Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / drivers / md / dm-snap.c
blob216035be566185e3108e562d58389fa2409b7df8
1 /*
2 * dm-snapshot.c
4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
6 * This file is released under the GPL.
7 */
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>
23 #include "dm.h"
25 #include "dm-exception-store.h"
27 #define DM_MSG_PREFIX "snapshots"
29 static const char dm_snapshot_merge_target_name[] = "snapshot-merge";
31 #define dm_target_is_snapshot_merge(ti) \
32 ((ti)->type->name == dm_snapshot_merge_target_name)
35 * The size of the mempool used to track chunks in use.
37 #define MIN_IOS 256
39 #define DM_TRACKED_CHUNK_HASH_SIZE 16
40 #define DM_TRACKED_CHUNK_HASH(x) ((unsigned long)(x) & \
41 (DM_TRACKED_CHUNK_HASH_SIZE - 1))
43 struct dm_exception_table {
44 uint32_t hash_mask;
45 unsigned hash_shift;
46 struct list_head *table;
49 struct dm_snapshot {
50 struct mutex lock;
52 struct dm_dev *origin;
53 struct dm_dev *cow;
55 struct dm_target *ti;
57 /* List of snapshots per Origin */
58 struct list_head list;
61 * You can't use a snapshot if this is 0 (e.g. if full).
62 * A snapshot-merge target never clears this.
64 int valid;
67 * The snapshot overflowed because of a write to the snapshot device.
68 * We don't have to invalidate the snapshot in this case, but we need
69 * to prevent further writes.
71 int snapshot_overflowed;
73 /* Origin writes don't trigger exceptions until this is set */
74 int active;
76 atomic_t pending_exceptions_count;
78 /* Protected by "lock" */
79 sector_t exception_start_sequence;
81 /* Protected by kcopyd single-threaded callback */
82 sector_t exception_complete_sequence;
85 * A list of pending exceptions that completed out of order.
86 * Protected by kcopyd single-threaded callback.
88 struct list_head out_of_order_list;
90 mempool_t *pending_pool;
92 struct dm_exception_table pending;
93 struct dm_exception_table complete;
96 * pe_lock protects all pending_exception operations and access
97 * as well as the snapshot_bios list.
99 spinlock_t pe_lock;
101 /* Chunks with outstanding reads */
102 spinlock_t tracked_chunk_lock;
103 struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE];
105 /* The on disk metadata handler */
106 struct dm_exception_store *store;
108 struct dm_kcopyd_client *kcopyd_client;
110 /* Wait for events based on state_bits */
111 unsigned long state_bits;
113 /* Range of chunks currently being merged. */
114 chunk_t first_merging_chunk;
115 int num_merging_chunks;
118 * The merge operation failed if this flag is set.
119 * Failure modes are handled as follows:
120 * - I/O error reading the header
121 * => don't load the target; abort.
122 * - Header does not have "valid" flag set
123 * => use the origin; forget about the snapshot.
124 * - I/O error when reading exceptions
125 * => don't load the target; abort.
126 * (We can't use the intermediate origin state.)
127 * - I/O error while merging
128 * => stop merging; set merge_failed; process I/O normally.
130 int merge_failed;
133 * Incoming bios that overlap with chunks being merged must wait
134 * for them to be committed.
136 struct bio_list bios_queued_during_merge;
140 * state_bits:
141 * RUNNING_MERGE - Merge operation is in progress.
142 * SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
143 * cleared afterwards.
145 #define RUNNING_MERGE 0
146 #define SHUTDOWN_MERGE 1
148 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
149 "A percentage of time allocated for copy on write");
151 struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
153 return s->origin;
155 EXPORT_SYMBOL(dm_snap_origin);
157 struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
159 return s->cow;
161 EXPORT_SYMBOL(dm_snap_cow);
163 static sector_t chunk_to_sector(struct dm_exception_store *store,
164 chunk_t chunk)
166 return chunk << store->chunk_shift;
169 static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
172 * There is only ever one instance of a particular block
173 * device so we can compare pointers safely.
175 return lhs == rhs;
178 struct dm_snap_pending_exception {
179 struct dm_exception e;
182 * Origin buffers waiting for this to complete are held
183 * in a bio list
185 struct bio_list origin_bios;
186 struct bio_list snapshot_bios;
188 /* Pointer back to snapshot context */
189 struct dm_snapshot *snap;
192 * 1 indicates the exception has already been sent to
193 * kcopyd.
195 int started;
197 /* There was copying error. */
198 int copy_error;
200 /* A sequence number, it is used for in-order completion. */
201 sector_t exception_sequence;
203 struct list_head out_of_order_entry;
206 * For writing a complete chunk, bypassing the copy.
208 struct bio *full_bio;
209 bio_end_io_t *full_bio_end_io;
213 * Hash table mapping origin volumes to lists of snapshots and
214 * a lock to protect it
216 static struct kmem_cache *exception_cache;
217 static struct kmem_cache *pending_cache;
219 struct dm_snap_tracked_chunk {
220 struct hlist_node node;
221 chunk_t chunk;
224 static void init_tracked_chunk(struct bio *bio)
226 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
227 INIT_HLIST_NODE(&c->node);
230 static bool is_bio_tracked(struct bio *bio)
232 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
233 return !hlist_unhashed(&c->node);
236 static void track_chunk(struct dm_snapshot *s, struct bio *bio, chunk_t chunk)
238 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
240 c->chunk = chunk;
242 spin_lock_irq(&s->tracked_chunk_lock);
243 hlist_add_head(&c->node,
244 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
245 spin_unlock_irq(&s->tracked_chunk_lock);
248 static void stop_tracking_chunk(struct dm_snapshot *s, struct bio *bio)
250 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
251 unsigned long flags;
253 spin_lock_irqsave(&s->tracked_chunk_lock, flags);
254 hlist_del(&c->node);
255 spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
258 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
260 struct dm_snap_tracked_chunk *c;
261 int found = 0;
263 spin_lock_irq(&s->tracked_chunk_lock);
265 hlist_for_each_entry(c,
266 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
267 if (c->chunk == chunk) {
268 found = 1;
269 break;
273 spin_unlock_irq(&s->tracked_chunk_lock);
275 return found;
279 * This conflicting I/O is extremely improbable in the caller,
280 * so msleep(1) is sufficient and there is no need for a wait queue.
282 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
284 while (__chunk_is_tracked(s, chunk))
285 msleep(1);
289 * One of these per registered origin, held in the snapshot_origins hash
291 struct origin {
292 /* The origin device */
293 struct block_device *bdev;
295 struct list_head hash_list;
297 /* List of snapshots for this origin */
298 struct list_head snapshots;
302 * This structure is allocated for each origin target
304 struct dm_origin {
305 struct dm_dev *dev;
306 struct dm_target *ti;
307 unsigned split_boundary;
308 struct list_head hash_list;
312 * Size of the hash table for origin volumes. If we make this
313 * the size of the minors list then it should be nearly perfect
315 #define ORIGIN_HASH_SIZE 256
316 #define ORIGIN_MASK 0xFF
317 static struct list_head *_origins;
318 static struct list_head *_dm_origins;
319 static struct rw_semaphore _origins_lock;
321 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
322 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
323 static uint64_t _pending_exceptions_done_count;
325 static int init_origin_hash(void)
327 int i;
329 _origins = kmalloc(ORIGIN_HASH_SIZE * sizeof(struct list_head),
330 GFP_KERNEL);
331 if (!_origins) {
332 DMERR("unable to allocate memory for _origins");
333 return -ENOMEM;
335 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
336 INIT_LIST_HEAD(_origins + i);
338 _dm_origins = kmalloc(ORIGIN_HASH_SIZE * sizeof(struct list_head),
339 GFP_KERNEL);
340 if (!_dm_origins) {
341 DMERR("unable to allocate memory for _dm_origins");
342 kfree(_origins);
343 return -ENOMEM;
345 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
346 INIT_LIST_HEAD(_dm_origins + i);
348 init_rwsem(&_origins_lock);
350 return 0;
353 static void exit_origin_hash(void)
355 kfree(_origins);
356 kfree(_dm_origins);
359 static unsigned origin_hash(struct block_device *bdev)
361 return bdev->bd_dev & ORIGIN_MASK;
364 static struct origin *__lookup_origin(struct block_device *origin)
366 struct list_head *ol;
367 struct origin *o;
369 ol = &_origins[origin_hash(origin)];
370 list_for_each_entry (o, ol, hash_list)
371 if (bdev_equal(o->bdev, origin))
372 return o;
374 return NULL;
377 static void __insert_origin(struct origin *o)
379 struct list_head *sl = &_origins[origin_hash(o->bdev)];
380 list_add_tail(&o->hash_list, sl);
383 static struct dm_origin *__lookup_dm_origin(struct block_device *origin)
385 struct list_head *ol;
386 struct dm_origin *o;
388 ol = &_dm_origins[origin_hash(origin)];
389 list_for_each_entry (o, ol, hash_list)
390 if (bdev_equal(o->dev->bdev, origin))
391 return o;
393 return NULL;
396 static void __insert_dm_origin(struct dm_origin *o)
398 struct list_head *sl = &_dm_origins[origin_hash(o->dev->bdev)];
399 list_add_tail(&o->hash_list, sl);
402 static void __remove_dm_origin(struct dm_origin *o)
404 list_del(&o->hash_list);
408 * _origins_lock must be held when calling this function.
409 * Returns number of snapshots registered using the supplied cow device, plus:
410 * snap_src - a snapshot suitable for use as a source of exception handover
411 * snap_dest - a snapshot capable of receiving exception handover.
412 * snap_merge - an existing snapshot-merge target linked to the same origin.
413 * There can be at most one snapshot-merge target. The parameter is optional.
415 * Possible return values and states of snap_src and snap_dest.
416 * 0: NULL, NULL - first new snapshot
417 * 1: snap_src, NULL - normal snapshot
418 * 2: snap_src, snap_dest - waiting for handover
419 * 2: snap_src, NULL - handed over, waiting for old to be deleted
420 * 1: NULL, snap_dest - source got destroyed without handover
422 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
423 struct dm_snapshot **snap_src,
424 struct dm_snapshot **snap_dest,
425 struct dm_snapshot **snap_merge)
427 struct dm_snapshot *s;
428 struct origin *o;
429 int count = 0;
430 int active;
432 o = __lookup_origin(snap->origin->bdev);
433 if (!o)
434 goto out;
436 list_for_each_entry(s, &o->snapshots, list) {
437 if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
438 *snap_merge = s;
439 if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
440 continue;
442 mutex_lock(&s->lock);
443 active = s->active;
444 mutex_unlock(&s->lock);
446 if (active) {
447 if (snap_src)
448 *snap_src = s;
449 } else if (snap_dest)
450 *snap_dest = s;
452 count++;
455 out:
456 return count;
460 * On success, returns 1 if this snapshot is a handover destination,
461 * otherwise returns 0.
463 static int __validate_exception_handover(struct dm_snapshot *snap)
465 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
466 struct dm_snapshot *snap_merge = NULL;
468 /* Does snapshot need exceptions handed over to it? */
469 if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
470 &snap_merge) == 2) ||
471 snap_dest) {
472 snap->ti->error = "Snapshot cow pairing for exception "
473 "table handover failed";
474 return -EINVAL;
478 * If no snap_src was found, snap cannot become a handover
479 * destination.
481 if (!snap_src)
482 return 0;
485 * Non-snapshot-merge handover?
487 if (!dm_target_is_snapshot_merge(snap->ti))
488 return 1;
491 * Do not allow more than one merging snapshot.
493 if (snap_merge) {
494 snap->ti->error = "A snapshot is already merging.";
495 return -EINVAL;
498 if (!snap_src->store->type->prepare_merge ||
499 !snap_src->store->type->commit_merge) {
500 snap->ti->error = "Snapshot exception store does not "
501 "support snapshot-merge.";
502 return -EINVAL;
505 return 1;
508 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
510 struct dm_snapshot *l;
512 /* Sort the list according to chunk size, largest-first smallest-last */
513 list_for_each_entry(l, &o->snapshots, list)
514 if (l->store->chunk_size < s->store->chunk_size)
515 break;
516 list_add_tail(&s->list, &l->list);
520 * Make a note of the snapshot and its origin so we can look it
521 * up when the origin has a write on it.
523 * Also validate snapshot exception store handovers.
524 * On success, returns 1 if this registration is a handover destination,
525 * otherwise returns 0.
527 static int register_snapshot(struct dm_snapshot *snap)
529 struct origin *o, *new_o = NULL;
530 struct block_device *bdev = snap->origin->bdev;
531 int r = 0;
533 new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
534 if (!new_o)
535 return -ENOMEM;
537 down_write(&_origins_lock);
539 r = __validate_exception_handover(snap);
540 if (r < 0) {
541 kfree(new_o);
542 goto out;
545 o = __lookup_origin(bdev);
546 if (o)
547 kfree(new_o);
548 else {
549 /* New origin */
550 o = new_o;
552 /* Initialise the struct */
553 INIT_LIST_HEAD(&o->snapshots);
554 o->bdev = bdev;
556 __insert_origin(o);
559 __insert_snapshot(o, snap);
561 out:
562 up_write(&_origins_lock);
564 return r;
568 * Move snapshot to correct place in list according to chunk size.
570 static void reregister_snapshot(struct dm_snapshot *s)
572 struct block_device *bdev = s->origin->bdev;
574 down_write(&_origins_lock);
576 list_del(&s->list);
577 __insert_snapshot(__lookup_origin(bdev), s);
579 up_write(&_origins_lock);
582 static void unregister_snapshot(struct dm_snapshot *s)
584 struct origin *o;
586 down_write(&_origins_lock);
587 o = __lookup_origin(s->origin->bdev);
589 list_del(&s->list);
590 if (o && list_empty(&o->snapshots)) {
591 list_del(&o->hash_list);
592 kfree(o);
595 up_write(&_origins_lock);
599 * Implementation of the exception hash tables.
600 * The lowest hash_shift bits of the chunk number are ignored, allowing
601 * some consecutive chunks to be grouped together.
603 static int dm_exception_table_init(struct dm_exception_table *et,
604 uint32_t size, unsigned hash_shift)
606 unsigned int i;
608 et->hash_shift = hash_shift;
609 et->hash_mask = size - 1;
610 et->table = dm_vcalloc(size, sizeof(struct list_head));
611 if (!et->table)
612 return -ENOMEM;
614 for (i = 0; i < size; i++)
615 INIT_LIST_HEAD(et->table + i);
617 return 0;
620 static void dm_exception_table_exit(struct dm_exception_table *et,
621 struct kmem_cache *mem)
623 struct list_head *slot;
624 struct dm_exception *ex, *next;
625 int i, size;
627 size = et->hash_mask + 1;
628 for (i = 0; i < size; i++) {
629 slot = et->table + i;
631 list_for_each_entry_safe (ex, next, slot, hash_list)
632 kmem_cache_free(mem, ex);
635 vfree(et->table);
638 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
640 return (chunk >> et->hash_shift) & et->hash_mask;
643 static void dm_remove_exception(struct dm_exception *e)
645 list_del(&e->hash_list);
649 * Return the exception data for a sector, or NULL if not
650 * remapped.
652 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
653 chunk_t chunk)
655 struct list_head *slot;
656 struct dm_exception *e;
658 slot = &et->table[exception_hash(et, chunk)];
659 list_for_each_entry (e, slot, hash_list)
660 if (chunk >= e->old_chunk &&
661 chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
662 return e;
664 return NULL;
667 static struct dm_exception *alloc_completed_exception(gfp_t gfp)
669 struct dm_exception *e;
671 e = kmem_cache_alloc(exception_cache, gfp);
672 if (!e && gfp == GFP_NOIO)
673 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
675 return e;
678 static void free_completed_exception(struct dm_exception *e)
680 kmem_cache_free(exception_cache, e);
683 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
685 struct dm_snap_pending_exception *pe = mempool_alloc(s->pending_pool,
686 GFP_NOIO);
688 atomic_inc(&s->pending_exceptions_count);
689 pe->snap = s;
691 return pe;
694 static void free_pending_exception(struct dm_snap_pending_exception *pe)
696 struct dm_snapshot *s = pe->snap;
698 mempool_free(pe, s->pending_pool);
699 smp_mb__before_atomic();
700 atomic_dec(&s->pending_exceptions_count);
703 static void dm_insert_exception(struct dm_exception_table *eh,
704 struct dm_exception *new_e)
706 struct list_head *l;
707 struct dm_exception *e = NULL;
709 l = &eh->table[exception_hash(eh, new_e->old_chunk)];
711 /* Add immediately if this table doesn't support consecutive chunks */
712 if (!eh->hash_shift)
713 goto out;
715 /* List is ordered by old_chunk */
716 list_for_each_entry_reverse(e, l, hash_list) {
717 /* Insert after an existing chunk? */
718 if (new_e->old_chunk == (e->old_chunk +
719 dm_consecutive_chunk_count(e) + 1) &&
720 new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
721 dm_consecutive_chunk_count(e) + 1)) {
722 dm_consecutive_chunk_count_inc(e);
723 free_completed_exception(new_e);
724 return;
727 /* Insert before an existing chunk? */
728 if (new_e->old_chunk == (e->old_chunk - 1) &&
729 new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
730 dm_consecutive_chunk_count_inc(e);
731 e->old_chunk--;
732 e->new_chunk--;
733 free_completed_exception(new_e);
734 return;
737 if (new_e->old_chunk > e->old_chunk)
738 break;
741 out:
742 list_add(&new_e->hash_list, e ? &e->hash_list : l);
746 * Callback used by the exception stores to load exceptions when
747 * initialising.
749 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
751 struct dm_snapshot *s = context;
752 struct dm_exception *e;
754 e = alloc_completed_exception(GFP_KERNEL);
755 if (!e)
756 return -ENOMEM;
758 e->old_chunk = old;
760 /* Consecutive_count is implicitly initialised to zero */
761 e->new_chunk = new;
763 dm_insert_exception(&s->complete, e);
765 return 0;
769 * Return a minimum chunk size of all snapshots that have the specified origin.
770 * Return zero if the origin has no snapshots.
772 static uint32_t __minimum_chunk_size(struct origin *o)
774 struct dm_snapshot *snap;
775 unsigned chunk_size = 0;
777 if (o)
778 list_for_each_entry(snap, &o->snapshots, list)
779 chunk_size = min_not_zero(chunk_size,
780 snap->store->chunk_size);
782 return (uint32_t) chunk_size;
786 * Hard coded magic.
788 static int calc_max_buckets(void)
790 /* use a fixed size of 2MB */
791 unsigned long mem = 2 * 1024 * 1024;
792 mem /= sizeof(struct list_head);
794 return mem;
798 * Allocate room for a suitable hash table.
800 static int init_hash_tables(struct dm_snapshot *s)
802 sector_t hash_size, cow_dev_size, max_buckets;
805 * Calculate based on the size of the original volume or
806 * the COW volume...
808 cow_dev_size = get_dev_size(s->cow->bdev);
809 max_buckets = calc_max_buckets();
811 hash_size = cow_dev_size >> s->store->chunk_shift;
812 hash_size = min(hash_size, max_buckets);
814 if (hash_size < 64)
815 hash_size = 64;
816 hash_size = rounddown_pow_of_two(hash_size);
817 if (dm_exception_table_init(&s->complete, hash_size,
818 DM_CHUNK_CONSECUTIVE_BITS))
819 return -ENOMEM;
822 * Allocate hash table for in-flight exceptions
823 * Make this smaller than the real hash table
825 hash_size >>= 3;
826 if (hash_size < 64)
827 hash_size = 64;
829 if (dm_exception_table_init(&s->pending, hash_size, 0)) {
830 dm_exception_table_exit(&s->complete, exception_cache);
831 return -ENOMEM;
834 return 0;
837 static void merge_shutdown(struct dm_snapshot *s)
839 clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
840 smp_mb__after_atomic();
841 wake_up_bit(&s->state_bits, RUNNING_MERGE);
844 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
846 s->first_merging_chunk = 0;
847 s->num_merging_chunks = 0;
849 return bio_list_get(&s->bios_queued_during_merge);
853 * Remove one chunk from the index of completed exceptions.
855 static int __remove_single_exception_chunk(struct dm_snapshot *s,
856 chunk_t old_chunk)
858 struct dm_exception *e;
860 e = dm_lookup_exception(&s->complete, old_chunk);
861 if (!e) {
862 DMERR("Corruption detected: exception for block %llu is "
863 "on disk but not in memory",
864 (unsigned long long)old_chunk);
865 return -EINVAL;
869 * If this is the only chunk using this exception, remove exception.
871 if (!dm_consecutive_chunk_count(e)) {
872 dm_remove_exception(e);
873 free_completed_exception(e);
874 return 0;
878 * The chunk may be either at the beginning or the end of a
879 * group of consecutive chunks - never in the middle. We are
880 * removing chunks in the opposite order to that in which they
881 * were added, so this should always be true.
882 * Decrement the consecutive chunk counter and adjust the
883 * starting point if necessary.
885 if (old_chunk == e->old_chunk) {
886 e->old_chunk++;
887 e->new_chunk++;
888 } else if (old_chunk != e->old_chunk +
889 dm_consecutive_chunk_count(e)) {
890 DMERR("Attempt to merge block %llu from the "
891 "middle of a chunk range [%llu - %llu]",
892 (unsigned long long)old_chunk,
893 (unsigned long long)e->old_chunk,
894 (unsigned long long)
895 e->old_chunk + dm_consecutive_chunk_count(e));
896 return -EINVAL;
899 dm_consecutive_chunk_count_dec(e);
901 return 0;
904 static void flush_bios(struct bio *bio);
906 static int remove_single_exception_chunk(struct dm_snapshot *s)
908 struct bio *b = NULL;
909 int r;
910 chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
912 mutex_lock(&s->lock);
915 * Process chunks (and associated exceptions) in reverse order
916 * so that dm_consecutive_chunk_count_dec() accounting works.
918 do {
919 r = __remove_single_exception_chunk(s, old_chunk);
920 if (r)
921 goto out;
922 } while (old_chunk-- > s->first_merging_chunk);
924 b = __release_queued_bios_after_merge(s);
926 out:
927 mutex_unlock(&s->lock);
928 if (b)
929 flush_bios(b);
931 return r;
934 static int origin_write_extent(struct dm_snapshot *merging_snap,
935 sector_t sector, unsigned chunk_size);
937 static void merge_callback(int read_err, unsigned long write_err,
938 void *context);
940 static uint64_t read_pending_exceptions_done_count(void)
942 uint64_t pending_exceptions_done;
944 spin_lock(&_pending_exceptions_done_spinlock);
945 pending_exceptions_done = _pending_exceptions_done_count;
946 spin_unlock(&_pending_exceptions_done_spinlock);
948 return pending_exceptions_done;
951 static void increment_pending_exceptions_done_count(void)
953 spin_lock(&_pending_exceptions_done_spinlock);
954 _pending_exceptions_done_count++;
955 spin_unlock(&_pending_exceptions_done_spinlock);
957 wake_up_all(&_pending_exceptions_done);
960 static void snapshot_merge_next_chunks(struct dm_snapshot *s)
962 int i, linear_chunks;
963 chunk_t old_chunk, new_chunk;
964 struct dm_io_region src, dest;
965 sector_t io_size;
966 uint64_t previous_count;
968 BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
969 if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
970 goto shut;
973 * valid flag never changes during merge, so no lock required.
975 if (!s->valid) {
976 DMERR("Snapshot is invalid: can't merge");
977 goto shut;
980 linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
981 &new_chunk);
982 if (linear_chunks <= 0) {
983 if (linear_chunks < 0) {
984 DMERR("Read error in exception store: "
985 "shutting down merge");
986 mutex_lock(&s->lock);
987 s->merge_failed = 1;
988 mutex_unlock(&s->lock);
990 goto shut;
993 /* Adjust old_chunk and new_chunk to reflect start of linear region */
994 old_chunk = old_chunk + 1 - linear_chunks;
995 new_chunk = new_chunk + 1 - linear_chunks;
998 * Use one (potentially large) I/O to copy all 'linear_chunks'
999 * from the exception store to the origin
1001 io_size = linear_chunks * s->store->chunk_size;
1003 dest.bdev = s->origin->bdev;
1004 dest.sector = chunk_to_sector(s->store, old_chunk);
1005 dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
1007 src.bdev = s->cow->bdev;
1008 src.sector = chunk_to_sector(s->store, new_chunk);
1009 src.count = dest.count;
1012 * Reallocate any exceptions needed in other snapshots then
1013 * wait for the pending exceptions to complete.
1014 * Each time any pending exception (globally on the system)
1015 * completes we are woken and repeat the process to find out
1016 * if we can proceed. While this may not seem a particularly
1017 * efficient algorithm, it is not expected to have any
1018 * significant impact on performance.
1020 previous_count = read_pending_exceptions_done_count();
1021 while (origin_write_extent(s, dest.sector, io_size)) {
1022 wait_event(_pending_exceptions_done,
1023 (read_pending_exceptions_done_count() !=
1024 previous_count));
1025 /* Retry after the wait, until all exceptions are done. */
1026 previous_count = read_pending_exceptions_done_count();
1029 mutex_lock(&s->lock);
1030 s->first_merging_chunk = old_chunk;
1031 s->num_merging_chunks = linear_chunks;
1032 mutex_unlock(&s->lock);
1034 /* Wait until writes to all 'linear_chunks' drain */
1035 for (i = 0; i < linear_chunks; i++)
1036 __check_for_conflicting_io(s, old_chunk + i);
1038 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
1039 return;
1041 shut:
1042 merge_shutdown(s);
1045 static void error_bios(struct bio *bio);
1047 static void merge_callback(int read_err, unsigned long write_err, void *context)
1049 struct dm_snapshot *s = context;
1050 struct bio *b = NULL;
1052 if (read_err || write_err) {
1053 if (read_err)
1054 DMERR("Read error: shutting down merge.");
1055 else
1056 DMERR("Write error: shutting down merge.");
1057 goto shut;
1060 if (s->store->type->commit_merge(s->store,
1061 s->num_merging_chunks) < 0) {
1062 DMERR("Write error in exception store: shutting down merge");
1063 goto shut;
1066 if (remove_single_exception_chunk(s) < 0)
1067 goto shut;
1069 snapshot_merge_next_chunks(s);
1071 return;
1073 shut:
1074 mutex_lock(&s->lock);
1075 s->merge_failed = 1;
1076 b = __release_queued_bios_after_merge(s);
1077 mutex_unlock(&s->lock);
1078 error_bios(b);
1080 merge_shutdown(s);
1083 static void start_merge(struct dm_snapshot *s)
1085 if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1086 snapshot_merge_next_chunks(s);
1090 * Stop the merging process and wait until it finishes.
1092 static void stop_merge(struct dm_snapshot *s)
1094 set_bit(SHUTDOWN_MERGE, &s->state_bits);
1095 wait_on_bit(&s->state_bits, RUNNING_MERGE, TASK_UNINTERRUPTIBLE);
1096 clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1100 * Construct a snapshot mapping: <origin_dev> <COW-dev> <p|po|n> <chunk-size>
1102 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1104 struct dm_snapshot *s;
1105 int i;
1106 int r = -EINVAL;
1107 char *origin_path, *cow_path;
1108 dev_t origin_dev, cow_dev;
1109 unsigned args_used, num_flush_bios = 1;
1110 fmode_t origin_mode = FMODE_READ;
1112 if (argc != 4) {
1113 ti->error = "requires exactly 4 arguments";
1114 r = -EINVAL;
1115 goto bad;
1118 if (dm_target_is_snapshot_merge(ti)) {
1119 num_flush_bios = 2;
1120 origin_mode = FMODE_WRITE;
1123 s = kmalloc(sizeof(*s), GFP_KERNEL);
1124 if (!s) {
1125 ti->error = "Cannot allocate private snapshot structure";
1126 r = -ENOMEM;
1127 goto bad;
1130 origin_path = argv[0];
1131 argv++;
1132 argc--;
1134 r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1135 if (r) {
1136 ti->error = "Cannot get origin device";
1137 goto bad_origin;
1139 origin_dev = s->origin->bdev->bd_dev;
1141 cow_path = argv[0];
1142 argv++;
1143 argc--;
1145 cow_dev = dm_get_dev_t(cow_path);
1146 if (cow_dev && cow_dev == origin_dev) {
1147 ti->error = "COW device cannot be the same as origin device";
1148 r = -EINVAL;
1149 goto bad_cow;
1152 r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1153 if (r) {
1154 ti->error = "Cannot get COW device";
1155 goto bad_cow;
1158 r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1159 if (r) {
1160 ti->error = "Couldn't create exception store";
1161 r = -EINVAL;
1162 goto bad_store;
1165 argv += args_used;
1166 argc -= args_used;
1168 s->ti = ti;
1169 s->valid = 1;
1170 s->snapshot_overflowed = 0;
1171 s->active = 0;
1172 atomic_set(&s->pending_exceptions_count, 0);
1173 s->exception_start_sequence = 0;
1174 s->exception_complete_sequence = 0;
1175 INIT_LIST_HEAD(&s->out_of_order_list);
1176 mutex_init(&s->lock);
1177 INIT_LIST_HEAD(&s->list);
1178 spin_lock_init(&s->pe_lock);
1179 s->state_bits = 0;
1180 s->merge_failed = 0;
1181 s->first_merging_chunk = 0;
1182 s->num_merging_chunks = 0;
1183 bio_list_init(&s->bios_queued_during_merge);
1185 /* Allocate hash table for COW data */
1186 if (init_hash_tables(s)) {
1187 ti->error = "Unable to allocate hash table space";
1188 r = -ENOMEM;
1189 goto bad_hash_tables;
1192 s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1193 if (IS_ERR(s->kcopyd_client)) {
1194 r = PTR_ERR(s->kcopyd_client);
1195 ti->error = "Could not create kcopyd client";
1196 goto bad_kcopyd;
1199 s->pending_pool = mempool_create_slab_pool(MIN_IOS, pending_cache);
1200 if (!s->pending_pool) {
1201 ti->error = "Could not allocate mempool for pending exceptions";
1202 r = -ENOMEM;
1203 goto bad_pending_pool;
1206 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1207 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1209 spin_lock_init(&s->tracked_chunk_lock);
1211 ti->private = s;
1212 ti->num_flush_bios = num_flush_bios;
1213 ti->per_io_data_size = sizeof(struct dm_snap_tracked_chunk);
1215 /* Add snapshot to the list of snapshots for this origin */
1216 /* Exceptions aren't triggered till snapshot_resume() is called */
1217 r = register_snapshot(s);
1218 if (r == -ENOMEM) {
1219 ti->error = "Snapshot origin struct allocation failed";
1220 goto bad_load_and_register;
1221 } else if (r < 0) {
1222 /* invalid handover, register_snapshot has set ti->error */
1223 goto bad_load_and_register;
1227 * Metadata must only be loaded into one table at once, so skip this
1228 * if metadata will be handed over during resume.
1229 * Chunk size will be set during the handover - set it to zero to
1230 * ensure it's ignored.
1232 if (r > 0) {
1233 s->store->chunk_size = 0;
1234 return 0;
1237 r = s->store->type->read_metadata(s->store, dm_add_exception,
1238 (void *)s);
1239 if (r < 0) {
1240 ti->error = "Failed to read snapshot metadata";
1241 goto bad_read_metadata;
1242 } else if (r > 0) {
1243 s->valid = 0;
1244 DMWARN("Snapshot is marked invalid.");
1247 if (!s->store->chunk_size) {
1248 ti->error = "Chunk size not set";
1249 goto bad_read_metadata;
1252 r = dm_set_target_max_io_len(ti, s->store->chunk_size);
1253 if (r)
1254 goto bad_read_metadata;
1256 return 0;
1258 bad_read_metadata:
1259 unregister_snapshot(s);
1261 bad_load_and_register:
1262 mempool_destroy(s->pending_pool);
1264 bad_pending_pool:
1265 dm_kcopyd_client_destroy(s->kcopyd_client);
1267 bad_kcopyd:
1268 dm_exception_table_exit(&s->pending, pending_cache);
1269 dm_exception_table_exit(&s->complete, exception_cache);
1271 bad_hash_tables:
1272 dm_exception_store_destroy(s->store);
1274 bad_store:
1275 dm_put_device(ti, s->cow);
1277 bad_cow:
1278 dm_put_device(ti, s->origin);
1280 bad_origin:
1281 kfree(s);
1283 bad:
1284 return r;
1287 static void __free_exceptions(struct dm_snapshot *s)
1289 dm_kcopyd_client_destroy(s->kcopyd_client);
1290 s->kcopyd_client = NULL;
1292 dm_exception_table_exit(&s->pending, pending_cache);
1293 dm_exception_table_exit(&s->complete, exception_cache);
1296 static void __handover_exceptions(struct dm_snapshot *snap_src,
1297 struct dm_snapshot *snap_dest)
1299 union {
1300 struct dm_exception_table table_swap;
1301 struct dm_exception_store *store_swap;
1302 } u;
1305 * Swap all snapshot context information between the two instances.
1307 u.table_swap = snap_dest->complete;
1308 snap_dest->complete = snap_src->complete;
1309 snap_src->complete = u.table_swap;
1311 u.store_swap = snap_dest->store;
1312 snap_dest->store = snap_src->store;
1313 snap_dest->store->userspace_supports_overflow = u.store_swap->userspace_supports_overflow;
1314 snap_src->store = u.store_swap;
1316 snap_dest->store->snap = snap_dest;
1317 snap_src->store->snap = snap_src;
1319 snap_dest->ti->max_io_len = snap_dest->store->chunk_size;
1320 snap_dest->valid = snap_src->valid;
1321 snap_dest->snapshot_overflowed = snap_src->snapshot_overflowed;
1324 * Set source invalid to ensure it receives no further I/O.
1326 snap_src->valid = 0;
1329 static void snapshot_dtr(struct dm_target *ti)
1331 #ifdef CONFIG_DM_DEBUG
1332 int i;
1333 #endif
1334 struct dm_snapshot *s = ti->private;
1335 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1337 down_read(&_origins_lock);
1338 /* Check whether exception handover must be cancelled */
1339 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1340 if (snap_src && snap_dest && (s == snap_src)) {
1341 mutex_lock(&snap_dest->lock);
1342 snap_dest->valid = 0;
1343 mutex_unlock(&snap_dest->lock);
1344 DMERR("Cancelling snapshot handover.");
1346 up_read(&_origins_lock);
1348 if (dm_target_is_snapshot_merge(ti))
1349 stop_merge(s);
1351 /* Prevent further origin writes from using this snapshot. */
1352 /* After this returns there can be no new kcopyd jobs. */
1353 unregister_snapshot(s);
1355 while (atomic_read(&s->pending_exceptions_count))
1356 msleep(1);
1358 * Ensure instructions in mempool_destroy aren't reordered
1359 * before atomic_read.
1361 smp_mb();
1363 #ifdef CONFIG_DM_DEBUG
1364 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1365 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1366 #endif
1368 __free_exceptions(s);
1370 mempool_destroy(s->pending_pool);
1372 dm_exception_store_destroy(s->store);
1374 mutex_destroy(&s->lock);
1376 dm_put_device(ti, s->cow);
1378 dm_put_device(ti, s->origin);
1380 kfree(s);
1384 * Flush a list of buffers.
1386 static void flush_bios(struct bio *bio)
1388 struct bio *n;
1390 while (bio) {
1391 n = bio->bi_next;
1392 bio->bi_next = NULL;
1393 generic_make_request(bio);
1394 bio = n;
1398 static int do_origin(struct dm_dev *origin, struct bio *bio);
1401 * Flush a list of buffers.
1403 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1405 struct bio *n;
1406 int r;
1408 while (bio) {
1409 n = bio->bi_next;
1410 bio->bi_next = NULL;
1411 r = do_origin(s->origin, bio);
1412 if (r == DM_MAPIO_REMAPPED)
1413 generic_make_request(bio);
1414 bio = n;
1419 * Error a list of buffers.
1421 static void error_bios(struct bio *bio)
1423 struct bio *n;
1425 while (bio) {
1426 n = bio->bi_next;
1427 bio->bi_next = NULL;
1428 bio_io_error(bio);
1429 bio = n;
1433 static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1435 if (!s->valid)
1436 return;
1438 if (err == -EIO)
1439 DMERR("Invalidating snapshot: Error reading/writing.");
1440 else if (err == -ENOMEM)
1441 DMERR("Invalidating snapshot: Unable to allocate exception.");
1443 if (s->store->type->drop_snapshot)
1444 s->store->type->drop_snapshot(s->store);
1446 s->valid = 0;
1448 dm_table_event(s->ti->table);
1451 static void pending_complete(void *context, int success)
1453 struct dm_snap_pending_exception *pe = context;
1454 struct dm_exception *e;
1455 struct dm_snapshot *s = pe->snap;
1456 struct bio *origin_bios = NULL;
1457 struct bio *snapshot_bios = NULL;
1458 struct bio *full_bio = NULL;
1459 int error = 0;
1461 if (!success) {
1462 /* Read/write error - snapshot is unusable */
1463 mutex_lock(&s->lock);
1464 __invalidate_snapshot(s, -EIO);
1465 error = 1;
1466 goto out;
1469 e = alloc_completed_exception(GFP_NOIO);
1470 if (!e) {
1471 mutex_lock(&s->lock);
1472 __invalidate_snapshot(s, -ENOMEM);
1473 error = 1;
1474 goto out;
1476 *e = pe->e;
1478 mutex_lock(&s->lock);
1479 if (!s->valid) {
1480 free_completed_exception(e);
1481 error = 1;
1482 goto out;
1485 /* Check for conflicting reads */
1486 __check_for_conflicting_io(s, pe->e.old_chunk);
1489 * Add a proper exception, and remove the
1490 * in-flight exception from the list.
1492 dm_insert_exception(&s->complete, e);
1494 out:
1495 dm_remove_exception(&pe->e);
1496 snapshot_bios = bio_list_get(&pe->snapshot_bios);
1497 origin_bios = bio_list_get(&pe->origin_bios);
1498 full_bio = pe->full_bio;
1499 if (full_bio)
1500 full_bio->bi_end_io = pe->full_bio_end_io;
1501 increment_pending_exceptions_done_count();
1503 mutex_unlock(&s->lock);
1505 /* Submit any pending write bios */
1506 if (error) {
1507 if (full_bio)
1508 bio_io_error(full_bio);
1509 error_bios(snapshot_bios);
1510 } else {
1511 if (full_bio)
1512 bio_endio(full_bio);
1513 flush_bios(snapshot_bios);
1516 retry_origin_bios(s, origin_bios);
1518 free_pending_exception(pe);
1521 static void complete_exception(struct dm_snap_pending_exception *pe)
1523 struct dm_snapshot *s = pe->snap;
1525 /* Update the metadata if we are persistent */
1526 s->store->type->commit_exception(s->store, &pe->e, !pe->copy_error,
1527 pending_complete, pe);
1531 * Called when the copy I/O has finished. kcopyd actually runs
1532 * this code so don't block.
1534 static void copy_callback(int read_err, unsigned long write_err, void *context)
1536 struct dm_snap_pending_exception *pe = context;
1537 struct dm_snapshot *s = pe->snap;
1539 pe->copy_error = read_err || write_err;
1541 if (pe->exception_sequence == s->exception_complete_sequence) {
1542 s->exception_complete_sequence++;
1543 complete_exception(pe);
1545 while (!list_empty(&s->out_of_order_list)) {
1546 pe = list_entry(s->out_of_order_list.next,
1547 struct dm_snap_pending_exception, out_of_order_entry);
1548 if (pe->exception_sequence != s->exception_complete_sequence)
1549 break;
1550 s->exception_complete_sequence++;
1551 list_del(&pe->out_of_order_entry);
1552 complete_exception(pe);
1554 } else {
1555 struct list_head *lh;
1556 struct dm_snap_pending_exception *pe2;
1558 list_for_each_prev(lh, &s->out_of_order_list) {
1559 pe2 = list_entry(lh, struct dm_snap_pending_exception, out_of_order_entry);
1560 if (pe2->exception_sequence < pe->exception_sequence)
1561 break;
1563 list_add(&pe->out_of_order_entry, lh);
1568 * Dispatches the copy operation to kcopyd.
1570 static void start_copy(struct dm_snap_pending_exception *pe)
1572 struct dm_snapshot *s = pe->snap;
1573 struct dm_io_region src, dest;
1574 struct block_device *bdev = s->origin->bdev;
1575 sector_t dev_size;
1577 dev_size = get_dev_size(bdev);
1579 src.bdev = bdev;
1580 src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1581 src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1583 dest.bdev = s->cow->bdev;
1584 dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1585 dest.count = src.count;
1587 /* Hand over to kcopyd */
1588 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1591 static void full_bio_end_io(struct bio *bio)
1593 void *callback_data = bio->bi_private;
1595 dm_kcopyd_do_callback(callback_data, 0, bio->bi_status ? 1 : 0);
1598 static void start_full_bio(struct dm_snap_pending_exception *pe,
1599 struct bio *bio)
1601 struct dm_snapshot *s = pe->snap;
1602 void *callback_data;
1604 pe->full_bio = bio;
1605 pe->full_bio_end_io = bio->bi_end_io;
1607 callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1608 copy_callback, pe);
1610 bio->bi_end_io = full_bio_end_io;
1611 bio->bi_private = callback_data;
1613 generic_make_request(bio);
1616 static struct dm_snap_pending_exception *
1617 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1619 struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1621 if (!e)
1622 return NULL;
1624 return container_of(e, struct dm_snap_pending_exception, e);
1628 * Looks to see if this snapshot already has a pending exception
1629 * for this chunk, otherwise it allocates a new one and inserts
1630 * it into the pending table.
1632 * NOTE: a write lock must be held on snap->lock before calling
1633 * this.
1635 static struct dm_snap_pending_exception *
1636 __find_pending_exception(struct dm_snapshot *s,
1637 struct dm_snap_pending_exception *pe, chunk_t chunk)
1639 struct dm_snap_pending_exception *pe2;
1641 pe2 = __lookup_pending_exception(s, chunk);
1642 if (pe2) {
1643 free_pending_exception(pe);
1644 return pe2;
1647 pe->e.old_chunk = chunk;
1648 bio_list_init(&pe->origin_bios);
1649 bio_list_init(&pe->snapshot_bios);
1650 pe->started = 0;
1651 pe->full_bio = NULL;
1653 if (s->store->type->prepare_exception(s->store, &pe->e)) {
1654 free_pending_exception(pe);
1655 return NULL;
1658 pe->exception_sequence = s->exception_start_sequence++;
1660 dm_insert_exception(&s->pending, &pe->e);
1662 return pe;
1665 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1666 struct bio *bio, chunk_t chunk)
1668 bio_set_dev(bio, s->cow->bdev);
1669 bio->bi_iter.bi_sector =
1670 chunk_to_sector(s->store, dm_chunk_number(e->new_chunk) +
1671 (chunk - e->old_chunk)) +
1672 (bio->bi_iter.bi_sector & s->store->chunk_mask);
1675 static int snapshot_map(struct dm_target *ti, struct bio *bio)
1677 struct dm_exception *e;
1678 struct dm_snapshot *s = ti->private;
1679 int r = DM_MAPIO_REMAPPED;
1680 chunk_t chunk;
1681 struct dm_snap_pending_exception *pe = NULL;
1683 init_tracked_chunk(bio);
1685 if (bio->bi_opf & REQ_PREFLUSH) {
1686 bio_set_dev(bio, s->cow->bdev);
1687 return DM_MAPIO_REMAPPED;
1690 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
1692 /* Full snapshots are not usable */
1693 /* To get here the table must be live so s->active is always set. */
1694 if (!s->valid)
1695 return DM_MAPIO_KILL;
1697 /* FIXME: should only take write lock if we need
1698 * to copy an exception */
1699 mutex_lock(&s->lock);
1701 if (!s->valid || (unlikely(s->snapshot_overflowed) &&
1702 bio_data_dir(bio) == WRITE)) {
1703 r = DM_MAPIO_KILL;
1704 goto out_unlock;
1707 /* If the block is already remapped - use that, else remap it */
1708 e = dm_lookup_exception(&s->complete, chunk);
1709 if (e) {
1710 remap_exception(s, e, bio, chunk);
1711 goto out_unlock;
1715 * Write to snapshot - higher level takes care of RW/RO
1716 * flags so we should only get this if we are
1717 * writeable.
1719 if (bio_data_dir(bio) == WRITE) {
1720 pe = __lookup_pending_exception(s, chunk);
1721 if (!pe) {
1722 mutex_unlock(&s->lock);
1723 pe = alloc_pending_exception(s);
1724 mutex_lock(&s->lock);
1726 if (!s->valid || s->snapshot_overflowed) {
1727 free_pending_exception(pe);
1728 r = DM_MAPIO_KILL;
1729 goto out_unlock;
1732 e = dm_lookup_exception(&s->complete, chunk);
1733 if (e) {
1734 free_pending_exception(pe);
1735 remap_exception(s, e, bio, chunk);
1736 goto out_unlock;
1739 pe = __find_pending_exception(s, pe, chunk);
1740 if (!pe) {
1741 if (s->store->userspace_supports_overflow) {
1742 s->snapshot_overflowed = 1;
1743 DMERR("Snapshot overflowed: Unable to allocate exception.");
1744 } else
1745 __invalidate_snapshot(s, -ENOMEM);
1746 r = DM_MAPIO_KILL;
1747 goto out_unlock;
1751 remap_exception(s, &pe->e, bio, chunk);
1753 r = DM_MAPIO_SUBMITTED;
1755 if (!pe->started &&
1756 bio->bi_iter.bi_size ==
1757 (s->store->chunk_size << SECTOR_SHIFT)) {
1758 pe->started = 1;
1759 mutex_unlock(&s->lock);
1760 start_full_bio(pe, bio);
1761 goto out;
1764 bio_list_add(&pe->snapshot_bios, bio);
1766 if (!pe->started) {
1767 /* this is protected by snap->lock */
1768 pe->started = 1;
1769 mutex_unlock(&s->lock);
1770 start_copy(pe);
1771 goto out;
1773 } else {
1774 bio_set_dev(bio, s->origin->bdev);
1775 track_chunk(s, bio, chunk);
1778 out_unlock:
1779 mutex_unlock(&s->lock);
1780 out:
1781 return r;
1785 * A snapshot-merge target behaves like a combination of a snapshot
1786 * target and a snapshot-origin target. It only generates new
1787 * exceptions in other snapshots and not in the one that is being
1788 * merged.
1790 * For each chunk, if there is an existing exception, it is used to
1791 * redirect I/O to the cow device. Otherwise I/O is sent to the origin,
1792 * which in turn might generate exceptions in other snapshots.
1793 * If merging is currently taking place on the chunk in question, the
1794 * I/O is deferred by adding it to s->bios_queued_during_merge.
1796 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio)
1798 struct dm_exception *e;
1799 struct dm_snapshot *s = ti->private;
1800 int r = DM_MAPIO_REMAPPED;
1801 chunk_t chunk;
1803 init_tracked_chunk(bio);
1805 if (bio->bi_opf & REQ_PREFLUSH) {
1806 if (!dm_bio_get_target_bio_nr(bio))
1807 bio_set_dev(bio, s->origin->bdev);
1808 else
1809 bio_set_dev(bio, s->cow->bdev);
1810 return DM_MAPIO_REMAPPED;
1813 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
1815 mutex_lock(&s->lock);
1817 /* Full merging snapshots are redirected to the origin */
1818 if (!s->valid)
1819 goto redirect_to_origin;
1821 /* If the block is already remapped - use that */
1822 e = dm_lookup_exception(&s->complete, chunk);
1823 if (e) {
1824 /* Queue writes overlapping with chunks being merged */
1825 if (bio_data_dir(bio) == WRITE &&
1826 chunk >= s->first_merging_chunk &&
1827 chunk < (s->first_merging_chunk +
1828 s->num_merging_chunks)) {
1829 bio_set_dev(bio, s->origin->bdev);
1830 bio_list_add(&s->bios_queued_during_merge, bio);
1831 r = DM_MAPIO_SUBMITTED;
1832 goto out_unlock;
1835 remap_exception(s, e, bio, chunk);
1837 if (bio_data_dir(bio) == WRITE)
1838 track_chunk(s, bio, chunk);
1839 goto out_unlock;
1842 redirect_to_origin:
1843 bio_set_dev(bio, s->origin->bdev);
1845 if (bio_data_dir(bio) == WRITE) {
1846 mutex_unlock(&s->lock);
1847 return do_origin(s->origin, bio);
1850 out_unlock:
1851 mutex_unlock(&s->lock);
1853 return r;
1856 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
1857 blk_status_t *error)
1859 struct dm_snapshot *s = ti->private;
1861 if (is_bio_tracked(bio))
1862 stop_tracking_chunk(s, bio);
1864 return DM_ENDIO_DONE;
1867 static void snapshot_merge_presuspend(struct dm_target *ti)
1869 struct dm_snapshot *s = ti->private;
1871 stop_merge(s);
1874 static int snapshot_preresume(struct dm_target *ti)
1876 int r = 0;
1877 struct dm_snapshot *s = ti->private;
1878 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1880 down_read(&_origins_lock);
1881 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1882 if (snap_src && snap_dest) {
1883 mutex_lock(&snap_src->lock);
1884 if (s == snap_src) {
1885 DMERR("Unable to resume snapshot source until "
1886 "handover completes.");
1887 r = -EINVAL;
1888 } else if (!dm_suspended(snap_src->ti)) {
1889 DMERR("Unable to perform snapshot handover until "
1890 "source is suspended.");
1891 r = -EINVAL;
1893 mutex_unlock(&snap_src->lock);
1895 up_read(&_origins_lock);
1897 return r;
1900 static void snapshot_resume(struct dm_target *ti)
1902 struct dm_snapshot *s = ti->private;
1903 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL, *snap_merging = NULL;
1904 struct dm_origin *o;
1905 struct mapped_device *origin_md = NULL;
1906 bool must_restart_merging = false;
1908 down_read(&_origins_lock);
1910 o = __lookup_dm_origin(s->origin->bdev);
1911 if (o)
1912 origin_md = dm_table_get_md(o->ti->table);
1913 if (!origin_md) {
1914 (void) __find_snapshots_sharing_cow(s, NULL, NULL, &snap_merging);
1915 if (snap_merging)
1916 origin_md = dm_table_get_md(snap_merging->ti->table);
1918 if (origin_md == dm_table_get_md(ti->table))
1919 origin_md = NULL;
1920 if (origin_md) {
1921 if (dm_hold(origin_md))
1922 origin_md = NULL;
1925 up_read(&_origins_lock);
1927 if (origin_md) {
1928 dm_internal_suspend_fast(origin_md);
1929 if (snap_merging && test_bit(RUNNING_MERGE, &snap_merging->state_bits)) {
1930 must_restart_merging = true;
1931 stop_merge(snap_merging);
1935 down_read(&_origins_lock);
1937 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1938 if (snap_src && snap_dest) {
1939 mutex_lock(&snap_src->lock);
1940 mutex_lock_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
1941 __handover_exceptions(snap_src, snap_dest);
1942 mutex_unlock(&snap_dest->lock);
1943 mutex_unlock(&snap_src->lock);
1946 up_read(&_origins_lock);
1948 if (origin_md) {
1949 if (must_restart_merging)
1950 start_merge(snap_merging);
1951 dm_internal_resume_fast(origin_md);
1952 dm_put(origin_md);
1955 /* Now we have correct chunk size, reregister */
1956 reregister_snapshot(s);
1958 mutex_lock(&s->lock);
1959 s->active = 1;
1960 mutex_unlock(&s->lock);
1963 static uint32_t get_origin_minimum_chunksize(struct block_device *bdev)
1965 uint32_t min_chunksize;
1967 down_read(&_origins_lock);
1968 min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
1969 up_read(&_origins_lock);
1971 return min_chunksize;
1974 static void snapshot_merge_resume(struct dm_target *ti)
1976 struct dm_snapshot *s = ti->private;
1979 * Handover exceptions from existing snapshot.
1981 snapshot_resume(ti);
1984 * snapshot-merge acts as an origin, so set ti->max_io_len
1986 ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev);
1988 start_merge(s);
1991 static void snapshot_status(struct dm_target *ti, status_type_t type,
1992 unsigned status_flags, char *result, unsigned maxlen)
1994 unsigned sz = 0;
1995 struct dm_snapshot *snap = ti->private;
1997 switch (type) {
1998 case STATUSTYPE_INFO:
2000 mutex_lock(&snap->lock);
2002 if (!snap->valid)
2003 DMEMIT("Invalid");
2004 else if (snap->merge_failed)
2005 DMEMIT("Merge failed");
2006 else if (snap->snapshot_overflowed)
2007 DMEMIT("Overflow");
2008 else {
2009 if (snap->store->type->usage) {
2010 sector_t total_sectors, sectors_allocated,
2011 metadata_sectors;
2012 snap->store->type->usage(snap->store,
2013 &total_sectors,
2014 &sectors_allocated,
2015 &metadata_sectors);
2016 DMEMIT("%llu/%llu %llu",
2017 (unsigned long long)sectors_allocated,
2018 (unsigned long long)total_sectors,
2019 (unsigned long long)metadata_sectors);
2021 else
2022 DMEMIT("Unknown");
2025 mutex_unlock(&snap->lock);
2027 break;
2029 case STATUSTYPE_TABLE:
2031 * kdevname returns a static pointer so we need
2032 * to make private copies if the output is to
2033 * make sense.
2035 DMEMIT("%s %s", snap->origin->name, snap->cow->name);
2036 snap->store->type->status(snap->store, type, result + sz,
2037 maxlen - sz);
2038 break;
2042 static int snapshot_iterate_devices(struct dm_target *ti,
2043 iterate_devices_callout_fn fn, void *data)
2045 struct dm_snapshot *snap = ti->private;
2046 int r;
2048 r = fn(ti, snap->origin, 0, ti->len, data);
2050 if (!r)
2051 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
2053 return r;
2057 /*-----------------------------------------------------------------
2058 * Origin methods
2059 *---------------------------------------------------------------*/
2062 * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
2063 * supplied bio was ignored. The caller may submit it immediately.
2064 * (No remapping actually occurs as the origin is always a direct linear
2065 * map.)
2067 * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
2068 * and any supplied bio is added to a list to be submitted once all
2069 * the necessary exceptions exist.
2071 static int __origin_write(struct list_head *snapshots, sector_t sector,
2072 struct bio *bio)
2074 int r = DM_MAPIO_REMAPPED;
2075 struct dm_snapshot *snap;
2076 struct dm_exception *e;
2077 struct dm_snap_pending_exception *pe;
2078 struct dm_snap_pending_exception *pe_to_start_now = NULL;
2079 struct dm_snap_pending_exception *pe_to_start_last = NULL;
2080 chunk_t chunk;
2082 /* Do all the snapshots on this origin */
2083 list_for_each_entry (snap, snapshots, list) {
2085 * Don't make new exceptions in a merging snapshot
2086 * because it has effectively been deleted
2088 if (dm_target_is_snapshot_merge(snap->ti))
2089 continue;
2091 mutex_lock(&snap->lock);
2093 /* Only deal with valid and active snapshots */
2094 if (!snap->valid || !snap->active)
2095 goto next_snapshot;
2097 /* Nothing to do if writing beyond end of snapshot */
2098 if (sector >= dm_table_get_size(snap->ti->table))
2099 goto next_snapshot;
2102 * Remember, different snapshots can have
2103 * different chunk sizes.
2105 chunk = sector_to_chunk(snap->store, sector);
2108 * Check exception table to see if block
2109 * is already remapped in this snapshot
2110 * and trigger an exception if not.
2112 e = dm_lookup_exception(&snap->complete, chunk);
2113 if (e)
2114 goto next_snapshot;
2116 pe = __lookup_pending_exception(snap, chunk);
2117 if (!pe) {
2118 mutex_unlock(&snap->lock);
2119 pe = alloc_pending_exception(snap);
2120 mutex_lock(&snap->lock);
2122 if (!snap->valid) {
2123 free_pending_exception(pe);
2124 goto next_snapshot;
2127 e = dm_lookup_exception(&snap->complete, chunk);
2128 if (e) {
2129 free_pending_exception(pe);
2130 goto next_snapshot;
2133 pe = __find_pending_exception(snap, pe, chunk);
2134 if (!pe) {
2135 __invalidate_snapshot(snap, -ENOMEM);
2136 goto next_snapshot;
2140 r = DM_MAPIO_SUBMITTED;
2143 * If an origin bio was supplied, queue it to wait for the
2144 * completion of this exception, and start this one last,
2145 * at the end of the function.
2147 if (bio) {
2148 bio_list_add(&pe->origin_bios, bio);
2149 bio = NULL;
2151 if (!pe->started) {
2152 pe->started = 1;
2153 pe_to_start_last = pe;
2157 if (!pe->started) {
2158 pe->started = 1;
2159 pe_to_start_now = pe;
2162 next_snapshot:
2163 mutex_unlock(&snap->lock);
2165 if (pe_to_start_now) {
2166 start_copy(pe_to_start_now);
2167 pe_to_start_now = NULL;
2172 * Submit the exception against which the bio is queued last,
2173 * to give the other exceptions a head start.
2175 if (pe_to_start_last)
2176 start_copy(pe_to_start_last);
2178 return r;
2182 * Called on a write from the origin driver.
2184 static int do_origin(struct dm_dev *origin, struct bio *bio)
2186 struct origin *o;
2187 int r = DM_MAPIO_REMAPPED;
2189 down_read(&_origins_lock);
2190 o = __lookup_origin(origin->bdev);
2191 if (o)
2192 r = __origin_write(&o->snapshots, bio->bi_iter.bi_sector, bio);
2193 up_read(&_origins_lock);
2195 return r;
2199 * Trigger exceptions in all non-merging snapshots.
2201 * The chunk size of the merging snapshot may be larger than the chunk
2202 * size of some other snapshot so we may need to reallocate multiple
2203 * chunks in other snapshots.
2205 * We scan all the overlapping exceptions in the other snapshots.
2206 * Returns 1 if anything was reallocated and must be waited for,
2207 * otherwise returns 0.
2209 * size must be a multiple of merging_snap's chunk_size.
2211 static int origin_write_extent(struct dm_snapshot *merging_snap,
2212 sector_t sector, unsigned size)
2214 int must_wait = 0;
2215 sector_t n;
2216 struct origin *o;
2219 * The origin's __minimum_chunk_size() got stored in max_io_len
2220 * by snapshot_merge_resume().
2222 down_read(&_origins_lock);
2223 o = __lookup_origin(merging_snap->origin->bdev);
2224 for (n = 0; n < size; n += merging_snap->ti->max_io_len)
2225 if (__origin_write(&o->snapshots, sector + n, NULL) ==
2226 DM_MAPIO_SUBMITTED)
2227 must_wait = 1;
2228 up_read(&_origins_lock);
2230 return must_wait;
2234 * Origin: maps a linear range of a device, with hooks for snapshotting.
2238 * Construct an origin mapping: <dev_path>
2239 * The context for an origin is merely a 'struct dm_dev *'
2240 * pointing to the real device.
2242 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2244 int r;
2245 struct dm_origin *o;
2247 if (argc != 1) {
2248 ti->error = "origin: incorrect number of arguments";
2249 return -EINVAL;
2252 o = kmalloc(sizeof(struct dm_origin), GFP_KERNEL);
2253 if (!o) {
2254 ti->error = "Cannot allocate private origin structure";
2255 r = -ENOMEM;
2256 goto bad_alloc;
2259 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &o->dev);
2260 if (r) {
2261 ti->error = "Cannot get target device";
2262 goto bad_open;
2265 o->ti = ti;
2266 ti->private = o;
2267 ti->num_flush_bios = 1;
2269 return 0;
2271 bad_open:
2272 kfree(o);
2273 bad_alloc:
2274 return r;
2277 static void origin_dtr(struct dm_target *ti)
2279 struct dm_origin *o = ti->private;
2281 dm_put_device(ti, o->dev);
2282 kfree(o);
2285 static int origin_map(struct dm_target *ti, struct bio *bio)
2287 struct dm_origin *o = ti->private;
2288 unsigned available_sectors;
2290 bio_set_dev(bio, o->dev->bdev);
2292 if (unlikely(bio->bi_opf & REQ_PREFLUSH))
2293 return DM_MAPIO_REMAPPED;
2295 if (bio_data_dir(bio) != WRITE)
2296 return DM_MAPIO_REMAPPED;
2298 available_sectors = o->split_boundary -
2299 ((unsigned)bio->bi_iter.bi_sector & (o->split_boundary - 1));
2301 if (bio_sectors(bio) > available_sectors)
2302 dm_accept_partial_bio(bio, available_sectors);
2304 /* Only tell snapshots if this is a write */
2305 return do_origin(o->dev, bio);
2308 static long origin_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
2309 long nr_pages, void **kaddr, pfn_t *pfn)
2311 DMWARN("device does not support dax.");
2312 return -EIO;
2316 * Set the target "max_io_len" field to the minimum of all the snapshots'
2317 * chunk sizes.
2319 static void origin_resume(struct dm_target *ti)
2321 struct dm_origin *o = ti->private;
2323 o->split_boundary = get_origin_minimum_chunksize(o->dev->bdev);
2325 down_write(&_origins_lock);
2326 __insert_dm_origin(o);
2327 up_write(&_origins_lock);
2330 static void origin_postsuspend(struct dm_target *ti)
2332 struct dm_origin *o = ti->private;
2334 down_write(&_origins_lock);
2335 __remove_dm_origin(o);
2336 up_write(&_origins_lock);
2339 static void origin_status(struct dm_target *ti, status_type_t type,
2340 unsigned status_flags, char *result, unsigned maxlen)
2342 struct dm_origin *o = ti->private;
2344 switch (type) {
2345 case STATUSTYPE_INFO:
2346 result[0] = '\0';
2347 break;
2349 case STATUSTYPE_TABLE:
2350 snprintf(result, maxlen, "%s", o->dev->name);
2351 break;
2355 static int origin_iterate_devices(struct dm_target *ti,
2356 iterate_devices_callout_fn fn, void *data)
2358 struct dm_origin *o = ti->private;
2360 return fn(ti, o->dev, 0, ti->len, data);
2363 static struct target_type origin_target = {
2364 .name = "snapshot-origin",
2365 .version = {1, 9, 0},
2366 .module = THIS_MODULE,
2367 .ctr = origin_ctr,
2368 .dtr = origin_dtr,
2369 .map = origin_map,
2370 .resume = origin_resume,
2371 .postsuspend = origin_postsuspend,
2372 .status = origin_status,
2373 .iterate_devices = origin_iterate_devices,
2374 .direct_access = origin_dax_direct_access,
2377 static struct target_type snapshot_target = {
2378 .name = "snapshot",
2379 .version = {1, 15, 0},
2380 .module = THIS_MODULE,
2381 .ctr = snapshot_ctr,
2382 .dtr = snapshot_dtr,
2383 .map = snapshot_map,
2384 .end_io = snapshot_end_io,
2385 .preresume = snapshot_preresume,
2386 .resume = snapshot_resume,
2387 .status = snapshot_status,
2388 .iterate_devices = snapshot_iterate_devices,
2391 static struct target_type merge_target = {
2392 .name = dm_snapshot_merge_target_name,
2393 .version = {1, 4, 0},
2394 .module = THIS_MODULE,
2395 .ctr = snapshot_ctr,
2396 .dtr = snapshot_dtr,
2397 .map = snapshot_merge_map,
2398 .end_io = snapshot_end_io,
2399 .presuspend = snapshot_merge_presuspend,
2400 .preresume = snapshot_preresume,
2401 .resume = snapshot_merge_resume,
2402 .status = snapshot_status,
2403 .iterate_devices = snapshot_iterate_devices,
2406 static int __init dm_snapshot_init(void)
2408 int r;
2410 r = dm_exception_store_init();
2411 if (r) {
2412 DMERR("Failed to initialize exception stores");
2413 return r;
2416 r = init_origin_hash();
2417 if (r) {
2418 DMERR("init_origin_hash failed.");
2419 goto bad_origin_hash;
2422 exception_cache = KMEM_CACHE(dm_exception, 0);
2423 if (!exception_cache) {
2424 DMERR("Couldn't create exception cache.");
2425 r = -ENOMEM;
2426 goto bad_exception_cache;
2429 pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2430 if (!pending_cache) {
2431 DMERR("Couldn't create pending cache.");
2432 r = -ENOMEM;
2433 goto bad_pending_cache;
2436 r = dm_register_target(&snapshot_target);
2437 if (r < 0) {
2438 DMERR("snapshot target register failed %d", r);
2439 goto bad_register_snapshot_target;
2442 r = dm_register_target(&origin_target);
2443 if (r < 0) {
2444 DMERR("Origin target register failed %d", r);
2445 goto bad_register_origin_target;
2448 r = dm_register_target(&merge_target);
2449 if (r < 0) {
2450 DMERR("Merge target register failed %d", r);
2451 goto bad_register_merge_target;
2454 return 0;
2456 bad_register_merge_target:
2457 dm_unregister_target(&origin_target);
2458 bad_register_origin_target:
2459 dm_unregister_target(&snapshot_target);
2460 bad_register_snapshot_target:
2461 kmem_cache_destroy(pending_cache);
2462 bad_pending_cache:
2463 kmem_cache_destroy(exception_cache);
2464 bad_exception_cache:
2465 exit_origin_hash();
2466 bad_origin_hash:
2467 dm_exception_store_exit();
2469 return r;
2472 static void __exit dm_snapshot_exit(void)
2474 dm_unregister_target(&snapshot_target);
2475 dm_unregister_target(&origin_target);
2476 dm_unregister_target(&merge_target);
2478 exit_origin_hash();
2479 kmem_cache_destroy(pending_cache);
2480 kmem_cache_destroy(exception_cache);
2482 dm_exception_store_exit();
2485 /* Module hooks */
2486 module_init(dm_snapshot_init);
2487 module_exit(dm_snapshot_exit);
2489 MODULE_DESCRIPTION(DM_NAME " snapshot target");
2490 MODULE_AUTHOR("Joe Thornber");
2491 MODULE_LICENSE("GPL");
2492 MODULE_ALIAS("dm-snapshot-origin");
2493 MODULE_ALIAS("dm-snapshot-merge");