xref: /openbmc/qemu/block/qcow2-cluster.c (revision 52f91c37)
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 <zlib.h>
26 
27 #include "qemu-common.h"
28 #include "block/block_int.h"
29 #include "block/qcow2.h"
30 #include "trace.h"
31 
32 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size,
33                         bool exact_size)
34 {
35     BDRVQcowState *s = bs->opaque;
36     int new_l1_size2, ret, i;
37     uint64_t *new_l1_table;
38     int64_t old_l1_table_offset, old_l1_size;
39     int64_t new_l1_table_offset, new_l1_size;
40     uint8_t data[12];
41 
42     if (min_size <= s->l1_size)
43         return 0;
44 
45     /* Do a sanity check on min_size before trying to calculate new_l1_size
46      * (this prevents overflows during the while loop for the calculation of
47      * new_l1_size) */
48     if (min_size > INT_MAX / sizeof(uint64_t)) {
49         return -EFBIG;
50     }
51 
52     if (exact_size) {
53         new_l1_size = min_size;
54     } else {
55         /* Bump size up to reduce the number of times we have to grow */
56         new_l1_size = s->l1_size;
57         if (new_l1_size == 0) {
58             new_l1_size = 1;
59         }
60         while (min_size > new_l1_size) {
61             new_l1_size = (new_l1_size * 3 + 1) / 2;
62         }
63     }
64 
65     if (new_l1_size > INT_MAX / sizeof(uint64_t)) {
66         return -EFBIG;
67     }
68 
69 #ifdef DEBUG_ALLOC2
70     fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n",
71             s->l1_size, new_l1_size);
72 #endif
73 
74     new_l1_size2 = sizeof(uint64_t) * new_l1_size;
75     new_l1_table = g_malloc0(align_offset(new_l1_size2, 512));
76     memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t));
77 
78     /* write new table (align to cluster) */
79     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE);
80     new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2);
81     if (new_l1_table_offset < 0) {
82         g_free(new_l1_table);
83         return new_l1_table_offset;
84     }
85 
86     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
87     if (ret < 0) {
88         goto fail;
89     }
90 
91     /* the L1 position has not yet been updated, so these clusters must
92      * indeed be completely free */
93     ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset,
94                                         new_l1_size2);
95     if (ret < 0) {
96         goto fail;
97     }
98 
99     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE);
100     for(i = 0; i < s->l1_size; i++)
101         new_l1_table[i] = cpu_to_be64(new_l1_table[i]);
102     ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_table, new_l1_size2);
103     if (ret < 0)
104         goto fail;
105     for(i = 0; i < s->l1_size; i++)
106         new_l1_table[i] = be64_to_cpu(new_l1_table[i]);
107 
108     /* set new table */
109     BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE);
110     cpu_to_be32w((uint32_t*)data, new_l1_size);
111     stq_be_p(data + 4, new_l1_table_offset);
112     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), data,sizeof(data));
113     if (ret < 0) {
114         goto fail;
115     }
116     g_free(s->l1_table);
117     old_l1_table_offset = s->l1_table_offset;
118     s->l1_table_offset = new_l1_table_offset;
119     s->l1_table = new_l1_table;
120     old_l1_size = s->l1_size;
121     s->l1_size = new_l1_size;
122     qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t),
123                         QCOW2_DISCARD_OTHER);
124     return 0;
125  fail:
126     g_free(new_l1_table);
127     qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2,
128                         QCOW2_DISCARD_OTHER);
129     return ret;
130 }
131 
132 /*
133  * l2_load
134  *
135  * Loads a L2 table into memory. If the table is in the cache, the cache
136  * is used; otherwise the L2 table is loaded from the image file.
137  *
138  * Returns a pointer to the L2 table on success, or NULL if the read from
139  * the image file failed.
140  */
141 
142 static int l2_load(BlockDriverState *bs, uint64_t l2_offset,
143     uint64_t **l2_table)
144 {
145     BDRVQcowState *s = bs->opaque;
146     int ret;
147 
148     ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, (void**) l2_table);
149 
150     return ret;
151 }
152 
153 /*
154  * Writes one sector of the L1 table to the disk (can't update single entries
155  * and we really don't want bdrv_pread to perform a read-modify-write)
156  */
157 #define L1_ENTRIES_PER_SECTOR (512 / 8)
158 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index)
159 {
160     BDRVQcowState *s = bs->opaque;
161     uint64_t buf[L1_ENTRIES_PER_SECTOR];
162     int l1_start_index;
163     int i, ret;
164 
165     l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1);
166     for (i = 0; i < L1_ENTRIES_PER_SECTOR; i++) {
167         buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]);
168     }
169 
170     ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1,
171             s->l1_table_offset + 8 * l1_start_index, sizeof(buf));
172     if (ret < 0) {
173         return ret;
174     }
175 
176     BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE);
177     ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset + 8 * l1_start_index,
178         buf, sizeof(buf));
179     if (ret < 0) {
180         return ret;
181     }
182 
183     return 0;
184 }
185 
186 /*
187  * l2_allocate
188  *
189  * Allocate a new l2 entry in the file. If l1_index points to an already
190  * used entry in the L2 table (i.e. we are doing a copy on write for the L2
191  * table) copy the contents of the old L2 table into the newly allocated one.
192  * Otherwise the new table is initialized with zeros.
193  *
194  */
195 
196 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table)
197 {
198     BDRVQcowState *s = bs->opaque;
199     uint64_t old_l2_offset;
200     uint64_t *l2_table = NULL;
201     int64_t l2_offset;
202     int ret;
203 
204     old_l2_offset = s->l1_table[l1_index];
205 
206     trace_qcow2_l2_allocate(bs, l1_index);
207 
208     /* allocate a new l2 entry */
209 
210     l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t));
211     if (l2_offset < 0) {
212         ret = l2_offset;
213         goto fail;
214     }
215 
216     ret = qcow2_cache_flush(bs, s->refcount_block_cache);
217     if (ret < 0) {
218         goto fail;
219     }
220 
221     /* allocate a new entry in the l2 cache */
222 
223     trace_qcow2_l2_allocate_get_empty(bs, l1_index);
224     ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table);
225     if (ret < 0) {
226         goto fail;
227     }
228 
229     l2_table = *table;
230 
231     if ((old_l2_offset & L1E_OFFSET_MASK) == 0) {
232         /* if there was no old l2 table, clear the new table */
233         memset(l2_table, 0, s->l2_size * sizeof(uint64_t));
234     } else {
235         uint64_t* old_table;
236 
237         /* if there was an old l2 table, read it from the disk */
238         BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ);
239         ret = qcow2_cache_get(bs, s->l2_table_cache,
240             old_l2_offset & L1E_OFFSET_MASK,
241             (void**) &old_table);
242         if (ret < 0) {
243             goto fail;
244         }
245 
246         memcpy(l2_table, old_table, s->cluster_size);
247 
248         ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &old_table);
249         if (ret < 0) {
250             goto fail;
251         }
252     }
253 
254     /* write the l2 table to the file */
255     BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE);
256 
257     trace_qcow2_l2_allocate_write_l2(bs, l1_index);
258     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
259     ret = qcow2_cache_flush(bs, s->l2_table_cache);
260     if (ret < 0) {
261         goto fail;
262     }
263 
264     /* update the L1 entry */
265     trace_qcow2_l2_allocate_write_l1(bs, l1_index);
266     s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED;
267     ret = qcow2_write_l1_entry(bs, l1_index);
268     if (ret < 0) {
269         goto fail;
270     }
271 
272     *table = l2_table;
273     trace_qcow2_l2_allocate_done(bs, l1_index, 0);
274     return 0;
275 
276 fail:
277     trace_qcow2_l2_allocate_done(bs, l1_index, ret);
278     if (l2_table != NULL) {
279         qcow2_cache_put(bs, s->l2_table_cache, (void**) table);
280     }
281     s->l1_table[l1_index] = old_l2_offset;
282     if (l2_offset > 0) {
283         qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
284                             QCOW2_DISCARD_ALWAYS);
285     }
286     return ret;
287 }
288 
289 /*
290  * Checks how many clusters in a given L2 table are contiguous in the image
291  * file. As soon as one of the flags in the bitmask stop_flags changes compared
292  * to the first cluster, the search is stopped and the cluster is not counted
293  * as contiguous. (This allows it, for example, to stop at the first compressed
294  * cluster which may require a different handling)
295  */
296 static int count_contiguous_clusters(uint64_t nb_clusters, int cluster_size,
297         uint64_t *l2_table, uint64_t stop_flags)
298 {
299     int i;
300     uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED;
301     uint64_t first_entry = be64_to_cpu(l2_table[0]);
302     uint64_t offset = first_entry & mask;
303 
304     if (!offset)
305         return 0;
306 
307     assert(qcow2_get_cluster_type(first_entry) != QCOW2_CLUSTER_COMPRESSED);
308 
309     for (i = 0; i < nb_clusters; i++) {
310         uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask;
311         if (offset + (uint64_t) i * cluster_size != l2_entry) {
312             break;
313         }
314     }
315 
316 	return i;
317 }
318 
319 static int count_contiguous_free_clusters(uint64_t nb_clusters, uint64_t *l2_table)
320 {
321     int i;
322 
323     for (i = 0; i < nb_clusters; i++) {
324         int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i]));
325 
326         if (type != QCOW2_CLUSTER_UNALLOCATED) {
327             break;
328         }
329     }
330 
331     return i;
332 }
333 
334 /* The crypt function is compatible with the linux cryptoloop
335    algorithm for < 4 GB images. NOTE: out_buf == in_buf is
336    supported */
337 void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num,
338                            uint8_t *out_buf, const uint8_t *in_buf,
339                            int nb_sectors, int enc,
340                            const AES_KEY *key)
341 {
342     union {
343         uint64_t ll[2];
344         uint8_t b[16];
345     } ivec;
346     int i;
347 
348     for(i = 0; i < nb_sectors; i++) {
349         ivec.ll[0] = cpu_to_le64(sector_num);
350         ivec.ll[1] = 0;
351         AES_cbc_encrypt(in_buf, out_buf, 512, key,
352                         ivec.b, enc);
353         sector_num++;
354         in_buf += 512;
355         out_buf += 512;
356     }
357 }
358 
359 static int coroutine_fn copy_sectors(BlockDriverState *bs,
360                                      uint64_t start_sect,
361                                      uint64_t cluster_offset,
362                                      int n_start, int n_end)
363 {
364     BDRVQcowState *s = bs->opaque;
365     QEMUIOVector qiov;
366     struct iovec iov;
367     int n, ret;
368 
369     n = n_end - n_start;
370     if (n <= 0) {
371         return 0;
372     }
373 
374     iov.iov_len = n * BDRV_SECTOR_SIZE;
375     iov.iov_base = qemu_blockalign(bs, iov.iov_len);
376 
377     qemu_iovec_init_external(&qiov, &iov, 1);
378 
379     BLKDBG_EVENT(bs->file, BLKDBG_COW_READ);
380 
381     if (!bs->drv) {
382         return -ENOMEDIUM;
383     }
384 
385     /* Call .bdrv_co_readv() directly instead of using the public block-layer
386      * interface.  This avoids double I/O throttling and request tracking,
387      * which can lead to deadlock when block layer copy-on-read is enabled.
388      */
389     ret = bs->drv->bdrv_co_readv(bs, start_sect + n_start, n, &qiov);
390     if (ret < 0) {
391         goto out;
392     }
393 
394     if (s->crypt_method) {
395         qcow2_encrypt_sectors(s, start_sect + n_start,
396                         iov.iov_base, iov.iov_base, n, 1,
397                         &s->aes_encrypt_key);
398     }
399 
400     ret = qcow2_pre_write_overlap_check(bs, 0,
401             cluster_offset + n_start * BDRV_SECTOR_SIZE, n * BDRV_SECTOR_SIZE);
402     if (ret < 0) {
403         goto out;
404     }
405 
406     BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE);
407     ret = bdrv_co_writev(bs->file, (cluster_offset >> 9) + n_start, n, &qiov);
408     if (ret < 0) {
409         goto out;
410     }
411 
412     ret = 0;
413 out:
414     qemu_vfree(iov.iov_base);
415     return ret;
416 }
417 
418 
419 /*
420  * get_cluster_offset
421  *
422  * For a given offset of the disk image, find the cluster offset in
423  * qcow2 file. The offset is stored in *cluster_offset.
424  *
425  * on entry, *num is the number of contiguous sectors we'd like to
426  * access following offset.
427  *
428  * on exit, *num is the number of contiguous sectors we can read.
429  *
430  * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
431  * cases.
432  */
433 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
434     int *num, uint64_t *cluster_offset)
435 {
436     BDRVQcowState *s = bs->opaque;
437     unsigned int l2_index;
438     uint64_t l1_index, l2_offset, *l2_table;
439     int l1_bits, c;
440     unsigned int index_in_cluster, nb_clusters;
441     uint64_t nb_available, nb_needed;
442     int ret;
443 
444     index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1);
445     nb_needed = *num + index_in_cluster;
446 
447     l1_bits = s->l2_bits + s->cluster_bits;
448 
449     /* compute how many bytes there are between the offset and
450      * the end of the l1 entry
451      */
452 
453     nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1));
454 
455     /* compute the number of available sectors */
456 
457     nb_available = (nb_available >> 9) + index_in_cluster;
458 
459     if (nb_needed > nb_available) {
460         nb_needed = nb_available;
461     }
462 
463     *cluster_offset = 0;
464 
465     /* seek the the l2 offset in the l1 table */
466 
467     l1_index = offset >> l1_bits;
468     if (l1_index >= s->l1_size) {
469         ret = QCOW2_CLUSTER_UNALLOCATED;
470         goto out;
471     }
472 
473     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
474     if (!l2_offset) {
475         ret = QCOW2_CLUSTER_UNALLOCATED;
476         goto out;
477     }
478 
479     /* load the l2 table in memory */
480 
481     ret = l2_load(bs, l2_offset, &l2_table);
482     if (ret < 0) {
483         return ret;
484     }
485 
486     /* find the cluster offset for the given disk offset */
487 
488     l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
489     *cluster_offset = be64_to_cpu(l2_table[l2_index]);
490     nb_clusters = size_to_clusters(s, nb_needed << 9);
491 
492     ret = qcow2_get_cluster_type(*cluster_offset);
493     switch (ret) {
494     case QCOW2_CLUSTER_COMPRESSED:
495         /* Compressed clusters can only be processed one by one */
496         c = 1;
497         *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK;
498         break;
499     case QCOW2_CLUSTER_ZERO:
500         if (s->qcow_version < 3) {
501             qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
502             return -EIO;
503         }
504         c = count_contiguous_clusters(nb_clusters, s->cluster_size,
505                 &l2_table[l2_index], QCOW_OFLAG_ZERO);
506         *cluster_offset = 0;
507         break;
508     case QCOW2_CLUSTER_UNALLOCATED:
509         /* how many empty clusters ? */
510         c = count_contiguous_free_clusters(nb_clusters, &l2_table[l2_index]);
511         *cluster_offset = 0;
512         break;
513     case QCOW2_CLUSTER_NORMAL:
514         /* how many allocated clusters ? */
515         c = count_contiguous_clusters(nb_clusters, s->cluster_size,
516                 &l2_table[l2_index], QCOW_OFLAG_ZERO);
517         *cluster_offset &= L2E_OFFSET_MASK;
518         break;
519     default:
520         abort();
521     }
522 
523     qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
524 
525     nb_available = (c * s->cluster_sectors);
526 
527 out:
528     if (nb_available > nb_needed)
529         nb_available = nb_needed;
530 
531     *num = nb_available - index_in_cluster;
532 
533     return ret;
534 }
535 
536 /*
537  * get_cluster_table
538  *
539  * for a given disk offset, load (and allocate if needed)
540  * the l2 table.
541  *
542  * the l2 table offset in the qcow2 file and the cluster index
543  * in the l2 table are given to the caller.
544  *
545  * Returns 0 on success, -errno in failure case
546  */
547 static int get_cluster_table(BlockDriverState *bs, uint64_t offset,
548                              uint64_t **new_l2_table,
549                              int *new_l2_index)
550 {
551     BDRVQcowState *s = bs->opaque;
552     unsigned int l2_index;
553     uint64_t l1_index, l2_offset;
554     uint64_t *l2_table = NULL;
555     int ret;
556 
557     /* seek the the l2 offset in the l1 table */
558 
559     l1_index = offset >> (s->l2_bits + s->cluster_bits);
560     if (l1_index >= s->l1_size) {
561         ret = qcow2_grow_l1_table(bs, l1_index + 1, false);
562         if (ret < 0) {
563             return ret;
564         }
565     }
566 
567     assert(l1_index < s->l1_size);
568     l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK;
569 
570     /* seek the l2 table of the given l2 offset */
571 
572     if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) {
573         /* load the l2 table in memory */
574         ret = l2_load(bs, l2_offset, &l2_table);
575         if (ret < 0) {
576             return ret;
577         }
578     } else {
579         /* First allocate a new L2 table (and do COW if needed) */
580         ret = l2_allocate(bs, l1_index, &l2_table);
581         if (ret < 0) {
582             return ret;
583         }
584 
585         /* Then decrease the refcount of the old table */
586         if (l2_offset) {
587             qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t),
588                                 QCOW2_DISCARD_OTHER);
589         }
590     }
591 
592     /* find the cluster offset for the given disk offset */
593 
594     l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1);
595 
596     *new_l2_table = l2_table;
597     *new_l2_index = l2_index;
598 
599     return 0;
600 }
601 
602 /*
603  * alloc_compressed_cluster_offset
604  *
605  * For a given offset of the disk image, return cluster offset in
606  * qcow2 file.
607  *
608  * If the offset is not found, allocate a new compressed cluster.
609  *
610  * Return the cluster offset if successful,
611  * Return 0, otherwise.
612  *
613  */
614 
615 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs,
616                                                uint64_t offset,
617                                                int compressed_size)
618 {
619     BDRVQcowState *s = bs->opaque;
620     int l2_index, ret;
621     uint64_t *l2_table;
622     int64_t cluster_offset;
623     int nb_csectors;
624 
625     ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
626     if (ret < 0) {
627         return 0;
628     }
629 
630     /* Compression can't overwrite anything. Fail if the cluster was already
631      * allocated. */
632     cluster_offset = be64_to_cpu(l2_table[l2_index]);
633     if (cluster_offset & L2E_OFFSET_MASK) {
634         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
635         return 0;
636     }
637 
638     cluster_offset = qcow2_alloc_bytes(bs, compressed_size);
639     if (cluster_offset < 0) {
640         qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
641         return 0;
642     }
643 
644     nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) -
645                   (cluster_offset >> 9);
646 
647     cluster_offset |= QCOW_OFLAG_COMPRESSED |
648                       ((uint64_t)nb_csectors << s->csize_shift);
649 
650     /* update L2 table */
651 
652     /* compressed clusters never have the copied flag */
653 
654     BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED);
655     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
656     l2_table[l2_index] = cpu_to_be64(cluster_offset);
657     ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
658     if (ret < 0) {
659         return 0;
660     }
661 
662     return cluster_offset;
663 }
664 
665 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r)
666 {
667     BDRVQcowState *s = bs->opaque;
668     int ret;
669 
670     if (r->nb_sectors == 0) {
671         return 0;
672     }
673 
674     qemu_co_mutex_unlock(&s->lock);
675     ret = copy_sectors(bs, m->offset / BDRV_SECTOR_SIZE, m->alloc_offset,
676                        r->offset / BDRV_SECTOR_SIZE,
677                        r->offset / BDRV_SECTOR_SIZE + r->nb_sectors);
678     qemu_co_mutex_lock(&s->lock);
679 
680     if (ret < 0) {
681         return ret;
682     }
683 
684     /*
685      * Before we update the L2 table to actually point to the new cluster, we
686      * need to be sure that the refcounts have been increased and COW was
687      * handled.
688      */
689     qcow2_cache_depends_on_flush(s->l2_table_cache);
690 
691     return 0;
692 }
693 
694 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m)
695 {
696     BDRVQcowState *s = bs->opaque;
697     int i, j = 0, l2_index, ret;
698     uint64_t *old_cluster, *l2_table;
699     uint64_t cluster_offset = m->alloc_offset;
700 
701     trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters);
702     assert(m->nb_clusters > 0);
703 
704     old_cluster = g_malloc(m->nb_clusters * sizeof(uint64_t));
705 
706     /* copy content of unmodified sectors */
707     ret = perform_cow(bs, m, &m->cow_start);
708     if (ret < 0) {
709         goto err;
710     }
711 
712     ret = perform_cow(bs, m, &m->cow_end);
713     if (ret < 0) {
714         goto err;
715     }
716 
717     /* Update L2 table. */
718     if (s->use_lazy_refcounts) {
719         qcow2_mark_dirty(bs);
720     }
721     if (qcow2_need_accurate_refcounts(s)) {
722         qcow2_cache_set_dependency(bs, s->l2_table_cache,
723                                    s->refcount_block_cache);
724     }
725 
726     ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index);
727     if (ret < 0) {
728         goto err;
729     }
730     qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
731 
732     assert(l2_index + m->nb_clusters <= s->l2_size);
733     for (i = 0; i < m->nb_clusters; i++) {
734         /* if two concurrent writes happen to the same unallocated cluster
735 	 * each write allocates separate cluster and writes data concurrently.
736 	 * The first one to complete updates l2 table with pointer to its
737 	 * cluster the second one has to do RMW (which is done above by
738 	 * copy_sectors()), update l2 table with its cluster pointer and free
739 	 * old cluster. This is what this loop does */
740         if(l2_table[l2_index + i] != 0)
741             old_cluster[j++] = l2_table[l2_index + i];
742 
743         l2_table[l2_index + i] = cpu_to_be64((cluster_offset +
744                     (i << s->cluster_bits)) | QCOW_OFLAG_COPIED);
745      }
746 
747 
748     ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
749     if (ret < 0) {
750         goto err;
751     }
752 
753     /*
754      * If this was a COW, we need to decrease the refcount of the old cluster.
755      * Also flush bs->file to get the right order for L2 and refcount update.
756      *
757      * Don't discard clusters that reach a refcount of 0 (e.g. compressed
758      * clusters), the next write will reuse them anyway.
759      */
760     if (j != 0) {
761         for (i = 0; i < j; i++) {
762             qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1,
763                                     QCOW2_DISCARD_NEVER);
764         }
765     }
766 
767     ret = 0;
768 err:
769     g_free(old_cluster);
770     return ret;
771  }
772 
773 /*
774  * Returns the number of contiguous clusters that can be used for an allocating
775  * write, but require COW to be performed (this includes yet unallocated space,
776  * which must copy from the backing file)
777  */
778 static int count_cow_clusters(BDRVQcowState *s, int nb_clusters,
779     uint64_t *l2_table, int l2_index)
780 {
781     int i;
782 
783     for (i = 0; i < nb_clusters; i++) {
784         uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]);
785         int cluster_type = qcow2_get_cluster_type(l2_entry);
786 
787         switch(cluster_type) {
788         case QCOW2_CLUSTER_NORMAL:
789             if (l2_entry & QCOW_OFLAG_COPIED) {
790                 goto out;
791             }
792             break;
793         case QCOW2_CLUSTER_UNALLOCATED:
794         case QCOW2_CLUSTER_COMPRESSED:
795         case QCOW2_CLUSTER_ZERO:
796             break;
797         default:
798             abort();
799         }
800     }
801 
802 out:
803     assert(i <= nb_clusters);
804     return i;
805 }
806 
807 /*
808  * Check if there already is an AIO write request in flight which allocates
809  * the same cluster. In this case we need to wait until the previous
810  * request has completed and updated the L2 table accordingly.
811  *
812  * Returns:
813  *   0       if there was no dependency. *cur_bytes indicates the number of
814  *           bytes from guest_offset that can be read before the next
815  *           dependency must be processed (or the request is complete)
816  *
817  *   -EAGAIN if we had to wait for another request, previously gathered
818  *           information on cluster allocation may be invalid now. The caller
819  *           must start over anyway, so consider *cur_bytes undefined.
820  */
821 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
822     uint64_t *cur_bytes, QCowL2Meta **m)
823 {
824     BDRVQcowState *s = bs->opaque;
825     QCowL2Meta *old_alloc;
826     uint64_t bytes = *cur_bytes;
827 
828     QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) {
829 
830         uint64_t start = guest_offset;
831         uint64_t end = start + bytes;
832         uint64_t old_start = l2meta_cow_start(old_alloc);
833         uint64_t old_end = l2meta_cow_end(old_alloc);
834 
835         if (end <= old_start || start >= old_end) {
836             /* No intersection */
837         } else {
838             if (start < old_start) {
839                 /* Stop at the start of a running allocation */
840                 bytes = old_start - start;
841             } else {
842                 bytes = 0;
843             }
844 
845             /* Stop if already an l2meta exists. After yielding, it wouldn't
846              * be valid any more, so we'd have to clean up the old L2Metas
847              * and deal with requests depending on them before starting to
848              * gather new ones. Not worth the trouble. */
849             if (bytes == 0 && *m) {
850                 *cur_bytes = 0;
851                 return 0;
852             }
853 
854             if (bytes == 0) {
855                 /* Wait for the dependency to complete. We need to recheck
856                  * the free/allocated clusters when we continue. */
857                 qemu_co_mutex_unlock(&s->lock);
858                 qemu_co_queue_wait(&old_alloc->dependent_requests);
859                 qemu_co_mutex_lock(&s->lock);
860                 return -EAGAIN;
861             }
862         }
863     }
864 
865     /* Make sure that existing clusters and new allocations are only used up to
866      * the next dependency if we shortened the request above */
867     *cur_bytes = bytes;
868 
869     return 0;
870 }
871 
872 /*
873  * Checks how many already allocated clusters that don't require a copy on
874  * write there are at the given guest_offset (up to *bytes). If
875  * *host_offset is not zero, only physically contiguous clusters beginning at
876  * this host offset are counted.
877  *
878  * Note that guest_offset may not be cluster aligned. In this case, the
879  * returned *host_offset points to exact byte referenced by guest_offset and
880  * therefore isn't cluster aligned as well.
881  *
882  * Returns:
883  *   0:     if no allocated clusters are available at the given offset.
884  *          *bytes is normally unchanged. It is set to 0 if the cluster
885  *          is allocated and doesn't need COW, but doesn't have the right
886  *          physical offset.
887  *
888  *   1:     if allocated clusters that don't require a COW are available at
889  *          the requested offset. *bytes may have decreased and describes
890  *          the length of the area that can be written to.
891  *
892  *  -errno: in error cases
893  */
894 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset,
895     uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
896 {
897     BDRVQcowState *s = bs->opaque;
898     int l2_index;
899     uint64_t cluster_offset;
900     uint64_t *l2_table;
901     unsigned int nb_clusters;
902     unsigned int keep_clusters;
903     int ret, pret;
904 
905     trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset,
906                               *bytes);
907 
908     assert(*host_offset == 0 ||    offset_into_cluster(s, guest_offset)
909                                 == offset_into_cluster(s, *host_offset));
910 
911     /*
912      * Calculate the number of clusters to look for. We stop at L2 table
913      * boundaries to keep things simple.
914      */
915     nb_clusters =
916         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
917 
918     l2_index = offset_to_l2_index(s, guest_offset);
919     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
920 
921     /* Find L2 entry for the first involved cluster */
922     ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
923     if (ret < 0) {
924         return ret;
925     }
926 
927     cluster_offset = be64_to_cpu(l2_table[l2_index]);
928 
929     /* Check how many clusters are already allocated and don't need COW */
930     if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL
931         && (cluster_offset & QCOW_OFLAG_COPIED))
932     {
933         /* If a specific host_offset is required, check it */
934         bool offset_matches =
935             (cluster_offset & L2E_OFFSET_MASK) == *host_offset;
936 
937         if (*host_offset != 0 && !offset_matches) {
938             *bytes = 0;
939             ret = 0;
940             goto out;
941         }
942 
943         /* We keep all QCOW_OFLAG_COPIED clusters */
944         keep_clusters =
945             count_contiguous_clusters(nb_clusters, s->cluster_size,
946                                       &l2_table[l2_index],
947                                       QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO);
948         assert(keep_clusters <= nb_clusters);
949 
950         *bytes = MIN(*bytes,
951                  keep_clusters * s->cluster_size
952                  - offset_into_cluster(s, guest_offset));
953 
954         ret = 1;
955     } else {
956         ret = 0;
957     }
958 
959     /* Cleanup */
960 out:
961     pret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
962     if (pret < 0) {
963         return pret;
964     }
965 
966     /* Only return a host offset if we actually made progress. Otherwise we
967      * would make requirements for handle_alloc() that it can't fulfill */
968     if (ret) {
969         *host_offset = (cluster_offset & L2E_OFFSET_MASK)
970                      + offset_into_cluster(s, guest_offset);
971     }
972 
973     return ret;
974 }
975 
976 /*
977  * Allocates new clusters for the given guest_offset.
978  *
979  * At most *nb_clusters are allocated, and on return *nb_clusters is updated to
980  * contain the number of clusters that have been allocated and are contiguous
981  * in the image file.
982  *
983  * If *host_offset is non-zero, it specifies the offset in the image file at
984  * which the new clusters must start. *nb_clusters can be 0 on return in this
985  * case if the cluster at host_offset is already in use. If *host_offset is
986  * zero, the clusters can be allocated anywhere in the image file.
987  *
988  * *host_offset is updated to contain the offset into the image file at which
989  * the first allocated cluster starts.
990  *
991  * Return 0 on success and -errno in error cases. -EAGAIN means that the
992  * function has been waiting for another request and the allocation must be
993  * restarted, but the whole request should not be failed.
994  */
995 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
996     uint64_t *host_offset, unsigned int *nb_clusters)
997 {
998     BDRVQcowState *s = bs->opaque;
999 
1000     trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
1001                                          *host_offset, *nb_clusters);
1002 
1003     /* Allocate new clusters */
1004     trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
1005     if (*host_offset == 0) {
1006         int64_t cluster_offset =
1007             qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size);
1008         if (cluster_offset < 0) {
1009             return cluster_offset;
1010         }
1011         *host_offset = cluster_offset;
1012         return 0;
1013     } else {
1014         int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
1015         if (ret < 0) {
1016             return ret;
1017         }
1018         *nb_clusters = ret;
1019         return 0;
1020     }
1021 }
1022 
1023 /*
1024  * Allocates new clusters for an area that either is yet unallocated or needs a
1025  * copy on write. If *host_offset is non-zero, clusters are only allocated if
1026  * the new allocation can match the specified host offset.
1027  *
1028  * Note that guest_offset may not be cluster aligned. In this case, the
1029  * returned *host_offset points to exact byte referenced by guest_offset and
1030  * therefore isn't cluster aligned as well.
1031  *
1032  * Returns:
1033  *   0:     if no clusters could be allocated. *bytes is set to 0,
1034  *          *host_offset is left unchanged.
1035  *
1036  *   1:     if new clusters were allocated. *bytes may be decreased if the
1037  *          new allocation doesn't cover all of the requested area.
1038  *          *host_offset is updated to contain the host offset of the first
1039  *          newly allocated cluster.
1040  *
1041  *  -errno: in error cases
1042  */
1043 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset,
1044     uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m)
1045 {
1046     BDRVQcowState *s = bs->opaque;
1047     int l2_index;
1048     uint64_t *l2_table;
1049     uint64_t entry;
1050     unsigned int nb_clusters;
1051     int ret;
1052 
1053     uint64_t alloc_cluster_offset;
1054 
1055     trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset,
1056                              *bytes);
1057     assert(*bytes > 0);
1058 
1059     /*
1060      * Calculate the number of clusters to look for. We stop at L2 table
1061      * boundaries to keep things simple.
1062      */
1063     nb_clusters =
1064         size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes);
1065 
1066     l2_index = offset_to_l2_index(s, guest_offset);
1067     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1068 
1069     /* Find L2 entry for the first involved cluster */
1070     ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index);
1071     if (ret < 0) {
1072         return ret;
1073     }
1074 
1075     entry = be64_to_cpu(l2_table[l2_index]);
1076 
1077     /* For the moment, overwrite compressed clusters one by one */
1078     if (entry & QCOW_OFLAG_COMPRESSED) {
1079         nb_clusters = 1;
1080     } else {
1081         nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index);
1082     }
1083 
1084     /* This function is only called when there were no non-COW clusters, so if
1085      * we can't find any unallocated or COW clusters either, something is
1086      * wrong with our code. */
1087     assert(nb_clusters > 0);
1088 
1089     ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1090     if (ret < 0) {
1091         return ret;
1092     }
1093 
1094     /* Allocate, if necessary at a given offset in the image file */
1095     alloc_cluster_offset = start_of_cluster(s, *host_offset);
1096     ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset,
1097                                   &nb_clusters);
1098     if (ret < 0) {
1099         goto fail;
1100     }
1101 
1102     /* Can't extend contiguous allocation */
1103     if (nb_clusters == 0) {
1104         *bytes = 0;
1105         return 0;
1106     }
1107 
1108     /*
1109      * Save info needed for meta data update.
1110      *
1111      * requested_sectors: Number of sectors from the start of the first
1112      * newly allocated cluster to the end of the (possibly shortened
1113      * before) write request.
1114      *
1115      * avail_sectors: Number of sectors from the start of the first
1116      * newly allocated to the end of the last newly allocated cluster.
1117      *
1118      * nb_sectors: The number of sectors from the start of the first
1119      * newly allocated cluster to the end of the area that the write
1120      * request actually writes to (excluding COW at the end)
1121      */
1122     int requested_sectors =
1123         (*bytes + offset_into_cluster(s, guest_offset))
1124         >> BDRV_SECTOR_BITS;
1125     int avail_sectors = nb_clusters
1126                         << (s->cluster_bits - BDRV_SECTOR_BITS);
1127     int alloc_n_start = offset_into_cluster(s, guest_offset)
1128                         >> BDRV_SECTOR_BITS;
1129     int nb_sectors = MIN(requested_sectors, avail_sectors);
1130     QCowL2Meta *old_m = *m;
1131 
1132     *m = g_malloc0(sizeof(**m));
1133 
1134     **m = (QCowL2Meta) {
1135         .next           = old_m,
1136 
1137         .alloc_offset   = alloc_cluster_offset,
1138         .offset         = start_of_cluster(s, guest_offset),
1139         .nb_clusters    = nb_clusters,
1140         .nb_available   = nb_sectors,
1141 
1142         .cow_start = {
1143             .offset     = 0,
1144             .nb_sectors = alloc_n_start,
1145         },
1146         .cow_end = {
1147             .offset     = nb_sectors * BDRV_SECTOR_SIZE,
1148             .nb_sectors = avail_sectors - nb_sectors,
1149         },
1150     };
1151     qemu_co_queue_init(&(*m)->dependent_requests);
1152     QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight);
1153 
1154     *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset);
1155     *bytes = MIN(*bytes, (nb_sectors * BDRV_SECTOR_SIZE)
1156                          - offset_into_cluster(s, guest_offset));
1157     assert(*bytes != 0);
1158 
1159     return 1;
1160 
1161 fail:
1162     if (*m && (*m)->nb_clusters > 0) {
1163         QLIST_REMOVE(*m, next_in_flight);
1164     }
1165     return ret;
1166 }
1167 
1168 /*
1169  * alloc_cluster_offset
1170  *
1171  * For a given offset on the virtual disk, find the cluster offset in qcow2
1172  * file. If the offset is not found, allocate a new cluster.
1173  *
1174  * If the cluster was already allocated, m->nb_clusters is set to 0 and
1175  * other fields in m are meaningless.
1176  *
1177  * If the cluster is newly allocated, m->nb_clusters is set to the number of
1178  * contiguous clusters that have been allocated. In this case, the other
1179  * fields of m are valid and contain information about the first allocated
1180  * cluster.
1181  *
1182  * If the request conflicts with another write request in flight, the coroutine
1183  * is queued and will be reentered when the dependency has completed.
1184  *
1185  * Return 0 on success and -errno in error cases
1186  */
1187 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
1188     int *num, uint64_t *host_offset, QCowL2Meta **m)
1189 {
1190     BDRVQcowState *s = bs->opaque;
1191     uint64_t start, remaining;
1192     uint64_t cluster_offset;
1193     uint64_t cur_bytes;
1194     int ret;
1195 
1196     trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *num);
1197 
1198     assert((offset & ~BDRV_SECTOR_MASK) == 0);
1199 
1200 again:
1201     start = offset;
1202     remaining = *num << BDRV_SECTOR_BITS;
1203     cluster_offset = 0;
1204     *host_offset = 0;
1205     cur_bytes = 0;
1206     *m = NULL;
1207 
1208     while (true) {
1209 
1210         if (!*host_offset) {
1211             *host_offset = start_of_cluster(s, cluster_offset);
1212         }
1213 
1214         assert(remaining >= cur_bytes);
1215 
1216         start           += cur_bytes;
1217         remaining       -= cur_bytes;
1218         cluster_offset  += cur_bytes;
1219 
1220         if (remaining == 0) {
1221             break;
1222         }
1223 
1224         cur_bytes = remaining;
1225 
1226         /*
1227          * Now start gathering as many contiguous clusters as possible:
1228          *
1229          * 1. Check for overlaps with in-flight allocations
1230          *
1231          *      a) Overlap not in the first cluster -> shorten this request and
1232          *         let the caller handle the rest in its next loop iteration.
1233          *
1234          *      b) Real overlaps of two requests. Yield and restart the search
1235          *         for contiguous clusters (the situation could have changed
1236          *         while we were sleeping)
1237          *
1238          *      c) TODO: Request starts in the same cluster as the in-flight
1239          *         allocation ends. Shorten the COW of the in-fight allocation,
1240          *         set cluster_offset to write to the same cluster and set up
1241          *         the right synchronisation between the in-flight request and
1242          *         the new one.
1243          */
1244         ret = handle_dependencies(bs, start, &cur_bytes, m);
1245         if (ret == -EAGAIN) {
1246             /* Currently handle_dependencies() doesn't yield if we already had
1247              * an allocation. If it did, we would have to clean up the L2Meta
1248              * structs before starting over. */
1249             assert(*m == NULL);
1250             goto again;
1251         } else if (ret < 0) {
1252             return ret;
1253         } else if (cur_bytes == 0) {
1254             break;
1255         } else {
1256             /* handle_dependencies() may have decreased cur_bytes (shortened
1257              * the allocations below) so that the next dependency is processed
1258              * correctly during the next loop iteration. */
1259         }
1260 
1261         /*
1262          * 2. Count contiguous COPIED clusters.
1263          */
1264         ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m);
1265         if (ret < 0) {
1266             return ret;
1267         } else if (ret) {
1268             continue;
1269         } else if (cur_bytes == 0) {
1270             break;
1271         }
1272 
1273         /*
1274          * 3. If the request still hasn't completed, allocate new clusters,
1275          *    considering any cluster_offset of steps 1c or 2.
1276          */
1277         ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m);
1278         if (ret < 0) {
1279             return ret;
1280         } else if (ret) {
1281             continue;
1282         } else {
1283             assert(cur_bytes == 0);
1284             break;
1285         }
1286     }
1287 
1288     *num -= remaining >> BDRV_SECTOR_BITS;
1289     assert(*num > 0);
1290     assert(*host_offset != 0);
1291 
1292     return 0;
1293 }
1294 
1295 static int decompress_buffer(uint8_t *out_buf, int out_buf_size,
1296                              const uint8_t *buf, int buf_size)
1297 {
1298     z_stream strm1, *strm = &strm1;
1299     int ret, out_len;
1300 
1301     memset(strm, 0, sizeof(*strm));
1302 
1303     strm->next_in = (uint8_t *)buf;
1304     strm->avail_in = buf_size;
1305     strm->next_out = out_buf;
1306     strm->avail_out = out_buf_size;
1307 
1308     ret = inflateInit2(strm, -12);
1309     if (ret != Z_OK)
1310         return -1;
1311     ret = inflate(strm, Z_FINISH);
1312     out_len = strm->next_out - out_buf;
1313     if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) ||
1314         out_len != out_buf_size) {
1315         inflateEnd(strm);
1316         return -1;
1317     }
1318     inflateEnd(strm);
1319     return 0;
1320 }
1321 
1322 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset)
1323 {
1324     BDRVQcowState *s = bs->opaque;
1325     int ret, csize, nb_csectors, sector_offset;
1326     uint64_t coffset;
1327 
1328     coffset = cluster_offset & s->cluster_offset_mask;
1329     if (s->cluster_cache_offset != coffset) {
1330         nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1;
1331         sector_offset = coffset & 511;
1332         csize = nb_csectors * 512 - sector_offset;
1333         BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED);
1334         ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, nb_csectors);
1335         if (ret < 0) {
1336             return ret;
1337         }
1338         if (decompress_buffer(s->cluster_cache, s->cluster_size,
1339                               s->cluster_data + sector_offset, csize) < 0) {
1340             return -EIO;
1341         }
1342         s->cluster_cache_offset = coffset;
1343     }
1344     return 0;
1345 }
1346 
1347 /*
1348  * This discards as many clusters of nb_clusters as possible at once (i.e.
1349  * all clusters in the same L2 table) and returns the number of discarded
1350  * clusters.
1351  */
1352 static int discard_single_l2(BlockDriverState *bs, uint64_t offset,
1353     unsigned int nb_clusters, enum qcow2_discard_type type)
1354 {
1355     BDRVQcowState *s = bs->opaque;
1356     uint64_t *l2_table;
1357     int l2_index;
1358     int ret;
1359     int i;
1360 
1361     ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1362     if (ret < 0) {
1363         return ret;
1364     }
1365 
1366     /* Limit nb_clusters to one L2 table */
1367     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1368 
1369     for (i = 0; i < nb_clusters; i++) {
1370         uint64_t old_l2_entry;
1371 
1372         old_l2_entry = be64_to_cpu(l2_table[l2_index + i]);
1373 
1374         /*
1375          * Make sure that a discarded area reads back as zeroes for v3 images
1376          * (we cannot do it for v2 without actually writing a zero-filled
1377          * buffer). We can skip the operation if the cluster is already marked
1378          * as zero, or if it's unallocated and we don't have a backing file.
1379          *
1380          * TODO We might want to use bdrv_get_block_status(bs) here, but we're
1381          * holding s->lock, so that doesn't work today.
1382          */
1383         switch (qcow2_get_cluster_type(old_l2_entry)) {
1384             case QCOW2_CLUSTER_UNALLOCATED:
1385                 if (!bs->backing_hd) {
1386                     continue;
1387                 }
1388                 break;
1389 
1390             case QCOW2_CLUSTER_ZERO:
1391                 continue;
1392 
1393             case QCOW2_CLUSTER_NORMAL:
1394             case QCOW2_CLUSTER_COMPRESSED:
1395                 break;
1396 
1397             default:
1398                 abort();
1399         }
1400 
1401         /* First remove L2 entries */
1402         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1403         if (s->qcow_version >= 3) {
1404             l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1405         } else {
1406             l2_table[l2_index + i] = cpu_to_be64(0);
1407         }
1408 
1409         /* Then decrease the refcount */
1410         qcow2_free_any_clusters(bs, old_l2_entry, 1, type);
1411     }
1412 
1413     ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1414     if (ret < 0) {
1415         return ret;
1416     }
1417 
1418     return nb_clusters;
1419 }
1420 
1421 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
1422     int nb_sectors, enum qcow2_discard_type type)
1423 {
1424     BDRVQcowState *s = bs->opaque;
1425     uint64_t end_offset;
1426     unsigned int nb_clusters;
1427     int ret;
1428 
1429     end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
1430 
1431     /* Round start up and end down */
1432     offset = align_offset(offset, s->cluster_size);
1433     end_offset = start_of_cluster(s, end_offset);
1434 
1435     if (offset > end_offset) {
1436         return 0;
1437     }
1438 
1439     nb_clusters = size_to_clusters(s, end_offset - offset);
1440 
1441     s->cache_discards = true;
1442 
1443     /* Each L2 table is handled by its own loop iteration */
1444     while (nb_clusters > 0) {
1445         ret = discard_single_l2(bs, offset, nb_clusters, type);
1446         if (ret < 0) {
1447             goto fail;
1448         }
1449 
1450         nb_clusters -= ret;
1451         offset += (ret * s->cluster_size);
1452     }
1453 
1454     ret = 0;
1455 fail:
1456     s->cache_discards = false;
1457     qcow2_process_discards(bs, ret);
1458 
1459     return ret;
1460 }
1461 
1462 /*
1463  * This zeroes as many clusters of nb_clusters as possible at once (i.e.
1464  * all clusters in the same L2 table) and returns the number of zeroed
1465  * clusters.
1466  */
1467 static int zero_single_l2(BlockDriverState *bs, uint64_t offset,
1468     unsigned int nb_clusters)
1469 {
1470     BDRVQcowState *s = bs->opaque;
1471     uint64_t *l2_table;
1472     int l2_index;
1473     int ret;
1474     int i;
1475 
1476     ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
1477     if (ret < 0) {
1478         return ret;
1479     }
1480 
1481     /* Limit nb_clusters to one L2 table */
1482     nb_clusters = MIN(nb_clusters, s->l2_size - l2_index);
1483 
1484     for (i = 0; i < nb_clusters; i++) {
1485         uint64_t old_offset;
1486 
1487         old_offset = be64_to_cpu(l2_table[l2_index + i]);
1488 
1489         /* Update L2 entries */
1490         qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1491         if (old_offset & QCOW_OFLAG_COMPRESSED) {
1492             l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
1493             qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST);
1494         } else {
1495             l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO);
1496         }
1497     }
1498 
1499     ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
1500     if (ret < 0) {
1501         return ret;
1502     }
1503 
1504     return nb_clusters;
1505 }
1506 
1507 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors)
1508 {
1509     BDRVQcowState *s = bs->opaque;
1510     unsigned int nb_clusters;
1511     int ret;
1512 
1513     /* The zero flag is only supported by version 3 and newer */
1514     if (s->qcow_version < 3) {
1515         return -ENOTSUP;
1516     }
1517 
1518     /* Each L2 table is handled by its own loop iteration */
1519     nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS);
1520 
1521     s->cache_discards = true;
1522 
1523     while (nb_clusters > 0) {
1524         ret = zero_single_l2(bs, offset, nb_clusters);
1525         if (ret < 0) {
1526             goto fail;
1527         }
1528 
1529         nb_clusters -= ret;
1530         offset += (ret * s->cluster_size);
1531     }
1532 
1533     ret = 0;
1534 fail:
1535     s->cache_discards = false;
1536     qcow2_process_discards(bs, ret);
1537 
1538     return ret;
1539 }
1540 
1541 /*
1542  * Expands all zero clusters in a specific L1 table (or deallocates them, for
1543  * non-backed non-pre-allocated zero clusters).
1544  *
1545  * expanded_clusters is a bitmap where every bit corresponds to one cluster in
1546  * the image file; a bit gets set if the corresponding cluster has been used for
1547  * zero expansion (i.e., has been filled with zeroes and is referenced from an
1548  * L2 table). nb_clusters contains the total cluster count of the image file,
1549  * i.e., the number of bits in expanded_clusters.
1550  */
1551 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table,
1552                                       int l1_size, uint8_t **expanded_clusters,
1553                                       uint64_t *nb_clusters)
1554 {
1555     BDRVQcowState *s = bs->opaque;
1556     bool is_active_l1 = (l1_table == s->l1_table);
1557     uint64_t *l2_table = NULL;
1558     int ret;
1559     int i, j;
1560 
1561     if (!is_active_l1) {
1562         /* inactive L2 tables require a buffer to be stored in when loading
1563          * them from disk */
1564         l2_table = qemu_blockalign(bs, s->cluster_size);
1565     }
1566 
1567     for (i = 0; i < l1_size; i++) {
1568         uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK;
1569         bool l2_dirty = false;
1570 
1571         if (!l2_offset) {
1572             /* unallocated */
1573             continue;
1574         }
1575 
1576         if (is_active_l1) {
1577             /* get active L2 tables from cache */
1578             ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset,
1579                     (void **)&l2_table);
1580         } else {
1581             /* load inactive L2 tables from disk */
1582             ret = bdrv_read(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1583                     (void *)l2_table, s->cluster_sectors);
1584         }
1585         if (ret < 0) {
1586             goto fail;
1587         }
1588 
1589         for (j = 0; j < s->l2_size; j++) {
1590             uint64_t l2_entry = be64_to_cpu(l2_table[j]);
1591             int64_t offset = l2_entry & L2E_OFFSET_MASK, cluster_index;
1592             int cluster_type = qcow2_get_cluster_type(l2_entry);
1593             bool preallocated = offset != 0;
1594 
1595             if (cluster_type == QCOW2_CLUSTER_NORMAL) {
1596                 cluster_index = offset >> s->cluster_bits;
1597                 assert((cluster_index >= 0) && (cluster_index < *nb_clusters));
1598                 if ((*expanded_clusters)[cluster_index / 8] &
1599                     (1 << (cluster_index % 8))) {
1600                     /* Probably a shared L2 table; this cluster was a zero
1601                      * cluster which has been expanded, its refcount
1602                      * therefore most likely requires an update. */
1603                     ret = qcow2_update_cluster_refcount(bs, cluster_index, 1,
1604                                                         QCOW2_DISCARD_NEVER);
1605                     if (ret < 0) {
1606                         goto fail;
1607                     }
1608                     /* Since we just increased the refcount, the COPIED flag may
1609                      * no longer be set. */
1610                     l2_table[j] = cpu_to_be64(l2_entry & ~QCOW_OFLAG_COPIED);
1611                     l2_dirty = true;
1612                 }
1613                 continue;
1614             }
1615             else if (qcow2_get_cluster_type(l2_entry) != QCOW2_CLUSTER_ZERO) {
1616                 continue;
1617             }
1618 
1619             if (!preallocated) {
1620                 if (!bs->backing_hd) {
1621                     /* not backed; therefore we can simply deallocate the
1622                      * cluster */
1623                     l2_table[j] = 0;
1624                     l2_dirty = true;
1625                     continue;
1626                 }
1627 
1628                 offset = qcow2_alloc_clusters(bs, s->cluster_size);
1629                 if (offset < 0) {
1630                     ret = offset;
1631                     goto fail;
1632                 }
1633             }
1634 
1635             ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size);
1636             if (ret < 0) {
1637                 if (!preallocated) {
1638                     qcow2_free_clusters(bs, offset, s->cluster_size,
1639                                         QCOW2_DISCARD_ALWAYS);
1640                 }
1641                 goto fail;
1642             }
1643 
1644             ret = bdrv_write_zeroes(bs->file, offset / BDRV_SECTOR_SIZE,
1645                                     s->cluster_sectors, 0);
1646             if (ret < 0) {
1647                 if (!preallocated) {
1648                     qcow2_free_clusters(bs, offset, s->cluster_size,
1649                                         QCOW2_DISCARD_ALWAYS);
1650                 }
1651                 goto fail;
1652             }
1653 
1654             l2_table[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED);
1655             l2_dirty = true;
1656 
1657             cluster_index = offset >> s->cluster_bits;
1658 
1659             if (cluster_index >= *nb_clusters) {
1660                 uint64_t old_bitmap_size = (*nb_clusters + 7) / 8;
1661                 uint64_t new_bitmap_size;
1662                 /* The offset may lie beyond the old end of the underlying image
1663                  * file for growable files only */
1664                 assert(bs->file->growable);
1665                 *nb_clusters = size_to_clusters(s, bs->file->total_sectors *
1666                                                 BDRV_SECTOR_SIZE);
1667                 new_bitmap_size = (*nb_clusters + 7) / 8;
1668                 *expanded_clusters = g_realloc(*expanded_clusters,
1669                                                new_bitmap_size);
1670                 /* clear the newly allocated space */
1671                 memset(&(*expanded_clusters)[old_bitmap_size], 0,
1672                        new_bitmap_size - old_bitmap_size);
1673             }
1674 
1675             assert((cluster_index >= 0) && (cluster_index < *nb_clusters));
1676             (*expanded_clusters)[cluster_index / 8] |= 1 << (cluster_index % 8);
1677         }
1678 
1679         if (is_active_l1) {
1680             if (l2_dirty) {
1681                 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table);
1682                 qcow2_cache_depends_on_flush(s->l2_table_cache);
1683             }
1684             ret = qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
1685             if (ret < 0) {
1686                 l2_table = NULL;
1687                 goto fail;
1688             }
1689         } else {
1690             if (l2_dirty) {
1691                 ret = qcow2_pre_write_overlap_check(bs,
1692                         QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2, l2_offset,
1693                         s->cluster_size);
1694                 if (ret < 0) {
1695                     goto fail;
1696                 }
1697 
1698                 ret = bdrv_write(bs->file, l2_offset / BDRV_SECTOR_SIZE,
1699                         (void *)l2_table, s->cluster_sectors);
1700                 if (ret < 0) {
1701                     goto fail;
1702                 }
1703             }
1704         }
1705     }
1706 
1707     ret = 0;
1708 
1709 fail:
1710     if (l2_table) {
1711         if (!is_active_l1) {
1712             qemu_vfree(l2_table);
1713         } else {
1714             if (ret < 0) {
1715                 qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table);
1716             } else {
1717                 ret = qcow2_cache_put(bs, s->l2_table_cache,
1718                         (void **)&l2_table);
1719             }
1720         }
1721     }
1722     return ret;
1723 }
1724 
1725 /*
1726  * For backed images, expands all zero clusters on the image. For non-backed
1727  * images, deallocates all non-pre-allocated zero clusters (and claims the
1728  * allocation for pre-allocated ones). This is important for downgrading to a
1729  * qcow2 version which doesn't yet support metadata zero clusters.
1730  */
1731 int qcow2_expand_zero_clusters(BlockDriverState *bs)
1732 {
1733     BDRVQcowState *s = bs->opaque;
1734     uint64_t *l1_table = NULL;
1735     uint64_t nb_clusters;
1736     uint8_t *expanded_clusters;
1737     int ret;
1738     int i, j;
1739 
1740     nb_clusters = size_to_clusters(s, bs->file->total_sectors *
1741                                    BDRV_SECTOR_SIZE);
1742     expanded_clusters = g_malloc0((nb_clusters + 7) / 8);
1743 
1744     ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size,
1745                                      &expanded_clusters, &nb_clusters);
1746     if (ret < 0) {
1747         goto fail;
1748     }
1749 
1750     /* Inactive L1 tables may point to active L2 tables - therefore it is
1751      * necessary to flush the L2 table cache before trying to access the L2
1752      * tables pointed to by inactive L1 entries (else we might try to expand
1753      * zero clusters that have already been expanded); furthermore, it is also
1754      * necessary to empty the L2 table cache, since it may contain tables which
1755      * are now going to be modified directly on disk, bypassing the cache.
1756      * qcow2_cache_empty() does both for us. */
1757     ret = qcow2_cache_empty(bs, s->l2_table_cache);
1758     if (ret < 0) {
1759         goto fail;
1760     }
1761 
1762     for (i = 0; i < s->nb_snapshots; i++) {
1763         int l1_sectors = (s->snapshots[i].l1_size * sizeof(uint64_t) +
1764                 BDRV_SECTOR_SIZE - 1) / BDRV_SECTOR_SIZE;
1765 
1766         l1_table = g_realloc(l1_table, l1_sectors * BDRV_SECTOR_SIZE);
1767 
1768         ret = bdrv_read(bs->file, s->snapshots[i].l1_table_offset /
1769                 BDRV_SECTOR_SIZE, (void *)l1_table, l1_sectors);
1770         if (ret < 0) {
1771             goto fail;
1772         }
1773 
1774         for (j = 0; j < s->snapshots[i].l1_size; j++) {
1775             be64_to_cpus(&l1_table[j]);
1776         }
1777 
1778         ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size,
1779                                          &expanded_clusters, &nb_clusters);
1780         if (ret < 0) {
1781             goto fail;
1782         }
1783     }
1784 
1785     ret = 0;
1786 
1787 fail:
1788     g_free(expanded_clusters);
1789     g_free(l1_table);
1790     return ret;
1791 }
1792