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_with_code(struct pool *pool,
314 struct dm_bio_prison_cell *cell, int error_code)
316 dm_cell_error(pool->prison, cell, error_code);
317 dm_bio_prison_free_cell(pool->prison, cell);
320 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
322 cell_error_with_code(pool, cell, -EIO);
325 /*----------------------------------------------------------------*/
328 * A global list of pools that uses a struct mapped_device as a key.
330 static struct dm_thin_pool_table {
332 struct list_head pools;
333 } dm_thin_pool_table;
335 static void pool_table_init(void)
337 mutex_init(&dm_thin_pool_table.mutex);
338 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
341 static void __pool_table_insert(struct pool *pool)
343 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
344 list_add(&pool->list, &dm_thin_pool_table.pools);
347 static void __pool_table_remove(struct pool *pool)
349 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
350 list_del(&pool->list);
353 static struct pool *__pool_table_lookup(struct mapped_device *md)
355 struct pool *pool = NULL, *tmp;
357 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
359 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
360 if (tmp->pool_md == md) {
369 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
371 struct pool *pool = NULL, *tmp;
373 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
375 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
376 if (tmp->md_dev == md_dev) {
385 /*----------------------------------------------------------------*/
387 struct dm_thin_endio_hook {
389 struct dm_deferred_entry *shared_read_entry;
390 struct dm_deferred_entry *all_io_entry;
391 struct dm_thin_new_mapping *overwrite_mapping;
392 struct rb_node rb_node;
395 static void requeue_bio_list(struct thin_c *tc, struct bio_list *master)
398 struct bio_list bios;
401 bio_list_init(&bios);
403 spin_lock_irqsave(&tc->lock, flags);
404 bio_list_merge(&bios, master);
405 bio_list_init(master);
406 spin_unlock_irqrestore(&tc->lock, flags);
408 while ((bio = bio_list_pop(&bios)))
409 bio_endio(bio, DM_ENDIO_REQUEUE);
412 static void requeue_io(struct thin_c *tc)
414 requeue_bio_list(tc, &tc->deferred_bio_list);
415 requeue_bio_list(tc, &tc->retry_on_resume_list);
418 static void error_thin_retry_list(struct thin_c *tc)
422 struct bio_list bios;
424 bio_list_init(&bios);
426 spin_lock_irqsave(&tc->lock, flags);
427 bio_list_merge(&bios, &tc->retry_on_resume_list);
428 bio_list_init(&tc->retry_on_resume_list);
429 spin_unlock_irqrestore(&tc->lock, flags);
431 while ((bio = bio_list_pop(&bios)))
435 static void error_retry_list(struct pool *pool)
440 list_for_each_entry_rcu(tc, &pool->active_thins, list)
441 error_thin_retry_list(tc);
446 * This section of code contains the logic for processing a thin device's IO.
447 * Much of the code depends on pool object resources (lists, workqueues, etc)
448 * but most is exclusively called from the thin target rather than the thin-pool
452 static bool block_size_is_power_of_two(struct pool *pool)
454 return pool->sectors_per_block_shift >= 0;
457 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
459 struct pool *pool = tc->pool;
460 sector_t block_nr = bio->bi_iter.bi_sector;
462 if (block_size_is_power_of_two(pool))
463 block_nr >>= pool->sectors_per_block_shift;
465 (void) sector_div(block_nr, pool->sectors_per_block);
470 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
472 struct pool *pool = tc->pool;
473 sector_t bi_sector = bio->bi_iter.bi_sector;
475 bio->bi_bdev = tc->pool_dev->bdev;
476 if (block_size_is_power_of_two(pool))
477 bio->bi_iter.bi_sector =
478 (block << pool->sectors_per_block_shift) |
479 (bi_sector & (pool->sectors_per_block - 1));
481 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
482 sector_div(bi_sector, pool->sectors_per_block);
485 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
487 bio->bi_bdev = tc->origin_dev->bdev;
490 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
492 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
493 dm_thin_changed_this_transaction(tc->td);
496 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
498 struct dm_thin_endio_hook *h;
500 if (bio->bi_rw & REQ_DISCARD)
503 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
504 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
507 static void issue(struct thin_c *tc, struct bio *bio)
509 struct pool *pool = tc->pool;
512 if (!bio_triggers_commit(tc, bio)) {
513 generic_make_request(bio);
518 * Complete bio with an error if earlier I/O caused changes to
519 * the metadata that can't be committed e.g, due to I/O errors
520 * on the metadata device.
522 if (dm_thin_aborted_changes(tc->td)) {
528 * Batch together any bios that trigger commits and then issue a
529 * single commit for them in process_deferred_bios().
531 spin_lock_irqsave(&pool->lock, flags);
532 bio_list_add(&pool->deferred_flush_bios, bio);
533 spin_unlock_irqrestore(&pool->lock, flags);
536 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
538 remap_to_origin(tc, bio);
542 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
545 remap(tc, bio, block);
549 /*----------------------------------------------------------------*/
552 * Bio endio functions.
554 struct dm_thin_new_mapping {
555 struct list_head list;
560 bool definitely_not_shared:1;
564 dm_block_t virt_block;
565 dm_block_t data_block;
566 struct dm_bio_prison_cell *cell, *cell2;
569 * If the bio covers the whole area of a block then we can avoid
570 * zeroing or copying. Instead this bio is hooked. The bio will
571 * still be in the cell, so care has to be taken to avoid issuing
575 bio_end_io_t *saved_bi_end_io;
578 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
580 struct pool *pool = m->tc->pool;
582 if (m->quiesced && m->prepared) {
583 list_add_tail(&m->list, &pool->prepared_mappings);
588 static void copy_complete(int read_err, unsigned long write_err, void *context)
591 struct dm_thin_new_mapping *m = context;
592 struct pool *pool = m->tc->pool;
594 m->err = read_err || write_err ? -EIO : 0;
596 spin_lock_irqsave(&pool->lock, flags);
598 __maybe_add_mapping(m);
599 spin_unlock_irqrestore(&pool->lock, flags);
602 static void overwrite_endio(struct bio *bio, int err)
605 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
606 struct dm_thin_new_mapping *m = h->overwrite_mapping;
607 struct pool *pool = m->tc->pool;
611 spin_lock_irqsave(&pool->lock, flags);
613 __maybe_add_mapping(m);
614 spin_unlock_irqrestore(&pool->lock, flags);
617 /*----------------------------------------------------------------*/
624 * Prepared mapping jobs.
628 * This sends the bios in the cell back to the deferred_bios list.
630 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
632 struct pool *pool = tc->pool;
635 spin_lock_irqsave(&tc->lock, flags);
636 cell_release(pool, cell, &tc->deferred_bio_list);
637 spin_unlock_irqrestore(&tc->lock, flags);
643 * Same as cell_defer above, except it omits the original holder of the cell.
645 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
647 struct pool *pool = tc->pool;
650 spin_lock_irqsave(&tc->lock, flags);
651 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
652 spin_unlock_irqrestore(&tc->lock, flags);
657 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
660 m->bio->bi_end_io = m->saved_bi_end_io;
661 atomic_inc(&m->bio->bi_remaining);
663 cell_error(m->tc->pool, m->cell);
665 mempool_free(m, m->tc->pool->mapping_pool);
668 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
670 struct thin_c *tc = m->tc;
671 struct pool *pool = tc->pool;
677 bio->bi_end_io = m->saved_bi_end_io;
678 atomic_inc(&bio->bi_remaining);
682 cell_error(pool, m->cell);
687 * Commit the prepared block into the mapping btree.
688 * Any I/O for this block arriving after this point will get
689 * remapped to it directly.
691 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
693 metadata_operation_failed(pool, "dm_thin_insert_block", r);
694 cell_error(pool, m->cell);
699 * Release any bios held while the block was being provisioned.
700 * If we are processing a write bio that completely covers the block,
701 * we already processed it so can ignore it now when processing
702 * the bios in the cell.
705 cell_defer_no_holder(tc, m->cell);
708 cell_defer(tc, m->cell);
712 mempool_free(m, pool->mapping_pool);
715 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
717 struct thin_c *tc = m->tc;
719 bio_io_error(m->bio);
720 cell_defer_no_holder(tc, m->cell);
721 cell_defer_no_holder(tc, m->cell2);
722 mempool_free(m, tc->pool->mapping_pool);
725 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
727 struct thin_c *tc = m->tc;
729 inc_all_io_entry(tc->pool, m->bio);
730 cell_defer_no_holder(tc, m->cell);
731 cell_defer_no_holder(tc, m->cell2);
734 if (m->definitely_not_shared)
735 remap_and_issue(tc, m->bio, m->data_block);
738 if (dm_pool_block_is_used(tc->pool->pmd, m->data_block, &used) || used)
739 bio_endio(m->bio, 0);
741 remap_and_issue(tc, m->bio, m->data_block);
744 bio_endio(m->bio, 0);
746 mempool_free(m, tc->pool->mapping_pool);
749 static void process_prepared_discard(struct dm_thin_new_mapping *m)
752 struct thin_c *tc = m->tc;
754 r = dm_thin_remove_block(tc->td, m->virt_block);
756 DMERR_LIMIT("dm_thin_remove_block() failed");
758 process_prepared_discard_passdown(m);
761 static void process_prepared(struct pool *pool, struct list_head *head,
762 process_mapping_fn *fn)
765 struct list_head maps;
766 struct dm_thin_new_mapping *m, *tmp;
768 INIT_LIST_HEAD(&maps);
769 spin_lock_irqsave(&pool->lock, flags);
770 list_splice_init(head, &maps);
771 spin_unlock_irqrestore(&pool->lock, flags);
773 list_for_each_entry_safe(m, tmp, &maps, list)
780 static int io_overlaps_block(struct pool *pool, struct bio *bio)
782 return bio->bi_iter.bi_size ==
783 (pool->sectors_per_block << SECTOR_SHIFT);
786 static int io_overwrites_block(struct pool *pool, struct bio *bio)
788 return (bio_data_dir(bio) == WRITE) &&
789 io_overlaps_block(pool, bio);
792 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
795 *save = bio->bi_end_io;
799 static int ensure_next_mapping(struct pool *pool)
801 if (pool->next_mapping)
804 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
806 return pool->next_mapping ? 0 : -ENOMEM;
809 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
811 struct dm_thin_new_mapping *m = pool->next_mapping;
813 BUG_ON(!pool->next_mapping);
815 memset(m, 0, sizeof(struct dm_thin_new_mapping));
816 INIT_LIST_HEAD(&m->list);
819 pool->next_mapping = NULL;
824 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
825 struct dm_dev *origin, dm_block_t data_origin,
826 dm_block_t data_dest,
827 struct dm_bio_prison_cell *cell, struct bio *bio)
830 struct pool *pool = tc->pool;
831 struct dm_thin_new_mapping *m = get_next_mapping(pool);
834 m->virt_block = virt_block;
835 m->data_block = data_dest;
838 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
842 * IO to pool_dev remaps to the pool target's data_dev.
844 * If the whole block of data is being overwritten, we can issue the
845 * bio immediately. Otherwise we use kcopyd to clone the data first.
847 if (io_overwrites_block(pool, bio)) {
848 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
850 h->overwrite_mapping = m;
852 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
853 inc_all_io_entry(pool, bio);
854 remap_and_issue(tc, bio, data_dest);
856 struct dm_io_region from, to;
858 from.bdev = origin->bdev;
859 from.sector = data_origin * pool->sectors_per_block;
860 from.count = pool->sectors_per_block;
862 to.bdev = tc->pool_dev->bdev;
863 to.sector = data_dest * pool->sectors_per_block;
864 to.count = pool->sectors_per_block;
866 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
867 0, copy_complete, m);
869 mempool_free(m, pool->mapping_pool);
870 DMERR_LIMIT("dm_kcopyd_copy() failed");
871 cell_error(pool, cell);
876 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
877 dm_block_t data_origin, dm_block_t data_dest,
878 struct dm_bio_prison_cell *cell, struct bio *bio)
880 schedule_copy(tc, virt_block, tc->pool_dev,
881 data_origin, data_dest, cell, bio);
884 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
885 dm_block_t data_dest,
886 struct dm_bio_prison_cell *cell, struct bio *bio)
888 schedule_copy(tc, virt_block, tc->origin_dev,
889 virt_block, data_dest, cell, bio);
892 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
893 dm_block_t data_block, struct dm_bio_prison_cell *cell,
896 struct pool *pool = tc->pool;
897 struct dm_thin_new_mapping *m = get_next_mapping(pool);
902 m->virt_block = virt_block;
903 m->data_block = data_block;
907 * If the whole block of data is being overwritten or we are not
908 * zeroing pre-existing data, we can issue the bio immediately.
909 * Otherwise we use kcopyd to zero the data first.
911 if (!pool->pf.zero_new_blocks)
912 process_prepared_mapping(m);
914 else if (io_overwrites_block(pool, bio)) {
915 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
917 h->overwrite_mapping = m;
919 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
920 inc_all_io_entry(pool, bio);
921 remap_and_issue(tc, bio, data_block);
924 struct dm_io_region to;
926 to.bdev = tc->pool_dev->bdev;
927 to.sector = data_block * pool->sectors_per_block;
928 to.count = pool->sectors_per_block;
930 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
932 mempool_free(m, pool->mapping_pool);
933 DMERR_LIMIT("dm_kcopyd_zero() failed");
934 cell_error(pool, cell);
940 * A non-zero return indicates read_only or fail_io mode.
941 * Many callers don't care about the return value.
943 static int commit(struct pool *pool)
947 if (get_pool_mode(pool) >= PM_READ_ONLY)
950 r = dm_pool_commit_metadata(pool->pmd);
952 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
957 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
961 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
962 DMWARN("%s: reached low water mark for data device: sending event.",
963 dm_device_name(pool->pool_md));
964 spin_lock_irqsave(&pool->lock, flags);
965 pool->low_water_triggered = true;
966 spin_unlock_irqrestore(&pool->lock, flags);
967 dm_table_event(pool->ti->table);
971 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
973 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
976 dm_block_t free_blocks;
977 struct pool *pool = tc->pool;
979 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
982 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
984 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
988 check_low_water_mark(pool, free_blocks);
992 * Try to commit to see if that will free up some
999 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1001 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1006 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1011 r = dm_pool_alloc_data_block(pool->pmd, result);
1013 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1021 * If we have run out of space, queue bios until the device is
1022 * resumed, presumably after having been reloaded with more space.
1024 static void retry_on_resume(struct bio *bio)
1026 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1027 struct thin_c *tc = h->tc;
1028 unsigned long flags;
1030 spin_lock_irqsave(&tc->lock, flags);
1031 bio_list_add(&tc->retry_on_resume_list, bio);
1032 spin_unlock_irqrestore(&tc->lock, flags);
1035 static int should_error_unserviceable_bio(struct pool *pool)
1037 enum pool_mode m = get_pool_mode(pool);
1041 /* Shouldn't get here */
1042 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1045 case PM_OUT_OF_DATA_SPACE:
1046 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1052 /* Shouldn't get here */
1053 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1058 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1060 int error = should_error_unserviceable_bio(pool);
1063 bio_endio(bio, error);
1065 retry_on_resume(bio);
1068 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1071 struct bio_list bios;
1074 error = should_error_unserviceable_bio(pool);
1076 cell_error_with_code(pool, cell, error);
1080 bio_list_init(&bios);
1081 cell_release(pool, cell, &bios);
1083 error = should_error_unserviceable_bio(pool);
1085 while ((bio = bio_list_pop(&bios)))
1086 bio_endio(bio, error);
1088 while ((bio = bio_list_pop(&bios)))
1089 retry_on_resume(bio);
1092 static void process_discard(struct thin_c *tc, struct bio *bio)
1095 unsigned long flags;
1096 struct pool *pool = tc->pool;
1097 struct dm_bio_prison_cell *cell, *cell2;
1098 struct dm_cell_key key, key2;
1099 dm_block_t block = get_bio_block(tc, bio);
1100 struct dm_thin_lookup_result lookup_result;
1101 struct dm_thin_new_mapping *m;
1103 build_virtual_key(tc->td, block, &key);
1104 if (bio_detain(tc->pool, &key, bio, &cell))
1107 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1111 * Check nobody is fiddling with this pool block. This can
1112 * happen if someone's in the process of breaking sharing
1115 build_data_key(tc->td, lookup_result.block, &key2);
1116 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1117 cell_defer_no_holder(tc, cell);
1121 if (io_overlaps_block(pool, bio)) {
1123 * IO may still be going to the destination block. We must
1124 * quiesce before we can do the removal.
1126 m = get_next_mapping(pool);
1128 m->pass_discard = pool->pf.discard_passdown;
1129 m->definitely_not_shared = !lookup_result.shared;
1130 m->virt_block = block;
1131 m->data_block = lookup_result.block;
1136 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
1137 spin_lock_irqsave(&pool->lock, flags);
1138 list_add_tail(&m->list, &pool->prepared_discards);
1139 spin_unlock_irqrestore(&pool->lock, flags);
1143 inc_all_io_entry(pool, bio);
1144 cell_defer_no_holder(tc, cell);
1145 cell_defer_no_holder(tc, cell2);
1148 * The DM core makes sure that the discard doesn't span
1149 * a block boundary. So we submit the discard of a
1150 * partial block appropriately.
1152 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1153 remap_and_issue(tc, bio, lookup_result.block);
1161 * It isn't provisioned, just forget it.
1163 cell_defer_no_holder(tc, cell);
1168 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1170 cell_defer_no_holder(tc, cell);
1176 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1177 struct dm_cell_key *key,
1178 struct dm_thin_lookup_result *lookup_result,
1179 struct dm_bio_prison_cell *cell)
1182 dm_block_t data_block;
1183 struct pool *pool = tc->pool;
1185 r = alloc_data_block(tc, &data_block);
1188 schedule_internal_copy(tc, block, lookup_result->block,
1189 data_block, cell, bio);
1193 retry_bios_on_resume(pool, cell);
1197 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1199 cell_error(pool, cell);
1204 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1206 struct dm_thin_lookup_result *lookup_result)
1208 struct dm_bio_prison_cell *cell;
1209 struct pool *pool = tc->pool;
1210 struct dm_cell_key key;
1213 * If cell is already occupied, then sharing is already in the process
1214 * of being broken so we have nothing further to do here.
1216 build_data_key(tc->td, lookup_result->block, &key);
1217 if (bio_detain(pool, &key, bio, &cell))
1220 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size)
1221 break_sharing(tc, bio, block, &key, lookup_result, cell);
1223 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1225 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1226 inc_all_io_entry(pool, bio);
1227 cell_defer_no_holder(tc, cell);
1229 remap_and_issue(tc, bio, lookup_result->block);
1233 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1234 struct dm_bio_prison_cell *cell)
1237 dm_block_t data_block;
1238 struct pool *pool = tc->pool;
1241 * Remap empty bios (flushes) immediately, without provisioning.
1243 if (!bio->bi_iter.bi_size) {
1244 inc_all_io_entry(pool, bio);
1245 cell_defer_no_holder(tc, cell);
1247 remap_and_issue(tc, bio, 0);
1252 * Fill read bios with zeroes and complete them immediately.
1254 if (bio_data_dir(bio) == READ) {
1256 cell_defer_no_holder(tc, cell);
1261 r = alloc_data_block(tc, &data_block);
1265 schedule_external_copy(tc, block, data_block, cell, bio);
1267 schedule_zero(tc, block, data_block, cell, bio);
1271 retry_bios_on_resume(pool, cell);
1275 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1277 cell_error(pool, cell);
1282 static void process_bio(struct thin_c *tc, struct bio *bio)
1285 struct pool *pool = tc->pool;
1286 dm_block_t block = get_bio_block(tc, bio);
1287 struct dm_bio_prison_cell *cell;
1288 struct dm_cell_key key;
1289 struct dm_thin_lookup_result lookup_result;
1292 * If cell is already occupied, then the block is already
1293 * being provisioned so we have nothing further to do here.
1295 build_virtual_key(tc->td, block, &key);
1296 if (bio_detain(pool, &key, bio, &cell))
1299 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1302 if (lookup_result.shared) {
1303 process_shared_bio(tc, bio, block, &lookup_result);
1304 cell_defer_no_holder(tc, cell); /* FIXME: pass this cell into process_shared? */
1306 inc_all_io_entry(pool, bio);
1307 cell_defer_no_holder(tc, cell);
1309 remap_and_issue(tc, bio, lookup_result.block);
1314 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1315 inc_all_io_entry(pool, bio);
1316 cell_defer_no_holder(tc, cell);
1318 remap_to_origin_and_issue(tc, bio);
1320 provision_block(tc, bio, block, cell);
1324 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1326 cell_defer_no_holder(tc, cell);
1332 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1335 int rw = bio_data_dir(bio);
1336 dm_block_t block = get_bio_block(tc, bio);
1337 struct dm_thin_lookup_result lookup_result;
1339 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1342 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size)
1343 handle_unserviceable_bio(tc->pool, bio);
1345 inc_all_io_entry(tc->pool, bio);
1346 remap_and_issue(tc, bio, lookup_result.block);
1352 handle_unserviceable_bio(tc->pool, bio);
1356 if (tc->origin_dev) {
1357 inc_all_io_entry(tc->pool, bio);
1358 remap_to_origin_and_issue(tc, bio);
1367 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1374 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1379 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1385 * FIXME: should we also commit due to size of transaction, measured in
1388 static int need_commit_due_to_time(struct pool *pool)
1390 return jiffies < pool->last_commit_jiffies ||
1391 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1394 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1395 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1397 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1399 struct rb_node **rbp, *parent;
1400 struct dm_thin_endio_hook *pbd;
1401 sector_t bi_sector = bio->bi_iter.bi_sector;
1403 rbp = &tc->sort_bio_list.rb_node;
1407 pbd = thin_pbd(parent);
1409 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1410 rbp = &(*rbp)->rb_left;
1412 rbp = &(*rbp)->rb_right;
1415 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1416 rb_link_node(&pbd->rb_node, parent, rbp);
1417 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1420 static void __extract_sorted_bios(struct thin_c *tc)
1422 struct rb_node *node;
1423 struct dm_thin_endio_hook *pbd;
1426 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1427 pbd = thin_pbd(node);
1428 bio = thin_bio(pbd);
1430 bio_list_add(&tc->deferred_bio_list, bio);
1431 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1434 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1437 static void __sort_thin_deferred_bios(struct thin_c *tc)
1440 struct bio_list bios;
1442 bio_list_init(&bios);
1443 bio_list_merge(&bios, &tc->deferred_bio_list);
1444 bio_list_init(&tc->deferred_bio_list);
1446 /* Sort deferred_bio_list using rb-tree */
1447 while ((bio = bio_list_pop(&bios)))
1448 __thin_bio_rb_add(tc, bio);
1451 * Transfer the sorted bios in sort_bio_list back to
1452 * deferred_bio_list to allow lockless submission of
1455 __extract_sorted_bios(tc);
1458 static void process_thin_deferred_bios(struct thin_c *tc)
1460 struct pool *pool = tc->pool;
1461 unsigned long flags;
1463 struct bio_list bios;
1464 struct blk_plug plug;
1466 if (tc->requeue_mode) {
1467 requeue_bio_list(tc, &tc->deferred_bio_list);
1471 bio_list_init(&bios);
1473 spin_lock_irqsave(&tc->lock, flags);
1475 if (bio_list_empty(&tc->deferred_bio_list)) {
1476 spin_unlock_irqrestore(&tc->lock, flags);
1480 __sort_thin_deferred_bios(tc);
1482 bio_list_merge(&bios, &tc->deferred_bio_list);
1483 bio_list_init(&tc->deferred_bio_list);
1485 spin_unlock_irqrestore(&tc->lock, flags);
1487 blk_start_plug(&plug);
1488 while ((bio = bio_list_pop(&bios))) {
1490 * If we've got no free new_mapping structs, and processing
1491 * this bio might require one, we pause until there are some
1492 * prepared mappings to process.
1494 if (ensure_next_mapping(pool)) {
1495 spin_lock_irqsave(&tc->lock, flags);
1496 bio_list_add(&tc->deferred_bio_list, bio);
1497 bio_list_merge(&tc->deferred_bio_list, &bios);
1498 spin_unlock_irqrestore(&tc->lock, flags);
1502 if (bio->bi_rw & REQ_DISCARD)
1503 pool->process_discard(tc, bio);
1505 pool->process_bio(tc, bio);
1507 blk_finish_plug(&plug);
1510 static void thin_get(struct thin_c *tc);
1511 static void thin_put(struct thin_c *tc);
1514 * We can't hold rcu_read_lock() around code that can block. So we
1515 * find a thin with the rcu lock held; bump a refcount; then drop
1518 static struct thin_c *get_first_thin(struct pool *pool)
1520 struct thin_c *tc = NULL;
1523 if (!list_empty(&pool->active_thins)) {
1524 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
1532 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
1534 struct thin_c *old_tc = tc;
1537 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
1549 static void process_deferred_bios(struct pool *pool)
1551 unsigned long flags;
1553 struct bio_list bios;
1556 tc = get_first_thin(pool);
1558 process_thin_deferred_bios(tc);
1559 tc = get_next_thin(pool, tc);
1563 * If there are any deferred flush bios, we must commit
1564 * the metadata before issuing them.
1566 bio_list_init(&bios);
1567 spin_lock_irqsave(&pool->lock, flags);
1568 bio_list_merge(&bios, &pool->deferred_flush_bios);
1569 bio_list_init(&pool->deferred_flush_bios);
1570 spin_unlock_irqrestore(&pool->lock, flags);
1572 if (bio_list_empty(&bios) &&
1573 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
1577 while ((bio = bio_list_pop(&bios)))
1581 pool->last_commit_jiffies = jiffies;
1583 while ((bio = bio_list_pop(&bios)))
1584 generic_make_request(bio);
1587 static void do_worker(struct work_struct *ws)
1589 struct pool *pool = container_of(ws, struct pool, worker);
1591 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1592 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1593 process_deferred_bios(pool);
1597 * We want to commit periodically so that not too much
1598 * unwritten data builds up.
1600 static void do_waker(struct work_struct *ws)
1602 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1604 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1608 * We're holding onto IO to allow userland time to react. After the
1609 * timeout either the pool will have been resized (and thus back in
1610 * PM_WRITE mode), or we degrade to PM_READ_ONLY and start erroring IO.
1612 static void do_no_space_timeout(struct work_struct *ws)
1614 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
1617 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space)
1618 set_pool_mode(pool, PM_READ_ONLY);
1621 /*----------------------------------------------------------------*/
1624 struct work_struct worker;
1625 struct completion complete;
1628 static struct pool_work *to_pool_work(struct work_struct *ws)
1630 return container_of(ws, struct pool_work, worker);
1633 static void pool_work_complete(struct pool_work *pw)
1635 complete(&pw->complete);
1638 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
1639 void (*fn)(struct work_struct *))
1641 INIT_WORK_ONSTACK(&pw->worker, fn);
1642 init_completion(&pw->complete);
1643 queue_work(pool->wq, &pw->worker);
1644 wait_for_completion(&pw->complete);
1647 /*----------------------------------------------------------------*/
1649 struct noflush_work {
1650 struct pool_work pw;
1654 static struct noflush_work *to_noflush(struct work_struct *ws)
1656 return container_of(to_pool_work(ws), struct noflush_work, pw);
1659 static void do_noflush_start(struct work_struct *ws)
1661 struct noflush_work *w = to_noflush(ws);
1662 w->tc->requeue_mode = true;
1664 pool_work_complete(&w->pw);
1667 static void do_noflush_stop(struct work_struct *ws)
1669 struct noflush_work *w = to_noflush(ws);
1670 w->tc->requeue_mode = false;
1671 pool_work_complete(&w->pw);
1674 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
1676 struct noflush_work w;
1679 pool_work_wait(&w.pw, tc->pool, fn);
1682 /*----------------------------------------------------------------*/
1684 static enum pool_mode get_pool_mode(struct pool *pool)
1686 return pool->pf.mode;
1689 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
1691 dm_table_event(pool->ti->table);
1692 DMINFO("%s: switching pool to %s mode",
1693 dm_device_name(pool->pool_md), new_mode);
1696 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
1698 struct pool_c *pt = pool->ti->private;
1699 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
1700 enum pool_mode old_mode = get_pool_mode(pool);
1701 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
1704 * Never allow the pool to transition to PM_WRITE mode if user
1705 * intervention is required to verify metadata and data consistency.
1707 if (new_mode == PM_WRITE && needs_check) {
1708 DMERR("%s: unable to switch pool to write mode until repaired.",
1709 dm_device_name(pool->pool_md));
1710 if (old_mode != new_mode)
1711 new_mode = old_mode;
1713 new_mode = PM_READ_ONLY;
1716 * If we were in PM_FAIL mode, rollback of metadata failed. We're
1717 * not going to recover without a thin_repair. So we never let the
1718 * pool move out of the old mode.
1720 if (old_mode == PM_FAIL)
1721 new_mode = old_mode;
1725 if (old_mode != new_mode)
1726 notify_of_pool_mode_change(pool, "failure");
1727 dm_pool_metadata_read_only(pool->pmd);
1728 pool->process_bio = process_bio_fail;
1729 pool->process_discard = process_bio_fail;
1730 pool->process_prepared_mapping = process_prepared_mapping_fail;
1731 pool->process_prepared_discard = process_prepared_discard_fail;
1733 error_retry_list(pool);
1737 if (old_mode != new_mode)
1738 notify_of_pool_mode_change(pool, "read-only");
1739 dm_pool_metadata_read_only(pool->pmd);
1740 pool->process_bio = process_bio_read_only;
1741 pool->process_discard = process_bio_success;
1742 pool->process_prepared_mapping = process_prepared_mapping_fail;
1743 pool->process_prepared_discard = process_prepared_discard_passdown;
1745 error_retry_list(pool);
1748 case PM_OUT_OF_DATA_SPACE:
1750 * Ideally we'd never hit this state; the low water mark
1751 * would trigger userland to extend the pool before we
1752 * completely run out of data space. However, many small
1753 * IOs to unprovisioned space can consume data space at an
1754 * alarming rate. Adjust your low water mark if you're
1755 * frequently seeing this mode.
1757 if (old_mode != new_mode)
1758 notify_of_pool_mode_change(pool, "out-of-data-space");
1759 pool->process_bio = process_bio_read_only;
1760 pool->process_discard = process_discard;
1761 pool->process_prepared_mapping = process_prepared_mapping;
1762 pool->process_prepared_discard = process_prepared_discard_passdown;
1764 if (!pool->pf.error_if_no_space && no_space_timeout)
1765 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
1769 if (old_mode != new_mode)
1770 notify_of_pool_mode_change(pool, "write");
1771 dm_pool_metadata_read_write(pool->pmd);
1772 pool->process_bio = process_bio;
1773 pool->process_discard = process_discard;
1774 pool->process_prepared_mapping = process_prepared_mapping;
1775 pool->process_prepared_discard = process_prepared_discard;
1779 pool->pf.mode = new_mode;
1781 * The pool mode may have changed, sync it so bind_control_target()
1782 * doesn't cause an unexpected mode transition on resume.
1784 pt->adjusted_pf.mode = new_mode;
1787 static void abort_transaction(struct pool *pool)
1789 const char *dev_name = dm_device_name(pool->pool_md);
1791 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1792 if (dm_pool_abort_metadata(pool->pmd)) {
1793 DMERR("%s: failed to abort metadata transaction", dev_name);
1794 set_pool_mode(pool, PM_FAIL);
1797 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
1798 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1799 set_pool_mode(pool, PM_FAIL);
1803 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
1805 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1806 dm_device_name(pool->pool_md), op, r);
1808 abort_transaction(pool);
1809 set_pool_mode(pool, PM_READ_ONLY);
1812 /*----------------------------------------------------------------*/
1815 * Mapping functions.
1819 * Called only while mapping a thin bio to hand it over to the workqueue.
1821 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1823 unsigned long flags;
1824 struct pool *pool = tc->pool;
1826 spin_lock_irqsave(&tc->lock, flags);
1827 bio_list_add(&tc->deferred_bio_list, bio);
1828 spin_unlock_irqrestore(&tc->lock, flags);
1833 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1835 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1838 h->shared_read_entry = NULL;
1839 h->all_io_entry = NULL;
1840 h->overwrite_mapping = NULL;
1844 * Non-blocking function called from the thin target's map function.
1846 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1849 struct thin_c *tc = ti->private;
1850 dm_block_t block = get_bio_block(tc, bio);
1851 struct dm_thin_device *td = tc->td;
1852 struct dm_thin_lookup_result result;
1853 struct dm_bio_prison_cell cell1, cell2;
1854 struct dm_bio_prison_cell *cell_result;
1855 struct dm_cell_key key;
1857 thin_hook_bio(tc, bio);
1859 if (tc->requeue_mode) {
1860 bio_endio(bio, DM_ENDIO_REQUEUE);
1861 return DM_MAPIO_SUBMITTED;
1864 if (get_pool_mode(tc->pool) == PM_FAIL) {
1866 return DM_MAPIO_SUBMITTED;
1869 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1870 thin_defer_bio(tc, bio);
1871 return DM_MAPIO_SUBMITTED;
1874 r = dm_thin_find_block(td, block, 0, &result);
1877 * Note that we defer readahead too.
1881 if (unlikely(result.shared)) {
1883 * We have a race condition here between the
1884 * result.shared value returned by the lookup and
1885 * snapshot creation, which may cause new
1888 * To avoid this always quiesce the origin before
1889 * taking the snap. You want to do this anyway to
1890 * ensure a consistent application view
1893 * More distant ancestors are irrelevant. The
1894 * shared flag will be set in their case.
1896 thin_defer_bio(tc, bio);
1897 return DM_MAPIO_SUBMITTED;
1900 build_virtual_key(tc->td, block, &key);
1901 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1, &cell_result))
1902 return DM_MAPIO_SUBMITTED;
1904 build_data_key(tc->td, result.block, &key);
1905 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2, &cell_result)) {
1906 cell_defer_no_holder_no_free(tc, &cell1);
1907 return DM_MAPIO_SUBMITTED;
1910 inc_all_io_entry(tc->pool, bio);
1911 cell_defer_no_holder_no_free(tc, &cell2);
1912 cell_defer_no_holder_no_free(tc, &cell1);
1914 remap(tc, bio, result.block);
1915 return DM_MAPIO_REMAPPED;
1918 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1920 * This block isn't provisioned, and we have no way
1923 handle_unserviceable_bio(tc->pool, bio);
1924 return DM_MAPIO_SUBMITTED;
1930 * In future, the failed dm_thin_find_block above could
1931 * provide the hint to load the metadata into cache.
1933 thin_defer_bio(tc, bio);
1934 return DM_MAPIO_SUBMITTED;
1938 * Must always call bio_io_error on failure.
1939 * dm_thin_find_block can fail with -EINVAL if the
1940 * pool is switched to fail-io mode.
1943 return DM_MAPIO_SUBMITTED;
1947 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1949 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1950 struct request_queue *q;
1952 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
1955 q = bdev_get_queue(pt->data_dev->bdev);
1956 return bdi_congested(&q->backing_dev_info, bdi_bits);
1959 static void requeue_bios(struct pool *pool)
1961 unsigned long flags;
1965 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
1966 spin_lock_irqsave(&tc->lock, flags);
1967 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
1968 bio_list_init(&tc->retry_on_resume_list);
1969 spin_unlock_irqrestore(&tc->lock, flags);
1974 /*----------------------------------------------------------------
1975 * Binding of control targets to a pool object
1976 *--------------------------------------------------------------*/
1977 static bool data_dev_supports_discard(struct pool_c *pt)
1979 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1981 return q && blk_queue_discard(q);
1984 static bool is_factor(sector_t block_size, uint32_t n)
1986 return !sector_div(block_size, n);
1990 * If discard_passdown was enabled verify that the data device
1991 * supports discards. Disable discard_passdown if not.
1993 static void disable_passdown_if_not_supported(struct pool_c *pt)
1995 struct pool *pool = pt->pool;
1996 struct block_device *data_bdev = pt->data_dev->bdev;
1997 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1998 sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1999 const char *reason = NULL;
2000 char buf[BDEVNAME_SIZE];
2002 if (!pt->adjusted_pf.discard_passdown)
2005 if (!data_dev_supports_discard(pt))
2006 reason = "discard unsupported";
2008 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2009 reason = "max discard sectors smaller than a block";
2011 else if (data_limits->discard_granularity > block_size)
2012 reason = "discard granularity larger than a block";
2014 else if (!is_factor(block_size, data_limits->discard_granularity))
2015 reason = "discard granularity not a factor of block size";
2018 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2019 pt->adjusted_pf.discard_passdown = false;
2023 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2025 struct pool_c *pt = ti->private;
2028 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2030 enum pool_mode old_mode = get_pool_mode(pool);
2031 enum pool_mode new_mode = pt->adjusted_pf.mode;
2034 * Don't change the pool's mode until set_pool_mode() below.
2035 * Otherwise the pool's process_* function pointers may
2036 * not match the desired pool mode.
2038 pt->adjusted_pf.mode = old_mode;
2041 pool->pf = pt->adjusted_pf;
2042 pool->low_water_blocks = pt->low_water_blocks;
2044 set_pool_mode(pool, new_mode);
2049 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2055 /*----------------------------------------------------------------
2057 *--------------------------------------------------------------*/
2058 /* Initialize pool features. */
2059 static void pool_features_init(struct pool_features *pf)
2061 pf->mode = PM_WRITE;
2062 pf->zero_new_blocks = true;
2063 pf->discard_enabled = true;
2064 pf->discard_passdown = true;
2065 pf->error_if_no_space = false;
2068 static void __pool_destroy(struct pool *pool)
2070 __pool_table_remove(pool);
2072 if (dm_pool_metadata_close(pool->pmd) < 0)
2073 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2075 dm_bio_prison_destroy(pool->prison);
2076 dm_kcopyd_client_destroy(pool->copier);
2079 destroy_workqueue(pool->wq);
2081 if (pool->next_mapping)
2082 mempool_free(pool->next_mapping, pool->mapping_pool);
2083 mempool_destroy(pool->mapping_pool);
2084 dm_deferred_set_destroy(pool->shared_read_ds);
2085 dm_deferred_set_destroy(pool->all_io_ds);
2089 static struct kmem_cache *_new_mapping_cache;
2091 static struct pool *pool_create(struct mapped_device *pool_md,
2092 struct block_device *metadata_dev,
2093 unsigned long block_size,
2094 int read_only, char **error)
2099 struct dm_pool_metadata *pmd;
2100 bool format_device = read_only ? false : true;
2102 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2104 *error = "Error creating metadata object";
2105 return (struct pool *)pmd;
2108 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2110 *error = "Error allocating memory for pool";
2111 err_p = ERR_PTR(-ENOMEM);
2116 pool->sectors_per_block = block_size;
2117 if (block_size & (block_size - 1))
2118 pool->sectors_per_block_shift = -1;
2120 pool->sectors_per_block_shift = __ffs(block_size);
2121 pool->low_water_blocks = 0;
2122 pool_features_init(&pool->pf);
2123 pool->prison = dm_bio_prison_create(PRISON_CELLS);
2124 if (!pool->prison) {
2125 *error = "Error creating pool's bio prison";
2126 err_p = ERR_PTR(-ENOMEM);
2130 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2131 if (IS_ERR(pool->copier)) {
2132 r = PTR_ERR(pool->copier);
2133 *error = "Error creating pool's kcopyd client";
2135 goto bad_kcopyd_client;
2139 * Create singlethreaded workqueue that will service all devices
2140 * that use this metadata.
2142 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2144 *error = "Error creating pool's workqueue";
2145 err_p = ERR_PTR(-ENOMEM);
2149 INIT_WORK(&pool->worker, do_worker);
2150 INIT_DELAYED_WORK(&pool->waker, do_waker);
2151 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2152 spin_lock_init(&pool->lock);
2153 bio_list_init(&pool->deferred_flush_bios);
2154 INIT_LIST_HEAD(&pool->prepared_mappings);
2155 INIT_LIST_HEAD(&pool->prepared_discards);
2156 INIT_LIST_HEAD(&pool->active_thins);
2157 pool->low_water_triggered = false;
2159 pool->shared_read_ds = dm_deferred_set_create();
2160 if (!pool->shared_read_ds) {
2161 *error = "Error creating pool's shared read deferred set";
2162 err_p = ERR_PTR(-ENOMEM);
2163 goto bad_shared_read_ds;
2166 pool->all_io_ds = dm_deferred_set_create();
2167 if (!pool->all_io_ds) {
2168 *error = "Error creating pool's all io deferred set";
2169 err_p = ERR_PTR(-ENOMEM);
2173 pool->next_mapping = NULL;
2174 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2175 _new_mapping_cache);
2176 if (!pool->mapping_pool) {
2177 *error = "Error creating pool's mapping mempool";
2178 err_p = ERR_PTR(-ENOMEM);
2179 goto bad_mapping_pool;
2182 pool->ref_count = 1;
2183 pool->last_commit_jiffies = jiffies;
2184 pool->pool_md = pool_md;
2185 pool->md_dev = metadata_dev;
2186 __pool_table_insert(pool);
2191 dm_deferred_set_destroy(pool->all_io_ds);
2193 dm_deferred_set_destroy(pool->shared_read_ds);
2195 destroy_workqueue(pool->wq);
2197 dm_kcopyd_client_destroy(pool->copier);
2199 dm_bio_prison_destroy(pool->prison);
2203 if (dm_pool_metadata_close(pmd))
2204 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2209 static void __pool_inc(struct pool *pool)
2211 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2215 static void __pool_dec(struct pool *pool)
2217 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2218 BUG_ON(!pool->ref_count);
2219 if (!--pool->ref_count)
2220 __pool_destroy(pool);
2223 static struct pool *__pool_find(struct mapped_device *pool_md,
2224 struct block_device *metadata_dev,
2225 unsigned long block_size, int read_only,
2226 char **error, int *created)
2228 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2231 if (pool->pool_md != pool_md) {
2232 *error = "metadata device already in use by a pool";
2233 return ERR_PTR(-EBUSY);
2238 pool = __pool_table_lookup(pool_md);
2240 if (pool->md_dev != metadata_dev) {
2241 *error = "different pool cannot replace a pool";
2242 return ERR_PTR(-EINVAL);
2247 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2255 /*----------------------------------------------------------------
2256 * Pool target methods
2257 *--------------------------------------------------------------*/
2258 static void pool_dtr(struct dm_target *ti)
2260 struct pool_c *pt = ti->private;
2262 mutex_lock(&dm_thin_pool_table.mutex);
2264 unbind_control_target(pt->pool, ti);
2265 __pool_dec(pt->pool);
2266 dm_put_device(ti, pt->metadata_dev);
2267 dm_put_device(ti, pt->data_dev);
2270 mutex_unlock(&dm_thin_pool_table.mutex);
2273 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2274 struct dm_target *ti)
2278 const char *arg_name;
2280 static struct dm_arg _args[] = {
2281 {0, 4, "Invalid number of pool feature arguments"},
2285 * No feature arguments supplied.
2290 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2294 while (argc && !r) {
2295 arg_name = dm_shift_arg(as);
2298 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2299 pf->zero_new_blocks = false;
2301 else if (!strcasecmp(arg_name, "ignore_discard"))
2302 pf->discard_enabled = false;
2304 else if (!strcasecmp(arg_name, "no_discard_passdown"))
2305 pf->discard_passdown = false;
2307 else if (!strcasecmp(arg_name, "read_only"))
2308 pf->mode = PM_READ_ONLY;
2310 else if (!strcasecmp(arg_name, "error_if_no_space"))
2311 pf->error_if_no_space = true;
2314 ti->error = "Unrecognised pool feature requested";
2323 static void metadata_low_callback(void *context)
2325 struct pool *pool = context;
2327 DMWARN("%s: reached low water mark for metadata device: sending event.",
2328 dm_device_name(pool->pool_md));
2330 dm_table_event(pool->ti->table);
2333 static sector_t get_dev_size(struct block_device *bdev)
2335 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
2338 static void warn_if_metadata_device_too_big(struct block_device *bdev)
2340 sector_t metadata_dev_size = get_dev_size(bdev);
2341 char buffer[BDEVNAME_SIZE];
2343 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
2344 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2345 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
2348 static sector_t get_metadata_dev_size(struct block_device *bdev)
2350 sector_t metadata_dev_size = get_dev_size(bdev);
2352 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
2353 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
2355 return metadata_dev_size;
2358 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
2360 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
2362 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
2364 return metadata_dev_size;
2368 * When a metadata threshold is crossed a dm event is triggered, and
2369 * userland should respond by growing the metadata device. We could let
2370 * userland set the threshold, like we do with the data threshold, but I'm
2371 * not sure they know enough to do this well.
2373 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
2376 * 4M is ample for all ops with the possible exception of thin
2377 * device deletion which is harmless if it fails (just retry the
2378 * delete after you've grown the device).
2380 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
2381 return min((dm_block_t)1024ULL /* 4M */, quarter);
2385 * thin-pool <metadata dev> <data dev>
2386 * <data block size (sectors)>
2387 * <low water mark (blocks)>
2388 * [<#feature args> [<arg>]*]
2390 * Optional feature arguments are:
2391 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
2392 * ignore_discard: disable discard
2393 * no_discard_passdown: don't pass discards down to the data device
2394 * read_only: Don't allow any changes to be made to the pool metadata.
2395 * error_if_no_space: error IOs, instead of queueing, if no space.
2397 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
2399 int r, pool_created = 0;
2402 struct pool_features pf;
2403 struct dm_arg_set as;
2404 struct dm_dev *data_dev;
2405 unsigned long block_size;
2406 dm_block_t low_water_blocks;
2407 struct dm_dev *metadata_dev;
2408 fmode_t metadata_mode;
2411 * FIXME Remove validation from scope of lock.
2413 mutex_lock(&dm_thin_pool_table.mutex);
2416 ti->error = "Invalid argument count";
2425 * Set default pool features.
2427 pool_features_init(&pf);
2429 dm_consume_args(&as, 4);
2430 r = parse_pool_features(&as, &pf, ti);
2434 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2435 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2437 ti->error = "Error opening metadata block device";
2440 warn_if_metadata_device_too_big(metadata_dev->bdev);
2442 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2444 ti->error = "Error getting data device";
2448 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2449 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2450 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2451 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2452 ti->error = "Invalid block size";
2457 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2458 ti->error = "Invalid low water mark";
2463 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2469 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2470 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2477 * 'pool_created' reflects whether this is the first table load.
2478 * Top level discard support is not allowed to be changed after
2479 * initial load. This would require a pool reload to trigger thin
2482 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2483 ti->error = "Discard support cannot be disabled once enabled";
2485 goto out_flags_changed;
2490 pt->metadata_dev = metadata_dev;
2491 pt->data_dev = data_dev;
2492 pt->low_water_blocks = low_water_blocks;
2493 pt->adjusted_pf = pt->requested_pf = pf;
2494 ti->num_flush_bios = 1;
2497 * Only need to enable discards if the pool should pass
2498 * them down to the data device. The thin device's discard
2499 * processing will cause mappings to be removed from the btree.
2501 ti->discard_zeroes_data_unsupported = true;
2502 if (pf.discard_enabled && pf.discard_passdown) {
2503 ti->num_discard_bios = 1;
2506 * Setting 'discards_supported' circumvents the normal
2507 * stacking of discard limits (this keeps the pool and
2508 * thin devices' discard limits consistent).
2510 ti->discards_supported = true;
2514 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2515 calc_metadata_threshold(pt),
2516 metadata_low_callback,
2521 pt->callbacks.congested_fn = pool_is_congested;
2522 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2524 mutex_unlock(&dm_thin_pool_table.mutex);
2533 dm_put_device(ti, data_dev);
2535 dm_put_device(ti, metadata_dev);
2537 mutex_unlock(&dm_thin_pool_table.mutex);
2542 static int pool_map(struct dm_target *ti, struct bio *bio)
2545 struct pool_c *pt = ti->private;
2546 struct pool *pool = pt->pool;
2547 unsigned long flags;
2550 * As this is a singleton target, ti->begin is always zero.
2552 spin_lock_irqsave(&pool->lock, flags);
2553 bio->bi_bdev = pt->data_dev->bdev;
2554 r = DM_MAPIO_REMAPPED;
2555 spin_unlock_irqrestore(&pool->lock, flags);
2560 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
2563 struct pool_c *pt = ti->private;
2564 struct pool *pool = pt->pool;
2565 sector_t data_size = ti->len;
2566 dm_block_t sb_data_size;
2568 *need_commit = false;
2570 (void) sector_div(data_size, pool->sectors_per_block);
2572 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2574 DMERR("%s: failed to retrieve data device size",
2575 dm_device_name(pool->pool_md));
2579 if (data_size < sb_data_size) {
2580 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
2581 dm_device_name(pool->pool_md),
2582 (unsigned long long)data_size, sb_data_size);
2585 } else if (data_size > sb_data_size) {
2586 if (dm_pool_metadata_needs_check(pool->pmd)) {
2587 DMERR("%s: unable to grow the data device until repaired.",
2588 dm_device_name(pool->pool_md));
2593 DMINFO("%s: growing the data device from %llu to %llu blocks",
2594 dm_device_name(pool->pool_md),
2595 sb_data_size, (unsigned long long)data_size);
2596 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2598 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
2602 *need_commit = true;
2608 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
2611 struct pool_c *pt = ti->private;
2612 struct pool *pool = pt->pool;
2613 dm_block_t metadata_dev_size, sb_metadata_dev_size;
2615 *need_commit = false;
2617 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
2619 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
2621 DMERR("%s: failed to retrieve metadata device size",
2622 dm_device_name(pool->pool_md));
2626 if (metadata_dev_size < sb_metadata_dev_size) {
2627 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
2628 dm_device_name(pool->pool_md),
2629 metadata_dev_size, sb_metadata_dev_size);
2632 } else if (metadata_dev_size > sb_metadata_dev_size) {
2633 if (dm_pool_metadata_needs_check(pool->pmd)) {
2634 DMERR("%s: unable to grow the metadata device until repaired.",
2635 dm_device_name(pool->pool_md));
2639 warn_if_metadata_device_too_big(pool->md_dev);
2640 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
2641 dm_device_name(pool->pool_md),
2642 sb_metadata_dev_size, metadata_dev_size);
2643 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
2645 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
2649 *need_commit = true;
2656 * Retrieves the number of blocks of the data device from
2657 * the superblock and compares it to the actual device size,
2658 * thus resizing the data device in case it has grown.
2660 * This both copes with opening preallocated data devices in the ctr
2661 * being followed by a resume
2663 * calling the resume method individually after userspace has
2664 * grown the data device in reaction to a table event.
2666 static int pool_preresume(struct dm_target *ti)
2669 bool need_commit1, need_commit2;
2670 struct pool_c *pt = ti->private;
2671 struct pool *pool = pt->pool;
2674 * Take control of the pool object.
2676 r = bind_control_target(pool, ti);
2680 r = maybe_resize_data_dev(ti, &need_commit1);
2684 r = maybe_resize_metadata_dev(ti, &need_commit2);
2688 if (need_commit1 || need_commit2)
2689 (void) commit(pool);
2694 static void pool_resume(struct dm_target *ti)
2696 struct pool_c *pt = ti->private;
2697 struct pool *pool = pt->pool;
2698 unsigned long flags;
2700 spin_lock_irqsave(&pool->lock, flags);
2701 pool->low_water_triggered = false;
2702 spin_unlock_irqrestore(&pool->lock, flags);
2705 do_waker(&pool->waker.work);
2708 static void pool_postsuspend(struct dm_target *ti)
2710 struct pool_c *pt = ti->private;
2711 struct pool *pool = pt->pool;
2713 cancel_delayed_work(&pool->waker);
2714 cancel_delayed_work(&pool->no_space_timeout);
2715 flush_workqueue(pool->wq);
2716 (void) commit(pool);
2719 static int check_arg_count(unsigned argc, unsigned args_required)
2721 if (argc != args_required) {
2722 DMWARN("Message received with %u arguments instead of %u.",
2723 argc, args_required);
2730 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2732 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2733 *dev_id <= MAX_DEV_ID)
2737 DMWARN("Message received with invalid device id: %s", arg);
2742 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2747 r = check_arg_count(argc, 2);
2751 r = read_dev_id(argv[1], &dev_id, 1);
2755 r = dm_pool_create_thin(pool->pmd, dev_id);
2757 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2765 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2768 dm_thin_id origin_dev_id;
2771 r = check_arg_count(argc, 3);
2775 r = read_dev_id(argv[1], &dev_id, 1);
2779 r = read_dev_id(argv[2], &origin_dev_id, 1);
2783 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2785 DMWARN("Creation of new snapshot %s of device %s failed.",
2793 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2798 r = check_arg_count(argc, 2);
2802 r = read_dev_id(argv[1], &dev_id, 1);
2806 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2808 DMWARN("Deletion of thin device %s failed.", argv[1]);
2813 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2815 dm_thin_id old_id, new_id;
2818 r = check_arg_count(argc, 3);
2822 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2823 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2827 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2828 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2832 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2834 DMWARN("Failed to change transaction id from %s to %s.",
2842 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2846 r = check_arg_count(argc, 1);
2850 (void) commit(pool);
2852 r = dm_pool_reserve_metadata_snap(pool->pmd);
2854 DMWARN("reserve_metadata_snap message failed.");
2859 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2863 r = check_arg_count(argc, 1);
2867 r = dm_pool_release_metadata_snap(pool->pmd);
2869 DMWARN("release_metadata_snap message failed.");
2875 * Messages supported:
2876 * create_thin <dev_id>
2877 * create_snap <dev_id> <origin_id>
2879 * trim <dev_id> <new_size_in_sectors>
2880 * set_transaction_id <current_trans_id> <new_trans_id>
2881 * reserve_metadata_snap
2882 * release_metadata_snap
2884 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2887 struct pool_c *pt = ti->private;
2888 struct pool *pool = pt->pool;
2890 if (!strcasecmp(argv[0], "create_thin"))
2891 r = process_create_thin_mesg(argc, argv, pool);
2893 else if (!strcasecmp(argv[0], "create_snap"))
2894 r = process_create_snap_mesg(argc, argv, pool);
2896 else if (!strcasecmp(argv[0], "delete"))
2897 r = process_delete_mesg(argc, argv, pool);
2899 else if (!strcasecmp(argv[0], "set_transaction_id"))
2900 r = process_set_transaction_id_mesg(argc, argv, pool);
2902 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2903 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2905 else if (!strcasecmp(argv[0], "release_metadata_snap"))
2906 r = process_release_metadata_snap_mesg(argc, argv, pool);
2909 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2912 (void) commit(pool);
2917 static void emit_flags(struct pool_features *pf, char *result,
2918 unsigned sz, unsigned maxlen)
2920 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2921 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
2922 pf->error_if_no_space;
2923 DMEMIT("%u ", count);
2925 if (!pf->zero_new_blocks)
2926 DMEMIT("skip_block_zeroing ");
2928 if (!pf->discard_enabled)
2929 DMEMIT("ignore_discard ");
2931 if (!pf->discard_passdown)
2932 DMEMIT("no_discard_passdown ");
2934 if (pf->mode == PM_READ_ONLY)
2935 DMEMIT("read_only ");
2937 if (pf->error_if_no_space)
2938 DMEMIT("error_if_no_space ");
2943 * <transaction id> <used metadata sectors>/<total metadata sectors>
2944 * <used data sectors>/<total data sectors> <held metadata root>
2946 static void pool_status(struct dm_target *ti, status_type_t type,
2947 unsigned status_flags, char *result, unsigned maxlen)
2951 uint64_t transaction_id;
2952 dm_block_t nr_free_blocks_data;
2953 dm_block_t nr_free_blocks_metadata;
2954 dm_block_t nr_blocks_data;
2955 dm_block_t nr_blocks_metadata;
2956 dm_block_t held_root;
2957 char buf[BDEVNAME_SIZE];
2958 char buf2[BDEVNAME_SIZE];
2959 struct pool_c *pt = ti->private;
2960 struct pool *pool = pt->pool;
2963 case STATUSTYPE_INFO:
2964 if (get_pool_mode(pool) == PM_FAIL) {
2969 /* Commit to ensure statistics aren't out-of-date */
2970 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2971 (void) commit(pool);
2973 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
2975 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
2976 dm_device_name(pool->pool_md), r);
2980 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
2982 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
2983 dm_device_name(pool->pool_md), r);
2987 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2989 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
2990 dm_device_name(pool->pool_md), r);
2994 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
2996 DMERR("%s: dm_pool_get_free_block_count returned %d",
2997 dm_device_name(pool->pool_md), r);
3001 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3003 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3004 dm_device_name(pool->pool_md), r);
3008 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3010 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3011 dm_device_name(pool->pool_md), r);
3015 DMEMIT("%llu %llu/%llu %llu/%llu ",
3016 (unsigned long long)transaction_id,
3017 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3018 (unsigned long long)nr_blocks_metadata,
3019 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3020 (unsigned long long)nr_blocks_data);
3023 DMEMIT("%llu ", held_root);
3027 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3028 DMEMIT("out_of_data_space ");
3029 else if (pool->pf.mode == PM_READ_ONLY)
3034 if (!pool->pf.discard_enabled)
3035 DMEMIT("ignore_discard ");
3036 else if (pool->pf.discard_passdown)
3037 DMEMIT("discard_passdown ");
3039 DMEMIT("no_discard_passdown ");
3041 if (pool->pf.error_if_no_space)
3042 DMEMIT("error_if_no_space ");
3044 DMEMIT("queue_if_no_space ");
3048 case STATUSTYPE_TABLE:
3049 DMEMIT("%s %s %lu %llu ",
3050 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3051 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3052 (unsigned long)pool->sectors_per_block,
3053 (unsigned long long)pt->low_water_blocks);
3054 emit_flags(&pt->requested_pf, result, sz, maxlen);
3063 static int pool_iterate_devices(struct dm_target *ti,
3064 iterate_devices_callout_fn fn, void *data)
3066 struct pool_c *pt = ti->private;
3068 return fn(ti, pt->data_dev, 0, ti->len, data);
3071 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3072 struct bio_vec *biovec, int max_size)
3074 struct pool_c *pt = ti->private;
3075 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3077 if (!q->merge_bvec_fn)
3080 bvm->bi_bdev = pt->data_dev->bdev;
3082 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3085 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
3087 struct pool *pool = pt->pool;
3088 struct queue_limits *data_limits;
3090 limits->max_discard_sectors = pool->sectors_per_block;
3093 * discard_granularity is just a hint, and not enforced.
3095 if (pt->adjusted_pf.discard_passdown) {
3096 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
3097 limits->discard_granularity = max(data_limits->discard_granularity,
3098 pool->sectors_per_block << SECTOR_SHIFT);
3100 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
3103 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3105 struct pool_c *pt = ti->private;
3106 struct pool *pool = pt->pool;
3107 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3110 * If the system-determined stacked limits are compatible with the
3111 * pool's blocksize (io_opt is a factor) do not override them.
3113 if (io_opt_sectors < pool->sectors_per_block ||
3114 do_div(io_opt_sectors, pool->sectors_per_block)) {
3115 blk_limits_io_min(limits, 0);
3116 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3120 * pt->adjusted_pf is a staging area for the actual features to use.
3121 * They get transferred to the live pool in bind_control_target()
3122 * called from pool_preresume().
3124 if (!pt->adjusted_pf.discard_enabled) {
3126 * Must explicitly disallow stacking discard limits otherwise the
3127 * block layer will stack them if pool's data device has support.
3128 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3129 * user to see that, so make sure to set all discard limits to 0.
3131 limits->discard_granularity = 0;
3135 disable_passdown_if_not_supported(pt);
3137 set_discard_limits(pt, limits);
3140 static struct target_type pool_target = {
3141 .name = "thin-pool",
3142 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3143 DM_TARGET_IMMUTABLE,
3144 .version = {1, 12, 0},
3145 .module = THIS_MODULE,
3149 .postsuspend = pool_postsuspend,
3150 .preresume = pool_preresume,
3151 .resume = pool_resume,
3152 .message = pool_message,
3153 .status = pool_status,
3154 .merge = pool_merge,
3155 .iterate_devices = pool_iterate_devices,
3156 .io_hints = pool_io_hints,
3159 /*----------------------------------------------------------------
3160 * Thin target methods
3161 *--------------------------------------------------------------*/
3162 static void thin_get(struct thin_c *tc)
3164 atomic_inc(&tc->refcount);
3167 static void thin_put(struct thin_c *tc)
3169 if (atomic_dec_and_test(&tc->refcount))
3170 complete(&tc->can_destroy);
3173 static void thin_dtr(struct dm_target *ti)
3175 struct thin_c *tc = ti->private;
3176 unsigned long flags;
3179 wait_for_completion(&tc->can_destroy);
3181 spin_lock_irqsave(&tc->pool->lock, flags);
3182 list_del_rcu(&tc->list);
3183 spin_unlock_irqrestore(&tc->pool->lock, flags);
3186 mutex_lock(&dm_thin_pool_table.mutex);
3188 __pool_dec(tc->pool);
3189 dm_pool_close_thin_device(tc->td);
3190 dm_put_device(ti, tc->pool_dev);
3192 dm_put_device(ti, tc->origin_dev);
3195 mutex_unlock(&dm_thin_pool_table.mutex);
3199 * Thin target parameters:
3201 * <pool_dev> <dev_id> [origin_dev]
3203 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3204 * dev_id: the internal device identifier
3205 * origin_dev: a device external to the pool that should act as the origin
3207 * If the pool device has discards disabled, they get disabled for the thin
3210 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3214 struct dm_dev *pool_dev, *origin_dev;
3215 struct mapped_device *pool_md;
3216 unsigned long flags;
3218 mutex_lock(&dm_thin_pool_table.mutex);
3220 if (argc != 2 && argc != 3) {
3221 ti->error = "Invalid argument count";
3226 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3228 ti->error = "Out of memory";
3232 spin_lock_init(&tc->lock);
3233 bio_list_init(&tc->deferred_bio_list);
3234 bio_list_init(&tc->retry_on_resume_list);
3235 tc->sort_bio_list = RB_ROOT;
3238 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
3240 ti->error = "Error opening origin device";
3241 goto bad_origin_dev;
3243 tc->origin_dev = origin_dev;
3246 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
3248 ti->error = "Error opening pool device";
3251 tc->pool_dev = pool_dev;
3253 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
3254 ti->error = "Invalid device id";
3259 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
3261 ti->error = "Couldn't get pool mapped device";
3266 tc->pool = __pool_table_lookup(pool_md);
3268 ti->error = "Couldn't find pool object";
3270 goto bad_pool_lookup;
3272 __pool_inc(tc->pool);
3274 if (get_pool_mode(tc->pool) == PM_FAIL) {
3275 ti->error = "Couldn't open thin device, Pool is in fail mode";
3280 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
3282 ti->error = "Couldn't open thin internal device";
3286 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
3288 goto bad_target_max_io_len;
3290 ti->num_flush_bios = 1;
3291 ti->flush_supported = true;
3292 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
3294 /* In case the pool supports discards, pass them on. */
3295 ti->discard_zeroes_data_unsupported = true;
3296 if (tc->pool->pf.discard_enabled) {
3297 ti->discards_supported = true;
3298 ti->num_discard_bios = 1;
3299 /* Discard bios must be split on a block boundary */
3300 ti->split_discard_bios = true;
3305 mutex_unlock(&dm_thin_pool_table.mutex);
3307 atomic_set(&tc->refcount, 1);
3308 init_completion(&tc->can_destroy);
3310 spin_lock_irqsave(&tc->pool->lock, flags);
3311 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
3312 spin_unlock_irqrestore(&tc->pool->lock, flags);
3314 * This synchronize_rcu() call is needed here otherwise we risk a
3315 * wake_worker() call finding no bios to process (because the newly
3316 * added tc isn't yet visible). So this reduces latency since we
3317 * aren't then dependent on the periodic commit to wake_worker().
3323 bad_target_max_io_len:
3324 dm_pool_close_thin_device(tc->td);
3326 __pool_dec(tc->pool);
3330 dm_put_device(ti, tc->pool_dev);
3333 dm_put_device(ti, tc->origin_dev);
3337 mutex_unlock(&dm_thin_pool_table.mutex);
3342 static int thin_map(struct dm_target *ti, struct bio *bio)
3344 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
3346 return thin_bio_map(ti, bio);
3349 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
3351 unsigned long flags;
3352 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
3353 struct list_head work;
3354 struct dm_thin_new_mapping *m, *tmp;
3355 struct pool *pool = h->tc->pool;
3357 if (h->shared_read_entry) {
3358 INIT_LIST_HEAD(&work);
3359 dm_deferred_entry_dec(h->shared_read_entry, &work);
3361 spin_lock_irqsave(&pool->lock, flags);
3362 list_for_each_entry_safe(m, tmp, &work, list) {
3365 __maybe_add_mapping(m);
3367 spin_unlock_irqrestore(&pool->lock, flags);
3370 if (h->all_io_entry) {
3371 INIT_LIST_HEAD(&work);
3372 dm_deferred_entry_dec(h->all_io_entry, &work);
3373 if (!list_empty(&work)) {
3374 spin_lock_irqsave(&pool->lock, flags);
3375 list_for_each_entry_safe(m, tmp, &work, list)
3376 list_add_tail(&m->list, &pool->prepared_discards);
3377 spin_unlock_irqrestore(&pool->lock, flags);
3385 static void thin_presuspend(struct dm_target *ti)
3387 struct thin_c *tc = ti->private;
3389 if (dm_noflush_suspending(ti))
3390 noflush_work(tc, do_noflush_start);
3393 static void thin_postsuspend(struct dm_target *ti)
3395 struct thin_c *tc = ti->private;
3398 * The dm_noflush_suspending flag has been cleared by now, so
3399 * unfortunately we must always run this.
3401 noflush_work(tc, do_noflush_stop);
3405 * <nr mapped sectors> <highest mapped sector>
3407 static void thin_status(struct dm_target *ti, status_type_t type,
3408 unsigned status_flags, char *result, unsigned maxlen)
3412 dm_block_t mapped, highest;
3413 char buf[BDEVNAME_SIZE];
3414 struct thin_c *tc = ti->private;
3416 if (get_pool_mode(tc->pool) == PM_FAIL) {
3425 case STATUSTYPE_INFO:
3426 r = dm_thin_get_mapped_count(tc->td, &mapped);
3428 DMERR("dm_thin_get_mapped_count returned %d", r);
3432 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
3434 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
3438 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
3440 DMEMIT("%llu", ((highest + 1) *
3441 tc->pool->sectors_per_block) - 1);
3446 case STATUSTYPE_TABLE:
3448 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
3449 (unsigned long) tc->dev_id);
3451 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
3462 static int thin_iterate_devices(struct dm_target *ti,
3463 iterate_devices_callout_fn fn, void *data)
3466 struct thin_c *tc = ti->private;
3467 struct pool *pool = tc->pool;
3470 * We can't call dm_pool_get_data_dev_size() since that blocks. So
3471 * we follow a more convoluted path through to the pool's target.
3474 return 0; /* nothing is bound */
3476 blocks = pool->ti->len;
3477 (void) sector_div(blocks, pool->sectors_per_block);
3479 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
3484 static struct target_type thin_target = {
3486 .version = {1, 12, 0},
3487 .module = THIS_MODULE,
3491 .end_io = thin_endio,
3492 .presuspend = thin_presuspend,
3493 .postsuspend = thin_postsuspend,
3494 .status = thin_status,
3495 .iterate_devices = thin_iterate_devices,
3498 /*----------------------------------------------------------------*/
3500 static int __init dm_thin_init(void)
3506 r = dm_register_target(&thin_target);
3510 r = dm_register_target(&pool_target);
3512 goto bad_pool_target;
3516 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
3517 if (!_new_mapping_cache)
3518 goto bad_new_mapping_cache;
3522 bad_new_mapping_cache:
3523 dm_unregister_target(&pool_target);
3525 dm_unregister_target(&thin_target);
3530 static void dm_thin_exit(void)
3532 dm_unregister_target(&thin_target);
3533 dm_unregister_target(&pool_target);
3535 kmem_cache_destroy(_new_mapping_cache);
3538 module_init(dm_thin_init);
3539 module_exit(dm_thin_exit);
3541 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
3542 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
3544 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
3545 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3546 MODULE_LICENSE("GPL");