4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
6 * This file is released under the GPL.
9 #include <linux/blkdev.h>
10 #include <linux/device-mapper.h>
11 #include <linux/delay.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>
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.
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 {
46 struct list_head *table;
52 struct dm_dev *origin;
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.
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 */
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 rb_root out_of_order_tree;
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.
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 unsigned in_progress;
109 struct wait_queue_head in_progress_wait;
111 struct dm_kcopyd_client *kcopyd_client;
113 /* Wait for events based on state_bits */
114 unsigned long state_bits;
116 /* Range of chunks currently being merged. */
117 chunk_t first_merging_chunk;
118 int num_merging_chunks;
121 * The merge operation failed if this flag is set.
122 * Failure modes are handled as follows:
123 * - I/O error reading the header
124 * => don't load the target; abort.
125 * - Header does not have "valid" flag set
126 * => use the origin; forget about the snapshot.
127 * - I/O error when reading exceptions
128 * => don't load the target; abort.
129 * (We can't use the intermediate origin state.)
130 * - I/O error while merging
131 * => stop merging; set merge_failed; process I/O normally.
136 * Incoming bios that overlap with chunks being merged must wait
137 * for them to be committed.
139 struct bio_list bios_queued_during_merge;
142 * Flush data after merge.
144 struct bio flush_bio;
149 * RUNNING_MERGE - Merge operation is in progress.
150 * SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
151 * cleared afterwards.
153 #define RUNNING_MERGE 0
154 #define SHUTDOWN_MERGE 1
157 * Maximum number of chunks being copied on write.
159 * The value was decided experimentally as a trade-off between memory
160 * consumption, stalling the kernel's workqueues and maintaining a high enough
163 #define DEFAULT_COW_THRESHOLD 2048
165 static unsigned cow_threshold = DEFAULT_COW_THRESHOLD;
166 module_param_named(snapshot_cow_threshold, cow_threshold, uint, 0644);
167 MODULE_PARM_DESC(snapshot_cow_threshold, "Maximum number of chunks being copied on write");
169 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
170 "A percentage of time allocated for copy on write");
172 struct dm_dev *dm_snap_origin(struct dm_snapshot *s)
176 EXPORT_SYMBOL(dm_snap_origin);
178 struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
182 EXPORT_SYMBOL(dm_snap_cow);
184 static sector_t chunk_to_sector(struct dm_exception_store *store,
187 return chunk << store->chunk_shift;
190 static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
193 * There is only ever one instance of a particular block
194 * device so we can compare pointers safely.
199 struct dm_snap_pending_exception {
200 struct dm_exception e;
203 * Origin buffers waiting for this to complete are held
206 struct bio_list origin_bios;
207 struct bio_list snapshot_bios;
209 /* Pointer back to snapshot context */
210 struct dm_snapshot *snap;
213 * 1 indicates the exception has already been sent to
218 /* There was copying error. */
221 /* A sequence number, it is used for in-order completion. */
222 sector_t exception_sequence;
224 struct rb_node out_of_order_node;
227 * For writing a complete chunk, bypassing the copy.
229 struct bio *full_bio;
230 bio_end_io_t *full_bio_end_io;
234 * Hash table mapping origin volumes to lists of snapshots and
235 * a lock to protect it
237 static struct kmem_cache *exception_cache;
238 static struct kmem_cache *pending_cache;
240 struct dm_snap_tracked_chunk {
241 struct hlist_node node;
245 static void init_tracked_chunk(struct bio *bio)
247 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
248 INIT_HLIST_NODE(&c->node);
251 static bool is_bio_tracked(struct bio *bio)
253 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
254 return !hlist_unhashed(&c->node);
257 static void track_chunk(struct dm_snapshot *s, struct bio *bio, chunk_t chunk)
259 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
263 spin_lock_irq(&s->tracked_chunk_lock);
264 hlist_add_head(&c->node,
265 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)]);
266 spin_unlock_irq(&s->tracked_chunk_lock);
269 static void stop_tracking_chunk(struct dm_snapshot *s, struct bio *bio)
271 struct dm_snap_tracked_chunk *c = dm_per_bio_data(bio, sizeof(struct dm_snap_tracked_chunk));
274 spin_lock_irqsave(&s->tracked_chunk_lock, flags);
276 spin_unlock_irqrestore(&s->tracked_chunk_lock, flags);
279 static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
281 struct dm_snap_tracked_chunk *c;
284 spin_lock_irq(&s->tracked_chunk_lock);
286 hlist_for_each_entry(c,
287 &s->tracked_chunk_hash[DM_TRACKED_CHUNK_HASH(chunk)], node) {
288 if (c->chunk == chunk) {
294 spin_unlock_irq(&s->tracked_chunk_lock);
300 * This conflicting I/O is extremely improbable in the caller,
301 * so msleep(1) is sufficient and there is no need for a wait queue.
303 static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
305 while (__chunk_is_tracked(s, chunk))
310 * One of these per registered origin, held in the snapshot_origins hash
313 /* The origin device */
314 struct block_device *bdev;
316 struct list_head hash_list;
318 /* List of snapshots for this origin */
319 struct list_head snapshots;
323 * This structure is allocated for each origin target
327 struct dm_target *ti;
328 unsigned split_boundary;
329 struct list_head hash_list;
333 * Size of the hash table for origin volumes. If we make this
334 * the size of the minors list then it should be nearly perfect
336 #define ORIGIN_HASH_SIZE 256
337 #define ORIGIN_MASK 0xFF
338 static struct list_head *_origins;
339 static struct list_head *_dm_origins;
340 static struct rw_semaphore _origins_lock;
342 static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
343 static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
344 static uint64_t _pending_exceptions_done_count;
346 static int init_origin_hash(void)
350 _origins = kmalloc_array(ORIGIN_HASH_SIZE, sizeof(struct list_head),
353 DMERR("unable to allocate memory for _origins");
356 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
357 INIT_LIST_HEAD(_origins + i);
359 _dm_origins = kmalloc_array(ORIGIN_HASH_SIZE,
360 sizeof(struct list_head),
363 DMERR("unable to allocate memory for _dm_origins");
367 for (i = 0; i < ORIGIN_HASH_SIZE; i++)
368 INIT_LIST_HEAD(_dm_origins + i);
370 init_rwsem(&_origins_lock);
375 static void exit_origin_hash(void)
381 static unsigned origin_hash(struct block_device *bdev)
383 return bdev->bd_dev & ORIGIN_MASK;
386 static struct origin *__lookup_origin(struct block_device *origin)
388 struct list_head *ol;
391 ol = &_origins[origin_hash(origin)];
392 list_for_each_entry (o, ol, hash_list)
393 if (bdev_equal(o->bdev, origin))
399 static void __insert_origin(struct origin *o)
401 struct list_head *sl = &_origins[origin_hash(o->bdev)];
402 list_add_tail(&o->hash_list, sl);
405 static struct dm_origin *__lookup_dm_origin(struct block_device *origin)
407 struct list_head *ol;
410 ol = &_dm_origins[origin_hash(origin)];
411 list_for_each_entry (o, ol, hash_list)
412 if (bdev_equal(o->dev->bdev, origin))
418 static void __insert_dm_origin(struct dm_origin *o)
420 struct list_head *sl = &_dm_origins[origin_hash(o->dev->bdev)];
421 list_add_tail(&o->hash_list, sl);
424 static void __remove_dm_origin(struct dm_origin *o)
426 list_del(&o->hash_list);
430 * _origins_lock must be held when calling this function.
431 * Returns number of snapshots registered using the supplied cow device, plus:
432 * snap_src - a snapshot suitable for use as a source of exception handover
433 * snap_dest - a snapshot capable of receiving exception handover.
434 * snap_merge - an existing snapshot-merge target linked to the same origin.
435 * There can be at most one snapshot-merge target. The parameter is optional.
437 * Possible return values and states of snap_src and snap_dest.
438 * 0: NULL, NULL - first new snapshot
439 * 1: snap_src, NULL - normal snapshot
440 * 2: snap_src, snap_dest - waiting for handover
441 * 2: snap_src, NULL - handed over, waiting for old to be deleted
442 * 1: NULL, snap_dest - source got destroyed without handover
444 static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
445 struct dm_snapshot **snap_src,
446 struct dm_snapshot **snap_dest,
447 struct dm_snapshot **snap_merge)
449 struct dm_snapshot *s;
454 o = __lookup_origin(snap->origin->bdev);
458 list_for_each_entry(s, &o->snapshots, list) {
459 if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
461 if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
464 mutex_lock(&s->lock);
466 mutex_unlock(&s->lock);
471 } else if (snap_dest)
482 * On success, returns 1 if this snapshot is a handover destination,
483 * otherwise returns 0.
485 static int __validate_exception_handover(struct dm_snapshot *snap)
487 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
488 struct dm_snapshot *snap_merge = NULL;
490 /* Does snapshot need exceptions handed over to it? */
491 if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
492 &snap_merge) == 2) ||
494 snap->ti->error = "Snapshot cow pairing for exception "
495 "table handover failed";
500 * If no snap_src was found, snap cannot become a handover
507 * Non-snapshot-merge handover?
509 if (!dm_target_is_snapshot_merge(snap->ti))
513 * Do not allow more than one merging snapshot.
516 snap->ti->error = "A snapshot is already merging.";
520 if (!snap_src->store->type->prepare_merge ||
521 !snap_src->store->type->commit_merge) {
522 snap->ti->error = "Snapshot exception store does not "
523 "support snapshot-merge.";
530 static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
532 struct dm_snapshot *l;
534 /* Sort the list according to chunk size, largest-first smallest-last */
535 list_for_each_entry(l, &o->snapshots, list)
536 if (l->store->chunk_size < s->store->chunk_size)
538 list_add_tail(&s->list, &l->list);
542 * Make a note of the snapshot and its origin so we can look it
543 * up when the origin has a write on it.
545 * Also validate snapshot exception store handovers.
546 * On success, returns 1 if this registration is a handover destination,
547 * otherwise returns 0.
549 static int register_snapshot(struct dm_snapshot *snap)
551 struct origin *o, *new_o = NULL;
552 struct block_device *bdev = snap->origin->bdev;
555 new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
559 down_write(&_origins_lock);
561 r = __validate_exception_handover(snap);
567 o = __lookup_origin(bdev);
574 /* Initialise the struct */
575 INIT_LIST_HEAD(&o->snapshots);
581 __insert_snapshot(o, snap);
584 up_write(&_origins_lock);
590 * Move snapshot to correct place in list according to chunk size.
592 static void reregister_snapshot(struct dm_snapshot *s)
594 struct block_device *bdev = s->origin->bdev;
596 down_write(&_origins_lock);
599 __insert_snapshot(__lookup_origin(bdev), s);
601 up_write(&_origins_lock);
604 static void unregister_snapshot(struct dm_snapshot *s)
608 down_write(&_origins_lock);
609 o = __lookup_origin(s->origin->bdev);
612 if (o && list_empty(&o->snapshots)) {
613 list_del(&o->hash_list);
617 up_write(&_origins_lock);
621 * Implementation of the exception hash tables.
622 * The lowest hash_shift bits of the chunk number are ignored, allowing
623 * some consecutive chunks to be grouped together.
625 static int dm_exception_table_init(struct dm_exception_table *et,
626 uint32_t size, unsigned hash_shift)
630 et->hash_shift = hash_shift;
631 et->hash_mask = size - 1;
632 et->table = dm_vcalloc(size, sizeof(struct list_head));
636 for (i = 0; i < size; i++)
637 INIT_LIST_HEAD(et->table + i);
642 static void dm_exception_table_exit(struct dm_exception_table *et,
643 struct kmem_cache *mem)
645 struct list_head *slot;
646 struct dm_exception *ex, *next;
649 size = et->hash_mask + 1;
650 for (i = 0; i < size; i++) {
651 slot = et->table + i;
653 list_for_each_entry_safe (ex, next, slot, hash_list)
654 kmem_cache_free(mem, ex);
660 static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
662 return (chunk >> et->hash_shift) & et->hash_mask;
665 static void dm_remove_exception(struct dm_exception *e)
667 list_del(&e->hash_list);
671 * Return the exception data for a sector, or NULL if not
674 static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
677 struct list_head *slot;
678 struct dm_exception *e;
680 slot = &et->table[exception_hash(et, chunk)];
681 list_for_each_entry (e, slot, hash_list)
682 if (chunk >= e->old_chunk &&
683 chunk <= e->old_chunk + dm_consecutive_chunk_count(e))
689 static struct dm_exception *alloc_completed_exception(gfp_t gfp)
691 struct dm_exception *e;
693 e = kmem_cache_alloc(exception_cache, gfp);
694 if (!e && gfp == GFP_NOIO)
695 e = kmem_cache_alloc(exception_cache, GFP_ATOMIC);
700 static void free_completed_exception(struct dm_exception *e)
702 kmem_cache_free(exception_cache, e);
705 static struct dm_snap_pending_exception *alloc_pending_exception(struct dm_snapshot *s)
707 struct dm_snap_pending_exception *pe = mempool_alloc(&s->pending_pool,
710 atomic_inc(&s->pending_exceptions_count);
716 static void free_pending_exception(struct dm_snap_pending_exception *pe)
718 struct dm_snapshot *s = pe->snap;
720 mempool_free(pe, &s->pending_pool);
721 smp_mb__before_atomic();
722 atomic_dec(&s->pending_exceptions_count);
725 static void dm_insert_exception(struct dm_exception_table *eh,
726 struct dm_exception *new_e)
729 struct dm_exception *e = NULL;
731 l = &eh->table[exception_hash(eh, new_e->old_chunk)];
733 /* Add immediately if this table doesn't support consecutive chunks */
737 /* List is ordered by old_chunk */
738 list_for_each_entry_reverse(e, l, hash_list) {
739 /* Insert after an existing chunk? */
740 if (new_e->old_chunk == (e->old_chunk +
741 dm_consecutive_chunk_count(e) + 1) &&
742 new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
743 dm_consecutive_chunk_count(e) + 1)) {
744 dm_consecutive_chunk_count_inc(e);
745 free_completed_exception(new_e);
749 /* Insert before an existing chunk? */
750 if (new_e->old_chunk == (e->old_chunk - 1) &&
751 new_e->new_chunk == (dm_chunk_number(e->new_chunk) - 1)) {
752 dm_consecutive_chunk_count_inc(e);
755 free_completed_exception(new_e);
759 if (new_e->old_chunk > e->old_chunk)
764 list_add(&new_e->hash_list, e ? &e->hash_list : l);
768 * Callback used by the exception stores to load exceptions when
771 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
773 struct dm_snapshot *s = context;
774 struct dm_exception *e;
776 e = alloc_completed_exception(GFP_KERNEL);
782 /* Consecutive_count is implicitly initialised to zero */
785 dm_insert_exception(&s->complete, e);
791 * Return a minimum chunk size of all snapshots that have the specified origin.
792 * Return zero if the origin has no snapshots.
794 static uint32_t __minimum_chunk_size(struct origin *o)
796 struct dm_snapshot *snap;
797 unsigned chunk_size = rounddown_pow_of_two(UINT_MAX);
800 list_for_each_entry(snap, &o->snapshots, list)
801 chunk_size = min_not_zero(chunk_size,
802 snap->store->chunk_size);
804 return (uint32_t) chunk_size;
810 static int calc_max_buckets(void)
812 /* use a fixed size of 2MB */
813 unsigned long mem = 2 * 1024 * 1024;
814 mem /= sizeof(struct list_head);
820 * Allocate room for a suitable hash table.
822 static int init_hash_tables(struct dm_snapshot *s)
824 sector_t hash_size, cow_dev_size, max_buckets;
827 * Calculate based on the size of the original volume or
830 cow_dev_size = get_dev_size(s->cow->bdev);
831 max_buckets = calc_max_buckets();
833 hash_size = cow_dev_size >> s->store->chunk_shift;
834 hash_size = min(hash_size, max_buckets);
838 hash_size = rounddown_pow_of_two(hash_size);
839 if (dm_exception_table_init(&s->complete, hash_size,
840 DM_CHUNK_CONSECUTIVE_BITS))
844 * Allocate hash table for in-flight exceptions
845 * Make this smaller than the real hash table
851 if (dm_exception_table_init(&s->pending, hash_size, 0)) {
852 dm_exception_table_exit(&s->complete, exception_cache);
859 static void merge_shutdown(struct dm_snapshot *s)
861 clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
862 smp_mb__after_atomic();
863 wake_up_bit(&s->state_bits, RUNNING_MERGE);
866 static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
868 s->first_merging_chunk = 0;
869 s->num_merging_chunks = 0;
871 return bio_list_get(&s->bios_queued_during_merge);
875 * Remove one chunk from the index of completed exceptions.
877 static int __remove_single_exception_chunk(struct dm_snapshot *s,
880 struct dm_exception *e;
882 e = dm_lookup_exception(&s->complete, old_chunk);
884 DMERR("Corruption detected: exception for block %llu is "
885 "on disk but not in memory",
886 (unsigned long long)old_chunk);
891 * If this is the only chunk using this exception, remove exception.
893 if (!dm_consecutive_chunk_count(e)) {
894 dm_remove_exception(e);
895 free_completed_exception(e);
900 * The chunk may be either at the beginning or the end of a
901 * group of consecutive chunks - never in the middle. We are
902 * removing chunks in the opposite order to that in which they
903 * were added, so this should always be true.
904 * Decrement the consecutive chunk counter and adjust the
905 * starting point if necessary.
907 if (old_chunk == e->old_chunk) {
910 } else if (old_chunk != e->old_chunk +
911 dm_consecutive_chunk_count(e)) {
912 DMERR("Attempt to merge block %llu from the "
913 "middle of a chunk range [%llu - %llu]",
914 (unsigned long long)old_chunk,
915 (unsigned long long)e->old_chunk,
917 e->old_chunk + dm_consecutive_chunk_count(e));
921 dm_consecutive_chunk_count_dec(e);
926 static void flush_bios(struct bio *bio);
928 static int remove_single_exception_chunk(struct dm_snapshot *s)
930 struct bio *b = NULL;
932 chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
934 mutex_lock(&s->lock);
937 * Process chunks (and associated exceptions) in reverse order
938 * so that dm_consecutive_chunk_count_dec() accounting works.
941 r = __remove_single_exception_chunk(s, old_chunk);
944 } while (old_chunk-- > s->first_merging_chunk);
946 b = __release_queued_bios_after_merge(s);
949 mutex_unlock(&s->lock);
956 static int origin_write_extent(struct dm_snapshot *merging_snap,
957 sector_t sector, unsigned chunk_size);
959 static void merge_callback(int read_err, unsigned long write_err,
962 static uint64_t read_pending_exceptions_done_count(void)
964 uint64_t pending_exceptions_done;
966 spin_lock(&_pending_exceptions_done_spinlock);
967 pending_exceptions_done = _pending_exceptions_done_count;
968 spin_unlock(&_pending_exceptions_done_spinlock);
970 return pending_exceptions_done;
973 static void increment_pending_exceptions_done_count(void)
975 spin_lock(&_pending_exceptions_done_spinlock);
976 _pending_exceptions_done_count++;
977 spin_unlock(&_pending_exceptions_done_spinlock);
979 wake_up_all(&_pending_exceptions_done);
982 static void snapshot_merge_next_chunks(struct dm_snapshot *s)
984 int i, linear_chunks;
985 chunk_t old_chunk, new_chunk;
986 struct dm_io_region src, dest;
988 uint64_t previous_count;
990 BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
991 if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
995 * valid flag never changes during merge, so no lock required.
998 DMERR("Snapshot is invalid: can't merge");
1002 linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
1004 if (linear_chunks <= 0) {
1005 if (linear_chunks < 0) {
1006 DMERR("Read error in exception store: "
1007 "shutting down merge");
1008 mutex_lock(&s->lock);
1009 s->merge_failed = 1;
1010 mutex_unlock(&s->lock);
1015 /* Adjust old_chunk and new_chunk to reflect start of linear region */
1016 old_chunk = old_chunk + 1 - linear_chunks;
1017 new_chunk = new_chunk + 1 - linear_chunks;
1020 * Use one (potentially large) I/O to copy all 'linear_chunks'
1021 * from the exception store to the origin
1023 io_size = linear_chunks * s->store->chunk_size;
1025 dest.bdev = s->origin->bdev;
1026 dest.sector = chunk_to_sector(s->store, old_chunk);
1027 dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
1029 src.bdev = s->cow->bdev;
1030 src.sector = chunk_to_sector(s->store, new_chunk);
1031 src.count = dest.count;
1034 * Reallocate any exceptions needed in other snapshots then
1035 * wait for the pending exceptions to complete.
1036 * Each time any pending exception (globally on the system)
1037 * completes we are woken and repeat the process to find out
1038 * if we can proceed. While this may not seem a particularly
1039 * efficient algorithm, it is not expected to have any
1040 * significant impact on performance.
1042 previous_count = read_pending_exceptions_done_count();
1043 while (origin_write_extent(s, dest.sector, io_size)) {
1044 wait_event(_pending_exceptions_done,
1045 (read_pending_exceptions_done_count() !=
1047 /* Retry after the wait, until all exceptions are done. */
1048 previous_count = read_pending_exceptions_done_count();
1051 mutex_lock(&s->lock);
1052 s->first_merging_chunk = old_chunk;
1053 s->num_merging_chunks = linear_chunks;
1054 mutex_unlock(&s->lock);
1056 /* Wait until writes to all 'linear_chunks' drain */
1057 for (i = 0; i < linear_chunks; i++)
1058 __check_for_conflicting_io(s, old_chunk + i);
1060 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
1067 static void error_bios(struct bio *bio);
1069 static int flush_data(struct dm_snapshot *s)
1071 struct bio *flush_bio = &s->flush_bio;
1073 bio_reset(flush_bio);
1074 bio_set_dev(flush_bio, s->origin->bdev);
1075 flush_bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1077 return submit_bio_wait(flush_bio);
1080 static void merge_callback(int read_err, unsigned long write_err, void *context)
1082 struct dm_snapshot *s = context;
1083 struct bio *b = NULL;
1085 if (read_err || write_err) {
1087 DMERR("Read error: shutting down merge.");
1089 DMERR("Write error: shutting down merge.");
1093 if (flush_data(s) < 0) {
1094 DMERR("Flush after merge failed: shutting down merge");
1098 if (s->store->type->commit_merge(s->store,
1099 s->num_merging_chunks) < 0) {
1100 DMERR("Write error in exception store: shutting down merge");
1104 if (remove_single_exception_chunk(s) < 0)
1107 snapshot_merge_next_chunks(s);
1112 mutex_lock(&s->lock);
1113 s->merge_failed = 1;
1114 b = __release_queued_bios_after_merge(s);
1115 mutex_unlock(&s->lock);
1121 static void start_merge(struct dm_snapshot *s)
1123 if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
1124 snapshot_merge_next_chunks(s);
1128 * Stop the merging process and wait until it finishes.
1130 static void stop_merge(struct dm_snapshot *s)
1132 set_bit(SHUTDOWN_MERGE, &s->state_bits);
1133 wait_on_bit(&s->state_bits, RUNNING_MERGE, TASK_UNINTERRUPTIBLE);
1134 clear_bit(SHUTDOWN_MERGE, &s->state_bits);
1138 * Construct a snapshot mapping: <origin_dev> <COW-dev> <p|po|n> <chunk-size>
1140 static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1142 struct dm_snapshot *s;
1145 char *origin_path, *cow_path;
1146 dev_t origin_dev, cow_dev;
1147 unsigned args_used, num_flush_bios = 1;
1148 fmode_t origin_mode = FMODE_READ;
1151 ti->error = "requires exactly 4 arguments";
1156 if (dm_target_is_snapshot_merge(ti)) {
1158 origin_mode = FMODE_WRITE;
1161 s = kzalloc(sizeof(*s), GFP_KERNEL);
1163 ti->error = "Cannot allocate private snapshot structure";
1168 origin_path = argv[0];
1172 r = dm_get_device(ti, origin_path, origin_mode, &s->origin);
1174 ti->error = "Cannot get origin device";
1177 origin_dev = s->origin->bdev->bd_dev;
1183 cow_dev = dm_get_dev_t(cow_path);
1184 if (cow_dev && cow_dev == origin_dev) {
1185 ti->error = "COW device cannot be the same as origin device";
1190 r = dm_get_device(ti, cow_path, dm_table_get_mode(ti->table), &s->cow);
1192 ti->error = "Cannot get COW device";
1196 r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
1198 ti->error = "Couldn't create exception store";
1208 s->snapshot_overflowed = 0;
1210 atomic_set(&s->pending_exceptions_count, 0);
1211 s->exception_start_sequence = 0;
1212 s->exception_complete_sequence = 0;
1213 s->out_of_order_tree = RB_ROOT;
1214 mutex_init(&s->lock);
1215 INIT_LIST_HEAD(&s->list);
1216 spin_lock_init(&s->pe_lock);
1218 s->merge_failed = 0;
1219 s->first_merging_chunk = 0;
1220 s->num_merging_chunks = 0;
1221 bio_list_init(&s->bios_queued_during_merge);
1222 bio_init(&s->flush_bio, NULL, 0);
1224 /* Allocate hash table for COW data */
1225 if (init_hash_tables(s)) {
1226 ti->error = "Unable to allocate hash table space";
1228 goto bad_hash_tables;
1231 init_waitqueue_head(&s->in_progress_wait);
1233 s->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1234 if (IS_ERR(s->kcopyd_client)) {
1235 r = PTR_ERR(s->kcopyd_client);
1236 ti->error = "Could not create kcopyd client";
1240 r = mempool_init_slab_pool(&s->pending_pool, MIN_IOS, pending_cache);
1242 ti->error = "Could not allocate mempool for pending exceptions";
1243 goto bad_pending_pool;
1246 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1247 INIT_HLIST_HEAD(&s->tracked_chunk_hash[i]);
1249 spin_lock_init(&s->tracked_chunk_lock);
1252 ti->num_flush_bios = num_flush_bios;
1253 ti->per_io_data_size = sizeof(struct dm_snap_tracked_chunk);
1255 /* Add snapshot to the list of snapshots for this origin */
1256 /* Exceptions aren't triggered till snapshot_resume() is called */
1257 r = register_snapshot(s);
1259 ti->error = "Snapshot origin struct allocation failed";
1260 goto bad_load_and_register;
1262 /* invalid handover, register_snapshot has set ti->error */
1263 goto bad_load_and_register;
1267 * Metadata must only be loaded into one table at once, so skip this
1268 * if metadata will be handed over during resume.
1269 * Chunk size will be set during the handover - set it to zero to
1270 * ensure it's ignored.
1273 s->store->chunk_size = 0;
1277 r = s->store->type->read_metadata(s->store, dm_add_exception,
1280 ti->error = "Failed to read snapshot metadata";
1281 goto bad_read_metadata;
1284 DMWARN("Snapshot is marked invalid.");
1287 if (!s->store->chunk_size) {
1288 ti->error = "Chunk size not set";
1290 goto bad_read_metadata;
1293 r = dm_set_target_max_io_len(ti, s->store->chunk_size);
1295 goto bad_read_metadata;
1300 unregister_snapshot(s);
1302 bad_load_and_register:
1303 mempool_exit(&s->pending_pool);
1306 dm_kcopyd_client_destroy(s->kcopyd_client);
1309 dm_exception_table_exit(&s->pending, pending_cache);
1310 dm_exception_table_exit(&s->complete, exception_cache);
1313 dm_exception_store_destroy(s->store);
1316 dm_put_device(ti, s->cow);
1319 dm_put_device(ti, s->origin);
1328 static void __free_exceptions(struct dm_snapshot *s)
1330 dm_kcopyd_client_destroy(s->kcopyd_client);
1331 s->kcopyd_client = NULL;
1333 dm_exception_table_exit(&s->pending, pending_cache);
1334 dm_exception_table_exit(&s->complete, exception_cache);
1337 static void __handover_exceptions(struct dm_snapshot *snap_src,
1338 struct dm_snapshot *snap_dest)
1341 struct dm_exception_table table_swap;
1342 struct dm_exception_store *store_swap;
1346 * Swap all snapshot context information between the two instances.
1348 u.table_swap = snap_dest->complete;
1349 snap_dest->complete = snap_src->complete;
1350 snap_src->complete = u.table_swap;
1352 u.store_swap = snap_dest->store;
1353 snap_dest->store = snap_src->store;
1354 snap_dest->store->userspace_supports_overflow = u.store_swap->userspace_supports_overflow;
1355 snap_src->store = u.store_swap;
1357 snap_dest->store->snap = snap_dest;
1358 snap_src->store->snap = snap_src;
1360 snap_dest->ti->max_io_len = snap_dest->store->chunk_size;
1361 snap_dest->valid = snap_src->valid;
1362 snap_dest->snapshot_overflowed = snap_src->snapshot_overflowed;
1365 * Set source invalid to ensure it receives no further I/O.
1367 snap_src->valid = 0;
1370 static void snapshot_dtr(struct dm_target *ti)
1372 #ifdef CONFIG_DM_DEBUG
1375 struct dm_snapshot *s = ti->private;
1376 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1378 down_read(&_origins_lock);
1379 /* Check whether exception handover must be cancelled */
1380 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1381 if (snap_src && snap_dest && (s == snap_src)) {
1382 mutex_lock(&snap_dest->lock);
1383 snap_dest->valid = 0;
1384 mutex_unlock(&snap_dest->lock);
1385 DMERR("Cancelling snapshot handover.");
1387 up_read(&_origins_lock);
1389 if (dm_target_is_snapshot_merge(ti))
1392 /* Prevent further origin writes from using this snapshot. */
1393 /* After this returns there can be no new kcopyd jobs. */
1394 unregister_snapshot(s);
1396 while (atomic_read(&s->pending_exceptions_count))
1399 * Ensure instructions in mempool_exit aren't reordered
1400 * before atomic_read.
1404 #ifdef CONFIG_DM_DEBUG
1405 for (i = 0; i < DM_TRACKED_CHUNK_HASH_SIZE; i++)
1406 BUG_ON(!hlist_empty(&s->tracked_chunk_hash[i]));
1409 __free_exceptions(s);
1411 mempool_exit(&s->pending_pool);
1413 dm_exception_store_destroy(s->store);
1415 mutex_destroy(&s->lock);
1417 bio_uninit(&s->flush_bio);
1419 dm_put_device(ti, s->cow);
1421 dm_put_device(ti, s->origin);
1423 WARN_ON(s->in_progress);
1428 static void account_start_copy(struct dm_snapshot *s)
1430 spin_lock(&s->in_progress_wait.lock);
1432 spin_unlock(&s->in_progress_wait.lock);
1435 static void account_end_copy(struct dm_snapshot *s)
1437 spin_lock(&s->in_progress_wait.lock);
1438 BUG_ON(!s->in_progress);
1440 if (likely(s->in_progress <= cow_threshold) &&
1441 unlikely(waitqueue_active(&s->in_progress_wait)))
1442 wake_up_locked(&s->in_progress_wait);
1443 spin_unlock(&s->in_progress_wait.lock);
1446 static bool wait_for_in_progress(struct dm_snapshot *s, bool unlock_origins)
1448 if (unlikely(s->in_progress > cow_threshold)) {
1449 spin_lock(&s->in_progress_wait.lock);
1450 if (likely(s->in_progress > cow_threshold)) {
1452 * NOTE: this throttle doesn't account for whether
1453 * the caller is servicing an IO that will trigger a COW
1454 * so excess throttling may result for chunks not required
1455 * to be COW'd. But if cow_threshold was reached, extra
1456 * throttling is unlikely to negatively impact performance.
1458 DECLARE_WAITQUEUE(wait, current);
1459 __add_wait_queue(&s->in_progress_wait, &wait);
1460 __set_current_state(TASK_UNINTERRUPTIBLE);
1461 spin_unlock(&s->in_progress_wait.lock);
1463 up_read(&_origins_lock);
1465 remove_wait_queue(&s->in_progress_wait, &wait);
1468 spin_unlock(&s->in_progress_wait.lock);
1474 * Flush a list of buffers.
1476 static void flush_bios(struct bio *bio)
1482 bio->bi_next = NULL;
1483 generic_make_request(bio);
1488 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit);
1491 * Flush a list of buffers.
1493 static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
1500 bio->bi_next = NULL;
1501 r = do_origin(s->origin, bio, false);
1502 if (r == DM_MAPIO_REMAPPED)
1503 generic_make_request(bio);
1509 * Error a list of buffers.
1511 static void error_bios(struct bio *bio)
1517 bio->bi_next = NULL;
1523 static void __invalidate_snapshot(struct dm_snapshot *s, int err)
1529 DMERR("Invalidating snapshot: Error reading/writing.");
1530 else if (err == -ENOMEM)
1531 DMERR("Invalidating snapshot: Unable to allocate exception.");
1533 if (s->store->type->drop_snapshot)
1534 s->store->type->drop_snapshot(s->store);
1538 dm_table_event(s->ti->table);
1541 static void pending_complete(void *context, int success)
1543 struct dm_snap_pending_exception *pe = context;
1544 struct dm_exception *e;
1545 struct dm_snapshot *s = pe->snap;
1546 struct bio *origin_bios = NULL;
1547 struct bio *snapshot_bios = NULL;
1548 struct bio *full_bio = NULL;
1552 /* Read/write error - snapshot is unusable */
1553 mutex_lock(&s->lock);
1554 __invalidate_snapshot(s, -EIO);
1559 e = alloc_completed_exception(GFP_NOIO);
1561 mutex_lock(&s->lock);
1562 __invalidate_snapshot(s, -ENOMEM);
1568 mutex_lock(&s->lock);
1570 free_completed_exception(e);
1575 /* Check for conflicting reads */
1576 __check_for_conflicting_io(s, pe->e.old_chunk);
1579 * Add a proper exception, and remove the
1580 * in-flight exception from the list.
1582 dm_insert_exception(&s->complete, e);
1585 dm_remove_exception(&pe->e);
1586 snapshot_bios = bio_list_get(&pe->snapshot_bios);
1587 origin_bios = bio_list_get(&pe->origin_bios);
1588 full_bio = pe->full_bio;
1590 full_bio->bi_end_io = pe->full_bio_end_io;
1591 increment_pending_exceptions_done_count();
1593 mutex_unlock(&s->lock);
1595 /* Submit any pending write bios */
1598 bio_io_error(full_bio);
1599 error_bios(snapshot_bios);
1602 bio_endio(full_bio);
1603 flush_bios(snapshot_bios);
1606 retry_origin_bios(s, origin_bios);
1608 free_pending_exception(pe);
1611 static void complete_exception(struct dm_snap_pending_exception *pe)
1613 struct dm_snapshot *s = pe->snap;
1615 /* Update the metadata if we are persistent */
1616 s->store->type->commit_exception(s->store, &pe->e, !pe->copy_error,
1617 pending_complete, pe);
1621 * Called when the copy I/O has finished. kcopyd actually runs
1622 * this code so don't block.
1624 static void copy_callback(int read_err, unsigned long write_err, void *context)
1626 struct dm_snap_pending_exception *pe = context;
1627 struct dm_snapshot *s = pe->snap;
1629 pe->copy_error = read_err || write_err;
1631 if (pe->exception_sequence == s->exception_complete_sequence) {
1632 struct rb_node *next;
1634 s->exception_complete_sequence++;
1635 complete_exception(pe);
1637 next = rb_first(&s->out_of_order_tree);
1639 pe = rb_entry(next, struct dm_snap_pending_exception,
1641 if (pe->exception_sequence != s->exception_complete_sequence)
1643 next = rb_next(next);
1644 s->exception_complete_sequence++;
1645 rb_erase(&pe->out_of_order_node, &s->out_of_order_tree);
1646 complete_exception(pe);
1650 struct rb_node *parent = NULL;
1651 struct rb_node **p = &s->out_of_order_tree.rb_node;
1652 struct dm_snap_pending_exception *pe2;
1655 pe2 = rb_entry(*p, struct dm_snap_pending_exception, out_of_order_node);
1658 BUG_ON(pe->exception_sequence == pe2->exception_sequence);
1659 if (pe->exception_sequence < pe2->exception_sequence)
1660 p = &((*p)->rb_left);
1662 p = &((*p)->rb_right);
1665 rb_link_node(&pe->out_of_order_node, parent, p);
1666 rb_insert_color(&pe->out_of_order_node, &s->out_of_order_tree);
1668 account_end_copy(s);
1672 * Dispatches the copy operation to kcopyd.
1674 static void start_copy(struct dm_snap_pending_exception *pe)
1676 struct dm_snapshot *s = pe->snap;
1677 struct dm_io_region src, dest;
1678 struct block_device *bdev = s->origin->bdev;
1681 dev_size = get_dev_size(bdev);
1684 src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
1685 src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
1687 dest.bdev = s->cow->bdev;
1688 dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
1689 dest.count = src.count;
1691 /* Hand over to kcopyd */
1692 account_start_copy(s);
1693 dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, copy_callback, pe);
1696 static void full_bio_end_io(struct bio *bio)
1698 void *callback_data = bio->bi_private;
1700 dm_kcopyd_do_callback(callback_data, 0, bio->bi_status ? 1 : 0);
1703 static void start_full_bio(struct dm_snap_pending_exception *pe,
1706 struct dm_snapshot *s = pe->snap;
1707 void *callback_data;
1710 pe->full_bio_end_io = bio->bi_end_io;
1712 account_start_copy(s);
1713 callback_data = dm_kcopyd_prepare_callback(s->kcopyd_client,
1716 bio->bi_end_io = full_bio_end_io;
1717 bio->bi_private = callback_data;
1719 generic_make_request(bio);
1722 static struct dm_snap_pending_exception *
1723 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
1725 struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
1730 return container_of(e, struct dm_snap_pending_exception, e);
1734 * Looks to see if this snapshot already has a pending exception
1735 * for this chunk, otherwise it allocates a new one and inserts
1736 * it into the pending table.
1738 * NOTE: a write lock must be held on snap->lock before calling
1741 static struct dm_snap_pending_exception *
1742 __find_pending_exception(struct dm_snapshot *s,
1743 struct dm_snap_pending_exception *pe, chunk_t chunk)
1745 struct dm_snap_pending_exception *pe2;
1747 pe2 = __lookup_pending_exception(s, chunk);
1749 free_pending_exception(pe);
1753 pe->e.old_chunk = chunk;
1754 bio_list_init(&pe->origin_bios);
1755 bio_list_init(&pe->snapshot_bios);
1757 pe->full_bio = NULL;
1759 if (s->store->type->prepare_exception(s->store, &pe->e)) {
1760 free_pending_exception(pe);
1764 pe->exception_sequence = s->exception_start_sequence++;
1766 dm_insert_exception(&s->pending, &pe->e);
1771 static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
1772 struct bio *bio, chunk_t chunk)
1774 bio_set_dev(bio, s->cow->bdev);
1775 bio->bi_iter.bi_sector =
1776 chunk_to_sector(s->store, dm_chunk_number(e->new_chunk) +
1777 (chunk - e->old_chunk)) +
1778 (bio->bi_iter.bi_sector & s->store->chunk_mask);
1781 static int snapshot_map(struct dm_target *ti, struct bio *bio)
1783 struct dm_exception *e;
1784 struct dm_snapshot *s = ti->private;
1785 int r = DM_MAPIO_REMAPPED;
1787 struct dm_snap_pending_exception *pe = NULL;
1789 init_tracked_chunk(bio);
1791 if (bio->bi_opf & REQ_PREFLUSH) {
1792 bio_set_dev(bio, s->cow->bdev);
1793 return DM_MAPIO_REMAPPED;
1796 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
1798 /* Full snapshots are not usable */
1799 /* To get here the table must be live so s->active is always set. */
1801 return DM_MAPIO_KILL;
1803 if (bio_data_dir(bio) == WRITE) {
1804 while (unlikely(!wait_for_in_progress(s, false)))
1805 ; /* wait_for_in_progress() has slept */
1808 mutex_lock(&s->lock);
1810 if (!s->valid || (unlikely(s->snapshot_overflowed) &&
1811 bio_data_dir(bio) == WRITE)) {
1816 /* If the block is already remapped - use that, else remap it */
1817 e = dm_lookup_exception(&s->complete, chunk);
1819 remap_exception(s, e, bio, chunk);
1824 * Write to snapshot - higher level takes care of RW/RO
1825 * flags so we should only get this if we are
1828 if (bio_data_dir(bio) == WRITE) {
1829 pe = __lookup_pending_exception(s, chunk);
1831 mutex_unlock(&s->lock);
1832 pe = alloc_pending_exception(s);
1833 mutex_lock(&s->lock);
1835 if (!s->valid || s->snapshot_overflowed) {
1836 free_pending_exception(pe);
1841 e = dm_lookup_exception(&s->complete, chunk);
1843 free_pending_exception(pe);
1844 remap_exception(s, e, bio, chunk);
1848 pe = __find_pending_exception(s, pe, chunk);
1850 if (s->store->userspace_supports_overflow) {
1851 s->snapshot_overflowed = 1;
1852 DMERR("Snapshot overflowed: Unable to allocate exception.");
1854 __invalidate_snapshot(s, -ENOMEM);
1860 remap_exception(s, &pe->e, bio, chunk);
1862 r = DM_MAPIO_SUBMITTED;
1865 bio->bi_iter.bi_size ==
1866 (s->store->chunk_size << SECTOR_SHIFT)) {
1868 mutex_unlock(&s->lock);
1869 start_full_bio(pe, bio);
1873 bio_list_add(&pe->snapshot_bios, bio);
1876 /* this is protected by snap->lock */
1878 mutex_unlock(&s->lock);
1883 bio_set_dev(bio, s->origin->bdev);
1884 track_chunk(s, bio, chunk);
1888 mutex_unlock(&s->lock);
1894 * A snapshot-merge target behaves like a combination of a snapshot
1895 * target and a snapshot-origin target. It only generates new
1896 * exceptions in other snapshots and not in the one that is being
1899 * For each chunk, if there is an existing exception, it is used to
1900 * redirect I/O to the cow device. Otherwise I/O is sent to the origin,
1901 * which in turn might generate exceptions in other snapshots.
1902 * If merging is currently taking place on the chunk in question, the
1903 * I/O is deferred by adding it to s->bios_queued_during_merge.
1905 static int snapshot_merge_map(struct dm_target *ti, struct bio *bio)
1907 struct dm_exception *e;
1908 struct dm_snapshot *s = ti->private;
1909 int r = DM_MAPIO_REMAPPED;
1912 init_tracked_chunk(bio);
1914 if (bio->bi_opf & REQ_PREFLUSH) {
1915 if (!dm_bio_get_target_bio_nr(bio))
1916 bio_set_dev(bio, s->origin->bdev);
1918 bio_set_dev(bio, s->cow->bdev);
1919 return DM_MAPIO_REMAPPED;
1922 chunk = sector_to_chunk(s->store, bio->bi_iter.bi_sector);
1924 mutex_lock(&s->lock);
1926 /* Full merging snapshots are redirected to the origin */
1928 goto redirect_to_origin;
1930 /* If the block is already remapped - use that */
1931 e = dm_lookup_exception(&s->complete, chunk);
1933 /* Queue writes overlapping with chunks being merged */
1934 if (bio_data_dir(bio) == WRITE &&
1935 chunk >= s->first_merging_chunk &&
1936 chunk < (s->first_merging_chunk +
1937 s->num_merging_chunks)) {
1938 bio_set_dev(bio, s->origin->bdev);
1939 bio_list_add(&s->bios_queued_during_merge, bio);
1940 r = DM_MAPIO_SUBMITTED;
1944 remap_exception(s, e, bio, chunk);
1946 if (bio_data_dir(bio) == WRITE)
1947 track_chunk(s, bio, chunk);
1952 bio_set_dev(bio, s->origin->bdev);
1954 if (bio_data_dir(bio) == WRITE) {
1955 mutex_unlock(&s->lock);
1956 return do_origin(s->origin, bio, false);
1960 mutex_unlock(&s->lock);
1965 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
1966 blk_status_t *error)
1968 struct dm_snapshot *s = ti->private;
1970 if (is_bio_tracked(bio))
1971 stop_tracking_chunk(s, bio);
1973 return DM_ENDIO_DONE;
1976 static void snapshot_merge_presuspend(struct dm_target *ti)
1978 struct dm_snapshot *s = ti->private;
1983 static int snapshot_preresume(struct dm_target *ti)
1986 struct dm_snapshot *s = ti->private;
1987 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
1989 down_read(&_origins_lock);
1990 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
1991 if (snap_src && snap_dest) {
1992 mutex_lock(&snap_src->lock);
1993 if (s == snap_src) {
1994 DMERR("Unable to resume snapshot source until "
1995 "handover completes.");
1997 } else if (!dm_suspended(snap_src->ti)) {
1998 DMERR("Unable to perform snapshot handover until "
1999 "source is suspended.");
2002 mutex_unlock(&snap_src->lock);
2004 up_read(&_origins_lock);
2009 static void snapshot_resume(struct dm_target *ti)
2011 struct dm_snapshot *s = ti->private;
2012 struct dm_snapshot *snap_src = NULL, *snap_dest = NULL, *snap_merging = NULL;
2013 struct dm_origin *o;
2014 struct mapped_device *origin_md = NULL;
2015 bool must_restart_merging = false;
2017 down_read(&_origins_lock);
2019 o = __lookup_dm_origin(s->origin->bdev);
2021 origin_md = dm_table_get_md(o->ti->table);
2023 (void) __find_snapshots_sharing_cow(s, NULL, NULL, &snap_merging);
2025 origin_md = dm_table_get_md(snap_merging->ti->table);
2027 if (origin_md == dm_table_get_md(ti->table))
2030 if (dm_hold(origin_md))
2034 up_read(&_origins_lock);
2037 dm_internal_suspend_fast(origin_md);
2038 if (snap_merging && test_bit(RUNNING_MERGE, &snap_merging->state_bits)) {
2039 must_restart_merging = true;
2040 stop_merge(snap_merging);
2044 down_read(&_origins_lock);
2046 (void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
2047 if (snap_src && snap_dest) {
2048 mutex_lock(&snap_src->lock);
2049 mutex_lock_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
2050 __handover_exceptions(snap_src, snap_dest);
2051 mutex_unlock(&snap_dest->lock);
2052 mutex_unlock(&snap_src->lock);
2055 up_read(&_origins_lock);
2058 if (must_restart_merging)
2059 start_merge(snap_merging);
2060 dm_internal_resume_fast(origin_md);
2064 /* Now we have correct chunk size, reregister */
2065 reregister_snapshot(s);
2067 mutex_lock(&s->lock);
2069 mutex_unlock(&s->lock);
2072 static uint32_t get_origin_minimum_chunksize(struct block_device *bdev)
2074 uint32_t min_chunksize;
2076 down_read(&_origins_lock);
2077 min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
2078 up_read(&_origins_lock);
2080 return min_chunksize;
2083 static void snapshot_merge_resume(struct dm_target *ti)
2085 struct dm_snapshot *s = ti->private;
2088 * Handover exceptions from existing snapshot.
2090 snapshot_resume(ti);
2093 * snapshot-merge acts as an origin, so set ti->max_io_len
2095 ti->max_io_len = get_origin_minimum_chunksize(s->origin->bdev);
2100 static void snapshot_status(struct dm_target *ti, status_type_t type,
2101 unsigned status_flags, char *result, unsigned maxlen)
2104 struct dm_snapshot *snap = ti->private;
2107 case STATUSTYPE_INFO:
2109 mutex_lock(&snap->lock);
2113 else if (snap->merge_failed)
2114 DMEMIT("Merge failed");
2115 else if (snap->snapshot_overflowed)
2118 if (snap->store->type->usage) {
2119 sector_t total_sectors, sectors_allocated,
2121 snap->store->type->usage(snap->store,
2125 DMEMIT("%llu/%llu %llu",
2126 (unsigned long long)sectors_allocated,
2127 (unsigned long long)total_sectors,
2128 (unsigned long long)metadata_sectors);
2134 mutex_unlock(&snap->lock);
2138 case STATUSTYPE_TABLE:
2140 * kdevname returns a static pointer so we need
2141 * to make private copies if the output is to
2144 DMEMIT("%s %s", snap->origin->name, snap->cow->name);
2145 snap->store->type->status(snap->store, type, result + sz,
2151 static int snapshot_iterate_devices(struct dm_target *ti,
2152 iterate_devices_callout_fn fn, void *data)
2154 struct dm_snapshot *snap = ti->private;
2157 r = fn(ti, snap->origin, 0, ti->len, data);
2160 r = fn(ti, snap->cow, 0, get_dev_size(snap->cow->bdev), data);
2166 /*-----------------------------------------------------------------
2168 *---------------------------------------------------------------*/
2171 * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
2172 * supplied bio was ignored. The caller may submit it immediately.
2173 * (No remapping actually occurs as the origin is always a direct linear
2176 * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
2177 * and any supplied bio is added to a list to be submitted once all
2178 * the necessary exceptions exist.
2180 static int __origin_write(struct list_head *snapshots, sector_t sector,
2183 int r = DM_MAPIO_REMAPPED;
2184 struct dm_snapshot *snap;
2185 struct dm_exception *e;
2186 struct dm_snap_pending_exception *pe;
2187 struct dm_snap_pending_exception *pe_to_start_now = NULL;
2188 struct dm_snap_pending_exception *pe_to_start_last = NULL;
2191 /* Do all the snapshots on this origin */
2192 list_for_each_entry (snap, snapshots, list) {
2194 * Don't make new exceptions in a merging snapshot
2195 * because it has effectively been deleted
2197 if (dm_target_is_snapshot_merge(snap->ti))
2200 mutex_lock(&snap->lock);
2202 /* Only deal with valid and active snapshots */
2203 if (!snap->valid || !snap->active)
2206 /* Nothing to do if writing beyond end of snapshot */
2207 if (sector >= dm_table_get_size(snap->ti->table))
2211 * Remember, different snapshots can have
2212 * different chunk sizes.
2214 chunk = sector_to_chunk(snap->store, sector);
2217 * Check exception table to see if block
2218 * is already remapped in this snapshot
2219 * and trigger an exception if not.
2221 e = dm_lookup_exception(&snap->complete, chunk);
2225 pe = __lookup_pending_exception(snap, chunk);
2227 mutex_unlock(&snap->lock);
2228 pe = alloc_pending_exception(snap);
2229 mutex_lock(&snap->lock);
2232 free_pending_exception(pe);
2236 e = dm_lookup_exception(&snap->complete, chunk);
2238 free_pending_exception(pe);
2242 pe = __find_pending_exception(snap, pe, chunk);
2244 __invalidate_snapshot(snap, -ENOMEM);
2249 r = DM_MAPIO_SUBMITTED;
2252 * If an origin bio was supplied, queue it to wait for the
2253 * completion of this exception, and start this one last,
2254 * at the end of the function.
2257 bio_list_add(&pe->origin_bios, bio);
2262 pe_to_start_last = pe;
2268 pe_to_start_now = pe;
2272 mutex_unlock(&snap->lock);
2274 if (pe_to_start_now) {
2275 start_copy(pe_to_start_now);
2276 pe_to_start_now = NULL;
2281 * Submit the exception against which the bio is queued last,
2282 * to give the other exceptions a head start.
2284 if (pe_to_start_last)
2285 start_copy(pe_to_start_last);
2291 * Called on a write from the origin driver.
2293 static int do_origin(struct dm_dev *origin, struct bio *bio, bool limit)
2296 int r = DM_MAPIO_REMAPPED;
2299 down_read(&_origins_lock);
2300 o = __lookup_origin(origin->bdev);
2303 struct dm_snapshot *s;
2304 list_for_each_entry(s, &o->snapshots, list)
2305 if (unlikely(!wait_for_in_progress(s, true)))
2309 r = __origin_write(&o->snapshots, bio->bi_iter.bi_sector, bio);
2311 up_read(&_origins_lock);
2317 * Trigger exceptions in all non-merging snapshots.
2319 * The chunk size of the merging snapshot may be larger than the chunk
2320 * size of some other snapshot so we may need to reallocate multiple
2321 * chunks in other snapshots.
2323 * We scan all the overlapping exceptions in the other snapshots.
2324 * Returns 1 if anything was reallocated and must be waited for,
2325 * otherwise returns 0.
2327 * size must be a multiple of merging_snap's chunk_size.
2329 static int origin_write_extent(struct dm_snapshot *merging_snap,
2330 sector_t sector, unsigned size)
2337 * The origin's __minimum_chunk_size() got stored in max_io_len
2338 * by snapshot_merge_resume().
2340 down_read(&_origins_lock);
2341 o = __lookup_origin(merging_snap->origin->bdev);
2342 for (n = 0; n < size; n += merging_snap->ti->max_io_len)
2343 if (__origin_write(&o->snapshots, sector + n, NULL) ==
2346 up_read(&_origins_lock);
2352 * Origin: maps a linear range of a device, with hooks for snapshotting.
2356 * Construct an origin mapping: <dev_path>
2357 * The context for an origin is merely a 'struct dm_dev *'
2358 * pointing to the real device.
2360 static int origin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2363 struct dm_origin *o;
2366 ti->error = "origin: incorrect number of arguments";
2370 o = kmalloc(sizeof(struct dm_origin), GFP_KERNEL);
2372 ti->error = "Cannot allocate private origin structure";
2377 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &o->dev);
2379 ti->error = "Cannot get target device";
2385 ti->num_flush_bios = 1;
2395 static void origin_dtr(struct dm_target *ti)
2397 struct dm_origin *o = ti->private;
2399 dm_put_device(ti, o->dev);
2403 static int origin_map(struct dm_target *ti, struct bio *bio)
2405 struct dm_origin *o = ti->private;
2406 unsigned available_sectors;
2408 bio_set_dev(bio, o->dev->bdev);
2410 if (unlikely(bio->bi_opf & REQ_PREFLUSH))
2411 return DM_MAPIO_REMAPPED;
2413 if (bio_data_dir(bio) != WRITE)
2414 return DM_MAPIO_REMAPPED;
2416 available_sectors = o->split_boundary -
2417 ((unsigned)bio->bi_iter.bi_sector & (o->split_boundary - 1));
2419 if (bio_sectors(bio) > available_sectors)
2420 dm_accept_partial_bio(bio, available_sectors);
2422 /* Only tell snapshots if this is a write */
2423 return do_origin(o->dev, bio, true);
2426 static long origin_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
2427 long nr_pages, void **kaddr, pfn_t *pfn)
2429 DMWARN("device does not support dax.");
2434 * Set the target "max_io_len" field to the minimum of all the snapshots'
2437 static void origin_resume(struct dm_target *ti)
2439 struct dm_origin *o = ti->private;
2441 o->split_boundary = get_origin_minimum_chunksize(o->dev->bdev);
2443 down_write(&_origins_lock);
2444 __insert_dm_origin(o);
2445 up_write(&_origins_lock);
2448 static void origin_postsuspend(struct dm_target *ti)
2450 struct dm_origin *o = ti->private;
2452 down_write(&_origins_lock);
2453 __remove_dm_origin(o);
2454 up_write(&_origins_lock);
2457 static void origin_status(struct dm_target *ti, status_type_t type,
2458 unsigned status_flags, char *result, unsigned maxlen)
2460 struct dm_origin *o = ti->private;
2463 case STATUSTYPE_INFO:
2467 case STATUSTYPE_TABLE:
2468 snprintf(result, maxlen, "%s", o->dev->name);
2473 static int origin_iterate_devices(struct dm_target *ti,
2474 iterate_devices_callout_fn fn, void *data)
2476 struct dm_origin *o = ti->private;
2478 return fn(ti, o->dev, 0, ti->len, data);
2481 static struct target_type origin_target = {
2482 .name = "snapshot-origin",
2483 .version = {1, 9, 0},
2484 .module = THIS_MODULE,
2488 .resume = origin_resume,
2489 .postsuspend = origin_postsuspend,
2490 .status = origin_status,
2491 .iterate_devices = origin_iterate_devices,
2492 .direct_access = origin_dax_direct_access,
2495 static struct target_type snapshot_target = {
2497 .version = {1, 15, 0},
2498 .module = THIS_MODULE,
2499 .ctr = snapshot_ctr,
2500 .dtr = snapshot_dtr,
2501 .map = snapshot_map,
2502 .end_io = snapshot_end_io,
2503 .preresume = snapshot_preresume,
2504 .resume = snapshot_resume,
2505 .status = snapshot_status,
2506 .iterate_devices = snapshot_iterate_devices,
2509 static struct target_type merge_target = {
2510 .name = dm_snapshot_merge_target_name,
2511 .version = {1, 4, 0},
2512 .module = THIS_MODULE,
2513 .ctr = snapshot_ctr,
2514 .dtr = snapshot_dtr,
2515 .map = snapshot_merge_map,
2516 .end_io = snapshot_end_io,
2517 .presuspend = snapshot_merge_presuspend,
2518 .preresume = snapshot_preresume,
2519 .resume = snapshot_merge_resume,
2520 .status = snapshot_status,
2521 .iterate_devices = snapshot_iterate_devices,
2524 static int __init dm_snapshot_init(void)
2528 r = dm_exception_store_init();
2530 DMERR("Failed to initialize exception stores");
2534 r = init_origin_hash();
2536 DMERR("init_origin_hash failed.");
2537 goto bad_origin_hash;
2540 exception_cache = KMEM_CACHE(dm_exception, 0);
2541 if (!exception_cache) {
2542 DMERR("Couldn't create exception cache.");
2544 goto bad_exception_cache;
2547 pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
2548 if (!pending_cache) {
2549 DMERR("Couldn't create pending cache.");
2551 goto bad_pending_cache;
2554 r = dm_register_target(&snapshot_target);
2556 DMERR("snapshot target register failed %d", r);
2557 goto bad_register_snapshot_target;
2560 r = dm_register_target(&origin_target);
2562 DMERR("Origin target register failed %d", r);
2563 goto bad_register_origin_target;
2566 r = dm_register_target(&merge_target);
2568 DMERR("Merge target register failed %d", r);
2569 goto bad_register_merge_target;
2574 bad_register_merge_target:
2575 dm_unregister_target(&origin_target);
2576 bad_register_origin_target:
2577 dm_unregister_target(&snapshot_target);
2578 bad_register_snapshot_target:
2579 kmem_cache_destroy(pending_cache);
2581 kmem_cache_destroy(exception_cache);
2582 bad_exception_cache:
2585 dm_exception_store_exit();
2590 static void __exit dm_snapshot_exit(void)
2592 dm_unregister_target(&snapshot_target);
2593 dm_unregister_target(&origin_target);
2594 dm_unregister_target(&merge_target);
2597 kmem_cache_destroy(pending_cache);
2598 kmem_cache_destroy(exception_cache);
2600 dm_exception_store_exit();
2604 module_init(dm_snapshot_init);
2605 module_exit(dm_snapshot_exit);
2607 MODULE_DESCRIPTION(DM_NAME " snapshot target");
2608 MODULE_AUTHOR("Joe Thornber");
2609 MODULE_LICENSE("GPL");
2610 MODULE_ALIAS("dm-snapshot-origin");
2611 MODULE_ALIAS("dm-snapshot-merge");