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/jiffies.h>
15 #include <linux/log2.h>
16 #include <linux/list.h>
17 #include <linux/rculist.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/sort.h>
22 #include <linux/rbtree.h>
24 #define DM_MSG_PREFIX "thin"
29 #define ENDIO_HOOK_POOL_SIZE 1024
30 #define MAPPING_POOL_SIZE 1024
31 #define COMMIT_PERIOD HZ
32 #define NO_SPACE_TIMEOUT_SECS 60
34 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
37 "A percentage of time allocated for copy on write");
40 * The block size of the device holding pool data must be
41 * between 64KB and 1GB.
43 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
44 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
47 * Device id is restricted to 24 bits.
49 #define MAX_DEV_ID ((1 << 24) - 1)
52 * How do we handle breaking sharing of data blocks?
53 * =================================================
55 * We use a standard copy-on-write btree to store the mappings for the
56 * devices (note I'm talking about copy-on-write of the metadata here, not
57 * the data). When you take an internal snapshot you clone the root node
58 * of the origin btree. After this there is no concept of an origin or a
59 * snapshot. They are just two device trees that happen to point to the
62 * When we get a write in we decide if it's to a shared data block using
63 * some timestamp magic. If it is, we have to break sharing.
65 * Let's say we write to a shared block in what was the origin. The
68 * i) plug io further to this physical block. (see bio_prison code).
70 * ii) quiesce any read io to that shared data block. Obviously
71 * including all devices that share this block. (see dm_deferred_set code)
73 * iii) copy the data block to a newly allocate block. This step can be
74 * missed out if the io covers the block. (schedule_copy).
76 * iv) insert the new mapping into the origin's btree
77 * (process_prepared_mapping). This act of inserting breaks some
78 * sharing of btree nodes between the two devices. Breaking sharing only
79 * effects the btree of that specific device. Btrees for the other
80 * devices that share the block never change. The btree for the origin
81 * device as it was after the last commit is untouched, ie. we're using
82 * persistent data structures in the functional programming sense.
84 * v) unplug io to this physical block, including the io that triggered
85 * the breaking of sharing.
87 * Steps (ii) and (iii) occur in parallel.
89 * The metadata _doesn't_ need to be committed before the io continues. We
90 * get away with this because the io is always written to a _new_ block.
91 * If there's a crash, then:
93 * - The origin mapping will point to the old origin block (the shared
94 * one). This will contain the data as it was before the io that triggered
95 * the breaking of sharing came in.
97 * - The snap mapping still points to the old block. As it would after
100 * The downside of this scheme is the timestamp magic isn't perfect, and
101 * will continue to think that data block in the snapshot device is shared
102 * even after the write to the origin has broken sharing. I suspect data
103 * blocks will typically be shared by many different devices, so we're
104 * breaking sharing n + 1 times, rather than n, where n is the number of
105 * devices that reference this data block. At the moment I think the
106 * benefits far, far outweigh the disadvantages.
109 /*----------------------------------------------------------------*/
114 static void build_data_key(struct dm_thin_device *td,
115 dm_block_t b, struct dm_cell_key *key)
118 key->dev = dm_thin_dev_id(td);
119 key->block_begin = b;
120 key->block_end = b + 1ULL;
123 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
124 struct dm_cell_key *key)
127 key->dev = dm_thin_dev_id(td);
128 key->block_begin = b;
129 key->block_end = b + 1ULL;
132 /*----------------------------------------------------------------*/
134 #define THROTTLE_THRESHOLD (1 * HZ)
137 struct rw_semaphore lock;
138 unsigned long threshold;
139 bool throttle_applied;
142 static void throttle_init(struct throttle *t)
144 init_rwsem(&t->lock);
145 t->throttle_applied = false;
148 static void throttle_work_start(struct throttle *t)
150 t->threshold = jiffies + THROTTLE_THRESHOLD;
153 static void throttle_work_update(struct throttle *t)
155 if (!t->throttle_applied && jiffies > t->threshold) {
156 down_write(&t->lock);
157 t->throttle_applied = true;
161 static void throttle_work_complete(struct throttle *t)
163 if (t->throttle_applied) {
164 t->throttle_applied = false;
169 static void throttle_lock(struct throttle *t)
174 static void throttle_unlock(struct throttle *t)
179 /*----------------------------------------------------------------*/
182 * A pool device ties together a metadata device and a data device. It
183 * also provides the interface for creating and destroying internal
186 struct dm_thin_new_mapping;
189 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
192 PM_WRITE, /* metadata may be changed */
193 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
194 PM_READ_ONLY, /* metadata may not be changed */
195 PM_FAIL, /* all I/O fails */
198 struct pool_features {
201 bool zero_new_blocks:1;
202 bool discard_enabled:1;
203 bool discard_passdown:1;
204 bool error_if_no_space:1;
208 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
209 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
210 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
212 #define CELL_SORT_ARRAY_SIZE 8192
215 struct list_head list;
216 struct dm_target *ti; /* Only set if a pool target is bound */
218 struct mapped_device *pool_md;
219 struct block_device *md_dev;
220 struct dm_pool_metadata *pmd;
222 dm_block_t low_water_blocks;
223 uint32_t sectors_per_block;
224 int sectors_per_block_shift;
226 struct pool_features pf;
227 bool low_water_triggered:1; /* A dm event has been sent */
230 struct dm_bio_prison *prison;
231 struct dm_kcopyd_client *copier;
233 struct workqueue_struct *wq;
234 struct throttle throttle;
235 struct work_struct worker;
236 struct delayed_work waker;
237 struct delayed_work no_space_timeout;
239 unsigned long last_commit_jiffies;
243 struct bio_list deferred_flush_bios;
244 struct list_head prepared_mappings;
245 struct list_head prepared_discards;
246 struct list_head active_thins;
248 struct dm_deferred_set *shared_read_ds;
249 struct dm_deferred_set *all_io_ds;
251 struct dm_thin_new_mapping *next_mapping;
252 mempool_t *mapping_pool;
254 process_bio_fn process_bio;
255 process_bio_fn process_discard;
257 process_cell_fn process_cell;
258 process_cell_fn process_discard_cell;
260 process_mapping_fn process_prepared_mapping;
261 process_mapping_fn process_prepared_discard;
263 struct dm_bio_prison_cell *cell_sort_array[CELL_SORT_ARRAY_SIZE];
266 static enum pool_mode get_pool_mode(struct pool *pool);
267 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
270 * Target context for a pool.
273 struct dm_target *ti;
275 struct dm_dev *data_dev;
276 struct dm_dev *metadata_dev;
277 struct dm_target_callbacks callbacks;
279 dm_block_t low_water_blocks;
280 struct pool_features requested_pf; /* Features requested during table load */
281 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
285 * Target context for a thin.
288 struct list_head list;
289 struct dm_dev *pool_dev;
290 struct dm_dev *origin_dev;
291 sector_t origin_size;
295 struct dm_thin_device *td;
296 struct mapped_device *thin_md;
300 struct list_head deferred_cells;
301 struct bio_list deferred_bio_list;
302 struct bio_list retry_on_resume_list;
303 struct rb_root sort_bio_list; /* sorted list of deferred bios */
306 * Ensures the thin is not destroyed until the worker has finished
307 * iterating the active_thins list.
310 struct completion can_destroy;
313 /*----------------------------------------------------------------*/
316 * wake_worker() is used when new work is queued and when pool_resume is
317 * ready to continue deferred IO processing.
319 static void wake_worker(struct pool *pool)
321 queue_work(pool->wq, &pool->worker);
324 /*----------------------------------------------------------------*/
326 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
327 struct dm_bio_prison_cell **cell_result)
330 struct dm_bio_prison_cell *cell_prealloc;
333 * Allocate a cell from the prison's mempool.
334 * This might block but it can't fail.
336 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
338 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
341 * We reused an old cell; we can get rid of
344 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
349 static void cell_release(struct pool *pool,
350 struct dm_bio_prison_cell *cell,
351 struct bio_list *bios)
353 dm_cell_release(pool->prison, cell, bios);
354 dm_bio_prison_free_cell(pool->prison, cell);
357 static void cell_visit_release(struct pool *pool,
358 void (*fn)(void *, struct dm_bio_prison_cell *),
360 struct dm_bio_prison_cell *cell)
362 dm_cell_visit_release(pool->prison, fn, context, cell);
363 dm_bio_prison_free_cell(pool->prison, cell);
366 static void cell_release_no_holder(struct pool *pool,
367 struct dm_bio_prison_cell *cell,
368 struct bio_list *bios)
370 dm_cell_release_no_holder(pool->prison, cell, bios);
371 dm_bio_prison_free_cell(pool->prison, cell);
374 static void cell_error_with_code(struct pool *pool,
375 struct dm_bio_prison_cell *cell, int error_code)
377 dm_cell_error(pool->prison, cell, error_code);
378 dm_bio_prison_free_cell(pool->prison, cell);
381 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
383 cell_error_with_code(pool, cell, -EIO);
386 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
388 cell_error_with_code(pool, cell, 0);
391 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
393 cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
396 /*----------------------------------------------------------------*/
399 * A global list of pools that uses a struct mapped_device as a key.
401 static struct dm_thin_pool_table {
403 struct list_head pools;
404 } dm_thin_pool_table;
406 static void pool_table_init(void)
408 mutex_init(&dm_thin_pool_table.mutex);
409 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
412 static void __pool_table_insert(struct pool *pool)
414 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
415 list_add(&pool->list, &dm_thin_pool_table.pools);
418 static void __pool_table_remove(struct pool *pool)
420 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
421 list_del(&pool->list);
424 static struct pool *__pool_table_lookup(struct mapped_device *md)
426 struct pool *pool = NULL, *tmp;
428 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
430 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
431 if (tmp->pool_md == md) {
440 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
442 struct pool *pool = NULL, *tmp;
444 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
446 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
447 if (tmp->md_dev == md_dev) {
456 /*----------------------------------------------------------------*/
458 struct dm_thin_endio_hook {
460 struct dm_deferred_entry *shared_read_entry;
461 struct dm_deferred_entry *all_io_entry;
462 struct dm_thin_new_mapping *overwrite_mapping;
463 struct rb_node rb_node;
466 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
468 bio_list_merge(bios, master);
469 bio_list_init(master);
472 static void error_bio_list(struct bio_list *bios, int error)
476 while ((bio = bio_list_pop(bios)))
477 bio_endio(bio, error);
480 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
482 struct bio_list bios;
485 bio_list_init(&bios);
487 spin_lock_irqsave(&tc->lock, flags);
488 __merge_bio_list(&bios, master);
489 spin_unlock_irqrestore(&tc->lock, flags);
491 error_bio_list(&bios, error);
494 static void requeue_deferred_cells(struct thin_c *tc)
496 struct pool *pool = tc->pool;
498 struct list_head cells;
499 struct dm_bio_prison_cell *cell, *tmp;
501 INIT_LIST_HEAD(&cells);
503 spin_lock_irqsave(&tc->lock, flags);
504 list_splice_init(&tc->deferred_cells, &cells);
505 spin_unlock_irqrestore(&tc->lock, flags);
507 list_for_each_entry_safe(cell, tmp, &cells, user_list)
508 cell_requeue(pool, cell);
511 static void requeue_io(struct thin_c *tc)
513 struct bio_list bios;
516 bio_list_init(&bios);
518 spin_lock_irqsave(&tc->lock, flags);
519 __merge_bio_list(&bios, &tc->deferred_bio_list);
520 __merge_bio_list(&bios, &tc->retry_on_resume_list);
521 spin_unlock_irqrestore(&tc->lock, flags);
523 error_bio_list(&bios, DM_ENDIO_REQUEUE);
524 requeue_deferred_cells(tc);
527 static void error_retry_list(struct pool *pool)
532 list_for_each_entry_rcu(tc, &pool->active_thins, list)
533 error_thin_bio_list(tc, &tc->retry_on_resume_list, -EIO);
538 * This section of code contains the logic for processing a thin device's IO.
539 * Much of the code depends on pool object resources (lists, workqueues, etc)
540 * but most is exclusively called from the thin target rather than the thin-pool
544 static bool block_size_is_power_of_two(struct pool *pool)
546 return pool->sectors_per_block_shift >= 0;
549 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
551 struct pool *pool = tc->pool;
552 sector_t block_nr = bio->bi_iter.bi_sector;
554 if (block_size_is_power_of_two(pool))
555 block_nr >>= pool->sectors_per_block_shift;
557 (void) sector_div(block_nr, pool->sectors_per_block);
562 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
564 struct pool *pool = tc->pool;
565 sector_t bi_sector = bio->bi_iter.bi_sector;
567 bio->bi_bdev = tc->pool_dev->bdev;
568 if (block_size_is_power_of_two(pool))
569 bio->bi_iter.bi_sector =
570 (block << pool->sectors_per_block_shift) |
571 (bi_sector & (pool->sectors_per_block - 1));
573 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
574 sector_div(bi_sector, pool->sectors_per_block);
577 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
579 bio->bi_bdev = tc->origin_dev->bdev;
582 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
584 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
585 dm_thin_changed_this_transaction(tc->td);
588 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
590 struct dm_thin_endio_hook *h;
592 if (bio->bi_rw & REQ_DISCARD)
595 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
596 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
599 static void issue(struct thin_c *tc, struct bio *bio)
601 struct pool *pool = tc->pool;
604 if (!bio_triggers_commit(tc, bio)) {
605 generic_make_request(bio);
610 * Complete bio with an error if earlier I/O caused changes to
611 * the metadata that can't be committed e.g, due to I/O errors
612 * on the metadata device.
614 if (dm_thin_aborted_changes(tc->td)) {
620 * Batch together any bios that trigger commits and then issue a
621 * single commit for them in process_deferred_bios().
623 spin_lock_irqsave(&pool->lock, flags);
624 bio_list_add(&pool->deferred_flush_bios, bio);
625 spin_unlock_irqrestore(&pool->lock, flags);
628 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
630 remap_to_origin(tc, bio);
634 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
637 remap(tc, bio, block);
641 /*----------------------------------------------------------------*/
644 * Bio endio functions.
646 struct dm_thin_new_mapping {
647 struct list_head list;
650 bool definitely_not_shared:1;
653 * Track quiescing, copying and zeroing preparation actions. When this
654 * counter hits zero the block is prepared and can be inserted into the
657 atomic_t prepare_actions;
661 dm_block_t virt_block;
662 dm_block_t data_block;
663 struct dm_bio_prison_cell *cell, *cell2;
666 * If the bio covers the whole area of a block then we can avoid
667 * zeroing or copying. Instead this bio is hooked. The bio will
668 * still be in the cell, so care has to be taken to avoid issuing
672 bio_end_io_t *saved_bi_end_io;
675 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
677 struct pool *pool = m->tc->pool;
679 if (atomic_dec_and_test(&m->prepare_actions)) {
680 list_add_tail(&m->list, &pool->prepared_mappings);
685 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
688 struct pool *pool = m->tc->pool;
690 spin_lock_irqsave(&pool->lock, flags);
691 __complete_mapping_preparation(m);
692 spin_unlock_irqrestore(&pool->lock, flags);
695 static void copy_complete(int read_err, unsigned long write_err, void *context)
697 struct dm_thin_new_mapping *m = context;
699 m->err = read_err || write_err ? -EIO : 0;
700 complete_mapping_preparation(m);
703 static void overwrite_endio(struct bio *bio, int err)
705 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
706 struct dm_thin_new_mapping *m = h->overwrite_mapping;
709 complete_mapping_preparation(m);
712 /*----------------------------------------------------------------*/
719 * Prepared mapping jobs.
723 * This sends the bios in the cell, except the original holder, back
724 * to the deferred_bios list.
726 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
728 struct pool *pool = tc->pool;
731 spin_lock_irqsave(&tc->lock, flags);
732 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
733 spin_unlock_irqrestore(&tc->lock, flags);
738 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
742 struct bio_list defer_bios;
743 struct bio_list issue_bios;
746 static void __inc_remap_and_issue_cell(void *context,
747 struct dm_bio_prison_cell *cell)
749 struct remap_info *info = context;
752 while ((bio = bio_list_pop(&cell->bios))) {
753 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA))
754 bio_list_add(&info->defer_bios, bio);
756 inc_all_io_entry(info->tc->pool, bio);
759 * We can't issue the bios with the bio prison lock
760 * held, so we add them to a list to issue on
761 * return from this function.
763 bio_list_add(&info->issue_bios, bio);
768 static void inc_remap_and_issue_cell(struct thin_c *tc,
769 struct dm_bio_prison_cell *cell,
773 struct remap_info info;
776 bio_list_init(&info.defer_bios);
777 bio_list_init(&info.issue_bios);
780 * We have to be careful to inc any bios we're about to issue
781 * before the cell is released, and avoid a race with new bios
782 * being added to the cell.
784 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
787 while ((bio = bio_list_pop(&info.defer_bios)))
788 thin_defer_bio(tc, bio);
790 while ((bio = bio_list_pop(&info.issue_bios)))
791 remap_and_issue(info.tc, bio, block);
794 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
797 m->bio->bi_end_io = m->saved_bi_end_io;
798 atomic_inc(&m->bio->bi_remaining);
800 cell_error(m->tc->pool, m->cell);
802 mempool_free(m, m->tc->pool->mapping_pool);
805 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
807 struct thin_c *tc = m->tc;
808 struct pool *pool = tc->pool;
814 bio->bi_end_io = m->saved_bi_end_io;
815 atomic_inc(&bio->bi_remaining);
819 cell_error(pool, m->cell);
824 * Commit the prepared block into the mapping btree.
825 * Any I/O for this block arriving after this point will get
826 * remapped to it directly.
828 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
830 metadata_operation_failed(pool, "dm_thin_insert_block", r);
831 cell_error(pool, m->cell);
836 * Release any bios held while the block was being provisioned.
837 * If we are processing a write bio that completely covers the block,
838 * we already processed it so can ignore it now when processing
839 * the bios in the cell.
842 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
845 inc_all_io_entry(tc->pool, m->cell->holder);
846 remap_and_issue(tc, m->cell->holder, m->data_block);
847 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
852 mempool_free(m, pool->mapping_pool);
855 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
857 struct thin_c *tc = m->tc;
859 bio_io_error(m->bio);
860 cell_defer_no_holder(tc, m->cell);
861 cell_defer_no_holder(tc, m->cell2);
862 mempool_free(m, tc->pool->mapping_pool);
865 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
867 struct thin_c *tc = m->tc;
869 inc_all_io_entry(tc->pool, m->bio);
870 cell_defer_no_holder(tc, m->cell);
871 cell_defer_no_holder(tc, m->cell2);
874 if (m->definitely_not_shared)
875 remap_and_issue(tc, m->bio, m->data_block);
878 if (dm_pool_block_is_used(tc->pool->pmd, m->data_block, &used) || used)
879 bio_endio(m->bio, 0);
881 remap_and_issue(tc, m->bio, m->data_block);
884 bio_endio(m->bio, 0);
886 mempool_free(m, tc->pool->mapping_pool);
889 static void process_prepared_discard(struct dm_thin_new_mapping *m)
892 struct thin_c *tc = m->tc;
894 r = dm_thin_remove_block(tc->td, m->virt_block);
896 DMERR_LIMIT("dm_thin_remove_block() failed");
898 process_prepared_discard_passdown(m);
901 static void process_prepared(struct pool *pool, struct list_head *head,
902 process_mapping_fn *fn)
905 struct list_head maps;
906 struct dm_thin_new_mapping *m, *tmp;
908 INIT_LIST_HEAD(&maps);
909 spin_lock_irqsave(&pool->lock, flags);
910 list_splice_init(head, &maps);
911 spin_unlock_irqrestore(&pool->lock, flags);
913 list_for_each_entry_safe(m, tmp, &maps, list)
920 static int io_overlaps_block(struct pool *pool, struct bio *bio)
922 return bio->bi_iter.bi_size ==
923 (pool->sectors_per_block << SECTOR_SHIFT);
926 static int io_overwrites_block(struct pool *pool, struct bio *bio)
928 return (bio_data_dir(bio) == WRITE) &&
929 io_overlaps_block(pool, bio);
932 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
935 *save = bio->bi_end_io;
939 static int ensure_next_mapping(struct pool *pool)
941 if (pool->next_mapping)
944 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
946 return pool->next_mapping ? 0 : -ENOMEM;
949 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
951 struct dm_thin_new_mapping *m = pool->next_mapping;
953 BUG_ON(!pool->next_mapping);
955 memset(m, 0, sizeof(struct dm_thin_new_mapping));
956 INIT_LIST_HEAD(&m->list);
959 pool->next_mapping = NULL;
964 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
965 sector_t begin, sector_t end)
968 struct dm_io_region to;
970 to.bdev = tc->pool_dev->bdev;
972 to.count = end - begin;
974 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
976 DMERR_LIMIT("dm_kcopyd_zero() failed");
977 copy_complete(1, 1, m);
981 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
982 dm_block_t data_block,
983 struct dm_thin_new_mapping *m)
985 struct pool *pool = tc->pool;
986 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
988 h->overwrite_mapping = m;
990 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
991 inc_all_io_entry(pool, bio);
992 remap_and_issue(tc, bio, data_block);
996 * A partial copy also needs to zero the uncopied region.
998 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
999 struct dm_dev *origin, dm_block_t data_origin,
1000 dm_block_t data_dest,
1001 struct dm_bio_prison_cell *cell, struct bio *bio,
1005 struct pool *pool = tc->pool;
1006 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1009 m->virt_block = virt_block;
1010 m->data_block = data_dest;
1014 * quiesce action + copy action + an extra reference held for the
1015 * duration of this function (we may need to inc later for a
1018 atomic_set(&m->prepare_actions, 3);
1020 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1021 complete_mapping_preparation(m); /* already quiesced */
1024 * IO to pool_dev remaps to the pool target's data_dev.
1026 * If the whole block of data is being overwritten, we can issue the
1027 * bio immediately. Otherwise we use kcopyd to clone the data first.
1029 if (io_overwrites_block(pool, bio))
1030 remap_and_issue_overwrite(tc, bio, data_dest, m);
1032 struct dm_io_region from, to;
1034 from.bdev = origin->bdev;
1035 from.sector = data_origin * pool->sectors_per_block;
1038 to.bdev = tc->pool_dev->bdev;
1039 to.sector = data_dest * pool->sectors_per_block;
1042 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1043 0, copy_complete, m);
1045 DMERR_LIMIT("dm_kcopyd_copy() failed");
1046 copy_complete(1, 1, m);
1049 * We allow the zero to be issued, to simplify the
1050 * error path. Otherwise we'd need to start
1051 * worrying about decrementing the prepare_actions
1057 * Do we need to zero a tail region?
1059 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1060 atomic_inc(&m->prepare_actions);
1062 data_dest * pool->sectors_per_block + len,
1063 (data_dest + 1) * pool->sectors_per_block);
1067 complete_mapping_preparation(m); /* drop our ref */
1070 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1071 dm_block_t data_origin, dm_block_t data_dest,
1072 struct dm_bio_prison_cell *cell, struct bio *bio)
1074 schedule_copy(tc, virt_block, tc->pool_dev,
1075 data_origin, data_dest, cell, bio,
1076 tc->pool->sectors_per_block);
1079 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1080 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1083 struct pool *pool = tc->pool;
1084 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1086 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1088 m->virt_block = virt_block;
1089 m->data_block = data_block;
1093 * If the whole block of data is being overwritten or we are not
1094 * zeroing pre-existing data, we can issue the bio immediately.
1095 * Otherwise we use kcopyd to zero the data first.
1097 if (!pool->pf.zero_new_blocks)
1098 process_prepared_mapping(m);
1100 else if (io_overwrites_block(pool, bio))
1101 remap_and_issue_overwrite(tc, bio, data_block, m);
1105 data_block * pool->sectors_per_block,
1106 (data_block + 1) * pool->sectors_per_block);
1109 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1110 dm_block_t data_dest,
1111 struct dm_bio_prison_cell *cell, struct bio *bio)
1113 struct pool *pool = tc->pool;
1114 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1115 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1117 if (virt_block_end <= tc->origin_size)
1118 schedule_copy(tc, virt_block, tc->origin_dev,
1119 virt_block, data_dest, cell, bio,
1120 pool->sectors_per_block);
1122 else if (virt_block_begin < tc->origin_size)
1123 schedule_copy(tc, virt_block, tc->origin_dev,
1124 virt_block, data_dest, cell, bio,
1125 tc->origin_size - virt_block_begin);
1128 schedule_zero(tc, virt_block, data_dest, cell, bio);
1131 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1133 static void check_for_space(struct pool *pool)
1138 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1141 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1146 set_pool_mode(pool, PM_WRITE);
1150 * A non-zero return indicates read_only or fail_io mode.
1151 * Many callers don't care about the return value.
1153 static int commit(struct pool *pool)
1157 if (get_pool_mode(pool) >= PM_READ_ONLY)
1160 r = dm_pool_commit_metadata(pool->pmd);
1162 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1164 check_for_space(pool);
1169 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1171 unsigned long flags;
1173 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1174 DMWARN("%s: reached low water mark for data device: sending event.",
1175 dm_device_name(pool->pool_md));
1176 spin_lock_irqsave(&pool->lock, flags);
1177 pool->low_water_triggered = true;
1178 spin_unlock_irqrestore(&pool->lock, flags);
1179 dm_table_event(pool->ti->table);
1183 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1186 dm_block_t free_blocks;
1187 struct pool *pool = tc->pool;
1189 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1192 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1194 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1198 check_low_water_mark(pool, free_blocks);
1202 * Try to commit to see if that will free up some
1209 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1211 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1216 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1221 r = dm_pool_alloc_data_block(pool->pmd, result);
1223 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1231 * If we have run out of space, queue bios until the device is
1232 * resumed, presumably after having been reloaded with more space.
1234 static void retry_on_resume(struct bio *bio)
1236 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1237 struct thin_c *tc = h->tc;
1238 unsigned long flags;
1240 spin_lock_irqsave(&tc->lock, flags);
1241 bio_list_add(&tc->retry_on_resume_list, bio);
1242 spin_unlock_irqrestore(&tc->lock, flags);
1245 static int should_error_unserviceable_bio(struct pool *pool)
1247 enum pool_mode m = get_pool_mode(pool);
1251 /* Shouldn't get here */
1252 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1255 case PM_OUT_OF_DATA_SPACE:
1256 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1262 /* Shouldn't get here */
1263 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1268 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1270 int error = should_error_unserviceable_bio(pool);
1273 bio_endio(bio, error);
1275 retry_on_resume(bio);
1278 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1281 struct bio_list bios;
1284 error = should_error_unserviceable_bio(pool);
1286 cell_error_with_code(pool, cell, error);
1290 bio_list_init(&bios);
1291 cell_release(pool, cell, &bios);
1293 while ((bio = bio_list_pop(&bios)))
1294 retry_on_resume(bio);
1297 static void process_discard_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1300 struct bio *bio = cell->holder;
1301 struct pool *pool = tc->pool;
1302 struct dm_bio_prison_cell *cell2;
1303 struct dm_cell_key key2;
1304 dm_block_t block = get_bio_block(tc, bio);
1305 struct dm_thin_lookup_result lookup_result;
1306 struct dm_thin_new_mapping *m;
1308 if (tc->requeue_mode) {
1309 cell_requeue(pool, cell);
1313 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1317 * Check nobody is fiddling with this pool block. This can
1318 * happen if someone's in the process of breaking sharing
1321 build_data_key(tc->td, lookup_result.block, &key2);
1322 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1323 cell_defer_no_holder(tc, cell);
1327 if (io_overlaps_block(pool, bio)) {
1329 * IO may still be going to the destination block. We must
1330 * quiesce before we can do the removal.
1332 m = get_next_mapping(pool);
1334 m->pass_discard = pool->pf.discard_passdown;
1335 m->definitely_not_shared = !lookup_result.shared;
1336 m->virt_block = block;
1337 m->data_block = lookup_result.block;
1342 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1343 pool->process_prepared_discard(m);
1346 inc_all_io_entry(pool, bio);
1347 cell_defer_no_holder(tc, cell);
1348 cell_defer_no_holder(tc, cell2);
1351 * The DM core makes sure that the discard doesn't span
1352 * a block boundary. So we submit the discard of a
1353 * partial block appropriately.
1355 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1356 remap_and_issue(tc, bio, lookup_result.block);
1364 * It isn't provisioned, just forget it.
1366 cell_defer_no_holder(tc, cell);
1371 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1373 cell_defer_no_holder(tc, cell);
1379 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1381 struct dm_bio_prison_cell *cell;
1382 struct dm_cell_key key;
1383 dm_block_t block = get_bio_block(tc, bio);
1385 build_virtual_key(tc->td, block, &key);
1386 if (bio_detain(tc->pool, &key, bio, &cell))
1389 process_discard_cell(tc, cell);
1392 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1393 struct dm_cell_key *key,
1394 struct dm_thin_lookup_result *lookup_result,
1395 struct dm_bio_prison_cell *cell)
1398 dm_block_t data_block;
1399 struct pool *pool = tc->pool;
1401 r = alloc_data_block(tc, &data_block);
1404 schedule_internal_copy(tc, block, lookup_result->block,
1405 data_block, cell, bio);
1409 retry_bios_on_resume(pool, cell);
1413 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1415 cell_error(pool, cell);
1420 static void __remap_and_issue_shared_cell(void *context,
1421 struct dm_bio_prison_cell *cell)
1423 struct remap_info *info = context;
1426 while ((bio = bio_list_pop(&cell->bios))) {
1427 if ((bio_data_dir(bio) == WRITE) ||
1428 (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)))
1429 bio_list_add(&info->defer_bios, bio);
1431 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1433 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1434 inc_all_io_entry(info->tc->pool, bio);
1435 bio_list_add(&info->issue_bios, bio);
1440 static void remap_and_issue_shared_cell(struct thin_c *tc,
1441 struct dm_bio_prison_cell *cell,
1445 struct remap_info info;
1448 bio_list_init(&info.defer_bios);
1449 bio_list_init(&info.issue_bios);
1451 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1454 while ((bio = bio_list_pop(&info.defer_bios)))
1455 thin_defer_bio(tc, bio);
1457 while ((bio = bio_list_pop(&info.issue_bios)))
1458 remap_and_issue(tc, bio, block);
1461 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1463 struct dm_thin_lookup_result *lookup_result,
1464 struct dm_bio_prison_cell *virt_cell)
1466 struct dm_bio_prison_cell *data_cell;
1467 struct pool *pool = tc->pool;
1468 struct dm_cell_key key;
1471 * If cell is already occupied, then sharing is already in the process
1472 * of being broken so we have nothing further to do here.
1474 build_data_key(tc->td, lookup_result->block, &key);
1475 if (bio_detain(pool, &key, bio, &data_cell)) {
1476 cell_defer_no_holder(tc, virt_cell);
1480 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1481 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1482 cell_defer_no_holder(tc, virt_cell);
1484 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1486 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1487 inc_all_io_entry(pool, bio);
1488 remap_and_issue(tc, bio, lookup_result->block);
1490 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1491 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1495 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1496 struct dm_bio_prison_cell *cell)
1499 dm_block_t data_block;
1500 struct pool *pool = tc->pool;
1503 * Remap empty bios (flushes) immediately, without provisioning.
1505 if (!bio->bi_iter.bi_size) {
1506 inc_all_io_entry(pool, bio);
1507 cell_defer_no_holder(tc, cell);
1509 remap_and_issue(tc, bio, 0);
1514 * Fill read bios with zeroes and complete them immediately.
1516 if (bio_data_dir(bio) == READ) {
1518 cell_defer_no_holder(tc, cell);
1523 r = alloc_data_block(tc, &data_block);
1527 schedule_external_copy(tc, block, data_block, cell, bio);
1529 schedule_zero(tc, block, data_block, cell, bio);
1533 retry_bios_on_resume(pool, cell);
1537 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1539 cell_error(pool, cell);
1544 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1547 struct pool *pool = tc->pool;
1548 struct bio *bio = cell->holder;
1549 dm_block_t block = get_bio_block(tc, bio);
1550 struct dm_thin_lookup_result lookup_result;
1552 if (tc->requeue_mode) {
1553 cell_requeue(pool, cell);
1557 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1560 if (lookup_result.shared)
1561 process_shared_bio(tc, bio, block, &lookup_result, cell);
1563 inc_all_io_entry(pool, bio);
1564 remap_and_issue(tc, bio, lookup_result.block);
1565 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1570 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1571 inc_all_io_entry(pool, bio);
1572 cell_defer_no_holder(tc, cell);
1574 if (bio_end_sector(bio) <= tc->origin_size)
1575 remap_to_origin_and_issue(tc, bio);
1577 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1579 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1580 remap_to_origin_and_issue(tc, bio);
1587 provision_block(tc, bio, block, cell);
1591 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1593 cell_defer_no_holder(tc, cell);
1599 static void process_bio(struct thin_c *tc, struct bio *bio)
1601 struct pool *pool = tc->pool;
1602 dm_block_t block = get_bio_block(tc, bio);
1603 struct dm_bio_prison_cell *cell;
1604 struct dm_cell_key key;
1607 * If cell is already occupied, then the block is already
1608 * being provisioned so we have nothing further to do here.
1610 build_virtual_key(tc->td, block, &key);
1611 if (bio_detain(pool, &key, bio, &cell))
1614 process_cell(tc, cell);
1617 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1618 struct dm_bio_prison_cell *cell)
1621 int rw = bio_data_dir(bio);
1622 dm_block_t block = get_bio_block(tc, bio);
1623 struct dm_thin_lookup_result lookup_result;
1625 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1628 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1629 handle_unserviceable_bio(tc->pool, bio);
1631 cell_defer_no_holder(tc, cell);
1633 inc_all_io_entry(tc->pool, bio);
1634 remap_and_issue(tc, bio, lookup_result.block);
1636 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1642 cell_defer_no_holder(tc, cell);
1644 handle_unserviceable_bio(tc->pool, bio);
1648 if (tc->origin_dev) {
1649 inc_all_io_entry(tc->pool, bio);
1650 remap_to_origin_and_issue(tc, bio);
1659 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1662 cell_defer_no_holder(tc, cell);
1668 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1670 __process_bio_read_only(tc, bio, NULL);
1673 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1675 __process_bio_read_only(tc, cell->holder, cell);
1678 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1683 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1688 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1690 cell_success(tc->pool, cell);
1693 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1695 cell_error(tc->pool, cell);
1699 * FIXME: should we also commit due to size of transaction, measured in
1702 static int need_commit_due_to_time(struct pool *pool)
1704 return !time_in_range(jiffies, pool->last_commit_jiffies,
1705 pool->last_commit_jiffies + COMMIT_PERIOD);
1708 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1709 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1711 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1713 struct rb_node **rbp, *parent;
1714 struct dm_thin_endio_hook *pbd;
1715 sector_t bi_sector = bio->bi_iter.bi_sector;
1717 rbp = &tc->sort_bio_list.rb_node;
1721 pbd = thin_pbd(parent);
1723 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1724 rbp = &(*rbp)->rb_left;
1726 rbp = &(*rbp)->rb_right;
1729 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1730 rb_link_node(&pbd->rb_node, parent, rbp);
1731 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1734 static void __extract_sorted_bios(struct thin_c *tc)
1736 struct rb_node *node;
1737 struct dm_thin_endio_hook *pbd;
1740 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1741 pbd = thin_pbd(node);
1742 bio = thin_bio(pbd);
1744 bio_list_add(&tc->deferred_bio_list, bio);
1745 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1748 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1751 static void __sort_thin_deferred_bios(struct thin_c *tc)
1754 struct bio_list bios;
1756 bio_list_init(&bios);
1757 bio_list_merge(&bios, &tc->deferred_bio_list);
1758 bio_list_init(&tc->deferred_bio_list);
1760 /* Sort deferred_bio_list using rb-tree */
1761 while ((bio = bio_list_pop(&bios)))
1762 __thin_bio_rb_add(tc, bio);
1765 * Transfer the sorted bios in sort_bio_list back to
1766 * deferred_bio_list to allow lockless submission of
1769 __extract_sorted_bios(tc);
1772 static void process_thin_deferred_bios(struct thin_c *tc)
1774 struct pool *pool = tc->pool;
1775 unsigned long flags;
1777 struct bio_list bios;
1778 struct blk_plug plug;
1781 if (tc->requeue_mode) {
1782 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
1786 bio_list_init(&bios);
1788 spin_lock_irqsave(&tc->lock, flags);
1790 if (bio_list_empty(&tc->deferred_bio_list)) {
1791 spin_unlock_irqrestore(&tc->lock, flags);
1795 __sort_thin_deferred_bios(tc);
1797 bio_list_merge(&bios, &tc->deferred_bio_list);
1798 bio_list_init(&tc->deferred_bio_list);
1800 spin_unlock_irqrestore(&tc->lock, flags);
1802 blk_start_plug(&plug);
1803 while ((bio = bio_list_pop(&bios))) {
1805 * If we've got no free new_mapping structs, and processing
1806 * this bio might require one, we pause until there are some
1807 * prepared mappings to process.
1809 if (ensure_next_mapping(pool)) {
1810 spin_lock_irqsave(&tc->lock, flags);
1811 bio_list_add(&tc->deferred_bio_list, bio);
1812 bio_list_merge(&tc->deferred_bio_list, &bios);
1813 spin_unlock_irqrestore(&tc->lock, flags);
1817 if (bio->bi_rw & REQ_DISCARD)
1818 pool->process_discard(tc, bio);
1820 pool->process_bio(tc, bio);
1822 if ((count++ & 127) == 0) {
1823 throttle_work_update(&pool->throttle);
1824 dm_pool_issue_prefetches(pool->pmd);
1827 blk_finish_plug(&plug);
1830 static int cmp_cells(const void *lhs, const void *rhs)
1832 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
1833 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
1835 BUG_ON(!lhs_cell->holder);
1836 BUG_ON(!rhs_cell->holder);
1838 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
1841 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
1847 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
1850 struct dm_bio_prison_cell *cell, *tmp;
1852 list_for_each_entry_safe(cell, tmp, cells, user_list) {
1853 if (count >= CELL_SORT_ARRAY_SIZE)
1856 pool->cell_sort_array[count++] = cell;
1857 list_del(&cell->user_list);
1860 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
1865 static void process_thin_deferred_cells(struct thin_c *tc)
1867 struct pool *pool = tc->pool;
1868 unsigned long flags;
1869 struct list_head cells;
1870 struct dm_bio_prison_cell *cell;
1871 unsigned i, j, count;
1873 INIT_LIST_HEAD(&cells);
1875 spin_lock_irqsave(&tc->lock, flags);
1876 list_splice_init(&tc->deferred_cells, &cells);
1877 spin_unlock_irqrestore(&tc->lock, flags);
1879 if (list_empty(&cells))
1883 count = sort_cells(tc->pool, &cells);
1885 for (i = 0; i < count; i++) {
1886 cell = pool->cell_sort_array[i];
1887 BUG_ON(!cell->holder);
1890 * If we've got no free new_mapping structs, and processing
1891 * this bio might require one, we pause until there are some
1892 * prepared mappings to process.
1894 if (ensure_next_mapping(pool)) {
1895 for (j = i; j < count; j++)
1896 list_add(&pool->cell_sort_array[j]->user_list, &cells);
1898 spin_lock_irqsave(&tc->lock, flags);
1899 list_splice(&cells, &tc->deferred_cells);
1900 spin_unlock_irqrestore(&tc->lock, flags);
1904 if (cell->holder->bi_rw & REQ_DISCARD)
1905 pool->process_discard_cell(tc, cell);
1907 pool->process_cell(tc, cell);
1909 } while (!list_empty(&cells));
1912 static void thin_get(struct thin_c *tc);
1913 static void thin_put(struct thin_c *tc);
1916 * We can't hold rcu_read_lock() around code that can block. So we
1917 * find a thin with the rcu lock held; bump a refcount; then drop
1920 static struct thin_c *get_first_thin(struct pool *pool)
1922 struct thin_c *tc = NULL;
1925 if (!list_empty(&pool->active_thins)) {
1926 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
1934 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
1936 struct thin_c *old_tc = tc;
1939 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
1951 static void process_deferred_bios(struct pool *pool)
1953 unsigned long flags;
1955 struct bio_list bios;
1958 tc = get_first_thin(pool);
1960 process_thin_deferred_cells(tc);
1961 process_thin_deferred_bios(tc);
1962 tc = get_next_thin(pool, tc);
1966 * If there are any deferred flush bios, we must commit
1967 * the metadata before issuing them.
1969 bio_list_init(&bios);
1970 spin_lock_irqsave(&pool->lock, flags);
1971 bio_list_merge(&bios, &pool->deferred_flush_bios);
1972 bio_list_init(&pool->deferred_flush_bios);
1973 spin_unlock_irqrestore(&pool->lock, flags);
1975 if (bio_list_empty(&bios) &&
1976 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
1980 while ((bio = bio_list_pop(&bios)))
1984 pool->last_commit_jiffies = jiffies;
1986 while ((bio = bio_list_pop(&bios)))
1987 generic_make_request(bio);
1990 static void do_worker(struct work_struct *ws)
1992 struct pool *pool = container_of(ws, struct pool, worker);
1994 throttle_work_start(&pool->throttle);
1995 dm_pool_issue_prefetches(pool->pmd);
1996 throttle_work_update(&pool->throttle);
1997 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1998 throttle_work_update(&pool->throttle);
1999 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2000 throttle_work_update(&pool->throttle);
2001 process_deferred_bios(pool);
2002 throttle_work_complete(&pool->throttle);
2006 * We want to commit periodically so that not too much
2007 * unwritten data builds up.
2009 static void do_waker(struct work_struct *ws)
2011 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2013 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2017 * We're holding onto IO to allow userland time to react. After the
2018 * timeout either the pool will have been resized (and thus back in
2019 * PM_WRITE mode), or we degrade to PM_READ_ONLY and start erroring IO.
2021 static void do_no_space_timeout(struct work_struct *ws)
2023 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2026 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space)
2027 set_pool_mode(pool, PM_READ_ONLY);
2030 /*----------------------------------------------------------------*/
2033 struct work_struct worker;
2034 struct completion complete;
2037 static struct pool_work *to_pool_work(struct work_struct *ws)
2039 return container_of(ws, struct pool_work, worker);
2042 static void pool_work_complete(struct pool_work *pw)
2044 complete(&pw->complete);
2047 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2048 void (*fn)(struct work_struct *))
2050 INIT_WORK_ONSTACK(&pw->worker, fn);
2051 init_completion(&pw->complete);
2052 queue_work(pool->wq, &pw->worker);
2053 wait_for_completion(&pw->complete);
2056 /*----------------------------------------------------------------*/
2058 struct noflush_work {
2059 struct pool_work pw;
2063 static struct noflush_work *to_noflush(struct work_struct *ws)
2065 return container_of(to_pool_work(ws), struct noflush_work, pw);
2068 static void do_noflush_start(struct work_struct *ws)
2070 struct noflush_work *w = to_noflush(ws);
2071 w->tc->requeue_mode = true;
2073 pool_work_complete(&w->pw);
2076 static void do_noflush_stop(struct work_struct *ws)
2078 struct noflush_work *w = to_noflush(ws);
2079 w->tc->requeue_mode = false;
2080 pool_work_complete(&w->pw);
2083 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2085 struct noflush_work w;
2088 pool_work_wait(&w.pw, tc->pool, fn);
2091 /*----------------------------------------------------------------*/
2093 static enum pool_mode get_pool_mode(struct pool *pool)
2095 return pool->pf.mode;
2098 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2100 dm_table_event(pool->ti->table);
2101 DMINFO("%s: switching pool to %s mode",
2102 dm_device_name(pool->pool_md), new_mode);
2105 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2107 struct pool_c *pt = pool->ti->private;
2108 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2109 enum pool_mode old_mode = get_pool_mode(pool);
2110 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2113 * Never allow the pool to transition to PM_WRITE mode if user
2114 * intervention is required to verify metadata and data consistency.
2116 if (new_mode == PM_WRITE && needs_check) {
2117 DMERR("%s: unable to switch pool to write mode until repaired.",
2118 dm_device_name(pool->pool_md));
2119 if (old_mode != new_mode)
2120 new_mode = old_mode;
2122 new_mode = PM_READ_ONLY;
2125 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2126 * not going to recover without a thin_repair. So we never let the
2127 * pool move out of the old mode.
2129 if (old_mode == PM_FAIL)
2130 new_mode = old_mode;
2134 if (old_mode != new_mode)
2135 notify_of_pool_mode_change(pool, "failure");
2136 dm_pool_metadata_read_only(pool->pmd);
2137 pool->process_bio = process_bio_fail;
2138 pool->process_discard = process_bio_fail;
2139 pool->process_cell = process_cell_fail;
2140 pool->process_discard_cell = process_cell_fail;
2141 pool->process_prepared_mapping = process_prepared_mapping_fail;
2142 pool->process_prepared_discard = process_prepared_discard_fail;
2144 error_retry_list(pool);
2148 if (old_mode != new_mode)
2149 notify_of_pool_mode_change(pool, "read-only");
2150 dm_pool_metadata_read_only(pool->pmd);
2151 pool->process_bio = process_bio_read_only;
2152 pool->process_discard = process_bio_success;
2153 pool->process_cell = process_cell_read_only;
2154 pool->process_discard_cell = process_cell_success;
2155 pool->process_prepared_mapping = process_prepared_mapping_fail;
2156 pool->process_prepared_discard = process_prepared_discard_passdown;
2158 error_retry_list(pool);
2161 case PM_OUT_OF_DATA_SPACE:
2163 * Ideally we'd never hit this state; the low water mark
2164 * would trigger userland to extend the pool before we
2165 * completely run out of data space. However, many small
2166 * IOs to unprovisioned space can consume data space at an
2167 * alarming rate. Adjust your low water mark if you're
2168 * frequently seeing this mode.
2170 if (old_mode != new_mode)
2171 notify_of_pool_mode_change(pool, "out-of-data-space");
2172 pool->process_bio = process_bio_read_only;
2173 pool->process_discard = process_discard_bio;
2174 pool->process_cell = process_cell_read_only;
2175 pool->process_discard_cell = process_discard_cell;
2176 pool->process_prepared_mapping = process_prepared_mapping;
2177 pool->process_prepared_discard = process_prepared_discard;
2179 if (!pool->pf.error_if_no_space && no_space_timeout)
2180 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2184 if (old_mode != new_mode)
2185 notify_of_pool_mode_change(pool, "write");
2186 dm_pool_metadata_read_write(pool->pmd);
2187 pool->process_bio = process_bio;
2188 pool->process_discard = process_discard_bio;
2189 pool->process_cell = process_cell;
2190 pool->process_discard_cell = process_discard_cell;
2191 pool->process_prepared_mapping = process_prepared_mapping;
2192 pool->process_prepared_discard = process_prepared_discard;
2196 pool->pf.mode = new_mode;
2198 * The pool mode may have changed, sync it so bind_control_target()
2199 * doesn't cause an unexpected mode transition on resume.
2201 pt->adjusted_pf.mode = new_mode;
2204 static void abort_transaction(struct pool *pool)
2206 const char *dev_name = dm_device_name(pool->pool_md);
2208 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2209 if (dm_pool_abort_metadata(pool->pmd)) {
2210 DMERR("%s: failed to abort metadata transaction", dev_name);
2211 set_pool_mode(pool, PM_FAIL);
2214 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2215 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2216 set_pool_mode(pool, PM_FAIL);
2220 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2222 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2223 dm_device_name(pool->pool_md), op, r);
2225 abort_transaction(pool);
2226 set_pool_mode(pool, PM_READ_ONLY);
2229 /*----------------------------------------------------------------*/
2232 * Mapping functions.
2236 * Called only while mapping a thin bio to hand it over to the workqueue.
2238 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2240 unsigned long flags;
2241 struct pool *pool = tc->pool;
2243 spin_lock_irqsave(&tc->lock, flags);
2244 bio_list_add(&tc->deferred_bio_list, bio);
2245 spin_unlock_irqrestore(&tc->lock, flags);
2250 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2252 struct pool *pool = tc->pool;
2254 throttle_lock(&pool->throttle);
2255 thin_defer_bio(tc, bio);
2256 throttle_unlock(&pool->throttle);
2259 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2261 unsigned long flags;
2262 struct pool *pool = tc->pool;
2264 throttle_lock(&pool->throttle);
2265 spin_lock_irqsave(&tc->lock, flags);
2266 list_add_tail(&cell->user_list, &tc->deferred_cells);
2267 spin_unlock_irqrestore(&tc->lock, flags);
2268 throttle_unlock(&pool->throttle);
2273 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2275 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2278 h->shared_read_entry = NULL;
2279 h->all_io_entry = NULL;
2280 h->overwrite_mapping = NULL;
2284 * Non-blocking function called from the thin target's map function.
2286 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2289 struct thin_c *tc = ti->private;
2290 dm_block_t block = get_bio_block(tc, bio);
2291 struct dm_thin_device *td = tc->td;
2292 struct dm_thin_lookup_result result;
2293 struct dm_bio_prison_cell *virt_cell, *data_cell;
2294 struct dm_cell_key key;
2296 thin_hook_bio(tc, bio);
2298 if (tc->requeue_mode) {
2299 bio_endio(bio, DM_ENDIO_REQUEUE);
2300 return DM_MAPIO_SUBMITTED;
2303 if (get_pool_mode(tc->pool) == PM_FAIL) {
2305 return DM_MAPIO_SUBMITTED;
2308 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
2309 thin_defer_bio_with_throttle(tc, bio);
2310 return DM_MAPIO_SUBMITTED;
2314 * We must hold the virtual cell before doing the lookup, otherwise
2315 * there's a race with discard.
2317 build_virtual_key(tc->td, block, &key);
2318 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2319 return DM_MAPIO_SUBMITTED;
2321 r = dm_thin_find_block(td, block, 0, &result);
2324 * Note that we defer readahead too.
2328 if (unlikely(result.shared)) {
2330 * We have a race condition here between the
2331 * result.shared value returned by the lookup and
2332 * snapshot creation, which may cause new
2335 * To avoid this always quiesce the origin before
2336 * taking the snap. You want to do this anyway to
2337 * ensure a consistent application view
2340 * More distant ancestors are irrelevant. The
2341 * shared flag will be set in their case.
2343 thin_defer_cell(tc, virt_cell);
2344 return DM_MAPIO_SUBMITTED;
2347 build_data_key(tc->td, result.block, &key);
2348 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2349 cell_defer_no_holder(tc, virt_cell);
2350 return DM_MAPIO_SUBMITTED;
2353 inc_all_io_entry(tc->pool, bio);
2354 cell_defer_no_holder(tc, data_cell);
2355 cell_defer_no_holder(tc, virt_cell);
2357 remap(tc, bio, result.block);
2358 return DM_MAPIO_REMAPPED;
2361 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
2363 * This block isn't provisioned, and we have no way
2366 handle_unserviceable_bio(tc->pool, bio);
2367 cell_defer_no_holder(tc, virt_cell);
2368 return DM_MAPIO_SUBMITTED;
2373 thin_defer_cell(tc, virt_cell);
2374 return DM_MAPIO_SUBMITTED;
2378 * Must always call bio_io_error on failure.
2379 * dm_thin_find_block can fail with -EINVAL if the
2380 * pool is switched to fail-io mode.
2383 cell_defer_no_holder(tc, virt_cell);
2384 return DM_MAPIO_SUBMITTED;
2388 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2390 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2391 struct request_queue *q;
2393 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2396 q = bdev_get_queue(pt->data_dev->bdev);
2397 return bdi_congested(&q->backing_dev_info, bdi_bits);
2400 static void requeue_bios(struct pool *pool)
2402 unsigned long flags;
2406 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2407 spin_lock_irqsave(&tc->lock, flags);
2408 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2409 bio_list_init(&tc->retry_on_resume_list);
2410 spin_unlock_irqrestore(&tc->lock, flags);
2415 /*----------------------------------------------------------------
2416 * Binding of control targets to a pool object
2417 *--------------------------------------------------------------*/
2418 static bool data_dev_supports_discard(struct pool_c *pt)
2420 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2422 return q && blk_queue_discard(q);
2425 static bool is_factor(sector_t block_size, uint32_t n)
2427 return !sector_div(block_size, n);
2431 * If discard_passdown was enabled verify that the data device
2432 * supports discards. Disable discard_passdown if not.
2434 static void disable_passdown_if_not_supported(struct pool_c *pt)
2436 struct pool *pool = pt->pool;
2437 struct block_device *data_bdev = pt->data_dev->bdev;
2438 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2439 sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
2440 const char *reason = NULL;
2441 char buf[BDEVNAME_SIZE];
2443 if (!pt->adjusted_pf.discard_passdown)
2446 if (!data_dev_supports_discard(pt))
2447 reason = "discard unsupported";
2449 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2450 reason = "max discard sectors smaller than a block";
2452 else if (data_limits->discard_granularity > block_size)
2453 reason = "discard granularity larger than a block";
2455 else if (!is_factor(block_size, data_limits->discard_granularity))
2456 reason = "discard granularity not a factor of block size";
2459 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2460 pt->adjusted_pf.discard_passdown = false;
2464 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2466 struct pool_c *pt = ti->private;
2469 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2471 enum pool_mode old_mode = get_pool_mode(pool);
2472 enum pool_mode new_mode = pt->adjusted_pf.mode;
2475 * Don't change the pool's mode until set_pool_mode() below.
2476 * Otherwise the pool's process_* function pointers may
2477 * not match the desired pool mode.
2479 pt->adjusted_pf.mode = old_mode;
2482 pool->pf = pt->adjusted_pf;
2483 pool->low_water_blocks = pt->low_water_blocks;
2485 set_pool_mode(pool, new_mode);
2490 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2496 /*----------------------------------------------------------------
2498 *--------------------------------------------------------------*/
2499 /* Initialize pool features. */
2500 static void pool_features_init(struct pool_features *pf)
2502 pf->mode = PM_WRITE;
2503 pf->zero_new_blocks = true;
2504 pf->discard_enabled = true;
2505 pf->discard_passdown = true;
2506 pf->error_if_no_space = false;
2509 static void __pool_destroy(struct pool *pool)
2511 __pool_table_remove(pool);
2513 if (dm_pool_metadata_close(pool->pmd) < 0)
2514 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2516 dm_bio_prison_destroy(pool->prison);
2517 dm_kcopyd_client_destroy(pool->copier);
2520 destroy_workqueue(pool->wq);
2522 if (pool->next_mapping)
2523 mempool_free(pool->next_mapping, pool->mapping_pool);
2524 mempool_destroy(pool->mapping_pool);
2525 dm_deferred_set_destroy(pool->shared_read_ds);
2526 dm_deferred_set_destroy(pool->all_io_ds);
2530 static struct kmem_cache *_new_mapping_cache;
2532 static struct pool *pool_create(struct mapped_device *pool_md,
2533 struct block_device *metadata_dev,
2534 unsigned long block_size,
2535 int read_only, char **error)
2540 struct dm_pool_metadata *pmd;
2541 bool format_device = read_only ? false : true;
2543 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2545 *error = "Error creating metadata object";
2546 return (struct pool *)pmd;
2549 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2551 *error = "Error allocating memory for pool";
2552 err_p = ERR_PTR(-ENOMEM);
2557 pool->sectors_per_block = block_size;
2558 if (block_size & (block_size - 1))
2559 pool->sectors_per_block_shift = -1;
2561 pool->sectors_per_block_shift = __ffs(block_size);
2562 pool->low_water_blocks = 0;
2563 pool_features_init(&pool->pf);
2564 pool->prison = dm_bio_prison_create();
2565 if (!pool->prison) {
2566 *error = "Error creating pool's bio prison";
2567 err_p = ERR_PTR(-ENOMEM);
2571 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2572 if (IS_ERR(pool->copier)) {
2573 r = PTR_ERR(pool->copier);
2574 *error = "Error creating pool's kcopyd client";
2576 goto bad_kcopyd_client;
2580 * Create singlethreaded workqueue that will service all devices
2581 * that use this metadata.
2583 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2585 *error = "Error creating pool's workqueue";
2586 err_p = ERR_PTR(-ENOMEM);
2590 throttle_init(&pool->throttle);
2591 INIT_WORK(&pool->worker, do_worker);
2592 INIT_DELAYED_WORK(&pool->waker, do_waker);
2593 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2594 spin_lock_init(&pool->lock);
2595 bio_list_init(&pool->deferred_flush_bios);
2596 INIT_LIST_HEAD(&pool->prepared_mappings);
2597 INIT_LIST_HEAD(&pool->prepared_discards);
2598 INIT_LIST_HEAD(&pool->active_thins);
2599 pool->low_water_triggered = false;
2600 pool->suspended = true;
2602 pool->shared_read_ds = dm_deferred_set_create();
2603 if (!pool->shared_read_ds) {
2604 *error = "Error creating pool's shared read deferred set";
2605 err_p = ERR_PTR(-ENOMEM);
2606 goto bad_shared_read_ds;
2609 pool->all_io_ds = dm_deferred_set_create();
2610 if (!pool->all_io_ds) {
2611 *error = "Error creating pool's all io deferred set";
2612 err_p = ERR_PTR(-ENOMEM);
2616 pool->next_mapping = NULL;
2617 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2618 _new_mapping_cache);
2619 if (!pool->mapping_pool) {
2620 *error = "Error creating pool's mapping mempool";
2621 err_p = ERR_PTR(-ENOMEM);
2622 goto bad_mapping_pool;
2625 pool->ref_count = 1;
2626 pool->last_commit_jiffies = jiffies;
2627 pool->pool_md = pool_md;
2628 pool->md_dev = metadata_dev;
2629 __pool_table_insert(pool);
2634 dm_deferred_set_destroy(pool->all_io_ds);
2636 dm_deferred_set_destroy(pool->shared_read_ds);
2638 destroy_workqueue(pool->wq);
2640 dm_kcopyd_client_destroy(pool->copier);
2642 dm_bio_prison_destroy(pool->prison);
2646 if (dm_pool_metadata_close(pmd))
2647 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2652 static void __pool_inc(struct pool *pool)
2654 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2658 static void __pool_dec(struct pool *pool)
2660 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2661 BUG_ON(!pool->ref_count);
2662 if (!--pool->ref_count)
2663 __pool_destroy(pool);
2666 static struct pool *__pool_find(struct mapped_device *pool_md,
2667 struct block_device *metadata_dev,
2668 unsigned long block_size, int read_only,
2669 char **error, int *created)
2671 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2674 if (pool->pool_md != pool_md) {
2675 *error = "metadata device already in use by a pool";
2676 return ERR_PTR(-EBUSY);
2681 pool = __pool_table_lookup(pool_md);
2683 if (pool->md_dev != metadata_dev) {
2684 *error = "different pool cannot replace a pool";
2685 return ERR_PTR(-EINVAL);
2690 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2698 /*----------------------------------------------------------------
2699 * Pool target methods
2700 *--------------------------------------------------------------*/
2701 static void pool_dtr(struct dm_target *ti)
2703 struct pool_c *pt = ti->private;
2705 mutex_lock(&dm_thin_pool_table.mutex);
2707 unbind_control_target(pt->pool, ti);
2708 __pool_dec(pt->pool);
2709 dm_put_device(ti, pt->metadata_dev);
2710 dm_put_device(ti, pt->data_dev);
2713 mutex_unlock(&dm_thin_pool_table.mutex);
2716 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2717 struct dm_target *ti)
2721 const char *arg_name;
2723 static struct dm_arg _args[] = {
2724 {0, 4, "Invalid number of pool feature arguments"},
2728 * No feature arguments supplied.
2733 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2737 while (argc && !r) {
2738 arg_name = dm_shift_arg(as);
2741 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2742 pf->zero_new_blocks = false;
2744 else if (!strcasecmp(arg_name, "ignore_discard"))
2745 pf->discard_enabled = false;
2747 else if (!strcasecmp(arg_name, "no_discard_passdown"))
2748 pf->discard_passdown = false;
2750 else if (!strcasecmp(arg_name, "read_only"))
2751 pf->mode = PM_READ_ONLY;
2753 else if (!strcasecmp(arg_name, "error_if_no_space"))
2754 pf->error_if_no_space = true;
2757 ti->error = "Unrecognised pool feature requested";
2766 static void metadata_low_callback(void *context)
2768 struct pool *pool = context;
2770 DMWARN("%s: reached low water mark for metadata device: sending event.",
2771 dm_device_name(pool->pool_md));
2773 dm_table_event(pool->ti->table);
2776 static sector_t get_dev_size(struct block_device *bdev)
2778 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
2781 static void warn_if_metadata_device_too_big(struct block_device *bdev)
2783 sector_t metadata_dev_size = get_dev_size(bdev);
2784 char buffer[BDEVNAME_SIZE];
2786 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
2787 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2788 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
2791 static sector_t get_metadata_dev_size(struct block_device *bdev)
2793 sector_t metadata_dev_size = get_dev_size(bdev);
2795 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
2796 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
2798 return metadata_dev_size;
2801 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
2803 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
2805 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
2807 return metadata_dev_size;
2811 * When a metadata threshold is crossed a dm event is triggered, and
2812 * userland should respond by growing the metadata device. We could let
2813 * userland set the threshold, like we do with the data threshold, but I'm
2814 * not sure they know enough to do this well.
2816 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
2819 * 4M is ample for all ops with the possible exception of thin
2820 * device deletion which is harmless if it fails (just retry the
2821 * delete after you've grown the device).
2823 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
2824 return min((dm_block_t)1024ULL /* 4M */, quarter);
2828 * thin-pool <metadata dev> <data dev>
2829 * <data block size (sectors)>
2830 * <low water mark (blocks)>
2831 * [<#feature args> [<arg>]*]
2833 * Optional feature arguments are:
2834 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
2835 * ignore_discard: disable discard
2836 * no_discard_passdown: don't pass discards down to the data device
2837 * read_only: Don't allow any changes to be made to the pool metadata.
2838 * error_if_no_space: error IOs, instead of queueing, if no space.
2840 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
2842 int r, pool_created = 0;
2845 struct pool_features pf;
2846 struct dm_arg_set as;
2847 struct dm_dev *data_dev;
2848 unsigned long block_size;
2849 dm_block_t low_water_blocks;
2850 struct dm_dev *metadata_dev;
2851 fmode_t metadata_mode;
2854 * FIXME Remove validation from scope of lock.
2856 mutex_lock(&dm_thin_pool_table.mutex);
2859 ti->error = "Invalid argument count";
2868 * Set default pool features.
2870 pool_features_init(&pf);
2872 dm_consume_args(&as, 4);
2873 r = parse_pool_features(&as, &pf, ti);
2877 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2878 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2880 ti->error = "Error opening metadata block device";
2883 warn_if_metadata_device_too_big(metadata_dev->bdev);
2885 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2887 ti->error = "Error getting data device";
2891 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2892 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2893 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2894 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2895 ti->error = "Invalid block size";
2900 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2901 ti->error = "Invalid low water mark";
2906 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2912 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2913 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2920 * 'pool_created' reflects whether this is the first table load.
2921 * Top level discard support is not allowed to be changed after
2922 * initial load. This would require a pool reload to trigger thin
2925 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2926 ti->error = "Discard support cannot be disabled once enabled";
2928 goto out_flags_changed;
2933 pt->metadata_dev = metadata_dev;
2934 pt->data_dev = data_dev;
2935 pt->low_water_blocks = low_water_blocks;
2936 pt->adjusted_pf = pt->requested_pf = pf;
2937 ti->num_flush_bios = 1;
2940 * Only need to enable discards if the pool should pass
2941 * them down to the data device. The thin device's discard
2942 * processing will cause mappings to be removed from the btree.
2944 ti->discard_zeroes_data_unsupported = true;
2945 if (pf.discard_enabled && pf.discard_passdown) {
2946 ti->num_discard_bios = 1;
2949 * Setting 'discards_supported' circumvents the normal
2950 * stacking of discard limits (this keeps the pool and
2951 * thin devices' discard limits consistent).
2953 ti->discards_supported = true;
2957 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2958 calc_metadata_threshold(pt),
2959 metadata_low_callback,
2964 pt->callbacks.congested_fn = pool_is_congested;
2965 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2967 mutex_unlock(&dm_thin_pool_table.mutex);
2976 dm_put_device(ti, data_dev);
2978 dm_put_device(ti, metadata_dev);
2980 mutex_unlock(&dm_thin_pool_table.mutex);
2985 static int pool_map(struct dm_target *ti, struct bio *bio)
2988 struct pool_c *pt = ti->private;
2989 struct pool *pool = pt->pool;
2990 unsigned long flags;
2993 * As this is a singleton target, ti->begin is always zero.
2995 spin_lock_irqsave(&pool->lock, flags);
2996 bio->bi_bdev = pt->data_dev->bdev;
2997 r = DM_MAPIO_REMAPPED;
2998 spin_unlock_irqrestore(&pool->lock, flags);
3003 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3006 struct pool_c *pt = ti->private;
3007 struct pool *pool = pt->pool;
3008 sector_t data_size = ti->len;
3009 dm_block_t sb_data_size;
3011 *need_commit = false;
3013 (void) sector_div(data_size, pool->sectors_per_block);
3015 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3017 DMERR("%s: failed to retrieve data device size",
3018 dm_device_name(pool->pool_md));
3022 if (data_size < sb_data_size) {
3023 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3024 dm_device_name(pool->pool_md),
3025 (unsigned long long)data_size, sb_data_size);
3028 } else if (data_size > sb_data_size) {
3029 if (dm_pool_metadata_needs_check(pool->pmd)) {
3030 DMERR("%s: unable to grow the data device until repaired.",
3031 dm_device_name(pool->pool_md));
3036 DMINFO("%s: growing the data device from %llu to %llu blocks",
3037 dm_device_name(pool->pool_md),
3038 sb_data_size, (unsigned long long)data_size);
3039 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3041 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3045 *need_commit = true;
3051 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3054 struct pool_c *pt = ti->private;
3055 struct pool *pool = pt->pool;
3056 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3058 *need_commit = false;
3060 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3062 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3064 DMERR("%s: failed to retrieve metadata device size",
3065 dm_device_name(pool->pool_md));
3069 if (metadata_dev_size < sb_metadata_dev_size) {
3070 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3071 dm_device_name(pool->pool_md),
3072 metadata_dev_size, sb_metadata_dev_size);
3075 } else if (metadata_dev_size > sb_metadata_dev_size) {
3076 if (dm_pool_metadata_needs_check(pool->pmd)) {
3077 DMERR("%s: unable to grow the metadata device until repaired.",
3078 dm_device_name(pool->pool_md));
3082 warn_if_metadata_device_too_big(pool->md_dev);
3083 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3084 dm_device_name(pool->pool_md),
3085 sb_metadata_dev_size, metadata_dev_size);
3086 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3088 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3092 *need_commit = true;
3099 * Retrieves the number of blocks of the data device from
3100 * the superblock and compares it to the actual device size,
3101 * thus resizing the data device in case it has grown.
3103 * This both copes with opening preallocated data devices in the ctr
3104 * being followed by a resume
3106 * calling the resume method individually after userspace has
3107 * grown the data device in reaction to a table event.
3109 static int pool_preresume(struct dm_target *ti)
3112 bool need_commit1, need_commit2;
3113 struct pool_c *pt = ti->private;
3114 struct pool *pool = pt->pool;
3117 * Take control of the pool object.
3119 r = bind_control_target(pool, ti);
3123 r = maybe_resize_data_dev(ti, &need_commit1);
3127 r = maybe_resize_metadata_dev(ti, &need_commit2);
3131 if (need_commit1 || need_commit2)
3132 (void) commit(pool);
3137 static void pool_suspend_active_thins(struct pool *pool)
3141 /* Suspend all active thin devices */
3142 tc = get_first_thin(pool);
3144 dm_internal_suspend_noflush(tc->thin_md);
3145 tc = get_next_thin(pool, tc);
3149 static void pool_resume_active_thins(struct pool *pool)
3153 /* Resume all active thin devices */
3154 tc = get_first_thin(pool);
3156 dm_internal_resume(tc->thin_md);
3157 tc = get_next_thin(pool, tc);
3161 static void pool_resume(struct dm_target *ti)
3163 struct pool_c *pt = ti->private;
3164 struct pool *pool = pt->pool;
3165 unsigned long flags;
3168 * Must requeue active_thins' bios and then resume
3169 * active_thins _before_ clearing 'suspend' flag.
3172 pool_resume_active_thins(pool);
3174 spin_lock_irqsave(&pool->lock, flags);
3175 pool->low_water_triggered = false;
3176 pool->suspended = false;
3177 spin_unlock_irqrestore(&pool->lock, flags);
3179 do_waker(&pool->waker.work);
3182 static void pool_presuspend(struct dm_target *ti)
3184 struct pool_c *pt = ti->private;
3185 struct pool *pool = pt->pool;
3186 unsigned long flags;
3188 spin_lock_irqsave(&pool->lock, flags);
3189 pool->suspended = true;
3190 spin_unlock_irqrestore(&pool->lock, flags);
3192 pool_suspend_active_thins(pool);
3195 static void pool_presuspend_undo(struct dm_target *ti)
3197 struct pool_c *pt = ti->private;
3198 struct pool *pool = pt->pool;
3199 unsigned long flags;
3201 pool_resume_active_thins(pool);
3203 spin_lock_irqsave(&pool->lock, flags);
3204 pool->suspended = false;
3205 spin_unlock_irqrestore(&pool->lock, flags);
3208 static void pool_postsuspend(struct dm_target *ti)
3210 struct pool_c *pt = ti->private;
3211 struct pool *pool = pt->pool;
3213 cancel_delayed_work(&pool->waker);
3214 cancel_delayed_work(&pool->no_space_timeout);
3215 flush_workqueue(pool->wq);
3216 (void) commit(pool);
3219 static int check_arg_count(unsigned argc, unsigned args_required)
3221 if (argc != args_required) {
3222 DMWARN("Message received with %u arguments instead of %u.",
3223 argc, args_required);
3230 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3232 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3233 *dev_id <= MAX_DEV_ID)
3237 DMWARN("Message received with invalid device id: %s", arg);
3242 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3247 r = check_arg_count(argc, 2);
3251 r = read_dev_id(argv[1], &dev_id, 1);
3255 r = dm_pool_create_thin(pool->pmd, dev_id);
3257 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3265 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3268 dm_thin_id origin_dev_id;
3271 r = check_arg_count(argc, 3);
3275 r = read_dev_id(argv[1], &dev_id, 1);
3279 r = read_dev_id(argv[2], &origin_dev_id, 1);
3283 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3285 DMWARN("Creation of new snapshot %s of device %s failed.",
3293 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3298 r = check_arg_count(argc, 2);
3302 r = read_dev_id(argv[1], &dev_id, 1);
3306 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3308 DMWARN("Deletion of thin device %s failed.", argv[1]);
3313 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3315 dm_thin_id old_id, new_id;
3318 r = check_arg_count(argc, 3);
3322 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3323 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3327 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3328 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3332 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3334 DMWARN("Failed to change transaction id from %s to %s.",
3342 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3346 r = check_arg_count(argc, 1);
3350 (void) commit(pool);
3352 r = dm_pool_reserve_metadata_snap(pool->pmd);
3354 DMWARN("reserve_metadata_snap message failed.");
3359 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3363 r = check_arg_count(argc, 1);
3367 r = dm_pool_release_metadata_snap(pool->pmd);
3369 DMWARN("release_metadata_snap message failed.");
3375 * Messages supported:
3376 * create_thin <dev_id>
3377 * create_snap <dev_id> <origin_id>
3379 * set_transaction_id <current_trans_id> <new_trans_id>
3380 * reserve_metadata_snap
3381 * release_metadata_snap
3383 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3386 struct pool_c *pt = ti->private;
3387 struct pool *pool = pt->pool;
3389 if (get_pool_mode(pool) >= PM_READ_ONLY) {
3390 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3391 dm_device_name(pool->pool_md));
3395 if (!strcasecmp(argv[0], "create_thin"))
3396 r = process_create_thin_mesg(argc, argv, pool);
3398 else if (!strcasecmp(argv[0], "create_snap"))
3399 r = process_create_snap_mesg(argc, argv, pool);
3401 else if (!strcasecmp(argv[0], "delete"))
3402 r = process_delete_mesg(argc, argv, pool);
3404 else if (!strcasecmp(argv[0], "set_transaction_id"))
3405 r = process_set_transaction_id_mesg(argc, argv, pool);
3407 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3408 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3410 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3411 r = process_release_metadata_snap_mesg(argc, argv, pool);
3414 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3417 (void) commit(pool);
3422 static void emit_flags(struct pool_features *pf, char *result,
3423 unsigned sz, unsigned maxlen)
3425 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3426 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3427 pf->error_if_no_space;
3428 DMEMIT("%u ", count);
3430 if (!pf->zero_new_blocks)
3431 DMEMIT("skip_block_zeroing ");
3433 if (!pf->discard_enabled)
3434 DMEMIT("ignore_discard ");
3436 if (!pf->discard_passdown)
3437 DMEMIT("no_discard_passdown ");
3439 if (pf->mode == PM_READ_ONLY)
3440 DMEMIT("read_only ");
3442 if (pf->error_if_no_space)
3443 DMEMIT("error_if_no_space ");
3448 * <transaction id> <used metadata sectors>/<total metadata sectors>
3449 * <used data sectors>/<total data sectors> <held metadata root>
3451 static void pool_status(struct dm_target *ti, status_type_t type,
3452 unsigned status_flags, char *result, unsigned maxlen)
3456 uint64_t transaction_id;
3457 dm_block_t nr_free_blocks_data;
3458 dm_block_t nr_free_blocks_metadata;
3459 dm_block_t nr_blocks_data;
3460 dm_block_t nr_blocks_metadata;
3461 dm_block_t held_root;
3462 char buf[BDEVNAME_SIZE];
3463 char buf2[BDEVNAME_SIZE];
3464 struct pool_c *pt = ti->private;
3465 struct pool *pool = pt->pool;
3468 case STATUSTYPE_INFO:
3469 if (get_pool_mode(pool) == PM_FAIL) {
3474 /* Commit to ensure statistics aren't out-of-date */
3475 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3476 (void) commit(pool);
3478 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3480 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3481 dm_device_name(pool->pool_md), r);
3485 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3487 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3488 dm_device_name(pool->pool_md), r);
3492 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3494 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3495 dm_device_name(pool->pool_md), r);
3499 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3501 DMERR("%s: dm_pool_get_free_block_count returned %d",
3502 dm_device_name(pool->pool_md), r);
3506 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3508 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3509 dm_device_name(pool->pool_md), r);
3513 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3515 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3516 dm_device_name(pool->pool_md), r);
3520 DMEMIT("%llu %llu/%llu %llu/%llu ",
3521 (unsigned long long)transaction_id,
3522 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3523 (unsigned long long)nr_blocks_metadata,
3524 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3525 (unsigned long long)nr_blocks_data);
3528 DMEMIT("%llu ", held_root);
3532 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3533 DMEMIT("out_of_data_space ");
3534 else if (pool->pf.mode == PM_READ_ONLY)
3539 if (!pool->pf.discard_enabled)
3540 DMEMIT("ignore_discard ");
3541 else if (pool->pf.discard_passdown)
3542 DMEMIT("discard_passdown ");
3544 DMEMIT("no_discard_passdown ");
3546 if (pool->pf.error_if_no_space)
3547 DMEMIT("error_if_no_space ");
3549 DMEMIT("queue_if_no_space ");
3553 case STATUSTYPE_TABLE:
3554 DMEMIT("%s %s %lu %llu ",
3555 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3556 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3557 (unsigned long)pool->sectors_per_block,
3558 (unsigned long long)pt->low_water_blocks);
3559 emit_flags(&pt->requested_pf, result, sz, maxlen);
3568 static int pool_iterate_devices(struct dm_target *ti,
3569 iterate_devices_callout_fn fn, void *data)
3571 struct pool_c *pt = ti->private;
3573 return fn(ti, pt->data_dev, 0, ti->len, data);
3576 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3577 struct bio_vec *biovec, int max_size)
3579 struct pool_c *pt = ti->private;
3580 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3582 if (!q->merge_bvec_fn)
3585 bvm->bi_bdev = pt->data_dev->bdev;
3587 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3590 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
3592 struct pool *pool = pt->pool;
3593 struct queue_limits *data_limits;
3595 limits->max_discard_sectors = pool->sectors_per_block;
3598 * discard_granularity is just a hint, and not enforced.
3600 if (pt->adjusted_pf.discard_passdown) {
3601 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
3602 limits->discard_granularity = max(data_limits->discard_granularity,
3603 pool->sectors_per_block << SECTOR_SHIFT);
3605 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
3608 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3610 struct pool_c *pt = ti->private;
3611 struct pool *pool = pt->pool;
3612 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3615 * If max_sectors is smaller than pool->sectors_per_block adjust it
3616 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3617 * This is especially beneficial when the pool's data device is a RAID
3618 * device that has a full stripe width that matches pool->sectors_per_block
3619 * -- because even though partial RAID stripe-sized IOs will be issued to a
3620 * single RAID stripe; when aggregated they will end on a full RAID stripe
3621 * boundary.. which avoids additional partial RAID stripe writes cascading
3623 if (limits->max_sectors < pool->sectors_per_block) {
3624 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3625 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3626 limits->max_sectors--;
3627 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3632 * If the system-determined stacked limits are compatible with the
3633 * pool's blocksize (io_opt is a factor) do not override them.
3635 if (io_opt_sectors < pool->sectors_per_block ||
3636 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3637 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3638 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3640 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3641 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3645 * pt->adjusted_pf is a staging area for the actual features to use.
3646 * They get transferred to the live pool in bind_control_target()
3647 * called from pool_preresume().
3649 if (!pt->adjusted_pf.discard_enabled) {
3651 * Must explicitly disallow stacking discard limits otherwise the
3652 * block layer will stack them if pool's data device has support.
3653 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3654 * user to see that, so make sure to set all discard limits to 0.
3656 limits->discard_granularity = 0;
3660 disable_passdown_if_not_supported(pt);
3662 set_discard_limits(pt, limits);
3665 static struct target_type pool_target = {
3666 .name = "thin-pool",
3667 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3668 DM_TARGET_IMMUTABLE,
3669 .version = {1, 14, 0},
3670 .module = THIS_MODULE,
3674 .presuspend = pool_presuspend,
3675 .presuspend_undo = pool_presuspend_undo,
3676 .postsuspend = pool_postsuspend,
3677 .preresume = pool_preresume,
3678 .resume = pool_resume,
3679 .message = pool_message,
3680 .status = pool_status,
3681 .merge = pool_merge,
3682 .iterate_devices = pool_iterate_devices,
3683 .io_hints = pool_io_hints,
3686 /*----------------------------------------------------------------
3687 * Thin target methods
3688 *--------------------------------------------------------------*/
3689 static void thin_get(struct thin_c *tc)
3691 atomic_inc(&tc->refcount);
3694 static void thin_put(struct thin_c *tc)
3696 if (atomic_dec_and_test(&tc->refcount))
3697 complete(&tc->can_destroy);
3700 static void thin_dtr(struct dm_target *ti)
3702 struct thin_c *tc = ti->private;
3703 unsigned long flags;
3705 spin_lock_irqsave(&tc->pool->lock, flags);
3706 list_del_rcu(&tc->list);
3707 spin_unlock_irqrestore(&tc->pool->lock, flags);
3711 wait_for_completion(&tc->can_destroy);
3713 mutex_lock(&dm_thin_pool_table.mutex);
3715 __pool_dec(tc->pool);
3716 dm_pool_close_thin_device(tc->td);
3717 dm_put_device(ti, tc->pool_dev);
3719 dm_put_device(ti, tc->origin_dev);
3722 mutex_unlock(&dm_thin_pool_table.mutex);
3726 * Thin target parameters:
3728 * <pool_dev> <dev_id> [origin_dev]
3730 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3731 * dev_id: the internal device identifier
3732 * origin_dev: a device external to the pool that should act as the origin
3734 * If the pool device has discards disabled, they get disabled for the thin
3737 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3741 struct dm_dev *pool_dev, *origin_dev;
3742 struct mapped_device *pool_md;
3743 unsigned long flags;
3745 mutex_lock(&dm_thin_pool_table.mutex);
3747 if (argc != 2 && argc != 3) {
3748 ti->error = "Invalid argument count";
3753 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3755 ti->error = "Out of memory";
3759 tc->thin_md = dm_table_get_md(ti->table);
3760 spin_lock_init(&tc->lock);
3761 INIT_LIST_HEAD(&tc->deferred_cells);
3762 bio_list_init(&tc->deferred_bio_list);
3763 bio_list_init(&tc->retry_on_resume_list);
3764 tc->sort_bio_list = RB_ROOT;
3767 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
3769 ti->error = "Error opening origin device";
3770 goto bad_origin_dev;
3772 tc->origin_dev = origin_dev;
3775 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
3777 ti->error = "Error opening pool device";
3780 tc->pool_dev = pool_dev;
3782 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
3783 ti->error = "Invalid device id";
3788 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
3790 ti->error = "Couldn't get pool mapped device";
3795 tc->pool = __pool_table_lookup(pool_md);
3797 ti->error = "Couldn't find pool object";
3799 goto bad_pool_lookup;
3801 __pool_inc(tc->pool);
3803 if (get_pool_mode(tc->pool) == PM_FAIL) {
3804 ti->error = "Couldn't open thin device, Pool is in fail mode";
3809 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
3811 ti->error = "Couldn't open thin internal device";
3815 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
3819 ti->num_flush_bios = 1;
3820 ti->flush_supported = true;
3821 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
3823 /* In case the pool supports discards, pass them on. */
3824 ti->discard_zeroes_data_unsupported = true;
3825 if (tc->pool->pf.discard_enabled) {
3826 ti->discards_supported = true;
3827 ti->num_discard_bios = 1;
3828 /* Discard bios must be split on a block boundary */
3829 ti->split_discard_bios = true;
3832 mutex_unlock(&dm_thin_pool_table.mutex);
3834 spin_lock_irqsave(&tc->pool->lock, flags);
3835 if (tc->pool->suspended) {
3836 spin_unlock_irqrestore(&tc->pool->lock, flags);
3837 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
3838 ti->error = "Unable to activate thin device while pool is suspended";
3842 atomic_set(&tc->refcount, 1);
3843 init_completion(&tc->can_destroy);
3844 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
3845 spin_unlock_irqrestore(&tc->pool->lock, flags);
3847 * This synchronize_rcu() call is needed here otherwise we risk a
3848 * wake_worker() call finding no bios to process (because the newly
3849 * added tc isn't yet visible). So this reduces latency since we
3850 * aren't then dependent on the periodic commit to wake_worker().
3859 dm_pool_close_thin_device(tc->td);
3861 __pool_dec(tc->pool);
3865 dm_put_device(ti, tc->pool_dev);
3868 dm_put_device(ti, tc->origin_dev);
3872 mutex_unlock(&dm_thin_pool_table.mutex);
3877 static int thin_map(struct dm_target *ti, struct bio *bio)
3879 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
3881 return thin_bio_map(ti, bio);
3884 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
3886 unsigned long flags;
3887 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
3888 struct list_head work;
3889 struct dm_thin_new_mapping *m, *tmp;
3890 struct pool *pool = h->tc->pool;
3892 if (h->shared_read_entry) {
3893 INIT_LIST_HEAD(&work);
3894 dm_deferred_entry_dec(h->shared_read_entry, &work);
3896 spin_lock_irqsave(&pool->lock, flags);
3897 list_for_each_entry_safe(m, tmp, &work, list) {
3899 __complete_mapping_preparation(m);
3901 spin_unlock_irqrestore(&pool->lock, flags);
3904 if (h->all_io_entry) {
3905 INIT_LIST_HEAD(&work);
3906 dm_deferred_entry_dec(h->all_io_entry, &work);
3907 if (!list_empty(&work)) {
3908 spin_lock_irqsave(&pool->lock, flags);
3909 list_for_each_entry_safe(m, tmp, &work, list)
3910 list_add_tail(&m->list, &pool->prepared_discards);
3911 spin_unlock_irqrestore(&pool->lock, flags);
3919 static void thin_presuspend(struct dm_target *ti)
3921 struct thin_c *tc = ti->private;
3923 if (dm_noflush_suspending(ti))
3924 noflush_work(tc, do_noflush_start);
3927 static void thin_postsuspend(struct dm_target *ti)
3929 struct thin_c *tc = ti->private;
3932 * The dm_noflush_suspending flag has been cleared by now, so
3933 * unfortunately we must always run this.
3935 noflush_work(tc, do_noflush_stop);
3938 static int thin_preresume(struct dm_target *ti)
3940 struct thin_c *tc = ti->private;
3943 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
3949 * <nr mapped sectors> <highest mapped sector>
3951 static void thin_status(struct dm_target *ti, status_type_t type,
3952 unsigned status_flags, char *result, unsigned maxlen)
3956 dm_block_t mapped, highest;
3957 char buf[BDEVNAME_SIZE];
3958 struct thin_c *tc = ti->private;
3960 if (get_pool_mode(tc->pool) == PM_FAIL) {
3969 case STATUSTYPE_INFO:
3970 r = dm_thin_get_mapped_count(tc->td, &mapped);
3972 DMERR("dm_thin_get_mapped_count returned %d", r);
3976 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
3978 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
3982 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
3984 DMEMIT("%llu", ((highest + 1) *
3985 tc->pool->sectors_per_block) - 1);
3990 case STATUSTYPE_TABLE:
3992 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
3993 (unsigned long) tc->dev_id);
3995 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4006 static int thin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
4007 struct bio_vec *biovec, int max_size)
4009 struct thin_c *tc = ti->private;
4010 struct request_queue *q = bdev_get_queue(tc->pool_dev->bdev);
4012 if (!q->merge_bvec_fn)
4015 bvm->bi_bdev = tc->pool_dev->bdev;
4016 bvm->bi_sector = dm_target_offset(ti, bvm->bi_sector);
4018 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
4021 static int thin_iterate_devices(struct dm_target *ti,
4022 iterate_devices_callout_fn fn, void *data)
4025 struct thin_c *tc = ti->private;
4026 struct pool *pool = tc->pool;
4029 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4030 * we follow a more convoluted path through to the pool's target.
4033 return 0; /* nothing is bound */
4035 blocks = pool->ti->len;
4036 (void) sector_div(blocks, pool->sectors_per_block);
4038 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4043 static struct target_type thin_target = {
4045 .version = {1, 14, 0},
4046 .module = THIS_MODULE,
4050 .end_io = thin_endio,
4051 .preresume = thin_preresume,
4052 .presuspend = thin_presuspend,
4053 .postsuspend = thin_postsuspend,
4054 .status = thin_status,
4055 .merge = thin_merge,
4056 .iterate_devices = thin_iterate_devices,
4059 /*----------------------------------------------------------------*/
4061 static int __init dm_thin_init(void)
4067 r = dm_register_target(&thin_target);
4071 r = dm_register_target(&pool_target);
4073 goto bad_pool_target;
4077 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4078 if (!_new_mapping_cache)
4079 goto bad_new_mapping_cache;
4083 bad_new_mapping_cache:
4084 dm_unregister_target(&pool_target);
4086 dm_unregister_target(&thin_target);
4091 static void dm_thin_exit(void)
4093 dm_unregister_target(&thin_target);
4094 dm_unregister_target(&pool_target);
4096 kmem_cache_destroy(_new_mapping_cache);
4099 module_init(dm_thin_init);
4100 module_exit(dm_thin_exit);
4102 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4103 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4105 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4106 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4107 MODULE_LICENSE("GPL");