xref: /openbmc/qemu/block/qcow2-cluster.c (revision 14776ab5)
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,
128                                        ROUND_UP(new_l1_size2, 512));
129     if (new_l1_table == NULL) {
130         return -ENOMEM;
131     }
132     memset(new_l1_table, 0, ROUND_UP(new_l1_size2, 512));
133 
134     if (s->l1_size) {
135         memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
136     }
137 
138     /* write new table (align to cluster) */
139     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
140     new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
141     if (new_l1_table_offset < 0) {
142         qemu_vfree(new_l1_table);
143         return new_l1_table_offset;
144     }
145 
146     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
147     if (ret < 0) {
148         goto fail;
149     }
150 
151     /* the L1 position has not yet been updated, so these clusters must
152      * indeed be completely free */
153     ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
154                                         new_l1_size2, false);
155     if (ret < 0) {
156         goto fail;
157     }
158 
159     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
160     for(i = 0; i < s->l1_size; i++)
161         new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
162     ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset,
163                            new_l1_table, new_l1_size2);
164     if (ret < 0)
165         goto fail;
166     for(i = 0; i < s->l1_size; i++)
167         new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
168 
169     /* set new table */
170     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
171     stl_be_p(data, new_l1_size);
172     stq_be_p(data + 4, new_l1_table_offset);
173     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size),
174                            data, sizeof(data));
175     if (ret < 0) {
176         goto fail;
177     }
178     qemu_vfree(s->l1_table);
179     old_l1_table_offset = s->l1_table_offset;
180     s->l1_table_offset = new_l1_table_offset;
181     s->l1_table = new_l1_table;
182     old_l1_size = s->l1_size;
183     s->l1_size = new_l1_size;
184     qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
185                         QCOW2_DISCARD_OTHER);
186     return 0;
187  fail:
188     qemu_vfree(new_l1_table);
189     qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
190                         QCOW2_DISCARD_OTHER);
191     return ret;
192 }
193 
194 /*
195  * l2_load
196  *
197  * @bs: The BlockDriverState
198  * @offset: A guest offset, used to calculate what slice of the L2
199  *          table to load.
200  * @l2_offset: Offset to the L2 table in the image file.
201  * @l2_slice: Location to store the pointer to the L2 slice.
202  *
203  * Loads a L2 slice into memory (L2 slices are the parts of L2 tables
204  * that are loaded by the qcow2 cache). If the slice is in the cache,
205  * the cache is used; otherwise the L2 slice is loaded from the image
206  * file.
207  */
208 static int l2_load(BlockDriverState *bs, uint64_t offset,
209                    uint64_t l2_offset, uint64_t **l2_slice)
210 {
211     BDRVQcow2State *s = bs->opaque;
212     int start_of_slice = sizeof(uint64_t) *
213         (offset_to_l2_index(s, offset) - offset_to_l2_slice_index(s, offset));
214 
215     return qcow2_cache_get(bs, s->l2_table_cache, l2_offset + start_of_slice,
216                            (void **)l2_slice);
217 }
218 
219 /*
220  * Writes one sector of the L1 table to the disk (can't update single entries
221  * and we really don't want bdrv_pread to perform a read-modify-write)
222  */
223 #define L1_ENTRIES_PER_SECTOR (512 / 8)
224 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
225 {
226     BDRVQcow2State *s = bs->opaque;
227     uint64_t buf[L1_ENTRIES_PER_SECTOR] = { 0 };
228     int l1_start_index;
229     int i, ret;
230 
231     l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
232     for (i = 0; i < L1_ENTRIES_PER_SECTOR && l1_start_index + i < s->l1_size;
233          i++)
234     {
235         buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
236     }
237 
238     ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
239             s->l1_table_offset + 8 * l1_start_index, sizeof(buf), false);
240     if (ret < 0) {
241         return ret;
242     }
243 
244     BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
245     ret = bdrv_pwrite_sync(bs->file,
246                            s->l1_table_offset + 8 * l1_start_index,
247                            buf, sizeof(buf));
248     if (ret < 0) {
249         return ret;
250     }
251 
252     return 0;
253 }
254 
255 /*
256  * l2_allocate
257  *
258  * Allocate a new l2 entry in the file. If l1_index points to an already
259  * used entry in the L2 table (i.e. we are doing a copy on write for the L2
260  * table) copy the contents of the old L2 table into the newly allocated one.
261  * Otherwise the new table is initialized with zeros.
262  *
263  */
264 
265 static int l2_allocate(BlockDriverState *bs, int l1_index)
266 {
267     BDRVQcow2State *s = bs->opaque;
268     uint64_t old_l2_offset;
269     uint64_t *l2_slice = NULL;
270     unsigned slice, slice_size2, n_slices;
271     int64_t l2_offset;
272     int ret;
273 
274     old_l2_offset = s->l1_table[l1_index];
275 
276     trace_qcow2_l2_allocate(bs, l1_index);
277 
278     /* allocate a new l2 entry */
279 
280     l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
281     if (l2_offset < 0) {
282         ret = l2_offset;
283         goto fail;
284     }
285 
286     /* The offset must fit in the offset field of the L1 table entry */
287     assert((l2_offset & L1E_OFFSET_MASK) == l2_offset);
288 
289     /* If we're allocating the table at offset 0 then something is wrong */
290     if (l2_offset == 0) {
291         qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid "
292                                 "allocation of L2 table at offset 0");
293         ret = -EIO;
294         goto fail;
295     }
296 
297     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
298     if (ret < 0) {
299         goto fail;
300     }
301 
302     /* allocate a new entry in the l2 cache */
303 
304     slice_size2 = s->l2_slice_size * sizeof(uint64_t);
305     n_slices = s->cluster_size / slice_size2;
306 
307     trace_qcow2_l2_allocate_get_empty(bs, l1_index);
308     for (slice = 0; slice < n_slices; slice++) {
309         ret = qcow2_cache_get_empty(bs, s->l2_table_cache,
310                                     l2_offset + slice * slice_size2,
311                                     (void **) &l2_slice);
312         if (ret < 0) {
313             goto fail;
314         }
315 
316         if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
317             /* if there was no old l2 table, clear the new slice */
318             memset(l2_slice, 0, slice_size2);
319         } else {
320             uint64_t *old_slice;
321             uint64_t old_l2_slice_offset =
322                 (old_l2_offset & L1E_OFFSET_MASK) + slice * slice_size2;
323 
324             /* if there was an old l2 table, read a slice from the disk */
325             BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
326             ret = qcow2_cache_get(bs, s->l2_table_cache, old_l2_slice_offset,
327                                   (void **) &old_slice);
328             if (ret < 0) {
329                 goto fail;
330             }
331 
332             memcpy(l2_slice, old_slice, slice_size2);
333 
334             qcow2_cache_put(s->l2_table_cache, (void **) &old_slice);
335         }
336 
337         /* write the l2 slice to the file */
338         BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
339 
340         trace_qcow2_l2_allocate_write_l2(bs, l1_index);
341         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
342         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
343     }
344 
345     ret = qcow2_cache_flush(bs, s->l2_table_cache);
346     if (ret < 0) {
347         goto fail;
348     }
349 
350     /* update the L1 entry */
351     trace_qcow2_l2_allocate_write_l1(bs, l1_index);
352     s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
353     ret = qcow2_write_l1_entry(bs, l1_index);
354     if (ret < 0) {
355         goto fail;
356     }
357 
358     trace_qcow2_l2_allocate_done(bs, l1_index, 0);
359     return 0;
360 
361 fail:
362     trace_qcow2_l2_allocate_done(bs, l1_index, ret);
363     if (l2_slice != NULL) {
364         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
365     }
366     s->l1_table[l1_index] = old_l2_offset;
367     if (l2_offset > 0) {
368         qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
369                             QCOW2_DISCARD_ALWAYS);
370     }
371     return ret;
372 }
373 
374 /*
375  * Checks how many clusters in a given L2 slice are contiguous in the image
376  * file. As soon as one of the flags in the bitmask stop_flags changes compared
377  * to the first cluster, the search is stopped and the cluster is not counted
378  * as contiguous. (This allows it, for example, to stop at the first compressed
379  * cluster which may require a different handling)
380  */
381 static int count_contiguous_clusters(BlockDriverState *bs, int nb_clusters,
382         int cluster_size, uint64_t *l2_slice, uint64_t stop_flags)
383 {
384     int i;
385     QCow2ClusterType first_cluster_type;
386     uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
387     uint64_t first_entry = be64_to_cpu(l2_slice[0]);
388     uint64_t offset = first_entry & mask;
389 
390     first_cluster_type = qcow2_get_cluster_type(bs, first_entry);
391     if (first_cluster_type == QCOW2_CLUSTER_UNALLOCATED) {
392         return 0;
393     }
394 
395     /* must be allocated */
396     assert(first_cluster_type == QCOW2_CLUSTER_NORMAL ||
397            first_cluster_type == QCOW2_CLUSTER_ZERO_ALLOC);
398 
399     for (i = 0; i < nb_clusters; i++) {
400         uint64_t l2_entry = be64_to_cpu(l2_slice[i]) & mask;
401         if (offset + (uint64_t) i * cluster_size != l2_entry) {
402             break;
403         }
404     }
405 
406         return i;
407 }
408 
409 /*
410  * Checks how many consecutive unallocated clusters in a given L2
411  * slice have the same cluster type.
412  */
413 static int count_contiguous_clusters_unallocated(BlockDriverState *bs,
414                                                  int nb_clusters,
415                                                  uint64_t *l2_slice,
416                                                  QCow2ClusterType wanted_type)
417 {
418     int i;
419 
420     assert(wanted_type == QCOW2_CLUSTER_ZERO_PLAIN ||
421            wanted_type == QCOW2_CLUSTER_UNALLOCATED);
422     for (i = 0; i < nb_clusters; i++) {
423         uint64_t entry = be64_to_cpu(l2_slice[i]);
424         QCow2ClusterType type = qcow2_get_cluster_type(bs, entry);
425 
426         if (type != wanted_type) {
427             break;
428         }
429     }
430 
431     return i;
432 }
433 
434 static int coroutine_fn do_perform_cow_read(BlockDriverState *bs,
435                                             uint64_t src_cluster_offset,
436                                             unsigned offset_in_cluster,
437                                             QEMUIOVector *qiov)
438 {
439     int ret;
440 
441     if (qiov->size == 0) {
442         return 0;
443     }
444 
445     BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
446 
447     if (!bs->drv) {
448         return -ENOMEDIUM;
449     }
450 
451     /* Call .bdrv_co_readv() directly instead of using the public block-layer
452      * interface.  This avoids double I/O throttling and request tracking,
453      * which can lead to deadlock when block layer copy-on-read is enabled.
454      */
455     ret = bs->drv->bdrv_co_preadv_part(bs,
456                                        src_cluster_offset + offset_in_cluster,
457                                        qiov->size, qiov, 0, 0);
458     if (ret < 0) {
459         return ret;
460     }
461 
462     return 0;
463 }
464 
465 static bool coroutine_fn do_perform_cow_encrypt(BlockDriverState *bs,
466                                                 uint64_t src_cluster_offset,
467                                                 uint64_t cluster_offset,
468                                                 unsigned offset_in_cluster,
469                                                 uint8_t *buffer,
470                                                 unsigned bytes)
471 {
472     if (bytes && bs->encrypted) {
473         BDRVQcow2State *s = bs->opaque;
474         assert((offset_in_cluster & ~BDRV_SECTOR_MASK) == 0);
475         assert((bytes & ~BDRV_SECTOR_MASK) == 0);
476         assert(s->crypto);
477         if (qcow2_co_encrypt(bs, cluster_offset,
478                              src_cluster_offset + offset_in_cluster,
479                              buffer, bytes) < 0) {
480             return false;
481         }
482     }
483     return true;
484 }
485 
486 static int coroutine_fn do_perform_cow_write(BlockDriverState *bs,
487                                              uint64_t cluster_offset,
488                                              unsigned offset_in_cluster,
489                                              QEMUIOVector *qiov)
490 {
491     BDRVQcow2State *s = bs->opaque;
492     int ret;
493 
494     if (qiov->size == 0) {
495         return 0;
496     }
497 
498     ret = qcow2_pre_write_overlap_check(bs, 0,
499             cluster_offset + offset_in_cluster, qiov->size, true);
500     if (ret < 0) {
501         return ret;
502     }
503 
504     BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
505     ret = bdrv_co_pwritev(s->data_file, cluster_offset + offset_in_cluster,
506                           qiov->size, qiov, 0);
507     if (ret < 0) {
508         return ret;
509     }
510 
511     return 0;
512 }
513 
514 
515 /*
516  * get_cluster_offset
517  *
518  * For a given offset of the virtual disk, find the cluster type and offset in
519  * the qcow2 file. The offset is stored in *cluster_offset.
520  *
521  * On entry, *bytes is the maximum number of contiguous bytes starting at
522  * offset that we are interested in.
523  *
524  * On exit, *bytes is the number of bytes starting at offset that have the same
525  * cluster type and (if applicable) are stored contiguously in the image file.
526  * Compressed clusters are always returned one by one.
527  *
528  * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
529  * cases.
530  */
531 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
532                              unsigned int *bytes, uint64_t *cluster_offset)
533 {
534     BDRVQcow2State *s = bs->opaque;
535     unsigned int l2_index;
536     uint64_t l1_index, l2_offset, *l2_slice;
537     int c;
538     unsigned int offset_in_cluster;
539     uint64_t bytes_available, bytes_needed, nb_clusters;
540     QCow2ClusterType type;
541     int ret;
542 
543     offset_in_cluster = offset_into_cluster(s, offset);
544     bytes_needed = (uint64_t) *bytes + offset_in_cluster;
545 
546     /* compute how many bytes there are between the start of the cluster
547      * containing offset and the end of the l2 slice that contains
548      * the entry pointing to it */
549     bytes_available =
550         ((uint64_t) (s->l2_slice_size - offset_to_l2_slice_index(s, offset)))
551         << s->cluster_bits;
552 
553     if (bytes_needed > bytes_available) {
554         bytes_needed = bytes_available;
555     }
556 
557     *cluster_offset = 0;
558 
559     /* seek to the l2 offset in the l1 table */
560 
561     l1_index = offset_to_l1_index(s, offset);
562     if (l1_index >= s->l1_size) {
563         type = QCOW2_CLUSTER_UNALLOCATED;
564         goto out;
565     }
566 
567     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
568     if (!l2_offset) {
569         type = QCOW2_CLUSTER_UNALLOCATED;
570         goto out;
571     }
572 
573     if (offset_into_cluster(s, l2_offset)) {
574         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
575                                 " unaligned (L1 index: %#" PRIx64 ")",
576                                 l2_offset, l1_index);
577         return -EIO;
578     }
579 
580     /* load the l2 slice in memory */
581 
582     ret = l2_load(bs, offset, l2_offset, &l2_slice);
583     if (ret < 0) {
584         return ret;
585     }
586 
587     /* find the cluster offset for the given disk offset */
588 
589     l2_index = offset_to_l2_slice_index(s, offset);
590     *cluster_offset = be64_to_cpu(l2_slice[l2_index]);
591 
592     nb_clusters = size_to_clusters(s, bytes_needed);
593     /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned
594      * integers; the minimum cluster size is 512, so this assertion is always
595      * true */
596     assert(nb_clusters <= INT_MAX);
597 
598     type = qcow2_get_cluster_type(bs, *cluster_offset);
599     if (s->qcow_version < 3 && (type == QCOW2_CLUSTER_ZERO_PLAIN ||
600                                 type == QCOW2_CLUSTER_ZERO_ALLOC)) {
601         qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found"
602                                 " in pre-v3 image (L2 offset: %#" PRIx64
603                                 ", L2 index: %#x)", l2_offset, l2_index);
604         ret = -EIO;
605         goto fail;
606     }
607     switch (type) {
608     case QCOW2_CLUSTER_COMPRESSED:
609         if (has_data_file(bs)) {
610             qcow2_signal_corruption(bs, true, -1, -1, "Compressed cluster "
611                                     "entry found in image with external data "
612                                     "file (L2 offset: %#" PRIx64 ", L2 index: "
613                                     "%#x)", l2_offset, l2_index);
614             ret = -EIO;
615             goto fail;
616         }
617         /* Compressed clusters can only be processed one by one */
618         c = 1;
619         *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
620         break;
621     case QCOW2_CLUSTER_ZERO_PLAIN:
622     case QCOW2_CLUSTER_UNALLOCATED:
623         /* how many empty clusters ? */
624         c = count_contiguous_clusters_unallocated(bs, nb_clusters,
625                                                   &l2_slice[l2_index], type);
626         *cluster_offset = 0;
627         break;
628     case QCOW2_CLUSTER_ZERO_ALLOC:
629     case QCOW2_CLUSTER_NORMAL:
630         /* how many allocated clusters ? */
631         c = count_contiguous_clusters(bs, nb_clusters, s->cluster_size,
632                                       &l2_slice[l2_index], QCOW_OFLAG_ZERO);
633         *cluster_offset &= L2E_OFFSET_MASK;
634         if (offset_into_cluster(s, *cluster_offset)) {
635             qcow2_signal_corruption(bs, true, -1, -1,
636                                     "Cluster allocation offset %#"
637                                     PRIx64 " unaligned (L2 offset: %#" PRIx64
638                                     ", L2 index: %#x)", *cluster_offset,
639                                     l2_offset, l2_index);
640             ret = -EIO;
641             goto fail;
642         }
643         if (has_data_file(bs) && *cluster_offset != offset - offset_in_cluster)
644         {
645             qcow2_signal_corruption(bs, true, -1, -1,
646                                     "External data file host cluster offset %#"
647                                     PRIx64 " does not match guest cluster "
648                                     "offset: %#" PRIx64
649                                     ", L2 index: %#x)", *cluster_offset,
650                                     offset - offset_in_cluster, l2_index);
651             ret = -EIO;
652             goto fail;
653         }
654         break;
655     default:
656         abort();
657     }
658 
659     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
660 
661     bytes_available = (int64_t)c * s->cluster_size;
662 
663 out:
664     if (bytes_available > bytes_needed) {
665         bytes_available = bytes_needed;
666     }
667 
668     /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster;
669      * subtracting offset_in_cluster will therefore definitely yield something
670      * not exceeding UINT_MAX */
671     assert(bytes_available - offset_in_cluster <= UINT_MAX);
672     *bytes = bytes_available - offset_in_cluster;
673 
674     return type;
675 
676 fail:
677     qcow2_cache_put(s->l2_table_cache, (void **)&l2_slice);
678     return ret;
679 }
680 
681 /*
682  * get_cluster_table
683  *
684  * for a given disk offset, load (and allocate if needed)
685  * the appropriate slice of its l2 table.
686  *
687  * the cluster index in the l2 slice is given to the caller.
688  *
689  * Returns 0 on success, -errno in failure case
690  */
691 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
692                              uint64_t **new_l2_slice,
693                              int *new_l2_index)
694 {
695     BDRVQcow2State *s = bs->opaque;
696     unsigned int l2_index;
697     uint64_t l1_index, l2_offset;
698     uint64_t *l2_slice = NULL;
699     int ret;
700 
701     /* seek to the l2 offset in the l1 table */
702 
703     l1_index = offset_to_l1_index(s, offset);
704     if (l1_index >= s->l1_size) {
705         ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
706         if (ret < 0) {
707             return ret;
708         }
709     }
710 
711     assert(l1_index < s->l1_size);
712     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
713     if (offset_into_cluster(s, l2_offset)) {
714         qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64
715                                 " unaligned (L1 index: %#" PRIx64 ")",
716                                 l2_offset, l1_index);
717         return -EIO;
718     }
719 
720     if (!(s->l1_table[l1_index] & QCOW_OFLAG_COPIED)) {
721         /* First allocate a new L2 table (and do COW if needed) */
722         ret = l2_allocate(bs, l1_index);
723         if (ret < 0) {
724             return ret;
725         }
726 
727         /* Then decrease the refcount of the old table */
728         if (l2_offset) {
729             qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
730                                 QCOW2_DISCARD_OTHER);
731         }
732 
733         /* Get the offset of the newly-allocated l2 table */
734         l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
735         assert(offset_into_cluster(s, l2_offset) == 0);
736     }
737 
738     /* load the l2 slice in memory */
739     ret = l2_load(bs, offset, l2_offset, &l2_slice);
740     if (ret < 0) {
741         return ret;
742     }
743 
744     /* find the cluster offset for the given disk offset */
745 
746     l2_index = offset_to_l2_slice_index(s, offset);
747 
748     *new_l2_slice = l2_slice;
749     *new_l2_index = l2_index;
750 
751     return 0;
752 }
753 
754 /*
755  * alloc_compressed_cluster_offset
756  *
757  * For a given offset on the virtual disk, allocate a new compressed cluster
758  * and put the host offset of the cluster into *host_offset. If a cluster is
759  * already allocated at the offset, return an error.
760  *
761  * Return 0 on success and -errno in error cases
762  */
763 int qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
764                                           uint64_t offset,
765                                           int compressed_size,
766                                           uint64_t *host_offset)
767 {
768     BDRVQcow2State *s = bs->opaque;
769     int l2_index, ret;
770     uint64_t *l2_slice;
771     int64_t cluster_offset;
772     int nb_csectors;
773 
774     if (has_data_file(bs)) {
775         return 0;
776     }
777 
778     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
779     if (ret < 0) {
780         return ret;
781     }
782 
783     /* Compression can't overwrite anything. Fail if the cluster was already
784      * allocated. */
785     cluster_offset = be64_to_cpu(l2_slice[l2_index]);
786     if (cluster_offset & L2E_OFFSET_MASK) {
787         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
788         return -EIO;
789     }
790 
791     cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
792     if (cluster_offset < 0) {
793         qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
794         return cluster_offset;
795     }
796 
797     nb_csectors =
798         (cluster_offset + compressed_size - 1) / QCOW2_COMPRESSED_SECTOR_SIZE -
799         (cluster_offset / QCOW2_COMPRESSED_SECTOR_SIZE);
800 
801     cluster_offset |= QCOW_OFLAG_COMPRESSED |
802                       ((uint64_t)nb_csectors << s->csize_shift);
803 
804     /* update L2 table */
805 
806     /* compressed clusters never have the copied flag */
807 
808     BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
809     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
810     l2_slice[l2_index] = cpu_to_be64(cluster_offset);
811     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
812 
813     *host_offset = cluster_offset & s->cluster_offset_mask;
814     return 0;
815 }
816 
817 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
818 {
819     BDRVQcow2State *s = bs->opaque;
820     Qcow2COWRegion *start = &m->cow_start;
821     Qcow2COWRegion *end = &m->cow_end;
822     unsigned buffer_size;
823     unsigned data_bytes = end->offset - (start->offset + start->nb_bytes);
824     bool merge_reads;
825     uint8_t *start_buffer, *end_buffer;
826     QEMUIOVector qiov;
827     int ret;
828 
829     assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
830     assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
831     assert(start->offset + start->nb_bytes <= end->offset);
832 
833     if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
834         return 0;
835     }
836 
837     /* If we have to read both the start and end COW regions and the
838      * middle region is not too large then perform just one read
839      * operation */
840     merge_reads = start->nb_bytes && end->nb_bytes && data_bytes <= 16384;
841     if (merge_reads) {
842         buffer_size = start->nb_bytes + data_bytes + end->nb_bytes;
843     } else {
844         /* If we have to do two reads, add some padding in the middle
845          * if necessary to make sure that the end region is optimally
846          * aligned. */
847         size_t align = bdrv_opt_mem_align(bs);
848         assert(align > 0 && align <= UINT_MAX);
849         assert(QEMU_ALIGN_UP(start->nb_bytes, align) <=
850                UINT_MAX - end->nb_bytes);
851         buffer_size = QEMU_ALIGN_UP(start->nb_bytes, align) + end->nb_bytes;
852     }
853 
854     /* Reserve a buffer large enough to store all the data that we're
855      * going to read */
856     start_buffer = qemu_try_blockalign(bs, buffer_size);
857     if (start_buffer == NULL) {
858         return -ENOMEM;
859     }
860     /* The part of the buffer where the end region is located */
861     end_buffer = start_buffer + buffer_size - end->nb_bytes;
862 
863     qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
864                                 qemu_iovec_subvec_niov(m->data_qiov,
865                                                        m->data_qiov_offset,
866                                                        data_bytes)
867                                 : 0));
868 
869     qemu_co_mutex_unlock(&s->lock);
870     /* First we read the existing data from both COW regions. We
871      * either read the whole region in one go, or the start and end
872      * regions separately. */
873     if (merge_reads) {
874         qemu_iovec_add(&qiov, start_buffer, buffer_size);
875         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
876     } else {
877         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
878         ret = do_perform_cow_read(bs, m->offset, start->offset, &qiov);
879         if (ret < 0) {
880             goto fail;
881         }
882 
883         qemu_iovec_reset(&qiov);
884         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
885         ret = do_perform_cow_read(bs, m->offset, end->offset, &qiov);
886     }
887     if (ret < 0) {
888         goto fail;
889     }
890 
891     /* Encrypt the data if necessary before writing it */
892     if (bs->encrypted) {
893         if (!do_perform_cow_encrypt(bs, m->offset, m->alloc_offset,
894                                     start->offset, start_buffer,
895                                     start->nb_bytes) ||
896             !do_perform_cow_encrypt(bs, m->offset, m->alloc_offset,
897                                     end->offset, end_buffer, end->nb_bytes)) {
898             ret = -EIO;
899             goto fail;
900         }
901     }
902 
903     /* And now we can write everything. If we have the guest data we
904      * can write everything in one single operation */
905     if (m->data_qiov) {
906         qemu_iovec_reset(&qiov);
907         if (start->nb_bytes) {
908             qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
909         }
910         qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
911         if (end->nb_bytes) {
912             qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
913         }
914         /* NOTE: we have a write_aio blkdebug event here followed by
915          * a cow_write one in do_perform_cow_write(), but there's only
916          * one single I/O operation */
917         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
918         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
919     } else {
920         /* If there's no guest data then write both COW regions separately */
921         qemu_iovec_reset(&qiov);
922         qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
923         ret = do_perform_cow_write(bs, m->alloc_offset, start->offset, &qiov);
924         if (ret < 0) {
925             goto fail;
926         }
927 
928         qemu_iovec_reset(&qiov);
929         qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
930         ret = do_perform_cow_write(bs, m->alloc_offset, end->offset, &qiov);
931     }
932 
933 fail:
934     qemu_co_mutex_lock(&s->lock);
935 
936     /*
937      * Before we update the L2 table to actually point to the new cluster, we
938      * need to be sure that the refcounts have been increased and COW was
939      * handled.
940      */
941     if (ret == 0) {
942         qcow2_cache_depends_on_flush(s->l2_table_cache);
943     }
944 
945     qemu_vfree(start_buffer);
946     qemu_iovec_destroy(&qiov);
947     return ret;
948 }
949 
950 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
951 {
952     BDRVQcow2State *s = bs->opaque;
953     int i, j = 0, l2_index, ret;
954     uint64_t *old_cluster, *l2_slice;
955     uint64_t cluster_offset = m->alloc_offset;
956 
957     trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
958     assert(m->nb_clusters > 0);
959 
960     old_cluster = g_try_new(uint64_t, m->nb_clusters);
961     if (old_cluster == NULL) {
962         ret = -ENOMEM;
963         goto err;
964     }
965 
966     /* copy content of unmodified sectors */
967     ret = perform_cow(bs, m);
968     if (ret < 0) {
969         goto err;
970     }
971 
972     /* Update L2 table. */
973     if (s->use_lazy_refcounts) {
974         qcow2_mark_dirty(bs);
975     }
976     if (qcow2_need_accurate_refcounts(s)) {
977         qcow2_cache_set_dependency(bs, s->l2_table_cache,
978                                    s->refcount_block_cache);
979     }
980 
981     ret = get_cluster_table(bs, m->offset, &l2_slice, &l2_index);
982     if (ret < 0) {
983         goto err;
984     }
985     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
986 
987     assert(l2_index + m->nb_clusters <= s->l2_slice_size);
988     for (i = 0; i < m->nb_clusters; i++) {
989         /* if two concurrent writes happen to the same unallocated cluster
990          * each write allocates separate cluster and writes data concurrently.
991          * The first one to complete updates l2 table with pointer to its
992          * cluster the second one has to do RMW (which is done above by
993          * perform_cow()), update l2 table with its cluster pointer and free
994          * old cluster. This is what this loop does */
995         if (l2_slice[l2_index + i] != 0) {
996             old_cluster[j++] = l2_slice[l2_index + i];
997         }
998 
999         l2_slice[l2_index + i] = cpu_to_be64((cluster_offset +
1000                     (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
1001      }
1002 
1003 
1004     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1005 
1006     /*
1007      * If this was a COW, we need to decrease the refcount of the old cluster.
1008      *
1009      * Don't discard clusters that reach a refcount of 0 (e.g. compressed
1010      * clusters), the next write will reuse them anyway.
1011      */
1012     if (!m->keep_old_clusters && j != 0) {
1013         for (i = 0; i < j; i++) {
1014             qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
1015                                     QCOW2_DISCARD_NEVER);
1016         }
1017     }
1018 
1019     ret = 0;
1020 err:
1021     g_free(old_cluster);
1022     return ret;
1023  }
1024 
1025 /**
1026  * Frees the allocated clusters because the request failed and they won't
1027  * actually be linked.
1028  */
1029 void qcow2_alloc_cluster_abort(BlockDriverState *bs, QCowL2Meta *m)
1030 {
1031     BDRVQcow2State *s = bs->opaque;
1032     qcow2_free_clusters(bs, m->alloc_offset, m->nb_clusters << s->cluster_bits,
1033                         QCOW2_DISCARD_NEVER);
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     /* Find L2 entry for the first involved cluster */
1348     ret = get_cluster_table(bs, guest_offset, &l2_slice, &l2_index);
1349     if (ret < 0) {
1350         return ret;
1351     }
1352 
1353     entry = be64_to_cpu(l2_slice[l2_index]);
1354 
1355     /* For the moment, overwrite compressed clusters one by one */
1356     if (entry & QCOW_OFLAG_COMPRESSED) {
1357         nb_clusters = 1;
1358     } else {
1359         nb_clusters = count_cow_clusters(bs, nb_clusters, l2_slice, l2_index);
1360     }
1361 
1362     /* This function is only called when there were no non-COW clusters, so if
1363      * we can't find any unallocated or COW clusters either, something is
1364      * wrong with our code. */
1365     assert(nb_clusters > 0);
1366 
1367     if (qcow2_get_cluster_type(bs, entry) == QCOW2_CLUSTER_ZERO_ALLOC &&
1368         (entry & QCOW_OFLAG_COPIED) &&
1369         (*host_offset == INV_OFFSET ||
1370          start_of_cluster(s, *host_offset) == (entry & L2E_OFFSET_MASK)))
1371     {
1372         int preallocated_nb_clusters;
1373 
1374         if (offset_into_cluster(s, entry & L2E_OFFSET_MASK)) {
1375             qcow2_signal_corruption(bs, true, -1, -1, "Preallocated zero "
1376                                     "cluster offset %#llx unaligned (guest "
1377                                     "offset: %#" PRIx64 ")",
1378                                     entry & L2E_OFFSET_MASK, guest_offset);
1379             ret = -EIO;
1380             goto fail;
1381         }
1382 
1383         /* Try to reuse preallocated zero clusters; contiguous normal clusters
1384          * would be fine, too, but count_cow_clusters() above has limited
1385          * nb_clusters already to a range of COW clusters */
1386         preallocated_nb_clusters =
1387             count_contiguous_clusters(bs, nb_clusters, s->cluster_size,
1388                                       &l2_slice[l2_index], QCOW_OFLAG_COPIED);
1389         assert(preallocated_nb_clusters > 0);
1390 
1391         nb_clusters = preallocated_nb_clusters;
1392         alloc_cluster_offset = entry & L2E_OFFSET_MASK;
1393 
1394         /* We want to reuse these clusters, so qcow2_alloc_cluster_link_l2()
1395          * should not free them. */
1396         keep_old_clusters = true;
1397     }
1398 
1399     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1400 
1401     if (alloc_cluster_offset == INV_OFFSET) {
1402         /* Allocate, if necessary at a given offset in the image file */
1403         alloc_cluster_offset = *host_offset == INV_OFFSET ? INV_OFFSET :
1404                                start_of_cluster(s, *host_offset);
1405         ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1406                                       &nb_clusters);
1407         if (ret < 0) {
1408             goto fail;
1409         }
1410 
1411         /* Can't extend contiguous allocation */
1412         if (nb_clusters == 0) {
1413             *bytes = 0;
1414             return 0;
1415         }
1416 
1417         assert(alloc_cluster_offset != INV_OFFSET);
1418     }
1419 
1420     /*
1421      * Save info needed for meta data update.
1422      *
1423      * requested_bytes: Number of bytes from the start of the first
1424      * newly allocated cluster to the end of the (possibly shortened
1425      * before) write request.
1426      *
1427      * avail_bytes: Number of bytes from the start of the first
1428      * newly allocated to the end of the last newly allocated cluster.
1429      *
1430      * nb_bytes: The number of bytes from the start of the first
1431      * newly allocated cluster to the end of the area that the write
1432      * request actually writes to (excluding COW at the end)
1433      */
1434     uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset);
1435     int avail_bytes = MIN(INT_MAX, nb_clusters << s->cluster_bits);
1436     int nb_bytes = MIN(requested_bytes, avail_bytes);
1437     QCowL2Meta *old_m = *m;
1438 
1439     *m = g_malloc0(sizeof(**m));
1440 
1441     **m = (QCowL2Meta) {
1442         .next           = old_m,
1443 
1444         .alloc_offset   = alloc_cluster_offset,
1445         .offset         = start_of_cluster(s, guest_offset),
1446         .nb_clusters    = nb_clusters,
1447 
1448         .keep_old_clusters  = keep_old_clusters,
1449 
1450         .cow_start = {
1451             .offset     = 0,
1452             .nb_bytes   = offset_into_cluster(s, guest_offset),
1453         },
1454         .cow_end = {
1455             .offset     = nb_bytes,
1456             .nb_bytes   = avail_bytes - nb_bytes,
1457         },
1458     };
1459     qemu_co_queue_init(&(*m)->dependent_requests);
1460     QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1461 
1462     *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1463     *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset));
1464     assert(*bytes != 0);
1465 
1466     return 1;
1467 
1468 fail:
1469     if (*m && (*m)->nb_clusters > 0) {
1470         QLIST_REMOVE(*m, next_in_flight);
1471     }
1472     return ret;
1473 }
1474 
1475 /*
1476  * alloc_cluster_offset
1477  *
1478  * For a given offset on the virtual disk, find the cluster offset in qcow2
1479  * file. If the offset is not found, allocate a new cluster.
1480  *
1481  * If the cluster was already allocated, m->nb_clusters is set to 0 and
1482  * other fields in m are meaningless.
1483  *
1484  * If the cluster is newly allocated, m->nb_clusters is set to the number of
1485  * contiguous clusters that have been allocated. In this case, the other
1486  * fields of m are valid and contain information about the first allocated
1487  * cluster.
1488  *
1489  * If the request conflicts with another write request in flight, the coroutine
1490  * is queued and will be reentered when the dependency has completed.
1491  *
1492  * Return 0 on success and -errno in error cases
1493  */
1494 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1495                                unsigned int *bytes, uint64_t *host_offset,
1496                                QCowL2Meta **m)
1497 {
1498     BDRVQcow2State *s = bs->opaque;
1499     uint64_t start, remaining;
1500     uint64_t cluster_offset;
1501     uint64_t cur_bytes;
1502     int ret;
1503 
1504     trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes);
1505 
1506 again:
1507     start = offset;
1508     remaining = *bytes;
1509     cluster_offset = INV_OFFSET;
1510     *host_offset = INV_OFFSET;
1511     cur_bytes = 0;
1512     *m = NULL;
1513 
1514     while (true) {
1515 
1516         if (*host_offset == INV_OFFSET && cluster_offset != INV_OFFSET) {
1517             *host_offset = start_of_cluster(s, cluster_offset);
1518         }
1519 
1520         assert(remaining >= cur_bytes);
1521 
1522         start           += cur_bytes;
1523         remaining       -= cur_bytes;
1524 
1525         if (cluster_offset != INV_OFFSET) {
1526             cluster_offset += cur_bytes;
1527         }
1528 
1529         if (remaining == 0) {
1530             break;
1531         }
1532 
1533         cur_bytes = remaining;
1534 
1535         /*
1536          * Now start gathering as many contiguous clusters as possible:
1537          *
1538          * 1. Check for overlaps with in-flight allocations
1539          *
1540          *      a) Overlap not in the first cluster -> shorten this request and
1541          *         let the caller handle the rest in its next loop iteration.
1542          *
1543          *      b) Real overlaps of two requests. Yield and restart the search
1544          *         for contiguous clusters (the situation could have changed
1545          *         while we were sleeping)
1546          *
1547          *      c) TODO: Request starts in the same cluster as the in-flight
1548          *         allocation ends. Shorten the COW of the in-fight allocation,
1549          *         set cluster_offset to write to the same cluster and set up
1550          *         the right synchronisation between the in-flight request and
1551          *         the new one.
1552          */
1553         ret = handle_dependencies(bs, start, &cur_bytes, m);
1554         if (ret == -EAGAIN) {
1555             /* Currently handle_dependencies() doesn't yield if we already had
1556              * an allocation. If it did, we would have to clean up the L2Meta
1557              * structs before starting over. */
1558             assert(*m == NULL);
1559             goto again;
1560         } else if (ret < 0) {
1561             return ret;
1562         } else if (cur_bytes == 0) {
1563             break;
1564         } else {
1565             /* handle_dependencies() may have decreased cur_bytes (shortened
1566              * the allocations below) so that the next dependency is processed
1567              * correctly during the next loop iteration. */
1568         }
1569 
1570         /*
1571          * 2. Count contiguous COPIED clusters.
1572          */
1573         ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1574         if (ret < 0) {
1575             return ret;
1576         } else if (ret) {
1577             continue;
1578         } else if (cur_bytes == 0) {
1579             break;
1580         }
1581 
1582         /*
1583          * 3. If the request still hasn't completed, allocate new clusters,
1584          *    considering any cluster_offset of steps 1c or 2.
1585          */
1586         ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1587         if (ret < 0) {
1588             return ret;
1589         } else if (ret) {
1590             continue;
1591         } else {
1592             assert(cur_bytes == 0);
1593             break;
1594         }
1595     }
1596 
1597     *bytes -= remaining;
1598     assert(*bytes > 0);
1599     assert(*host_offset != INV_OFFSET);
1600 
1601     return 0;
1602 }
1603 
1604 /*
1605  * This discards as many clusters of nb_clusters as possible at once (i.e.
1606  * all clusters in the same L2 slice) and returns the number of discarded
1607  * clusters.
1608  */
1609 static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1610                                uint64_t nb_clusters,
1611                                enum qcow2_discard_type type, bool full_discard)
1612 {
1613     BDRVQcow2State *s = bs->opaque;
1614     uint64_t *l2_slice;
1615     int l2_index;
1616     int ret;
1617     int i;
1618 
1619     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1620     if (ret < 0) {
1621         return ret;
1622     }
1623 
1624     /* Limit nb_clusters to one L2 slice */
1625     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1626     assert(nb_clusters <= INT_MAX);
1627 
1628     for (i = 0; i < nb_clusters; i++) {
1629         uint64_t old_l2_entry;
1630 
1631         old_l2_entry = be64_to_cpu(l2_slice[l2_index + i]);
1632 
1633         /*
1634          * If full_discard is false, make sure that a discarded area reads back
1635          * as zeroes for v3 images (we cannot do it for v2 without actually
1636          * writing a zero-filled buffer). We can skip the operation if the
1637          * cluster is already marked as zero, or if it's unallocated and we
1638          * don't have a backing file.
1639          *
1640          * TODO We might want to use bdrv_block_status(bs) here, but we're
1641          * holding s->lock, so that doesn't work today.
1642          *
1643          * If full_discard is true, the sector should not read back as zeroes,
1644          * but rather fall through to the backing file.
1645          */
1646         switch (qcow2_get_cluster_type(bs, old_l2_entry)) {
1647         case QCOW2_CLUSTER_UNALLOCATED:
1648             if (full_discard || !bs->backing) {
1649                 continue;
1650             }
1651             break;
1652 
1653         case QCOW2_CLUSTER_ZERO_PLAIN:
1654             if (!full_discard) {
1655                 continue;
1656             }
1657             break;
1658 
1659         case QCOW2_CLUSTER_ZERO_ALLOC:
1660         case QCOW2_CLUSTER_NORMAL:
1661         case QCOW2_CLUSTER_COMPRESSED:
1662             break;
1663 
1664         default:
1665             abort();
1666         }
1667 
1668         /* First remove L2 entries */
1669         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1670         if (!full_discard && s->qcow_version >= 3) {
1671             l2_slice[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1672         } else {
1673             l2_slice[l2_index + i] = cpu_to_be64(0);
1674         }
1675 
1676         /* Then decrease the refcount */
1677         qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1678     }
1679 
1680     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1681 
1682     return nb_clusters;
1683 }
1684 
1685 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset,
1686                           uint64_t bytes, enum qcow2_discard_type type,
1687                           bool full_discard)
1688 {
1689     BDRVQcow2State *s = bs->opaque;
1690     uint64_t end_offset = offset + bytes;
1691     uint64_t nb_clusters;
1692     int64_t cleared;
1693     int ret;
1694 
1695     /* Caller must pass aligned values, except at image end */
1696     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1697     assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1698            end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
1699 
1700     nb_clusters = size_to_clusters(s, bytes);
1701 
1702     s->cache_discards = true;
1703 
1704     /* Each L2 slice is handled by its own loop iteration */
1705     while (nb_clusters > 0) {
1706         cleared = discard_in_l2_slice(bs, offset, nb_clusters, type,
1707                                       full_discard);
1708         if (cleared < 0) {
1709             ret = cleared;
1710             goto fail;
1711         }
1712 
1713         nb_clusters -= cleared;
1714         offset += (cleared * s->cluster_size);
1715     }
1716 
1717     ret = 0;
1718 fail:
1719     s->cache_discards = false;
1720     qcow2_process_discards(bs, ret);
1721 
1722     return ret;
1723 }
1724 
1725 /*
1726  * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1727  * all clusters in the same L2 slice) and returns the number of zeroed
1728  * clusters.
1729  */
1730 static int zero_in_l2_slice(BlockDriverState *bs, uint64_t offset,
1731                             uint64_t nb_clusters, int flags)
1732 {
1733     BDRVQcow2State *s = bs->opaque;
1734     uint64_t *l2_slice;
1735     int l2_index;
1736     int ret;
1737     int i;
1738     bool unmap = !!(flags & BDRV_REQ_MAY_UNMAP);
1739 
1740     ret = get_cluster_table(bs, offset, &l2_slice, &l2_index);
1741     if (ret < 0) {
1742         return ret;
1743     }
1744 
1745     /* Limit nb_clusters to one L2 slice */
1746     nb_clusters = MIN(nb_clusters, s->l2_slice_size - l2_index);
1747     assert(nb_clusters <= INT_MAX);
1748 
1749     for (i = 0; i < nb_clusters; i++) {
1750         uint64_t old_offset;
1751         QCow2ClusterType cluster_type;
1752 
1753         old_offset = be64_to_cpu(l2_slice[l2_index + i]);
1754 
1755         /*
1756          * Minimize L2 changes if the cluster already reads back as
1757          * zeroes with correct allocation.
1758          */
1759         cluster_type = qcow2_get_cluster_type(bs, old_offset);
1760         if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN ||
1761             (cluster_type == QCOW2_CLUSTER_ZERO_ALLOC && !unmap)) {
1762             continue;
1763         }
1764 
1765         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_slice);
1766         if (cluster_type == QCOW2_CLUSTER_COMPRESSED || unmap) {
1767             l2_slice[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1768             qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1769         } else {
1770             l2_slice[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1771         }
1772     }
1773 
1774     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
1775 
1776     return nb_clusters;
1777 }
1778 
1779 int qcow2_cluster_zeroize(BlockDriverState *bs, uint64_t offset,
1780                           uint64_t bytes, int flags)
1781 {
1782     BDRVQcow2State *s = bs->opaque;
1783     uint64_t end_offset = offset + bytes;
1784     uint64_t nb_clusters;
1785     int64_t cleared;
1786     int ret;
1787 
1788     /* If we have to stay in sync with an external data file, zero out
1789      * s->data_file first. */
1790     if (data_file_is_raw(bs)) {
1791         assert(has_data_file(bs));
1792         ret = bdrv_co_pwrite_zeroes(s->data_file, offset, bytes, flags);
1793         if (ret < 0) {
1794             return ret;
1795         }
1796     }
1797 
1798     /* Caller must pass aligned values, except at image end */
1799     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
1800     assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) ||
1801            end_offset == bs->total_sectors << BDRV_SECTOR_BITS);
1802 
1803     /* The zero flag is only supported by version 3 and newer */
1804     if (s->qcow_version < 3) {
1805         return -ENOTSUP;
1806     }
1807 
1808     /* Each L2 slice is handled by its own loop iteration */
1809     nb_clusters = size_to_clusters(s, bytes);
1810 
1811     s->cache_discards = true;
1812 
1813     while (nb_clusters > 0) {
1814         cleared = zero_in_l2_slice(bs, offset, nb_clusters, flags);
1815         if (cleared < 0) {
1816             ret = cleared;
1817             goto fail;
1818         }
1819 
1820         nb_clusters -= cleared;
1821         offset += (cleared * s->cluster_size);
1822     }
1823 
1824     ret = 0;
1825 fail:
1826     s->cache_discards = false;
1827     qcow2_process_discards(bs, ret);
1828 
1829     return ret;
1830 }
1831 
1832 /*
1833  * Expands all zero clusters in a specific L1 table (or deallocates them, for
1834  * non-backed non-pre-allocated zero clusters).
1835  *
1836  * l1_entries and *visited_l1_entries are used to keep track of progress for
1837  * status_cb(). l1_entries contains the total number of L1 entries and
1838  * *visited_l1_entries counts all visited L1 entries.
1839  */
1840 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1841                                       int l1_size, int64_t *visited_l1_entries,
1842                                       int64_t l1_entries,
1843                                       BlockDriverAmendStatusCB *status_cb,
1844                                       void *cb_opaque)
1845 {
1846     BDRVQcow2State *s = bs->opaque;
1847     bool is_active_l1 = (l1_table == s->l1_table);
1848     uint64_t *l2_slice = NULL;
1849     unsigned slice, slice_size2, n_slices;
1850     int ret;
1851     int i, j;
1852 
1853     slice_size2 = s->l2_slice_size * sizeof(uint64_t);
1854     n_slices = s->cluster_size / slice_size2;
1855 
1856     if (!is_active_l1) {
1857         /* inactive L2 tables require a buffer to be stored in when loading
1858          * them from disk */
1859         l2_slice = qemu_try_blockalign(bs->file->bs, slice_size2);
1860         if (l2_slice == NULL) {
1861             return -ENOMEM;
1862         }
1863     }
1864 
1865     for (i = 0; i < l1_size; i++) {
1866         uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1867         uint64_t l2_refcount;
1868 
1869         if (!l2_offset) {
1870             /* unallocated */
1871             (*visited_l1_entries)++;
1872             if (status_cb) {
1873                 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque);
1874             }
1875             continue;
1876         }
1877 
1878         if (offset_into_cluster(s, l2_offset)) {
1879             qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#"
1880                                     PRIx64 " unaligned (L1 index: %#x)",
1881                                     l2_offset, i);
1882             ret = -EIO;
1883             goto fail;
1884         }
1885 
1886         ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits,
1887                                  &l2_refcount);
1888         if (ret < 0) {
1889             goto fail;
1890         }
1891 
1892         for (slice = 0; slice < n_slices; slice++) {
1893             uint64_t slice_offset = l2_offset + slice * slice_size2;
1894             bool l2_dirty = false;
1895             if (is_active_l1) {
1896                 /* get active L2 tables from cache */
1897                 ret = qcow2_cache_get(bs, s->l2_table_cache, slice_offset,
1898                                       (void **)&l2_slice);
1899             } else {
1900                 /* load inactive L2 tables from disk */
1901                 ret = bdrv_pread(bs->file, slice_offset, l2_slice, slice_size2);
1902             }
1903             if (ret < 0) {
1904                 goto fail;
1905             }
1906 
1907             for (j = 0; j < s->l2_slice_size; j++) {
1908                 uint64_t l2_entry = be64_to_cpu(l2_slice[j]);
1909                 int64_t offset = l2_entry & L2E_OFFSET_MASK;
1910                 QCow2ClusterType cluster_type =
1911                     qcow2_get_cluster_type(bs, l2_entry);
1912 
1913                 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN &&
1914                     cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) {
1915                     continue;
1916                 }
1917 
1918                 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) {
1919                     if (!bs->backing) {
1920                         /* not backed; therefore we can simply deallocate the
1921                          * cluster */
1922                         l2_slice[j] = 0;
1923                         l2_dirty = true;
1924                         continue;
1925                     }
1926 
1927                     offset = qcow2_alloc_clusters(bs, s->cluster_size);
1928                     if (offset < 0) {
1929                         ret = offset;
1930                         goto fail;
1931                     }
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