]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/md/dm-snap.c
223e7ebf24f7e7f4198a49bba48615885c719e20
[~shefty/rdma-dev.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
202         c->chunk = chunk;
203
204         spin_lock_irq(&s->tracked_chunk_lock);
205         hlist_add_head(&c->node,
206                        &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
207         spin_unlock_irq(&s->tracked_chunk_lock);
208
209         return c;
210 }
211
212 static void stop_tracking_chunk(struct dm_snapshot *s,
213                                 struct dm_snap_tracked_chunk *c)
214 {
215         unsigned long flags;
216
217         spin_lock_irqsave(&s->tracked_chunk_lock, flags);
218         hlist_del(&c->node);
219         spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
220
221         mempool_free(c, s->tracked_chunk_pool);
222 }
223
224 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
225 {
226         struct dm_snap_tracked_chunk *c;
227         struct hlist_node *hn;
228         int found = 0;
229
230         spin_lock_irq(&s->tracked_chunk_lock);
231
232         hlist_for_each_entry(c, hn,
233             &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
234                 if (c->chunk == chunk) {
235                         found = 1;
236                         break;
237                 }
238         }
239
240         spin_unlock_irq(&s->tracked_chunk_lock);
241
242         return found;
243 }
244
245 /*
246  * This conflicting I/O is extremely improbable in the caller,
247  * so msleep(1) is sufficient and there is no need for a wait queue.
248  */
249 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
250 {
251         while (__chunk_is_tracked(s, chunk))
252                 msleep(1);
253 }
254
255 /*
256  * One of these per registered origin, held in the snapshot_origins hash
257  */
258 struct origin {
259         /* The origin device */
260         struct block_device *bdev;
261
262         struct list_head hash_list;
263
264         /* List of snapshots for this origin */
265         struct list_head snapshots;
266 };
267
268 /*
269  * Size of the hash table for origin volumes. If we make this
270  * the size of the minors list then it should be nearly perfect
271  */
272 #define ORIGIN_HASH_SIZE 256
273 #define ORIGIN_MASK      0xFF
274 static struct list_head *_origins;
275 static struct rw_semaphore _origins_lock;
276
277 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
278 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
279 static uint64_t _pending_exceptions_done_count;
280
281 static int init_origin_hash(void)
282 {
283         int i;
284
285         _origins = kmalloc(ORIGIN_HASH_SIZE * sizeof(struct list_head),
286                            GFP_KERNEL);
287         if (!_origins) {
288                 DMERR("unable to allocate memory");
289                 return -ENOMEM;
290         }
291
292         for (i = 0; i < ORIGIN_HASH_SIZE; i++)
293                 INIT_LIST_HEAD(_origins + i);
294         init_rwsem(&_origins_lock);
295
296         return 0;
297 }
298
299 static void exit_origin_hash(void)
300 {
301         kfree(_origins);
302 }
303
304 static unsigned origin_hash(struct block_device *bdev)
305 {
306         return bdev->bd_dev & ORIGIN_MASK;
307 }
308
309 static struct origin *__lookup_origin(struct block_device *origin)
310 {
311         struct list_head *ol;
312         struct origin *o;
313
314         ol = &_origins[origin_hash(origin)];
315         list_for_each_entry (o, ol, hash_list)
316                 if (bdev_equal(o->bdev, origin))
317                         return o;
318
319         return NULL;
320 }
321
322 static void __insert_origin(struct origin *o)
323 {
324         struct list_head *sl = &_origins[origin_hash(o->bdev)];
325         list_add_tail(&o->hash_list, sl);
326 }
327
328 /*
329  * _origins_lock must be held when calling this function.
330  * Returns number of snapshots registered using the supplied cow device, plus:
331  * snap_src - a snapshot suitable for use as a source of exception handover
332  * snap_dest - a snapshot capable of receiving exception handover.
333  * snap_merge - an existing snapshot-merge target linked to the same origin.
334  *   There can be at most one snapshot-merge target. The parameter is optional.
335  *
336  * Possible return values and states of snap_src and snap_dest.
337  *   0: NULL, NULL  - first new snapshot
338  *   1: snap_src, NULL - normal snapshot
339  *   2: snap_src, snap_dest  - waiting for handover
340  *   2: snap_src, NULL - handed over, waiting for old to be deleted
341  *   1: NULL, snap_dest - source got destroyed without handover
342  */
343 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
344                                         struct dm_snapshot **snap_src,
345                                         struct dm_snapshot **snap_dest,
346                                         struct dm_snapshot **snap_merge)
347 {
348         struct dm_snapshot *s;
349         struct origin *o;
350         int count = 0;
351         int active;
352
353         o = __lookup_origin(snap->origin->bdev);
354         if (!o)
355                 goto out;
356
357         list_for_each_entry(s, &o->snapshots, list) {
358                 if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
359                         *snap_merge = s;
360                 if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
361                         continue;
362
363                 down_read(&s->lock);
364                 active = s->active;
365                 up_read(&s->lock);
366
367                 if (active) {
368                         if (snap_src)
369                                 *snap_src = s;
370                 } else if (snap_dest)
371                         *snap_dest = s;
372
373                 count++;
374         }
375
376 out:
377         return count;
378 }
379
380 /*
381  * On success, returns 1 if this snapshot is a handover destination,
382  * otherwise returns 0.
383  */
384 static int __validate_exception_handover(struct dm_snapshot *snap)
385 {
386         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
387         struct dm_snapshot *snap_merge = NULL;
388
389         /* Does snapshot need exceptions handed over to it? */
390         if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
391                                           &snap_merge) == 2) ||
392             snap_dest) {
393                 snap->ti->error = "Snapshot cow pairing for exception "
394                                   "table handover failed";
395                 return -EINVAL;
396         }
397
398         /*
399          * If no snap_src was found, snap cannot become a handover
400          * destination.
401          */
402         if (!snap_src)
403                 return 0;
404
405         /*
406          * Non-snapshot-merge handover?
407          */
408         if (!dm_target_is_snapshot_merge(snap->ti))
409                 return 1;
410
411         /*
412          * Do not allow more than one merging snapshot.
413          */
414         if (snap_merge) {
415                 snap->ti->error = "A snapshot is already merging.";
416                 return -EINVAL;
417         }
418
419         if (!snap_src->store->type->prepare_merge ||
420             !snap_src->store->type->commit_merge) {
421                 snap->ti->error = "Snapshot exception store does not "
422                                   "support snapshot-merge.";
423                 return -EINVAL;
424         }
425
426         return 1;
427 }
428
429 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
430 {
431         struct dm_snapshot *l;
432
433         /* Sort the list according to chunk size, largest-first smallest-last */
434         list_for_each_entry(l, &o->snapshots, list)
435                 if (l->store->chunk_size < s->store->chunk_size)
436                         break;
437         list_add_tail(&s->list, &l->list);
438 }
439
440 /*
441  * Make a note of the snapshot and its origin so we can look it
442  * up when the origin has a write on it.
443  *
444  * Also validate snapshot exception store handovers.
445  * On success, returns 1 if this registration is a handover destination,
446  * otherwise returns 0.
447  */
448 static int register_snapshot(struct dm_snapshot *snap)
449 {
450         struct origin *o, *new_o = NULL;
451         struct block_device *bdev = snap->origin->bdev;
452         int r = 0;
453
454         new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
455         if (!new_o)
456                 return -ENOMEM;
457
458         down_write(&_origins_lock);
459
460         r = __validate_exception_handover(snap);
461         if (r < 0) {
462                 kfree(new_o);
463                 goto out;
464         }
465
466         o = __lookup_origin(bdev);
467         if (o)
468                 kfree(new_o);
469         else {
470                 /* New origin */
471                 o = new_o;
472
473                 /* Initialise the struct */
474                 INIT_LIST_HEAD(&o->snapshots);
475                 o->bdev = bdev;
476
477                 __insert_origin(o);
478         }
479
480         __insert_snapshot(o, snap);
481
482 out:
483         up_write(&_origins_lock);
484
485         return r;
486 }
487
488 /*
489  * Move snapshot to correct place in list according to chunk size.
490  */
491 static void reregister_snapshot(struct dm_snapshot *s)
492 {
493         struct block_device *bdev = s->origin->bdev;
494
495         down_write(&_origins_lock);
496
497         list_del(&s->list);
498         __insert_snapshot(__lookup_origin(bdev), s);
499
500         up_write(&_origins_lock);
501 }
502
503 static void unregister_snapshot(struct dm_snapshot *s)
504 {
505         struct origin *o;
506
507         down_write(&_origins_lock);
508         o = __lookup_origin(s->origin->bdev);
509
510         list_del(&s->list);
511         if (o && list_empty(&o->snapshots)) {
512                 list_del(&o->hash_list);
513                 kfree(o);
514         }
515
516         up_write(&_origins_lock);
517 }
518
519 /*
520  * Implementation of the exception hash tables.
521  * The lowest hash_shift bits of the chunk number are ignored, allowing
522  * some consecutive chunks to be grouped together.
523  */
524 static int dm_exception_table_init(struct dm_exception_table *et,
525                                    uint32_t size, unsigned hash_shift)
526 {
527         unsigned int i;
528
529         et->hash_shift = hash_shift;
530         et->hash_mask = size - 1;
531         et->table = dm_vcalloc(size, sizeof(struct list_head));
532         if (!et->table)
533                 return -ENOMEM;
534
535         for (i = 0; i < size; i++)
536                 INIT_LIST_HEAD(et->table + i);
537
538         return 0;
539 }
540
541 static void dm_exception_table_exit(struct dm_exception_table *et,
542                                     struct kmem_cache *mem)
543 {
544         struct list_head *slot;
545         struct dm_exception *ex, *next;
546         int i, size;
547
548         size = et->hash_mask + 1;
549         for (i = 0; i < size; i++) {
550                 slot = et->table + i;
551
552                 list_for_each_entry_safe (ex, next, slot, hash_list)
553                         kmem_cache_free(mem, ex);
554         }
555
556         vfree(et->table);
557 }
558
559 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
560 {
561         return (chunk >> et->hash_shift) & et->hash_mask;
562 }
563
564 static void dm_remove_exception(struct dm_exception *e)
565 {
566         list_del(&e->hash_list);
567 }
568
569 /*
570  * Return the exception data for a sector, or NULL if not
571  * remapped.
572  */
573 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
574                                                 chunk_t chunk)
575 {
576         struct list_head *slot;
577         struct dm_exception *e;
578
579         slot = &et->table[exception_hash(et, chunk)];
580         list_for_each_entry (e, slot, hash_list)
581                 if (chunk >= e->old_chunk &&
582                     chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
583                         return e;
584
585         return NULL;
586 }
587
588 static struct dm_exception *alloc_completed_exception(void)
589 {
590         struct dm_exception *e;
591
592         e = kmem_cache_alloc(exception_cache, GFP_NOIO);
593         if (!e)
594                 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
595
596         return e;
597 }
598
599 static void free_completed_exception(struct dm_exception *e)
600 {
601         kmem_cache_free(exception_cache, e);
602 }
603
604 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
605 {
606         struct dm_snap_pending_exception *pe = mempool_alloc(s->pending_pool,
607                                                              GFP_NOIO);
608
609         atomic_inc(&s->pending_exceptions_count);
610         pe->snap = s;
611
612         return pe;
613 }
614
615 static void free_pending_exception(struct dm_snap_pending_exception *pe)
616 {
617         struct dm_snapshot *s = pe->snap;
618
619         mempool_free(pe, s->pending_pool);
620         smp_mb__before_atomic_dec();
621         atomic_dec(&s->pending_exceptions_count);
622 }
623
624 static void dm_insert_exception(struct dm_exception_table *eh,
625                                 struct dm_exception *new_e)
626 {
627         struct list_head *l;
628         struct dm_exception *e = NULL;
629
630         l = &eh->table[exception_hash(eh, new_e->old_chunk)];
631
632         /* Add immediately if this table doesn't support consecutive chunks */
633         if (!eh->hash_shift)
634                 goto out;
635
636         /* List is ordered by old_chunk */
637         list_for_each_entry_reverse(e, l, hash_list) {
638                 /* Insert after an existing chunk? */
639                 if (new_e->old_chunk == (e->old_chunk +
640                                          dm_consecutive_chunk_count(e) + 1) &&
641                     new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
642                                          dm_consecutive_chunk_count(e) + 1)) {
643                         dm_consecutive_chunk_count_inc(e);
644                         free_completed_exception(new_e);
645                         return;
646                 }
647
648                 /* Insert before an existing chunk? */
649                 if (new_e->old_chunk == (e->old_chunk - 1) &&
650                     new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
651                         dm_consecutive_chunk_count_inc(e);
652                         e->old_chunk--;
653                         e->new_chunk--;
654                         free_completed_exception(new_e);
655                         return;
656                 }
657
658                 if (new_e->old_chunk > e->old_chunk)
659                         break;
660         }
661
662 out:
663         list_add(&new_e->hash_list, e ? &e->hash_list : l);
664 }
665
666 /*
667  * Callback used by the exception stores to load exceptions when
668  * initialising.
669  */
670 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
671 {
672         struct dm_snapshot *s = context;
673         struct dm_exception *e;
674
675         e = alloc_completed_exception();
676         if (!e)
677                 return -ENOMEM;
678
679         e->old_chunk = old;
680
681         /* Consecutive_count is implicitly initialised to zero */
682         e->new_chunk = new;
683
684         dm_insert_exception(&s->complete, e);
685
686         return 0;
687 }
688
689 /*
690  * Return a minimum chunk size of all snapshots that have the specified origin.
691  * Return zero if the origin has no snapshots.
692  */
693 static uint32_t __minimum_chunk_size(struct origin *o)
694 {
695         struct dm_snapshot *snap;
696         unsigned chunk_size = 0;
697
698         if (o)
699                 list_for_each_entry(snap, &o->snapshots, list)
700                         chunk_size = min_not_zero(chunk_size,
701                                                   snap->store->chunk_size);
702
703         return (uint32_t) chunk_size;
704 }
705
706 /*
707  * Hard coded magic.
708  */
709 static int calc_max_buckets(void)
710 {
711         /* use a fixed size of 2MB */
712         unsigned long mem = 2 * 1024 * 1024;
713         mem /= sizeof(struct list_head);
714
715         return mem;
716 }
717
718 /*
719  * Allocate room for a suitable hash table.
720  */
721 static int init_hash_tables(struct dm_snapshot *s)
722 {
723         sector_t hash_size, cow_dev_size, origin_dev_size, max_buckets;
724
725         /*
726          * Calculate based on the size of the original volume or
727          * the COW volume...
728          */
729         cow_dev_size = get_dev_size(s->cow->bdev);
730         origin_dev_size = get_dev_size(s->origin->bdev);
731         max_buckets = calc_max_buckets();
732
733         hash_size = min(origin_dev_size, cow_dev_size) >> s->store->chunk_shift;
734         hash_size = min(hash_size, max_buckets);
735
736         if (hash_size < 64)
737                 hash_size = 64;
738         hash_size = rounddown_pow_of_two(hash_size);
739         if (dm_exception_table_init(&s->complete, hash_size,
740                                     DM_CHUNK_CONSECUTIVE_BITS))
741                 return -ENOMEM;
742
743         /*
744          * Allocate hash table for in-flight exceptions
745          * Make this smaller than the real hash table
746          */
747         hash_size >>= 3;
748         if (hash_size < 64)
749                 hash_size = 64;
750
751         if (dm_exception_table_init(&s->pending, hash_size, 0)) {
752                 dm_exception_table_exit(&s->complete, exception_cache);
753                 return -ENOMEM;
754         }
755
756         return 0;
757 }
758
759 static void merge_shutdown(struct dm_snapshot *s)
760 {
761         clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
762         smp_mb__after_clear_bit();
763         wake_up_bit(&s->state_bits, RUNNING_MERGE);
764 }
765
766 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
767 {
768         s->first_merging_chunk = 0;
769         s->num_merging_chunks = 0;
770
771         return bio_list_get(&s->bios_queued_during_merge);
772 }
773
774 /*
775  * Remove one chunk from the index of completed exceptions.
776  */
777 static int __remove_single_exception_chunk(struct dm_snapshot *s,
778                                            chunk_t old_chunk)
779 {
780         struct dm_exception *e;
781
782         e = dm_lookup_exception(&s->complete, old_chunk);
783         if (!e) {
784                 DMERR("Corruption detected: exception for block %llu is "
785                       "on disk but not in memory",
786                       (unsigned long long)old_chunk);
787                 return -EINVAL;
788         }
789
790         /*
791          * If this is the only chunk using this exception, remove exception.
792          */
793         if (!dm_consecutive_chunk_count(e)) {
794                 dm_remove_exception(e);
795                 free_completed_exception(e);
796                 return 0;
797         }
798
799         /*
800          * The chunk may be either at the beginning or the end of a
801          * group of consecutive chunks - never in the middle.  We are
802          * removing chunks in the opposite order to that in which they
803          * were added, so this should always be true.
804          * Decrement the consecutive chunk counter and adjust the
805          * starting point if necessary.
806          */
807         if (old_chunk == e->old_chunk) {
808                 e->old_chunk++;
809                 e->new_chunk++;
810         } else if (old_chunk != e->old_chunk +
811                    dm_consecutive_chunk_count(e)) {
812                 DMERR("Attempt to merge block %llu from the "
813                       "middle of a chunk range [%llu - %llu]",
814                       (unsigned long long)old_chunk,
815                       (unsigned long long)e->old_chunk,
816                       (unsigned long long)
817                       e->old_chunk + dm_consecutive_chunk_count(e));
818                 return -EINVAL;
819         }
820
821         dm_consecutive_chunk_count_dec(e);
822
823         return 0;
824 }
825
826 static void flush_bios(struct bio *bio);
827
828 static int remove_single_exception_chunk(struct dm_snapshot *s)
829 {
830         struct bio *b = NULL;
831         int r;
832         chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
833
834         down_write(&s->lock);
835
836         /*
837          * Process chunks (and associated exceptions) in reverse order
838          * so that dm_consecutive_chunk_count_dec() accounting works.
839          */
840         do {
841                 r = __remove_single_exception_chunk(s, old_chunk);
842                 if (r)
843                         goto out;
844         } while (old_chunk-- > s->first_merging_chunk);
845
846         b = __release_queued_bios_after_merge(s);
847
848 out:
849         up_write(&s->lock);
850         if (b)
851                 flush_bios(b);
852
853         return r;
854 }
855
856 static int origin_write_extent(struct dm_snapshot *merging_snap,
857                                sector_t sector, unsigned chunk_size);
858
859 static void merge_callback(int read_err, unsigned long write_err,
860                            void *context);
861
862 static uint64_t read_pending_exceptions_done_count(void)
863 {
864         uint64_t pending_exceptions_done;
865
866         spin_lock(&_pending_exceptions_done_spinlock);
867         pending_exceptions_done = _pending_exceptions_done_count;
868         spin_unlock(&_pending_exceptions_done_spinlock);
869
870         return pending_exceptions_done;
871 }
872
873 static void increment_pending_exceptions_done_count(void)
874 {
875         spin_lock(&_pending_exceptions_done_spinlock);
876         _pending_exceptions_done_count++;
877         spin_unlock(&_pending_exceptions_done_spinlock);
878
879         wake_up_all(&_pending_exceptions_done);
880 }
881
882 static void snapshot_merge_next_chunks(struct dm_snapshot *s)
883 {
884         int i, linear_chunks;
885         chunk_t old_chunk, new_chunk;
886         struct dm_io_region src, dest;
887         sector_t io_size;
888         uint64_t previous_count;
889
890         BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
891         if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
892                 goto shut;
893
894         /*
895          * valid flag never changes during merge, so no lock required.
896          */
897         if (!s->valid) {
898                 DMERR("Snapshot is invalid: can't merge");
899                 goto shut;
900         }
901
902         linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
903                                                       &new_chunk);
904         if (linear_chunks <= 0) {
905                 if (linear_chunks < 0) {
906                         DMERR("Read error in exception store: "
907                               "shutting down merge");
908                         down_write(&s->lock);
909                         s->merge_failed = 1;
910                         up_write(&s->lock);
911                 }
912                 goto shut;
913         }
914
915         /* Adjust old_chunk and new_chunk to reflect start of linear region */
916         old_chunk = old_chunk + 1 - linear_chunks;
917         new_chunk = new_chunk + 1 - linear_chunks;
918
919         /*
920          * Use one (potentially large) I/O to copy all 'linear_chunks'
921          * from the exception store to the origin
922          */
923         io_size = linear_chunks * s->store->chunk_size;
924
925         dest.bdev = s->origin->bdev;
926         dest.sector = chunk_to_sector(s->store, old_chunk);
927         dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
928
929         src.bdev = s->cow->bdev;
930         src.sector = chunk_to_sector(s->store, new_chunk);
931         src.count = dest.count;
932
933         /*
934          * Reallocate any exceptions needed in other snapshots then
935          * wait for the pending exceptions to complete.
936          * Each time any pending exception (globally on the system)
937          * completes we are woken and repeat the process to find out
938          * if we can proceed.  While this may not seem a particularly
939          * efficient algorithm, it is not expected to have any
940          * significant impact on performance.
941          */
942         previous_count = read_pending_exceptions_done_count();
943         while (origin_write_extent(s, dest.sector, io_size)) {
944                 wait_event(_pending_exceptions_done,
945                            (read_pending_exceptions_done_count() !=
946                             previous_count));
947                 /* Retry after the wait, until all exceptions are done. */
948                 previous_count = read_pending_exceptions_done_count();
949         }
950
951         down_write(&s->lock);
952         s->first_merging_chunk = old_chunk;
953         s->num_merging_chunks = linear_chunks;
954         up_write(&s->lock);
955
956         /* Wait until writes to all 'linear_chunks' drain */
957         for (i = 0; i < linear_chunks; i++)
958                 __check_for_conflicting_io(s, old_chunk + i);
959
960         dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
961         return;
962
963 shut:
964         merge_shutdown(s);
965 }
966
967 static void error_bios(struct bio *bio);
968
969 static void merge_callback(int read_err, unsigned long write_err, void *context)
970 {
971         struct dm_snapshot *s = context;
972         struct bio *b = NULL;
973
974         if (read_err || write_err) {
975                 if (read_err)
976                         DMERR("Read error: shutting down merge.");
977                 else
978                         DMERR("Write error: shutting down merge.");
979                 goto shut;
980         }
981
982         if (s->store->type->commit_merge(s->store,
983                                          s->num_merging_chunks) < 0) {
984                 DMERR("Write error in exception store: shutting down merge");
985                 goto shut;
986         }
987
988         if (remove_single_exception_chunk(s) < 0)
989                 goto shut;
990
991         snapshot_merge_next_chunks(s);
992
993         return;
994
995 shut:
996         down_write(&s->lock);
997         s->merge_failed = 1;
998         b = __release_queued_bios_after_merge(s);
999         up_write(&s->lock);
1000         error_bios(b);
1001
1002         merge_shutdown(s);
1003 }
1004
1005 static void start_merge(struct dm_snapshot *s)
1006 {
1007         if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1008                 snapshot_merge_next_chunks(s);
1009 }
1010
1011 static int wait_schedule(void *ptr)
1012 {
1013         schedule();
1014
1015         return 0;
1016 }
1017
1018 /*
1019  * Stop the merging process and wait until it finishes.
1020  */
1021 static void stop_merge(struct dm_snapshot *s)
1022 {
1023         set_bit(SHUTDOWN_MERGE, &s->state_bits);
1024         wait_on_bit(&s->state_bits, RUNNING_MERGE, wait_schedule,
1025                     TASK_UNINTERRUPTIBLE);
1026         clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1027 }
1028
1029 /*
1030  * Construct a snapshot mapping: <origin_dev> <COW-dev> <p/n> <chunk-size>
1031  */
1032 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1033 {
1034         struct dm_snapshot *s;
1035         int i;
1036         int r = -EINVAL;
1037         char *origin_path, *cow_path;
1038         unsigned args_used, num_flush_requests = 1;
1039         fmode_t origin_mode = FMODE_READ;
1040
1041         if (argc != 4) {
1042                 ti->error = "requires exactly 4 arguments";
1043                 r = -EINVAL;
1044                 goto bad;
1045         }
1046
1047         if (dm_target_is_snapshot_merge(ti)) {
1048                 num_flush_requests = 2;
1049                 origin_mode = FMODE_WRITE;
1050         }
1051
1052         s = kmalloc(sizeof(*s), GFP_KERNEL);
1053         if (!s) {
1054                 ti->error = "Cannot allocate private snapshot structure";
1055                 r = -ENOMEM;
1056                 goto bad;
1057         }
1058
1059         origin_path = argv[0];
1060         argv++;
1061         argc--;
1062
1063         r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1064         if (r) {
1065                 ti->error = "Cannot get origin device";
1066                 goto bad_origin;
1067         }
1068
1069         cow_path = argv[0];
1070         argv++;
1071         argc--;
1072
1073         r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1074         if (r) {
1075                 ti->error = "Cannot get COW device";
1076                 goto bad_cow;
1077         }
1078
1079         r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1080         if (r) {
1081                 ti->error = "Couldn't create exception store";
1082                 r = -EINVAL;
1083                 goto bad_store;
1084         }
1085
1086         argv += args_used;
1087         argc -= args_used;
1088
1089         s->ti = ti;
1090         s->valid = 1;
1091         s->active = 0;
1092         atomic_set(&s->pending_exceptions_count, 0);
1093         init_rwsem(&s->lock);
1094         INIT_LIST_HEAD(&s->list);
1095         spin_lock_init(&s->pe_lock);
1096         s->state_bits = 0;
1097         s->merge_failed = 0;
1098         s->first_merging_chunk = 0;
1099         s->num_merging_chunks = 0;
1100         bio_list_init(&s->bios_queued_during_merge);
1101
1102         /* Allocate hash table for COW data */
1103         if (init_hash_tables(s)) {
1104                 ti->error = "Unable to allocate hash table space";
1105                 r = -ENOMEM;
1106                 goto bad_hash_tables;
1107         }
1108
1109         s->kcopyd_client = dm_kcopyd_client_create();
1110         if (IS_ERR(s->kcopyd_client)) {
1111                 r = PTR_ERR(s->kcopyd_client);
1112                 ti->error = "Could not create kcopyd client";
1113                 goto bad_kcopyd;
1114         }
1115
1116         s->pending_pool = mempool_create_slab_pool(MIN_IOS, pending_cache);
1117         if (!s->pending_pool) {
1118                 ti->error = "Could not allocate mempool for pending exceptions";
1119                 goto bad_pending_pool;
1120         }
1121
1122         s->tracked_chunk_pool = mempool_create_slab_pool(MIN_IOS,
1123                                                          tracked_chunk_cache);
1124         if (!s->tracked_chunk_pool) {
1125                 ti->error = "Could not allocate tracked_chunk mempool for "
1126                             "tracking reads";
1127                 goto bad_tracked_chunk_pool;
1128         }
1129
1130         for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1131                 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1132
1133         spin_lock_init(&s->tracked_chunk_lock);
1134
1135         ti->private = s;
1136         ti->num_flush_requests = num_flush_requests;
1137
1138         /* Add snapshot to the list of snapshots for this origin */
1139         /* Exceptions aren't triggered till snapshot_resume() is called */
1140         r = register_snapshot(s);
1141         if (r == -ENOMEM) {
1142                 ti->error = "Snapshot origin struct allocation failed";
1143                 goto bad_load_and_register;
1144         } else if (r < 0) {
1145                 /* invalid handover, register_snapshot has set ti->error */
1146                 goto bad_load_and_register;
1147         }
1148
1149         /*
1150          * Metadata must only be loaded into one table at once, so skip this
1151          * if metadata will be handed over during resume.
1152          * Chunk size will be set during the handover - set it to zero to
1153          * ensure it's ignored.
1154          */
1155         if (r > 0) {
1156                 s->store->chunk_size = 0;
1157                 return 0;
1158         }
1159
1160         r = s->store->type->read_metadata(s->store, dm_add_exception,
1161                                           (void *)s);
1162         if (r < 0) {
1163                 ti->error = "Failed to read snapshot metadata";
1164                 goto bad_read_metadata;
1165         } else if (r > 0) {
1166                 s->valid = 0;
1167                 DMWARN("Snapshot is marked invalid.");
1168         }
1169
1170         if (!s->store->chunk_size) {
1171                 ti->error = "Chunk size not set";
1172                 goto bad_read_metadata;
1173         }
1174
1175         r = dm_set_target_max_io_len(ti, s->store->chunk_size);
1176         if (r)
1177                 goto bad_read_metadata;
1178
1179         return 0;
1180
1181 bad_read_metadata:
1182         unregister_snapshot(s);
1183
1184 bad_load_and_register:
1185         mempool_destroy(s->tracked_chunk_pool);
1186
1187 bad_tracked_chunk_pool:
1188         mempool_destroy(s->pending_pool);
1189
1190 bad_pending_pool:
1191         dm_kcopyd_client_destroy(s->kcopyd_client);
1192
1193 bad_kcopyd:
1194         dm_exception_table_exit(&s->pending, pending_cache);
1195         dm_exception_table_exit(&s->complete, exception_cache);
1196
1197 bad_hash_tables:
1198         dm_exception_store_destroy(s->store);
1199
1200 bad_store:
1201         dm_put_device(ti, s->cow);
1202
1203 bad_cow:
1204         dm_put_device(ti, s->origin);
1205
1206 bad_origin:
1207         kfree(s);
1208
1209 bad:
1210         return r;
1211 }
1212
1213 static void __free_exceptions(struct dm_snapshot *s)
1214 {
1215         dm_kcopyd_client_destroy(s->kcopyd_client);
1216         s->kcopyd_client = NULL;
1217
1218         dm_exception_table_exit(&s->pending, pending_cache);
1219         dm_exception_table_exit(&s->complete, exception_cache);
1220 }
1221
1222 static void __handover_exceptions(struct dm_snapshot *snap_src,
1223                                   struct dm_snapshot *snap_dest)
1224 {
1225         union {
1226                 struct dm_exception_table table_swap;
1227                 struct dm_exception_store *store_swap;
1228         } u;
1229
1230         /*
1231          * Swap all snapshot context information between the two instances.
1232          */
1233         u.table_swap = snap_dest->complete;
1234         snap_dest->complete = snap_src->complete;
1235         snap_src->complete = u.table_swap;
1236
1237         u.store_swap = snap_dest->store;
1238         snap_dest->store = snap_src->store;
1239         snap_src->store = u.store_swap;
1240
1241         snap_dest->store->snap = snap_dest;
1242         snap_src->store->snap = snap_src;
1243
1244         snap_dest->ti->max_io_len = snap_dest->store->chunk_size;
1245         snap_dest->valid = snap_src->valid;
1246
1247         /*
1248          * Set source invalid to ensure it receives no further I/O.
1249          */
1250         snap_src->valid = 0;
1251 }
1252
1253 static void snapshot_dtr(struct dm_target *ti)
1254 {
1255 #ifdef CONFIG_DM_DEBUG
1256         int i;
1257 #endif
1258         struct dm_snapshot *s = ti->private;
1259         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1260
1261         down_read(&_origins_lock);
1262         /* Check whether exception handover must be cancelled */
1263         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1264         if (snap_src && snap_dest && (s == snap_src)) {
1265                 down_write(&snap_dest->lock);
1266                 snap_dest->valid = 0;
1267                 up_write(&snap_dest->lock);
1268                 DMERR("Cancelling snapshot handover.");
1269         }
1270         up_read(&_origins_lock);
1271
1272         if (dm_target_is_snapshot_merge(ti))
1273                 stop_merge(s);
1274
1275         /* Prevent further origin writes from using this snapshot. */
1276         /* After this returns there can be no new kcopyd jobs. */
1277         unregister_snapshot(s);
1278
1279         while (atomic_read(&s->pending_exceptions_count))
1280                 msleep(1);
1281         /*
1282          * Ensure instructions in mempool_destroy aren't reordered
1283          * before atomic_read.
1284          */
1285         smp_mb();
1286
1287 #ifdef CONFIG_DM_DEBUG
1288         for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1289                 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1290 #endif
1291
1292         mempool_destroy(s->tracked_chunk_pool);
1293
1294         __free_exceptions(s);
1295
1296         mempool_destroy(s->pending_pool);
1297
1298         dm_exception_store_destroy(s->store);
1299
1300         dm_put_device(ti, s->cow);
1301
1302         dm_put_device(ti, s->origin);
1303
1304         kfree(s);
1305 }
1306
1307 /*
1308  * Flush a list of buffers.
1309  */
1310 static void flush_bios(struct bio *bio)
1311 {
1312         struct bio *n;
1313
1314         while (bio) {
1315                 n = bio->bi_next;
1316                 bio->bi_next = NULL;
1317                 generic_make_request(bio);
1318                 bio = n;
1319         }
1320 }
1321
1322 static int do_origin(struct dm_dev *origin, struct bio *bio);
1323
1324 /*
1325  * Flush a list of buffers.
1326  */
1327 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1328 {
1329         struct bio *n;
1330         int r;
1331
1332         while (bio) {
1333                 n = bio->bi_next;
1334                 bio->bi_next = NULL;
1335                 r = do_origin(s->origin, bio);
1336                 if (r == DM_MAPIO_REMAPPED)
1337                         generic_make_request(bio);
1338                 bio = n;
1339         }
1340 }
1341
1342 /*
1343  * Error a list of buffers.
1344  */
1345 static void error_bios(struct bio *bio)
1346 {
1347         struct bio *n;
1348
1349         while (bio) {
1350                 n = bio->bi_next;
1351                 bio->bi_next = NULL;
1352                 bio_io_error(bio);
1353                 bio = n;
1354         }
1355 }
1356
1357 static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1358 {
1359         if (!s->valid)
1360                 return;
1361
1362         if (err == -EIO)
1363                 DMERR("Invalidating snapshot: Error reading/writing.");
1364         else if (err == -ENOMEM)
1365                 DMERR("Invalidating snapshot: Unable to allocate exception.");
1366
1367         if (s->store->type->drop_snapshot)
1368                 s->store->type->drop_snapshot(s->store);
1369
1370         s->valid = 0;
1371
1372         dm_table_event(s->ti->table);
1373 }
1374
1375 static void pending_complete(struct dm_snap_pending_exception *pe, int success)
1376 {
1377         struct dm_exception *e;
1378         struct dm_snapshot *s = pe->snap;
1379         struct bio *origin_bios = NULL;
1380         struct bio *snapshot_bios = NULL;
1381         struct bio *full_bio = NULL;
1382         int error = 0;
1383
1384         if (!success) {
1385                 /* Read/write error - snapshot is unusable */
1386                 down_write(&s->lock);
1387                 __invalidate_snapshot(s, -EIO);
1388                 error = 1;
1389                 goto out;
1390         }
1391
1392         e = alloc_completed_exception();
1393         if (!e) {
1394                 down_write(&s->lock);
1395                 __invalidate_snapshot(s, -ENOMEM);
1396                 error = 1;
1397                 goto out;
1398         }
1399         *e = pe->e;
1400
1401         down_write(&s->lock);
1402         if (!s->valid) {
1403                 free_completed_exception(e);
1404                 error = 1;
1405                 goto out;
1406         }
1407
1408         /* Check for conflicting reads */
1409         __check_for_conflicting_io(s, pe->e.old_chunk);
1410
1411         /*
1412          * Add a proper exception, and remove the
1413          * in-flight exception from the list.
1414          */
1415         dm_insert_exception(&s->complete, e);
1416
1417 out:
1418         dm_remove_exception(&pe->e);
1419         snapshot_bios = bio_list_get(&pe->snapshot_bios);
1420         origin_bios = bio_list_get(&pe->origin_bios);
1421         full_bio = pe->full_bio;
1422         if (full_bio) {
1423                 full_bio->bi_end_io = pe->full_bio_end_io;
1424                 full_bio->bi_private = pe->full_bio_private;
1425         }
1426         free_pending_exception(pe);
1427
1428         increment_pending_exceptions_done_count();
1429
1430         up_write(&s->lock);
1431
1432         /* Submit any pending write bios */
1433         if (error) {
1434                 if (full_bio)
1435                         bio_io_error(full_bio);
1436                 error_bios(snapshot_bios);
1437         } else {
1438                 if (full_bio)
1439                         bio_endio(full_bio, 0);
1440                 flush_bios(snapshot_bios);
1441         }
1442
1443         retry_origin_bios(s, origin_bios);
1444 }
1445
1446 static void commit_callback(void *context, int success)
1447 {
1448         struct dm_snap_pending_exception *pe = context;
1449
1450         pending_complete(pe, success);
1451 }
1452
1453 /*
1454  * Called when the copy I/O has finished.  kcopyd actually runs
1455  * this code so don't block.
1456  */
1457 static void copy_callback(int read_err, unsigned long write_err, void *context)
1458 {
1459         struct dm_snap_pending_exception *pe = context;
1460         struct dm_snapshot *s = pe->snap;
1461
1462         if (read_err || write_err)
1463                 pending_complete(pe, 0);
1464
1465         else
1466                 /* Update the metadata if we are persistent */
1467                 s->store->type->commit_exception(s->store, &pe->e,
1468                                                  commit_callback, pe);
1469 }
1470
1471 /*
1472  * Dispatches the copy operation to kcopyd.
1473  */
1474 static void start_copy(struct dm_snap_pending_exception *pe)
1475 {
1476         struct dm_snapshot *s = pe->snap;
1477         struct dm_io_region src, dest;
1478         struct block_device *bdev = s->origin->bdev;
1479         sector_t dev_size;
1480
1481         dev_size = get_dev_size(bdev);
1482
1483         src.bdev = bdev;
1484         src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1485         src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1486
1487         dest.bdev = s->cow->bdev;
1488         dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1489         dest.count = src.count;
1490
1491         /* Hand over to kcopyd */
1492         dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1493 }
1494
1495 static void full_bio_end_io(struct bio *bio, int error)
1496 {
1497         void *callback_data = bio->bi_private;
1498
1499         dm_kcopyd_do_callback(callback_data, 0, error ? 1 : 0);
1500 }
1501
1502 static void start_full_bio(struct dm_snap_pending_exception *pe,
1503                            struct bio *bio)
1504 {
1505         struct dm_snapshot *s = pe->snap;
1506         void *callback_data;
1507
1508         pe->full_bio = bio;
1509         pe->full_bio_end_io = bio->bi_end_io;
1510         pe->full_bio_private = bio->bi_private;
1511
1512         callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1513                                                    copy_callback, pe);
1514
1515         bio->bi_end_io = full_bio_end_io;
1516         bio->bi_private = callback_data;
1517
1518         generic_make_request(bio);
1519 }
1520
1521 static struct dm_snap_pending_exception *
1522 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1523 {
1524         struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1525
1526         if (!e)
1527                 return NULL;
1528
1529         return container_of(e, struct dm_snap_pending_exception, e);
1530 }
1531
1532 /*
1533  * Looks to see if this snapshot already has a pending exception
1534  * for this chunk, otherwise it allocates a new one and inserts
1535  * it into the pending table.
1536  *
1537  * NOTE: a write lock must be held on snap->lock before calling
1538  * this.
1539  */
1540 static struct dm_snap_pending_exception *
1541 __find_pending_exception(struct dm_snapshot *s,
1542                          struct dm_snap_pending_exception *pe, chunk_t chunk)
1543 {
1544         struct dm_snap_pending_exception *pe2;
1545
1546         pe2 = __lookup_pending_exception(s, chunk);
1547         if (pe2) {
1548                 free_pending_exception(pe);
1549                 return pe2;
1550         }
1551
1552         pe->e.old_chunk = chunk;
1553         bio_list_init(&pe->origin_bios);
1554         bio_list_init(&pe->snapshot_bios);
1555         pe->started = 0;
1556         pe->full_bio = NULL;
1557
1558         if (s->store->type->prepare_exception(s->store, &pe->e)) {
1559                 free_pending_exception(pe);
1560                 return NULL;
1561         }
1562
1563         dm_insert_exception(&s->pending, &pe->e);
1564
1565         return pe;
1566 }
1567
1568 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1569                             struct bio *bio, chunk_t chunk)
1570 {
1571         bio->bi_bdev = s->cow->bdev;
1572         bio->bi_sector = chunk_to_sector(s->store,
1573                                          dm_chunk_number(e->new_chunk) +
1574                                          (chunk - e->old_chunk)) +
1575                                          (bio->bi_sector &
1576                                           s->store->chunk_mask);
1577 }
1578
1579 static int snapshot_map(struct dm_target *ti, struct bio *bio,
1580                         union map_info *map_context)
1581 {
1582         struct dm_exception *e;
1583         struct dm_snapshot *s = ti->private;
1584         int r = DM_MAPIO_REMAPPED;
1585         chunk_t chunk;
1586         struct dm_snap_pending_exception *pe = NULL;
1587
1588         if (bio->bi_rw & REQ_FLUSH) {
1589                 bio->bi_bdev = s->cow->bdev;
1590                 return DM_MAPIO_REMAPPED;
1591         }
1592
1593         chunk = sector_to_chunk(s->store, bio->bi_sector);
1594
1595         /* Full snapshots are not usable */
1596         /* To get here the table must be live so s->active is always set. */
1597         if (!s->valid)
1598                 return -EIO;
1599
1600         /* FIXME: should only take write lock if we need
1601          * to copy an exception */
1602         down_write(&s->lock);
1603
1604         if (!s->valid) {
1605                 r = -EIO;
1606                 goto out_unlock;
1607         }
1608
1609         /* If the block is already remapped - use that, else remap it */
1610         e = dm_lookup_exception(&s->complete, chunk);
1611         if (e) {
1612                 remap_exception(s, e, bio, chunk);
1613                 goto out_unlock;
1614         }
1615
1616         /*
1617          * Write to snapshot - higher level takes care of RW/RO
1618          * flags so we should only get this if we are
1619          * writeable.
1620          */
1621         if (bio_rw(bio) == WRITE) {
1622                 pe = __lookup_pending_exception(s, chunk);
1623                 if (!pe) {
1624                         up_write(&s->lock);
1625                         pe = alloc_pending_exception(s);
1626                         down_write(&s->lock);
1627
1628                         if (!s->valid) {
1629                                 free_pending_exception(pe);
1630                                 r = -EIO;
1631                                 goto out_unlock;
1632                         }
1633
1634                         e = dm_lookup_exception(&s->complete, chunk);
1635                         if (e) {
1636                                 free_pending_exception(pe);
1637                                 remap_exception(s, e, bio, chunk);
1638                                 goto out_unlock;
1639                         }
1640
1641                         pe = __find_pending_exception(s, pe, chunk);
1642                         if (!pe) {
1643                                 __invalidate_snapshot(s, -ENOMEM);
1644                                 r = -EIO;
1645                                 goto out_unlock;
1646                         }
1647                 }
1648
1649                 remap_exception(s, &pe->e, bio, chunk);
1650
1651                 r = DM_MAPIO_SUBMITTED;
1652
1653                 if (!pe->started &&
1654                     bio->bi_size == (s->store->chunk_size << SECTOR_SHIFT)) {
1655                         pe->started = 1;
1656                         up_write(&s->lock);
1657                         start_full_bio(pe, bio);
1658                         goto out;
1659                 }
1660
1661                 bio_list_add(&pe->snapshot_bios, bio);
1662
1663                 if (!pe->started) {
1664                         /* this is protected by snap->lock */
1665                         pe->started = 1;
1666                         up_write(&s->lock);
1667                         start_copy(pe);
1668                         goto out;
1669                 }
1670         } else {
1671                 bio->bi_bdev = s->origin->bdev;
1672                 map_context->ptr = track_chunk(s, chunk);
1673         }
1674
1675 out_unlock:
1676         up_write(&s->lock);
1677 out:
1678         return r;
1679 }
1680
1681 /*
1682  * A snapshot-merge target behaves like a combination of a snapshot
1683  * target and a snapshot-origin target.  It only generates new
1684  * exceptions in other snapshots and not in the one that is being
1685  * merged.
1686  *
1687  * For each chunk, if there is an existing exception, it is used to
1688  * redirect I/O to the cow device.  Otherwise I/O is sent to the origin,
1689  * which in turn might generate exceptions in other snapshots.
1690  * If merging is currently taking place on the chunk in question, the
1691  * I/O is deferred by adding it to s->bios_queued_during_merge.
1692  */
1693 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio,
1694                               union map_info *map_context)
1695 {
1696         struct dm_exception *e;
1697         struct dm_snapshot *s = ti->private;
1698         int r = DM_MAPIO_REMAPPED;
1699         chunk_t chunk;
1700
1701         if (bio->bi_rw & REQ_FLUSH) {
1702                 if (!map_context->target_request_nr)
1703                         bio->bi_bdev = s->origin->bdev;
1704                 else
1705                         bio->bi_bdev = s->cow->bdev;
1706                 map_context->ptr = NULL;
1707                 return DM_MAPIO_REMAPPED;
1708         }
1709
1710         chunk = sector_to_chunk(s->store, bio->bi_sector);
1711
1712         down_write(&s->lock);
1713
1714         /* Full merging snapshots are redirected to the origin */
1715         if (!s->valid)
1716                 goto redirect_to_origin;
1717
1718         /* If the block is already remapped - use that */
1719         e = dm_lookup_exception(&s->complete, chunk);
1720         if (e) {
1721                 /* Queue writes overlapping with chunks being merged */
1722                 if (bio_rw(bio) == WRITE &&
1723                     chunk >= s->first_merging_chunk &&
1724                     chunk < (s->first_merging_chunk +
1725                              s->num_merging_chunks)) {
1726                         bio->bi_bdev = s->origin->bdev;
1727                         bio_list_add(&s->bios_queued_during_merge, bio);
1728                         r = DM_MAPIO_SUBMITTED;
1729                         goto out_unlock;
1730                 }
1731
1732                 remap_exception(s, e, bio, chunk);
1733
1734                 if (bio_rw(bio) == WRITE)
1735                         map_context->ptr = track_chunk(s, chunk);
1736                 goto out_unlock;
1737         }
1738
1739 redirect_to_origin:
1740         bio->bi_bdev = s->origin->bdev;
1741
1742         if (bio_rw(bio) == WRITE) {
1743                 up_write(&s->lock);
1744                 return do_origin(s->origin, bio);
1745         }
1746
1747 out_unlock:
1748         up_write(&s->lock);
1749
1750         return r;
1751 }
1752
1753 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
1754                            int error, union map_info *map_context)
1755 {
1756         struct dm_snapshot *s = ti->private;
1757         struct dm_snap_tracked_chunk *c = map_context->ptr;
1758
1759         if (c)
1760                 stop_tracking_chunk(s, c);
1761
1762         return 0;
1763 }
1764
1765 static void snapshot_merge_presuspend(struct dm_target *ti)
1766 {
1767         struct dm_snapshot *s = ti->private;
1768
1769         stop_merge(s);
1770 }
1771
1772 static int snapshot_preresume(struct dm_target *ti)
1773 {
1774         int r = 0;
1775         struct dm_snapshot *s = ti->private;
1776         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1777
1778         down_read(&_origins_lock);
1779         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1780         if (snap_src && snap_dest) {
1781                 down_read(&snap_src->lock);
1782                 if (s == snap_src) {
1783                         DMERR("Unable to resume snapshot source until "
1784                               "handover completes.");
1785                         r = -EINVAL;
1786                 } else if (!dm_suspended(snap_src->ti)) {
1787                         DMERR("Unable to perform snapshot handover until "
1788                               "source is suspended.");
1789                         r = -EINVAL;
1790                 }
1791                 up_read(&snap_src->lock);
1792         }
1793         up_read(&_origins_lock);
1794
1795         return r;
1796 }
1797
1798 static void snapshot_resume(struct dm_target *ti)
1799 {
1800         struct dm_snapshot *s = ti->private;
1801         struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1802
1803         down_read(&_origins_lock);
1804         (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1805         if (snap_src && snap_dest) {
1806                 down_write(&snap_src->lock);
1807                 down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
1808                 __handover_exceptions(snap_src, snap_dest);
1809                 up_write(&snap_dest->lock);
1810                 up_write(&snap_src->lock);
1811         }
1812         up_read(&_origins_lock);
1813
1814         /* Now we have correct chunk size, reregister */
1815         reregister_snapshot(s);
1816
1817         down_write(&s->lock);
1818         s->active = 1;
1819         up_write(&s->lock);
1820 }
1821
1822 static uint32_t get_origin_minimum_chunksize(struct block_device *bdev)
1823 {
1824         uint32_t min_chunksize;
1825
1826         down_read(&_origins_lock);
1827         min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
1828         up_read(&_origins_lock);
1829
1830         return min_chunksize;
1831 }
1832
1833 static void snapshot_merge_resume(struct dm_target *ti)
1834 {
1835         struct dm_snapshot *s = ti->private;
1836
1837         /*
1838          * Handover exceptions from existing snapshot.
1839          */
1840         snapshot_resume(ti);
1841
1842         /*
1843          * snapshot-merge acts as an origin, so set ti->max_io_len
1844          */
1845         ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev);
1846
1847         start_merge(s);
1848 }
1849
1850 static int snapshot_status(struct dm_target *ti, status_type_t type,
1851                            unsigned status_flags, char *result, unsigned maxlen)
1852 {
1853         unsigned sz = 0;
1854         struct dm_snapshot *snap = ti->private;
1855
1856         switch (type) {
1857         case STATUSTYPE_INFO:
1858
1859                 down_write(&snap->lock);
1860
1861                 if (!snap->valid)
1862                         DMEMIT("Invalid");
1863                 else if (snap->merge_failed)
1864                         DMEMIT("Merge failed");
1865                 else {
1866                         if (snap->store->type->usage) {
1867                                 sector_t total_sectors, sectors_allocated,
1868                                          metadata_sectors;
1869                                 snap->store->type->usage(snap->store,
1870                                                          &total_sectors,
1871                                                          &sectors_allocated,
1872                                                          &metadata_sectors);
1873                                 DMEMIT("%llu/%llu %llu",
1874                                        (unsigned long long)sectors_allocated,
1875                                        (unsigned long long)total_sectors,
1876                                        (unsigned long long)metadata_sectors);
1877                         }
1878                         else
1879                                 DMEMIT("Unknown");
1880                 }
1881
1882                 up_write(&snap->lock);
1883
1884                 break;
1885
1886         case STATUSTYPE_TABLE:
1887                 /*
1888                  * kdevname returns a static pointer so we need
1889                  * to make private copies if the output is to
1890                  * make sense.
1891                  */
1892                 DMEMIT("%s %s", snap->origin->name, snap->cow->name);
1893                 snap->store->type->status(snap->store, type, result + sz,
1894                                           maxlen - sz);
1895                 break;
1896         }
1897
1898         return 0;
1899 }
1900
1901 static int snapshot_iterate_devices(struct dm_target *ti,
1902                                     iterate_devices_callout_fn fn, void *data)
1903 {
1904         struct dm_snapshot *snap = ti->private;
1905         int r;
1906
1907         r = fn(ti, snap->origin, 0, ti->len, data);
1908
1909         if (!r)
1910                 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
1911
1912         return r;
1913 }
1914
1915
1916 /*-----------------------------------------------------------------
1917  * Origin methods
1918  *---------------------------------------------------------------*/
1919
1920 /*
1921  * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
1922  * supplied bio was ignored.  The caller may submit it immediately.
1923  * (No remapping actually occurs as the origin is always a direct linear
1924  * map.)
1925  *
1926  * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
1927  * and any supplied bio is added to a list to be submitted once all
1928  * the necessary exceptions exist.
1929  */
1930 static int __origin_write(struct list_head *snapshots, sector_t sector,
1931                           struct bio *bio)
1932 {
1933         int r = DM_MAPIO_REMAPPED;
1934         struct dm_snapshot *snap;
1935         struct dm_exception *e;
1936         struct dm_snap_pending_exception *pe;
1937         struct dm_snap_pending_exception *pe_to_start_now = NULL;
1938         struct dm_snap_pending_exception *pe_to_start_last = NULL;
1939         chunk_t chunk;
1940
1941         /* Do all the snapshots on this origin */
1942         list_for_each_entry (snap, snapshots, list) {
1943                 /*
1944                  * Don't make new exceptions in a merging snapshot
1945                  * because it has effectively been deleted
1946                  */
1947                 if (dm_target_is_snapshot_merge(snap->ti))
1948                         continue;
1949
1950                 down_write(&snap->lock);
1951
1952                 /* Only deal with valid and active snapshots */
1953                 if (!snap->valid || !snap->active)
1954                         goto next_snapshot;
1955
1956                 /* Nothing to do if writing beyond end of snapshot */
1957                 if (sector >= dm_table_get_size(snap->ti->table))
1958                         goto next_snapshot;
1959
1960                 /*
1961                  * Remember, different snapshots can have
1962                  * different chunk sizes.
1963                  */
1964                 chunk = sector_to_chunk(snap->store, sector);
1965
1966                 /*
1967                  * Check exception table to see if block
1968                  * is already remapped in this snapshot
1969                  * and trigger an exception if not.
1970                  */
1971                 e = dm_lookup_exception(&snap->complete, chunk);
1972                 if (e)
1973                         goto next_snapshot;
1974
1975                 pe = __lookup_pending_exception(snap, chunk);
1976                 if (!pe) {
1977                         up_write(&snap->lock);
1978                         pe = alloc_pending_exception(snap);
1979                         down_write(&snap->lock);
1980
1981                         if (!snap->valid) {
1982                                 free_pending_exception(pe);
1983                                 goto next_snapshot;
1984                         }
1985
1986                         e = dm_lookup_exception(&snap->complete, chunk);
1987                         if (e) {
1988                                 free_pending_exception(pe);
1989                                 goto next_snapshot;
1990                         }
1991
1992                         pe = __find_pending_exception(snap, pe, chunk);
1993                         if (!pe) {
1994                                 __invalidate_snapshot(snap, -ENOMEM);
1995                                 goto next_snapshot;
1996                         }
1997                 }
1998
1999                 r = DM_MAPIO_SUBMITTED;
2000
2001                 /*
2002                  * If an origin bio was supplied, queue it to wait for the
2003                  * completion of this exception, and start this one last,
2004                  * at the end of the function.
2005                  */
2006                 if (bio) {
2007                         bio_list_add(&pe->origin_bios, bio);
2008                         bio = NULL;
2009
2010                         if (!pe->started) {
2011                                 pe->started = 1;
2012                                 pe_to_start_last = pe;
2013                         }
2014                 }
2015
2016                 if (!pe->started) {
2017                         pe->started = 1;
2018                         pe_to_start_now = pe;
2019                 }
2020
2021 next_snapshot:
2022                 up_write(&snap->lock);
2023
2024                 if (pe_to_start_now) {
2025                         start_copy(pe_to_start_now);
2026                         pe_to_start_now = NULL;
2027                 }
2028         }
2029
2030         /*
2031          * Submit the exception against which the bio is queued last,
2032          * to give the other exceptions a head start.
2033          */
2034         if (pe_to_start_last)
2035                 start_copy(pe_to_start_last);
2036
2037         return r;
2038 }
2039
2040 /*
2041  * Called on a write from the origin driver.
2042  */
2043 static int do_origin(struct dm_dev *origin, struct bio *bio)
2044 {
2045         struct origin *o;
2046         int r = DM_MAPIO_REMAPPED;
2047
2048         down_read(&_origins_lock);
2049         o = __lookup_origin(origin->bdev);
2050         if (o)
2051                 r = __origin_write(&o->snapshots, bio->bi_sector, bio);
2052         up_read(&_origins_lock);
2053
2054         return r;
2055 }
2056
2057 /*
2058  * Trigger exceptions in all non-merging snapshots.
2059  *
2060  * The chunk size of the merging snapshot may be larger than the chunk
2061  * size of some other snapshot so we may need to reallocate multiple
2062  * chunks in other snapshots.
2063  *
2064  * We scan all the overlapping exceptions in the other snapshots.
2065  * Returns 1 if anything was reallocated and must be waited for,
2066  * otherwise returns 0.
2067  *
2068  * size must be a multiple of merging_snap's chunk_size.
2069  */
2070 static int origin_write_extent(struct dm_snapshot *merging_snap,
2071                                sector_t sector, unsigned size)
2072 {
2073         int must_wait = 0;
2074         sector_t n;
2075         struct origin *o;
2076
2077         /*
2078          * The origin's __minimum_chunk_size() got stored in max_io_len
2079          * by snapshot_merge_resume().
2080          */
2081         down_read(&_origins_lock);
2082         o = __lookup_origin(merging_snap->origin->bdev);
2083         for (n = 0; n < size; n += merging_snap->ti->max_io_len)
2084                 if (__origin_write(&o->snapshots, sector + n, NULL) ==
2085                     DM_MAPIO_SUBMITTED)
2086                         must_wait = 1;
2087         up_read(&_origins_lock);
2088
2089         return must_wait;
2090 }
2091
2092 /*
2093  * Origin: maps a linear range of a device, with hooks for snapshotting.
2094  */
2095
2096 /*
2097  * Construct an origin mapping: <dev_path>
2098  * The context for an origin is merely a 'struct dm_dev *'
2099  * pointing to the real device.
2100  */
2101 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2102 {
2103         int r;
2104         struct dm_dev *dev;
2105
2106         if (argc != 1) {
2107                 ti->error = "origin: incorrect number of arguments";
2108                 return -EINVAL;
2109         }
2110
2111         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &dev);
2112         if (r) {
2113                 ti->error = "Cannot get target device";
2114                 return r;
2115         }
2116
2117         ti->private = dev;
2118         ti->num_flush_requests = 1;
2119
2120         return 0;
2121 }
2122
2123 static void origin_dtr(struct dm_target *ti)
2124 {
2125         struct dm_dev *dev = ti->private;
2126         dm_put_device(ti, dev);
2127 }
2128
2129 static int origin_map(struct dm_target *ti, struct bio *bio,
2130                       union map_info *map_context)
2131 {
2132         struct dm_dev *dev = ti->private;
2133         bio->bi_bdev = dev->bdev;
2134
2135         if (bio->bi_rw & REQ_FLUSH)
2136                 return DM_MAPIO_REMAPPED;
2137
2138         /* Only tell snapshots if this is a write */
2139         return (bio_rw(bio) == WRITE) ? do_origin(dev, bio) : DM_MAPIO_REMAPPED;
2140 }
2141
2142 /*
2143  * Set the target "max_io_len" field to the minimum of all the snapshots'
2144  * chunk sizes.
2145  */
2146 static void origin_resume(struct dm_target *ti)
2147 {
2148         struct dm_dev *dev = ti->private;
2149
2150         ti->max_io_len = get_origin_minimum_chunksize(dev->bdev);
2151 }
2152
2153 static int origin_status(struct dm_target *ti, status_type_t type,
2154                          unsigned status_flags, char *result, unsigned maxlen)
2155 {
2156         struct dm_dev *dev = ti->private;
2157
2158         switch (type) {
2159         case STATUSTYPE_INFO:
2160                 result[0] = '\0';
2161                 break;
2162
2163         case STATUSTYPE_TABLE:
2164                 snprintf(result, maxlen, "%s", dev->name);
2165                 break;
2166         }
2167
2168         return 0;
2169 }
2170
2171 static int origin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2172                         struct bio_vec *biovec, int max_size)
2173 {
2174         struct dm_dev *dev = ti->private;
2175         struct request_queue *q = bdev_get_queue(dev->bdev);
2176
2177         if (!q->merge_bvec_fn)
2178                 return max_size;
2179
2180         bvm->bi_bdev = dev->bdev;
2181
2182         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2183 }
2184
2185 static int origin_iterate_devices(struct dm_target *ti,
2186                                   iterate_devices_callout_fn fn, void *data)
2187 {
2188         struct dm_dev *dev = ti->private;
2189
2190         return fn(ti, dev, 0, ti->len, data);
2191 }
2192
2193 static struct target_type origin_target = {
2194         .name    = "snapshot-origin",
2195         .version = {1, 7, 1},
2196         .module  = THIS_MODULE,
2197         .ctr     = origin_ctr,
2198         .dtr     = origin_dtr,
2199         .map     = origin_map,
2200         .resume  = origin_resume,
2201         .status  = origin_status,
2202         .merge   = origin_merge,
2203         .iterate_devices = origin_iterate_devices,
2204 };
2205
2206 static struct target_type snapshot_target = {
2207         .name    = "snapshot",
2208         .version = {1, 10, 0},
2209         .module  = THIS_MODULE,
2210         .ctr     = snapshot_ctr,
2211         .dtr     = snapshot_dtr,
2212         .map     = snapshot_map,
2213         .end_io  = snapshot_end_io,
2214         .preresume  = snapshot_preresume,
2215         .resume  = snapshot_resume,
2216         .status  = snapshot_status,
2217         .iterate_devices = snapshot_iterate_devices,
2218 };
2219
2220 static struct target_type merge_target = {
2221         .name    = dm_snapshot_merge_target_name,
2222         .version = {1, 1, 0},
2223         .module  = THIS_MODULE,
2224         .ctr     = snapshot_ctr,
2225         .dtr     = snapshot_dtr,
2226         .map     = snapshot_merge_map,
2227         .end_io  = snapshot_end_io,
2228         .presuspend = snapshot_merge_presuspend,
2229         .preresume  = snapshot_preresume,
2230         .resume  = snapshot_merge_resume,
2231         .status  = snapshot_status,
2232         .iterate_devices = snapshot_iterate_devices,
2233 };
2234
2235 static int __init dm_snapshot_init(void)
2236 {
2237         int r;
2238
2239         r = dm_exception_store_init();
2240         if (r) {
2241                 DMERR("Failed to initialize exception stores");
2242                 return r;
2243         }
2244
2245         r = dm_register_target(&snapshot_target);
2246         if (r < 0) {
2247                 DMERR("snapshot target register failed %d", r);
2248                 goto bad_register_snapshot_target;
2249         }
2250
2251         r = dm_register_target(&origin_target);
2252         if (r < 0) {
2253                 DMERR("Origin target register failed %d", r);
2254                 goto bad_register_origin_target;
2255         }
2256
2257         r = dm_register_target(&merge_target);
2258         if (r < 0) {
2259                 DMERR("Merge target register failed %d", r);
2260                 goto bad_register_merge_target;
2261         }
2262
2263         r = init_origin_hash();
2264         if (r) {
2265                 DMERR("init_origin_hash failed.");
2266                 goto bad_origin_hash;
2267         }
2268
2269         exception_cache = KMEM_CACHE(dm_exception, 0);
2270         if (!exception_cache) {
2271                 DMERR("Couldn't create exception cache.");
2272                 r = -ENOMEM;
2273                 goto bad_exception_cache;
2274         }
2275
2276         pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2277         if (!pending_cache) {
2278                 DMERR("Couldn't create pending cache.");
2279                 r = -ENOMEM;
2280                 goto bad_pending_cache;
2281         }
2282
2283         tracked_chunk_cache = KMEM_CACHE(dm_snap_tracked_chunk, 0);
2284         if (!tracked_chunk_cache) {
2285                 DMERR("Couldn't create cache to track chunks in use.");
2286                 r = -ENOMEM;
2287                 goto bad_tracked_chunk_cache;
2288         }
2289
2290         return 0;
2291
2292 bad_tracked_chunk_cache:
2293         kmem_cache_destroy(pending_cache);
2294 bad_pending_cache:
2295         kmem_cache_destroy(exception_cache);
2296 bad_exception_cache:
2297         exit_origin_hash();
2298 bad_origin_hash:
2299         dm_unregister_target(&merge_target);
2300 bad_register_merge_target:
2301         dm_unregister_target(&origin_target);
2302 bad_register_origin_target:
2303         dm_unregister_target(&snapshot_target);
2304 bad_register_snapshot_target:
2305         dm_exception_store_exit();
2306
2307         return r;
2308 }
2309
2310 static void __exit dm_snapshot_exit(void)
2311 {
2312         dm_unregister_target(&snapshot_target);
2313         dm_unregister_target(&origin_target);
2314         dm_unregister_target(&merge_target);
2315
2316         exit_origin_hash();
2317         kmem_cache_destroy(pending_cache);
2318         kmem_cache_destroy(exception_cache);
2319         kmem_cache_destroy(tracked_chunk_cache);
2320
2321         dm_exception_store_exit();
2322 }
2323
2324 /* Module hooks */
2325 module_init(dm_snapshot_init);
2326 module_exit(dm_snapshot_exit);
2327
2328 MODULE_DESCRIPTION(DM_NAME " snapshot target");
2329 MODULE_AUTHOR("Joe Thornber");
2330 MODULE_LICENSE("GPL");