2 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
4 * This file is released under the GPL.
7 #include <linux/blkdev.h>
8 #include <linux/device-mapper.h>
9 #include <linux/delay.h>
11 #include <linux/init.h>
12 #include <linux/kdev_t.h>
13 #include <linux/list.h>
14 #include <linux/list_bl.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/slab.h>
18 #include <linux/vmalloc.h>
19 #include <linux/log2.h>
20 #include <linux/dm-kcopyd.h>
24 #include "dm-exception-store.h"
26 #define DM_MSG_PREFIX "snapshots"
28 static const char dm_snapshot_merge_target_name[] = "snapshot-merge";
30 #define dm_target_is_snapshot_merge(ti) \
31 ((ti)->type->name == dm_snapshot_merge_target_name)
34 * The size of the mempool used to track chunks in use.
38 #define DM_TRACKED_CHUNK_HASH_SIZE 16
39 #define DM_TRACKED_CHUNK_HASH(x) ((unsigned long)(x) & \
40 (DM_TRACKED_CHUNK_HASH_SIZE - 1))
42 struct dm_exception_table {
45 struct hlist_bl_head *table;
49 struct rw_semaphore lock;
51 struct dm_dev *origin;
56 /* List of snapshots per Origin */
57 struct list_head list;
60 * You can't use a snapshot if this is 0 (e.g. if full).
61 * A snapshot-merge target never clears this.
66 * The snapshot overflowed because of a write to the snapshot device.
67 * We don't have to invalidate the snapshot in this case, but we need
68 * to prevent further writes.
70 int snapshot_overflowed;
72 /* Origin writes don't trigger exceptions until this is set */
75 atomic_t pending_exceptions_count;
77 spinlock_t pe_allocation_lock;
79 /* Protected by "pe_allocation_lock" */
80 sector_t exception_start_sequence;
82 /* Protected by kcopyd single-threaded callback */
83 sector_t exception_complete_sequence;
86 * A list of pending exceptions that completed out of order.
87 * Protected by kcopyd single-threaded callback.
89 struct rb_root out_of_order_tree;
91 mempool_t pending_pool;
93 struct dm_exception_table pending;
94 struct dm_exception_table complete;
97 * pe_lock protects all pending_exception operations and access
98 * as well as the snapshot_bios list.
102 /* Chunks with outstanding reads */
103 spinlock_t tracked_chunk_lock;
104 struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE];
106 /* The on disk metadata handler */
107 struct dm_exception_store *store;
109 unsigned in_progress;
110 struct wait_queue_head in_progress_wait;
112 struct dm_kcopyd_client *kcopyd_client;
114 /* Wait for events based on state_bits */
115 unsigned long state_bits;
117 /* Range of chunks currently being merged. */
118 chunk_t first_merging_chunk;
119 int num_merging_chunks;
122 * The merge operation failed if this flag is set.
123 * Failure modes are handled as follows:
124 * - I/O error reading the header
125 * => don't load the target; abort.
126 * - Header does not have "valid" flag set
127 * => use the origin; forget about the snapshot.
128 * - I/O error when reading exceptions
129 * => don't load the target; abort.
130 * (We can't use the intermediate origin state.)
131 * - I/O error while merging
132 * => stop merging; set merge_failed; process I/O normally.
136 bool discard_zeroes_cow:1;
137 bool discard_passdown_origin:1;
140 * Incoming bios that overlap with chunks being merged must wait
141 * for them to be committed.
143 struct bio_list bios_queued_during_merge;
148 * RUNNING_MERGE - Merge operation is in progress.
149 * SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
150 * cleared afterwards.
152 #define RUNNING_MERGE 0
153 #define SHUTDOWN_MERGE 1
156 * Maximum number of chunks being copied on write.
158 * The value was decided experimentally as a trade-off between memory
159 * consumption, stalling the kernel's workqueues and maintaining a high enough
162 #define DEFAULT_COW_THRESHOLD 2048
164 static unsigned cow_threshold = DEFAULT_COW_THRESHOLD;
165 module_param_named(snapshot_cow_threshold, cow_threshold, uint, 0644);
166 MODULE_PARM_DESC(snapshot_cow_threshold, "Maximum number of chunks being copied on write");
168 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
169 "A percentage of time allocated for copy on write");
171 struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
175 EXPORT_SYMBOL(dm_snap_origin);
177 struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
181 EXPORT_SYMBOL(dm_snap_cow);
183 static sector_t chunk_to_sector(struct dm_exception_store *store,
186 return chunk << store->chunk_shift;
189 static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
192 * There is only ever one instance of a particular block
193 * device so we can compare pointers safely.
198 struct dm_snap_pending_exception {
199 struct dm_exception e;
202 * Origin buffers waiting for this to complete are held
205 struct bio_list origin_bios;
206 struct bio_list snapshot_bios;
208 /* Pointer back to snapshot context */
209 struct dm_snapshot *snap;
212 * 1 indicates the exception has already been sent to
217 /* There was copying error. */
220 /* A sequence number, it is used for in-order completion. */
221 sector_t exception_sequence;
223 struct rb_node out_of_order_node;
226 * For writing a complete chunk, bypassing the copy.
228 struct bio *full_bio;
229 bio_end_io_t *full_bio_end_io;
233 * Hash table mapping origin volumes to lists of snapshots and
234 * a lock to protect it
236 static struct kmem_cache *exception_cache;
237 static struct kmem_cache *pending_cache;
239 struct dm_snap_tracked_chunk {
240 struct hlist_node node;
244 static void init_tracked_chunk(struct bio *bio)
246 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
247 INIT_HLIST_NODE(&c->node);
250 static bool is_bio_tracked(struct bio *bio)
252 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
253 return !hlist_unhashed(&c->node);
256 static void track_chunk(struct dm_snapshot *s, struct bio *bio, chunk_t chunk)
258 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
262 spin_lock_irq(&s->tracked_chunk_lock);
263 hlist_add_head(&c->node,
264 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
265 spin_unlock_irq(&s->tracked_chunk_lock);
268 static void stop_tracking_chunk(struct dm_snapshot *s, struct bio *bio)
270 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
273 spin_lock_irqsave(&s->tracked_chunk_lock, flags);
275 spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
278 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
280 struct dm_snap_tracked_chunk *c;
283 spin_lock_irq(&s->tracked_chunk_lock);
285 hlist_for_each_entry(c,
286 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
287 if (c->chunk == chunk) {
293 spin_unlock_irq(&s->tracked_chunk_lock);
299 * This conflicting I/O is extremely improbable in the caller,
300 * so msleep(1) is sufficient and there is no need for a wait queue.
302 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
304 while (__chunk_is_tracked(s, chunk))
309 * One of these per registered origin, held in the snapshot_origins hash
312 /* The origin device */
313 struct block_device *bdev;
315 struct list_head hash_list;
317 /* List of snapshots for this origin */
318 struct list_head snapshots;
322 * This structure is allocated for each origin target
326 struct dm_target *ti;
327 unsigned split_boundary;
328 struct list_head hash_list;
332 * Size of the hash table for origin volumes. If we make this
333 * the size of the minors list then it should be nearly perfect
335 #define ORIGIN_HASH_SIZE 256
336 #define ORIGIN_MASK 0xFF
337 static struct list_head *_origins;
338 static struct list_head *_dm_origins;
339 static struct rw_semaphore _origins_lock;
341 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
342 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
343 static uint64_t _pending_exceptions_done_count;
345 static int init_origin_hash(void)
349 _origins = kmalloc_array(ORIGIN_HASH_SIZE, sizeof(struct list_head),
352 DMERR("unable to allocate memory for _origins");
355 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
356 INIT_LIST_HEAD(_origins + i);
358 _dm_origins = kmalloc_array(ORIGIN_HASH_SIZE,
359 sizeof(struct list_head),
362 DMERR("unable to allocate memory for _dm_origins");
366 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
367 INIT_LIST_HEAD(_dm_origins + i);
369 init_rwsem(&_origins_lock);
374 static void exit_origin_hash(void)
380 static unsigned origin_hash(struct block_device *bdev)
382 return bdev->bd_dev & ORIGIN_MASK;
385 static struct origin *__lookup_origin(struct block_device *origin)
387 struct list_head *ol;
390 ol = &_origins[origin_hash(origin)];
391 list_for_each_entry (o, ol, hash_list)
392 if (bdev_equal(o->bdev, origin))
398 static void __insert_origin(struct origin *o)
400 struct list_head *sl = &_origins[origin_hash(o->bdev)];
401 list_add_tail(&o->hash_list, sl);
404 static struct dm_origin *__lookup_dm_origin(struct block_device *origin)
406 struct list_head *ol;
409 ol = &_dm_origins[origin_hash(origin)];
410 list_for_each_entry (o, ol, hash_list)
411 if (bdev_equal(o->dev->bdev, origin))
417 static void __insert_dm_origin(struct dm_origin *o)
419 struct list_head *sl = &_dm_origins[origin_hash(o->dev->bdev)];
420 list_add_tail(&o->hash_list, sl);
423 static void __remove_dm_origin(struct dm_origin *o)
425 list_del(&o->hash_list);
429 * _origins_lock must be held when calling this function.
430 * Returns number of snapshots registered using the supplied cow device, plus:
431 * snap_src - a snapshot suitable for use as a source of exception handover
432 * snap_dest - a snapshot capable of receiving exception handover.
433 * snap_merge - an existing snapshot-merge target linked to the same origin.
434 * There can be at most one snapshot-merge target. The parameter is optional.
436 * Possible return values and states of snap_src and snap_dest.
437 * 0: NULL, NULL - first new snapshot
438 * 1: snap_src, NULL - normal snapshot
439 * 2: snap_src, snap_dest - waiting for handover
440 * 2: snap_src, NULL - handed over, waiting for old to be deleted
441 * 1: NULL, snap_dest - source got destroyed without handover
443 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
444 struct dm_snapshot **snap_src,
445 struct dm_snapshot **snap_dest,
446 struct dm_snapshot **snap_merge)
448 struct dm_snapshot *s;
453 o = __lookup_origin(snap->origin->bdev);
457 list_for_each_entry(s, &o->snapshots, list) {
458 if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
460 if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
470 } else if (snap_dest)
481 * On success, returns 1 if this snapshot is a handover destination,
482 * otherwise returns 0.
484 static int __validate_exception_handover(struct dm_snapshot *snap)
486 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
487 struct dm_snapshot *snap_merge = NULL;
489 /* Does snapshot need exceptions handed over to it? */
490 if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
491 &snap_merge) == 2) ||
493 snap->ti->error = "Snapshot cow pairing for exception "
494 "table handover failed";
499 * If no snap_src was found, snap cannot become a handover
506 * Non-snapshot-merge handover?
508 if (!dm_target_is_snapshot_merge(snap->ti))
512 * Do not allow more than one merging snapshot.
515 snap->ti->error = "A snapshot is already merging.";
519 if (!snap_src->store->type->prepare_merge ||
520 !snap_src->store->type->commit_merge) {
521 snap->ti->error = "Snapshot exception store does not "
522 "support snapshot-merge.";
529 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
531 struct dm_snapshot *l;
533 /* Sort the list according to chunk size, largest-first smallest-last */
534 list_for_each_entry(l, &o->snapshots, list)
535 if (l->store->chunk_size < s->store->chunk_size)
537 list_add_tail(&s->list, &l->list);
541 * Make a note of the snapshot and its origin so we can look it
542 * up when the origin has a write on it.
544 * Also validate snapshot exception store handovers.
545 * On success, returns 1 if this registration is a handover destination,
546 * otherwise returns 0.
548 static int register_snapshot(struct dm_snapshot *snap)
550 struct origin *o, *new_o = NULL;
551 struct block_device *bdev = snap->origin->bdev;
554 new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
558 down_write(&_origins_lock);
560 r = __validate_exception_handover(snap);
566 o = __lookup_origin(bdev);
573 /* Initialise the struct */
574 INIT_LIST_HEAD(&o->snapshots);
580 __insert_snapshot(o, snap);
583 up_write(&_origins_lock);
589 * Move snapshot to correct place in list according to chunk size.
591 static void reregister_snapshot(struct dm_snapshot *s)
593 struct block_device *bdev = s->origin->bdev;
595 down_write(&_origins_lock);
598 __insert_snapshot(__lookup_origin(bdev), s);
600 up_write(&_origins_lock);
603 static void unregister_snapshot(struct dm_snapshot *s)
607 down_write(&_origins_lock);
608 o = __lookup_origin(s->origin->bdev);
611 if (o && list_empty(&o->snapshots)) {
612 list_del(&o->hash_list);
616 up_write(&_origins_lock);
620 * Implementation of the exception hash tables.
621 * The lowest hash_shift bits of the chunk number are ignored, allowing
622 * some consecutive chunks to be grouped together.
624 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk);
626 /* Lock to protect access to the completed and pending exception hash tables. */
627 struct dm_exception_table_lock {
628 struct hlist_bl_head *complete_slot;
629 struct hlist_bl_head *pending_slot;
632 static void dm_exception_table_lock_init(struct dm_snapshot *s, chunk_t chunk,
633 struct dm_exception_table_lock *lock)
635 struct dm_exception_table *complete = &s->complete;
636 struct dm_exception_table *pending = &s->pending;
638 lock->complete_slot = &complete->table[exception_hash(complete, chunk)];
639 lock->pending_slot = &pending->table[exception_hash(pending, chunk)];
642 static void dm_exception_table_lock(struct dm_exception_table_lock *lock)
644 hlist_bl_lock(lock->complete_slot);
645 hlist_bl_lock(lock->pending_slot);
648 static void dm_exception_table_unlock(struct dm_exception_table_lock *lock)
650 hlist_bl_unlock(lock->pending_slot);
651 hlist_bl_unlock(lock->complete_slot);
654 static int dm_exception_table_init(struct dm_exception_table *et,
655 uint32_t size, unsigned hash_shift)
659 et->hash_shift = hash_shift;
660 et->hash_mask = size - 1;
661 et->table = dm_vcalloc(size, sizeof(struct hlist_bl_head));
665 for (i = 0; i < size; i++)
666 INIT_HLIST_BL_HEAD(et->table + i);
671 static void dm_exception_table_exit(struct dm_exception_table *et,
672 struct kmem_cache *mem)
674 struct hlist_bl_head *slot;
675 struct dm_exception *ex;
676 struct hlist_bl_node *pos, *n;
679 size = et->hash_mask + 1;
680 for (i = 0; i < size; i++) {
681 slot = et->table + i;
683 hlist_bl_for_each_entry_safe(ex, pos, n, slot, hash_list)
684 kmem_cache_free(mem, ex);
690 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
692 return (chunk >> et->hash_shift) & et->hash_mask;
695 static void dm_remove_exception(struct dm_exception *e)
697 hlist_bl_del(&e->hash_list);
701 * Return the exception data for a sector, or NULL if not
704 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
707 struct hlist_bl_head *slot;
708 struct hlist_bl_node *pos;
709 struct dm_exception *e;
711 slot = &et->table[exception_hash(et, chunk)];
712 hlist_bl_for_each_entry(e, pos, slot, hash_list)
713 if (chunk >= e->old_chunk &&
714 chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
720 static struct dm_exception *alloc_completed_exception(gfp_t gfp)
722 struct dm_exception *e;
724 e = kmem_cache_alloc(exception_cache, gfp);
725 if (!e && gfp == GFP_NOIO)
726 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
731 static void free_completed_exception(struct dm_exception *e)
733 kmem_cache_free(exception_cache, e);
736 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
738 struct dm_snap_pending_exception *pe = mempool_alloc(&s->pending_pool,
741 atomic_inc(&s->pending_exceptions_count);
747 static void free_pending_exception(struct dm_snap_pending_exception *pe)
749 struct dm_snapshot *s = pe->snap;
751 mempool_free(pe, &s->pending_pool);
752 smp_mb__before_atomic();
753 atomic_dec(&s->pending_exceptions_count);
756 static void dm_insert_exception(struct dm_exception_table *eh,
757 struct dm_exception *new_e)
759 struct hlist_bl_head *l;
760 struct hlist_bl_node *pos;
761 struct dm_exception *e = NULL;
763 l = &eh->table[exception_hash(eh, new_e->old_chunk)];
765 /* Add immediately if this table doesn't support consecutive chunks */
769 /* List is ordered by old_chunk */
770 hlist_bl_for_each_entry(e, pos, l, hash_list) {
771 /* Insert after an existing chunk? */
772 if (new_e->old_chunk == (e->old_chunk +
773 dm_consecutive_chunk_count(e) + 1) &&
774 new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
775 dm_consecutive_chunk_count(e) + 1)) {
776 dm_consecutive_chunk_count_inc(e);
777 free_completed_exception(new_e);
781 /* Insert before an existing chunk? */
782 if (new_e->old_chunk == (e->old_chunk - 1) &&
783 new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
784 dm_consecutive_chunk_count_inc(e);
787 free_completed_exception(new_e);
791 if (new_e->old_chunk < e->old_chunk)
798 * Either the table doesn't support consecutive chunks or slot
801 hlist_bl_add_head(&new_e->hash_list, l);
802 } else if (new_e->old_chunk < e->old_chunk) {
803 /* Add before an existing exception */
804 hlist_bl_add_before(&new_e->hash_list, &e->hash_list);
806 /* Add to l's tail: e is the last exception in this slot */
807 hlist_bl_add_behind(&new_e->hash_list, &e->hash_list);
812 * Callback used by the exception stores to load exceptions when
815 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
817 struct dm_exception_table_lock lock;
818 struct dm_snapshot *s = context;
819 struct dm_exception *e;
821 e = alloc_completed_exception(GFP_KERNEL);
827 /* Consecutive_count is implicitly initialised to zero */
831 * Although there is no need to lock access to the exception tables
832 * here, if we don't then hlist_bl_add_head(), called by
833 * dm_insert_exception(), will complain about accessing the
834 * corresponding list without locking it first.
836 dm_exception_table_lock_init(s, old, &lock);
838 dm_exception_table_lock(&lock);
839 dm_insert_exception(&s->complete, e);
840 dm_exception_table_unlock(&lock);
846 * Return a minimum chunk size of all snapshots that have the specified origin.
847 * Return zero if the origin has no snapshots.
849 static uint32_t __minimum_chunk_size(struct origin *o)
851 struct dm_snapshot *snap;
852 unsigned chunk_size = 0;
855 list_for_each_entry(snap, &o->snapshots, list)
856 chunk_size = min_not_zero(chunk_size,
857 snap->store->chunk_size);
859 return (uint32_t) chunk_size;
865 static int calc_max_buckets(void)
867 /* use a fixed size of 2MB */
868 unsigned long mem = 2 * 1024 * 1024;
869 mem /= sizeof(struct hlist_bl_head);
875 * Allocate room for a suitable hash table.
877 static int init_hash_tables(struct dm_snapshot *s)
879 sector_t hash_size, cow_dev_size, max_buckets;
882 * Calculate based on the size of the original volume or
885 cow_dev_size = get_dev_size(s->cow->bdev);
886 max_buckets = calc_max_buckets();
888 hash_size = cow_dev_size >> s->store->chunk_shift;
889 hash_size = min(hash_size, max_buckets);
893 hash_size = rounddown_pow_of_two(hash_size);
894 if (dm_exception_table_init(&s->complete, hash_size,
895 DM_CHUNK_CONSECUTIVE_BITS))
899 * Allocate hash table for in-flight exceptions
900 * Make this smaller than the real hash table
906 if (dm_exception_table_init(&s->pending, hash_size, 0)) {
907 dm_exception_table_exit(&s->complete, exception_cache);
914 static void merge_shutdown(struct dm_snapshot *s)
916 clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
917 smp_mb__after_atomic();
918 wake_up_bit(&s->state_bits, RUNNING_MERGE);
921 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
923 s->first_merging_chunk = 0;
924 s->num_merging_chunks = 0;
926 return bio_list_get(&s->bios_queued_during_merge);
930 * Remove one chunk from the index of completed exceptions.
932 static int __remove_single_exception_chunk(struct dm_snapshot *s,
935 struct dm_exception *e;
937 e = dm_lookup_exception(&s->complete, old_chunk);
939 DMERR("Corruption detected: exception for block %llu is "
940 "on disk but not in memory",
941 (unsigned long long)old_chunk);
946 * If this is the only chunk using this exception, remove exception.
948 if (!dm_consecutive_chunk_count(e)) {
949 dm_remove_exception(e);
950 free_completed_exception(e);
955 * The chunk may be either at the beginning or the end of a
956 * group of consecutive chunks - never in the middle. We are
957 * removing chunks in the opposite order to that in which they
958 * were added, so this should always be true.
959 * Decrement the consecutive chunk counter and adjust the
960 * starting point if necessary.
962 if (old_chunk == e->old_chunk) {
965 } else if (old_chunk != e->old_chunk +
966 dm_consecutive_chunk_count(e)) {
967 DMERR("Attempt to merge block %llu from the "
968 "middle of a chunk range [%llu - %llu]",
969 (unsigned long long)old_chunk,
970 (unsigned long long)e->old_chunk,
972 e->old_chunk + dm_consecutive_chunk_count(e));
976 dm_consecutive_chunk_count_dec(e);
981 static void flush_bios(struct bio *bio);
983 static int remove_single_exception_chunk(struct dm_snapshot *s)
985 struct bio *b = NULL;
987 chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
989 down_write(&s->lock);
992 * Process chunks (and associated exceptions) in reverse order
993 * so that dm_consecutive_chunk_count_dec() accounting works.
996 r = __remove_single_exception_chunk(s, old_chunk);
999 } while (old_chunk-- > s->first_merging_chunk);
1001 b = __release_queued_bios_after_merge(s);
1011 static int origin_write_extent(struct dm_snapshot *merging_snap,
1012 sector_t sector, unsigned chunk_size);
1014 static void merge_callback(int read_err, unsigned long write_err,
1017 static uint64_t read_pending_exceptions_done_count(void)
1019 uint64_t pending_exceptions_done;
1021 spin_lock(&_pending_exceptions_done_spinlock);
1022 pending_exceptions_done = _pending_exceptions_done_count;
1023 spin_unlock(&_pending_exceptions_done_spinlock);
1025 return pending_exceptions_done;
1028 static void increment_pending_exceptions_done_count(void)
1030 spin_lock(&_pending_exceptions_done_spinlock);
1031 _pending_exceptions_done_count++;
1032 spin_unlock(&_pending_exceptions_done_spinlock);
1034 wake_up_all(&_pending_exceptions_done);
1037 static void snapshot_merge_next_chunks(struct dm_snapshot *s)
1039 int i, linear_chunks;
1040 chunk_t old_chunk, new_chunk;
1041 struct dm_io_region src, dest;
1043 uint64_t previous_count;
1045 BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
1046 if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
1050 * valid flag never changes during merge, so no lock required.
1053 DMERR("Snapshot is invalid: can't merge");
1057 linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
1059 if (linear_chunks <= 0) {
1060 if (linear_chunks < 0) {
1061 DMERR("Read error in exception store: "
1062 "shutting down merge");
1063 down_write(&s->lock);
1064 s->merge_failed = 1;
1070 /* Adjust old_chunk and new_chunk to reflect start of linear region */
1071 old_chunk = old_chunk + 1 - linear_chunks;
1072 new_chunk = new_chunk + 1 - linear_chunks;
1075 * Use one (potentially large) I/O to copy all 'linear_chunks'
1076 * from the exception store to the origin
1078 io_size = linear_chunks * s->store->chunk_size;
1080 dest.bdev = s->origin->bdev;
1081 dest.sector = chunk_to_sector(s->store, old_chunk);
1082 dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
1084 src.bdev = s->cow->bdev;
1085 src.sector = chunk_to_sector(s->store, new_chunk);
1086 src.count = dest.count;
1089 * Reallocate any exceptions needed in other snapshots then
1090 * wait for the pending exceptions to complete.
1091 * Each time any pending exception (globally on the system)
1092 * completes we are woken and repeat the process to find out
1093 * if we can proceed. While this may not seem a particularly
1094 * efficient algorithm, it is not expected to have any
1095 * significant impact on performance.
1097 previous_count = read_pending_exceptions_done_count();
1098 while (origin_write_extent(s, dest.sector, io_size)) {
1099 wait_event(_pending_exceptions_done,
1100 (read_pending_exceptions_done_count() !=
1102 /* Retry after the wait, until all exceptions are done. */
1103 previous_count = read_pending_exceptions_done_count();
1106 down_write(&s->lock);
1107 s->first_merging_chunk = old_chunk;
1108 s->num_merging_chunks = linear_chunks;
1111 /* Wait until writes to all 'linear_chunks' drain */
1112 for (i = 0; i < linear_chunks; i++)
1113 __check_for_conflicting_io(s, old_chunk + i);
1115 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
1122 static void error_bios(struct bio *bio);
1124 static void merge_callback(int read_err, unsigned long write_err, void *context)
1126 struct dm_snapshot *s = context;
1127 struct bio *b = NULL;
1129 if (read_err || write_err) {
1131 DMERR("Read error: shutting down merge.");
1133 DMERR("Write error: shutting down merge.");
1137 if (s->store->type->commit_merge(s->store,
1138 s->num_merging_chunks) < 0) {
1139 DMERR("Write error in exception store: shutting down merge");
1143 if (remove_single_exception_chunk(s) < 0)
1146 snapshot_merge_next_chunks(s);
1151 down_write(&s->lock);
1152 s->merge_failed = 1;
1153 b = __release_queued_bios_after_merge(s);
1160 static void start_merge(struct dm_snapshot *s)
1162 if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1163 snapshot_merge_next_chunks(s);
1167 * Stop the merging process and wait until it finishes.
1169 static void stop_merge(struct dm_snapshot *s)
1171 set_bit(SHUTDOWN_MERGE, &s->state_bits);
1172 wait_on_bit(&s->state_bits, RUNNING_MERGE, TASK_UNINTERRUPTIBLE);
1173 clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1176 static int parse_snapshot_features(struct dm_arg_set *as, struct dm_snapshot *s,
1177 struct dm_target *ti)
1181 const char *arg_name;
1183 static const struct dm_arg _args[] = {
1184 {0, 2, "Invalid number of feature arguments"},
1188 * No feature arguments supplied.
1193 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1197 while (argc && !r) {
1198 arg_name = dm_shift_arg(as);
1201 if (!strcasecmp(arg_name, "discard_zeroes_cow"))
1202 s->discard_zeroes_cow = true;
1204 else if (!strcasecmp(arg_name, "discard_passdown_origin"))
1205 s->discard_passdown_origin = true;
1208 ti->error = "Unrecognised feature requested";
1214 if (!s->discard_zeroes_cow && s->discard_passdown_origin) {
1216 * TODO: really these are disjoint.. but ti->num_discard_bios
1217 * and dm_bio_get_target_bio_nr() require rigid constraints.
1219 ti->error = "discard_passdown_origin feature depends on discard_zeroes_cow";
1227 * Construct a snapshot mapping:
1228 * <origin_dev> <COW-dev> <p|po|n> <chunk-size> [<# feature args> [<arg>]*]
1230 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1232 struct dm_snapshot *s;
1233 struct dm_arg_set as;
1236 char *origin_path, *cow_path;
1237 dev_t origin_dev, cow_dev;
1238 unsigned args_used, num_flush_bios = 1;
1239 fmode_t origin_mode = FMODE_READ;
1242 ti->error = "requires 4 or more arguments";
1247 if (dm_target_is_snapshot_merge(ti)) {
1249 origin_mode = FMODE_WRITE;
1252 s = kzalloc(sizeof(*s), GFP_KERNEL);
1254 ti->error = "Cannot allocate private snapshot structure";
1261 dm_consume_args(&as, 4);
1262 r = parse_snapshot_features(&as, s, ti);
1266 origin_path = argv[0];
1270 r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1272 ti->error = "Cannot get origin device";
1275 origin_dev = s->origin->bdev->bd_dev;
1281 cow_dev = dm_get_dev_t(cow_path);
1282 if (cow_dev && cow_dev == origin_dev) {
1283 ti->error = "COW device cannot be the same as origin device";
1288 r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1290 ti->error = "Cannot get COW device";
1294 r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1296 ti->error = "Couldn't create exception store";
1306 s->snapshot_overflowed = 0;
1308 atomic_set(&s->pending_exceptions_count, 0);
1309 spin_lock_init(&s->pe_allocation_lock);
1310 s->exception_start_sequence = 0;
1311 s->exception_complete_sequence = 0;
1312 s->out_of_order_tree = RB_ROOT;
1313 init_rwsem(&s->lock);
1314 INIT_LIST_HEAD(&s->list);
1315 spin_lock_init(&s->pe_lock);
1317 s->merge_failed = 0;
1318 s->first_merging_chunk = 0;
1319 s->num_merging_chunks = 0;
1320 bio_list_init(&s->bios_queued_during_merge);
1322 /* Allocate hash table for COW data */
1323 if (init_hash_tables(s)) {
1324 ti->error = "Unable to allocate hash table space";
1326 goto bad_hash_tables;
1329 init_waitqueue_head(&s->in_progress_wait);
1331 s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1332 if (IS_ERR(s->kcopyd_client)) {
1333 r = PTR_ERR(s->kcopyd_client);
1334 ti->error = "Could not create kcopyd client";
1338 r = mempool_init_slab_pool(&s->pending_pool, MIN_IOS, pending_cache);
1340 ti->error = "Could not allocate mempool for pending exceptions";
1341 goto bad_pending_pool;
1344 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1345 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1347 spin_lock_init(&s->tracked_chunk_lock);
1350 ti->num_flush_bios = num_flush_bios;
1351 if (s->discard_zeroes_cow)
1352 ti->num_discard_bios = (s->discard_passdown_origin ? 2 : 1);
1353 ti->per_io_data_size = sizeof(struct dm_snap_tracked_chunk);
1355 /* Add snapshot to the list of snapshots for this origin */
1356 /* Exceptions aren't triggered till snapshot_resume() is called */
1357 r = register_snapshot(s);
1359 ti->error = "Snapshot origin struct allocation failed";
1360 goto bad_load_and_register;
1362 /* invalid handover, register_snapshot has set ti->error */
1363 goto bad_load_and_register;
1367 * Metadata must only be loaded into one table at once, so skip this
1368 * if metadata will be handed over during resume.
1369 * Chunk size will be set during the handover - set it to zero to
1370 * ensure it's ignored.
1373 s->store->chunk_size = 0;
1377 r = s->store->type->read_metadata(s->store, dm_add_exception,
1380 ti->error = "Failed to read snapshot metadata";
1381 goto bad_read_metadata;
1384 DMWARN("Snapshot is marked invalid.");
1387 if (!s->store->chunk_size) {
1388 ti->error = "Chunk size not set";
1389 goto bad_read_metadata;
1392 r = dm_set_target_max_io_len(ti, s->store->chunk_size);
1394 goto bad_read_metadata;
1399 unregister_snapshot(s);
1400 bad_load_and_register:
1401 mempool_exit(&s->pending_pool);
1403 dm_kcopyd_client_destroy(s->kcopyd_client);
1405 dm_exception_table_exit(&s->pending, pending_cache);
1406 dm_exception_table_exit(&s->complete, exception_cache);
1408 dm_exception_store_destroy(s->store);
1410 dm_put_device(ti, s->cow);
1412 dm_put_device(ti, s->origin);
1420 static void __free_exceptions(struct dm_snapshot *s)
1422 dm_kcopyd_client_destroy(s->kcopyd_client);
1423 s->kcopyd_client = NULL;
1425 dm_exception_table_exit(&s->pending, pending_cache);
1426 dm_exception_table_exit(&s->complete, exception_cache);
1429 static void __handover_exceptions(struct dm_snapshot *snap_src,
1430 struct dm_snapshot *snap_dest)
1433 struct dm_exception_table table_swap;
1434 struct dm_exception_store *store_swap;
1438 * Swap all snapshot context information between the two instances.
1440 u.table_swap = snap_dest->complete;
1441 snap_dest->complete = snap_src->complete;
1442 snap_src->complete = u.table_swap;
1444 u.store_swap = snap_dest->store;
1445 snap_dest->store = snap_src->store;
1446 snap_dest->store->userspace_supports_overflow = u.store_swap->userspace_supports_overflow;
1447 snap_src->store = u.store_swap;
1449 snap_dest->store->snap = snap_dest;
1450 snap_src->store->snap = snap_src;
1452 snap_dest->ti->max_io_len = snap_dest->store->chunk_size;
1453 snap_dest->valid = snap_src->valid;
1454 snap_dest->snapshot_overflowed = snap_src->snapshot_overflowed;
1457 * Set source invalid to ensure it receives no further I/O.
1459 snap_src->valid = 0;
1462 static void snapshot_dtr(struct dm_target *ti)
1464 #ifdef CONFIG_DM_DEBUG
1467 struct dm_snapshot *s = ti->private;
1468 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1470 down_read(&_origins_lock);
1471 /* Check whether exception handover must be cancelled */
1472 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1473 if (snap_src && snap_dest && (s == snap_src)) {
1474 down_write(&snap_dest->lock);
1475 snap_dest->valid = 0;
1476 up_write(&snap_dest->lock);
1477 DMERR("Cancelling snapshot handover.");
1479 up_read(&_origins_lock);
1481 if (dm_target_is_snapshot_merge(ti))
1484 /* Prevent further origin writes from using this snapshot. */
1485 /* After this returns there can be no new kcopyd jobs. */
1486 unregister_snapshot(s);
1488 while (atomic_read(&s->pending_exceptions_count))
1491 * Ensure instructions in mempool_exit aren't reordered
1492 * before atomic_read.
1496 #ifdef CONFIG_DM_DEBUG
1497 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1498 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1501 __free_exceptions(s);
1503 mempool_exit(&s->pending_pool);
1505 dm_exception_store_destroy(s->store);
1507 dm_put_device(ti, s->cow);
1509 dm_put_device(ti, s->origin);
1511 WARN_ON(s->in_progress);
1516 static void account_start_copy(struct dm_snapshot *s)
1518 spin_lock(&s->in_progress_wait.lock);
1520 spin_unlock(&s->in_progress_wait.lock);
1523 static void account_end_copy(struct dm_snapshot *s)
1525 spin_lock(&s->in_progress_wait.lock);
1526 BUG_ON(!s->in_progress);
1528 if (likely(s->in_progress <= cow_threshold) &&
1529 unlikely(waitqueue_active(&s->in_progress_wait)))
1530 wake_up_locked(&s->in_progress_wait);
1531 spin_unlock(&s->in_progress_wait.lock);
1534 static bool wait_for_in_progress(struct dm_snapshot *s, bool unlock_origins)
1536 if (unlikely(s->in_progress > cow_threshold)) {
1537 spin_lock(&s->in_progress_wait.lock);
1538 if (likely(s->in_progress > cow_threshold)) {
1540 * NOTE: this throttle doesn't account for whether
1541 * the caller is servicing an IO that will trigger a COW
1542 * so excess throttling may result for chunks not required
1543 * to be COW'd. But if cow_threshold was reached, extra
1544 * throttling is unlikely to negatively impact performance.
1546 DECLARE_WAITQUEUE(wait, current);
1547 __add_wait_queue(&s->in_progress_wait, &wait);
1548 __set_current_state(TASK_UNINTERRUPTIBLE);
1549 spin_unlock(&s->in_progress_wait.lock);
1551 up_read(&_origins_lock);
1553 remove_wait_queue(&s->in_progress_wait, &wait);
1556 spin_unlock(&s->in_progress_wait.lock);
1562 * Flush a list of buffers.
1564 static void flush_bios(struct bio *bio)
1570 bio->bi_next = NULL;
1571 generic_make_request(bio);
1576 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit);
1579 * Flush a list of buffers.
1581 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1588 bio->bi_next = NULL;
1589 r = do_origin(s->origin, bio, false);
1590 if (r == DM_MAPIO_REMAPPED)
1591 generic_make_request(bio);
1597 * Error a list of buffers.
1599 static void error_bios(struct bio *bio)
1605 bio->bi_next = NULL;
1611 static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1617 DMERR("Invalidating snapshot: Error reading/writing.");
1618 else if (err == -ENOMEM)
1619 DMERR("Invalidating snapshot: Unable to allocate exception.");
1621 if (s->store->type->drop_snapshot)
1622 s->store->type->drop_snapshot(s->store);
1626 dm_table_event(s->ti->table);
1629 static void invalidate_snapshot(struct dm_snapshot *s, int err)
1631 down_write(&s->lock);
1632 __invalidate_snapshot(s, err);
1636 static void pending_complete(void *context, int success)
1638 struct dm_snap_pending_exception *pe = context;
1639 struct dm_exception *e;
1640 struct dm_snapshot *s = pe->snap;
1641 struct bio *origin_bios = NULL;
1642 struct bio *snapshot_bios = NULL;
1643 struct bio *full_bio = NULL;
1644 struct dm_exception_table_lock lock;
1647 dm_exception_table_lock_init(s, pe->e.old_chunk, &lock);
1650 /* Read/write error - snapshot is unusable */
1651 invalidate_snapshot(s, -EIO);
1654 dm_exception_table_lock(&lock);
1658 e = alloc_completed_exception(GFP_NOIO);
1660 invalidate_snapshot(s, -ENOMEM);
1663 dm_exception_table_lock(&lock);
1668 down_read(&s->lock);
1669 dm_exception_table_lock(&lock);
1672 free_completed_exception(e);
1679 * Add a proper exception. After inserting the completed exception all
1680 * subsequent snapshot reads to this chunk will be redirected to the
1681 * COW device. This ensures that we do not starve. Moreover, as long
1682 * as the pending exception exists, neither origin writes nor snapshot
1683 * merging can overwrite the chunk in origin.
1685 dm_insert_exception(&s->complete, e);
1688 /* Wait for conflicting reads to drain */
1689 if (__chunk_is_tracked(s, pe->e.old_chunk)) {
1690 dm_exception_table_unlock(&lock);
1691 __check_for_conflicting_io(s, pe->e.old_chunk);
1692 dm_exception_table_lock(&lock);
1696 /* Remove the in-flight exception from the list */
1697 dm_remove_exception(&pe->e);
1699 dm_exception_table_unlock(&lock);
1701 snapshot_bios = bio_list_get(&pe->snapshot_bios);
1702 origin_bios = bio_list_get(&pe->origin_bios);
1703 full_bio = pe->full_bio;
1705 full_bio->bi_end_io = pe->full_bio_end_io;
1706 increment_pending_exceptions_done_count();
1708 /* Submit any pending write bios */
1711 bio_io_error(full_bio);
1712 error_bios(snapshot_bios);
1715 bio_endio(full_bio);
1716 flush_bios(snapshot_bios);
1719 retry_origin_bios(s, origin_bios);
1721 free_pending_exception(pe);
1724 static void complete_exception(struct dm_snap_pending_exception *pe)
1726 struct dm_snapshot *s = pe->snap;
1728 /* Update the metadata if we are persistent */
1729 s->store->type->commit_exception(s->store, &pe->e, !pe->copy_error,
1730 pending_complete, pe);
1734 * Called when the copy I/O has finished. kcopyd actually runs
1735 * this code so don't block.
1737 static void copy_callback(int read_err, unsigned long write_err, void *context)
1739 struct dm_snap_pending_exception *pe = context;
1740 struct dm_snapshot *s = pe->snap;
1742 pe->copy_error = read_err || write_err;
1744 if (pe->exception_sequence == s->exception_complete_sequence) {
1745 struct rb_node *next;
1747 s->exception_complete_sequence++;
1748 complete_exception(pe);
1750 next = rb_first(&s->out_of_order_tree);
1752 pe = rb_entry(next, struct dm_snap_pending_exception,
1754 if (pe->exception_sequence != s->exception_complete_sequence)
1756 next = rb_next(next);
1757 s->exception_complete_sequence++;
1758 rb_erase(&pe->out_of_order_node, &s->out_of_order_tree);
1759 complete_exception(pe);
1763 struct rb_node *parent = NULL;
1764 struct rb_node **p = &s->out_of_order_tree.rb_node;
1765 struct dm_snap_pending_exception *pe2;
1768 pe2 = rb_entry(*p, struct dm_snap_pending_exception, out_of_order_node);
1771 BUG_ON(pe->exception_sequence == pe2->exception_sequence);
1772 if (pe->exception_sequence < pe2->exception_sequence)
1773 p = &((*p)->rb_left);
1775 p = &((*p)->rb_right);
1778 rb_link_node(&pe->out_of_order_node, parent, p);
1779 rb_insert_color(&pe->out_of_order_node, &s->out_of_order_tree);
1781 account_end_copy(s);
1785 * Dispatches the copy operation to kcopyd.
1787 static void start_copy(struct dm_snap_pending_exception *pe)
1789 struct dm_snapshot *s = pe->snap;
1790 struct dm_io_region src, dest;
1791 struct block_device *bdev = s->origin->bdev;
1794 dev_size = get_dev_size(bdev);
1797 src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1798 src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1800 dest.bdev = s->cow->bdev;
1801 dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1802 dest.count = src.count;
1804 /* Hand over to kcopyd */
1805 account_start_copy(s);
1806 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1809 static void full_bio_end_io(struct bio *bio)
1811 void *callback_data = bio->bi_private;
1813 dm_kcopyd_do_callback(callback_data, 0, bio->bi_status ? 1 : 0);
1816 static void start_full_bio(struct dm_snap_pending_exception *pe,
1819 struct dm_snapshot *s = pe->snap;
1820 void *callback_data;
1823 pe->full_bio_end_io = bio->bi_end_io;
1825 account_start_copy(s);
1826 callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1829 bio->bi_end_io = full_bio_end_io;
1830 bio->bi_private = callback_data;
1832 generic_make_request(bio);
1835 static struct dm_snap_pending_exception *
1836 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1838 struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1843 return container_of(e, struct dm_snap_pending_exception, e);
1847 * Inserts a pending exception into the pending table.
1849 * NOTE: a write lock must be held on the chunk's pending exception table slot
1850 * before calling this.
1852 static struct dm_snap_pending_exception *
1853 __insert_pending_exception(struct dm_snapshot *s,
1854 struct dm_snap_pending_exception *pe, chunk_t chunk)
1856 pe->e.old_chunk = chunk;
1857 bio_list_init(&pe->origin_bios);
1858 bio_list_init(&pe->snapshot_bios);
1860 pe->full_bio = NULL;
1862 spin_lock(&s->pe_allocation_lock);
1863 if (s->store->type->prepare_exception(s->store, &pe->e)) {
1864 spin_unlock(&s->pe_allocation_lock);
1865 free_pending_exception(pe);
1869 pe->exception_sequence = s->exception_start_sequence++;
1870 spin_unlock(&s->pe_allocation_lock);
1872 dm_insert_exception(&s->pending, &pe->e);
1878 * Looks to see if this snapshot already has a pending exception
1879 * for this chunk, otherwise it allocates a new one and inserts
1880 * it into the pending table.
1882 * NOTE: a write lock must be held on the chunk's pending exception table slot
1883 * before calling this.
1885 static struct dm_snap_pending_exception *
1886 __find_pending_exception(struct dm_snapshot *s,
1887 struct dm_snap_pending_exception *pe, chunk_t chunk)
1889 struct dm_snap_pending_exception *pe2;
1891 pe2 = __lookup_pending_exception(s, chunk);
1893 free_pending_exception(pe);
1897 return __insert_pending_exception(s, pe, chunk);
1900 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1901 struct bio *bio, chunk_t chunk)
1903 bio_set_dev(bio, s->cow->bdev);
1904 bio->bi_iter.bi_sector =
1905 chunk_to_sector(s->store, dm_chunk_number(e->new_chunk) +
1906 (chunk - e->old_chunk)) +
1907 (bio->bi_iter.bi_sector & s->store->chunk_mask);
1910 static void zero_callback(int read_err, unsigned long write_err, void *context)
1912 struct bio *bio = context;
1913 struct dm_snapshot *s = bio->bi_private;
1915 account_end_copy(s);
1916 bio->bi_status = write_err ? BLK_STS_IOERR : 0;
1920 static void zero_exception(struct dm_snapshot *s, struct dm_exception *e,
1921 struct bio *bio, chunk_t chunk)
1923 struct dm_io_region dest;
1925 dest.bdev = s->cow->bdev;
1926 dest.sector = bio->bi_iter.bi_sector;
1927 dest.count = s->store->chunk_size;
1929 account_start_copy(s);
1930 WARN_ON_ONCE(bio->bi_private);
1931 bio->bi_private = s;
1932 dm_kcopyd_zero(s->kcopyd_client, 1, &dest, 0, zero_callback, bio);
1935 static bool io_overlaps_chunk(struct dm_snapshot *s, struct bio *bio)
1937 return bio->bi_iter.bi_size ==
1938 (s->store->chunk_size << SECTOR_SHIFT);
1941 static int snapshot_map(struct dm_target *ti, struct bio *bio)
1943 struct dm_exception *e;
1944 struct dm_snapshot *s = ti->private;
1945 int r = DM_MAPIO_REMAPPED;
1947 struct dm_snap_pending_exception *pe = NULL;
1948 struct dm_exception_table_lock lock;
1950 init_tracked_chunk(bio);
1952 if (bio->bi_opf & REQ_PREFLUSH) {
1953 bio_set_dev(bio, s->cow->bdev);
1954 return DM_MAPIO_REMAPPED;
1957 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
1958 dm_exception_table_lock_init(s, chunk, &lock);
1960 /* Full snapshots are not usable */
1961 /* To get here the table must be live so s->active is always set. */
1963 return DM_MAPIO_KILL;
1965 if (bio_data_dir(bio) == WRITE) {
1966 while (unlikely(!wait_for_in_progress(s, false)))
1967 ; /* wait_for_in_progress() has slept */
1970 down_read(&s->lock);
1971 dm_exception_table_lock(&lock);
1973 if (!s->valid || (unlikely(s->snapshot_overflowed) &&
1974 bio_data_dir(bio) == WRITE)) {
1979 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
1980 if (s->discard_passdown_origin && dm_bio_get_target_bio_nr(bio)) {
1982 * passdown discard to origin (without triggering
1983 * snapshot exceptions via do_origin; doing so would
1984 * defeat the goal of freeing space in origin that is
1985 * implied by the "discard_passdown_origin" feature)
1987 bio_set_dev(bio, s->origin->bdev);
1988 track_chunk(s, bio, chunk);
1991 /* discard to snapshot (target_bio_nr == 0) zeroes exceptions */
1994 /* If the block is already remapped - use that, else remap it */
1995 e = dm_lookup_exception(&s->complete, chunk);
1997 remap_exception(s, e, bio, chunk);
1998 if (unlikely(bio_op(bio) == REQ_OP_DISCARD) &&
1999 io_overlaps_chunk(s, bio)) {
2000 dm_exception_table_unlock(&lock);
2002 zero_exception(s, e, bio, chunk);
2003 r = DM_MAPIO_SUBMITTED; /* discard is not issued */
2009 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
2011 * If no exception exists, complete discard immediately
2012 * otherwise it'll trigger copy-out.
2015 r = DM_MAPIO_SUBMITTED;
2020 * Write to snapshot - higher level takes care of RW/RO
2021 * flags so we should only get this if we are
2024 if (bio_data_dir(bio) == WRITE) {
2025 pe = __lookup_pending_exception(s, chunk);
2027 dm_exception_table_unlock(&lock);
2028 pe = alloc_pending_exception(s);
2029 dm_exception_table_lock(&lock);
2031 e = dm_lookup_exception(&s->complete, chunk);
2033 free_pending_exception(pe);
2034 remap_exception(s, e, bio, chunk);
2038 pe = __find_pending_exception(s, pe, chunk);
2040 dm_exception_table_unlock(&lock);
2043 down_write(&s->lock);
2045 if (s->store->userspace_supports_overflow) {
2046 if (s->valid && !s->snapshot_overflowed) {
2047 s->snapshot_overflowed = 1;
2048 DMERR("Snapshot overflowed: Unable to allocate exception.");
2051 __invalidate_snapshot(s, -ENOMEM);
2059 remap_exception(s, &pe->e, bio, chunk);
2061 r = DM_MAPIO_SUBMITTED;
2063 if (!pe->started && io_overlaps_chunk(s, bio)) {
2066 dm_exception_table_unlock(&lock);
2069 start_full_bio(pe, bio);
2073 bio_list_add(&pe->snapshot_bios, bio);
2076 /* this is protected by the exception table lock */
2079 dm_exception_table_unlock(&lock);
2086 bio_set_dev(bio, s->origin->bdev);
2087 track_chunk(s, bio, chunk);
2091 dm_exception_table_unlock(&lock);
2098 * A snapshot-merge target behaves like a combination of a snapshot
2099 * target and a snapshot-origin target. It only generates new
2100 * exceptions in other snapshots and not in the one that is being
2103 * For each chunk, if there is an existing exception, it is used to
2104 * redirect I/O to the cow device. Otherwise I/O is sent to the origin,
2105 * which in turn might generate exceptions in other snapshots.
2106 * If merging is currently taking place on the chunk in question, the
2107 * I/O is deferred by adding it to s->bios_queued_during_merge.
2109 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio)
2111 struct dm_exception *e;
2112 struct dm_snapshot *s = ti->private;
2113 int r = DM_MAPIO_REMAPPED;
2116 init_tracked_chunk(bio);
2118 if (bio->bi_opf & REQ_PREFLUSH) {
2119 if (!dm_bio_get_target_bio_nr(bio))
2120 bio_set_dev(bio, s->origin->bdev);
2122 bio_set_dev(bio, s->cow->bdev);
2123 return DM_MAPIO_REMAPPED;
2126 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) {
2127 /* Once merging, discards no longer effect change */
2129 return DM_MAPIO_SUBMITTED;
2132 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
2134 down_write(&s->lock);
2136 /* Full merging snapshots are redirected to the origin */
2138 goto redirect_to_origin;
2140 /* If the block is already remapped - use that */
2141 e = dm_lookup_exception(&s->complete, chunk);
2143 /* Queue writes overlapping with chunks being merged */
2144 if (bio_data_dir(bio) == WRITE &&
2145 chunk >= s->first_merging_chunk &&
2146 chunk < (s->first_merging_chunk +
2147 s->num_merging_chunks)) {
2148 bio_set_dev(bio, s->origin->bdev);
2149 bio_list_add(&s->bios_queued_during_merge, bio);
2150 r = DM_MAPIO_SUBMITTED;
2154 remap_exception(s, e, bio, chunk);
2156 if (bio_data_dir(bio) == WRITE)
2157 track_chunk(s, bio, chunk);
2162 bio_set_dev(bio, s->origin->bdev);
2164 if (bio_data_dir(bio) == WRITE) {
2166 return do_origin(s->origin, bio, false);
2175 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
2176 blk_status_t *error)
2178 struct dm_snapshot *s = ti->private;
2180 if (is_bio_tracked(bio))
2181 stop_tracking_chunk(s, bio);
2183 return DM_ENDIO_DONE;
2186 static void snapshot_merge_presuspend(struct dm_target *ti)
2188 struct dm_snapshot *s = ti->private;
2193 static int snapshot_preresume(struct dm_target *ti)
2196 struct dm_snapshot *s = ti->private;
2197 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
2199 down_read(&_origins_lock);
2200 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
2201 if (snap_src && snap_dest) {
2202 down_read(&snap_src->lock);
2203 if (s == snap_src) {
2204 DMERR("Unable to resume snapshot source until "
2205 "handover completes.");
2207 } else if (!dm_suspended(snap_src->ti)) {
2208 DMERR("Unable to perform snapshot handover until "
2209 "source is suspended.");
2212 up_read(&snap_src->lock);
2214 up_read(&_origins_lock);
2219 static void snapshot_resume(struct dm_target *ti)
2221 struct dm_snapshot *s = ti->private;
2222 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL, *snap_merging = NULL;
2223 struct dm_origin *o;
2224 struct mapped_device *origin_md = NULL;
2225 bool must_restart_merging = false;
2227 down_read(&_origins_lock);
2229 o = __lookup_dm_origin(s->origin->bdev);
2231 origin_md = dm_table_get_md(o->ti->table);
2233 (void) __find_snapshots_sharing_cow(s, NULL, NULL, &snap_merging);
2235 origin_md = dm_table_get_md(snap_merging->ti->table);
2237 if (origin_md == dm_table_get_md(ti->table))
2240 if (dm_hold(origin_md))
2244 up_read(&_origins_lock);
2247 dm_internal_suspend_fast(origin_md);
2248 if (snap_merging && test_bit(RUNNING_MERGE, &snap_merging->state_bits)) {
2249 must_restart_merging = true;
2250 stop_merge(snap_merging);
2254 down_read(&_origins_lock);
2256 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
2257 if (snap_src && snap_dest) {
2258 down_write(&snap_src->lock);
2259 down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
2260 __handover_exceptions(snap_src, snap_dest);
2261 up_write(&snap_dest->lock);
2262 up_write(&snap_src->lock);
2265 up_read(&_origins_lock);
2268 if (must_restart_merging)
2269 start_merge(snap_merging);
2270 dm_internal_resume_fast(origin_md);
2274 /* Now we have correct chunk size, reregister */
2275 reregister_snapshot(s);
2277 down_write(&s->lock);
2282 static uint32_t get_origin_minimum_chunksize(struct block_device *bdev)
2284 uint32_t min_chunksize;
2286 down_read(&_origins_lock);
2287 min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
2288 up_read(&_origins_lock);
2290 return min_chunksize;
2293 static void snapshot_merge_resume(struct dm_target *ti)
2295 struct dm_snapshot *s = ti->private;
2298 * Handover exceptions from existing snapshot.
2300 snapshot_resume(ti);
2303 * snapshot-merge acts as an origin, so set ti->max_io_len
2305 ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev);
2310 static void snapshot_status(struct dm_target *ti, status_type_t type,
2311 unsigned status_flags, char *result, unsigned maxlen)
2314 struct dm_snapshot *snap = ti->private;
2315 unsigned num_features;
2318 case STATUSTYPE_INFO:
2320 down_write(&snap->lock);
2324 else if (snap->merge_failed)
2325 DMEMIT("Merge failed");
2326 else if (snap->snapshot_overflowed)
2329 if (snap->store->type->usage) {
2330 sector_t total_sectors, sectors_allocated,
2332 snap->store->type->usage(snap->store,
2336 DMEMIT("%llu/%llu %llu",
2337 (unsigned long long)sectors_allocated,
2338 (unsigned long long)total_sectors,
2339 (unsigned long long)metadata_sectors);
2345 up_write(&snap->lock);
2349 case STATUSTYPE_TABLE:
2351 * kdevname returns a static pointer so we need
2352 * to make private copies if the output is to
2355 DMEMIT("%s %s", snap->origin->name, snap->cow->name);
2356 sz += snap->store->type->status(snap->store, type, result + sz,
2358 num_features = snap->discard_zeroes_cow + snap->discard_passdown_origin;
2360 DMEMIT(" %u", num_features);
2361 if (snap->discard_zeroes_cow)
2362 DMEMIT(" discard_zeroes_cow");
2363 if (snap->discard_passdown_origin)
2364 DMEMIT(" discard_passdown_origin");
2370 static int snapshot_iterate_devices(struct dm_target *ti,
2371 iterate_devices_callout_fn fn, void *data)
2373 struct dm_snapshot *snap = ti->private;
2376 r = fn(ti, snap->origin, 0, ti->len, data);
2379 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
2384 static void snapshot_io_hints(struct dm_target *ti, struct queue_limits *limits)
2386 struct dm_snapshot *snap = ti->private;
2388 if (snap->discard_zeroes_cow) {
2389 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
2391 down_read(&_origins_lock);
2393 (void) __find_snapshots_sharing_cow(snap, &snap_src, &snap_dest, NULL);
2394 if (snap_src && snap_dest)
2397 /* All discards are split on chunk_size boundary */
2398 limits->discard_granularity = snap->store->chunk_size;
2399 limits->max_discard_sectors = snap->store->chunk_size;
2401 up_read(&_origins_lock);
2405 /*-----------------------------------------------------------------
2407 *---------------------------------------------------------------*/
2410 * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
2411 * supplied bio was ignored. The caller may submit it immediately.
2412 * (No remapping actually occurs as the origin is always a direct linear
2415 * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
2416 * and any supplied bio is added to a list to be submitted once all
2417 * the necessary exceptions exist.
2419 static int __origin_write(struct list_head *snapshots, sector_t sector,
2422 int r = DM_MAPIO_REMAPPED;
2423 struct dm_snapshot *snap;
2424 struct dm_exception *e;
2425 struct dm_snap_pending_exception *pe, *pe2;
2426 struct dm_snap_pending_exception *pe_to_start_now = NULL;
2427 struct dm_snap_pending_exception *pe_to_start_last = NULL;
2428 struct dm_exception_table_lock lock;
2431 /* Do all the snapshots on this origin */
2432 list_for_each_entry (snap, snapshots, list) {
2434 * Don't make new exceptions in a merging snapshot
2435 * because it has effectively been deleted
2437 if (dm_target_is_snapshot_merge(snap->ti))
2440 /* Nothing to do if writing beyond end of snapshot */
2441 if (sector >= dm_table_get_size(snap->ti->table))
2445 * Remember, different snapshots can have
2446 * different chunk sizes.
2448 chunk = sector_to_chunk(snap->store, sector);
2449 dm_exception_table_lock_init(snap, chunk, &lock);
2451 down_read(&snap->lock);
2452 dm_exception_table_lock(&lock);
2454 /* Only deal with valid and active snapshots */
2455 if (!snap->valid || !snap->active)
2458 pe = __lookup_pending_exception(snap, chunk);
2461 * Check exception table to see if block is already
2462 * remapped in this snapshot and trigger an exception
2465 e = dm_lookup_exception(&snap->complete, chunk);
2469 dm_exception_table_unlock(&lock);
2470 pe = alloc_pending_exception(snap);
2471 dm_exception_table_lock(&lock);
2473 pe2 = __lookup_pending_exception(snap, chunk);
2476 e = dm_lookup_exception(&snap->complete, chunk);
2478 free_pending_exception(pe);
2482 pe = __insert_pending_exception(snap, pe, chunk);
2484 dm_exception_table_unlock(&lock);
2485 up_read(&snap->lock);
2487 invalidate_snapshot(snap, -ENOMEM);
2491 free_pending_exception(pe);
2496 r = DM_MAPIO_SUBMITTED;
2499 * If an origin bio was supplied, queue it to wait for the
2500 * completion of this exception, and start this one last,
2501 * at the end of the function.
2504 bio_list_add(&pe->origin_bios, bio);
2509 pe_to_start_last = pe;
2515 pe_to_start_now = pe;
2519 dm_exception_table_unlock(&lock);
2520 up_read(&snap->lock);
2522 if (pe_to_start_now) {
2523 start_copy(pe_to_start_now);
2524 pe_to_start_now = NULL;
2529 * Submit the exception against which the bio is queued last,
2530 * to give the other exceptions a head start.
2532 if (pe_to_start_last)
2533 start_copy(pe_to_start_last);
2539 * Called on a write from the origin driver.
2541 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit)
2544 int r = DM_MAPIO_REMAPPED;
2547 down_read(&_origins_lock);
2548 o = __lookup_origin(origin->bdev);
2551 struct dm_snapshot *s;
2552 list_for_each_entry(s, &o->snapshots, list)
2553 if (unlikely(!wait_for_in_progress(s, true)))
2557 r = __origin_write(&o->snapshots, bio->bi_iter.bi_sector, bio);
2559 up_read(&_origins_lock);
2565 * Trigger exceptions in all non-merging snapshots.
2567 * The chunk size of the merging snapshot may be larger than the chunk
2568 * size of some other snapshot so we may need to reallocate multiple
2569 * chunks in other snapshots.
2571 * We scan all the overlapping exceptions in the other snapshots.
2572 * Returns 1 if anything was reallocated and must be waited for,
2573 * otherwise returns 0.
2575 * size must be a multiple of merging_snap's chunk_size.
2577 static int origin_write_extent(struct dm_snapshot *merging_snap,
2578 sector_t sector, unsigned size)
2585 * The origin's __minimum_chunk_size() got stored in max_io_len
2586 * by snapshot_merge_resume().
2588 down_read(&_origins_lock);
2589 o = __lookup_origin(merging_snap->origin->bdev);
2590 for (n = 0; n < size; n += merging_snap->ti->max_io_len)
2591 if (__origin_write(&o->snapshots, sector + n, NULL) ==
2594 up_read(&_origins_lock);
2600 * Origin: maps a linear range of a device, with hooks for snapshotting.
2604 * Construct an origin mapping: <dev_path>
2605 * The context for an origin is merely a 'struct dm_dev *'
2606 * pointing to the real device.
2608 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2611 struct dm_origin *o;
2614 ti->error = "origin: incorrect number of arguments";
2618 o = kmalloc(sizeof(struct dm_origin), GFP_KERNEL);
2620 ti->error = "Cannot allocate private origin structure";
2625 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &o->dev);
2627 ti->error = "Cannot get target device";
2633 ti->num_flush_bios = 1;
2643 static void origin_dtr(struct dm_target *ti)
2645 struct dm_origin *o = ti->private;
2647 dm_put_device(ti, o->dev);
2651 static int origin_map(struct dm_target *ti, struct bio *bio)
2653 struct dm_origin *o = ti->private;
2654 unsigned available_sectors;
2656 bio_set_dev(bio, o->dev->bdev);
2658 if (unlikely(bio->bi_opf & REQ_PREFLUSH))
2659 return DM_MAPIO_REMAPPED;
2661 if (bio_data_dir(bio) != WRITE)
2662 return DM_MAPIO_REMAPPED;
2664 available_sectors = o->split_boundary -
2665 ((unsigned)bio->bi_iter.bi_sector & (o->split_boundary - 1));
2667 if (bio_sectors(bio) > available_sectors)
2668 dm_accept_partial_bio(bio, available_sectors);
2670 /* Only tell snapshots if this is a write */
2671 return do_origin(o->dev, bio, true);
2675 * Set the target "max_io_len" field to the minimum of all the snapshots'
2678 static void origin_resume(struct dm_target *ti)
2680 struct dm_origin *o = ti->private;
2682 o->split_boundary = get_origin_minimum_chunksize(o->dev->bdev);
2684 down_write(&_origins_lock);
2685 __insert_dm_origin(o);
2686 up_write(&_origins_lock);
2689 static void origin_postsuspend(struct dm_target *ti)
2691 struct dm_origin *o = ti->private;
2693 down_write(&_origins_lock);
2694 __remove_dm_origin(o);
2695 up_write(&_origins_lock);
2698 static void origin_status(struct dm_target *ti, status_type_t type,
2699 unsigned status_flags, char *result, unsigned maxlen)
2701 struct dm_origin *o = ti->private;
2704 case STATUSTYPE_INFO:
2708 case STATUSTYPE_TABLE:
2709 snprintf(result, maxlen, "%s", o->dev->name);
2714 static int origin_iterate_devices(struct dm_target *ti,
2715 iterate_devices_callout_fn fn, void *data)
2717 struct dm_origin *o = ti->private;
2719 return fn(ti, o->dev, 0, ti->len, data);
2722 static struct target_type origin_target = {
2723 .name = "snapshot-origin",
2724 .version = {1, 9, 0},
2725 .module = THIS_MODULE,
2729 .resume = origin_resume,
2730 .postsuspend = origin_postsuspend,
2731 .status = origin_status,
2732 .iterate_devices = origin_iterate_devices,
2735 static struct target_type snapshot_target = {
2737 .version = {1, 16, 0},
2738 .module = THIS_MODULE,
2739 .ctr = snapshot_ctr,
2740 .dtr = snapshot_dtr,
2741 .map = snapshot_map,
2742 .end_io = snapshot_end_io,
2743 .preresume = snapshot_preresume,
2744 .resume = snapshot_resume,
2745 .status = snapshot_status,
2746 .iterate_devices = snapshot_iterate_devices,
2747 .io_hints = snapshot_io_hints,
2750 static struct target_type merge_target = {
2751 .name = dm_snapshot_merge_target_name,
2752 .version = {1, 5, 0},
2753 .module = THIS_MODULE,
2754 .ctr = snapshot_ctr,
2755 .dtr = snapshot_dtr,
2756 .map = snapshot_merge_map,
2757 .end_io = snapshot_end_io,
2758 .presuspend = snapshot_merge_presuspend,
2759 .preresume = snapshot_preresume,
2760 .resume = snapshot_merge_resume,
2761 .status = snapshot_status,
2762 .iterate_devices = snapshot_iterate_devices,
2763 .io_hints = snapshot_io_hints,
2766 static int __init dm_snapshot_init(void)
2770 r = dm_exception_store_init();
2772 DMERR("Failed to initialize exception stores");
2776 r = init_origin_hash();
2778 DMERR("init_origin_hash failed.");
2779 goto bad_origin_hash;
2782 exception_cache = KMEM_CACHE(dm_exception, 0);
2783 if (!exception_cache) {
2784 DMERR("Couldn't create exception cache.");
2786 goto bad_exception_cache;
2789 pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2790 if (!pending_cache) {
2791 DMERR("Couldn't create pending cache.");
2793 goto bad_pending_cache;
2796 r = dm_register_target(&snapshot_target);
2798 DMERR("snapshot target register failed %d", r);
2799 goto bad_register_snapshot_target;
2802 r = dm_register_target(&origin_target);
2804 DMERR("Origin target register failed %d", r);
2805 goto bad_register_origin_target;
2808 r = dm_register_target(&merge_target);
2810 DMERR("Merge target register failed %d", r);
2811 goto bad_register_merge_target;
2816 bad_register_merge_target:
2817 dm_unregister_target(&origin_target);
2818 bad_register_origin_target:
2819 dm_unregister_target(&snapshot_target);
2820 bad_register_snapshot_target:
2821 kmem_cache_destroy(pending_cache);
2823 kmem_cache_destroy(exception_cache);
2824 bad_exception_cache:
2827 dm_exception_store_exit();
2832 static void __exit dm_snapshot_exit(void)
2834 dm_unregister_target(&snapshot_target);
2835 dm_unregister_target(&origin_target);
2836 dm_unregister_target(&merge_target);
2839 kmem_cache_destroy(pending_cache);
2840 kmem_cache_destroy(exception_cache);
2842 dm_exception_store_exit();
2846 module_init(dm_snapshot_init);
2847 module_exit(dm_snapshot_exit);
2849 MODULE_DESCRIPTION(DM_NAME " snapshot target");
2850 MODULE_AUTHOR("Joe Thornber");
2851 MODULE_LICENSE("GPL");
2852 MODULE_ALIAS("dm-snapshot-origin");
2853 MODULE_ALIAS("dm-snapshot-merge");