2 * Copyright (C) 2011 Red Hat UK.
4 * This file is released under the GPL.
7 #include "dm-thin-metadata.h"
9 #include <linux/device-mapper.h>
10 #include <linux/dm-io.h>
11 #include <linux/dm-kcopyd.h>
12 #include <linux/list.h>
13 #include <linux/init.h>
14 #include <linux/module.h>
15 #include <linux/slab.h>
17 #define DM_MSG_PREFIX "thin"
22 #define ENDIO_HOOK_POOL_SIZE 10240
23 #define DEFERRED_SET_SIZE 64
24 #define MAPPING_POOL_SIZE 1024
25 #define PRISON_CELLS 1024
26 #define COMMIT_PERIOD HZ
29 * The block size of the device holding pool data must be
30 * between 64KB and 1GB.
32 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
33 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
36 * Device id is restricted to 24 bits.
38 #define MAX_DEV_ID ((1 << 24) - 1)
41 * How do we handle breaking sharing of data blocks?
42 * =================================================
44 * We use a standard copy-on-write btree to store the mappings for the
45 * devices (note I'm talking about copy-on-write of the metadata here, not
46 * the data). When you take an internal snapshot you clone the root node
47 * of the origin btree. After this there is no concept of an origin or a
48 * snapshot. They are just two device trees that happen to point to the
51 * When we get a write in we decide if it's to a shared data block using
52 * some timestamp magic. If it is, we have to break sharing.
54 * Let's say we write to a shared block in what was the origin. The
57 * i) plug io further to this physical block. (see bio_prison code).
59 * ii) quiesce any read io to that shared data block. Obviously
60 * including all devices that share this block. (see deferred_set code)
62 * iii) copy the data block to a newly allocate block. This step can be
63 * missed out if the io covers the block. (schedule_copy).
65 * iv) insert the new mapping into the origin's btree
66 * (process_prepared_mapping). This act of inserting breaks some
67 * sharing of btree nodes between the two devices. Breaking sharing only
68 * effects the btree of that specific device. Btrees for the other
69 * devices that share the block never change. The btree for the origin
70 * device as it was after the last commit is untouched, ie. we're using
71 * persistent data structures in the functional programming sense.
73 * v) unplug io to this physical block, including the io that triggered
74 * the breaking of sharing.
76 * Steps (ii) and (iii) occur in parallel.
78 * The metadata _doesn't_ need to be committed before the io continues. We
79 * get away with this because the io is always written to a _new_ block.
80 * If there's a crash, then:
82 * - The origin mapping will point to the old origin block (the shared
83 * one). This will contain the data as it was before the io that triggered
84 * the breaking of sharing came in.
86 * - The snap mapping still points to the old block. As it would after
89 * The downside of this scheme is the timestamp magic isn't perfect, and
90 * will continue to think that data block in the snapshot device is shared
91 * even after the write to the origin has broken sharing. I suspect data
92 * blocks will typically be shared by many different devices, so we're
93 * breaking sharing n + 1 times, rather than n, where n is the number of
94 * devices that reference this data block. At the moment I think the
95 * benefits far, far outweigh the disadvantages.
98 /*----------------------------------------------------------------*/
101 * Sometimes we can't deal with a bio straight away. We put them in prison
102 * where they can't cause any mischief. Bios are put in a cell identified
103 * by a key, multiple bios can be in the same cell. When the cell is
104 * subsequently unlocked the bios become available.
115 struct hlist_node list;
116 struct bio_prison *prison;
119 struct bio_list bios;
124 mempool_t *cell_pool;
128 struct hlist_head *cells;
131 static uint32_t calc_nr_buckets(unsigned nr_cells)
136 nr_cells = min(nr_cells, 8192u);
145 * @nr_cells should be the number of cells you want in use _concurrently_.
146 * Don't confuse it with the number of distinct keys.
148 static struct bio_prison *prison_create(unsigned nr_cells)
151 uint32_t nr_buckets = calc_nr_buckets(nr_cells);
152 size_t len = sizeof(struct bio_prison) +
153 (sizeof(struct hlist_head) * nr_buckets);
154 struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
159 spin_lock_init(&prison->lock);
160 prison->cell_pool = mempool_create_kmalloc_pool(nr_cells,
161 sizeof(struct cell));
162 if (!prison->cell_pool) {
167 prison->nr_buckets = nr_buckets;
168 prison->hash_mask = nr_buckets - 1;
169 prison->cells = (struct hlist_head *) (prison + 1);
170 for (i = 0; i < nr_buckets; i++)
171 INIT_HLIST_HEAD(prison->cells + i);
176 static void prison_destroy(struct bio_prison *prison)
178 mempool_destroy(prison->cell_pool);
182 static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
184 const unsigned long BIG_PRIME = 4294967291UL;
185 uint64_t hash = key->block * BIG_PRIME;
187 return (uint32_t) (hash & prison->hash_mask);
190 static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
192 return (lhs->virtual == rhs->virtual) &&
193 (lhs->dev == rhs->dev) &&
194 (lhs->block == rhs->block);
197 static struct cell *__search_bucket(struct hlist_head *bucket,
198 struct cell_key *key)
201 struct hlist_node *tmp;
203 hlist_for_each_entry(cell, tmp, bucket, list)
204 if (keys_equal(&cell->key, key))
211 * This may block if a new cell needs allocating. You must ensure that
212 * cells will be unlocked even if the calling thread is blocked.
214 * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
216 static int bio_detain(struct bio_prison *prison, struct cell_key *key,
217 struct bio *inmate, struct cell **ref)
221 uint32_t hash = hash_key(prison, key);
222 struct cell *cell, *cell2;
224 BUG_ON(hash > prison->nr_buckets);
226 spin_lock_irqsave(&prison->lock, flags);
228 cell = __search_bucket(prison->cells + hash, key);
230 bio_list_add(&cell->bios, inmate);
235 * Allocate a new cell
237 spin_unlock_irqrestore(&prison->lock, flags);
238 cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
239 spin_lock_irqsave(&prison->lock, flags);
242 * We've been unlocked, so we have to double check that
243 * nobody else has inserted this cell in the meantime.
245 cell = __search_bucket(prison->cells + hash, key);
247 mempool_free(cell2, prison->cell_pool);
248 bio_list_add(&cell->bios, inmate);
257 cell->prison = prison;
258 memcpy(&cell->key, key, sizeof(cell->key));
259 cell->holder = inmate;
260 bio_list_init(&cell->bios);
261 hlist_add_head(&cell->list, prison->cells + hash);
266 spin_unlock_irqrestore(&prison->lock, flags);
274 * @inmates must have been initialised prior to this call
276 static void __cell_release(struct cell *cell, struct bio_list *inmates)
278 struct bio_prison *prison = cell->prison;
280 hlist_del(&cell->list);
283 bio_list_add(inmates, cell->holder);
284 bio_list_merge(inmates, &cell->bios);
287 mempool_free(cell, prison->cell_pool);
290 static void cell_release(struct cell *cell, struct bio_list *bios)
293 struct bio_prison *prison = cell->prison;
295 spin_lock_irqsave(&prison->lock, flags);
296 __cell_release(cell, bios);
297 spin_unlock_irqrestore(&prison->lock, flags);
301 * There are a couple of places where we put a bio into a cell briefly
302 * before taking it out again. In these situations we know that no other
303 * bio may be in the cell. This function releases the cell, and also does
306 static void __cell_release_singleton(struct cell *cell, struct bio *bio)
308 BUG_ON(cell->holder != bio);
309 BUG_ON(!bio_list_empty(&cell->bios));
311 __cell_release(cell, NULL);
314 static void cell_release_singleton(struct cell *cell, struct bio *bio)
317 struct bio_prison *prison = cell->prison;
319 spin_lock_irqsave(&prison->lock, flags);
320 __cell_release_singleton(cell, bio);
321 spin_unlock_irqrestore(&prison->lock, flags);
325 * Sometimes we don't want the holder, just the additional bios.
327 static void __cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
329 struct bio_prison *prison = cell->prison;
331 hlist_del(&cell->list);
332 bio_list_merge(inmates, &cell->bios);
334 mempool_free(cell, prison->cell_pool);
337 static void cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
340 struct bio_prison *prison = cell->prison;
342 spin_lock_irqsave(&prison->lock, flags);
343 __cell_release_no_holder(cell, inmates);
344 spin_unlock_irqrestore(&prison->lock, flags);
347 static void cell_error(struct cell *cell)
349 struct bio_prison *prison = cell->prison;
350 struct bio_list bios;
354 bio_list_init(&bios);
356 spin_lock_irqsave(&prison->lock, flags);
357 __cell_release(cell, &bios);
358 spin_unlock_irqrestore(&prison->lock, flags);
360 while ((bio = bio_list_pop(&bios)))
364 /*----------------------------------------------------------------*/
367 * We use the deferred set to keep track of pending reads to shared blocks.
368 * We do this to ensure the new mapping caused by a write isn't performed
369 * until these prior reads have completed. Otherwise the insertion of the
370 * new mapping could free the old block that the read bios are mapped to.
374 struct deferred_entry {
375 struct deferred_set *ds;
377 struct list_head work_items;
380 struct deferred_set {
382 unsigned current_entry;
384 struct deferred_entry entries[DEFERRED_SET_SIZE];
387 static void ds_init(struct deferred_set *ds)
391 spin_lock_init(&ds->lock);
392 ds->current_entry = 0;
394 for (i = 0; i < DEFERRED_SET_SIZE; i++) {
395 ds->entries[i].ds = ds;
396 ds->entries[i].count = 0;
397 INIT_LIST_HEAD(&ds->entries[i].work_items);
401 static struct deferred_entry *ds_inc(struct deferred_set *ds)
404 struct deferred_entry *entry;
406 spin_lock_irqsave(&ds->lock, flags);
407 entry = ds->entries + ds->current_entry;
409 spin_unlock_irqrestore(&ds->lock, flags);
414 static unsigned ds_next(unsigned index)
416 return (index + 1) % DEFERRED_SET_SIZE;
419 static void __sweep(struct deferred_set *ds, struct list_head *head)
421 while ((ds->sweeper != ds->current_entry) &&
422 !ds->entries[ds->sweeper].count) {
423 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
424 ds->sweeper = ds_next(ds->sweeper);
427 if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
428 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
431 static void ds_dec(struct deferred_entry *entry, struct list_head *head)
435 spin_lock_irqsave(&entry->ds->lock, flags);
436 BUG_ON(!entry->count);
438 __sweep(entry->ds, head);
439 spin_unlock_irqrestore(&entry->ds->lock, flags);
443 * Returns 1 if deferred or 0 if no pending items to delay job.
445 static int ds_add_work(struct deferred_set *ds, struct list_head *work)
451 spin_lock_irqsave(&ds->lock, flags);
452 if ((ds->sweeper == ds->current_entry) &&
453 !ds->entries[ds->current_entry].count)
456 list_add(work, &ds->entries[ds->current_entry].work_items);
457 next_entry = ds_next(ds->current_entry);
458 if (!ds->entries[next_entry].count)
459 ds->current_entry = next_entry;
461 spin_unlock_irqrestore(&ds->lock, flags);
466 /*----------------------------------------------------------------*/
471 static void build_data_key(struct dm_thin_device *td,
472 dm_block_t b, struct cell_key *key)
475 key->dev = dm_thin_dev_id(td);
479 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
480 struct cell_key *key)
483 key->dev = dm_thin_dev_id(td);
487 /*----------------------------------------------------------------*/
490 * A pool device ties together a metadata device and a data device. It
491 * also provides the interface for creating and destroying internal
496 struct pool_features {
497 unsigned zero_new_blocks:1;
498 unsigned discard_enabled:1;
499 unsigned discard_passdown:1;
503 struct list_head list;
504 struct dm_target *ti; /* Only set if a pool target is bound */
506 struct mapped_device *pool_md;
507 struct block_device *md_dev;
508 struct dm_pool_metadata *pmd;
510 uint32_t sectors_per_block;
511 unsigned block_shift;
512 dm_block_t offset_mask;
513 dm_block_t low_water_blocks;
515 struct pool_features pf;
516 unsigned low_water_triggered:1; /* A dm event has been sent */
517 unsigned no_free_space:1; /* A -ENOSPC warning has been issued */
519 struct bio_prison *prison;
520 struct dm_kcopyd_client *copier;
522 struct workqueue_struct *wq;
523 struct work_struct worker;
524 struct delayed_work waker;
527 unsigned long last_commit_jiffies;
530 struct bio_list deferred_bios;
531 struct bio_list deferred_flush_bios;
532 struct list_head prepared_mappings;
533 struct list_head prepared_discards;
535 struct bio_list retry_on_resume_list;
537 struct deferred_set shared_read_ds;
538 struct deferred_set all_io_ds;
540 struct new_mapping *next_mapping;
541 mempool_t *mapping_pool;
542 mempool_t *endio_hook_pool;
546 * Target context for a pool.
549 struct dm_target *ti;
551 struct dm_dev *data_dev;
552 struct dm_dev *metadata_dev;
553 struct dm_target_callbacks callbacks;
555 dm_block_t low_water_blocks;
556 struct pool_features pf;
560 * Target context for a thin.
563 struct dm_dev *pool_dev;
564 struct dm_dev *origin_dev;
568 struct dm_thin_device *td;
571 /*----------------------------------------------------------------*/
574 * A global list of pools that uses a struct mapped_device as a key.
576 static struct dm_thin_pool_table {
578 struct list_head pools;
579 } dm_thin_pool_table;
581 static void pool_table_init(void)
583 mutex_init(&dm_thin_pool_table.mutex);
584 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
587 static void __pool_table_insert(struct pool *pool)
589 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
590 list_add(&pool->list, &dm_thin_pool_table.pools);
593 static void __pool_table_remove(struct pool *pool)
595 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
596 list_del(&pool->list);
599 static struct pool *__pool_table_lookup(struct mapped_device *md)
601 struct pool *pool = NULL, *tmp;
603 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
605 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
606 if (tmp->pool_md == md) {
615 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
617 struct pool *pool = NULL, *tmp;
619 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
621 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
622 if (tmp->md_dev == md_dev) {
631 /*----------------------------------------------------------------*/
635 struct deferred_entry *shared_read_entry;
636 struct deferred_entry *all_io_entry;
637 struct new_mapping *overwrite_mapping;
640 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
643 struct bio_list bios;
645 bio_list_init(&bios);
646 bio_list_merge(&bios, master);
647 bio_list_init(master);
649 while ((bio = bio_list_pop(&bios))) {
650 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
652 bio_endio(bio, DM_ENDIO_REQUEUE);
654 bio_list_add(master, bio);
658 static void requeue_io(struct thin_c *tc)
660 struct pool *pool = tc->pool;
663 spin_lock_irqsave(&pool->lock, flags);
664 __requeue_bio_list(tc, &pool->deferred_bios);
665 __requeue_bio_list(tc, &pool->retry_on_resume_list);
666 spin_unlock_irqrestore(&pool->lock, flags);
670 * This section of code contains the logic for processing a thin device's IO.
671 * Much of the code depends on pool object resources (lists, workqueues, etc)
672 * but most is exclusively called from the thin target rather than the thin-pool
676 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
678 return bio->bi_sector >> tc->pool->block_shift;
681 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
683 struct pool *pool = tc->pool;
685 bio->bi_bdev = tc->pool_dev->bdev;
686 bio->bi_sector = (block << pool->block_shift) +
687 (bio->bi_sector & pool->offset_mask);
690 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
692 bio->bi_bdev = tc->origin_dev->bdev;
695 static void issue(struct thin_c *tc, struct bio *bio)
697 struct pool *pool = tc->pool;
701 * Batch together any FUA/FLUSH bios we find and then issue
702 * a single commit for them in process_deferred_bios().
704 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
705 spin_lock_irqsave(&pool->lock, flags);
706 bio_list_add(&pool->deferred_flush_bios, bio);
707 spin_unlock_irqrestore(&pool->lock, flags);
709 generic_make_request(bio);
712 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
714 remap_to_origin(tc, bio);
718 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
721 remap(tc, bio, block);
726 * wake_worker() is used when new work is queued and when pool_resume is
727 * ready to continue deferred IO processing.
729 static void wake_worker(struct pool *pool)
731 queue_work(pool->wq, &pool->worker);
734 /*----------------------------------------------------------------*/
737 * Bio endio functions.
740 struct list_head list;
744 unsigned pass_discard:1;
747 dm_block_t virt_block;
748 dm_block_t data_block;
749 struct cell *cell, *cell2;
753 * If the bio covers the whole area of a block then we can avoid
754 * zeroing or copying. Instead this bio is hooked. The bio will
755 * still be in the cell, so care has to be taken to avoid issuing
759 bio_end_io_t *saved_bi_end_io;
762 static void __maybe_add_mapping(struct new_mapping *m)
764 struct pool *pool = m->tc->pool;
766 if (m->quiesced && m->prepared) {
767 list_add(&m->list, &pool->prepared_mappings);
772 static void copy_complete(int read_err, unsigned long write_err, void *context)
775 struct new_mapping *m = context;
776 struct pool *pool = m->tc->pool;
778 m->err = read_err || write_err ? -EIO : 0;
780 spin_lock_irqsave(&pool->lock, flags);
782 __maybe_add_mapping(m);
783 spin_unlock_irqrestore(&pool->lock, flags);
786 static void overwrite_endio(struct bio *bio, int err)
789 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
790 struct new_mapping *m = h->overwrite_mapping;
791 struct pool *pool = m->tc->pool;
795 spin_lock_irqsave(&pool->lock, flags);
797 __maybe_add_mapping(m);
798 spin_unlock_irqrestore(&pool->lock, flags);
801 /*----------------------------------------------------------------*/
808 * Prepared mapping jobs.
812 * This sends the bios in the cell back to the deferred_bios list.
814 static void cell_defer(struct thin_c *tc, struct cell *cell,
815 dm_block_t data_block)
817 struct pool *pool = tc->pool;
820 spin_lock_irqsave(&pool->lock, flags);
821 cell_release(cell, &pool->deferred_bios);
822 spin_unlock_irqrestore(&tc->pool->lock, flags);
828 * Same as cell_defer above, except it omits one particular detainee,
829 * a write bio that covers the block and has already been processed.
831 static void cell_defer_except(struct thin_c *tc, struct cell *cell)
833 struct bio_list bios;
834 struct pool *pool = tc->pool;
837 bio_list_init(&bios);
839 spin_lock_irqsave(&pool->lock, flags);
840 cell_release_no_holder(cell, &pool->deferred_bios);
841 spin_unlock_irqrestore(&pool->lock, flags);
846 static void process_prepared_mapping(struct new_mapping *m)
848 struct thin_c *tc = m->tc;
854 bio->bi_end_io = m->saved_bi_end_io;
862 * Commit the prepared block into the mapping btree.
863 * Any I/O for this block arriving after this point will get
864 * remapped to it directly.
866 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
868 DMERR("dm_thin_insert_block() failed");
874 * Release any bios held while the block was being provisioned.
875 * If we are processing a write bio that completely covers the block,
876 * we already processed it so can ignore it now when processing
877 * the bios in the cell.
880 cell_defer_except(tc, m->cell);
883 cell_defer(tc, m->cell, m->data_block);
886 mempool_free(m, tc->pool->mapping_pool);
889 static void process_prepared_discard(struct new_mapping *m)
892 struct thin_c *tc = m->tc;
894 r = dm_thin_remove_block(tc->td, m->virt_block);
896 DMERR("dm_thin_remove_block() failed");
899 * Pass the discard down to the underlying device?
902 remap_and_issue(tc, m->bio, m->data_block);
904 bio_endio(m->bio, 0);
906 cell_defer_except(tc, m->cell);
907 cell_defer_except(tc, m->cell2);
908 mempool_free(m, tc->pool->mapping_pool);
911 static void process_prepared(struct pool *pool, struct list_head *head,
912 void (*fn)(struct new_mapping *))
915 struct list_head maps;
916 struct new_mapping *m, *tmp;
918 INIT_LIST_HEAD(&maps);
919 spin_lock_irqsave(&pool->lock, flags);
920 list_splice_init(head, &maps);
921 spin_unlock_irqrestore(&pool->lock, flags);
923 list_for_each_entry_safe(m, tmp, &maps, list)
930 static int io_overlaps_block(struct pool *pool, struct bio *bio)
932 return !(bio->bi_sector & pool->offset_mask) &&
933 (bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
937 static int io_overwrites_block(struct pool *pool, struct bio *bio)
939 return (bio_data_dir(bio) == WRITE) &&
940 io_overlaps_block(pool, bio);
943 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
946 *save = bio->bi_end_io;
950 static int ensure_next_mapping(struct pool *pool)
952 if (pool->next_mapping)
955 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
957 return pool->next_mapping ? 0 : -ENOMEM;
960 static struct new_mapping *get_next_mapping(struct pool *pool)
962 struct new_mapping *r = pool->next_mapping;
964 BUG_ON(!pool->next_mapping);
966 pool->next_mapping = NULL;
971 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
972 struct dm_dev *origin, dm_block_t data_origin,
973 dm_block_t data_dest,
974 struct cell *cell, struct bio *bio)
977 struct pool *pool = tc->pool;
978 struct new_mapping *m = get_next_mapping(pool);
980 INIT_LIST_HEAD(&m->list);
984 m->virt_block = virt_block;
985 m->data_block = data_dest;
990 if (!ds_add_work(&pool->shared_read_ds, &m->list))
994 * IO to pool_dev remaps to the pool target's data_dev.
996 * If the whole block of data is being overwritten, we can issue the
997 * bio immediately. Otherwise we use kcopyd to clone the data first.
999 if (io_overwrites_block(pool, bio)) {
1000 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1001 h->overwrite_mapping = m;
1003 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1004 remap_and_issue(tc, bio, data_dest);
1006 struct dm_io_region from, to;
1008 from.bdev = origin->bdev;
1009 from.sector = data_origin * pool->sectors_per_block;
1010 from.count = pool->sectors_per_block;
1012 to.bdev = tc->pool_dev->bdev;
1013 to.sector = data_dest * pool->sectors_per_block;
1014 to.count = pool->sectors_per_block;
1016 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1017 0, copy_complete, m);
1019 mempool_free(m, pool->mapping_pool);
1020 DMERR("dm_kcopyd_copy() failed");
1026 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1027 dm_block_t data_origin, dm_block_t data_dest,
1028 struct cell *cell, struct bio *bio)
1030 schedule_copy(tc, virt_block, tc->pool_dev,
1031 data_origin, data_dest, cell, bio);
1034 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1035 dm_block_t data_dest,
1036 struct cell *cell, struct bio *bio)
1038 schedule_copy(tc, virt_block, tc->origin_dev,
1039 virt_block, data_dest, cell, bio);
1042 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1043 dm_block_t data_block, struct cell *cell,
1046 struct pool *pool = tc->pool;
1047 struct new_mapping *m = get_next_mapping(pool);
1049 INIT_LIST_HEAD(&m->list);
1053 m->virt_block = virt_block;
1054 m->data_block = data_block;
1060 * If the whole block of data is being overwritten or we are not
1061 * zeroing pre-existing data, we can issue the bio immediately.
1062 * Otherwise we use kcopyd to zero the data first.
1064 if (!pool->pf.zero_new_blocks)
1065 process_prepared_mapping(m);
1067 else if (io_overwrites_block(pool, bio)) {
1068 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1069 h->overwrite_mapping = m;
1071 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1072 remap_and_issue(tc, bio, data_block);
1076 struct dm_io_region to;
1078 to.bdev = tc->pool_dev->bdev;
1079 to.sector = data_block * pool->sectors_per_block;
1080 to.count = pool->sectors_per_block;
1082 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
1084 mempool_free(m, pool->mapping_pool);
1085 DMERR("dm_kcopyd_zero() failed");
1091 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1094 dm_block_t free_blocks;
1095 unsigned long flags;
1096 struct pool *pool = tc->pool;
1098 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1102 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1103 DMWARN("%s: reached low water mark, sending event.",
1104 dm_device_name(pool->pool_md));
1105 spin_lock_irqsave(&pool->lock, flags);
1106 pool->low_water_triggered = 1;
1107 spin_unlock_irqrestore(&pool->lock, flags);
1108 dm_table_event(pool->ti->table);
1112 if (pool->no_free_space)
1116 * Try to commit to see if that will free up some
1119 r = dm_pool_commit_metadata(pool->pmd);
1121 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1126 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1131 * If we still have no space we set a flag to avoid
1132 * doing all this checking and return -ENOSPC.
1135 DMWARN("%s: no free space available.",
1136 dm_device_name(pool->pool_md));
1137 spin_lock_irqsave(&pool->lock, flags);
1138 pool->no_free_space = 1;
1139 spin_unlock_irqrestore(&pool->lock, flags);
1145 r = dm_pool_alloc_data_block(pool->pmd, result);
1153 * If we have run out of space, queue bios until the device is
1154 * resumed, presumably after having been reloaded with more space.
1156 static void retry_on_resume(struct bio *bio)
1158 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1159 struct thin_c *tc = h->tc;
1160 struct pool *pool = tc->pool;
1161 unsigned long flags;
1163 spin_lock_irqsave(&pool->lock, flags);
1164 bio_list_add(&pool->retry_on_resume_list, bio);
1165 spin_unlock_irqrestore(&pool->lock, flags);
1168 static void no_space(struct cell *cell)
1171 struct bio_list bios;
1173 bio_list_init(&bios);
1174 cell_release(cell, &bios);
1176 while ((bio = bio_list_pop(&bios)))
1177 retry_on_resume(bio);
1180 static void process_discard(struct thin_c *tc, struct bio *bio)
1183 struct pool *pool = tc->pool;
1184 struct cell *cell, *cell2;
1185 struct cell_key key, key2;
1186 dm_block_t block = get_bio_block(tc, bio);
1187 struct dm_thin_lookup_result lookup_result;
1188 struct new_mapping *m;
1190 build_virtual_key(tc->td, block, &key);
1191 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1194 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1198 * Check nobody is fiddling with this pool block. This can
1199 * happen if someone's in the process of breaking sharing
1202 build_data_key(tc->td, lookup_result.block, &key2);
1203 if (bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
1204 cell_release_singleton(cell, bio);
1208 if (io_overlaps_block(pool, bio)) {
1210 * IO may still be going to the destination block. We must
1211 * quiesce before we can do the removal.
1213 m = get_next_mapping(pool);
1215 m->pass_discard = (!lookup_result.shared) & pool->pf.discard_passdown;
1216 m->virt_block = block;
1217 m->data_block = lookup_result.block;
1223 if (!ds_add_work(&pool->all_io_ds, &m->list)) {
1224 list_add(&m->list, &pool->prepared_discards);
1229 * This path is hit if people are ignoring
1230 * limits->discard_granularity. It ignores any
1231 * part of the discard that is in a subsequent
1234 sector_t offset = bio->bi_sector - (block << pool->block_shift);
1235 unsigned remaining = (pool->sectors_per_block - offset) << 9;
1236 bio->bi_size = min(bio->bi_size, remaining);
1238 cell_release_singleton(cell, bio);
1239 cell_release_singleton(cell2, bio);
1240 remap_and_issue(tc, bio, lookup_result.block);
1246 * It isn't provisioned, just forget it.
1248 cell_release_singleton(cell, bio);
1253 DMERR("discard: find block unexpectedly returned %d", r);
1254 cell_release_singleton(cell, bio);
1260 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1261 struct cell_key *key,
1262 struct dm_thin_lookup_result *lookup_result,
1266 dm_block_t data_block;
1268 r = alloc_data_block(tc, &data_block);
1271 schedule_internal_copy(tc, block, lookup_result->block,
1272 data_block, cell, bio);
1280 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1286 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1288 struct dm_thin_lookup_result *lookup_result)
1291 struct pool *pool = tc->pool;
1292 struct cell_key key;
1295 * If cell is already occupied, then sharing is already in the process
1296 * of being broken so we have nothing further to do here.
1298 build_data_key(tc->td, lookup_result->block, &key);
1299 if (bio_detain(pool->prison, &key, bio, &cell))
1302 if (bio_data_dir(bio) == WRITE)
1303 break_sharing(tc, bio, block, &key, lookup_result, cell);
1305 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1307 h->shared_read_entry = ds_inc(&pool->shared_read_ds);
1309 cell_release_singleton(cell, bio);
1310 remap_and_issue(tc, bio, lookup_result->block);
1314 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1318 dm_block_t data_block;
1321 * Remap empty bios (flushes) immediately, without provisioning.
1323 if (!bio->bi_size) {
1324 cell_release_singleton(cell, bio);
1325 remap_and_issue(tc, bio, 0);
1330 * Fill read bios with zeroes and complete them immediately.
1332 if (bio_data_dir(bio) == READ) {
1334 cell_release_singleton(cell, bio);
1339 r = alloc_data_block(tc, &data_block);
1343 schedule_external_copy(tc, block, data_block, cell, bio);
1345 schedule_zero(tc, block, data_block, cell, bio);
1353 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1359 static void process_bio(struct thin_c *tc, struct bio *bio)
1362 dm_block_t block = get_bio_block(tc, bio);
1364 struct cell_key key;
1365 struct dm_thin_lookup_result lookup_result;
1368 * If cell is already occupied, then the block is already
1369 * being provisioned so we have nothing further to do here.
1371 build_virtual_key(tc->td, block, &key);
1372 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1375 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1379 * We can release this cell now. This thread is the only
1380 * one that puts bios into a cell, and we know there were
1381 * no preceding bios.
1384 * TODO: this will probably have to change when discard goes
1387 cell_release_singleton(cell, bio);
1389 if (lookup_result.shared)
1390 process_shared_bio(tc, bio, block, &lookup_result);
1392 remap_and_issue(tc, bio, lookup_result.block);
1396 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1397 cell_release_singleton(cell, bio);
1398 remap_to_origin_and_issue(tc, bio);
1400 provision_block(tc, bio, block, cell);
1404 DMERR("dm_thin_find_block() failed, error = %d", r);
1405 cell_release_singleton(cell, bio);
1411 static int need_commit_due_to_time(struct pool *pool)
1413 return jiffies < pool->last_commit_jiffies ||
1414 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1417 static void process_deferred_bios(struct pool *pool)
1419 unsigned long flags;
1421 struct bio_list bios;
1424 bio_list_init(&bios);
1426 spin_lock_irqsave(&pool->lock, flags);
1427 bio_list_merge(&bios, &pool->deferred_bios);
1428 bio_list_init(&pool->deferred_bios);
1429 spin_unlock_irqrestore(&pool->lock, flags);
1431 while ((bio = bio_list_pop(&bios))) {
1432 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1433 struct thin_c *tc = h->tc;
1436 * If we've got no free new_mapping structs, and processing
1437 * this bio might require one, we pause until there are some
1438 * prepared mappings to process.
1440 if (ensure_next_mapping(pool)) {
1441 spin_lock_irqsave(&pool->lock, flags);
1442 bio_list_merge(&pool->deferred_bios, &bios);
1443 spin_unlock_irqrestore(&pool->lock, flags);
1448 if (bio->bi_rw & REQ_DISCARD)
1449 process_discard(tc, bio);
1451 process_bio(tc, bio);
1455 * If there are any deferred flush bios, we must commit
1456 * the metadata before issuing them.
1458 bio_list_init(&bios);
1459 spin_lock_irqsave(&pool->lock, flags);
1460 bio_list_merge(&bios, &pool->deferred_flush_bios);
1461 bio_list_init(&pool->deferred_flush_bios);
1462 spin_unlock_irqrestore(&pool->lock, flags);
1464 if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1467 r = dm_pool_commit_metadata(pool->pmd);
1469 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1471 while ((bio = bio_list_pop(&bios)))
1475 pool->last_commit_jiffies = jiffies;
1477 while ((bio = bio_list_pop(&bios)))
1478 generic_make_request(bio);
1481 static void do_worker(struct work_struct *ws)
1483 struct pool *pool = container_of(ws, struct pool, worker);
1485 process_prepared(pool, &pool->prepared_mappings, process_prepared_mapping);
1486 process_prepared(pool, &pool->prepared_discards, process_prepared_discard);
1487 process_deferred_bios(pool);
1491 * We want to commit periodically so that not too much
1492 * unwritten data builds up.
1494 static void do_waker(struct work_struct *ws)
1496 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1498 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1501 /*----------------------------------------------------------------*/
1504 * Mapping functions.
1508 * Called only while mapping a thin bio to hand it over to the workqueue.
1510 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1512 unsigned long flags;
1513 struct pool *pool = tc->pool;
1515 spin_lock_irqsave(&pool->lock, flags);
1516 bio_list_add(&pool->deferred_bios, bio);
1517 spin_unlock_irqrestore(&pool->lock, flags);
1522 static struct endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1524 struct pool *pool = tc->pool;
1525 struct endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1528 h->shared_read_entry = NULL;
1529 h->all_io_entry = bio->bi_rw & REQ_DISCARD ? NULL : ds_inc(&pool->all_io_ds);
1530 h->overwrite_mapping = NULL;
1536 * Non-blocking function called from the thin target's map function.
1538 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1539 union map_info *map_context)
1542 struct thin_c *tc = ti->private;
1543 dm_block_t block = get_bio_block(tc, bio);
1544 struct dm_thin_device *td = tc->td;
1545 struct dm_thin_lookup_result result;
1547 map_context->ptr = thin_hook_bio(tc, bio);
1548 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1549 thin_defer_bio(tc, bio);
1550 return DM_MAPIO_SUBMITTED;
1553 r = dm_thin_find_block(td, block, 0, &result);
1556 * Note that we defer readahead too.
1560 if (unlikely(result.shared)) {
1562 * We have a race condition here between the
1563 * result.shared value returned by the lookup and
1564 * snapshot creation, which may cause new
1567 * To avoid this always quiesce the origin before
1568 * taking the snap. You want to do this anyway to
1569 * ensure a consistent application view
1572 * More distant ancestors are irrelevant. The
1573 * shared flag will be set in their case.
1575 thin_defer_bio(tc, bio);
1576 r = DM_MAPIO_SUBMITTED;
1578 remap(tc, bio, result.block);
1579 r = DM_MAPIO_REMAPPED;
1585 * In future, the failed dm_thin_find_block above could
1586 * provide the hint to load the metadata into cache.
1589 thin_defer_bio(tc, bio);
1590 r = DM_MAPIO_SUBMITTED;
1597 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1600 unsigned long flags;
1601 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1603 spin_lock_irqsave(&pt->pool->lock, flags);
1604 r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1605 spin_unlock_irqrestore(&pt->pool->lock, flags);
1608 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1609 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1615 static void __requeue_bios(struct pool *pool)
1617 bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1618 bio_list_init(&pool->retry_on_resume_list);
1621 /*----------------------------------------------------------------
1622 * Binding of control targets to a pool object
1623 *--------------------------------------------------------------*/
1624 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1626 struct pool_c *pt = ti->private;
1629 pool->low_water_blocks = pt->low_water_blocks;
1635 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1641 /*----------------------------------------------------------------
1643 *--------------------------------------------------------------*/
1644 /* Initialize pool features. */
1645 static void pool_features_init(struct pool_features *pf)
1647 pf->zero_new_blocks = 1;
1648 pf->discard_enabled = 1;
1649 pf->discard_passdown = 1;
1652 static void __pool_destroy(struct pool *pool)
1654 __pool_table_remove(pool);
1656 if (dm_pool_metadata_close(pool->pmd) < 0)
1657 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1659 prison_destroy(pool->prison);
1660 dm_kcopyd_client_destroy(pool->copier);
1663 destroy_workqueue(pool->wq);
1665 if (pool->next_mapping)
1666 mempool_free(pool->next_mapping, pool->mapping_pool);
1667 mempool_destroy(pool->mapping_pool);
1668 mempool_destroy(pool->endio_hook_pool);
1672 static struct pool *pool_create(struct mapped_device *pool_md,
1673 struct block_device *metadata_dev,
1674 unsigned long block_size, char **error)
1679 struct dm_pool_metadata *pmd;
1681 pmd = dm_pool_metadata_open(metadata_dev, block_size);
1683 *error = "Error creating metadata object";
1684 return (struct pool *)pmd;
1687 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1689 *error = "Error allocating memory for pool";
1690 err_p = ERR_PTR(-ENOMEM);
1695 pool->sectors_per_block = block_size;
1696 pool->block_shift = ffs(block_size) - 1;
1697 pool->offset_mask = block_size - 1;
1698 pool->low_water_blocks = 0;
1699 pool_features_init(&pool->pf);
1700 pool->prison = prison_create(PRISON_CELLS);
1701 if (!pool->prison) {
1702 *error = "Error creating pool's bio prison";
1703 err_p = ERR_PTR(-ENOMEM);
1707 pool->copier = dm_kcopyd_client_create();
1708 if (IS_ERR(pool->copier)) {
1709 r = PTR_ERR(pool->copier);
1710 *error = "Error creating pool's kcopyd client";
1712 goto bad_kcopyd_client;
1716 * Create singlethreaded workqueue that will service all devices
1717 * that use this metadata.
1719 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1721 *error = "Error creating pool's workqueue";
1722 err_p = ERR_PTR(-ENOMEM);
1726 INIT_WORK(&pool->worker, do_worker);
1727 INIT_DELAYED_WORK(&pool->waker, do_waker);
1728 spin_lock_init(&pool->lock);
1729 bio_list_init(&pool->deferred_bios);
1730 bio_list_init(&pool->deferred_flush_bios);
1731 INIT_LIST_HEAD(&pool->prepared_mappings);
1732 INIT_LIST_HEAD(&pool->prepared_discards);
1733 pool->low_water_triggered = 0;
1734 pool->no_free_space = 0;
1735 bio_list_init(&pool->retry_on_resume_list);
1736 ds_init(&pool->shared_read_ds);
1737 ds_init(&pool->all_io_ds);
1739 pool->next_mapping = NULL;
1740 pool->mapping_pool =
1741 mempool_create_kmalloc_pool(MAPPING_POOL_SIZE, sizeof(struct new_mapping));
1742 if (!pool->mapping_pool) {
1743 *error = "Error creating pool's mapping mempool";
1744 err_p = ERR_PTR(-ENOMEM);
1745 goto bad_mapping_pool;
1748 pool->endio_hook_pool =
1749 mempool_create_kmalloc_pool(ENDIO_HOOK_POOL_SIZE, sizeof(struct endio_hook));
1750 if (!pool->endio_hook_pool) {
1751 *error = "Error creating pool's endio_hook mempool";
1752 err_p = ERR_PTR(-ENOMEM);
1753 goto bad_endio_hook_pool;
1755 pool->ref_count = 1;
1756 pool->last_commit_jiffies = jiffies;
1757 pool->pool_md = pool_md;
1758 pool->md_dev = metadata_dev;
1759 __pool_table_insert(pool);
1763 bad_endio_hook_pool:
1764 mempool_destroy(pool->mapping_pool);
1766 destroy_workqueue(pool->wq);
1768 dm_kcopyd_client_destroy(pool->copier);
1770 prison_destroy(pool->prison);
1774 if (dm_pool_metadata_close(pmd))
1775 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1780 static void __pool_inc(struct pool *pool)
1782 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1786 static void __pool_dec(struct pool *pool)
1788 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1789 BUG_ON(!pool->ref_count);
1790 if (!--pool->ref_count)
1791 __pool_destroy(pool);
1794 static struct pool *__pool_find(struct mapped_device *pool_md,
1795 struct block_device *metadata_dev,
1796 unsigned long block_size, char **error,
1799 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1802 if (pool->pool_md != pool_md)
1803 return ERR_PTR(-EBUSY);
1807 pool = __pool_table_lookup(pool_md);
1809 if (pool->md_dev != metadata_dev)
1810 return ERR_PTR(-EINVAL);
1814 pool = pool_create(pool_md, metadata_dev, block_size, error);
1822 /*----------------------------------------------------------------
1823 * Pool target methods
1824 *--------------------------------------------------------------*/
1825 static void pool_dtr(struct dm_target *ti)
1827 struct pool_c *pt = ti->private;
1829 mutex_lock(&dm_thin_pool_table.mutex);
1831 unbind_control_target(pt->pool, ti);
1832 __pool_dec(pt->pool);
1833 dm_put_device(ti, pt->metadata_dev);
1834 dm_put_device(ti, pt->data_dev);
1837 mutex_unlock(&dm_thin_pool_table.mutex);
1840 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1841 struct dm_target *ti)
1845 const char *arg_name;
1847 static struct dm_arg _args[] = {
1848 {0, 3, "Invalid number of pool feature arguments"},
1852 * No feature arguments supplied.
1857 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1861 while (argc && !r) {
1862 arg_name = dm_shift_arg(as);
1865 if (!strcasecmp(arg_name, "skip_block_zeroing")) {
1866 pf->zero_new_blocks = 0;
1868 } else if (!strcasecmp(arg_name, "ignore_discard")) {
1869 pf->discard_enabled = 0;
1871 } else if (!strcasecmp(arg_name, "no_discard_passdown")) {
1872 pf->discard_passdown = 0;
1876 ti->error = "Unrecognised pool feature requested";
1884 * thin-pool <metadata dev> <data dev>
1885 * <data block size (sectors)>
1886 * <low water mark (blocks)>
1887 * [<#feature args> [<arg>]*]
1889 * Optional feature arguments are:
1890 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1891 * ignore_discard: disable discard
1892 * no_discard_passdown: don't pass discards down to the data device
1894 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1896 int r, pool_created = 0;
1899 struct pool_features pf;
1900 struct dm_arg_set as;
1901 struct dm_dev *data_dev;
1902 unsigned long block_size;
1903 dm_block_t low_water_blocks;
1904 struct dm_dev *metadata_dev;
1905 sector_t metadata_dev_size;
1906 char b[BDEVNAME_SIZE];
1909 * FIXME Remove validation from scope of lock.
1911 mutex_lock(&dm_thin_pool_table.mutex);
1914 ti->error = "Invalid argument count";
1921 r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1923 ti->error = "Error opening metadata block device";
1927 metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1928 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1929 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1930 bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1932 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1934 ti->error = "Error getting data device";
1938 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1939 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1940 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1941 !is_power_of_2(block_size)) {
1942 ti->error = "Invalid block size";
1947 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1948 ti->error = "Invalid low water mark";
1954 * Set default pool features.
1956 pool_features_init(&pf);
1958 dm_consume_args(&as, 4);
1959 r = parse_pool_features(&as, &pf, ti);
1963 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1969 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1970 block_size, &ti->error, &pool_created);
1977 * 'pool_created' reflects whether this is the first table load.
1978 * Top level discard support is not allowed to be changed after
1979 * initial load. This would require a pool reload to trigger thin
1982 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1983 ti->error = "Discard support cannot be disabled once enabled";
1985 goto out_flags_changed;
1989 * If discard_passdown was enabled verify that the data device
1990 * supports discards. Disable discard_passdown if not; otherwise
1991 * -EOPNOTSUPP will be returned.
1993 if (pf.discard_passdown) {
1994 struct request_queue *q = bdev_get_queue(data_dev->bdev);
1995 if (!q || !blk_queue_discard(q)) {
1996 DMWARN("Discard unsupported by data device: Disabling discard passdown.");
1997 pf.discard_passdown = 0;
2003 pt->metadata_dev = metadata_dev;
2004 pt->data_dev = data_dev;
2005 pt->low_water_blocks = low_water_blocks;
2007 ti->num_flush_requests = 1;
2009 * Only need to enable discards if the pool should pass
2010 * them down to the data device. The thin device's discard
2011 * processing will cause mappings to be removed from the btree.
2013 if (pf.discard_enabled && pf.discard_passdown) {
2014 ti->num_discard_requests = 1;
2016 * Setting 'discards_supported' circumvents the normal
2017 * stacking of discard limits (this keeps the pool and
2018 * thin devices' discard limits consistent).
2020 ti->discards_supported = 1;
2024 pt->callbacks.congested_fn = pool_is_congested;
2025 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2027 mutex_unlock(&dm_thin_pool_table.mutex);
2036 dm_put_device(ti, data_dev);
2038 dm_put_device(ti, metadata_dev);
2040 mutex_unlock(&dm_thin_pool_table.mutex);
2045 static int pool_map(struct dm_target *ti, struct bio *bio,
2046 union map_info *map_context)
2049 struct pool_c *pt = ti->private;
2050 struct pool *pool = pt->pool;
2051 unsigned long flags;
2054 * As this is a singleton target, ti->begin is always zero.
2056 spin_lock_irqsave(&pool->lock, flags);
2057 bio->bi_bdev = pt->data_dev->bdev;
2058 r = DM_MAPIO_REMAPPED;
2059 spin_unlock_irqrestore(&pool->lock, flags);
2065 * Retrieves the number of blocks of the data device from
2066 * the superblock and compares it to the actual device size,
2067 * thus resizing the data device in case it has grown.
2069 * This both copes with opening preallocated data devices in the ctr
2070 * being followed by a resume
2072 * calling the resume method individually after userspace has
2073 * grown the data device in reaction to a table event.
2075 static int pool_preresume(struct dm_target *ti)
2078 struct pool_c *pt = ti->private;
2079 struct pool *pool = pt->pool;
2080 dm_block_t data_size, sb_data_size;
2083 * Take control of the pool object.
2085 r = bind_control_target(pool, ti);
2089 data_size = ti->len >> pool->block_shift;
2090 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2092 DMERR("failed to retrieve data device size");
2096 if (data_size < sb_data_size) {
2097 DMERR("pool target too small, is %llu blocks (expected %llu)",
2098 data_size, sb_data_size);
2101 } else if (data_size > sb_data_size) {
2102 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2104 DMERR("failed to resize data device");
2108 r = dm_pool_commit_metadata(pool->pmd);
2110 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2119 static void pool_resume(struct dm_target *ti)
2121 struct pool_c *pt = ti->private;
2122 struct pool *pool = pt->pool;
2123 unsigned long flags;
2125 spin_lock_irqsave(&pool->lock, flags);
2126 pool->low_water_triggered = 0;
2127 pool->no_free_space = 0;
2128 __requeue_bios(pool);
2129 spin_unlock_irqrestore(&pool->lock, flags);
2131 do_waker(&pool->waker.work);
2134 static void pool_postsuspend(struct dm_target *ti)
2137 struct pool_c *pt = ti->private;
2138 struct pool *pool = pt->pool;
2140 cancel_delayed_work(&pool->waker);
2141 flush_workqueue(pool->wq);
2143 r = dm_pool_commit_metadata(pool->pmd);
2145 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2147 /* FIXME: invalidate device? error the next FUA or FLUSH bio ?*/
2151 static int check_arg_count(unsigned argc, unsigned args_required)
2153 if (argc != args_required) {
2154 DMWARN("Message received with %u arguments instead of %u.",
2155 argc, args_required);
2162 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2164 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2165 *dev_id <= MAX_DEV_ID)
2169 DMWARN("Message received with invalid device id: %s", arg);
2174 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2179 r = check_arg_count(argc, 2);
2183 r = read_dev_id(argv[1], &dev_id, 1);
2187 r = dm_pool_create_thin(pool->pmd, dev_id);
2189 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2197 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2200 dm_thin_id origin_dev_id;
2203 r = check_arg_count(argc, 3);
2207 r = read_dev_id(argv[1], &dev_id, 1);
2211 r = read_dev_id(argv[2], &origin_dev_id, 1);
2215 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2217 DMWARN("Creation of new snapshot %s of device %s failed.",
2225 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2230 r = check_arg_count(argc, 2);
2234 r = read_dev_id(argv[1], &dev_id, 1);
2238 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2240 DMWARN("Deletion of thin device %s failed.", argv[1]);
2245 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2247 dm_thin_id old_id, new_id;
2250 r = check_arg_count(argc, 3);
2254 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2255 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2259 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2260 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2264 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2266 DMWARN("Failed to change transaction id from %s to %s.",
2275 * Messages supported:
2276 * create_thin <dev_id>
2277 * create_snap <dev_id> <origin_id>
2279 * trim <dev_id> <new_size_in_sectors>
2280 * set_transaction_id <current_trans_id> <new_trans_id>
2282 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2285 struct pool_c *pt = ti->private;
2286 struct pool *pool = pt->pool;
2288 if (!strcasecmp(argv[0], "create_thin"))
2289 r = process_create_thin_mesg(argc, argv, pool);
2291 else if (!strcasecmp(argv[0], "create_snap"))
2292 r = process_create_snap_mesg(argc, argv, pool);
2294 else if (!strcasecmp(argv[0], "delete"))
2295 r = process_delete_mesg(argc, argv, pool);
2297 else if (!strcasecmp(argv[0], "set_transaction_id"))
2298 r = process_set_transaction_id_mesg(argc, argv, pool);
2301 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2304 r = dm_pool_commit_metadata(pool->pmd);
2306 DMERR("%s message: dm_pool_commit_metadata() failed, error = %d",
2315 * <transaction id> <used metadata sectors>/<total metadata sectors>
2316 * <used data sectors>/<total data sectors> <held metadata root>
2318 static int pool_status(struct dm_target *ti, status_type_t type,
2319 char *result, unsigned maxlen)
2323 uint64_t transaction_id;
2324 dm_block_t nr_free_blocks_data;
2325 dm_block_t nr_free_blocks_metadata;
2326 dm_block_t nr_blocks_data;
2327 dm_block_t nr_blocks_metadata;
2328 dm_block_t held_root;
2329 char buf[BDEVNAME_SIZE];
2330 char buf2[BDEVNAME_SIZE];
2331 struct pool_c *pt = ti->private;
2332 struct pool *pool = pt->pool;
2335 case STATUSTYPE_INFO:
2336 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2341 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2342 &nr_free_blocks_metadata);
2346 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2350 r = dm_pool_get_free_block_count(pool->pmd,
2351 &nr_free_blocks_data);
2355 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2359 r = dm_pool_get_held_metadata_root(pool->pmd, &held_root);
2363 DMEMIT("%llu %llu/%llu %llu/%llu ",
2364 (unsigned long long)transaction_id,
2365 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2366 (unsigned long long)nr_blocks_metadata,
2367 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2368 (unsigned long long)nr_blocks_data);
2371 DMEMIT("%llu", held_root);
2377 case STATUSTYPE_TABLE:
2378 DMEMIT("%s %s %lu %llu ",
2379 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2380 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2381 (unsigned long)pool->sectors_per_block,
2382 (unsigned long long)pt->low_water_blocks);
2384 count = !pool->pf.zero_new_blocks + !pool->pf.discard_enabled +
2385 !pool->pf.discard_passdown;
2386 DMEMIT("%u ", count);
2388 if (!pool->pf.zero_new_blocks)
2389 DMEMIT("skip_block_zeroing ");
2391 if (!pool->pf.discard_enabled)
2392 DMEMIT("ignore_discard ");
2394 if (!pool->pf.discard_passdown)
2395 DMEMIT("no_discard_passdown ");
2403 static int pool_iterate_devices(struct dm_target *ti,
2404 iterate_devices_callout_fn fn, void *data)
2406 struct pool_c *pt = ti->private;
2408 return fn(ti, pt->data_dev, 0, ti->len, data);
2411 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2412 struct bio_vec *biovec, int max_size)
2414 struct pool_c *pt = ti->private;
2415 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2417 if (!q->merge_bvec_fn)
2420 bvm->bi_bdev = pt->data_dev->bdev;
2422 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2425 static void set_discard_limits(struct pool *pool, struct queue_limits *limits)
2428 * FIXME: these limits may be incompatible with the pool's data device
2430 limits->max_discard_sectors = pool->sectors_per_block;
2433 * This is just a hint, and not enforced. We have to cope with
2434 * bios that overlap 2 blocks.
2436 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2437 limits->discard_zeroes_data = pool->pf.zero_new_blocks;
2440 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2442 struct pool_c *pt = ti->private;
2443 struct pool *pool = pt->pool;
2445 blk_limits_io_min(limits, 0);
2446 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2447 if (pool->pf.discard_enabled)
2448 set_discard_limits(pool, limits);
2451 static struct target_type pool_target = {
2452 .name = "thin-pool",
2453 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2454 DM_TARGET_IMMUTABLE,
2455 .version = {1, 1, 0},
2456 .module = THIS_MODULE,
2460 .postsuspend = pool_postsuspend,
2461 .preresume = pool_preresume,
2462 .resume = pool_resume,
2463 .message = pool_message,
2464 .status = pool_status,
2465 .merge = pool_merge,
2466 .iterate_devices = pool_iterate_devices,
2467 .io_hints = pool_io_hints,
2470 /*----------------------------------------------------------------
2471 * Thin target methods
2472 *--------------------------------------------------------------*/
2473 static void thin_dtr(struct dm_target *ti)
2475 struct thin_c *tc = ti->private;
2477 mutex_lock(&dm_thin_pool_table.mutex);
2479 __pool_dec(tc->pool);
2480 dm_pool_close_thin_device(tc->td);
2481 dm_put_device(ti, tc->pool_dev);
2483 dm_put_device(ti, tc->origin_dev);
2486 mutex_unlock(&dm_thin_pool_table.mutex);
2490 * Thin target parameters:
2492 * <pool_dev> <dev_id> [origin_dev]
2494 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2495 * dev_id: the internal device identifier
2496 * origin_dev: a device external to the pool that should act as the origin
2498 * If the pool device has discards disabled, they get disabled for the thin
2501 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2505 struct dm_dev *pool_dev, *origin_dev;
2506 struct mapped_device *pool_md;
2508 mutex_lock(&dm_thin_pool_table.mutex);
2510 if (argc != 2 && argc != 3) {
2511 ti->error = "Invalid argument count";
2516 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2518 ti->error = "Out of memory";
2524 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2526 ti->error = "Error opening origin device";
2527 goto bad_origin_dev;
2529 tc->origin_dev = origin_dev;
2532 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2534 ti->error = "Error opening pool device";
2537 tc->pool_dev = pool_dev;
2539 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2540 ti->error = "Invalid device id";
2545 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2547 ti->error = "Couldn't get pool mapped device";
2552 tc->pool = __pool_table_lookup(pool_md);
2554 ti->error = "Couldn't find pool object";
2556 goto bad_pool_lookup;
2558 __pool_inc(tc->pool);
2560 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2562 ti->error = "Couldn't open thin internal device";
2566 ti->split_io = tc->pool->sectors_per_block;
2567 ti->num_flush_requests = 1;
2569 /* In case the pool supports discards, pass them on. */
2570 if (tc->pool->pf.discard_enabled) {
2571 ti->discards_supported = 1;
2572 ti->num_discard_requests = 1;
2577 mutex_unlock(&dm_thin_pool_table.mutex);
2582 __pool_dec(tc->pool);
2586 dm_put_device(ti, tc->pool_dev);
2589 dm_put_device(ti, tc->origin_dev);
2593 mutex_unlock(&dm_thin_pool_table.mutex);
2598 static int thin_map(struct dm_target *ti, struct bio *bio,
2599 union map_info *map_context)
2601 bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2603 return thin_bio_map(ti, bio, map_context);
2606 static int thin_endio(struct dm_target *ti,
2607 struct bio *bio, int err,
2608 union map_info *map_context)
2610 unsigned long flags;
2611 struct endio_hook *h = map_context->ptr;
2612 struct list_head work;
2613 struct new_mapping *m, *tmp;
2614 struct pool *pool = h->tc->pool;
2616 if (h->shared_read_entry) {
2617 INIT_LIST_HEAD(&work);
2618 ds_dec(h->shared_read_entry, &work);
2620 spin_lock_irqsave(&pool->lock, flags);
2621 list_for_each_entry_safe(m, tmp, &work, list) {
2624 __maybe_add_mapping(m);
2626 spin_unlock_irqrestore(&pool->lock, flags);
2629 if (h->all_io_entry) {
2630 INIT_LIST_HEAD(&work);
2631 ds_dec(h->all_io_entry, &work);
2632 list_for_each_entry_safe(m, tmp, &work, list)
2633 list_add(&m->list, &pool->prepared_discards);
2636 mempool_free(h, pool->endio_hook_pool);
2641 static void thin_postsuspend(struct dm_target *ti)
2643 if (dm_noflush_suspending(ti))
2644 requeue_io((struct thin_c *)ti->private);
2648 * <nr mapped sectors> <highest mapped sector>
2650 static int thin_status(struct dm_target *ti, status_type_t type,
2651 char *result, unsigned maxlen)
2655 dm_block_t mapped, highest;
2656 char buf[BDEVNAME_SIZE];
2657 struct thin_c *tc = ti->private;
2663 case STATUSTYPE_INFO:
2664 r = dm_thin_get_mapped_count(tc->td, &mapped);
2668 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2672 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2674 DMEMIT("%llu", ((highest + 1) *
2675 tc->pool->sectors_per_block) - 1);
2680 case STATUSTYPE_TABLE:
2682 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2683 (unsigned long) tc->dev_id);
2685 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2693 static int thin_iterate_devices(struct dm_target *ti,
2694 iterate_devices_callout_fn fn, void *data)
2697 struct thin_c *tc = ti->private;
2700 * We can't call dm_pool_get_data_dev_size() since that blocks. So
2701 * we follow a more convoluted path through to the pool's target.
2704 return 0; /* nothing is bound */
2706 blocks = tc->pool->ti->len >> tc->pool->block_shift;
2708 return fn(ti, tc->pool_dev, 0, tc->pool->sectors_per_block * blocks, data);
2713 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2715 struct thin_c *tc = ti->private;
2716 struct pool *pool = tc->pool;
2718 blk_limits_io_min(limits, 0);
2719 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2720 set_discard_limits(pool, limits);
2723 static struct target_type thin_target = {
2725 .version = {1, 1, 0},
2726 .module = THIS_MODULE,
2730 .end_io = thin_endio,
2731 .postsuspend = thin_postsuspend,
2732 .status = thin_status,
2733 .iterate_devices = thin_iterate_devices,
2734 .io_hints = thin_io_hints,
2737 /*----------------------------------------------------------------*/
2739 static int __init dm_thin_init(void)
2745 r = dm_register_target(&thin_target);
2749 r = dm_register_target(&pool_target);
2751 dm_unregister_target(&thin_target);
2756 static void dm_thin_exit(void)
2758 dm_unregister_target(&thin_target);
2759 dm_unregister_target(&pool_target);
2762 module_init(dm_thin_init);
2763 module_exit(dm_thin_exit);
2765 MODULE_DESCRIPTION(DM_NAME "device-mapper thin provisioning target");
2766 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2767 MODULE_LICENSE("GPL");