379 
1380         /* Try to reuse preallocated zero clusters; contiguous normal clusters
1381          * would be fine, too, but count_cow_clusters() above has limited
1382          * nb_clusters already to a range of COW clusters */
1383         preallocated_nb_clusters =
1384             count_contiguous_clusters(bs, nb_clusters, s->cluster_size,
1385                                       &l2_slice[l2_index], QCOW_OFLAG_COPIED);
1386         assert(preallocated_nb_clusters > 0);
1387 
1388         nb_clusters = preallocated_nb_clusters;
1389         alloc_cluster_offset = entry & L2E_OFFSET_MASK;
1390 
1391         /* We want to reuse these clusters, so qcow2_alloc_cluster_link_l2()
1392          * should not free them. */
1393         keep_old_clusters = true;
1394     }
1395 
1396     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1397 
1398     if (alloc_cluster_offset == INV_OFFSET) {
1399         /* Allocate, if necessary at a given offset in the image file */
1400         alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1401                                start_of_cluster(s, *host_offset);
1402         ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1403                                       &nb_clusters);
1404         if (ret < 0) {
1405             goto fail;
1406         }
1407 
1408         /* Can't extend contiguous allocation */
1409         if (nb_clusters == 0) {
1410             *bytes = 0;
1411             return 0;
1412         }
1413 
1414         assert(alloc_cluster_offset != INV_OFFSET);
1415     }
1416 
1417     /*
1418      * Save info needed for meta data update.
1419      *
1420      * requested_bytes: Number of bytes from the start of the first
1421      * newly allocated cluster to the end of the (possibly shortened
1422      * before) write request.
1423      *
1424      * avail_bytes: Number of bytes from the start of the first
1425      * newly allocated to the end of the last newly allocated cluster.
1426      *
1427      * nb_bytes: The number of bytes from the start of the first
1428      * newly allocated cluster to the end of the area that the write
1429      * request actually writes to (excluding COW at the end)
1430      */
1431     uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1432     int avail_bytes = nb_clusters << s->cluster_bits;
1433     int nb_bytes = MIN(requested_bytes, avail_bytes);
1434     QCowL2Meta *old_m = *m;
1435 
1436     *m = g_malloc0(sizeof(**m));
1437 
1438     **m = (QCowL2Meta) {
1439         .next           = old_m,
1440 
1441         .alloc_offset   = alloc_cluster_offset,
1442         .offset         = start_of_cluster(s, guest_offset),
1443         .nb_clusters    = nb_clusters,
1444 
1445         .keep_old_clusters  = keep_old_clusters,
1446 
1447         .cow_start = {
1448             .offset     = 0,
1449             .nb_bytes   = offset_into_cluster(s, guest_offset),
1450         },
1451         .cow_end = {
1452             .offset     = nb_bytes,
1453             .nb_bytes   = avail_bytes - nb_bytes,
1454         },
1455     };
1456     qemu_co_queue_init(&(*m)->dependent_requests);
1457     QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1458 
1459     *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1460     *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1461     assert(*bytes != 0);
1462 
1463     return 1;
1464 
1465 fail:
1466     if (*m && (*m)->nb_clusters > 0) {
1467         QLIST_REMOVE(*m, next_in_flight);
1468     }
1469     return ret;
1470 }
1471 
1472 /*
1473  * alloc_cluster_offset
1474  *
1475  * For a given offset on the virtual disk, find the cluster offset in qcow2
1476  * file. If the offset is not found, allocate a new cluster.
1477  *
1478  * If the cluster was already allocated, m->nb_clusters is set to 0 and
1479  * other fields in m are meaningless.
1480  *
1481  * If the cluster is newly allocated, m->nb_clusters is set to the number of
1482  * contiguous clusters that have been allocated. In this case, the other
1483  * fields of m are valid and contain information about the first allocated
1484  * cluster.
1485  *
1486  * If the request conflicts with another write request in flight, the coroutine
1487  * is queued and will be reentered when the dependency has completed.
1488  *
1489  * Return 0 on success and -errno in error cases
1490  */
1491 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1492                                unsigned int *bytes, uint64_t *host_offset,
1493                                QCowL2Meta **m)
1494 {
1495     BDRVQcow2State *s = bs->opaque;
1496     uint64_t start, remaining;
1497     uint64_t cluster_offset;
1498     uint64_t cur_bytes;
1499     int ret;
1500 
1501     trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1502 
1503 again:
1504     start = offset;
1505     remaining = *bytes;
1506     cluster_offset = INV_OFFSET;
1507     *host_offset = INV_OFFSET;
1508     cur_bytes = 0;
1509     *m = NULL;
1510 
1511     while (true) {
1512 
1513         if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1514             *host_offset = start_of_cluster(s, cluster_offset);
1515         }
1516 
1517         assert(remaining >= cur_bytes);
1518 
1519         start           += cur_bytes;
1520         remaining       -= cur_bytes;
1521 
1522         if (cluster_offset != INV_OFFSET) {
1523             cluster_offset += cur_bytes;
1524         }
1525 
1526         if (remaining == 0) {
1527             break;
1528         }
1529 
1530         cur_bytes = remaining;
1531 
1532         /*
1533          * Now start gathering as many contiguous clusters as possible:
1534          *
1535          * 1. Check for overlaps with in-flight allocations
1536          *
1537          *      a) Overlap not in the first cluster -> shorten this request and
1538          *         let the caller handle the rest in its next loop iteration.
1539          *
1540          *      b) Real overlaps of two requests. Yield and restart the search
1541          *         for contiguous clusters (the situation could have changed
1542          *         while we were sleeping)
1543          *
1544          *      c) TODO: Request starts in the same cluster as the in-flight
1545          *         allocation ends. Shorten the COW of the in-fight allocation,
1546          *         set cluster_offset to write to the same cluster and set up
1547          *         the right synchronisation between the in-flight request and
1548          *         the new one.
1549          */
1550         ret = handle_dependencies(bs, start, &cur_bytes, m);
1551         if (ret == -EAGAIN) {
1552             /* Currently handle_dependencies() doesn't yield if we already had
1553              * an allocation. If it did, we would have to clean up the L2Meta
1554              * structs before starting over. */
1555             assert(*m == NULL);
1556             goto again;
1557         } else if (ret < 0) {
1558             return ret;
1559         } else if (cur_bytes == 0) {
1560             break;
1561         } else {
1562             /* handle_dependencies() may have decreased cur_bytes (shortened
1563              * the allocations below) so that the next dependency is processed
1564              * correctly during the next loop iteration. */
1565         }
1566 
1567         /*
1568          * 2. Count contiguous COPIED clusters.
1569          */
1570         ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1571         if (ret < 0) {
1572             return ret;
1573         } else if (ret) {
1574             continue;
1575         } else if (cur_bytes == 0) {
1576             break;
1577         }
1578 
1579         /*
1580          * 3. If the request still hasn't completed, allocate new clusters,
1581          *    considering any cluster_offset of steps 1c or 2.
1582          */
1583         ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1584         if (ret < 0) {
1585             return ret;
1586         } else if (ret) {
1587             continue;
1588         } else {
1589             assert(cur_bytes == 0);
1590             break;
1591         }
1592     }
1593 
1594     *bytes -= remaining;
1595     assert(*bytes > 0);
1596     assert(*host_offset != INV_OFFSET);
1597 
1598     return 0;
1599 }
1600 
1601 /*
1602  * This discards as many clusters of nb_clusters as possible at once (i.e.
1603  * all clusters in the same L2 slice) and returns the number of discarded
1604  * clusters.
1605  */
1606 static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1607                                uint64_t nb_clusters,
1608                                enum qcow2_discard_type type, bool full_discard)
1609 {
1610     BDRVQcow2State *s = bs->opaque;
1611     uint64_t *l2_slice;
1612     int l2_index;
1613     int ret;
1614     int i;
1615 
1616     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1617     if (ret < 0) {
1618         return ret;
1619     }
1620 
1621     /* Limit nb_clusters to one L2 slice */
1622     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1623     assert(nb_clusters <= INT_MAX);
1624 
1625     for (i = 0; i < nb_clusters; i++) {
1626         uint64_t old_l2_entry;
1627 
1628         old_l2_entry = be64_to_cpu(l2_slice[l2_index + i]);
1629 
1630         /*
1631          * If full_discard is false, make sure that a discarded area reads back
1632          * as zeroes for v3 images (we cannot do it for v2 without actually
1633          * writing a zero-filled buffer). We can skip the operation if the
1634          * cluster is already marked as zero, or if it's unallocated and we
1635          * don't have a backing file.
1636          *
1637          * TODO We might want to use bdrv_block_status(bs) here, but we're
1638          * holding s->lock, so that doesn't work today.
1639          *
1640          * If full_discard is true, the sector should not read back as zeroes,
1641          * but rather fall through to the backing file.
1642          */
1643         switch (qcow2_get_cluster_type(bs, old_l2_entry)) {
1644         case QCOW2_CLUSTER_UNALLOCATED:
1645             if (full_discard || !bs->backing) {
1646                 continue;
1647             }
1648             break;
1649 
1650         case QCOW2_CLUSTER_ZERO_PLAIN:
1651             if (!full_discard) {
1652                 continue;
1653             }
1654             break;
1655 
1656         case QCOW2_CLUSTER_ZERO_ALLOC:
1657         case QCOW2_CLUSTER_NORMAL:
1658         case QCOW2_CLUSTER_COMPRESSED:
1659             break;
1660 
1661         default:
1662             abort();
1663         }
1664 
1665         /* First remove L2 entries */
1666         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1667         if (!full_discard && s->qcow_version >= 3) {
1668             l2_slice[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1669         } else {
1670             l2_slice[l2_index + i] = cpu_to_be64(0);
1671         }
1672 
1673         /* Then decrease the refcount */
1674         qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1675     }
1676 
1677     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1678 
1679     return nb_clusters;
1680 }
1681 
1682 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1683                           uint64_t bytes, enum qcow2_discard_type type,
1684                           bool full_discard)
1685 {
1686     BDRVQcow2State *s = bs->opaque;
1687     uint64_t end_offset = offset + bytes;
1688     uint64_t nb_clusters;
1689     int64_t cleared;
1690     int ret;
1691 
1692     /* Caller must pass aligned values, except at image end */
1693     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1694     assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1695            end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
1696 
1697     nb_clusters = size_to_clusters(s, bytes);
1698 
1699     s->cache_discards = true;
1700 
1701     /* Each L2 slice is handled by its own loop iteration */
1702     while (nb_clusters > 0) {
1703         cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
1704                                       full_discard);
1705         if (cleared < 0) {
1706             ret = cleared;
1707             goto fail;
1708         }
1709 
1710         nb_clusters -= cleared;
1711         offset += (cleared * s->cluster_size);
1712     }
1713 
1714     ret = 0;
1715 fail:
1716     s->cache_discards = false;
1717     qcow2_process_discards(bs, ret);
1718 
1719     return ret;
1720 }
1721 
1722 /*
1723  * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1724  * all clusters in the same L2 slice) and returns the number of zeroed
1725  * clusters.
1726  */
1727 static int zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1728                             uint64_t nb_clusters, int flags)
1729 {
1730     BDRVQcow2State *s = bs->opaque;
1731     uint64_t *l2_slice;
1732     int l2_index;
1733     int ret;
1734     int i;
1735     bool unmap = !!(flags & BDRV_REQ_MAY_UNMAP);
1736 
1737     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1738     if (ret < 0) {
1739         return ret;
1740     }
1741 
1742     /* Limit nb_clusters to one L2 slice */
1743     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1744     assert(nb_clusters <= INT_MAX);
1745 
1746     for (i = 0; i < nb_clusters; i++) {
1747         uint64_t old_offset;
1748         QCow2ClusterType cluster_type;
1749 
1750         old_offset = be64_to_cpu(l2_slice[l2_index + i]);
1751 
1752         /*
1753          * Minimize L2 changes if the cluster already reads back as
1754          * zeroes with correct allocation.
1755          */
1756         cluster_type = qcow2_get_cluster_type(bs, old_offset);
1757         if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN ||
1758             (cluster_type == QCOW2_CLUSTER_ZERO_ALLOC && !unmap)) {
1759             continue;
1760         }
1761 
1762         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1763         if (cluster_type == QCOW2_CLUSTER_COMPRESSED || unmap) {
1764             l2_slice[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1765             qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1766         } else {
1767             l2_slice[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1768         }
1769     }
1770 
1771     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1772 
1773     return nb_clusters;
1774 }
1775 
1776 int qcow2_cluster_zeroize(BlockDriverState *bs, uint64_t offset,
1777                           uint64_t bytes, int flags)
1778 {
1779     BDRVQcow2State *s = bs->opaque;
1780     uint64_t end_offset = offset + bytes;
1781     uint64_t nb_clusters;
1782     int64_t cleared;
1783     int ret;
1784 
1785     /* If we have to stay in sync with an external data file, zero out
1786      * s->data_file first. */
1787     if (data_file_is_raw(bs)) {
1788         assert(has_data_file(bs));
1789         ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
1790         if (ret < 0) {
1791             return ret;
1792         }
1793     }
1794 
1795     /* Caller must pass aligned values, except at image end */
1796     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1797     assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1798            end_offset >= bs->total_sectors << BDRV_SECTOR_BITS);
1799 
1800     /* The zero flag is only supported by version 3 and newer */
1801     if (s->qcow_version < 3) {
1802         return -ENOTSUP;
1803     }
1804 
1805     /* Each L2 slice is handled by its own loop iteration */
1806     nb_clusters = size_to_clusters(s, bytes);
1807 
1808     s->cache_discards = true;
1809 
1810     while (nb_clusters > 0) {
1811         cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
1812         if (cleared < 0) {
1813             ret = cleared;
1814             goto fail;
1815         }
1816 
1817         nb_clusters -= cleared;
1818         offset += (cleared * s->cluster_size);
1819     }
1820 
1821     ret = 0;
1822 fail:
1823     s->cache_discards = false;
1824     qcow2_process_discards(bs, ret);
1825 
1826     return ret;
1827 }
1828 
1829 /*
1830  * Expands all zero clusters in a specific L1 table (or deallocates them, for
1831  * non-backed non-pre-allocated zero clusters).
1832  *
1833  * l1_entries and *visited_l1_entries are used to keep track of progress for
1834  * status_cb(). l1_entries contains the total number of L1 entries and
1835  * *visited_l1_entries counts all visited L1 entries.
1836  */
1837 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1838                                       int l1_size, int64_t *visited_l1_entries,
1839                                       int64_t l1_entries,
1840                                       BlockDriverAmendStatusCB *status_cb,
1841                                       void *cb_opaque)
1842 {
1843     BDRVQcow2State *s = bs->opaque;
1844     bool is_active_l1 = (l1_table == s->l1_table);
1845     uint64_t *l2_slice = NULL;
1846     unsigned slice, slice_size2, n_slices;
1847     int ret;
1848     int i, j;
1849 
1850     slice_size2 = s->l2_slice_size * sizeof(uint64_t);
1851     n_slices = s->cluster_size / slice_size2;
1852 
1853     if (!is_active_l1) {
1854         /* inactive L2 tables require a buffer to be stored in when loading
1855          * them from disk */
1856         l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
1857         if (l2_slice == NULL) {
1858             return -ENOMEM;
1859         }
1860     }
1861 
1862     for (i = 0; i < l1_size; i++) {
1863         uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1864         uint64_t l2_refcount;
1865 
1866         if (!l2_offset) {
1867             /* unallocated */
1868             (*visited_l1_entries)++;
1869             if (status_cb) {
1870                 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
1871             }
1872             continue;
1873         }
1874 
1875         if (offset_into_cluster(s, l2_offset)) {
1876             qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1877                                     PRIx64 " unaligned (L1 index: %#x)",
1878                                     l2_offset, i);
1879             ret = -EIO;
1880             goto fail;
1881         }
1882 
1883         ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1884                                  &l2_refcount);
1885         if (ret < 0) {
1886             goto fail;
1887         }
1888 
1889         for (slice = 0; slice < n_slices; slice++) {
1890             uint64_t slice_offset = l2_offset + slice * slice_size2;
1891             bool l2_dirty = false;
1892             if (is_active_l1) {
1893                 /* get active L2 tables from cache */
1894                 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
1895                                       (void **)&l2_slice);
1896             } else {
1897                 /* load inactive L2 tables from disk */
1898                 ret = bdrv_pread(bs->file, slice_offset, l2_slice, slice_size2);
1899             }
1900             if (ret < 0) {
1901                 goto fail;
1902             }
1903 
1904             for (j = 0; j < s->l2_slice_size; j++) {
1905                 uint64_t l2_entry = be64_to_cpu(l2_slice[j]);
1906                 int64_t offset = l2_entry & L2E_OFFSET_MASK;
1907                 QCow2ClusterType cluster_type =
1908                     qcow2_get_cluster_type(bs, l2_entry);
1909 
1910                 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
1911                     cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
1912                     continue;
1913                 }
1914 
1915                 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
1916                     if (!bs->backing) {
1917                         /* not backed; therefore we can simply deallocate the
1918                          * cluster */
1919                         l2_slice[j] = 0;
1920                         l2_dirty = true;
1921                         continue;
1922                     }
1923 
1924                     offset = qcow2_alloc_clusters(bs, s->cluster_size);
1925                     if (offset < 0) {
1926                         ret = offset;
1927                         goto fail;
1928                     }
1929 
1930                     /* The offset must fit in the offset field */
1931                     assert((offset & L2E_OFFSET_MASK) == offset);
1932 
1933                     if (l2_refcount > 1) {
1934                         /* For shared L2 tables, set the refcount accordingly
1935                          * (it is already 1 and needs to be l2_refcount) */
1936                         ret = qcow2_update_cluster_refcount(
1937                             bs, offset >> s->cluster_bits,
1938                             refcount_diff(1, l2_refcount), false,
1939                             QCOW2_DISCARD_OTHER);
1940                         if (ret < 0) {
1941                             qcow2_free_clusters(bs, offset, s->cluster_size,
1942                                                 QCOW2_DISCARD_OTHER);
1943                             goto fail;
1944                         }
1945                     }
1946                 }
1947 
1948                 if (offset_into_cluster(s, offset)) {
1949                     int l2_index = slice * s->l2_slice_size + j;
1950                     qcow2_signal_corruption(
1951                         bs, true, -1, -1,
1952                         "Cluster allocation offset "
1953                         "%#" PRIx64 " unaligned (L2 offset: %#"
1954                         PRIx64 ", L2 index: %#x)", offset,
1955                         l2_offset, l2_index);
1956                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
1957                         qcow2_free_clusters(bs, offset, s->cluster_size,
1958                                             QCOW2_DISCARD_ALWAYS);
1959                     }
1960                     ret = -EIO;
1961                     goto fail;
1962                 }
1963 
1964                 ret = qcow2_pre_write_overlap_check(bs, 0, offset,
1965                                                     s->cluster_size, true);
1966                 if (ret < 0) {
1967                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
1968                         qcow2_free_clusters(bs, offset, s->cluster_size,
1969                                             QCOW2_DISCARD_ALWAYS);
1970                     }
1971                     goto fail;
1972                 }
1973 
1974                 ret = bdrv_pwrite_zeroes(s->data_file, offset,
1975                                          s->cluster_size, 0);
1976                 if (ret < 0) {
1977                     if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
1978                         qcow2_free_clusters(bs, offset, s->cluster_size,
1979                                             QCOW2_DISCARD_ALWAYS);
1980                     }
1981                     goto fail;
1982                 }
1983 
1984                 if (l2_refcount == 1) {
1985                     l2_slice[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
1986                 } else {
1987                     l2_slice[j] = cpu_to_be64(offset);
1988                 }
1989                 l2_dirty = true;
1990             }
1991 
1992             if (is_active_l1) {
1993                 if (l2_dirty) {
1994                     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1995                     qcow2_cache_depends_on_flush(s->l2_table_cache);
1996                 }
1997                 qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1998             } else {
1999                 if (l2_dirty) {
2000                     ret = qcow2_pre_write_overlap_check(
2001                         bs, QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2,
2002                         slice_offset, slice_size2, false);
2003                     if (ret < 0) {
2004                         goto fail;
2005                     }
2006 
2007                     ret = bdrv_pwrite(bs->file, slice_offset,
2008                                       l2_slice, slice_size2);
2009                     if (ret < 0) {
2010                         goto fail;
2011                     }
2012                 }
2013             }
2014         }
2015 
2016         (*visited_l1_entries)++;
2017         if (status_cb) {
2018             status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
2019         }
2020     }
2021 
2022     ret = 0;
2023 
2024 fail:
2025     if (l2_slice) {
2026         if (!is_active_l1) {
2027             qemu_vfree(l2_slice);
2028         } else {
2029             qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
2030         }
2031     }
2032     return ret;
2033 }
2034 
2035 /*
2036  * For backed images, expands all zero clusters on the image. For non-backed
2037  * images, deallocates all non-pre-allocated zero clusters (and claims the
2038  * allocation for pre-allocated ones). This is important for downgrading to a
2039  * qcow2 version which doesn't yet support metadata zero clusters.
2040  */
2041 int qcow2_expand_zero_clusters(BlockDriverState *bs,
2042                                BlockDriverAmendStatusCB *status_cb,
2043                                void *cb_opaque)
2044 {
2045     BDRVQcow2State *s = bs->opaque;
2046     uint64_t *l1_table = NULL;
2047     int64_t l1_entries = 0, visited_l1_entries = 0;
2048     int ret;
2049     int i, j;
2050 
2051     if (status_cb) {
2052         l1_entries = s->l1_size;
2053         for (i = 0; i < s->nb_snapshots; i++) {
2054             l1_entries += s->snapshots[i].l1_size;
2055         }
2056     }
2057 
2058     ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
2059                                      &visited_l1_entries, l1_entries,
2060                                      status_cb, cb_opaque);
2061     if (ret < 0) {
2062         goto fail;
2063     }
2064 
2065     /* Inactive L1 tables may point to active L2 tables - therefore it is
2066      * necessary to flush the L2 table cache before trying to access the L2
2067      * tables pointed to by inactive L1 entries (else we might try to expand
2068      * zero clusters that have already been expanded); furthermore, it is also
2069      * necessary to empty the L2 table cache, since it may contain tables which
2070      * are now going to be modified directly on disk, bypassing the cache.
2071      * qcow2_cache_empty() does both for us. */
2072     ret = qcow2_cache_empty(bs, s->l2_table_cache);
2073     if (ret < 0) {
2074         goto fail;
2075     }
2076 
2077     for (i = 0; i < s->nb_snapshots; i++) {
2078         int l1_size2;
2079         uint64_t *new_l1_table;
2080         Error *local_err = NULL;
2081 
2082         ret = qcow2_validate_table(bs, s->snapshots[i].l1_table_offset,
2083                                    s->snapshots[i].l1_size, sizeof(uint64_t),
2084                                    QCOW_MAX_L1_SIZE, "Snapshot L1 table",
2085                                    &local_err);
2086         if (ret < 0) {
2087             error_report_err(local_err);
2088             goto fail;
2089         }
2090 
2091         l1_size2 = s->snapshots[i].l1_size * sizeof(uint64_t);
2092         new_l1_table = g_try_realloc(l1_table, l1_size2);
2093 
2094         if (!new_l1_table) {
2095             ret = -ENOMEM;
2096             goto fail;
2097         }
2098 
2099         l1_table = new_l1_table;
2100 
2101         ret = bdrv_pread(bs->file, s->snapshots[i].l1_table_offset,
2102                          l1_table, l1_size2);
2103         if (ret < 0) {
2104             goto fail;
2105         }
2106 
2107         for (j = 0; j < s->snapshots[i].l1_size; j++) {
2108             be64_to_cpus(&l1_table[j]);
2109         }
2110 
2111         ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
2112                                          &visited_l1_entries, l1_entries,
2113                                          status_cb, cb_opaque);
2114         if (ret < 0) {
2115             goto fail;
2116         }
2117     }
2118 
2119     ret = 0;
2120 
2121 fail:
2122     g_free(l1_table);
2123     return ret;
2124 }
2125