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