6f758870fc19cf0e66db3055dab36ab96e038fee
[linux-2.6.git] / drivers / md / dm-snap.c
1 /*
2  * dm-snapshot.c
3  *
4  * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
5  *
6  * This file is released under the GPL.
7  */
8
9 #include <linux/blkdev.h>
10 #include <linux/device-mapper.h>
11 #include <linux/delay.h>
12 #include <linux/fs.h>
13 #include <linux/init.h>
14 #include <linux/kdev_t.h>
15 #include <linux/list.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/vmalloc.h>
20 #include <linux/log2.h>
21 #include <linux/dm-kcopyd.h>
22
23 #include "dm-exception-store.h"
24
25 #define DM_MSG_PREFIX "snapshots"
26
27 static const char dm_snapshot_merge_target_name[] = "snapshot-merge";
28
29 #define dm_target_is_snapshot_merge(ti) \
30         ((ti)->type->name == dm_snapshot_merge_target_name)
31
32 /*
33  * The size of the mempool used to track chunks in use.
34  */
35 #define MIN_IOS 256
36
37 #define DM_TRACKED_CHUNK_HASH_SIZE      16
38 #define DM_TRACKED_CHUNK_HASH(x)        ((unsigned long)(x) & \
39                                          (DM_TRACKED_CHUNK_HASH_SIZE - 1))
40
41 struct dm_exception_table {
42         uint32_t hash_mask;
43         unsigned hash_shift;
44         struct list_head *table;
45 };
46
47 struct dm_snapshot {
48         struct rw_semaphore lock;
49
50         struct dm_dev *origin;
51         struct dm_dev *cow;
52
53         struct dm_target *ti;
54
55         /* List of snapshots per Origin */
56         struct list_head list;
57
58         /*
59          * You can't use a snapshot if this is 0 (e.g. if full).
60          * A snapshot-merge target never clears this.
61          */
62         int valid;
63
64         /* Origin writes don't trigger exceptions until this is set */
65         int active;
66
67         atomic_t pending_exceptions_count;
68
69         mempool_t *pending_pool;
70
71         struct dm_exception_table pending;
72         struct dm_exception_table complete;
73
74         /*
75          * pe_lock protects all pending_exception operations and access
76          * as well as the snapshot_bios list.
77          */
78         spinlock_t pe_lock;
79
80         /* Chunks with outstanding reads */
81         spinlock_t tracked_chunk_lock;
82         mempool_t *tracked_chunk_pool;
83         struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE];
84
85         /* The on disk metadata handler */
86         struct dm_exception_store *store;
87
88         struct dm_kcopyd_client *kcopyd_client;
89
90         /* Wait for events based on state_bits */
91         unsigned long state_bits;
92
93         /* Range of chunks currently being merged. */
94         chunk_t first_merging_chunk;
95         int num_merging_chunks;
96
97         /*
98          * The merge operation failed if this flag is set.
99          * Failure modes are handled as follows:
100          * - I/O error reading the header
101          *      => don't load the target; abort.
102          * - Header does not have "valid" flag set
103          *      => use the origin; forget about the snapshot.
104          * - I/O error when reading exceptions
105          *      => don't load the target; abort.
106          *         (We can't use the intermediate origin state.)
107          * - I/O error while merging
108          *      => stop merging; set merge_failed; process I/O normally.
109          */
110         int merge_failed;
111
112         /*
113          * Incoming bios that overlap with chunks being merged must wait
114          * for them to be committed.
115          */
116         struct bio_list bios_queued_during_merge;
117 };
118
119 /*
120  * state_bits:
121  *   RUNNING_MERGE  - Merge operation is in progress.
122  *   SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
123  *                    cleared afterwards.
124  */
125 #define RUNNING_MERGE          0
126 #define SHUTDOWN_MERGE         1
127
128 struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
129 {
130         return s->origin;
131 }
132 EXPORT_SYMBOL(dm_snap_origin);
133
134 struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
135 {
136         return s->cow;
137 }
138 EXPORT_SYMBOL(dm_snap_cow);
139
140 static sector_t chunk_to_sector(struct dm_exception_store *store,
141                                 chunk_t chunk)
142 {
143         return chunk << store->chunk_shift;
144 }
145
146 static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
147 {
148         /*
149          * There is only ever one instance of a particular block
150          * device so we can compare pointers safely.
151          */
152         return lhs == rhs;
153 }
154
155 struct dm_snap_pending_exception {
156         struct dm_exception e;
157
158         /*
159          * Origin buffers waiting for this to complete are held
160          * in a bio list
161          */
162         struct bio_list origin_bios;
163         struct bio_list snapshot_bios;
164
165         /* Pointer back to snapshot context */
166         struct dm_snapshot *snap;
167
168         /*
169          * 1 indicates the exception has already been sent to
170          * kcopyd.
171          */
172         int started;
173
174         /*
175          * For writing a complete chunk, bypassing the copy.
176          */
177         struct bio *full_bio;
178         bio_end_io_t *full_bio_end_io;
179         void *full_bio_private;
180 };
181
182 /*
183  * Hash table mapping origin volumes to lists of snapshots and
184  * a lock to protect it
185  */
186 static struct kmem_cache *exception_cache;
187 static struct kmem_cache *pending_cache;
188
189 struct dm_snap_tracked_chunk {
190         struct hlist_node node;
191         chunk_t chunk;
192 };
193
194 static struct kmem_cache *tracked_chunk_cache;
195
196 static struct dm_snap_tracked_chunk *track_chunk(struct dm_snapshot *s,
197                                                  chunk_t chunk)
198 {
199         struct dm_snap_tracked_chunk *c = mempool_alloc(s->tracked_chunk_pool,
200                                                         GFP_NOIO);
201         unsigned long flags;
202
203         c->chunk = chunk;
204
205         spin_lock_irqsave(&s->tracked_chunk_lock, flags);
206         hlist_add_head(&c->node,
207                        &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
208         spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
209
210         return c;
211 }
212
213 static void stop_tracking_chunk(struct dm_snapshot *s,
214                                 struct dm_snap_tracked_chunk *c)
215 {
216         unsigned long flags;
217
218         spin_lock_irqsave(&s->tracked_chunk_lock, flags);
219         hlist_del(&c->node);
220         spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
221
222         mempool_free(c, s->tracked_chunk_pool);
223 }
224
225 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
226 {
227         struct dm_snap_tracked_chunk *c;
228         struct hlist_node *hn;
229         int found = 0;
230
231         spin_lock_irq(&s->tracked_chunk_lock);
232
233         hlist_for_each_entry(c, hn,
234             &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
235                 if (c->chunk == chunk) {
236                         found = 1;
237                         break;
238                 }
239         }
240
241         spin_unlock_irq(&s->tracked_chunk_lock);
242
243         return found;
244 }
245
246 /*
247  * This conflicting I/O is extremely improbable in the caller,
248  * so msleep(1) is sufficient and there is no need for a wait queue.
249  */
250 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
251 {
252         while (__chunk_is_tracked(s, chunk))
253                 msleep(1);
254 }
255
256 /*
257  * One of these per registered origin, held in the snapshot_origins hash
258  */
259 struct origin {
260         /* The origin device */
261         struct block_device *bdev;
262
263         struct list_head hash_list;
264
265         /* List of snapshots for this origin */
266         struct list_head snapshots;
267 };
268
269 /*
270  * Size of the hash table for origin volumes. If we make this
271  * the size of the minors list then it should be nearly perfect
272  */
273 #define ORIGIN_HASH_SIZE 256
274 #define ORIGIN_MASK      0xFF
275 static struct list_head *_origins;
276 static struct rw_semaphore _origins_lock;
277
278 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
279 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
280 static uint64_t _pending_exceptions_done_count;
281
282 static int init_origin_hash(void)
283 {
284         int i;
285
286         _origins = kmalloc(ORIGIN_HASH_SIZE * sizeof(struct list_head),
287                            GFP_KERNEL);
288         if (!_origins) {
289                 DMERR("unable to allocate memory");
290                 return -ENOMEM;
291         }
292
293         for (i = 0; i < ORIGIN_HASH_SIZE; i++)
294                 INIT_LIST_HEAD(_origins + i);
295         init_rwsem(&_origins_lock);
296
297         return 0;
298 }
299
300 static void exit_origin_hash(void)
301 {
302         kfree(_origins);
303 }
304
305 static unsigned origin_hash(struct block_device *bdev)
306 {
307         return bdev->bd_dev & ORIGIN_MASK;
308 }
309
310 static struct origin *__lookup_origin(struct block_device *origin)
311 {
312         struct list_head *ol;
313         struct origin *o;
314
315         ol = &_origins[origin_hash(origin)];
316         list_for_each_entry (o, ol, hash_list)
317                 if (bdev_equal(o->bdev, origin))
318                         return o;
319
320         return NULL;
321 }
322
323 static void __insert_origin(struct origin *o)
324 {
325         struct list_head *sl = &_origins[origin_hash(o->bdev)];
326         list_add_tail(&o->hash_list, sl);
327 }
328
329 /*
330  * _origins_lock must be held when calling this function.
331  * Returns number of snapshots registered using the supplied cow device, plus:
332  * snap_src - a snapshot suitable for use as a source of exception handover
333  * snap_dest - a snapshot capable of receiving exception handover.
334  * snap_merge - an existing snapshot-merge target linked to the same origin.
335  *   There can be at most one snapshot-merge target. The parameter is optional.
336  *
337  * Possible return values and states of snap_src and snap_dest.
338  *   0: NULL, NULL  - first new snapshot
339  *   1: snap_src, NULL - normal snapshot
340  *   2: snap_src, snap_dest  - waiting for handover
341  *   2: snap_src, NULL - handed over, waiting for old to be deleted
342  *   1: NULL, snap_dest - source got destroyed without handover
343  */
344 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
345                                         struct dm_snapshot **snap_src,
346                                         struct dm_snapshot **snap_dest,
347                                         struct dm_snapshot **snap_merge)
348 {
349         struct dm_snapshot *s;
350         struct origin *o;
351         int count = 0;
352         int active;
353
354         o = __lookup_origin(snap->origin->bdev);
355         if (!o)
356                 goto out;
357
358         list_for_each_entry(s, &o->snapshots, list) {
359                 if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
360                         *snap_merge = s;
361                 if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
362                         continue;
363
364                 down_read(&s->lock);
365                 active = s->active;
366                 up_read(&s->lock);
367
368                 if (active) {
369                         if (snap_src)
370                                 *snap_src = s;
371                 } else if (snap_dest)
372                         *snap_dest = s;
373
374                 count++;
375         }
376
377 out:
378         return count;
379 }
380
381 /*
382  * On success, returns 1 if this snapshot is a handover destination,
383  * otherwise returns 0.
384  */
385 static int __validate_exception_handover(struct dm_snapshot *snap)
386 {
387         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
388         struct dm_snapshot *snap_merge = NULL;
389
390         /* Does snapshot need exceptions handed over to it? */
391         if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
392                                           &snap_merge) == 2) ||
393             snap_dest) {
394                 snap->ti->error = "Snapshot cow pairing for exception "
395                                   "table handover failed";
396                 return -EINVAL;
397         }
398
399         /*
400          * If no snap_src was found, snap cannot become a handover
401          * destination.
402          */
403         if (!snap_src)
404                 return 0;
405
406         /*
407          * Non-snapshot-merge handover?
408          */
409         if (!dm_target_is_snapshot_merge(snap->ti))
410                 return 1;
411
412         /*
413          * Do not allow more than one merging snapshot.
414          */
415         if (snap_merge) {
416                 snap->ti->error = "A snapshot is already merging.";
417                 return -EINVAL;
418         }
419
420         if (!snap_src->store->type->prepare_merge ||
421             !snap_src->store->type->commit_merge) {
422                 snap->ti->error = "Snapshot exception store does not "
423                                   "support snapshot-merge.";
424                 return -EINVAL;
425         }
426
427         return 1;
428 }
429
430 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
431 {
432         struct dm_snapshot *l;
433
434         /* Sort the list according to chunk size, largest-first smallest-last */
435         list_for_each_entry(l, &o->snapshots, list)
436                 if (l->store->chunk_size < s->store->chunk_size)
437                         break;
438         list_add_tail(&s->list, &l->list);
439 }
440
441 /*
442  * Make a note of the snapshot and its origin so we can look it
443  * up when the origin has a write on it.
444  *
445  * Also validate snapshot exception store handovers.
446  * On success, returns 1 if this registration is a handover destination,
447  * otherwise returns 0.
448  */
449 static int register_snapshot(struct dm_snapshot *snap)
450 {
451         struct origin *o, *new_o = NULL;
452         struct block_device *bdev = snap->origin->bdev;
453         int r = 0;
454
455         new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
456         if (!new_o)
457                 return -ENOMEM;
458
459         down_write(&_origins_lock);
460
461         r = __validate_exception_handover(snap);
462         if (r < 0) {
463                 kfree(new_o);
464                 goto out;
465         }
466
467         o = __lookup_origin(bdev);
468         if (o)
469                 kfree(new_o);
470         else {
471                 /* New origin */
472                 o = new_o;
473
474                 /* Initialise the struct */
475                 INIT_LIST_HEAD(&o->snapshots);
476                 o->bdev = bdev;
477
478                 __insert_origin(o);
479         }
480
481         __insert_snapshot(o, snap);
482
483 out:
484         up_write(&_origins_lock);
485
486         return r;
487 }
488
489 /*
490  * Move snapshot to correct place in list according to chunk size.
491  */
492 static void reregister_snapshot(struct dm_snapshot *s)
493 {
494         struct block_device *bdev = s->origin->bdev;
495
496         down_write(&_origins_lock);
497
498         list_del(&s->list);
499         __insert_snapshot(__lookup_origin(bdev), s);
500
501         up_write(&_origins_lock);
502 }
503
504 static void unregister_snapshot(struct dm_snapshot *s)
505 {
506         struct origin *o;
507
508         down_write(&_origins_lock);
509         o = __lookup_origin(s->origin->bdev);
510
511         list_del(&s->list);
512         if (o && list_empty(&o->snapshots)) {
513                 list_del(&o->hash_list);
514                 kfree(o);
515         }
516
517         up_write(&_origins_lock);
518 }
519
520 /*
521  * Implementation of the exception hash tables.
522  * The lowest hash_shift bits of the chunk number are ignored, allowing
523  * some consecutive chunks to be grouped together.
524  */
525 static int dm_exception_table_init(struct dm_exception_table *et,
526                                    uint32_t size, unsigned hash_shift)
527 {
528         unsigned int i;
529
530         et->hash_shift = hash_shift;
531         et->hash_mask = size - 1;
532         et->table = dm_vcalloc(size, sizeof(struct list_head));
533         if (!et->table)
534                 return -ENOMEM;
535
536         for (i = 0; i < size; i++)
537                 INIT_LIST_HEAD(et->table + i);
538
539         return 0;
540 }
541
542 static void dm_exception_table_exit(struct dm_exception_table *et,
543                                     struct kmem_cache *mem)
544 {
545         struct list_head *slot;
546         struct dm_exception *ex, *next;
547         int i, size;
548
549         size = et->hash_mask + 1;
550         for (i = 0; i < size; i++) {
551                 slot = et->table + i;
552
553                 list_for_each_entry_safe (ex, next, slot, hash_list)
554                         kmem_cache_free(mem, ex);
555         }
556
557         vfree(et->table);
558 }
559
560 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
561 {
562         return (chunk >> et->hash_shift) & et->hash_mask;
563 }
564
565 static void dm_remove_exception(struct dm_exception *e)
566 {
567         list_del(&e->hash_list);
568 }
569
570 /*
571  * Return the exception data for a sector, or NULL if not
572  * remapped.
573  */
574 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
575                                                 chunk_t chunk)
576 {
577         struct list_head *slot;
578         struct dm_exception *e;
579
580         slot = &et->table[exception_hash(et, chunk)];
581         list_for_each_entry (e, slot, hash_list)
582                 if (chunk >= e->old_chunk &&
583                     chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
584                         return e;
585
586         return NULL;
587 }
588
589 static struct dm_exception *alloc_completed_exception(void)
590 {
591         struct dm_exception *e;
592
593         e = kmem_cache_alloc(exception_cache, GFP_NOIO);
594         if (!e)
595                 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
596
597         return e;
598 }
599
600 static void free_completed_exception(struct dm_exception *e)
601 {
602         kmem_cache_free(exception_cache, e);
603 }
604
605 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
606 {
607         struct dm_snap_pending_exception *pe = mempool_alloc(s->pending_pool,
608                                                              GFP_NOIO);
609
610         atomic_inc(&s->pending_exceptions_count);
611         pe->snap = s;
612
613         return pe;
614 }
615
616 static void free_pending_exception(struct dm_snap_pending_exception *pe)
617 {
618         struct dm_snapshot *s = pe->snap;
619
620         mempool_free(pe, s->pending_pool);
621         smp_mb__before_atomic_dec();
622         atomic_dec(&s->pending_exceptions_count);
623 }
624
625 static void dm_insert_exception(struct dm_exception_table *eh,
626                                 struct dm_exception *new_e)
627 {
628         struct list_head *l;
629         struct dm_exception *e = NULL;
630
631         l = &eh->table[exception_hash(eh, new_e->old_chunk)];
632
633         /* Add immediately if this table doesn't support consecutive chunks */
634         if (!eh->hash_shift)
635                 goto out;
636
637         /* List is ordered by old_chunk */
638         list_for_each_entry_reverse(e, l, hash_list) {
639                 /* Insert after an existing chunk? */
640                 if (new_e->old_chunk == (e->old_chunk +
641                                          dm_consecutive_chunk_count(e) + 1) &&
642                     new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
643                                          dm_consecutive_chunk_count(e) + 1)) {
644                         dm_consecutive_chunk_count_inc(e);
645                         free_completed_exception(new_e);
646                         return;
647                 }
648
649                 /* Insert before an existing chunk? */
650                 if (new_e->old_chunk == (e->old_chunk - 1) &&
651                     new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
652                         dm_consecutive_chunk_count_inc(e);
653                         e->old_chunk--;
654                         e->new_chunk--;
655                         free_completed_exception(new_e);
656                         return;
657                 }
658
659                 if (new_e->old_chunk > e->old_chunk)
660                         break;
661         }
662
663 out:
664         list_add(&new_e->hash_list, e ? &e->hash_list : l);
665 }
666
667 /*
668  * Callback used by the exception stores to load exceptions when
669  * initialising.
670  */
671 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
672 {
673         struct dm_snapshot *s = context;
674         struct dm_exception *e;
675
676         e = alloc_completed_exception();
677         if (!e)
678                 return -ENOMEM;
679
680         e->old_chunk = old;
681
682         /* Consecutive_count is implicitly initialised to zero */
683         e->new_chunk = new;
684
685         dm_insert_exception(&s->complete, e);
686
687         return 0;
688 }
689
690 /*
691  * Return a minimum chunk size of all snapshots that have the specified origin.
692  * Return zero if the origin has no snapshots.
693  */
694 static sector_t __minimum_chunk_size(struct origin *o)
695 {
696         struct dm_snapshot *snap;
697         unsigned chunk_size = 0;
698
699         if (o)
700                 list_for_each_entry(snap, &o->snapshots, list)
701                         chunk_size = min_not_zero(chunk_size,
702                                                   snap->store->chunk_size);
703
704         return chunk_size;
705 }
706
707 /*
708  * Hard coded magic.
709  */
710 static int calc_max_buckets(void)
711 {
712         /* use a fixed size of 2MB */
713         unsigned long mem = 2 * 1024 * 1024;
714         mem /= sizeof(struct list_head);
715
716         return mem;
717 }
718
719 /*
720  * Allocate room for a suitable hash table.
721  */
722 static int init_hash_tables(struct dm_snapshot *s)
723 {
724         sector_t hash_size, cow_dev_size, origin_dev_size, max_buckets;
725
726         /*
727          * Calculate based on the size of the original volume or
728          * the COW volume...
729          */
730         cow_dev_size = get_dev_size(s->cow->bdev);
731         origin_dev_size = get_dev_size(s->origin->bdev);
732         max_buckets = calc_max_buckets();
733
734         hash_size = min(origin_dev_size, cow_dev_size) >> s->store->chunk_shift;
735         hash_size = min(hash_size, max_buckets);
736
737         if (hash_size < 64)
738                 hash_size = 64;
739         hash_size = rounddown_pow_of_two(hash_size);
740         if (dm_exception_table_init(&s->complete, hash_size,
741                                     DM_CHUNK_CONSECUTIVE_BITS))
742                 return -ENOMEM;
743
744         /*
745          * Allocate hash table for in-flight exceptions
746          * Make this smaller than the real hash table
747          */
748         hash_size >>= 3;
749         if (hash_size < 64)
750                 hash_size = 64;
751
752         if (dm_exception_table_init(&s->pending, hash_size, 0)) {
753                 dm_exception_table_exit(&s->complete, exception_cache);
754                 return -ENOMEM;
755         }
756
757         return 0;
758 }
759
760 static void merge_shutdown(struct dm_snapshot *s)
761 {
762         clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
763         smp_mb__after_clear_bit();
764         wake_up_bit(&s->state_bits, RUNNING_MERGE);
765 }
766
767 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
768 {
769         s->first_merging_chunk = 0;
770         s->num_merging_chunks = 0;
771
772         return bio_list_get(&s->bios_queued_during_merge);
773 }
774
775 /*
776  * Remove one chunk from the index of completed exceptions.
777  */
778 static int __remove_single_exception_chunk(struct dm_snapshot *s,
779                                            chunk_t old_chunk)
780 {
781         struct dm_exception *e;
782
783         e = dm_lookup_exception(&s->complete, old_chunk);
784         if (!e) {
785                 DMERR("Corruption detected: exception for block %llu is "
786                       "on disk but not in memory",
787                       (unsigned long long)old_chunk);
788                 return -EINVAL;
789         }
790
791         /*
792          * If this is the only chunk using this exception, remove exception.
793          */
794         if (!dm_consecutive_chunk_count(e)) {
795                 dm_remove_exception(e);
796                 free_completed_exception(e);
797                 return 0;
798         }
799
800         /*
801          * The chunk may be either at the beginning or the end of a
802          * group of consecutive chunks - never in the middle.  We are
803          * removing chunks in the opposite order to that in which they
804          * were added, so this should always be true.
805          * Decrement the consecutive chunk counter and adjust the
806          * starting point if necessary.
807          */
808         if (old_chunk == e->old_chunk) {
809                 e->old_chunk++;
810                 e->new_chunk++;
811         } else if (old_chunk != e->old_chunk +
812                    dm_consecutive_chunk_count(e)) {
813                 DMERR("Attempt to merge block %llu from the "
814                       "middle of a chunk range [%llu - %llu]",
815                       (unsigned long long)old_chunk,
816                       (unsigned long long)e->old_chunk,
817                       (unsigned long long)
818                       e->old_chunk + dm_consecutive_chunk_count(e));
819                 return -EINVAL;
820         }
821
822         dm_consecutive_chunk_count_dec(e);
823
824         return 0;
825 }
826
827 static void flush_bios(struct bio *bio);
828
829 static int remove_single_exception_chunk(struct dm_snapshot *s)
830 {
831         struct bio *b = NULL;
832         int r;
833         chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
834
835         down_write(&s->lock);
836
837         /*
838          * Process chunks (and associated exceptions) in reverse order
839          * so that dm_consecutive_chunk_count_dec() accounting works.
840          */
841         do {
842                 r = __remove_single_exception_chunk(s, old_chunk);
843                 if (r)
844                         goto out;
845         } while (old_chunk-- > s->first_merging_chunk);
846
847         b = __release_queued_bios_after_merge(s);
848
849 out:
850         up_write(&s->lock);
851         if (b)
852                 flush_bios(b);
853
854         return r;
855 }
856
857 static int origin_write_extent(struct dm_snapshot *merging_snap,
858                                sector_t sector, unsigned chunk_size);
859
860 static void merge_callback(int read_err, unsigned long write_err,
861                            void *context);
862
863 static uint64_t read_pending_exceptions_done_count(void)
864 {
865         uint64_t pending_exceptions_done;
866
867         spin_lock(&_pending_exceptions_done_spinlock);
868         pending_exceptions_done = _pending_exceptions_done_count;
869         spin_unlock(&_pending_exceptions_done_spinlock);
870
871         return pending_exceptions_done;
872 }
873
874 static void increment_pending_exceptions_done_count(void)
875 {
876         spin_lock(&_pending_exceptions_done_spinlock);
877         _pending_exceptions_done_count++;
878         spin_unlock(&_pending_exceptions_done_spinlock);
879
880         wake_up_all(&_pending_exceptions_done);
881 }
882
883 static void snapshot_merge_next_chunks(struct dm_snapshot *s)
884 {
885         int i, linear_chunks;
886         chunk_t old_chunk, new_chunk;
887         struct dm_io_region src, dest;
888         sector_t io_size;
889         uint64_t previous_count;
890
891         BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
892         if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
893                 goto shut;
894
895         /*
896          * valid flag never changes during merge, so no lock required.
897          */
898         if (!s->valid) {
899                 DMERR("Snapshot is invalid: can't merge");
900                 goto shut;
901         }
902
903         linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
904                                                       &new_chunk);
905         if (linear_chunks <= 0) {
906                 if (linear_chunks < 0) {
907                         DMERR("Read error in exception store: "
908                               "shutting down merge");
909                         down_write(&s->lock);
910                         s->merge_failed = 1;
911                         up_write(&s->lock);
912                 }
913                 goto shut;
914         }
915
916         /* Adjust old_chunk and new_chunk to reflect start of linear region */
917         old_chunk = old_chunk + 1 - linear_chunks;
918         new_chunk = new_chunk + 1 - linear_chunks;
919
920         /*
921          * Use one (potentially large) I/O to copy all 'linear_chunks'
922          * from the exception store to the origin
923          */
924         io_size = linear_chunks * s->store->chunk_size;
925
926         dest.bdev = s->origin->bdev;
927         dest.sector = chunk_to_sector(s->store, old_chunk);
928         dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
929
930         src.bdev = s->cow->bdev;
931         src.sector = chunk_to_sector(s->store, new_chunk);
932         src.count = dest.count;
933
934         /*
935          * Reallocate any exceptions needed in other snapshots then
936          * wait for the pending exceptions to complete.
937          * Each time any pending exception (globally on the system)
938          * completes we are woken and repeat the process to find out
939          * if we can proceed.  While this may not seem a particularly
940          * efficient algorithm, it is not expected to have any
941          * significant impact on performance.
942          */
943         previous_count = read_pending_exceptions_done_count();
944         while (origin_write_extent(s, dest.sector, io_size)) {
945                 wait_event(_pending_exceptions_done,
946                            (read_pending_exceptions_done_count() !=
947                             previous_count));
948                 /* Retry after the wait, until all exceptions are done. */
949                 previous_count = read_pending_exceptions_done_count();
950         }
951
952         down_write(&s->lock);
953         s->first_merging_chunk = old_chunk;
954         s->num_merging_chunks = linear_chunks;
955         up_write(&s->lock);
956
957         /* Wait until writes to all 'linear_chunks' drain */
958         for (i = 0; i < linear_chunks; i++)
959                 __check_for_conflicting_io(s, old_chunk + i);
960
961         dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
962         return;
963
964 shut:
965         merge_shutdown(s);
966 }
967
968 static void error_bios(struct bio *bio);
969
970 static void merge_callback(int read_err, unsigned long write_err, void *context)
971 {
972         struct dm_snapshot *s = context;
973         struct bio *b = NULL;
974
975         if (read_err || write_err) {
976                 if (read_err)
977                         DMERR("Read error: shutting down merge.");
978                 else
979                         DMERR("Write error: shutting down merge.");
980                 goto shut;
981         }
982
983         if (s->store->type->commit_merge(s->store,
984                                          s->num_merging_chunks) < 0) {
985                 DMERR("Write error in exception store: shutting down merge");
986                 goto shut;
987         }
988
989         if (remove_single_exception_chunk(s) < 0)
990                 goto shut;
991
992         snapshot_merge_next_chunks(s);
993
994         return;
995
996 shut:
997         down_write(&s->lock);
998         s->merge_failed = 1;
999         b = __release_queued_bios_after_merge(s);
1000         up_write(&s->lock);
1001         error_bios(b);
1002
1003         merge_shutdown(s);
1004 }
1005
1006 static void start_merge(struct dm_snapshot *s)
1007 {
1008         if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1009                 snapshot_merge_next_chunks(s);
1010 }
1011
1012 static int wait_schedule(void *ptr)
1013 {
1014         schedule();
1015
1016         return 0;
1017 }
1018
1019 /*
1020  * Stop the merging process and wait until it finishes.
1021  */
1022 static void stop_merge(struct dm_snapshot *s)
1023 {
1024         set_bit(SHUTDOWN_MERGE, &s->state_bits);
1025         wait_on_bit(&s->state_bits, RUNNING_MERGE, wait_schedule,
1026                     TASK_UNINTERRUPTIBLE);
1027         clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1028 }
1029
1030 /*
1031  * Construct a snapshot mapping: <origin_dev> <COW-dev> <p/n> <chunk-size>
1032  */
1033 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1034 {
1035         struct dm_snapshot *s;
1036         int i;
1037         int r = -EINVAL;
1038         char *origin_path, *cow_path;
1039         unsigned args_used, num_flush_requests = 1;
1040         fmode_t origin_mode = FMODE_READ;
1041
1042         if (argc != 4) {
1043                 ti->error = "requires exactly 4 arguments";
1044                 r = -EINVAL;
1045                 goto bad;
1046         }
1047
1048         if (dm_target_is_snapshot_merge(ti)) {
1049                 num_flush_requests = 2;
1050                 origin_mode = FMODE_WRITE;
1051         }
1052
1053         s = kmalloc(sizeof(*s), GFP_KERNEL);
1054         if (!s) {
1055                 ti->error = "Cannot allocate private snapshot structure";
1056                 r = -ENOMEM;
1057                 goto bad;
1058         }
1059
1060         origin_path = argv[0];
1061         argv++;
1062         argc--;
1063
1064         r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1065         if (r) {
1066                 ti->error = "Cannot get origin device";
1067                 goto bad_origin;
1068         }
1069
1070         cow_path = argv[0];
1071         argv++;
1072         argc--;
1073
1074         r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1075         if (r) {
1076                 ti->error = "Cannot get COW device";
1077                 goto bad_cow;
1078         }
1079
1080         r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1081         if (r) {
1082                 ti->error = "Couldn't create exception store";
1083                 r = -EINVAL;
1084                 goto bad_store;
1085         }
1086
1087         argv += args_used;
1088         argc -= args_used;
1089
1090         s->ti = ti;
1091         s->valid = 1;
1092         s->active = 0;
1093         atomic_set(&s->pending_exceptions_count, 0);
1094         init_rwsem(&s->lock);
1095         INIT_LIST_HEAD(&s->list);
1096         spin_lock_init(&s->pe_lock);
1097         s->state_bits = 0;
1098         s->merge_failed = 0;
1099         s->first_merging_chunk = 0;
1100         s->num_merging_chunks = 0;
1101         bio_list_init(&s->bios_queued_during_merge);
1102
1103         /* Allocate hash table for COW data */
1104         if (init_hash_tables(s)) {
1105                 ti->error = "Unable to allocate hash table space";
1106                 r = -ENOMEM;
1107                 goto bad_hash_tables;
1108         }
1109
1110         s->kcopyd_client = dm_kcopyd_client_create();
1111         if (IS_ERR(s->kcopyd_client)) {
1112                 r = PTR_ERR(s->kcopyd_client);
1113                 ti->error = "Could not create kcopyd client";
1114                 goto bad_kcopyd;
1115         }
1116
1117         s->pending_pool = mempool_create_slab_pool(MIN_IOS, pending_cache);
1118         if (!s->pending_pool) {
1119                 ti->error = "Could not allocate mempool for pending exceptions";
1120                 goto bad_pending_pool;
1121         }
1122
1123         s->tracked_chunk_pool = mempool_create_slab_pool(MIN_IOS,
1124                                                          tracked_chunk_cache);
1125         if (!s->tracked_chunk_pool) {
1126                 ti->error = "Could not allocate tracked_chunk mempool for "
1127                             "tracking reads";
1128                 goto bad_tracked_chunk_pool;
1129         }
1130
1131         for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1132                 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1133
1134         spin_lock_init(&s->tracked_chunk_lock);
1135
1136         ti->private = s;
1137         ti->num_flush_requests = num_flush_requests;
1138
1139         /* Add snapshot to the list of snapshots for this origin */
1140         /* Exceptions aren't triggered till snapshot_resume() is called */
1141         r = register_snapshot(s);
1142         if (r == -ENOMEM) {
1143                 ti->error = "Snapshot origin struct allocation failed";
1144                 goto bad_load_and_register;
1145         } else if (r < 0) {
1146                 /* invalid handover, register_snapshot has set ti->error */
1147                 goto bad_load_and_register;
1148         }
1149
1150         /*
1151          * Metadata must only be loaded into one table at once, so skip this
1152          * if metadata will be handed over during resume.
1153          * Chunk size will be set during the handover - set it to zero to
1154          * ensure it's ignored.
1155          */
1156         if (r > 0) {
1157                 s->store->chunk_size = 0;
1158                 return 0;
1159         }
1160
1161         r = s->store->type->read_metadata(s->store, dm_add_exception,
1162                                           (void *)s);
1163         if (r < 0) {
1164                 ti->error = "Failed to read snapshot metadata";
1165                 goto bad_read_metadata;
1166         } else if (r > 0) {
1167                 s->valid = 0;
1168                 DMWARN("Snapshot is marked invalid.");
1169         }
1170
1171         if (!s->store->chunk_size) {
1172                 ti->error = "Chunk size not set";
1173                 goto bad_read_metadata;
1174         }
1175         ti->split_io = s->store->chunk_size;
1176
1177         return 0;
1178
1179 bad_read_metadata:
1180         unregister_snapshot(s);
1181
1182 bad_load_and_register:
1183         mempool_destroy(s->tracked_chunk_pool);
1184
1185 bad_tracked_chunk_pool:
1186         mempool_destroy(s->pending_pool);
1187
1188 bad_pending_pool:
1189         dm_kcopyd_client_destroy(s->kcopyd_client);
1190
1191 bad_kcopyd:
1192         dm_exception_table_exit(&s->pending, pending_cache);
1193         dm_exception_table_exit(&s->complete, exception_cache);
1194
1195 bad_hash_tables:
1196         dm_exception_store_destroy(s->store);
1197
1198 bad_store:
1199         dm_put_device(ti, s->cow);
1200
1201 bad_cow:
1202         dm_put_device(ti, s->origin);
1203
1204 bad_origin:
1205         kfree(s);
1206
1207 bad:
1208         return r;
1209 }
1210
1211 static void __free_exceptions(struct dm_snapshot *s)
1212 {
1213         dm_kcopyd_client_destroy(s->kcopyd_client);
1214         s->kcopyd_client = NULL;
1215
1216         dm_exception_table_exit(&s->pending, pending_cache);
1217         dm_exception_table_exit(&s->complete, exception_cache);
1218 }
1219
1220 static void __handover_exceptions(struct dm_snapshot *snap_src,
1221                                   struct dm_snapshot *snap_dest)
1222 {
1223         union {
1224                 struct dm_exception_table table_swap;
1225                 struct dm_exception_store *store_swap;
1226         } u;
1227
1228         /*
1229          * Swap all snapshot context information between the two instances.
1230          */
1231         u.table_swap = snap_dest->complete;
1232         snap_dest->complete = snap_src->complete;
1233         snap_src->complete = u.table_swap;
1234
1235         u.store_swap = snap_dest->store;
1236         snap_dest->store = snap_src->store;
1237         snap_src->store = u.store_swap;
1238
1239         snap_dest->store->snap = snap_dest;
1240         snap_src->store->snap = snap_src;
1241
1242         snap_dest->ti->split_io = snap_dest->store->chunk_size;
1243         snap_dest->valid = snap_src->valid;
1244
1245         /*
1246          * Set source invalid to ensure it receives no further I/O.
1247          */
1248         snap_src->valid = 0;
1249 }
1250
1251 static void snapshot_dtr(struct dm_target *ti)
1252 {
1253 #ifdef CONFIG_DM_DEBUG
1254         int i;
1255 #endif
1256         struct dm_snapshot *s = ti->private;
1257         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1258
1259         down_read(&_origins_lock);
1260         /* Check whether exception handover must be cancelled */
1261         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1262         if (snap_src && snap_dest && (s == snap_src)) {
1263                 down_write(&snap_dest->lock);
1264                 snap_dest->valid = 0;
1265                 up_write(&snap_dest->lock);
1266                 DMERR("Cancelling snapshot handover.");
1267         }
1268         up_read(&_origins_lock);
1269
1270         if (dm_target_is_snapshot_merge(ti))
1271                 stop_merge(s);
1272
1273         /* Prevent further origin writes from using this snapshot. */
1274         /* After this returns there can be no new kcopyd jobs. */
1275         unregister_snapshot(s);
1276
1277         while (atomic_read(&s->pending_exceptions_count))
1278                 msleep(1);
1279         /*
1280          * Ensure instructions in mempool_destroy aren't reordered
1281          * before atomic_read.
1282          */
1283         smp_mb();
1284
1285 #ifdef CONFIG_DM_DEBUG
1286         for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1287                 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1288 #endif
1289
1290         mempool_destroy(s->tracked_chunk_pool);
1291
1292         __free_exceptions(s);
1293
1294         mempool_destroy(s->pending_pool);
1295
1296         dm_exception_store_destroy(s->store);
1297
1298         dm_put_device(ti, s->cow);
1299
1300         dm_put_device(ti, s->origin);
1301
1302         kfree(s);
1303 }
1304
1305 /*
1306  * Flush a list of buffers.
1307  */
1308 static void flush_bios(struct bio *bio)
1309 {
1310         struct bio *n;
1311
1312         while (bio) {
1313                 n = bio->bi_next;
1314                 bio->bi_next = NULL;
1315                 generic_make_request(bio);
1316                 bio = n;
1317         }
1318 }
1319
1320 static int do_origin(struct dm_dev *origin, struct bio *bio);
1321
1322 /*
1323  * Flush a list of buffers.
1324  */
1325 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1326 {
1327         struct bio *n;
1328         int r;
1329
1330         while (bio) {
1331                 n = bio->bi_next;
1332                 bio->bi_next = NULL;
1333                 r = do_origin(s->origin, bio);
1334                 if (r == DM_MAPIO_REMAPPED)
1335                         generic_make_request(bio);
1336                 bio = n;
1337         }
1338 }
1339
1340 /*
1341  * Error a list of buffers.
1342  */
1343 static void error_bios(struct bio *bio)
1344 {
1345         struct bio *n;
1346
1347         while (bio) {
1348                 n = bio->bi_next;
1349                 bio->bi_next = NULL;
1350                 bio_io_error(bio);
1351                 bio = n;
1352         }
1353 }
1354
1355 static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1356 {
1357         if (!s->valid)
1358                 return;
1359
1360         if (err == -EIO)
1361                 DMERR("Invalidating snapshot: Error reading/writing.");
1362         else if (err == -ENOMEM)
1363                 DMERR("Invalidating snapshot: Unable to allocate exception.");
1364
1365         if (s->store->type->drop_snapshot)
1366                 s->store->type->drop_snapshot(s->store);
1367
1368         s->valid = 0;
1369
1370         dm_table_event(s->ti->table);
1371 }
1372
1373 static void pending_complete(struct dm_snap_pending_exception *pe, int success)
1374 {
1375         struct dm_exception *e;
1376         struct dm_snapshot *s = pe->snap;
1377         struct bio *origin_bios = NULL;
1378         struct bio *snapshot_bios = NULL;
1379         struct bio *full_bio = NULL;
1380         int error = 0;
1381
1382         if (!success) {
1383                 /* Read/write error - snapshot is unusable */
1384                 down_write(&s->lock);
1385                 __invalidate_snapshot(s, -EIO);
1386                 error = 1;
1387                 goto out;
1388         }
1389
1390         e = alloc_completed_exception();
1391         if (!e) {
1392                 down_write(&s->lock);
1393                 __invalidate_snapshot(s, -ENOMEM);
1394                 error = 1;
1395                 goto out;
1396         }
1397         *e = pe->e;
1398
1399         down_write(&s->lock);
1400         if (!s->valid) {
1401                 free_completed_exception(e);
1402                 error = 1;
1403                 goto out;
1404         }
1405
1406         /* Check for conflicting reads */
1407         __check_for_conflicting_io(s, pe->e.old_chunk);
1408
1409         /*
1410          * Add a proper exception, and remove the
1411          * in-flight exception from the list.
1412          */
1413         dm_insert_exception(&s->complete, e);
1414
1415 out:
1416         dm_remove_exception(&pe->e);
1417         snapshot_bios = bio_list_get(&pe->snapshot_bios);
1418         origin_bios = bio_list_get(&pe->origin_bios);
1419         full_bio = pe->full_bio;
1420         if (full_bio) {
1421                 full_bio->bi_end_io = pe->full_bio_end_io;
1422                 full_bio->bi_private = pe->full_bio_private;
1423         }
1424         free_pending_exception(pe);
1425
1426         increment_pending_exceptions_done_count();
1427
1428         up_write(&s->lock);
1429
1430         /* Submit any pending write bios */
1431         if (error) {
1432                 if (full_bio)
1433                         bio_io_error(full_bio);
1434                 error_bios(snapshot_bios);
1435         } else {
1436                 if (full_bio)
1437                         bio_endio(full_bio, 0);
1438                 flush_bios(snapshot_bios);
1439         }
1440
1441         retry_origin_bios(s, origin_bios);
1442 }
1443
1444 static void commit_callback(void *context, int success)
1445 {
1446         struct dm_snap_pending_exception *pe = context;
1447
1448         pending_complete(pe, success);
1449 }
1450
1451 /*
1452  * Called when the copy I/O has finished.  kcopyd actually runs
1453  * this code so don't block.
1454  */
1455 static void copy_callback(int read_err, unsigned long write_err, void *context)
1456 {
1457         struct dm_snap_pending_exception *pe = context;
1458         struct dm_snapshot *s = pe->snap;
1459
1460         if (read_err || write_err)
1461                 pending_complete(pe, 0);
1462
1463         else
1464                 /* Update the metadata if we are persistent */
1465                 s->store->type->commit_exception(s->store, &pe->e,
1466                                                  commit_callback, pe);
1467 }
1468
1469 /*
1470  * Dispatches the copy operation to kcopyd.
1471  */
1472 static void start_copy(struct dm_snap_pending_exception *pe)
1473 {
1474         struct dm_snapshot *s = pe->snap;
1475         struct dm_io_region src, dest;
1476         struct block_device *bdev = s->origin->bdev;
1477         sector_t dev_size;
1478
1479         dev_size = get_dev_size(bdev);
1480
1481         src.bdev = bdev;
1482         src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1483         src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1484
1485         dest.bdev = s->cow->bdev;
1486         dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1487         dest.count = src.count;
1488
1489         /* Hand over to kcopyd */
1490         dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1491 }
1492
1493 static void full_bio_end_io(struct bio *bio, int error)
1494 {
1495         void *callback_data = bio->bi_private;
1496
1497         dm_kcopyd_do_callback(callback_data, 0, error ? 1 : 0);
1498 }
1499
1500 static void start_full_bio(struct dm_snap_pending_exception *pe,
1501                            struct bio *bio)
1502 {
1503         struct dm_snapshot *s = pe->snap;
1504         void *callback_data;
1505
1506         pe->full_bio = bio;
1507         pe->full_bio_end_io = bio->bi_end_io;
1508         pe->full_bio_private = bio->bi_private;
1509
1510         callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1511                                                    copy_callback, pe);
1512
1513         bio->bi_end_io = full_bio_end_io;
1514         bio->bi_private = callback_data;
1515
1516         generic_make_request(bio);
1517 }
1518
1519 static struct dm_snap_pending_exception *
1520 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1521 {
1522         struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1523
1524         if (!e)
1525                 return NULL;
1526
1527         return container_of(e, struct dm_snap_pending_exception, e);
1528 }
1529
1530 /*
1531  * Looks to see if this snapshot already has a pending exception
1532  * for this chunk, otherwise it allocates a new one and inserts
1533  * it into the pending table.
1534  *
1535  * NOTE: a write lock must be held on snap->lock before calling
1536  * this.
1537  */
1538 static struct dm_snap_pending_exception *
1539 __find_pending_exception(struct dm_snapshot *s,
1540                          struct dm_snap_pending_exception *pe, chunk_t chunk)
1541 {
1542         struct dm_snap_pending_exception *pe2;
1543
1544         pe2 = __lookup_pending_exception(s, chunk);
1545         if (pe2) {
1546                 free_pending_exception(pe);
1547                 return pe2;
1548         }
1549
1550         pe->e.old_chunk = chunk;
1551         bio_list_init(&pe->origin_bios);
1552         bio_list_init(&pe->snapshot_bios);
1553         pe->started = 0;
1554         pe->full_bio = NULL;
1555
1556         if (s->store->type->prepare_exception(s->store, &pe->e)) {
1557                 free_pending_exception(pe);
1558                 return NULL;
1559         }
1560
1561         dm_insert_exception(&s->pending, &pe->e);
1562
1563         return pe;
1564 }
1565
1566 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1567                             struct bio *bio, chunk_t chunk)
1568 {
1569         bio->bi_bdev = s->cow->bdev;
1570         bio->bi_sector = chunk_to_sector(s->store,
1571                                          dm_chunk_number(e->new_chunk) +
1572                                          (chunk - e->old_chunk)) +
1573                                          (bio->bi_sector &
1574                                           s->store->chunk_mask);
1575 }
1576
1577 static int snapshot_map(struct dm_target *ti, struct bio *bio,
1578                         union map_info *map_context)
1579 {
1580         struct dm_exception *e;
1581         struct dm_snapshot *s = ti->private;
1582         int r = DM_MAPIO_REMAPPED;
1583         chunk_t chunk;
1584         struct dm_snap_pending_exception *pe = NULL;
1585
1586         if (bio->bi_rw & REQ_FLUSH) {
1587                 bio->bi_bdev = s->cow->bdev;
1588                 return DM_MAPIO_REMAPPED;
1589         }
1590
1591         chunk = sector_to_chunk(s->store, bio->bi_sector);
1592
1593         /* Full snapshots are not usable */
1594         /* To get here the table must be live so s->active is always set. */
1595         if (!s->valid)
1596                 return -EIO;
1597
1598         /* FIXME: should only take write lock if we need
1599          * to copy an exception */
1600         down_write(&s->lock);
1601
1602         if (!s->valid) {
1603                 r = -EIO;
1604                 goto out_unlock;
1605         }
1606
1607         /* If the block is already remapped - use that, else remap it */
1608         e = dm_lookup_exception(&s->complete, chunk);
1609         if (e) {
1610                 remap_exception(s, e, bio, chunk);
1611                 goto out_unlock;
1612         }
1613
1614         /*
1615          * Write to snapshot - higher level takes care of RW/RO
1616          * flags so we should only get this if we are
1617          * writeable.
1618          */
1619         if (bio_rw(bio) == WRITE) {
1620                 pe = __lookup_pending_exception(s, chunk);
1621                 if (!pe) {
1622                         up_write(&s->lock);
1623                         pe = alloc_pending_exception(s);
1624                         down_write(&s->lock);
1625
1626                         if (!s->valid) {
1627                                 free_pending_exception(pe);
1628                                 r = -EIO;
1629                                 goto out_unlock;
1630                         }
1631
1632                         e = dm_lookup_exception(&s->complete, chunk);
1633                         if (e) {
1634                                 free_pending_exception(pe);
1635                                 remap_exception(s, e, bio, chunk);
1636                                 goto out_unlock;
1637                         }
1638
1639                         pe = __find_pending_exception(s, pe, chunk);
1640                         if (!pe) {
1641                                 __invalidate_snapshot(s, -ENOMEM);
1642                                 r = -EIO;
1643                                 goto out_unlock;
1644                         }
1645                 }
1646
1647                 remap_exception(s, &pe->e, bio, chunk);
1648
1649                 r = DM_MAPIO_SUBMITTED;
1650
1651                 if (!pe->started &&
1652                     bio->bi_size == (s->store->chunk_size << SECTOR_SHIFT)) {
1653                         pe->started = 1;
1654                         up_write(&s->lock);
1655                         start_full_bio(pe, bio);
1656                         goto out;
1657                 }
1658
1659                 bio_list_add(&pe->snapshot_bios, bio);
1660
1661                 if (!pe->started) {
1662                         /* this is protected by snap->lock */
1663                         pe->started = 1;
1664                         up_write(&s->lock);
1665                         start_copy(pe);
1666                         goto out;
1667                 }
1668         } else {
1669                 bio->bi_bdev = s->origin->bdev;
1670                 map_context->ptr = track_chunk(s, chunk);
1671         }
1672
1673 out_unlock:
1674         up_write(&s->lock);
1675 out:
1676         return r;
1677 }
1678
1679 /*
1680  * A snapshot-merge target behaves like a combination of a snapshot
1681  * target and a snapshot-origin target.  It only generates new
1682  * exceptions in other snapshots and not in the one that is being
1683  * merged.
1684  *
1685  * For each chunk, if there is an existing exception, it is used to
1686  * redirect I/O to the cow device.  Otherwise I/O is sent to the origin,
1687  * which in turn might generate exceptions in other snapshots.
1688  * If merging is currently taking place on the chunk in question, the
1689  * I/O is deferred by adding it to s->bios_queued_during_merge.
1690  */
1691 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio,
1692                               union map_info *map_context)
1693 {
1694         struct dm_exception *e;
1695         struct dm_snapshot *s = ti->private;
1696         int r = DM_MAPIO_REMAPPED;
1697         chunk_t chunk;
1698
1699         if (bio->bi_rw & REQ_FLUSH) {
1700                 if (!map_context->target_request_nr)
1701                         bio->bi_bdev = s->origin->bdev;
1702                 else
1703                         bio->bi_bdev = s->cow->bdev;
1704                 map_context->ptr = NULL;
1705                 return DM_MAPIO_REMAPPED;
1706         }
1707
1708         chunk = sector_to_chunk(s->store, bio->bi_sector);
1709
1710         down_write(&s->lock);
1711
1712         /* Full merging snapshots are redirected to the origin */
1713         if (!s->valid)
1714                 goto redirect_to_origin;
1715
1716         /* If the block is already remapped - use that */
1717         e = dm_lookup_exception(&s->complete, chunk);
1718         if (e) {
1719                 /* Queue writes overlapping with chunks being merged */
1720                 if (bio_rw(bio) == WRITE &&
1721                     chunk >= s->first_merging_chunk &&
1722                     chunk < (s->first_merging_chunk +
1723                              s->num_merging_chunks)) {
1724                         bio->bi_bdev = s->origin->bdev;
1725                         bio_list_add(&s->bios_queued_during_merge, bio);
1726                         r = DM_MAPIO_SUBMITTED;
1727                         goto out_unlock;
1728                 }
1729
1730                 remap_exception(s, e, bio, chunk);
1731
1732                 if (bio_rw(bio) == WRITE)
1733                         map_context->ptr = track_chunk(s, chunk);
1734                 goto out_unlock;
1735         }
1736
1737 redirect_to_origin:
1738         bio->bi_bdev = s->origin->bdev;
1739
1740         if (bio_rw(bio) == WRITE) {
1741                 up_write(&s->lock);
1742                 return do_origin(s->origin, bio);
1743         }
1744
1745 out_unlock:
1746         up_write(&s->lock);
1747
1748         return r;
1749 }
1750
1751 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
1752                            int error, union map_info *map_context)
1753 {
1754         struct dm_snapshot *s = ti->private;
1755         struct dm_snap_tracked_chunk *c = map_context->ptr;
1756
1757         if (c)
1758                 stop_tracking_chunk(s, c);
1759
1760         return 0;
1761 }
1762
1763 static void snapshot_merge_presuspend(struct dm_target *ti)
1764 {
1765         struct dm_snapshot *s = ti->private;
1766
1767         stop_merge(s);
1768 }
1769
1770 static int snapshot_preresume(struct dm_target *ti)
1771 {
1772         int r = 0;
1773         struct dm_snapshot *s = ti->private;
1774         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1775
1776         down_read(&_origins_lock);
1777         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1778         if (snap_src && snap_dest) {
1779                 down_read(&snap_src->lock);
1780                 if (s == snap_src) {
1781                         DMERR("Unable to resume snapshot source until "
1782                               "handover completes.");
1783                         r = -EINVAL;
1784                 } else if (!dm_suspended(snap_src->ti)) {
1785                         DMERR("Unable to perform snapshot handover until "
1786                               "source is suspended.");
1787                         r = -EINVAL;
1788                 }
1789                 up_read(&snap_src->lock);
1790         }
1791         up_read(&_origins_lock);
1792
1793         return r;
1794 }
1795
1796 static void snapshot_resume(struct dm_target *ti)
1797 {
1798         struct dm_snapshot *s = ti->private;
1799         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1800
1801         down_read(&_origins_lock);
1802         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1803         if (snap_src && snap_dest) {
1804                 down_write(&snap_src->lock);
1805                 down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
1806                 __handover_exceptions(snap_src, snap_dest);
1807                 up_write(&snap_dest->lock);
1808                 up_write(&snap_src->lock);
1809         }
1810         up_read(&_origins_lock);
1811
1812         /* Now we have correct chunk size, reregister */
1813         reregister_snapshot(s);
1814
1815         down_write(&s->lock);
1816         s->active = 1;
1817         up_write(&s->lock);
1818 }
1819
1820 static sector_t get_origin_minimum_chunksize(struct block_device *bdev)
1821 {
1822         sector_t min_chunksize;
1823
1824         down_read(&_origins_lock);
1825         min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
1826         up_read(&_origins_lock);
1827
1828         return min_chunksize;
1829 }
1830
1831 static void snapshot_merge_resume(struct dm_target *ti)
1832 {
1833         struct dm_snapshot *s = ti->private;
1834
1835         /*
1836          * Handover exceptions from existing snapshot.
1837          */
1838         snapshot_resume(ti);
1839
1840         /*
1841          * snapshot-merge acts as an origin, so set ti->split_io
1842          */
1843         ti->split_io = get_origin_minimum_chunksize(s->origin->bdev);
1844
1845         start_merge(s);
1846 }
1847
1848 static int snapshot_status(struct dm_target *ti, status_type_t type,
1849                            char *result, unsigned int maxlen)
1850 {
1851         unsigned sz = 0;
1852         struct dm_snapshot *snap = ti->private;
1853
1854         switch (type) {
1855         case STATUSTYPE_INFO:
1856
1857                 down_write(&snap->lock);
1858
1859                 if (!snap->valid)
1860                         DMEMIT("Invalid");
1861                 else if (snap->merge_failed)
1862                         DMEMIT("Merge failed");
1863                 else {
1864                         if (snap->store->type->usage) {
1865                                 sector_t total_sectors, sectors_allocated,
1866                                          metadata_sectors;
1867                                 snap->store->type->usage(snap->store,
1868                                                          &total_sectors,
1869                                                          &sectors_allocated,
1870                                                          &metadata_sectors);
1871                                 DMEMIT("%llu/%llu %llu",
1872                                        (unsigned long long)sectors_allocated,
1873                                        (unsigned long long)total_sectors,
1874                                        (unsigned long long)metadata_sectors);
1875                         }
1876                         else
1877                                 DMEMIT("Unknown");
1878                 }
1879
1880                 up_write(&snap->lock);
1881
1882                 break;
1883
1884         case STATUSTYPE_TABLE:
1885                 /*
1886                  * kdevname returns a static pointer so we need
1887                  * to make private copies if the output is to
1888                  * make sense.
1889                  */
1890                 DMEMIT("%s %s", snap->origin->name, snap->cow->name);
1891                 snap->store->type->status(snap->store, type, result + sz,
1892                                           maxlen - sz);
1893                 break;
1894         }
1895
1896         return 0;
1897 }
1898
1899 static int snapshot_iterate_devices(struct dm_target *ti,
1900                                     iterate_devices_callout_fn fn, void *data)
1901 {
1902         struct dm_snapshot *snap = ti->private;
1903         int r;
1904
1905         r = fn(ti, snap->origin, 0, ti->len, data);
1906
1907         if (!r)
1908                 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
1909
1910         return r;
1911 }
1912
1913
1914 /*-----------------------------------------------------------------
1915  * Origin methods
1916  *---------------------------------------------------------------*/
1917
1918 /*
1919  * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
1920  * supplied bio was ignored.  The caller may submit it immediately.
1921  * (No remapping actually occurs as the origin is always a direct linear
1922  * map.)
1923  *
1924  * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
1925  * and any supplied bio is added to a list to be submitted once all
1926  * the necessary exceptions exist.
1927  */
1928 static int __origin_write(struct list_head *snapshots, sector_t sector,
1929                           struct bio *bio)
1930 {
1931         int r = DM_MAPIO_REMAPPED;
1932         struct dm_snapshot *snap;
1933         struct dm_exception *e;
1934         struct dm_snap_pending_exception *pe;
1935         struct dm_snap_pending_exception *pe_to_start_now = NULL;
1936         struct dm_snap_pending_exception *pe_to_start_last = NULL;
1937         chunk_t chunk;
1938
1939         /* Do all the snapshots on this origin */
1940         list_for_each_entry (snap, snapshots, list) {
1941                 /*
1942                  * Don't make new exceptions in a merging snapshot
1943                  * because it has effectively been deleted
1944                  */
1945                 if (dm_target_is_snapshot_merge(snap->ti))
1946                         continue;
1947
1948                 down_write(&snap->lock);
1949
1950                 /* Only deal with valid and active snapshots */
1951                 if (!snap->valid || !snap->active)
1952                         goto next_snapshot;
1953
1954                 /* Nothing to do if writing beyond end of snapshot */
1955                 if (sector >= dm_table_get_size(snap->ti->table))
1956                         goto next_snapshot;
1957
1958                 /*
1959                  * Remember, different snapshots can have
1960                  * different chunk sizes.
1961                  */
1962                 chunk = sector_to_chunk(snap->store, sector);
1963
1964                 /*
1965                  * Check exception table to see if block
1966                  * is already remapped in this snapshot
1967                  * and trigger an exception if not.
1968                  */
1969                 e = dm_lookup_exception(&snap->complete, chunk);
1970                 if (e)
1971                         goto next_snapshot;
1972
1973                 pe = __lookup_pending_exception(snap, chunk);
1974                 if (!pe) {
1975                         up_write(&snap->lock);
1976                         pe = alloc_pending_exception(snap);
1977                         down_write(&snap->lock);
1978
1979                         if (!snap->valid) {
1980                                 free_pending_exception(pe);
1981                                 goto next_snapshot;
1982                         }
1983
1984                         e = dm_lookup_exception(&snap->complete, chunk);
1985                         if (e) {
1986                                 free_pending_exception(pe);
1987                                 goto next_snapshot;
1988                         }
1989
1990                         pe = __find_pending_exception(snap, pe, chunk);
1991                         if (!pe) {
1992                                 __invalidate_snapshot(snap, -ENOMEM);
1993                                 goto next_snapshot;
1994                         }
1995                 }
1996
1997                 r = DM_MAPIO_SUBMITTED;
1998
1999                 /*
2000                  * If an origin bio was supplied, queue it to wait for the
2001                  * completion of this exception, and start this one last,
2002                  * at the end of the function.
2003                  */
2004                 if (bio) {
2005                         bio_list_add(&pe->origin_bios, bio);
2006                         bio = NULL;
2007
2008                         if (!pe->started) {
2009                                 pe->started = 1;
2010                                 pe_to_start_last = pe;
2011                         }
2012                 }
2013
2014                 if (!pe->started) {
2015                         pe->started = 1;
2016                         pe_to_start_now = pe;
2017                 }
2018
2019 next_snapshot:
2020                 up_write(&snap->lock);
2021
2022                 if (pe_to_start_now) {
2023                         start_copy(pe_to_start_now);
2024                         pe_to_start_now = NULL;
2025                 }
2026         }
2027
2028         /*
2029          * Submit the exception against which the bio is queued last,
2030          * to give the other exceptions a head start.
2031          */
2032         if (pe_to_start_last)
2033                 start_copy(pe_to_start_last);
2034
2035         return r;
2036 }
2037
2038 /*
2039  * Called on a write from the origin driver.
2040  */
2041 static int do_origin(struct dm_dev *origin, struct bio *bio)
2042 {
2043         struct origin *o;
2044         int r = DM_MAPIO_REMAPPED;
2045
2046         down_read(&_origins_lock);
2047         o = __lookup_origin(origin->bdev);
2048         if (o)
2049                 r = __origin_write(&o->snapshots, bio->bi_sector, bio);
2050         up_read(&_origins_lock);
2051
2052         return r;
2053 }
2054
2055 /*
2056  * Trigger exceptions in all non-merging snapshots.
2057  *
2058  * The chunk size of the merging snapshot may be larger than the chunk
2059  * size of some other snapshot so we may need to reallocate multiple
2060  * chunks in other snapshots.
2061  *
2062  * We scan all the overlapping exceptions in the other snapshots.
2063  * Returns 1 if anything was reallocated and must be waited for,
2064  * otherwise returns 0.
2065  *
2066  * size must be a multiple of merging_snap's chunk_size.
2067  */
2068 static int origin_write_extent(struct dm_snapshot *merging_snap,
2069                                sector_t sector, unsigned size)
2070 {
2071         int must_wait = 0;
2072         sector_t n;
2073         struct origin *o;
2074
2075         /*
2076          * The origin's __minimum_chunk_size() got stored in split_io
2077          * by snapshot_merge_resume().
2078          */
2079         down_read(&_origins_lock);
2080         o = __lookup_origin(merging_snap->origin->bdev);
2081         for (n = 0; n < size; n += merging_snap->ti->split_io)
2082                 if (__origin_write(&o->snapshots, sector + n, NULL) ==
2083                     DM_MAPIO_SUBMITTED)
2084                         must_wait = 1;
2085         up_read(&_origins_lock);
2086
2087         return must_wait;
2088 }
2089
2090 /*
2091  * Origin: maps a linear range of a device, with hooks for snapshotting.
2092  */
2093
2094 /*
2095  * Construct an origin mapping: <dev_path>
2096  * The context for an origin is merely a 'struct dm_dev *'
2097  * pointing to the real device.
2098  */
2099 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2100 {
2101         int r;
2102         struct dm_dev *dev;
2103
2104         if (argc != 1) {
2105                 ti->error = "origin: incorrect number of arguments";
2106                 return -EINVAL;
2107         }
2108
2109         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &dev);
2110         if (r) {
2111                 ti->error = "Cannot get target device";
2112                 return r;
2113         }
2114
2115         ti->private = dev;
2116         ti->num_flush_requests = 1;
2117
2118         return 0;
2119 }
2120
2121 static void origin_dtr(struct dm_target *ti)
2122 {
2123         struct dm_dev *dev = ti->private;
2124         dm_put_device(ti, dev);
2125 }
2126
2127 static int origin_map(struct dm_target *ti, struct bio *bio,
2128                       union map_info *map_context)
2129 {
2130         struct dm_dev *dev = ti->private;
2131         bio->bi_bdev = dev->bdev;
2132
2133         if (bio->bi_rw & REQ_FLUSH)
2134                 return DM_MAPIO_REMAPPED;
2135
2136         /* Only tell snapshots if this is a write */
2137         return (bio_rw(bio) == WRITE) ? do_origin(dev, bio) : DM_MAPIO_REMAPPED;
2138 }
2139
2140 /*
2141  * Set the target "split_io" field to the minimum of all the snapshots'
2142  * chunk sizes.
2143  */
2144 static void origin_resume(struct dm_target *ti)
2145 {
2146         struct dm_dev *dev = ti->private;
2147
2148         ti->split_io = get_origin_minimum_chunksize(dev->bdev);
2149 }
2150
2151 static int origin_status(struct dm_target *ti, status_type_t type, char *result,
2152                          unsigned int maxlen)
2153 {
2154         struct dm_dev *dev = ti->private;
2155
2156         switch (type) {
2157         case STATUSTYPE_INFO:
2158                 result[0] = '\0';
2159                 break;
2160
2161         case STATUSTYPE_TABLE:
2162                 snprintf(result, maxlen, "%s", dev->name);
2163                 break;
2164         }
2165
2166         return 0;
2167 }
2168
2169 static int origin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2170                         struct bio_vec *biovec, int max_size)
2171 {
2172         struct dm_dev *dev = ti->private;
2173         struct request_queue *q = bdev_get_queue(dev->bdev);
2174
2175         if (!q->merge_bvec_fn)
2176                 return max_size;
2177
2178         bvm->bi_bdev = dev->bdev;
2179         bvm->bi_sector = bvm->bi_sector;
2180
2181         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2182 }
2183
2184 static int origin_iterate_devices(struct dm_target *ti,
2185                                   iterate_devices_callout_fn fn, void *data)
2186 {
2187         struct dm_dev *dev = ti->private;
2188
2189         return fn(ti, dev, 0, ti->len, data);
2190 }
2191
2192 static struct target_type origin_target = {
2193         .name    = "snapshot-origin",
2194         .version = {1, 7, 1},
2195         .module  = THIS_MODULE,
2196         .ctr     = origin_ctr,
2197         .dtr     = origin_dtr,
2198         .map     = origin_map,
2199         .resume  = origin_resume,
2200         .status  = origin_status,
2201         .merge   = origin_merge,
2202         .iterate_devices = origin_iterate_devices,
2203 };
2204
2205 static struct target_type snapshot_target = {
2206         .name    = "snapshot",
2207         .version = {1, 10, 0},
2208         .module  = THIS_MODULE,
2209         .ctr     = snapshot_ctr,
2210         .dtr     = snapshot_dtr,
2211         .map     = snapshot_map,
2212         .end_io  = snapshot_end_io,
2213         .preresume  = snapshot_preresume,
2214         .resume  = snapshot_resume,
2215         .status  = snapshot_status,
2216         .iterate_devices = snapshot_iterate_devices,
2217 };
2218
2219 static struct target_type merge_target = {
2220         .name    = dm_snapshot_merge_target_name,
2221         .version = {1, 1, 0},
2222         .module  = THIS_MODULE,
2223         .ctr     = snapshot_ctr,
2224         .dtr     = snapshot_dtr,
2225         .map     = snapshot_merge_map,
2226         .end_io  = snapshot_end_io,
2227         .presuspend = snapshot_merge_presuspend,
2228         .preresume  = snapshot_preresume,
2229         .resume  = snapshot_merge_resume,
2230         .status  = snapshot_status,
2231         .iterate_devices = snapshot_iterate_devices,
2232 };
2233
2234 static int __init dm_snapshot_init(void)
2235 {
2236         int r;
2237
2238         r = dm_exception_store_init();
2239         if (r) {
2240                 DMERR("Failed to initialize exception stores");
2241                 return r;
2242         }
2243
2244         r = dm_register_target(&snapshot_target);
2245         if (r < 0) {
2246                 DMERR("snapshot target register failed %d", r);
2247                 goto bad_register_snapshot_target;
2248         }
2249
2250         r = dm_register_target(&origin_target);
2251         if (r < 0) {
2252                 DMERR("Origin target register failed %d", r);
2253                 goto bad_register_origin_target;
2254         }
2255
2256         r = dm_register_target(&merge_target);
2257         if (r < 0) {
2258                 DMERR("Merge target register failed %d", r);
2259                 goto bad_register_merge_target;
2260         }
2261
2262         r = init_origin_hash();
2263         if (r) {
2264                 DMERR("init_origin_hash failed.");
2265                 goto bad_origin_hash;
2266         }
2267
2268         exception_cache = KMEM_CACHE(dm_exception, 0);
2269         if (!exception_cache) {
2270                 DMERR("Couldn't create exception cache.");
2271                 r = -ENOMEM;
2272                 goto bad_exception_cache;
2273         }
2274
2275         pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2276         if (!pending_cache) {
2277                 DMERR("Couldn't create pending cache.");
2278                 r = -ENOMEM;
2279                 goto bad_pending_cache;
2280         }
2281
2282         tracked_chunk_cache = KMEM_CACHE(dm_snap_tracked_chunk, 0);
2283         if (!tracked_chunk_cache) {
2284                 DMERR("Couldn't create cache to track chunks in use.");
2285                 r = -ENOMEM;
2286                 goto bad_tracked_chunk_cache;
2287         }
2288
2289         return 0;
2290
2291 bad_tracked_chunk_cache:
2292         kmem_cache_destroy(pending_cache);
2293 bad_pending_cache:
2294         kmem_cache_destroy(exception_cache);
2295 bad_exception_cache:
2296         exit_origin_hash();
2297 bad_origin_hash:
2298         dm_unregister_target(&merge_target);
2299 bad_register_merge_target:
2300         dm_unregister_target(&origin_target);
2301 bad_register_origin_target:
2302         dm_unregister_target(&snapshot_target);
2303 bad_register_snapshot_target:
2304         dm_exception_store_exit();
2305
2306         return r;
2307 }
2308
2309 static void __exit dm_snapshot_exit(void)
2310 {
2311         dm_unregister_target(&snapshot_target);
2312         dm_unregister_target(&origin_target);
2313         dm_unregister_target(&merge_target);
2314
2315         exit_origin_hash();
2316         kmem_cache_destroy(pending_cache);
2317         kmem_cache_destroy(exception_cache);
2318         kmem_cache_destroy(tracked_chunk_cache);
2319
2320         dm_exception_store_exit();
2321 }
2322
2323 /* Module hooks */
2324 module_init(dm_snapshot_init);
2325 module_exit(dm_snapshot_exit);
2326
2327 MODULE_DESCRIPTION(DM_NAME " snapshot target");
2328 MODULE_AUTHOR("Joe Thornber");
2329 MODULE_LICENSE("GPL");