2 * Copyright (C) 2011-2012 Red Hat UK.
4 * This file is released under the GPL.
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison.h"
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/list.h>
15 #include <linux/rculist.h>
16 #include <linux/init.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/rbtree.h>
21 #define DM_MSG_PREFIX "thin"
26 #define ENDIO_HOOK_POOL_SIZE 1024
27 #define MAPPING_POOL_SIZE 1024
28 #define PRISON_CELLS 1024
29 #define COMMIT_PERIOD HZ
30 #define NO_SPACE_TIMEOUT_SECS 60
32 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
34 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
35 "A percentage of time allocated for copy on write");
38 * The block size of the device holding pool data must be
39 * between 64KB and 1GB.
41 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
42 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
45 * Device id is restricted to 24 bits.
47 #define MAX_DEV_ID ((1 << 24) - 1)
50 * How do we handle breaking sharing of data blocks?
51 * =================================================
53 * We use a standard copy-on-write btree to store the mappings for the
54 * devices (note I'm talking about copy-on-write of the metadata here, not
55 * the data). When you take an internal snapshot you clone the root node
56 * of the origin btree. After this there is no concept of an origin or a
57 * snapshot. They are just two device trees that happen to point to the
60 * When we get a write in we decide if it's to a shared data block using
61 * some timestamp magic. If it is, we have to break sharing.
63 * Let's say we write to a shared block in what was the origin. The
66 * i) plug io further to this physical block. (see bio_prison code).
68 * ii) quiesce any read io to that shared data block. Obviously
69 * including all devices that share this block. (see dm_deferred_set code)
71 * iii) copy the data block to a newly allocate block. This step can be
72 * missed out if the io covers the block. (schedule_copy).
74 * iv) insert the new mapping into the origin's btree
75 * (process_prepared_mapping). This act of inserting breaks some
76 * sharing of btree nodes between the two devices. Breaking sharing only
77 * effects the btree of that specific device. Btrees for the other
78 * devices that share the block never change. The btree for the origin
79 * device as it was after the last commit is untouched, ie. we're using
80 * persistent data structures in the functional programming sense.
82 * v) unplug io to this physical block, including the io that triggered
83 * the breaking of sharing.
85 * Steps (ii) and (iii) occur in parallel.
87 * The metadata _doesn't_ need to be committed before the io continues. We
88 * get away with this because the io is always written to a _new_ block.
89 * If there's a crash, then:
91 * - The origin mapping will point to the old origin block (the shared
92 * one). This will contain the data as it was before the io that triggered
93 * the breaking of sharing came in.
95 * - The snap mapping still points to the old block. As it would after
98 * The downside of this scheme is the timestamp magic isn't perfect, and
99 * will continue to think that data block in the snapshot device is shared
100 * even after the write to the origin has broken sharing. I suspect data
101 * blocks will typically be shared by many different devices, so we're
102 * breaking sharing n + 1 times, rather than n, where n is the number of
103 * devices that reference this data block. At the moment I think the
104 * benefits far, far outweigh the disadvantages.
107 /*----------------------------------------------------------------*/
112 static void build_data_key(struct dm_thin_device *td,
113 dm_block_t b, struct dm_cell_key *key)
116 key->dev = dm_thin_dev_id(td);
120 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
121 struct dm_cell_key *key)
124 key->dev = dm_thin_dev_id(td);
128 /*----------------------------------------------------------------*/
131 * A pool device ties together a metadata device and a data device. It
132 * also provides the interface for creating and destroying internal
135 struct dm_thin_new_mapping;
138 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
141 PM_WRITE, /* metadata may be changed */
142 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
143 PM_READ_ONLY, /* metadata may not be changed */
144 PM_FAIL, /* all I/O fails */
147 struct pool_features {
150 bool zero_new_blocks:1;
151 bool discard_enabled:1;
152 bool discard_passdown:1;
153 bool error_if_no_space:1;
157 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
158 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
161 struct list_head list;
162 struct dm_target *ti; /* Only set if a pool target is bound */
164 struct mapped_device *pool_md;
165 struct block_device *md_dev;
166 struct dm_pool_metadata *pmd;
168 dm_block_t low_water_blocks;
169 uint32_t sectors_per_block;
170 int sectors_per_block_shift;
172 struct pool_features pf;
173 bool low_water_triggered:1; /* A dm event has been sent */
175 struct dm_bio_prison *prison;
176 struct dm_kcopyd_client *copier;
178 struct workqueue_struct *wq;
179 struct work_struct worker;
180 struct delayed_work waker;
181 struct delayed_work no_space_timeout;
183 unsigned long last_commit_jiffies;
187 struct bio_list deferred_flush_bios;
188 struct list_head prepared_mappings;
189 struct list_head prepared_discards;
190 struct list_head active_thins;
192 struct dm_deferred_set *shared_read_ds;
193 struct dm_deferred_set *all_io_ds;
195 struct dm_thin_new_mapping *next_mapping;
196 mempool_t *mapping_pool;
198 process_bio_fn process_bio;
199 process_bio_fn process_discard;
201 process_mapping_fn process_prepared_mapping;
202 process_mapping_fn process_prepared_discard;
205 static enum pool_mode get_pool_mode(struct pool *pool);
206 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
209 * Target context for a pool.
212 struct dm_target *ti;
214 struct dm_dev *data_dev;
215 struct dm_dev *metadata_dev;
216 struct dm_target_callbacks callbacks;
218 dm_block_t low_water_blocks;
219 struct pool_features requested_pf; /* Features requested during table load */
220 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
224 * Target context for a thin.
227 struct list_head list;
228 struct dm_dev *pool_dev;
229 struct dm_dev *origin_dev;
233 struct dm_thin_device *td;
236 struct bio_list deferred_bio_list;
237 struct bio_list retry_on_resume_list;
238 struct rb_root sort_bio_list; /* sorted list of deferred bios */
241 * Ensures the thin is not destroyed until the worker has finished
242 * iterating the active_thins list.
245 struct completion can_destroy;
248 /*----------------------------------------------------------------*/
251 * wake_worker() is used when new work is queued and when pool_resume is
252 * ready to continue deferred IO processing.
254 static void wake_worker(struct pool *pool)
256 queue_work(pool->wq, &pool->worker);
259 /*----------------------------------------------------------------*/
261 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
262 struct dm_bio_prison_cell **cell_result)
265 struct dm_bio_prison_cell *cell_prealloc;
268 * Allocate a cell from the prison's mempool.
269 * This might block but it can't fail.
271 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
273 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
276 * We reused an old cell; we can get rid of
279 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
284 static void cell_release(struct pool *pool,
285 struct dm_bio_prison_cell *cell,
286 struct bio_list *bios)
288 dm_cell_release(pool->prison, cell, bios);
289 dm_bio_prison_free_cell(pool->prison, cell);
292 static void cell_release_no_holder(struct pool *pool,
293 struct dm_bio_prison_cell *cell,
294 struct bio_list *bios)
296 dm_cell_release_no_holder(pool->prison, cell, bios);
297 dm_bio_prison_free_cell(pool->prison, cell);
300 static void cell_defer_no_holder_no_free(struct thin_c *tc,
301 struct dm_bio_prison_cell *cell)
303 struct pool *pool = tc->pool;
306 spin_lock_irqsave(&tc->lock, flags);
307 dm_cell_release_no_holder(pool->prison, cell, &tc->deferred_bio_list);
308 spin_unlock_irqrestore(&tc->lock, flags);
313 static void cell_error(struct pool *pool,
314 struct dm_bio_prison_cell *cell)
316 dm_cell_error(pool->prison, cell);
317 dm_bio_prison_free_cell(pool->prison, cell);
320 /*----------------------------------------------------------------*/
323 * A global list of pools that uses a struct mapped_device as a key.
325 static struct dm_thin_pool_table {
327 struct list_head pools;
328 } dm_thin_pool_table;
330 static void pool_table_init(void)
332 mutex_init(&dm_thin_pool_table.mutex);
333 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
336 static void __pool_table_insert(struct pool *pool)
338 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
339 list_add(&pool->list, &dm_thin_pool_table.pools);
342 static void __pool_table_remove(struct pool *pool)
344 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
345 list_del(&pool->list);
348 static struct pool *__pool_table_lookup(struct mapped_device *md)
350 struct pool *pool = NULL, *tmp;
352 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
354 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
355 if (tmp->pool_md == md) {
364 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
366 struct pool *pool = NULL, *tmp;
368 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
370 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
371 if (tmp->md_dev == md_dev) {
380 /*----------------------------------------------------------------*/
382 struct dm_thin_endio_hook {
384 struct dm_deferred_entry *shared_read_entry;
385 struct dm_deferred_entry *all_io_entry;
386 struct dm_thin_new_mapping *overwrite_mapping;
387 struct rb_node rb_node;
390 static void requeue_bio_list(struct thin_c *tc, struct bio_list *master)
393 struct bio_list bios;
396 bio_list_init(&bios);
398 spin_lock_irqsave(&tc->lock, flags);
399 bio_list_merge(&bios, master);
400 bio_list_init(master);
401 spin_unlock_irqrestore(&tc->lock, flags);
403 while ((bio = bio_list_pop(&bios)))
404 bio_endio(bio, DM_ENDIO_REQUEUE);
407 static void requeue_io(struct thin_c *tc)
409 requeue_bio_list(tc, &tc->deferred_bio_list);
410 requeue_bio_list(tc, &tc->retry_on_resume_list);
413 static void error_thin_retry_list(struct thin_c *tc)
417 struct bio_list bios;
419 bio_list_init(&bios);
421 spin_lock_irqsave(&tc->lock, flags);
422 bio_list_merge(&bios, &tc->retry_on_resume_list);
423 bio_list_init(&tc->retry_on_resume_list);
424 spin_unlock_irqrestore(&tc->lock, flags);
426 while ((bio = bio_list_pop(&bios)))
430 static void error_retry_list(struct pool *pool)
435 list_for_each_entry_rcu(tc, &pool->active_thins, list)
436 error_thin_retry_list(tc);
441 * This section of code contains the logic for processing a thin device's IO.
442 * Much of the code depends on pool object resources (lists, workqueues, etc)
443 * but most is exclusively called from the thin target rather than the thin-pool
447 static bool block_size_is_power_of_two(struct pool *pool)
449 return pool->sectors_per_block_shift >= 0;
452 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
454 struct pool *pool = tc->pool;
455 sector_t block_nr = bio->bi_iter.bi_sector;
457 if (block_size_is_power_of_two(pool))
458 block_nr >>= pool->sectors_per_block_shift;
460 (void) sector_div(block_nr, pool->sectors_per_block);
465 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
467 struct pool *pool = tc->pool;
468 sector_t bi_sector = bio->bi_iter.bi_sector;
470 bio->bi_bdev = tc->pool_dev->bdev;
471 if (block_size_is_power_of_two(pool))
472 bio->bi_iter.bi_sector =
473 (block << pool->sectors_per_block_shift) |
474 (bi_sector & (pool->sectors_per_block - 1));
476 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
477 sector_div(bi_sector, pool->sectors_per_block);
480 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
482 bio->bi_bdev = tc->origin_dev->bdev;
485 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
487 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
488 dm_thin_changed_this_transaction(tc->td);
491 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
493 struct dm_thin_endio_hook *h;
495 if (bio->bi_rw & REQ_DISCARD)
498 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
499 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
502 static void issue(struct thin_c *tc, struct bio *bio)
504 struct pool *pool = tc->pool;
507 if (!bio_triggers_commit(tc, bio)) {
508 generic_make_request(bio);
513 * Complete bio with an error if earlier I/O caused changes to
514 * the metadata that can't be committed e.g, due to I/O errors
515 * on the metadata device.
517 if (dm_thin_aborted_changes(tc->td)) {
523 * Batch together any bios that trigger commits and then issue a
524 * single commit for them in process_deferred_bios().
526 spin_lock_irqsave(&pool->lock, flags);
527 bio_list_add(&pool->deferred_flush_bios, bio);
528 spin_unlock_irqrestore(&pool->lock, flags);
531 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
533 remap_to_origin(tc, bio);
537 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
540 remap(tc, bio, block);
544 /*----------------------------------------------------------------*/
547 * Bio endio functions.
549 struct dm_thin_new_mapping {
550 struct list_head list;
555 bool definitely_not_shared:1;
559 dm_block_t virt_block;
560 dm_block_t data_block;
561 struct dm_bio_prison_cell *cell, *cell2;
564 * If the bio covers the whole area of a block then we can avoid
565 * zeroing or copying. Instead this bio is hooked. The bio will
566 * still be in the cell, so care has to be taken to avoid issuing
570 bio_end_io_t *saved_bi_end_io;
573 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
575 struct pool *pool = m->tc->pool;
577 if (m->quiesced && m->prepared) {
578 list_add_tail(&m->list, &pool->prepared_mappings);
583 static void copy_complete(int read_err, unsigned long write_err, void *context)
586 struct dm_thin_new_mapping *m = context;
587 struct pool *pool = m->tc->pool;
589 m->err = read_err || write_err ? -EIO : 0;
591 spin_lock_irqsave(&pool->lock, flags);
593 __maybe_add_mapping(m);
594 spin_unlock_irqrestore(&pool->lock, flags);
597 static void overwrite_endio(struct bio *bio, int err)
600 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
601 struct dm_thin_new_mapping *m = h->overwrite_mapping;
602 struct pool *pool = m->tc->pool;
606 spin_lock_irqsave(&pool->lock, flags);
608 __maybe_add_mapping(m);
609 spin_unlock_irqrestore(&pool->lock, flags);
612 /*----------------------------------------------------------------*/
619 * Prepared mapping jobs.
623 * This sends the bios in the cell back to the deferred_bios list.
625 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
627 struct pool *pool = tc->pool;
630 spin_lock_irqsave(&tc->lock, flags);
631 cell_release(pool, cell, &tc->deferred_bio_list);
632 spin_unlock_irqrestore(&tc->lock, flags);
638 * Same as cell_defer above, except it omits the original holder of the cell.
640 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
642 struct pool *pool = tc->pool;
645 spin_lock_irqsave(&tc->lock, flags);
646 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
647 spin_unlock_irqrestore(&tc->lock, flags);
652 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
655 m->bio->bi_end_io = m->saved_bi_end_io;
656 atomic_inc(&m->bio->bi_remaining);
658 cell_error(m->tc->pool, m->cell);
660 mempool_free(m, m->tc->pool->mapping_pool);
663 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
665 struct thin_c *tc = m->tc;
666 struct pool *pool = tc->pool;
672 bio->bi_end_io = m->saved_bi_end_io;
673 atomic_inc(&bio->bi_remaining);
677 cell_error(pool, m->cell);
682 * Commit the prepared block into the mapping btree.
683 * Any I/O for this block arriving after this point will get
684 * remapped to it directly.
686 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
688 metadata_operation_failed(pool, "dm_thin_insert_block", r);
689 cell_error(pool, m->cell);
694 * Release any bios held while the block was being provisioned.
695 * If we are processing a write bio that completely covers the block,
696 * we already processed it so can ignore it now when processing
697 * the bios in the cell.
700 cell_defer_no_holder(tc, m->cell);
703 cell_defer(tc, m->cell);
707 mempool_free(m, pool->mapping_pool);
710 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
712 struct thin_c *tc = m->tc;
714 bio_io_error(m->bio);
715 cell_defer_no_holder(tc, m->cell);
716 cell_defer_no_holder(tc, m->cell2);
717 mempool_free(m, tc->pool->mapping_pool);
720 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
722 struct thin_c *tc = m->tc;
724 inc_all_io_entry(tc->pool, m->bio);
725 cell_defer_no_holder(tc, m->cell);
726 cell_defer_no_holder(tc, m->cell2);
729 if (m->definitely_not_shared)
730 remap_and_issue(tc, m->bio, m->data_block);
733 if (dm_pool_block_is_used(tc->pool->pmd, m->data_block, &used) || used)
734 bio_endio(m->bio, 0);
736 remap_and_issue(tc, m->bio, m->data_block);
739 bio_endio(m->bio, 0);
741 mempool_free(m, tc->pool->mapping_pool);
744 static void process_prepared_discard(struct dm_thin_new_mapping *m)
747 struct thin_c *tc = m->tc;
749 r = dm_thin_remove_block(tc->td, m->virt_block);
751 DMERR_LIMIT("dm_thin_remove_block() failed");
753 process_prepared_discard_passdown(m);
756 static void process_prepared(struct pool *pool, struct list_head *head,
757 process_mapping_fn *fn)
760 struct list_head maps;
761 struct dm_thin_new_mapping *m, *tmp;
763 INIT_LIST_HEAD(&maps);
764 spin_lock_irqsave(&pool->lock, flags);
765 list_splice_init(head, &maps);
766 spin_unlock_irqrestore(&pool->lock, flags);
768 list_for_each_entry_safe(m, tmp, &maps, list)
775 static int io_overlaps_block(struct pool *pool, struct bio *bio)
777 return bio->bi_iter.bi_size ==
778 (pool->sectors_per_block << SECTOR_SHIFT);
781 static int io_overwrites_block(struct pool *pool, struct bio *bio)
783 return (bio_data_dir(bio) == WRITE) &&
784 io_overlaps_block(pool, bio);
787 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
790 *save = bio->bi_end_io;
794 static int ensure_next_mapping(struct pool *pool)
796 if (pool->next_mapping)
799 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
801 return pool->next_mapping ? 0 : -ENOMEM;
804 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
806 struct dm_thin_new_mapping *m = pool->next_mapping;
808 BUG_ON(!pool->next_mapping);
810 memset(m, 0, sizeof(struct dm_thin_new_mapping));
811 INIT_LIST_HEAD(&m->list);
814 pool->next_mapping = NULL;
819 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
820 struct dm_dev *origin, dm_block_t data_origin,
821 dm_block_t data_dest,
822 struct dm_bio_prison_cell *cell, struct bio *bio)
825 struct pool *pool = tc->pool;
826 struct dm_thin_new_mapping *m = get_next_mapping(pool);
829 m->virt_block = virt_block;
830 m->data_block = data_dest;
833 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
837 * IO to pool_dev remaps to the pool target's data_dev.
839 * If the whole block of data is being overwritten, we can issue the
840 * bio immediately. Otherwise we use kcopyd to clone the data first.
842 if (io_overwrites_block(pool, bio)) {
843 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
845 h->overwrite_mapping = m;
847 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
848 inc_all_io_entry(pool, bio);
849 remap_and_issue(tc, bio, data_dest);
851 struct dm_io_region from, to;
853 from.bdev = origin->bdev;
854 from.sector = data_origin * pool->sectors_per_block;
855 from.count = pool->sectors_per_block;
857 to.bdev = tc->pool_dev->bdev;
858 to.sector = data_dest * pool->sectors_per_block;
859 to.count = pool->sectors_per_block;
861 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
862 0, copy_complete, m);
864 mempool_free(m, pool->mapping_pool);
865 DMERR_LIMIT("dm_kcopyd_copy() failed");
866 cell_error(pool, cell);
871 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
872 dm_block_t data_origin, dm_block_t data_dest,
873 struct dm_bio_prison_cell *cell, struct bio *bio)
875 schedule_copy(tc, virt_block, tc->pool_dev,
876 data_origin, data_dest, cell, bio);
879 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
880 dm_block_t data_dest,
881 struct dm_bio_prison_cell *cell, struct bio *bio)
883 schedule_copy(tc, virt_block, tc->origin_dev,
884 virt_block, data_dest, cell, bio);
887 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
888 dm_block_t data_block, struct dm_bio_prison_cell *cell,
891 struct pool *pool = tc->pool;
892 struct dm_thin_new_mapping *m = get_next_mapping(pool);
897 m->virt_block = virt_block;
898 m->data_block = data_block;
902 * If the whole block of data is being overwritten or we are not
903 * zeroing pre-existing data, we can issue the bio immediately.
904 * Otherwise we use kcopyd to zero the data first.
906 if (!pool->pf.zero_new_blocks)
907 process_prepared_mapping(m);
909 else if (io_overwrites_block(pool, bio)) {
910 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
912 h->overwrite_mapping = m;
914 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
915 inc_all_io_entry(pool, bio);
916 remap_and_issue(tc, bio, data_block);
919 struct dm_io_region to;
921 to.bdev = tc->pool_dev->bdev;
922 to.sector = data_block * pool->sectors_per_block;
923 to.count = pool->sectors_per_block;
925 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
927 mempool_free(m, pool->mapping_pool);
928 DMERR_LIMIT("dm_kcopyd_zero() failed");
929 cell_error(pool, cell);
935 * A non-zero return indicates read_only or fail_io mode.
936 * Many callers don't care about the return value.
938 static int commit(struct pool *pool)
942 if (get_pool_mode(pool) >= PM_READ_ONLY)
945 r = dm_pool_commit_metadata(pool->pmd);
947 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
952 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
956 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
957 DMWARN("%s: reached low water mark for data device: sending event.",
958 dm_device_name(pool->pool_md));
959 spin_lock_irqsave(&pool->lock, flags);
960 pool->low_water_triggered = true;
961 spin_unlock_irqrestore(&pool->lock, flags);
962 dm_table_event(pool->ti->table);
966 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
968 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
971 dm_block_t free_blocks;
972 struct pool *pool = tc->pool;
974 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
977 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
979 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
983 check_low_water_mark(pool, free_blocks);
987 * Try to commit to see if that will free up some
994 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
996 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1001 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1006 r = dm_pool_alloc_data_block(pool->pmd, result);
1008 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1016 * If we have run out of space, queue bios until the device is
1017 * resumed, presumably after having been reloaded with more space.
1019 static void retry_on_resume(struct bio *bio)
1021 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1022 struct thin_c *tc = h->tc;
1023 unsigned long flags;
1025 spin_lock_irqsave(&tc->lock, flags);
1026 bio_list_add(&tc->retry_on_resume_list, bio);
1027 spin_unlock_irqrestore(&tc->lock, flags);
1030 static bool should_error_unserviceable_bio(struct pool *pool)
1032 enum pool_mode m = get_pool_mode(pool);
1036 /* Shouldn't get here */
1037 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1040 case PM_OUT_OF_DATA_SPACE:
1041 return pool->pf.error_if_no_space;
1047 /* Shouldn't get here */
1048 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1053 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1055 if (should_error_unserviceable_bio(pool))
1058 retry_on_resume(bio);
1061 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1064 struct bio_list bios;
1066 if (should_error_unserviceable_bio(pool)) {
1067 cell_error(pool, cell);
1071 bio_list_init(&bios);
1072 cell_release(pool, cell, &bios);
1074 if (should_error_unserviceable_bio(pool))
1075 while ((bio = bio_list_pop(&bios)))
1078 while ((bio = bio_list_pop(&bios)))
1079 retry_on_resume(bio);
1082 static void process_discard(struct thin_c *tc, struct bio *bio)
1085 unsigned long flags;
1086 struct pool *pool = tc->pool;
1087 struct dm_bio_prison_cell *cell, *cell2;
1088 struct dm_cell_key key, key2;
1089 dm_block_t block = get_bio_block(tc, bio);
1090 struct dm_thin_lookup_result lookup_result;
1091 struct dm_thin_new_mapping *m;
1093 build_virtual_key(tc->td, block, &key);
1094 if (bio_detain(tc->pool, &key, bio, &cell))
1097 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1101 * Check nobody is fiddling with this pool block. This can
1102 * happen if someone's in the process of breaking sharing
1105 build_data_key(tc->td, lookup_result.block, &key2);
1106 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1107 cell_defer_no_holder(tc, cell);
1111 if (io_overlaps_block(pool, bio)) {
1113 * IO may still be going to the destination block. We must
1114 * quiesce before we can do the removal.
1116 m = get_next_mapping(pool);
1118 m->pass_discard = pool->pf.discard_passdown;
1119 m->definitely_not_shared = !lookup_result.shared;
1120 m->virt_block = block;
1121 m->data_block = lookup_result.block;
1126 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
1127 spin_lock_irqsave(&pool->lock, flags);
1128 list_add_tail(&m->list, &pool->prepared_discards);
1129 spin_unlock_irqrestore(&pool->lock, flags);
1133 inc_all_io_entry(pool, bio);
1134 cell_defer_no_holder(tc, cell);
1135 cell_defer_no_holder(tc, cell2);
1138 * The DM core makes sure that the discard doesn't span
1139 * a block boundary. So we submit the discard of a
1140 * partial block appropriately.
1142 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1143 remap_and_issue(tc, bio, lookup_result.block);
1151 * It isn't provisioned, just forget it.
1153 cell_defer_no_holder(tc, cell);
1158 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1160 cell_defer_no_holder(tc, cell);
1166 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1167 struct dm_cell_key *key,
1168 struct dm_thin_lookup_result *lookup_result,
1169 struct dm_bio_prison_cell *cell)
1172 dm_block_t data_block;
1173 struct pool *pool = tc->pool;
1175 r = alloc_data_block(tc, &data_block);
1178 schedule_internal_copy(tc, block, lookup_result->block,
1179 data_block, cell, bio);
1183 retry_bios_on_resume(pool, cell);
1187 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1189 cell_error(pool, cell);
1194 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1196 struct dm_thin_lookup_result *lookup_result)
1198 struct dm_bio_prison_cell *cell;
1199 struct pool *pool = tc->pool;
1200 struct dm_cell_key key;
1203 * If cell is already occupied, then sharing is already in the process
1204 * of being broken so we have nothing further to do here.
1206 build_data_key(tc->td, lookup_result->block, &key);
1207 if (bio_detain(pool, &key, bio, &cell))
1210 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size)
1211 break_sharing(tc, bio, block, &key, lookup_result, cell);
1213 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1215 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1216 inc_all_io_entry(pool, bio);
1217 cell_defer_no_holder(tc, cell);
1219 remap_and_issue(tc, bio, lookup_result->block);
1223 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1224 struct dm_bio_prison_cell *cell)
1227 dm_block_t data_block;
1228 struct pool *pool = tc->pool;
1231 * Remap empty bios (flushes) immediately, without provisioning.
1233 if (!bio->bi_iter.bi_size) {
1234 inc_all_io_entry(pool, bio);
1235 cell_defer_no_holder(tc, cell);
1237 remap_and_issue(tc, bio, 0);
1242 * Fill read bios with zeroes and complete them immediately.
1244 if (bio_data_dir(bio) == READ) {
1246 cell_defer_no_holder(tc, cell);
1251 r = alloc_data_block(tc, &data_block);
1255 schedule_external_copy(tc, block, data_block, cell, bio);
1257 schedule_zero(tc, block, data_block, cell, bio);
1261 retry_bios_on_resume(pool, cell);
1265 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1267 cell_error(pool, cell);
1272 static void process_bio(struct thin_c *tc, struct bio *bio)
1275 struct pool *pool = tc->pool;
1276 dm_block_t block = get_bio_block(tc, bio);
1277 struct dm_bio_prison_cell *cell;
1278 struct dm_cell_key key;
1279 struct dm_thin_lookup_result lookup_result;
1282 * If cell is already occupied, then the block is already
1283 * being provisioned so we have nothing further to do here.
1285 build_virtual_key(tc->td, block, &key);
1286 if (bio_detain(pool, &key, bio, &cell))
1289 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1292 if (lookup_result.shared) {
1293 process_shared_bio(tc, bio, block, &lookup_result);
1294 cell_defer_no_holder(tc, cell); /* FIXME: pass this cell into process_shared? */
1296 inc_all_io_entry(pool, bio);
1297 cell_defer_no_holder(tc, cell);
1299 remap_and_issue(tc, bio, lookup_result.block);
1304 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1305 inc_all_io_entry(pool, bio);
1306 cell_defer_no_holder(tc, cell);
1308 remap_to_origin_and_issue(tc, bio);
1310 provision_block(tc, bio, block, cell);
1314 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1316 cell_defer_no_holder(tc, cell);
1322 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1325 int rw = bio_data_dir(bio);
1326 dm_block_t block = get_bio_block(tc, bio);
1327 struct dm_thin_lookup_result lookup_result;
1329 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1332 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size)
1333 handle_unserviceable_bio(tc->pool, bio);
1335 inc_all_io_entry(tc->pool, bio);
1336 remap_and_issue(tc, bio, lookup_result.block);
1342 handle_unserviceable_bio(tc->pool, bio);
1346 if (tc->origin_dev) {
1347 inc_all_io_entry(tc->pool, bio);
1348 remap_to_origin_and_issue(tc, bio);
1357 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1364 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1369 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1375 * FIXME: should we also commit due to size of transaction, measured in
1378 static int need_commit_due_to_time(struct pool *pool)
1380 return jiffies < pool->last_commit_jiffies ||
1381 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1384 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1385 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1387 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1389 struct rb_node **rbp, *parent;
1390 struct dm_thin_endio_hook *pbd;
1391 sector_t bi_sector = bio->bi_iter.bi_sector;
1393 rbp = &tc->sort_bio_list.rb_node;
1397 pbd = thin_pbd(parent);
1399 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1400 rbp = &(*rbp)->rb_left;
1402 rbp = &(*rbp)->rb_right;
1405 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1406 rb_link_node(&pbd->rb_node, parent, rbp);
1407 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1410 static void __extract_sorted_bios(struct thin_c *tc)
1412 struct rb_node *node;
1413 struct dm_thin_endio_hook *pbd;
1416 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1417 pbd = thin_pbd(node);
1418 bio = thin_bio(pbd);
1420 bio_list_add(&tc->deferred_bio_list, bio);
1421 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1424 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1427 static void __sort_thin_deferred_bios(struct thin_c *tc)
1430 struct bio_list bios;
1432 bio_list_init(&bios);
1433 bio_list_merge(&bios, &tc->deferred_bio_list);
1434 bio_list_init(&tc->deferred_bio_list);
1436 /* Sort deferred_bio_list using rb-tree */
1437 while ((bio = bio_list_pop(&bios)))
1438 __thin_bio_rb_add(tc, bio);
1441 * Transfer the sorted bios in sort_bio_list back to
1442 * deferred_bio_list to allow lockless submission of
1445 __extract_sorted_bios(tc);
1448 static void process_thin_deferred_bios(struct thin_c *tc)
1450 struct pool *pool = tc->pool;
1451 unsigned long flags;
1453 struct bio_list bios;
1454 struct blk_plug plug;
1456 if (tc->requeue_mode) {
1457 requeue_bio_list(tc, &tc->deferred_bio_list);
1461 bio_list_init(&bios);
1463 spin_lock_irqsave(&tc->lock, flags);
1465 if (bio_list_empty(&tc->deferred_bio_list)) {
1466 spin_unlock_irqrestore(&tc->lock, flags);
1470 __sort_thin_deferred_bios(tc);
1472 bio_list_merge(&bios, &tc->deferred_bio_list);
1473 bio_list_init(&tc->deferred_bio_list);
1475 spin_unlock_irqrestore(&tc->lock, flags);
1477 blk_start_plug(&plug);
1478 while ((bio = bio_list_pop(&bios))) {
1480 * If we've got no free new_mapping structs, and processing
1481 * this bio might require one, we pause until there are some
1482 * prepared mappings to process.
1484 if (ensure_next_mapping(pool)) {
1485 spin_lock_irqsave(&tc->lock, flags);
1486 bio_list_add(&tc->deferred_bio_list, bio);
1487 bio_list_merge(&tc->deferred_bio_list, &bios);
1488 spin_unlock_irqrestore(&tc->lock, flags);
1492 if (bio->bi_rw & REQ_DISCARD)
1493 pool->process_discard(tc, bio);
1495 pool->process_bio(tc, bio);
1497 blk_finish_plug(&plug);
1500 static void thin_get(struct thin_c *tc);
1501 static void thin_put(struct thin_c *tc);
1504 * We can't hold rcu_read_lock() around code that can block. So we
1505 * find a thin with the rcu lock held; bump a refcount; then drop
1508 static struct thin_c *get_first_thin(struct pool *pool)
1510 struct thin_c *tc = NULL;
1513 if (!list_empty(&pool->active_thins)) {
1514 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
1522 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
1524 struct thin_c *old_tc = tc;
1527 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
1539 static void process_deferred_bios(struct pool *pool)
1541 unsigned long flags;
1543 struct bio_list bios;
1546 tc = get_first_thin(pool);
1548 process_thin_deferred_bios(tc);
1549 tc = get_next_thin(pool, tc);
1553 * If there are any deferred flush bios, we must commit
1554 * the metadata before issuing them.
1556 bio_list_init(&bios);
1557 spin_lock_irqsave(&pool->lock, flags);
1558 bio_list_merge(&bios, &pool->deferred_flush_bios);
1559 bio_list_init(&pool->deferred_flush_bios);
1560 spin_unlock_irqrestore(&pool->lock, flags);
1562 if (bio_list_empty(&bios) &&
1563 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
1567 while ((bio = bio_list_pop(&bios)))
1571 pool->last_commit_jiffies = jiffies;
1573 while ((bio = bio_list_pop(&bios)))
1574 generic_make_request(bio);
1577 static void do_worker(struct work_struct *ws)
1579 struct pool *pool = container_of(ws, struct pool, worker);
1581 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1582 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1583 process_deferred_bios(pool);
1587 * We want to commit periodically so that not too much
1588 * unwritten data builds up.
1590 static void do_waker(struct work_struct *ws)
1592 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1594 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1598 * We're holding onto IO to allow userland time to react. After the
1599 * timeout either the pool will have been resized (and thus back in
1600 * PM_WRITE mode), or we degrade to PM_READ_ONLY and start erroring IO.
1602 static void do_no_space_timeout(struct work_struct *ws)
1604 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
1607 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space)
1608 set_pool_mode(pool, PM_READ_ONLY);
1611 /*----------------------------------------------------------------*/
1613 struct noflush_work {
1614 struct work_struct worker;
1618 wait_queue_head_t wait;
1621 static void complete_noflush_work(struct noflush_work *w)
1623 atomic_set(&w->complete, 1);
1627 static void do_noflush_start(struct work_struct *ws)
1629 struct noflush_work *w = container_of(ws, struct noflush_work, worker);
1630 w->tc->requeue_mode = true;
1632 complete_noflush_work(w);
1635 static void do_noflush_stop(struct work_struct *ws)
1637 struct noflush_work *w = container_of(ws, struct noflush_work, worker);
1638 w->tc->requeue_mode = false;
1639 complete_noflush_work(w);
1642 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
1644 struct noflush_work w;
1646 INIT_WORK_ONSTACK(&w.worker, fn);
1648 atomic_set(&w.complete, 0);
1649 init_waitqueue_head(&w.wait);
1651 queue_work(tc->pool->wq, &w.worker);
1653 wait_event(w.wait, atomic_read(&w.complete));
1656 /*----------------------------------------------------------------*/
1658 static enum pool_mode get_pool_mode(struct pool *pool)
1660 return pool->pf.mode;
1663 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
1665 dm_table_event(pool->ti->table);
1666 DMINFO("%s: switching pool to %s mode",
1667 dm_device_name(pool->pool_md), new_mode);
1670 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
1672 struct pool_c *pt = pool->ti->private;
1673 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
1674 enum pool_mode old_mode = get_pool_mode(pool);
1675 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
1678 * Never allow the pool to transition to PM_WRITE mode if user
1679 * intervention is required to verify metadata and data consistency.
1681 if (new_mode == PM_WRITE && needs_check) {
1682 DMERR("%s: unable to switch pool to write mode until repaired.",
1683 dm_device_name(pool->pool_md));
1684 if (old_mode != new_mode)
1685 new_mode = old_mode;
1687 new_mode = PM_READ_ONLY;
1690 * If we were in PM_FAIL mode, rollback of metadata failed. We're
1691 * not going to recover without a thin_repair. So we never let the
1692 * pool move out of the old mode.
1694 if (old_mode == PM_FAIL)
1695 new_mode = old_mode;
1699 if (old_mode != new_mode)
1700 notify_of_pool_mode_change(pool, "failure");
1701 dm_pool_metadata_read_only(pool->pmd);
1702 pool->process_bio = process_bio_fail;
1703 pool->process_discard = process_bio_fail;
1704 pool->process_prepared_mapping = process_prepared_mapping_fail;
1705 pool->process_prepared_discard = process_prepared_discard_fail;
1707 error_retry_list(pool);
1711 if (old_mode != new_mode)
1712 notify_of_pool_mode_change(pool, "read-only");
1713 dm_pool_metadata_read_only(pool->pmd);
1714 pool->process_bio = process_bio_read_only;
1715 pool->process_discard = process_bio_success;
1716 pool->process_prepared_mapping = process_prepared_mapping_fail;
1717 pool->process_prepared_discard = process_prepared_discard_passdown;
1719 error_retry_list(pool);
1722 case PM_OUT_OF_DATA_SPACE:
1724 * Ideally we'd never hit this state; the low water mark
1725 * would trigger userland to extend the pool before we
1726 * completely run out of data space. However, many small
1727 * IOs to unprovisioned space can consume data space at an
1728 * alarming rate. Adjust your low water mark if you're
1729 * frequently seeing this mode.
1731 if (old_mode != new_mode)
1732 notify_of_pool_mode_change(pool, "out-of-data-space");
1733 pool->process_bio = process_bio_read_only;
1734 pool->process_discard = process_discard;
1735 pool->process_prepared_mapping = process_prepared_mapping;
1736 pool->process_prepared_discard = process_prepared_discard_passdown;
1738 if (!pool->pf.error_if_no_space && no_space_timeout)
1739 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
1743 if (old_mode != new_mode)
1744 notify_of_pool_mode_change(pool, "write");
1745 dm_pool_metadata_read_write(pool->pmd);
1746 pool->process_bio = process_bio;
1747 pool->process_discard = process_discard;
1748 pool->process_prepared_mapping = process_prepared_mapping;
1749 pool->process_prepared_discard = process_prepared_discard;
1753 pool->pf.mode = new_mode;
1755 * The pool mode may have changed, sync it so bind_control_target()
1756 * doesn't cause an unexpected mode transition on resume.
1758 pt->adjusted_pf.mode = new_mode;
1761 static void abort_transaction(struct pool *pool)
1763 const char *dev_name = dm_device_name(pool->pool_md);
1765 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1766 if (dm_pool_abort_metadata(pool->pmd)) {
1767 DMERR("%s: failed to abort metadata transaction", dev_name);
1768 set_pool_mode(pool, PM_FAIL);
1771 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
1772 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1773 set_pool_mode(pool, PM_FAIL);
1777 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
1779 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1780 dm_device_name(pool->pool_md), op, r);
1782 abort_transaction(pool);
1783 set_pool_mode(pool, PM_READ_ONLY);
1786 /*----------------------------------------------------------------*/
1789 * Mapping functions.
1793 * Called only while mapping a thin bio to hand it over to the workqueue.
1795 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1797 unsigned long flags;
1798 struct pool *pool = tc->pool;
1800 spin_lock_irqsave(&tc->lock, flags);
1801 bio_list_add(&tc->deferred_bio_list, bio);
1802 spin_unlock_irqrestore(&tc->lock, flags);
1807 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1809 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1812 h->shared_read_entry = NULL;
1813 h->all_io_entry = NULL;
1814 h->overwrite_mapping = NULL;
1818 * Non-blocking function called from the thin target's map function.
1820 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1823 struct thin_c *tc = ti->private;
1824 dm_block_t block = get_bio_block(tc, bio);
1825 struct dm_thin_device *td = tc->td;
1826 struct dm_thin_lookup_result result;
1827 struct dm_bio_prison_cell cell1, cell2;
1828 struct dm_bio_prison_cell *cell_result;
1829 struct dm_cell_key key;
1831 thin_hook_bio(tc, bio);
1833 if (tc->requeue_mode) {
1834 bio_endio(bio, DM_ENDIO_REQUEUE);
1835 return DM_MAPIO_SUBMITTED;
1838 if (get_pool_mode(tc->pool) == PM_FAIL) {
1840 return DM_MAPIO_SUBMITTED;
1843 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1844 thin_defer_bio(tc, bio);
1845 return DM_MAPIO_SUBMITTED;
1848 r = dm_thin_find_block(td, block, 0, &result);
1851 * Note that we defer readahead too.
1855 if (unlikely(result.shared)) {
1857 * We have a race condition here between the
1858 * result.shared value returned by the lookup and
1859 * snapshot creation, which may cause new
1862 * To avoid this always quiesce the origin before
1863 * taking the snap. You want to do this anyway to
1864 * ensure a consistent application view
1867 * More distant ancestors are irrelevant. The
1868 * shared flag will be set in their case.
1870 thin_defer_bio(tc, bio);
1871 return DM_MAPIO_SUBMITTED;
1874 build_virtual_key(tc->td, block, &key);
1875 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1, &cell_result))
1876 return DM_MAPIO_SUBMITTED;
1878 build_data_key(tc->td, result.block, &key);
1879 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2, &cell_result)) {
1880 cell_defer_no_holder_no_free(tc, &cell1);
1881 return DM_MAPIO_SUBMITTED;
1884 inc_all_io_entry(tc->pool, bio);
1885 cell_defer_no_holder_no_free(tc, &cell2);
1886 cell_defer_no_holder_no_free(tc, &cell1);
1888 remap(tc, bio, result.block);
1889 return DM_MAPIO_REMAPPED;
1892 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1894 * This block isn't provisioned, and we have no way
1897 handle_unserviceable_bio(tc->pool, bio);
1898 return DM_MAPIO_SUBMITTED;
1904 * In future, the failed dm_thin_find_block above could
1905 * provide the hint to load the metadata into cache.
1907 thin_defer_bio(tc, bio);
1908 return DM_MAPIO_SUBMITTED;
1912 * Must always call bio_io_error on failure.
1913 * dm_thin_find_block can fail with -EINVAL if the
1914 * pool is switched to fail-io mode.
1917 return DM_MAPIO_SUBMITTED;
1921 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1923 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1924 struct request_queue *q;
1926 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
1929 q = bdev_get_queue(pt->data_dev->bdev);
1930 return bdi_congested(&q->backing_dev_info, bdi_bits);
1933 static void requeue_bios(struct pool *pool)
1935 unsigned long flags;
1939 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
1940 spin_lock_irqsave(&tc->lock, flags);
1941 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
1942 bio_list_init(&tc->retry_on_resume_list);
1943 spin_unlock_irqrestore(&tc->lock, flags);
1948 /*----------------------------------------------------------------
1949 * Binding of control targets to a pool object
1950 *--------------------------------------------------------------*/
1951 static bool data_dev_supports_discard(struct pool_c *pt)
1953 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1955 return q && blk_queue_discard(q);
1958 static bool is_factor(sector_t block_size, uint32_t n)
1960 return !sector_div(block_size, n);
1964 * If discard_passdown was enabled verify that the data device
1965 * supports discards. Disable discard_passdown if not.
1967 static void disable_passdown_if_not_supported(struct pool_c *pt)
1969 struct pool *pool = pt->pool;
1970 struct block_device *data_bdev = pt->data_dev->bdev;
1971 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1972 sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1973 const char *reason = NULL;
1974 char buf[BDEVNAME_SIZE];
1976 if (!pt->adjusted_pf.discard_passdown)
1979 if (!data_dev_supports_discard(pt))
1980 reason = "discard unsupported";
1982 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1983 reason = "max discard sectors smaller than a block";
1985 else if (data_limits->discard_granularity > block_size)
1986 reason = "discard granularity larger than a block";
1988 else if (!is_factor(block_size, data_limits->discard_granularity))
1989 reason = "discard granularity not a factor of block size";
1992 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1993 pt->adjusted_pf.discard_passdown = false;
1997 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1999 struct pool_c *pt = ti->private;
2002 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2004 enum pool_mode old_mode = get_pool_mode(pool);
2005 enum pool_mode new_mode = pt->adjusted_pf.mode;
2008 * Don't change the pool's mode until set_pool_mode() below.
2009 * Otherwise the pool's process_* function pointers may
2010 * not match the desired pool mode.
2012 pt->adjusted_pf.mode = old_mode;
2015 pool->pf = pt->adjusted_pf;
2016 pool->low_water_blocks = pt->low_water_blocks;
2018 set_pool_mode(pool, new_mode);
2023 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2029 /*----------------------------------------------------------------
2031 *--------------------------------------------------------------*/
2032 /* Initialize pool features. */
2033 static void pool_features_init(struct pool_features *pf)
2035 pf->mode = PM_WRITE;
2036 pf->zero_new_blocks = true;
2037 pf->discard_enabled = true;
2038 pf->discard_passdown = true;
2039 pf->error_if_no_space = false;
2042 static void __pool_destroy(struct pool *pool)
2044 __pool_table_remove(pool);
2046 if (dm_pool_metadata_close(pool->pmd) < 0)
2047 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2049 dm_bio_prison_destroy(pool->prison);
2050 dm_kcopyd_client_destroy(pool->copier);
2053 destroy_workqueue(pool->wq);
2055 if (pool->next_mapping)
2056 mempool_free(pool->next_mapping, pool->mapping_pool);
2057 mempool_destroy(pool->mapping_pool);
2058 dm_deferred_set_destroy(pool->shared_read_ds);
2059 dm_deferred_set_destroy(pool->all_io_ds);
2063 static struct kmem_cache *_new_mapping_cache;
2065 static struct pool *pool_create(struct mapped_device *pool_md,
2066 struct block_device *metadata_dev,
2067 unsigned long block_size,
2068 int read_only, char **error)
2073 struct dm_pool_metadata *pmd;
2074 bool format_device = read_only ? false : true;
2076 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2078 *error = "Error creating metadata object";
2079 return (struct pool *)pmd;
2082 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2084 *error = "Error allocating memory for pool";
2085 err_p = ERR_PTR(-ENOMEM);
2090 pool->sectors_per_block = block_size;
2091 if (block_size & (block_size - 1))
2092 pool->sectors_per_block_shift = -1;
2094 pool->sectors_per_block_shift = __ffs(block_size);
2095 pool->low_water_blocks = 0;
2096 pool_features_init(&pool->pf);
2097 pool->prison = dm_bio_prison_create(PRISON_CELLS);
2098 if (!pool->prison) {
2099 *error = "Error creating pool's bio prison";
2100 err_p = ERR_PTR(-ENOMEM);
2104 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2105 if (IS_ERR(pool->copier)) {
2106 r = PTR_ERR(pool->copier);
2107 *error = "Error creating pool's kcopyd client";
2109 goto bad_kcopyd_client;
2113 * Create singlethreaded workqueue that will service all devices
2114 * that use this metadata.
2116 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2118 *error = "Error creating pool's workqueue";
2119 err_p = ERR_PTR(-ENOMEM);
2123 INIT_WORK(&pool->worker, do_worker);
2124 INIT_DELAYED_WORK(&pool->waker, do_waker);
2125 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2126 spin_lock_init(&pool->lock);
2127 bio_list_init(&pool->deferred_flush_bios);
2128 INIT_LIST_HEAD(&pool->prepared_mappings);
2129 INIT_LIST_HEAD(&pool->prepared_discards);
2130 INIT_LIST_HEAD(&pool->active_thins);
2131 pool->low_water_triggered = false;
2133 pool->shared_read_ds = dm_deferred_set_create();
2134 if (!pool->shared_read_ds) {
2135 *error = "Error creating pool's shared read deferred set";
2136 err_p = ERR_PTR(-ENOMEM);
2137 goto bad_shared_read_ds;
2140 pool->all_io_ds = dm_deferred_set_create();
2141 if (!pool->all_io_ds) {
2142 *error = "Error creating pool's all io deferred set";
2143 err_p = ERR_PTR(-ENOMEM);
2147 pool->next_mapping = NULL;
2148 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2149 _new_mapping_cache);
2150 if (!pool->mapping_pool) {
2151 *error = "Error creating pool's mapping mempool";
2152 err_p = ERR_PTR(-ENOMEM);
2153 goto bad_mapping_pool;
2156 pool->ref_count = 1;
2157 pool->last_commit_jiffies = jiffies;
2158 pool->pool_md = pool_md;
2159 pool->md_dev = metadata_dev;
2160 __pool_table_insert(pool);
2165 dm_deferred_set_destroy(pool->all_io_ds);
2167 dm_deferred_set_destroy(pool->shared_read_ds);
2169 destroy_workqueue(pool->wq);
2171 dm_kcopyd_client_destroy(pool->copier);
2173 dm_bio_prison_destroy(pool->prison);
2177 if (dm_pool_metadata_close(pmd))
2178 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2183 static void __pool_inc(struct pool *pool)
2185 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2189 static void __pool_dec(struct pool *pool)
2191 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2192 BUG_ON(!pool->ref_count);
2193 if (!--pool->ref_count)
2194 __pool_destroy(pool);
2197 static struct pool *__pool_find(struct mapped_device *pool_md,
2198 struct block_device *metadata_dev,
2199 unsigned long block_size, int read_only,
2200 char **error, int *created)
2202 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2205 if (pool->pool_md != pool_md) {
2206 *error = "metadata device already in use by a pool";
2207 return ERR_PTR(-EBUSY);
2212 pool = __pool_table_lookup(pool_md);
2214 if (pool->md_dev != metadata_dev) {
2215 *error = "different pool cannot replace a pool";
2216 return ERR_PTR(-EINVAL);
2221 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2229 /*----------------------------------------------------------------
2230 * Pool target methods
2231 *--------------------------------------------------------------*/
2232 static void pool_dtr(struct dm_target *ti)
2234 struct pool_c *pt = ti->private;
2236 mutex_lock(&dm_thin_pool_table.mutex);
2238 unbind_control_target(pt->pool, ti);
2239 __pool_dec(pt->pool);
2240 dm_put_device(ti, pt->metadata_dev);
2241 dm_put_device(ti, pt->data_dev);
2244 mutex_unlock(&dm_thin_pool_table.mutex);
2247 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2248 struct dm_target *ti)
2252 const char *arg_name;
2254 static struct dm_arg _args[] = {
2255 {0, 4, "Invalid number of pool feature arguments"},
2259 * No feature arguments supplied.
2264 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2268 while (argc && !r) {
2269 arg_name = dm_shift_arg(as);
2272 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2273 pf->zero_new_blocks = false;
2275 else if (!strcasecmp(arg_name, "ignore_discard"))
2276 pf->discard_enabled = false;
2278 else if (!strcasecmp(arg_name, "no_discard_passdown"))
2279 pf->discard_passdown = false;
2281 else if (!strcasecmp(arg_name, "read_only"))
2282 pf->mode = PM_READ_ONLY;
2284 else if (!strcasecmp(arg_name, "error_if_no_space"))
2285 pf->error_if_no_space = true;
2288 ti->error = "Unrecognised pool feature requested";
2297 static void metadata_low_callback(void *context)
2299 struct pool *pool = context;
2301 DMWARN("%s: reached low water mark for metadata device: sending event.",
2302 dm_device_name(pool->pool_md));
2304 dm_table_event(pool->ti->table);
2307 static sector_t get_dev_size(struct block_device *bdev)
2309 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
2312 static void warn_if_metadata_device_too_big(struct block_device *bdev)
2314 sector_t metadata_dev_size = get_dev_size(bdev);
2315 char buffer[BDEVNAME_SIZE];
2317 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
2318 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2319 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
2322 static sector_t get_metadata_dev_size(struct block_device *bdev)
2324 sector_t metadata_dev_size = get_dev_size(bdev);
2326 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
2327 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
2329 return metadata_dev_size;
2332 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
2334 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
2336 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
2338 return metadata_dev_size;
2342 * When a metadata threshold is crossed a dm event is triggered, and
2343 * userland should respond by growing the metadata device. We could let
2344 * userland set the threshold, like we do with the data threshold, but I'm
2345 * not sure they know enough to do this well.
2347 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
2350 * 4M is ample for all ops with the possible exception of thin
2351 * device deletion which is harmless if it fails (just retry the
2352 * delete after you've grown the device).
2354 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
2355 return min((dm_block_t)1024ULL /* 4M */, quarter);
2359 * thin-pool <metadata dev> <data dev>
2360 * <data block size (sectors)>
2361 * <low water mark (blocks)>
2362 * [<#feature args> [<arg>]*]
2364 * Optional feature arguments are:
2365 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
2366 * ignore_discard: disable discard
2367 * no_discard_passdown: don't pass discards down to the data device
2368 * read_only: Don't allow any changes to be made to the pool metadata.
2369 * error_if_no_space: error IOs, instead of queueing, if no space.
2371 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
2373 int r, pool_created = 0;
2376 struct pool_features pf;
2377 struct dm_arg_set as;
2378 struct dm_dev *data_dev;
2379 unsigned long block_size;
2380 dm_block_t low_water_blocks;
2381 struct dm_dev *metadata_dev;
2382 fmode_t metadata_mode;
2385 * FIXME Remove validation from scope of lock.
2387 mutex_lock(&dm_thin_pool_table.mutex);
2390 ti->error = "Invalid argument count";
2399 * Set default pool features.
2401 pool_features_init(&pf);
2403 dm_consume_args(&as, 4);
2404 r = parse_pool_features(&as, &pf, ti);
2408 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2409 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2411 ti->error = "Error opening metadata block device";
2414 warn_if_metadata_device_too_big(metadata_dev->bdev);
2416 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2418 ti->error = "Error getting data device";
2422 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2423 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2424 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2425 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2426 ti->error = "Invalid block size";
2431 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2432 ti->error = "Invalid low water mark";
2437 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2443 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2444 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2451 * 'pool_created' reflects whether this is the first table load.
2452 * Top level discard support is not allowed to be changed after
2453 * initial load. This would require a pool reload to trigger thin
2456 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2457 ti->error = "Discard support cannot be disabled once enabled";
2459 goto out_flags_changed;
2464 pt->metadata_dev = metadata_dev;
2465 pt->data_dev = data_dev;
2466 pt->low_water_blocks = low_water_blocks;
2467 pt->adjusted_pf = pt->requested_pf = pf;
2468 ti->num_flush_bios = 1;
2471 * Only need to enable discards if the pool should pass
2472 * them down to the data device. The thin device's discard
2473 * processing will cause mappings to be removed from the btree.
2475 ti->discard_zeroes_data_unsupported = true;
2476 if (pf.discard_enabled && pf.discard_passdown) {
2477 ti->num_discard_bios = 1;
2480 * Setting 'discards_supported' circumvents the normal
2481 * stacking of discard limits (this keeps the pool and
2482 * thin devices' discard limits consistent).
2484 ti->discards_supported = true;
2488 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2489 calc_metadata_threshold(pt),
2490 metadata_low_callback,
2495 pt->callbacks.congested_fn = pool_is_congested;
2496 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2498 mutex_unlock(&dm_thin_pool_table.mutex);
2507 dm_put_device(ti, data_dev);
2509 dm_put_device(ti, metadata_dev);
2511 mutex_unlock(&dm_thin_pool_table.mutex);
2516 static int pool_map(struct dm_target *ti, struct bio *bio)
2519 struct pool_c *pt = ti->private;
2520 struct pool *pool = pt->pool;
2521 unsigned long flags;
2524 * As this is a singleton target, ti->begin is always zero.
2526 spin_lock_irqsave(&pool->lock, flags);
2527 bio->bi_bdev = pt->data_dev->bdev;
2528 r = DM_MAPIO_REMAPPED;
2529 spin_unlock_irqrestore(&pool->lock, flags);
2534 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
2537 struct pool_c *pt = ti->private;
2538 struct pool *pool = pt->pool;
2539 sector_t data_size = ti->len;
2540 dm_block_t sb_data_size;
2542 *need_commit = false;
2544 (void) sector_div(data_size, pool->sectors_per_block);
2546 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2548 DMERR("%s: failed to retrieve data device size",
2549 dm_device_name(pool->pool_md));
2553 if (data_size < sb_data_size) {
2554 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
2555 dm_device_name(pool->pool_md),
2556 (unsigned long long)data_size, sb_data_size);
2559 } else if (data_size > sb_data_size) {
2560 if (dm_pool_metadata_needs_check(pool->pmd)) {
2561 DMERR("%s: unable to grow the data device until repaired.",
2562 dm_device_name(pool->pool_md));
2567 DMINFO("%s: growing the data device from %llu to %llu blocks",
2568 dm_device_name(pool->pool_md),
2569 sb_data_size, (unsigned long long)data_size);
2570 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2572 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
2576 *need_commit = true;
2582 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
2585 struct pool_c *pt = ti->private;
2586 struct pool *pool = pt->pool;
2587 dm_block_t metadata_dev_size, sb_metadata_dev_size;
2589 *need_commit = false;
2591 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
2593 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
2595 DMERR("%s: failed to retrieve metadata device size",
2596 dm_device_name(pool->pool_md));
2600 if (metadata_dev_size < sb_metadata_dev_size) {
2601 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
2602 dm_device_name(pool->pool_md),
2603 metadata_dev_size, sb_metadata_dev_size);
2606 } else if (metadata_dev_size > sb_metadata_dev_size) {
2607 if (dm_pool_metadata_needs_check(pool->pmd)) {
2608 DMERR("%s: unable to grow the metadata device until repaired.",
2609 dm_device_name(pool->pool_md));
2613 warn_if_metadata_device_too_big(pool->md_dev);
2614 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
2615 dm_device_name(pool->pool_md),
2616 sb_metadata_dev_size, metadata_dev_size);
2617 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
2619 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
2623 *need_commit = true;
2630 * Retrieves the number of blocks of the data device from
2631 * the superblock and compares it to the actual device size,
2632 * thus resizing the data device in case it has grown.
2634 * This both copes with opening preallocated data devices in the ctr
2635 * being followed by a resume
2637 * calling the resume method individually after userspace has
2638 * grown the data device in reaction to a table event.
2640 static int pool_preresume(struct dm_target *ti)
2643 bool need_commit1, need_commit2;
2644 struct pool_c *pt = ti->private;
2645 struct pool *pool = pt->pool;
2648 * Take control of the pool object.
2650 r = bind_control_target(pool, ti);
2654 r = maybe_resize_data_dev(ti, &need_commit1);
2658 r = maybe_resize_metadata_dev(ti, &need_commit2);
2662 if (need_commit1 || need_commit2)
2663 (void) commit(pool);
2668 static void pool_resume(struct dm_target *ti)
2670 struct pool_c *pt = ti->private;
2671 struct pool *pool = pt->pool;
2672 unsigned long flags;
2674 spin_lock_irqsave(&pool->lock, flags);
2675 pool->low_water_triggered = false;
2676 spin_unlock_irqrestore(&pool->lock, flags);
2679 do_waker(&pool->waker.work);
2682 static void pool_postsuspend(struct dm_target *ti)
2684 struct pool_c *pt = ti->private;
2685 struct pool *pool = pt->pool;
2687 cancel_delayed_work(&pool->waker);
2688 cancel_delayed_work(&pool->no_space_timeout);
2689 flush_workqueue(pool->wq);
2690 (void) commit(pool);
2693 static int check_arg_count(unsigned argc, unsigned args_required)
2695 if (argc != args_required) {
2696 DMWARN("Message received with %u arguments instead of %u.",
2697 argc, args_required);
2704 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2706 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2707 *dev_id <= MAX_DEV_ID)
2711 DMWARN("Message received with invalid device id: %s", arg);
2716 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2721 r = check_arg_count(argc, 2);
2725 r = read_dev_id(argv[1], &dev_id, 1);
2729 r = dm_pool_create_thin(pool->pmd, dev_id);
2731 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2739 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2742 dm_thin_id origin_dev_id;
2745 r = check_arg_count(argc, 3);
2749 r = read_dev_id(argv[1], &dev_id, 1);
2753 r = read_dev_id(argv[2], &origin_dev_id, 1);
2757 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2759 DMWARN("Creation of new snapshot %s of device %s failed.",
2767 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2772 r = check_arg_count(argc, 2);
2776 r = read_dev_id(argv[1], &dev_id, 1);
2780 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2782 DMWARN("Deletion of thin device %s failed.", argv[1]);
2787 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2789 dm_thin_id old_id, new_id;
2792 r = check_arg_count(argc, 3);
2796 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2797 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2801 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2802 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2806 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2808 DMWARN("Failed to change transaction id from %s to %s.",
2816 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2820 r = check_arg_count(argc, 1);
2824 (void) commit(pool);
2826 r = dm_pool_reserve_metadata_snap(pool->pmd);
2828 DMWARN("reserve_metadata_snap message failed.");
2833 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2837 r = check_arg_count(argc, 1);
2841 r = dm_pool_release_metadata_snap(pool->pmd);
2843 DMWARN("release_metadata_snap message failed.");
2849 * Messages supported:
2850 * create_thin <dev_id>
2851 * create_snap <dev_id> <origin_id>
2853 * trim <dev_id> <new_size_in_sectors>
2854 * set_transaction_id <current_trans_id> <new_trans_id>
2855 * reserve_metadata_snap
2856 * release_metadata_snap
2858 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2861 struct pool_c *pt = ti->private;
2862 struct pool *pool = pt->pool;
2864 if (!strcasecmp(argv[0], "create_thin"))
2865 r = process_create_thin_mesg(argc, argv, pool);
2867 else if (!strcasecmp(argv[0], "create_snap"))
2868 r = process_create_snap_mesg(argc, argv, pool);
2870 else if (!strcasecmp(argv[0], "delete"))
2871 r = process_delete_mesg(argc, argv, pool);
2873 else if (!strcasecmp(argv[0], "set_transaction_id"))
2874 r = process_set_transaction_id_mesg(argc, argv, pool);
2876 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2877 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2879 else if (!strcasecmp(argv[0], "release_metadata_snap"))
2880 r = process_release_metadata_snap_mesg(argc, argv, pool);
2883 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2886 (void) commit(pool);
2891 static void emit_flags(struct pool_features *pf, char *result,
2892 unsigned sz, unsigned maxlen)
2894 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2895 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
2896 pf->error_if_no_space;
2897 DMEMIT("%u ", count);
2899 if (!pf->zero_new_blocks)
2900 DMEMIT("skip_block_zeroing ");
2902 if (!pf->discard_enabled)
2903 DMEMIT("ignore_discard ");
2905 if (!pf->discard_passdown)
2906 DMEMIT("no_discard_passdown ");
2908 if (pf->mode == PM_READ_ONLY)
2909 DMEMIT("read_only ");
2911 if (pf->error_if_no_space)
2912 DMEMIT("error_if_no_space ");
2917 * <transaction id> <used metadata sectors>/<total metadata sectors>
2918 * <used data sectors>/<total data sectors> <held metadata root>
2920 static void pool_status(struct dm_target *ti, status_type_t type,
2921 unsigned status_flags, char *result, unsigned maxlen)
2925 uint64_t transaction_id;
2926 dm_block_t nr_free_blocks_data;
2927 dm_block_t nr_free_blocks_metadata;
2928 dm_block_t nr_blocks_data;
2929 dm_block_t nr_blocks_metadata;
2930 dm_block_t held_root;
2931 char buf[BDEVNAME_SIZE];
2932 char buf2[BDEVNAME_SIZE];
2933 struct pool_c *pt = ti->private;
2934 struct pool *pool = pt->pool;
2937 case STATUSTYPE_INFO:
2938 if (get_pool_mode(pool) == PM_FAIL) {
2943 /* Commit to ensure statistics aren't out-of-date */
2944 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2945 (void) commit(pool);
2947 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
2949 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
2950 dm_device_name(pool->pool_md), r);
2954 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
2956 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
2957 dm_device_name(pool->pool_md), r);
2961 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2963 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
2964 dm_device_name(pool->pool_md), r);
2968 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
2970 DMERR("%s: dm_pool_get_free_block_count returned %d",
2971 dm_device_name(pool->pool_md), r);
2975 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2977 DMERR("%s: dm_pool_get_data_dev_size returned %d",
2978 dm_device_name(pool->pool_md), r);
2982 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2984 DMERR("%s: dm_pool_get_metadata_snap returned %d",
2985 dm_device_name(pool->pool_md), r);
2989 DMEMIT("%llu %llu/%llu %llu/%llu ",
2990 (unsigned long long)transaction_id,
2991 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2992 (unsigned long long)nr_blocks_metadata,
2993 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2994 (unsigned long long)nr_blocks_data);
2997 DMEMIT("%llu ", held_root);
3001 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3002 DMEMIT("out_of_data_space ");
3003 else if (pool->pf.mode == PM_READ_ONLY)
3008 if (!pool->pf.discard_enabled)
3009 DMEMIT("ignore_discard ");
3010 else if (pool->pf.discard_passdown)
3011 DMEMIT("discard_passdown ");
3013 DMEMIT("no_discard_passdown ");
3015 if (pool->pf.error_if_no_space)
3016 DMEMIT("error_if_no_space ");
3018 DMEMIT("queue_if_no_space ");
3022 case STATUSTYPE_TABLE:
3023 DMEMIT("%s %s %lu %llu ",
3024 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3025 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3026 (unsigned long)pool->sectors_per_block,
3027 (unsigned long long)pt->low_water_blocks);
3028 emit_flags(&pt->requested_pf, result, sz, maxlen);
3037 static int pool_iterate_devices(struct dm_target *ti,
3038 iterate_devices_callout_fn fn, void *data)
3040 struct pool_c *pt = ti->private;
3042 return fn(ti, pt->data_dev, 0, ti->len, data);
3045 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3046 struct bio_vec *biovec, int max_size)
3048 struct pool_c *pt = ti->private;
3049 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3051 if (!q->merge_bvec_fn)
3054 bvm->bi_bdev = pt->data_dev->bdev;
3056 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3059 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
3061 struct pool *pool = pt->pool;
3062 struct queue_limits *data_limits;
3064 limits->max_discard_sectors = pool->sectors_per_block;
3067 * discard_granularity is just a hint, and not enforced.
3069 if (pt->adjusted_pf.discard_passdown) {
3070 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
3071 limits->discard_granularity = data_limits->discard_granularity;
3073 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
3076 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3078 struct pool_c *pt = ti->private;
3079 struct pool *pool = pt->pool;
3080 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3083 * If the system-determined stacked limits are compatible with the
3084 * pool's blocksize (io_opt is a factor) do not override them.
3086 if (io_opt_sectors < pool->sectors_per_block ||
3087 do_div(io_opt_sectors, pool->sectors_per_block)) {
3088 blk_limits_io_min(limits, 0);
3089 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3093 * pt->adjusted_pf is a staging area for the actual features to use.
3094 * They get transferred to the live pool in bind_control_target()
3095 * called from pool_preresume().
3097 if (!pt->adjusted_pf.discard_enabled) {
3099 * Must explicitly disallow stacking discard limits otherwise the
3100 * block layer will stack them if pool's data device has support.
3101 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3102 * user to see that, so make sure to set all discard limits to 0.
3104 limits->discard_granularity = 0;
3108 disable_passdown_if_not_supported(pt);
3110 set_discard_limits(pt, limits);
3113 static struct target_type pool_target = {
3114 .name = "thin-pool",
3115 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3116 DM_TARGET_IMMUTABLE,
3117 .version = {1, 12, 0},
3118 .module = THIS_MODULE,
3122 .postsuspend = pool_postsuspend,
3123 .preresume = pool_preresume,
3124 .resume = pool_resume,
3125 .message = pool_message,
3126 .status = pool_status,
3127 .merge = pool_merge,
3128 .iterate_devices = pool_iterate_devices,
3129 .io_hints = pool_io_hints,
3132 /*----------------------------------------------------------------
3133 * Thin target methods
3134 *--------------------------------------------------------------*/
3135 static void thin_get(struct thin_c *tc)
3137 atomic_inc(&tc->refcount);
3140 static void thin_put(struct thin_c *tc)
3142 if (atomic_dec_and_test(&tc->refcount))
3143 complete(&tc->can_destroy);
3146 static void thin_dtr(struct dm_target *ti)
3148 struct thin_c *tc = ti->private;
3149 unsigned long flags;
3152 wait_for_completion(&tc->can_destroy);
3154 spin_lock_irqsave(&tc->pool->lock, flags);
3155 list_del_rcu(&tc->list);
3156 spin_unlock_irqrestore(&tc->pool->lock, flags);
3159 mutex_lock(&dm_thin_pool_table.mutex);
3161 __pool_dec(tc->pool);
3162 dm_pool_close_thin_device(tc->td);
3163 dm_put_device(ti, tc->pool_dev);
3165 dm_put_device(ti, tc->origin_dev);
3168 mutex_unlock(&dm_thin_pool_table.mutex);
3172 * Thin target parameters:
3174 * <pool_dev> <dev_id> [origin_dev]
3176 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3177 * dev_id: the internal device identifier
3178 * origin_dev: a device external to the pool that should act as the origin
3180 * If the pool device has discards disabled, they get disabled for the thin
3183 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3187 struct dm_dev *pool_dev, *origin_dev;
3188 struct mapped_device *pool_md;
3189 unsigned long flags;
3191 mutex_lock(&dm_thin_pool_table.mutex);
3193 if (argc != 2 && argc != 3) {
3194 ti->error = "Invalid argument count";
3199 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3201 ti->error = "Out of memory";
3205 spin_lock_init(&tc->lock);
3206 bio_list_init(&tc->deferred_bio_list);
3207 bio_list_init(&tc->retry_on_resume_list);
3208 tc->sort_bio_list = RB_ROOT;
3211 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
3213 ti->error = "Error opening origin device";
3214 goto bad_origin_dev;
3216 tc->origin_dev = origin_dev;
3219 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
3221 ti->error = "Error opening pool device";
3224 tc->pool_dev = pool_dev;
3226 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
3227 ti->error = "Invalid device id";
3232 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
3234 ti->error = "Couldn't get pool mapped device";
3239 tc->pool = __pool_table_lookup(pool_md);
3241 ti->error = "Couldn't find pool object";
3243 goto bad_pool_lookup;
3245 __pool_inc(tc->pool);
3247 if (get_pool_mode(tc->pool) == PM_FAIL) {
3248 ti->error = "Couldn't open thin device, Pool is in fail mode";
3253 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
3255 ti->error = "Couldn't open thin internal device";
3259 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
3261 goto bad_target_max_io_len;
3263 ti->num_flush_bios = 1;
3264 ti->flush_supported = true;
3265 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
3267 /* In case the pool supports discards, pass them on. */
3268 ti->discard_zeroes_data_unsupported = true;
3269 if (tc->pool->pf.discard_enabled) {
3270 ti->discards_supported = true;
3271 ti->num_discard_bios = 1;
3272 /* Discard bios must be split on a block boundary */
3273 ti->split_discard_bios = true;
3278 mutex_unlock(&dm_thin_pool_table.mutex);
3280 atomic_set(&tc->refcount, 1);
3281 init_completion(&tc->can_destroy);
3283 spin_lock_irqsave(&tc->pool->lock, flags);
3284 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
3285 spin_unlock_irqrestore(&tc->pool->lock, flags);
3287 * This synchronize_rcu() call is needed here otherwise we risk a
3288 * wake_worker() call finding no bios to process (because the newly
3289 * added tc isn't yet visible). So this reduces latency since we
3290 * aren't then dependent on the periodic commit to wake_worker().
3296 bad_target_max_io_len:
3297 dm_pool_close_thin_device(tc->td);
3299 __pool_dec(tc->pool);
3303 dm_put_device(ti, tc->pool_dev);
3306 dm_put_device(ti, tc->origin_dev);
3310 mutex_unlock(&dm_thin_pool_table.mutex);
3315 static int thin_map(struct dm_target *ti, struct bio *bio)
3317 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
3319 return thin_bio_map(ti, bio);
3322 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
3324 unsigned long flags;
3325 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
3326 struct list_head work;
3327 struct dm_thin_new_mapping *m, *tmp;
3328 struct pool *pool = h->tc->pool;
3330 if (h->shared_read_entry) {
3331 INIT_LIST_HEAD(&work);
3332 dm_deferred_entry_dec(h->shared_read_entry, &work);
3334 spin_lock_irqsave(&pool->lock, flags);
3335 list_for_each_entry_safe(m, tmp, &work, list) {
3338 __maybe_add_mapping(m);
3340 spin_unlock_irqrestore(&pool->lock, flags);
3343 if (h->all_io_entry) {
3344 INIT_LIST_HEAD(&work);
3345 dm_deferred_entry_dec(h->all_io_entry, &work);
3346 if (!list_empty(&work)) {
3347 spin_lock_irqsave(&pool->lock, flags);
3348 list_for_each_entry_safe(m, tmp, &work, list)
3349 list_add_tail(&m->list, &pool->prepared_discards);
3350 spin_unlock_irqrestore(&pool->lock, flags);
3358 static void thin_presuspend(struct dm_target *ti)
3360 struct thin_c *tc = ti->private;
3362 if (dm_noflush_suspending(ti))
3363 noflush_work(tc, do_noflush_start);
3366 static void thin_postsuspend(struct dm_target *ti)
3368 struct thin_c *tc = ti->private;
3371 * The dm_noflush_suspending flag has been cleared by now, so
3372 * unfortunately we must always run this.
3374 noflush_work(tc, do_noflush_stop);
3378 * <nr mapped sectors> <highest mapped sector>
3380 static void thin_status(struct dm_target *ti, status_type_t type,
3381 unsigned status_flags, char *result, unsigned maxlen)
3385 dm_block_t mapped, highest;
3386 char buf[BDEVNAME_SIZE];
3387 struct thin_c *tc = ti->private;
3389 if (get_pool_mode(tc->pool) == PM_FAIL) {
3398 case STATUSTYPE_INFO:
3399 r = dm_thin_get_mapped_count(tc->td, &mapped);
3401 DMERR("dm_thin_get_mapped_count returned %d", r);
3405 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
3407 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
3411 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
3413 DMEMIT("%llu", ((highest + 1) *
3414 tc->pool->sectors_per_block) - 1);
3419 case STATUSTYPE_TABLE:
3421 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
3422 (unsigned long) tc->dev_id);
3424 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
3435 static int thin_iterate_devices(struct dm_target *ti,
3436 iterate_devices_callout_fn fn, void *data)
3439 struct thin_c *tc = ti->private;
3440 struct pool *pool = tc->pool;
3443 * We can't call dm_pool_get_data_dev_size() since that blocks. So
3444 * we follow a more convoluted path through to the pool's target.
3447 return 0; /* nothing is bound */
3449 blocks = pool->ti->len;
3450 (void) sector_div(blocks, pool->sectors_per_block);
3452 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
3457 static struct target_type thin_target = {
3459 .version = {1, 12, 0},
3460 .module = THIS_MODULE,
3464 .end_io = thin_endio,
3465 .presuspend = thin_presuspend,
3466 .postsuspend = thin_postsuspend,
3467 .status = thin_status,
3468 .iterate_devices = thin_iterate_devices,
3471 /*----------------------------------------------------------------*/
3473 static int __init dm_thin_init(void)
3479 r = dm_register_target(&thin_target);
3483 r = dm_register_target(&pool_target);
3485 goto bad_pool_target;
3489 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
3490 if (!_new_mapping_cache)
3491 goto bad_new_mapping_cache;
3495 bad_new_mapping_cache:
3496 dm_unregister_target(&pool_target);
3498 dm_unregister_target(&thin_target);
3503 static void dm_thin_exit(void)
3505 dm_unregister_target(&thin_target);
3506 dm_unregister_target(&pool_target);
3508 kmem_cache_destroy(_new_mapping_cache);
3511 module_init(dm_thin_init);
3512 module_exit(dm_thin_exit);
3514 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
3515 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
3517 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
3518 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3519 MODULE_LICENSE("GPL");