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.
114 struct dm_bio_prison_cell {
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);
144 static struct kmem_cache *_cell_cache;
147 * @nr_cells should be the number of cells you want in use _concurrently_.
148 * Don't confuse it with the number of distinct keys.
150 static struct bio_prison *prison_create(unsigned nr_cells)
153 uint32_t nr_buckets = calc_nr_buckets(nr_cells);
154 size_t len = sizeof(struct bio_prison) +
155 (sizeof(struct hlist_head) * nr_buckets);
156 struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
161 spin_lock_init(&prison->lock);
162 prison->cell_pool = mempool_create_slab_pool(nr_cells, _cell_cache);
163 if (!prison->cell_pool) {
168 prison->nr_buckets = nr_buckets;
169 prison->hash_mask = nr_buckets - 1;
170 prison->cells = (struct hlist_head *) (prison + 1);
171 for (i = 0; i < nr_buckets; i++)
172 INIT_HLIST_HEAD(prison->cells + i);
177 static void prison_destroy(struct bio_prison *prison)
179 mempool_destroy(prison->cell_pool);
183 static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
185 const unsigned long BIG_PRIME = 4294967291UL;
186 uint64_t hash = key->block * BIG_PRIME;
188 return (uint32_t) (hash & prison->hash_mask);
191 static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
193 return (lhs->virtual == rhs->virtual) &&
194 (lhs->dev == rhs->dev) &&
195 (lhs->block == rhs->block);
198 static struct dm_bio_prison_cell *__search_bucket(struct hlist_head *bucket,
199 struct cell_key *key)
201 struct dm_bio_prison_cell *cell;
202 struct hlist_node *tmp;
204 hlist_for_each_entry(cell, tmp, bucket, list)
205 if (keys_equal(&cell->key, key))
212 * This may block if a new cell needs allocating. You must ensure that
213 * cells will be unlocked even if the calling thread is blocked.
215 * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
217 static int bio_detain(struct bio_prison *prison, struct cell_key *key,
218 struct bio *inmate, struct dm_bio_prison_cell **ref)
222 uint32_t hash = hash_key(prison, key);
223 struct dm_bio_prison_cell *cell, *cell2;
225 BUG_ON(hash > prison->nr_buckets);
227 spin_lock_irqsave(&prison->lock, flags);
229 cell = __search_bucket(prison->cells + hash, key);
231 bio_list_add(&cell->bios, inmate);
236 * Allocate a new cell
238 spin_unlock_irqrestore(&prison->lock, flags);
239 cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
240 spin_lock_irqsave(&prison->lock, flags);
243 * We've been unlocked, so we have to double check that
244 * nobody else has inserted this cell in the meantime.
246 cell = __search_bucket(prison->cells + hash, key);
248 mempool_free(cell2, prison->cell_pool);
249 bio_list_add(&cell->bios, inmate);
258 cell->prison = prison;
259 memcpy(&cell->key, key, sizeof(cell->key));
260 cell->holder = inmate;
261 bio_list_init(&cell->bios);
262 hlist_add_head(&cell->list, prison->cells + hash);
267 spin_unlock_irqrestore(&prison->lock, flags);
275 * @inmates must have been initialised prior to this call
277 static void __cell_release(struct dm_bio_prison_cell *cell, struct bio_list *inmates)
279 struct bio_prison *prison = cell->prison;
281 hlist_del(&cell->list);
284 bio_list_add(inmates, cell->holder);
285 bio_list_merge(inmates, &cell->bios);
288 mempool_free(cell, prison->cell_pool);
291 static void cell_release(struct dm_bio_prison_cell *cell, struct bio_list *bios)
294 struct bio_prison *prison = cell->prison;
296 spin_lock_irqsave(&prison->lock, flags);
297 __cell_release(cell, bios);
298 spin_unlock_irqrestore(&prison->lock, flags);
302 * There are a couple of places where we put a bio into a cell briefly
303 * before taking it out again. In these situations we know that no other
304 * bio may be in the cell. This function releases the cell, and also does
307 static void __cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
309 BUG_ON(cell->holder != bio);
310 BUG_ON(!bio_list_empty(&cell->bios));
312 __cell_release(cell, NULL);
315 static void cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
318 struct bio_prison *prison = cell->prison;
320 spin_lock_irqsave(&prison->lock, flags);
321 __cell_release_singleton(cell, bio);
322 spin_unlock_irqrestore(&prison->lock, flags);
326 * Sometimes we don't want the holder, just the additional bios.
328 static void __cell_release_no_holder(struct dm_bio_prison_cell *cell,
329 struct bio_list *inmates)
331 struct bio_prison *prison = cell->prison;
333 hlist_del(&cell->list);
334 bio_list_merge(inmates, &cell->bios);
336 mempool_free(cell, prison->cell_pool);
339 static void cell_release_no_holder(struct dm_bio_prison_cell *cell,
340 struct bio_list *inmates)
343 struct bio_prison *prison = cell->prison;
345 spin_lock_irqsave(&prison->lock, flags);
346 __cell_release_no_holder(cell, inmates);
347 spin_unlock_irqrestore(&prison->lock, flags);
350 static void cell_error(struct dm_bio_prison_cell *cell)
352 struct bio_prison *prison = cell->prison;
353 struct bio_list bios;
357 bio_list_init(&bios);
359 spin_lock_irqsave(&prison->lock, flags);
360 __cell_release(cell, &bios);
361 spin_unlock_irqrestore(&prison->lock, flags);
363 while ((bio = bio_list_pop(&bios)))
367 /*----------------------------------------------------------------*/
370 * We use the deferred set to keep track of pending reads to shared blocks.
371 * We do this to ensure the new mapping caused by a write isn't performed
372 * until these prior reads have completed. Otherwise the insertion of the
373 * new mapping could free the old block that the read bios are mapped to.
377 struct deferred_entry {
378 struct deferred_set *ds;
380 struct list_head work_items;
383 struct deferred_set {
385 unsigned current_entry;
387 struct deferred_entry entries[DEFERRED_SET_SIZE];
390 static void ds_init(struct deferred_set *ds)
394 spin_lock_init(&ds->lock);
395 ds->current_entry = 0;
397 for (i = 0; i < DEFERRED_SET_SIZE; i++) {
398 ds->entries[i].ds = ds;
399 ds->entries[i].count = 0;
400 INIT_LIST_HEAD(&ds->entries[i].work_items);
404 static struct deferred_entry *ds_inc(struct deferred_set *ds)
407 struct deferred_entry *entry;
409 spin_lock_irqsave(&ds->lock, flags);
410 entry = ds->entries + ds->current_entry;
412 spin_unlock_irqrestore(&ds->lock, flags);
417 static unsigned ds_next(unsigned index)
419 return (index + 1) % DEFERRED_SET_SIZE;
422 static void __sweep(struct deferred_set *ds, struct list_head *head)
424 while ((ds->sweeper != ds->current_entry) &&
425 !ds->entries[ds->sweeper].count) {
426 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
427 ds->sweeper = ds_next(ds->sweeper);
430 if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
431 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
434 static void ds_dec(struct deferred_entry *entry, struct list_head *head)
438 spin_lock_irqsave(&entry->ds->lock, flags);
439 BUG_ON(!entry->count);
441 __sweep(entry->ds, head);
442 spin_unlock_irqrestore(&entry->ds->lock, flags);
446 * Returns 1 if deferred or 0 if no pending items to delay job.
448 static int ds_add_work(struct deferred_set *ds, struct list_head *work)
454 spin_lock_irqsave(&ds->lock, flags);
455 if ((ds->sweeper == ds->current_entry) &&
456 !ds->entries[ds->current_entry].count)
459 list_add(work, &ds->entries[ds->current_entry].work_items);
460 next_entry = ds_next(ds->current_entry);
461 if (!ds->entries[next_entry].count)
462 ds->current_entry = next_entry;
464 spin_unlock_irqrestore(&ds->lock, flags);
469 /*----------------------------------------------------------------*/
474 static void build_data_key(struct dm_thin_device *td,
475 dm_block_t b, struct cell_key *key)
478 key->dev = dm_thin_dev_id(td);
482 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
483 struct cell_key *key)
486 key->dev = dm_thin_dev_id(td);
490 /*----------------------------------------------------------------*/
493 * A pool device ties together a metadata device and a data device. It
494 * also provides the interface for creating and destroying internal
497 struct dm_thin_new_mapping;
499 struct pool_features {
500 unsigned zero_new_blocks:1;
501 unsigned discard_enabled:1;
502 unsigned discard_passdown:1;
506 struct list_head list;
507 struct dm_target *ti; /* Only set if a pool target is bound */
509 struct mapped_device *pool_md;
510 struct block_device *md_dev;
511 struct dm_pool_metadata *pmd;
513 uint32_t sectors_per_block;
514 unsigned block_shift;
515 dm_block_t offset_mask;
516 dm_block_t low_water_blocks;
518 struct pool_features pf;
519 unsigned low_water_triggered:1; /* A dm event has been sent */
520 unsigned no_free_space:1; /* A -ENOSPC warning has been issued */
522 struct bio_prison *prison;
523 struct dm_kcopyd_client *copier;
525 struct workqueue_struct *wq;
526 struct work_struct worker;
527 struct delayed_work waker;
530 unsigned long last_commit_jiffies;
533 struct bio_list deferred_bios;
534 struct bio_list deferred_flush_bios;
535 struct list_head prepared_mappings;
536 struct list_head prepared_discards;
538 struct bio_list retry_on_resume_list;
540 struct deferred_set shared_read_ds;
541 struct deferred_set all_io_ds;
543 struct dm_thin_new_mapping *next_mapping;
544 mempool_t *mapping_pool;
545 mempool_t *endio_hook_pool;
549 * Target context for a pool.
552 struct dm_target *ti;
554 struct dm_dev *data_dev;
555 struct dm_dev *metadata_dev;
556 struct dm_target_callbacks callbacks;
558 dm_block_t low_water_blocks;
559 struct pool_features pf;
563 * Target context for a thin.
566 struct dm_dev *pool_dev;
567 struct dm_dev *origin_dev;
571 struct dm_thin_device *td;
574 /*----------------------------------------------------------------*/
577 * A global list of pools that uses a struct mapped_device as a key.
579 static struct dm_thin_pool_table {
581 struct list_head pools;
582 } dm_thin_pool_table;
584 static void pool_table_init(void)
586 mutex_init(&dm_thin_pool_table.mutex);
587 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
590 static void __pool_table_insert(struct pool *pool)
592 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
593 list_add(&pool->list, &dm_thin_pool_table.pools);
596 static void __pool_table_remove(struct pool *pool)
598 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
599 list_del(&pool->list);
602 static struct pool *__pool_table_lookup(struct mapped_device *md)
604 struct pool *pool = NULL, *tmp;
606 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
608 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
609 if (tmp->pool_md == md) {
618 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
620 struct pool *pool = NULL, *tmp;
622 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
624 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
625 if (tmp->md_dev == md_dev) {
634 /*----------------------------------------------------------------*/
636 struct dm_thin_endio_hook {
638 struct deferred_entry *shared_read_entry;
639 struct deferred_entry *all_io_entry;
640 struct dm_thin_new_mapping *overwrite_mapping;
643 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
646 struct bio_list bios;
648 bio_list_init(&bios);
649 bio_list_merge(&bios, master);
650 bio_list_init(master);
652 while ((bio = bio_list_pop(&bios))) {
653 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
656 bio_endio(bio, DM_ENDIO_REQUEUE);
658 bio_list_add(master, bio);
662 static void requeue_io(struct thin_c *tc)
664 struct pool *pool = tc->pool;
667 spin_lock_irqsave(&pool->lock, flags);
668 __requeue_bio_list(tc, &pool->deferred_bios);
669 __requeue_bio_list(tc, &pool->retry_on_resume_list);
670 spin_unlock_irqrestore(&pool->lock, flags);
674 * This section of code contains the logic for processing a thin device's IO.
675 * Much of the code depends on pool object resources (lists, workqueues, etc)
676 * but most is exclusively called from the thin target rather than the thin-pool
680 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
682 return bio->bi_sector >> tc->pool->block_shift;
685 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
687 struct pool *pool = tc->pool;
689 bio->bi_bdev = tc->pool_dev->bdev;
690 bio->bi_sector = (block << pool->block_shift) +
691 (bio->bi_sector & pool->offset_mask);
694 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
696 bio->bi_bdev = tc->origin_dev->bdev;
699 static void issue(struct thin_c *tc, struct bio *bio)
701 struct pool *pool = tc->pool;
705 * Batch together any FUA/FLUSH bios we find and then issue
706 * a single commit for them in process_deferred_bios().
708 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
709 spin_lock_irqsave(&pool->lock, flags);
710 bio_list_add(&pool->deferred_flush_bios, bio);
711 spin_unlock_irqrestore(&pool->lock, flags);
713 generic_make_request(bio);
716 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
718 remap_to_origin(tc, bio);
722 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
725 remap(tc, bio, block);
730 * wake_worker() is used when new work is queued and when pool_resume is
731 * ready to continue deferred IO processing.
733 static void wake_worker(struct pool *pool)
735 queue_work(pool->wq, &pool->worker);
738 /*----------------------------------------------------------------*/
741 * Bio endio functions.
743 struct dm_thin_new_mapping {
744 struct list_head list;
748 unsigned pass_discard:1;
751 dm_block_t virt_block;
752 dm_block_t data_block;
753 struct dm_bio_prison_cell *cell, *cell2;
757 * If the bio covers the whole area of a block then we can avoid
758 * zeroing or copying. Instead this bio is hooked. The bio will
759 * still be in the cell, so care has to be taken to avoid issuing
763 bio_end_io_t *saved_bi_end_io;
766 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
768 struct pool *pool = m->tc->pool;
770 if (m->quiesced && m->prepared) {
771 list_add(&m->list, &pool->prepared_mappings);
776 static void copy_complete(int read_err, unsigned long write_err, void *context)
779 struct dm_thin_new_mapping *m = context;
780 struct pool *pool = m->tc->pool;
782 m->err = read_err || write_err ? -EIO : 0;
784 spin_lock_irqsave(&pool->lock, flags);
786 __maybe_add_mapping(m);
787 spin_unlock_irqrestore(&pool->lock, flags);
790 static void overwrite_endio(struct bio *bio, int err)
793 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
794 struct dm_thin_new_mapping *m = h->overwrite_mapping;
795 struct pool *pool = m->tc->pool;
799 spin_lock_irqsave(&pool->lock, flags);
801 __maybe_add_mapping(m);
802 spin_unlock_irqrestore(&pool->lock, flags);
805 /*----------------------------------------------------------------*/
812 * Prepared mapping jobs.
816 * This sends the bios in the cell back to the deferred_bios list.
818 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell,
819 dm_block_t data_block)
821 struct pool *pool = tc->pool;
824 spin_lock_irqsave(&pool->lock, flags);
825 cell_release(cell, &pool->deferred_bios);
826 spin_unlock_irqrestore(&tc->pool->lock, flags);
832 * Same as cell_defer above, except it omits one particular detainee,
833 * a write bio that covers the block and has already been processed.
835 static void cell_defer_except(struct thin_c *tc, struct dm_bio_prison_cell *cell)
837 struct bio_list bios;
838 struct pool *pool = tc->pool;
841 bio_list_init(&bios);
843 spin_lock_irqsave(&pool->lock, flags);
844 cell_release_no_holder(cell, &pool->deferred_bios);
845 spin_unlock_irqrestore(&pool->lock, flags);
850 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
852 struct thin_c *tc = m->tc;
858 bio->bi_end_io = m->saved_bi_end_io;
866 * Commit the prepared block into the mapping btree.
867 * Any I/O for this block arriving after this point will get
868 * remapped to it directly.
870 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
872 DMERR("dm_thin_insert_block() failed");
878 * Release any bios held while the block was being provisioned.
879 * If we are processing a write bio that completely covers the block,
880 * we already processed it so can ignore it now when processing
881 * the bios in the cell.
884 cell_defer_except(tc, m->cell);
887 cell_defer(tc, m->cell, m->data_block);
890 mempool_free(m, tc->pool->mapping_pool);
893 static void process_prepared_discard(struct dm_thin_new_mapping *m)
896 struct thin_c *tc = m->tc;
898 r = dm_thin_remove_block(tc->td, m->virt_block);
900 DMERR("dm_thin_remove_block() failed");
903 * Pass the discard down to the underlying device?
906 remap_and_issue(tc, m->bio, m->data_block);
908 bio_endio(m->bio, 0);
910 cell_defer_except(tc, m->cell);
911 cell_defer_except(tc, m->cell2);
912 mempool_free(m, tc->pool->mapping_pool);
915 static void process_prepared(struct pool *pool, struct list_head *head,
916 void (*fn)(struct dm_thin_new_mapping *))
919 struct list_head maps;
920 struct dm_thin_new_mapping *m, *tmp;
922 INIT_LIST_HEAD(&maps);
923 spin_lock_irqsave(&pool->lock, flags);
924 list_splice_init(head, &maps);
925 spin_unlock_irqrestore(&pool->lock, flags);
927 list_for_each_entry_safe(m, tmp, &maps, list)
934 static int io_overlaps_block(struct pool *pool, struct bio *bio)
936 return !(bio->bi_sector & pool->offset_mask) &&
937 (bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
941 static int io_overwrites_block(struct pool *pool, struct bio *bio)
943 return (bio_data_dir(bio) == WRITE) &&
944 io_overlaps_block(pool, bio);
947 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
950 *save = bio->bi_end_io;
954 static int ensure_next_mapping(struct pool *pool)
956 if (pool->next_mapping)
959 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
961 return pool->next_mapping ? 0 : -ENOMEM;
964 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
966 struct dm_thin_new_mapping *r = pool->next_mapping;
968 BUG_ON(!pool->next_mapping);
970 pool->next_mapping = NULL;
975 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
976 struct dm_dev *origin, dm_block_t data_origin,
977 dm_block_t data_dest,
978 struct dm_bio_prison_cell *cell, struct bio *bio)
981 struct pool *pool = tc->pool;
982 struct dm_thin_new_mapping *m = get_next_mapping(pool);
984 INIT_LIST_HEAD(&m->list);
988 m->virt_block = virt_block;
989 m->data_block = data_dest;
994 if (!ds_add_work(&pool->shared_read_ds, &m->list))
998 * IO to pool_dev remaps to the pool target's data_dev.
1000 * If the whole block of data is being overwritten, we can issue the
1001 * bio immediately. Otherwise we use kcopyd to clone the data first.
1003 if (io_overwrites_block(pool, bio)) {
1004 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1006 h->overwrite_mapping = m;
1008 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1009 remap_and_issue(tc, bio, data_dest);
1011 struct dm_io_region from, to;
1013 from.bdev = origin->bdev;
1014 from.sector = data_origin * pool->sectors_per_block;
1015 from.count = pool->sectors_per_block;
1017 to.bdev = tc->pool_dev->bdev;
1018 to.sector = data_dest * pool->sectors_per_block;
1019 to.count = pool->sectors_per_block;
1021 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1022 0, copy_complete, m);
1024 mempool_free(m, pool->mapping_pool);
1025 DMERR("dm_kcopyd_copy() failed");
1031 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1032 dm_block_t data_origin, dm_block_t data_dest,
1033 struct dm_bio_prison_cell *cell, struct bio *bio)
1035 schedule_copy(tc, virt_block, tc->pool_dev,
1036 data_origin, data_dest, cell, bio);
1039 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1040 dm_block_t data_dest,
1041 struct dm_bio_prison_cell *cell, struct bio *bio)
1043 schedule_copy(tc, virt_block, tc->origin_dev,
1044 virt_block, data_dest, cell, bio);
1047 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1048 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1051 struct pool *pool = tc->pool;
1052 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1054 INIT_LIST_HEAD(&m->list);
1058 m->virt_block = virt_block;
1059 m->data_block = data_block;
1065 * If the whole block of data is being overwritten or we are not
1066 * zeroing pre-existing data, we can issue the bio immediately.
1067 * Otherwise we use kcopyd to zero the data first.
1069 if (!pool->pf.zero_new_blocks)
1070 process_prepared_mapping(m);
1072 else if (io_overwrites_block(pool, bio)) {
1073 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1075 h->overwrite_mapping = m;
1077 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1078 remap_and_issue(tc, bio, data_block);
1081 struct dm_io_region to;
1083 to.bdev = tc->pool_dev->bdev;
1084 to.sector = data_block * pool->sectors_per_block;
1085 to.count = pool->sectors_per_block;
1087 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
1089 mempool_free(m, pool->mapping_pool);
1090 DMERR("dm_kcopyd_zero() failed");
1096 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1099 dm_block_t free_blocks;
1100 unsigned long flags;
1101 struct pool *pool = tc->pool;
1103 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1107 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1108 DMWARN("%s: reached low water mark, sending event.",
1109 dm_device_name(pool->pool_md));
1110 spin_lock_irqsave(&pool->lock, flags);
1111 pool->low_water_triggered = 1;
1112 spin_unlock_irqrestore(&pool->lock, flags);
1113 dm_table_event(pool->ti->table);
1117 if (pool->no_free_space)
1121 * Try to commit to see if that will free up some
1124 r = dm_pool_commit_metadata(pool->pmd);
1126 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1131 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1136 * If we still have no space we set a flag to avoid
1137 * doing all this checking and return -ENOSPC.
1140 DMWARN("%s: no free space available.",
1141 dm_device_name(pool->pool_md));
1142 spin_lock_irqsave(&pool->lock, flags);
1143 pool->no_free_space = 1;
1144 spin_unlock_irqrestore(&pool->lock, flags);
1150 r = dm_pool_alloc_data_block(pool->pmd, result);
1158 * If we have run out of space, queue bios until the device is
1159 * resumed, presumably after having been reloaded with more space.
1161 static void retry_on_resume(struct bio *bio)
1163 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1164 struct thin_c *tc = h->tc;
1165 struct pool *pool = tc->pool;
1166 unsigned long flags;
1168 spin_lock_irqsave(&pool->lock, flags);
1169 bio_list_add(&pool->retry_on_resume_list, bio);
1170 spin_unlock_irqrestore(&pool->lock, flags);
1173 static void no_space(struct dm_bio_prison_cell *cell)
1176 struct bio_list bios;
1178 bio_list_init(&bios);
1179 cell_release(cell, &bios);
1181 while ((bio = bio_list_pop(&bios)))
1182 retry_on_resume(bio);
1185 static void process_discard(struct thin_c *tc, struct bio *bio)
1188 unsigned long flags;
1189 struct pool *pool = tc->pool;
1190 struct dm_bio_prison_cell *cell, *cell2;
1191 struct cell_key key, key2;
1192 dm_block_t block = get_bio_block(tc, bio);
1193 struct dm_thin_lookup_result lookup_result;
1194 struct dm_thin_new_mapping *m;
1196 build_virtual_key(tc->td, block, &key);
1197 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1200 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1204 * Check nobody is fiddling with this pool block. This can
1205 * happen if someone's in the process of breaking sharing
1208 build_data_key(tc->td, lookup_result.block, &key2);
1209 if (bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
1210 cell_release_singleton(cell, bio);
1214 if (io_overlaps_block(pool, bio)) {
1216 * IO may still be going to the destination block. We must
1217 * quiesce before we can do the removal.
1219 m = get_next_mapping(pool);
1221 m->pass_discard = (!lookup_result.shared) & pool->pf.discard_passdown;
1222 m->virt_block = block;
1223 m->data_block = lookup_result.block;
1229 if (!ds_add_work(&pool->all_io_ds, &m->list)) {
1230 spin_lock_irqsave(&pool->lock, flags);
1231 list_add(&m->list, &pool->prepared_discards);
1232 spin_unlock_irqrestore(&pool->lock, flags);
1237 * This path is hit if people are ignoring
1238 * limits->discard_granularity. It ignores any
1239 * part of the discard that is in a subsequent
1242 sector_t offset = bio->bi_sector - (block << pool->block_shift);
1243 unsigned remaining = (pool->sectors_per_block - offset) << 9;
1244 bio->bi_size = min(bio->bi_size, remaining);
1246 cell_release_singleton(cell, bio);
1247 cell_release_singleton(cell2, bio);
1248 remap_and_issue(tc, bio, lookup_result.block);
1254 * It isn't provisioned, just forget it.
1256 cell_release_singleton(cell, bio);
1261 DMERR("discard: find block unexpectedly returned %d", r);
1262 cell_release_singleton(cell, bio);
1268 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1269 struct cell_key *key,
1270 struct dm_thin_lookup_result *lookup_result,
1271 struct dm_bio_prison_cell *cell)
1274 dm_block_t data_block;
1276 r = alloc_data_block(tc, &data_block);
1279 schedule_internal_copy(tc, block, lookup_result->block,
1280 data_block, cell, bio);
1288 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1294 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1296 struct dm_thin_lookup_result *lookup_result)
1298 struct dm_bio_prison_cell *cell;
1299 struct pool *pool = tc->pool;
1300 struct cell_key key;
1303 * If cell is already occupied, then sharing is already in the process
1304 * of being broken so we have nothing further to do here.
1306 build_data_key(tc->td, lookup_result->block, &key);
1307 if (bio_detain(pool->prison, &key, bio, &cell))
1310 if (bio_data_dir(bio) == WRITE)
1311 break_sharing(tc, bio, block, &key, lookup_result, cell);
1313 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1315 h->shared_read_entry = ds_inc(&pool->shared_read_ds);
1317 cell_release_singleton(cell, bio);
1318 remap_and_issue(tc, bio, lookup_result->block);
1322 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1323 struct dm_bio_prison_cell *cell)
1326 dm_block_t data_block;
1329 * Remap empty bios (flushes) immediately, without provisioning.
1331 if (!bio->bi_size) {
1332 cell_release_singleton(cell, bio);
1333 remap_and_issue(tc, bio, 0);
1338 * Fill read bios with zeroes and complete them immediately.
1340 if (bio_data_dir(bio) == READ) {
1342 cell_release_singleton(cell, bio);
1347 r = alloc_data_block(tc, &data_block);
1351 schedule_external_copy(tc, block, data_block, cell, bio);
1353 schedule_zero(tc, block, data_block, cell, bio);
1361 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1367 static void process_bio(struct thin_c *tc, struct bio *bio)
1370 dm_block_t block = get_bio_block(tc, bio);
1371 struct dm_bio_prison_cell *cell;
1372 struct cell_key key;
1373 struct dm_thin_lookup_result lookup_result;
1376 * If cell is already occupied, then the block is already
1377 * being provisioned so we have nothing further to do here.
1379 build_virtual_key(tc->td, block, &key);
1380 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1383 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1387 * We can release this cell now. This thread is the only
1388 * one that puts bios into a cell, and we know there were
1389 * no preceding bios.
1392 * TODO: this will probably have to change when discard goes
1395 cell_release_singleton(cell, bio);
1397 if (lookup_result.shared)
1398 process_shared_bio(tc, bio, block, &lookup_result);
1400 remap_and_issue(tc, bio, lookup_result.block);
1404 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1405 cell_release_singleton(cell, bio);
1406 remap_to_origin_and_issue(tc, bio);
1408 provision_block(tc, bio, block, cell);
1412 DMERR("dm_thin_find_block() failed, error = %d", r);
1413 cell_release_singleton(cell, bio);
1419 static int need_commit_due_to_time(struct pool *pool)
1421 return jiffies < pool->last_commit_jiffies ||
1422 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1425 static void process_deferred_bios(struct pool *pool)
1427 unsigned long flags;
1429 struct bio_list bios;
1432 bio_list_init(&bios);
1434 spin_lock_irqsave(&pool->lock, flags);
1435 bio_list_merge(&bios, &pool->deferred_bios);
1436 bio_list_init(&pool->deferred_bios);
1437 spin_unlock_irqrestore(&pool->lock, flags);
1439 while ((bio = bio_list_pop(&bios))) {
1440 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1441 struct thin_c *tc = h->tc;
1444 * If we've got no free new_mapping structs, and processing
1445 * this bio might require one, we pause until there are some
1446 * prepared mappings to process.
1448 if (ensure_next_mapping(pool)) {
1449 spin_lock_irqsave(&pool->lock, flags);
1450 bio_list_merge(&pool->deferred_bios, &bios);
1451 spin_unlock_irqrestore(&pool->lock, flags);
1456 if (bio->bi_rw & REQ_DISCARD)
1457 process_discard(tc, bio);
1459 process_bio(tc, bio);
1463 * If there are any deferred flush bios, we must commit
1464 * the metadata before issuing them.
1466 bio_list_init(&bios);
1467 spin_lock_irqsave(&pool->lock, flags);
1468 bio_list_merge(&bios, &pool->deferred_flush_bios);
1469 bio_list_init(&pool->deferred_flush_bios);
1470 spin_unlock_irqrestore(&pool->lock, flags);
1472 if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1475 r = dm_pool_commit_metadata(pool->pmd);
1477 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1479 while ((bio = bio_list_pop(&bios)))
1483 pool->last_commit_jiffies = jiffies;
1485 while ((bio = bio_list_pop(&bios)))
1486 generic_make_request(bio);
1489 static void do_worker(struct work_struct *ws)
1491 struct pool *pool = container_of(ws, struct pool, worker);
1493 process_prepared(pool, &pool->prepared_mappings, process_prepared_mapping);
1494 process_prepared(pool, &pool->prepared_discards, process_prepared_discard);
1495 process_deferred_bios(pool);
1499 * We want to commit periodically so that not too much
1500 * unwritten data builds up.
1502 static void do_waker(struct work_struct *ws)
1504 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1506 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1509 /*----------------------------------------------------------------*/
1512 * Mapping functions.
1516 * Called only while mapping a thin bio to hand it over to the workqueue.
1518 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1520 unsigned long flags;
1521 struct pool *pool = tc->pool;
1523 spin_lock_irqsave(&pool->lock, flags);
1524 bio_list_add(&pool->deferred_bios, bio);
1525 spin_unlock_irqrestore(&pool->lock, flags);
1530 static struct dm_thin_endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1532 struct pool *pool = tc->pool;
1533 struct dm_thin_endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1536 h->shared_read_entry = NULL;
1537 h->all_io_entry = bio->bi_rw & REQ_DISCARD ? NULL : ds_inc(&pool->all_io_ds);
1538 h->overwrite_mapping = NULL;
1544 * Non-blocking function called from the thin target's map function.
1546 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1547 union map_info *map_context)
1550 struct thin_c *tc = ti->private;
1551 dm_block_t block = get_bio_block(tc, bio);
1552 struct dm_thin_device *td = tc->td;
1553 struct dm_thin_lookup_result result;
1555 map_context->ptr = thin_hook_bio(tc, bio);
1556 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1557 thin_defer_bio(tc, bio);
1558 return DM_MAPIO_SUBMITTED;
1561 r = dm_thin_find_block(td, block, 0, &result);
1564 * Note that we defer readahead too.
1568 if (unlikely(result.shared)) {
1570 * We have a race condition here between the
1571 * result.shared value returned by the lookup and
1572 * snapshot creation, which may cause new
1575 * To avoid this always quiesce the origin before
1576 * taking the snap. You want to do this anyway to
1577 * ensure a consistent application view
1580 * More distant ancestors are irrelevant. The
1581 * shared flag will be set in their case.
1583 thin_defer_bio(tc, bio);
1584 r = DM_MAPIO_SUBMITTED;
1586 remap(tc, bio, result.block);
1587 r = DM_MAPIO_REMAPPED;
1593 * In future, the failed dm_thin_find_block above could
1594 * provide the hint to load the metadata into cache.
1597 thin_defer_bio(tc, bio);
1598 r = DM_MAPIO_SUBMITTED;
1605 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1608 unsigned long flags;
1609 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1611 spin_lock_irqsave(&pt->pool->lock, flags);
1612 r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1613 spin_unlock_irqrestore(&pt->pool->lock, flags);
1616 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1617 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1623 static void __requeue_bios(struct pool *pool)
1625 bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1626 bio_list_init(&pool->retry_on_resume_list);
1629 /*----------------------------------------------------------------
1630 * Binding of control targets to a pool object
1631 *--------------------------------------------------------------*/
1632 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1634 struct pool_c *pt = ti->private;
1637 pool->low_water_blocks = pt->low_water_blocks;
1641 * If discard_passdown was enabled verify that the data device
1642 * supports discards. Disable discard_passdown if not; otherwise
1643 * -EOPNOTSUPP will be returned.
1645 if (pt->pf.discard_passdown) {
1646 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1647 if (!q || !blk_queue_discard(q)) {
1648 char buf[BDEVNAME_SIZE];
1649 DMWARN("Discard unsupported by data device (%s): Disabling discard passdown.",
1650 bdevname(pt->data_dev->bdev, buf));
1651 pool->pf.discard_passdown = 0;
1658 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1664 /*----------------------------------------------------------------
1666 *--------------------------------------------------------------*/
1667 /* Initialize pool features. */
1668 static void pool_features_init(struct pool_features *pf)
1670 pf->zero_new_blocks = 1;
1671 pf->discard_enabled = 1;
1672 pf->discard_passdown = 1;
1675 static void __pool_destroy(struct pool *pool)
1677 __pool_table_remove(pool);
1679 if (dm_pool_metadata_close(pool->pmd) < 0)
1680 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1682 prison_destroy(pool->prison);
1683 dm_kcopyd_client_destroy(pool->copier);
1686 destroy_workqueue(pool->wq);
1688 if (pool->next_mapping)
1689 mempool_free(pool->next_mapping, pool->mapping_pool);
1690 mempool_destroy(pool->mapping_pool);
1691 mempool_destroy(pool->endio_hook_pool);
1695 static struct kmem_cache *_new_mapping_cache;
1696 static struct kmem_cache *_endio_hook_cache;
1698 static struct pool *pool_create(struct mapped_device *pool_md,
1699 struct block_device *metadata_dev,
1700 unsigned long block_size, char **error)
1705 struct dm_pool_metadata *pmd;
1707 pmd = dm_pool_metadata_open(metadata_dev, block_size);
1709 *error = "Error creating metadata object";
1710 return (struct pool *)pmd;
1713 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1715 *error = "Error allocating memory for pool";
1716 err_p = ERR_PTR(-ENOMEM);
1721 pool->sectors_per_block = block_size;
1722 pool->block_shift = ffs(block_size) - 1;
1723 pool->offset_mask = block_size - 1;
1724 pool->low_water_blocks = 0;
1725 pool_features_init(&pool->pf);
1726 pool->prison = prison_create(PRISON_CELLS);
1727 if (!pool->prison) {
1728 *error = "Error creating pool's bio prison";
1729 err_p = ERR_PTR(-ENOMEM);
1733 pool->copier = dm_kcopyd_client_create();
1734 if (IS_ERR(pool->copier)) {
1735 r = PTR_ERR(pool->copier);
1736 *error = "Error creating pool's kcopyd client";
1738 goto bad_kcopyd_client;
1742 * Create singlethreaded workqueue that will service all devices
1743 * that use this metadata.
1745 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1747 *error = "Error creating pool's workqueue";
1748 err_p = ERR_PTR(-ENOMEM);
1752 INIT_WORK(&pool->worker, do_worker);
1753 INIT_DELAYED_WORK(&pool->waker, do_waker);
1754 spin_lock_init(&pool->lock);
1755 bio_list_init(&pool->deferred_bios);
1756 bio_list_init(&pool->deferred_flush_bios);
1757 INIT_LIST_HEAD(&pool->prepared_mappings);
1758 INIT_LIST_HEAD(&pool->prepared_discards);
1759 pool->low_water_triggered = 0;
1760 pool->no_free_space = 0;
1761 bio_list_init(&pool->retry_on_resume_list);
1762 ds_init(&pool->shared_read_ds);
1763 ds_init(&pool->all_io_ds);
1765 pool->next_mapping = NULL;
1766 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1767 _new_mapping_cache);
1768 if (!pool->mapping_pool) {
1769 *error = "Error creating pool's mapping mempool";
1770 err_p = ERR_PTR(-ENOMEM);
1771 goto bad_mapping_pool;
1774 pool->endio_hook_pool = mempool_create_slab_pool(ENDIO_HOOK_POOL_SIZE,
1776 if (!pool->endio_hook_pool) {
1777 *error = "Error creating pool's endio_hook mempool";
1778 err_p = ERR_PTR(-ENOMEM);
1779 goto bad_endio_hook_pool;
1781 pool->ref_count = 1;
1782 pool->last_commit_jiffies = jiffies;
1783 pool->pool_md = pool_md;
1784 pool->md_dev = metadata_dev;
1785 __pool_table_insert(pool);
1789 bad_endio_hook_pool:
1790 mempool_destroy(pool->mapping_pool);
1792 destroy_workqueue(pool->wq);
1794 dm_kcopyd_client_destroy(pool->copier);
1796 prison_destroy(pool->prison);
1800 if (dm_pool_metadata_close(pmd))
1801 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1806 static void __pool_inc(struct pool *pool)
1808 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1812 static void __pool_dec(struct pool *pool)
1814 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1815 BUG_ON(!pool->ref_count);
1816 if (!--pool->ref_count)
1817 __pool_destroy(pool);
1820 static struct pool *__pool_find(struct mapped_device *pool_md,
1821 struct block_device *metadata_dev,
1822 unsigned long block_size, char **error,
1825 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1828 if (pool->pool_md != pool_md)
1829 return ERR_PTR(-EBUSY);
1833 pool = __pool_table_lookup(pool_md);
1835 if (pool->md_dev != metadata_dev)
1836 return ERR_PTR(-EINVAL);
1840 pool = pool_create(pool_md, metadata_dev, block_size, error);
1848 /*----------------------------------------------------------------
1849 * Pool target methods
1850 *--------------------------------------------------------------*/
1851 static void pool_dtr(struct dm_target *ti)
1853 struct pool_c *pt = ti->private;
1855 mutex_lock(&dm_thin_pool_table.mutex);
1857 unbind_control_target(pt->pool, ti);
1858 __pool_dec(pt->pool);
1859 dm_put_device(ti, pt->metadata_dev);
1860 dm_put_device(ti, pt->data_dev);
1863 mutex_unlock(&dm_thin_pool_table.mutex);
1866 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1867 struct dm_target *ti)
1871 const char *arg_name;
1873 static struct dm_arg _args[] = {
1874 {0, 3, "Invalid number of pool feature arguments"},
1878 * No feature arguments supplied.
1883 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1887 while (argc && !r) {
1888 arg_name = dm_shift_arg(as);
1891 if (!strcasecmp(arg_name, "skip_block_zeroing")) {
1892 pf->zero_new_blocks = 0;
1894 } else if (!strcasecmp(arg_name, "ignore_discard")) {
1895 pf->discard_enabled = 0;
1897 } else if (!strcasecmp(arg_name, "no_discard_passdown")) {
1898 pf->discard_passdown = 0;
1902 ti->error = "Unrecognised pool feature requested";
1910 * thin-pool <metadata dev> <data dev>
1911 * <data block size (sectors)>
1912 * <low water mark (blocks)>
1913 * [<#feature args> [<arg>]*]
1915 * Optional feature arguments are:
1916 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1917 * ignore_discard: disable discard
1918 * no_discard_passdown: don't pass discards down to the data device
1920 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1922 int r, pool_created = 0;
1925 struct pool_features pf;
1926 struct dm_arg_set as;
1927 struct dm_dev *data_dev;
1928 unsigned long block_size;
1929 dm_block_t low_water_blocks;
1930 struct dm_dev *metadata_dev;
1931 sector_t metadata_dev_size;
1932 char b[BDEVNAME_SIZE];
1935 * FIXME Remove validation from scope of lock.
1937 mutex_lock(&dm_thin_pool_table.mutex);
1940 ti->error = "Invalid argument count";
1947 r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1949 ti->error = "Error opening metadata block device";
1953 metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1954 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1955 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1956 bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1958 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1960 ti->error = "Error getting data device";
1964 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1965 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1966 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1967 !is_power_of_2(block_size)) {
1968 ti->error = "Invalid block size";
1973 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1974 ti->error = "Invalid low water mark";
1980 * Set default pool features.
1982 pool_features_init(&pf);
1984 dm_consume_args(&as, 4);
1985 r = parse_pool_features(&as, &pf, ti);
1989 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1995 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1996 block_size, &ti->error, &pool_created);
2003 * 'pool_created' reflects whether this is the first table load.
2004 * Top level discard support is not allowed to be changed after
2005 * initial load. This would require a pool reload to trigger thin
2008 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2009 ti->error = "Discard support cannot be disabled once enabled";
2011 goto out_flags_changed;
2016 pt->metadata_dev = metadata_dev;
2017 pt->data_dev = data_dev;
2018 pt->low_water_blocks = low_water_blocks;
2020 ti->num_flush_requests = 1;
2022 * Only need to enable discards if the pool should pass
2023 * them down to the data device. The thin device's discard
2024 * processing will cause mappings to be removed from the btree.
2026 if (pf.discard_enabled && pf.discard_passdown) {
2027 ti->num_discard_requests = 1;
2029 * Setting 'discards_supported' circumvents the normal
2030 * stacking of discard limits (this keeps the pool and
2031 * thin devices' discard limits consistent).
2033 ti->discards_supported = 1;
2037 pt->callbacks.congested_fn = pool_is_congested;
2038 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2040 mutex_unlock(&dm_thin_pool_table.mutex);
2049 dm_put_device(ti, data_dev);
2051 dm_put_device(ti, metadata_dev);
2053 mutex_unlock(&dm_thin_pool_table.mutex);
2058 static int pool_map(struct dm_target *ti, struct bio *bio,
2059 union map_info *map_context)
2062 struct pool_c *pt = ti->private;
2063 struct pool *pool = pt->pool;
2064 unsigned long flags;
2067 * As this is a singleton target, ti->begin is always zero.
2069 spin_lock_irqsave(&pool->lock, flags);
2070 bio->bi_bdev = pt->data_dev->bdev;
2071 r = DM_MAPIO_REMAPPED;
2072 spin_unlock_irqrestore(&pool->lock, flags);
2078 * Retrieves the number of blocks of the data device from
2079 * the superblock and compares it to the actual device size,
2080 * thus resizing the data device in case it has grown.
2082 * This both copes with opening preallocated data devices in the ctr
2083 * being followed by a resume
2085 * calling the resume method individually after userspace has
2086 * grown the data device in reaction to a table event.
2088 static int pool_preresume(struct dm_target *ti)
2091 struct pool_c *pt = ti->private;
2092 struct pool *pool = pt->pool;
2093 dm_block_t data_size, sb_data_size;
2096 * Take control of the pool object.
2098 r = bind_control_target(pool, ti);
2102 data_size = ti->len >> pool->block_shift;
2103 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2105 DMERR("failed to retrieve data device size");
2109 if (data_size < sb_data_size) {
2110 DMERR("pool target too small, is %llu blocks (expected %llu)",
2111 data_size, sb_data_size);
2114 } else if (data_size > sb_data_size) {
2115 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2117 DMERR("failed to resize data device");
2121 r = dm_pool_commit_metadata(pool->pmd);
2123 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2132 static void pool_resume(struct dm_target *ti)
2134 struct pool_c *pt = ti->private;
2135 struct pool *pool = pt->pool;
2136 unsigned long flags;
2138 spin_lock_irqsave(&pool->lock, flags);
2139 pool->low_water_triggered = 0;
2140 pool->no_free_space = 0;
2141 __requeue_bios(pool);
2142 spin_unlock_irqrestore(&pool->lock, flags);
2144 do_waker(&pool->waker.work);
2147 static void pool_postsuspend(struct dm_target *ti)
2150 struct pool_c *pt = ti->private;
2151 struct pool *pool = pt->pool;
2153 cancel_delayed_work(&pool->waker);
2154 flush_workqueue(pool->wq);
2156 r = dm_pool_commit_metadata(pool->pmd);
2158 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2160 /* FIXME: invalidate device? error the next FUA or FLUSH bio ?*/
2164 static int check_arg_count(unsigned argc, unsigned args_required)
2166 if (argc != args_required) {
2167 DMWARN("Message received with %u arguments instead of %u.",
2168 argc, args_required);
2175 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2177 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2178 *dev_id <= MAX_DEV_ID)
2182 DMWARN("Message received with invalid device id: %s", arg);
2187 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2192 r = check_arg_count(argc, 2);
2196 r = read_dev_id(argv[1], &dev_id, 1);
2200 r = dm_pool_create_thin(pool->pmd, dev_id);
2202 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2210 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2213 dm_thin_id origin_dev_id;
2216 r = check_arg_count(argc, 3);
2220 r = read_dev_id(argv[1], &dev_id, 1);
2224 r = read_dev_id(argv[2], &origin_dev_id, 1);
2228 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2230 DMWARN("Creation of new snapshot %s of device %s failed.",
2238 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2243 r = check_arg_count(argc, 2);
2247 r = read_dev_id(argv[1], &dev_id, 1);
2251 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2253 DMWARN("Deletion of thin device %s failed.", argv[1]);
2258 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2260 dm_thin_id old_id, new_id;
2263 r = check_arg_count(argc, 3);
2267 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2268 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2272 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2273 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2277 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2279 DMWARN("Failed to change transaction id from %s to %s.",
2287 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2291 r = check_arg_count(argc, 1);
2295 r = dm_pool_reserve_metadata_snap(pool->pmd);
2297 DMWARN("reserve_metadata_snap message failed.");
2302 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2306 r = check_arg_count(argc, 1);
2310 r = dm_pool_release_metadata_snap(pool->pmd);
2312 DMWARN("release_metadata_snap message failed.");
2318 * Messages supported:
2319 * create_thin <dev_id>
2320 * create_snap <dev_id> <origin_id>
2322 * trim <dev_id> <new_size_in_sectors>
2323 * set_transaction_id <current_trans_id> <new_trans_id>
2324 * reserve_metadata_snap
2325 * release_metadata_snap
2327 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2330 struct pool_c *pt = ti->private;
2331 struct pool *pool = pt->pool;
2333 if (!strcasecmp(argv[0], "create_thin"))
2334 r = process_create_thin_mesg(argc, argv, pool);
2336 else if (!strcasecmp(argv[0], "create_snap"))
2337 r = process_create_snap_mesg(argc, argv, pool);
2339 else if (!strcasecmp(argv[0], "delete"))
2340 r = process_delete_mesg(argc, argv, pool);
2342 else if (!strcasecmp(argv[0], "set_transaction_id"))
2343 r = process_set_transaction_id_mesg(argc, argv, pool);
2345 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2346 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2348 else if (!strcasecmp(argv[0], "release_metadata_snap"))
2349 r = process_release_metadata_snap_mesg(argc, argv, pool);
2352 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2355 r = dm_pool_commit_metadata(pool->pmd);
2357 DMERR("%s message: dm_pool_commit_metadata() failed, error = %d",
2366 * <transaction id> <used metadata sectors>/<total metadata sectors>
2367 * <used data sectors>/<total data sectors> <held metadata root>
2369 static int pool_status(struct dm_target *ti, status_type_t type,
2370 char *result, unsigned maxlen)
2374 uint64_t transaction_id;
2375 dm_block_t nr_free_blocks_data;
2376 dm_block_t nr_free_blocks_metadata;
2377 dm_block_t nr_blocks_data;
2378 dm_block_t nr_blocks_metadata;
2379 dm_block_t held_root;
2380 char buf[BDEVNAME_SIZE];
2381 char buf2[BDEVNAME_SIZE];
2382 struct pool_c *pt = ti->private;
2383 struct pool *pool = pt->pool;
2386 case STATUSTYPE_INFO:
2387 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2392 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2393 &nr_free_blocks_metadata);
2397 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2401 r = dm_pool_get_free_block_count(pool->pmd,
2402 &nr_free_blocks_data);
2406 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2410 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2414 DMEMIT("%llu %llu/%llu %llu/%llu ",
2415 (unsigned long long)transaction_id,
2416 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2417 (unsigned long long)nr_blocks_metadata,
2418 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2419 (unsigned long long)nr_blocks_data);
2422 DMEMIT("%llu", held_root);
2428 case STATUSTYPE_TABLE:
2429 DMEMIT("%s %s %lu %llu ",
2430 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2431 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2432 (unsigned long)pool->sectors_per_block,
2433 (unsigned long long)pt->low_water_blocks);
2435 count = !pool->pf.zero_new_blocks + !pool->pf.discard_enabled +
2436 !pt->pf.discard_passdown;
2437 DMEMIT("%u ", count);
2439 if (!pool->pf.zero_new_blocks)
2440 DMEMIT("skip_block_zeroing ");
2442 if (!pool->pf.discard_enabled)
2443 DMEMIT("ignore_discard ");
2445 if (!pt->pf.discard_passdown)
2446 DMEMIT("no_discard_passdown ");
2454 static int pool_iterate_devices(struct dm_target *ti,
2455 iterate_devices_callout_fn fn, void *data)
2457 struct pool_c *pt = ti->private;
2459 return fn(ti, pt->data_dev, 0, ti->len, data);
2462 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2463 struct bio_vec *biovec, int max_size)
2465 struct pool_c *pt = ti->private;
2466 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2468 if (!q->merge_bvec_fn)
2471 bvm->bi_bdev = pt->data_dev->bdev;
2473 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2476 static void set_discard_limits(struct pool *pool, struct queue_limits *limits)
2479 * FIXME: these limits may be incompatible with the pool's data device
2481 limits->max_discard_sectors = pool->sectors_per_block;
2484 * This is just a hint, and not enforced. We have to cope with
2485 * bios that overlap 2 blocks.
2487 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2488 limits->discard_zeroes_data = pool->pf.zero_new_blocks;
2491 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2493 struct pool_c *pt = ti->private;
2494 struct pool *pool = pt->pool;
2496 blk_limits_io_min(limits, 0);
2497 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2498 if (pool->pf.discard_enabled)
2499 set_discard_limits(pool, limits);
2502 static struct target_type pool_target = {
2503 .name = "thin-pool",
2504 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2505 DM_TARGET_IMMUTABLE,
2506 .version = {1, 2, 0},
2507 .module = THIS_MODULE,
2511 .postsuspend = pool_postsuspend,
2512 .preresume = pool_preresume,
2513 .resume = pool_resume,
2514 .message = pool_message,
2515 .status = pool_status,
2516 .merge = pool_merge,
2517 .iterate_devices = pool_iterate_devices,
2518 .io_hints = pool_io_hints,
2521 /*----------------------------------------------------------------
2522 * Thin target methods
2523 *--------------------------------------------------------------*/
2524 static void thin_dtr(struct dm_target *ti)
2526 struct thin_c *tc = ti->private;
2528 mutex_lock(&dm_thin_pool_table.mutex);
2530 __pool_dec(tc->pool);
2531 dm_pool_close_thin_device(tc->td);
2532 dm_put_device(ti, tc->pool_dev);
2534 dm_put_device(ti, tc->origin_dev);
2537 mutex_unlock(&dm_thin_pool_table.mutex);
2541 * Thin target parameters:
2543 * <pool_dev> <dev_id> [origin_dev]
2545 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2546 * dev_id: the internal device identifier
2547 * origin_dev: a device external to the pool that should act as the origin
2549 * If the pool device has discards disabled, they get disabled for the thin
2552 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2556 struct dm_dev *pool_dev, *origin_dev;
2557 struct mapped_device *pool_md;
2559 mutex_lock(&dm_thin_pool_table.mutex);
2561 if (argc != 2 && argc != 3) {
2562 ti->error = "Invalid argument count";
2567 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2569 ti->error = "Out of memory";
2575 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2577 ti->error = "Error opening origin device";
2578 goto bad_origin_dev;
2580 tc->origin_dev = origin_dev;
2583 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2585 ti->error = "Error opening pool device";
2588 tc->pool_dev = pool_dev;
2590 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2591 ti->error = "Invalid device id";
2596 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2598 ti->error = "Couldn't get pool mapped device";
2603 tc->pool = __pool_table_lookup(pool_md);
2605 ti->error = "Couldn't find pool object";
2607 goto bad_pool_lookup;
2609 __pool_inc(tc->pool);
2611 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2613 ti->error = "Couldn't open thin internal device";
2617 ti->split_io = tc->pool->sectors_per_block;
2618 ti->num_flush_requests = 1;
2620 /* In case the pool supports discards, pass them on. */
2621 if (tc->pool->pf.discard_enabled) {
2622 ti->discards_supported = 1;
2623 ti->num_discard_requests = 1;
2628 mutex_unlock(&dm_thin_pool_table.mutex);
2633 __pool_dec(tc->pool);
2637 dm_put_device(ti, tc->pool_dev);
2640 dm_put_device(ti, tc->origin_dev);
2644 mutex_unlock(&dm_thin_pool_table.mutex);
2649 static int thin_map(struct dm_target *ti, struct bio *bio,
2650 union map_info *map_context)
2652 bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2654 return thin_bio_map(ti, bio, map_context);
2657 static int thin_endio(struct dm_target *ti,
2658 struct bio *bio, int err,
2659 union map_info *map_context)
2661 unsigned long flags;
2662 struct dm_thin_endio_hook *h = map_context->ptr;
2663 struct list_head work;
2664 struct dm_thin_new_mapping *m, *tmp;
2665 struct pool *pool = h->tc->pool;
2667 if (h->shared_read_entry) {
2668 INIT_LIST_HEAD(&work);
2669 ds_dec(h->shared_read_entry, &work);
2671 spin_lock_irqsave(&pool->lock, flags);
2672 list_for_each_entry_safe(m, tmp, &work, list) {
2675 __maybe_add_mapping(m);
2677 spin_unlock_irqrestore(&pool->lock, flags);
2680 if (h->all_io_entry) {
2681 INIT_LIST_HEAD(&work);
2682 ds_dec(h->all_io_entry, &work);
2683 spin_lock_irqsave(&pool->lock, flags);
2684 list_for_each_entry_safe(m, tmp, &work, list)
2685 list_add(&m->list, &pool->prepared_discards);
2686 spin_unlock_irqrestore(&pool->lock, flags);
2689 mempool_free(h, pool->endio_hook_pool);
2694 static void thin_postsuspend(struct dm_target *ti)
2696 if (dm_noflush_suspending(ti))
2697 requeue_io((struct thin_c *)ti->private);
2701 * <nr mapped sectors> <highest mapped sector>
2703 static int thin_status(struct dm_target *ti, status_type_t type,
2704 char *result, unsigned maxlen)
2708 dm_block_t mapped, highest;
2709 char buf[BDEVNAME_SIZE];
2710 struct thin_c *tc = ti->private;
2716 case STATUSTYPE_INFO:
2717 r = dm_thin_get_mapped_count(tc->td, &mapped);
2721 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2725 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2727 DMEMIT("%llu", ((highest + 1) *
2728 tc->pool->sectors_per_block) - 1);
2733 case STATUSTYPE_TABLE:
2735 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2736 (unsigned long) tc->dev_id);
2738 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2746 static int thin_iterate_devices(struct dm_target *ti,
2747 iterate_devices_callout_fn fn, void *data)
2750 struct thin_c *tc = ti->private;
2753 * We can't call dm_pool_get_data_dev_size() since that blocks. So
2754 * we follow a more convoluted path through to the pool's target.
2757 return 0; /* nothing is bound */
2759 blocks = tc->pool->ti->len >> tc->pool->block_shift;
2761 return fn(ti, tc->pool_dev, 0, tc->pool->sectors_per_block * blocks, data);
2766 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2768 struct thin_c *tc = ti->private;
2769 struct pool *pool = tc->pool;
2771 blk_limits_io_min(limits, 0);
2772 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2773 set_discard_limits(pool, limits);
2776 static struct target_type thin_target = {
2778 .version = {1, 1, 0},
2779 .module = THIS_MODULE,
2783 .end_io = thin_endio,
2784 .postsuspend = thin_postsuspend,
2785 .status = thin_status,
2786 .iterate_devices = thin_iterate_devices,
2787 .io_hints = thin_io_hints,
2790 /*----------------------------------------------------------------*/
2792 static int __init dm_thin_init(void)
2798 r = dm_register_target(&thin_target);
2802 r = dm_register_target(&pool_target);
2804 goto bad_pool_target;
2808 _cell_cache = KMEM_CACHE(dm_bio_prison_cell, 0);
2810 goto bad_cell_cache;
2812 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2813 if (!_new_mapping_cache)
2814 goto bad_new_mapping_cache;
2816 _endio_hook_cache = KMEM_CACHE(dm_thin_endio_hook, 0);
2817 if (!_endio_hook_cache)
2818 goto bad_endio_hook_cache;
2822 bad_endio_hook_cache:
2823 kmem_cache_destroy(_new_mapping_cache);
2824 bad_new_mapping_cache:
2825 kmem_cache_destroy(_cell_cache);
2827 dm_unregister_target(&pool_target);
2829 dm_unregister_target(&thin_target);
2834 static void dm_thin_exit(void)
2836 dm_unregister_target(&thin_target);
2837 dm_unregister_target(&pool_target);
2839 kmem_cache_destroy(_cell_cache);
2840 kmem_cache_destroy(_new_mapping_cache);
2841 kmem_cache_destroy(_endio_hook_cache);
2844 module_init(dm_thin_init);
2845 module_exit(dm_thin_exit);
2847 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2848 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2849 MODULE_LICENSE("GPL");