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/log2.h>
15 #include <linux/list.h>
16 #include <linux/rculist.h>
17 #include <linux/init.h>
18 #include <linux/module.h>
19 #include <linux/slab.h>
20 #include <linux/sort.h>
21 #include <linux/rbtree.h>
23 #define DM_MSG_PREFIX "thin"
28 #define ENDIO_HOOK_POOL_SIZE 1024
29 #define MAPPING_POOL_SIZE 1024
30 #define COMMIT_PERIOD HZ
31 #define NO_SPACE_TIMEOUT_SECS 60
33 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
35 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
36 "A percentage of time allocated for copy on write");
39 * The block size of the device holding pool data must be
40 * between 64KB and 1GB.
42 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
43 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
46 * Device id is restricted to 24 bits.
48 #define MAX_DEV_ID ((1 << 24) - 1)
51 * How do we handle breaking sharing of data blocks?
52 * =================================================
54 * We use a standard copy-on-write btree to store the mappings for the
55 * devices (note I'm talking about copy-on-write of the metadata here, not
56 * the data). When you take an internal snapshot you clone the root node
57 * of the origin btree. After this there is no concept of an origin or a
58 * snapshot. They are just two device trees that happen to point to the
61 * When we get a write in we decide if it's to a shared data block using
62 * some timestamp magic. If it is, we have to break sharing.
64 * Let's say we write to a shared block in what was the origin. The
67 * i) plug io further to this physical block. (see bio_prison code).
69 * ii) quiesce any read io to that shared data block. Obviously
70 * including all devices that share this block. (see dm_deferred_set code)
72 * iii) copy the data block to a newly allocate block. This step can be
73 * missed out if the io covers the block. (schedule_copy).
75 * iv) insert the new mapping into the origin's btree
76 * (process_prepared_mapping). This act of inserting breaks some
77 * sharing of btree nodes between the two devices. Breaking sharing only
78 * effects the btree of that specific device. Btrees for the other
79 * devices that share the block never change. The btree for the origin
80 * device as it was after the last commit is untouched, ie. we're using
81 * persistent data structures in the functional programming sense.
83 * v) unplug io to this physical block, including the io that triggered
84 * the breaking of sharing.
86 * Steps (ii) and (iii) occur in parallel.
88 * The metadata _doesn't_ need to be committed before the io continues. We
89 * get away with this because the io is always written to a _new_ block.
90 * If there's a crash, then:
92 * - The origin mapping will point to the old origin block (the shared
93 * one). This will contain the data as it was before the io that triggered
94 * the breaking of sharing came in.
96 * - The snap mapping still points to the old block. As it would after
99 * The downside of this scheme is the timestamp magic isn't perfect, and
100 * will continue to think that data block in the snapshot device is shared
101 * even after the write to the origin has broken sharing. I suspect data
102 * blocks will typically be shared by many different devices, so we're
103 * breaking sharing n + 1 times, rather than n, where n is the number of
104 * devices that reference this data block. At the moment I think the
105 * benefits far, far outweigh the disadvantages.
108 /*----------------------------------------------------------------*/
113 static void build_data_key(struct dm_thin_device *td,
114 dm_block_t b, struct dm_cell_key *key)
117 key->dev = dm_thin_dev_id(td);
118 key->block_begin = b;
119 key->block_end = b + 1ULL;
122 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
123 struct dm_cell_key *key)
126 key->dev = dm_thin_dev_id(td);
127 key->block_begin = b;
128 key->block_end = b + 1ULL;
131 /*----------------------------------------------------------------*/
133 #define THROTTLE_THRESHOLD (1 * HZ)
136 struct rw_semaphore lock;
137 unsigned long threshold;
138 bool throttle_applied;
141 static void throttle_init(struct throttle *t)
143 init_rwsem(&t->lock);
144 t->throttle_applied = false;
147 static void throttle_work_start(struct throttle *t)
149 t->threshold = jiffies + THROTTLE_THRESHOLD;
152 static void throttle_work_update(struct throttle *t)
154 if (!t->throttle_applied && jiffies > t->threshold) {
155 down_write(&t->lock);
156 t->throttle_applied = true;
160 static void throttle_work_complete(struct throttle *t)
162 if (t->throttle_applied) {
163 t->throttle_applied = false;
168 static void throttle_lock(struct throttle *t)
173 static void throttle_unlock(struct throttle *t)
178 /*----------------------------------------------------------------*/
181 * A pool device ties together a metadata device and a data device. It
182 * also provides the interface for creating and destroying internal
185 struct dm_thin_new_mapping;
188 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
191 PM_WRITE, /* metadata may be changed */
192 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
193 PM_READ_ONLY, /* metadata may not be changed */
194 PM_FAIL, /* all I/O fails */
197 struct pool_features {
200 bool zero_new_blocks:1;
201 bool discard_enabled:1;
202 bool discard_passdown:1;
203 bool error_if_no_space:1;
207 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
208 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
209 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
211 #define CELL_SORT_ARRAY_SIZE 8192
214 struct list_head list;
215 struct dm_target *ti; /* Only set if a pool target is bound */
217 struct mapped_device *pool_md;
218 struct block_device *md_dev;
219 struct dm_pool_metadata *pmd;
221 dm_block_t low_water_blocks;
222 uint32_t sectors_per_block;
223 int sectors_per_block_shift;
225 struct pool_features pf;
226 bool low_water_triggered:1; /* A dm event has been sent */
229 struct dm_bio_prison *prison;
230 struct dm_kcopyd_client *copier;
232 struct workqueue_struct *wq;
233 struct throttle throttle;
234 struct work_struct worker;
235 struct delayed_work waker;
236 struct delayed_work no_space_timeout;
238 unsigned long last_commit_jiffies;
242 struct bio_list deferred_flush_bios;
243 struct list_head prepared_mappings;
244 struct list_head prepared_discards;
245 struct list_head active_thins;
247 struct dm_deferred_set *shared_read_ds;
248 struct dm_deferred_set *all_io_ds;
250 struct dm_thin_new_mapping *next_mapping;
251 mempool_t *mapping_pool;
253 process_bio_fn process_bio;
254 process_bio_fn process_discard;
256 process_cell_fn process_cell;
257 process_cell_fn process_discard_cell;
259 process_mapping_fn process_prepared_mapping;
260 process_mapping_fn process_prepared_discard;
262 struct dm_bio_prison_cell *cell_sort_array[CELL_SORT_ARRAY_SIZE];
265 static enum pool_mode get_pool_mode(struct pool *pool);
266 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
269 * Target context for a pool.
272 struct dm_target *ti;
274 struct dm_dev *data_dev;
275 struct dm_dev *metadata_dev;
276 struct dm_target_callbacks callbacks;
278 dm_block_t low_water_blocks;
279 struct pool_features requested_pf; /* Features requested during table load */
280 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
284 * Target context for a thin.
287 struct list_head list;
288 struct dm_dev *pool_dev;
289 struct dm_dev *origin_dev;
290 sector_t origin_size;
294 struct dm_thin_device *td;
295 struct mapped_device *thin_md;
299 struct list_head deferred_cells;
300 struct bio_list deferred_bio_list;
301 struct bio_list retry_on_resume_list;
302 struct rb_root sort_bio_list; /* sorted list of deferred bios */
305 * Ensures the thin is not destroyed until the worker has finished
306 * iterating the active_thins list.
309 struct completion can_destroy;
312 /*----------------------------------------------------------------*/
315 * wake_worker() is used when new work is queued and when pool_resume is
316 * ready to continue deferred IO processing.
318 static void wake_worker(struct pool *pool)
320 queue_work(pool->wq, &pool->worker);
323 /*----------------------------------------------------------------*/
325 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
326 struct dm_bio_prison_cell **cell_result)
329 struct dm_bio_prison_cell *cell_prealloc;
332 * Allocate a cell from the prison's mempool.
333 * This might block but it can't fail.
335 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
337 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
340 * We reused an old cell; we can get rid of
343 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
348 static void cell_release(struct pool *pool,
349 struct dm_bio_prison_cell *cell,
350 struct bio_list *bios)
352 dm_cell_release(pool->prison, cell, bios);
353 dm_bio_prison_free_cell(pool->prison, cell);
356 static void cell_visit_release(struct pool *pool,
357 void (*fn)(void *, struct dm_bio_prison_cell *),
359 struct dm_bio_prison_cell *cell)
361 dm_cell_visit_release(pool->prison, fn, context, cell);
362 dm_bio_prison_free_cell(pool->prison, cell);
365 static void cell_release_no_holder(struct pool *pool,
366 struct dm_bio_prison_cell *cell,
367 struct bio_list *bios)
369 dm_cell_release_no_holder(pool->prison, cell, bios);
370 dm_bio_prison_free_cell(pool->prison, cell);
373 static void cell_error_with_code(struct pool *pool,
374 struct dm_bio_prison_cell *cell, int error_code)
376 dm_cell_error(pool->prison, cell, error_code);
377 dm_bio_prison_free_cell(pool->prison, cell);
380 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
382 cell_error_with_code(pool, cell, -EIO);
385 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
387 cell_error_with_code(pool, cell, 0);
390 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
392 cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
395 /*----------------------------------------------------------------*/
398 * A global list of pools that uses a struct mapped_device as a key.
400 static struct dm_thin_pool_table {
402 struct list_head pools;
403 } dm_thin_pool_table;
405 static void pool_table_init(void)
407 mutex_init(&dm_thin_pool_table.mutex);
408 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
411 static void __pool_table_insert(struct pool *pool)
413 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
414 list_add(&pool->list, &dm_thin_pool_table.pools);
417 static void __pool_table_remove(struct pool *pool)
419 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
420 list_del(&pool->list);
423 static struct pool *__pool_table_lookup(struct mapped_device *md)
425 struct pool *pool = NULL, *tmp;
427 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
429 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
430 if (tmp->pool_md == md) {
439 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
441 struct pool *pool = NULL, *tmp;
443 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
445 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
446 if (tmp->md_dev == md_dev) {
455 /*----------------------------------------------------------------*/
457 struct dm_thin_endio_hook {
459 struct dm_deferred_entry *shared_read_entry;
460 struct dm_deferred_entry *all_io_entry;
461 struct dm_thin_new_mapping *overwrite_mapping;
462 struct rb_node rb_node;
465 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
467 bio_list_merge(bios, master);
468 bio_list_init(master);
471 static void error_bio_list(struct bio_list *bios, int error)
475 while ((bio = bio_list_pop(bios)))
476 bio_endio(bio, error);
479 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
481 struct bio_list bios;
484 bio_list_init(&bios);
486 spin_lock_irqsave(&tc->lock, flags);
487 __merge_bio_list(&bios, master);
488 spin_unlock_irqrestore(&tc->lock, flags);
490 error_bio_list(&bios, error);
493 static void requeue_deferred_cells(struct thin_c *tc)
495 struct pool *pool = tc->pool;
497 struct list_head cells;
498 struct dm_bio_prison_cell *cell, *tmp;
500 INIT_LIST_HEAD(&cells);
502 spin_lock_irqsave(&tc->lock, flags);
503 list_splice_init(&tc->deferred_cells, &cells);
504 spin_unlock_irqrestore(&tc->lock, flags);
506 list_for_each_entry_safe(cell, tmp, &cells, user_list)
507 cell_requeue(pool, cell);
510 static void requeue_io(struct thin_c *tc)
512 struct bio_list bios;
515 bio_list_init(&bios);
517 spin_lock_irqsave(&tc->lock, flags);
518 __merge_bio_list(&bios, &tc->deferred_bio_list);
519 __merge_bio_list(&bios, &tc->retry_on_resume_list);
520 spin_unlock_irqrestore(&tc->lock, flags);
522 error_bio_list(&bios, DM_ENDIO_REQUEUE);
523 requeue_deferred_cells(tc);
526 static void error_retry_list(struct pool *pool)
531 list_for_each_entry_rcu(tc, &pool->active_thins, list)
532 error_thin_bio_list(tc, &tc->retry_on_resume_list, -EIO);
537 * This section of code contains the logic for processing a thin device's IO.
538 * Much of the code depends on pool object resources (lists, workqueues, etc)
539 * but most is exclusively called from the thin target rather than the thin-pool
543 static bool block_size_is_power_of_two(struct pool *pool)
545 return pool->sectors_per_block_shift >= 0;
548 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
550 struct pool *pool = tc->pool;
551 sector_t block_nr = bio->bi_iter.bi_sector;
553 if (block_size_is_power_of_two(pool))
554 block_nr >>= pool->sectors_per_block_shift;
556 (void) sector_div(block_nr, pool->sectors_per_block);
561 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
563 struct pool *pool = tc->pool;
564 sector_t bi_sector = bio->bi_iter.bi_sector;
566 bio->bi_bdev = tc->pool_dev->bdev;
567 if (block_size_is_power_of_two(pool))
568 bio->bi_iter.bi_sector =
569 (block << pool->sectors_per_block_shift) |
570 (bi_sector & (pool->sectors_per_block - 1));
572 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
573 sector_div(bi_sector, pool->sectors_per_block);
576 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
578 bio->bi_bdev = tc->origin_dev->bdev;
581 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
583 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
584 dm_thin_changed_this_transaction(tc->td);
587 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
589 struct dm_thin_endio_hook *h;
591 if (bio->bi_rw & REQ_DISCARD)
594 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
595 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
598 static void issue(struct thin_c *tc, struct bio *bio)
600 struct pool *pool = tc->pool;
603 if (!bio_triggers_commit(tc, bio)) {
604 generic_make_request(bio);
609 * Complete bio with an error if earlier I/O caused changes to
610 * the metadata that can't be committed e.g, due to I/O errors
611 * on the metadata device.
613 if (dm_thin_aborted_changes(tc->td)) {
619 * Batch together any bios that trigger commits and then issue a
620 * single commit for them in process_deferred_bios().
622 spin_lock_irqsave(&pool->lock, flags);
623 bio_list_add(&pool->deferred_flush_bios, bio);
624 spin_unlock_irqrestore(&pool->lock, flags);
627 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
629 remap_to_origin(tc, bio);
633 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
636 remap(tc, bio, block);
640 /*----------------------------------------------------------------*/
643 * Bio endio functions.
645 struct dm_thin_new_mapping {
646 struct list_head list;
649 bool definitely_not_shared:1;
652 * Track quiescing, copying and zeroing preparation actions. When this
653 * counter hits zero the block is prepared and can be inserted into the
656 atomic_t prepare_actions;
660 dm_block_t virt_block;
661 dm_block_t data_block;
662 struct dm_bio_prison_cell *cell, *cell2;
665 * If the bio covers the whole area of a block then we can avoid
666 * zeroing or copying. Instead this bio is hooked. The bio will
667 * still be in the cell, so care has to be taken to avoid issuing
671 bio_end_io_t *saved_bi_end_io;
674 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
676 struct pool *pool = m->tc->pool;
678 if (atomic_dec_and_test(&m->prepare_actions)) {
679 list_add_tail(&m->list, &pool->prepared_mappings);
684 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
687 struct pool *pool = m->tc->pool;
689 spin_lock_irqsave(&pool->lock, flags);
690 __complete_mapping_preparation(m);
691 spin_unlock_irqrestore(&pool->lock, flags);
694 static void copy_complete(int read_err, unsigned long write_err, void *context)
696 struct dm_thin_new_mapping *m = context;
698 m->err = read_err || write_err ? -EIO : 0;
699 complete_mapping_preparation(m);
702 static void overwrite_endio(struct bio *bio, int err)
704 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
705 struct dm_thin_new_mapping *m = h->overwrite_mapping;
708 complete_mapping_preparation(m);
711 /*----------------------------------------------------------------*/
718 * Prepared mapping jobs.
722 * This sends the bios in the cell, except the original holder, back
723 * to the deferred_bios list.
725 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
727 struct pool *pool = tc->pool;
730 spin_lock_irqsave(&tc->lock, flags);
731 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
732 spin_unlock_irqrestore(&tc->lock, flags);
737 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
741 struct bio_list defer_bios;
742 struct bio_list issue_bios;
745 static void __inc_remap_and_issue_cell(void *context,
746 struct dm_bio_prison_cell *cell)
748 struct remap_info *info = context;
751 while ((bio = bio_list_pop(&cell->bios))) {
752 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA))
753 bio_list_add(&info->defer_bios, bio);
755 inc_all_io_entry(info->tc->pool, bio);
758 * We can't issue the bios with the bio prison lock
759 * held, so we add them to a list to issue on
760 * return from this function.
762 bio_list_add(&info->issue_bios, bio);
767 static void inc_remap_and_issue_cell(struct thin_c *tc,
768 struct dm_bio_prison_cell *cell,
772 struct remap_info info;
775 bio_list_init(&info.defer_bios);
776 bio_list_init(&info.issue_bios);
779 * We have to be careful to inc any bios we're about to issue
780 * before the cell is released, and avoid a race with new bios
781 * being added to the cell.
783 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
786 while ((bio = bio_list_pop(&info.defer_bios)))
787 thin_defer_bio(tc, bio);
789 while ((bio = bio_list_pop(&info.issue_bios)))
790 remap_and_issue(info.tc, bio, block);
793 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
796 m->bio->bi_end_io = m->saved_bi_end_io;
797 atomic_inc(&m->bio->bi_remaining);
799 cell_error(m->tc->pool, m->cell);
801 mempool_free(m, m->tc->pool->mapping_pool);
804 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
806 struct thin_c *tc = m->tc;
807 struct pool *pool = tc->pool;
813 bio->bi_end_io = m->saved_bi_end_io;
814 atomic_inc(&bio->bi_remaining);
818 cell_error(pool, m->cell);
823 * Commit the prepared block into the mapping btree.
824 * Any I/O for this block arriving after this point will get
825 * remapped to it directly.
827 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
829 metadata_operation_failed(pool, "dm_thin_insert_block", r);
830 cell_error(pool, m->cell);
835 * Release any bios held while the block was being provisioned.
836 * If we are processing a write bio that completely covers the block,
837 * we already processed it so can ignore it now when processing
838 * the bios in the cell.
841 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
844 inc_all_io_entry(tc->pool, m->cell->holder);
845 remap_and_issue(tc, m->cell->holder, m->data_block);
846 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
851 mempool_free(m, pool->mapping_pool);
854 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
856 struct thin_c *tc = m->tc;
858 bio_io_error(m->bio);
859 cell_defer_no_holder(tc, m->cell);
860 cell_defer_no_holder(tc, m->cell2);
861 mempool_free(m, tc->pool->mapping_pool);
864 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
866 struct thin_c *tc = m->tc;
868 inc_all_io_entry(tc->pool, m->bio);
869 cell_defer_no_holder(tc, m->cell);
870 cell_defer_no_holder(tc, m->cell2);
873 if (m->definitely_not_shared)
874 remap_and_issue(tc, m->bio, m->data_block);
877 if (dm_pool_block_is_used(tc->pool->pmd, m->data_block, &used) || used)
878 bio_endio(m->bio, 0);
880 remap_and_issue(tc, m->bio, m->data_block);
883 bio_endio(m->bio, 0);
885 mempool_free(m, tc->pool->mapping_pool);
888 static void process_prepared_discard(struct dm_thin_new_mapping *m)
891 struct thin_c *tc = m->tc;
893 r = dm_thin_remove_block(tc->td, m->virt_block);
895 DMERR_LIMIT("dm_thin_remove_block() failed");
897 process_prepared_discard_passdown(m);
900 static void process_prepared(struct pool *pool, struct list_head *head,
901 process_mapping_fn *fn)
904 struct list_head maps;
905 struct dm_thin_new_mapping *m, *tmp;
907 INIT_LIST_HEAD(&maps);
908 spin_lock_irqsave(&pool->lock, flags);
909 list_splice_init(head, &maps);
910 spin_unlock_irqrestore(&pool->lock, flags);
912 list_for_each_entry_safe(m, tmp, &maps, list)
919 static int io_overlaps_block(struct pool *pool, struct bio *bio)
921 return bio->bi_iter.bi_size ==
922 (pool->sectors_per_block << SECTOR_SHIFT);
925 static int io_overwrites_block(struct pool *pool, struct bio *bio)
927 return (bio_data_dir(bio) == WRITE) &&
928 io_overlaps_block(pool, bio);
931 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
934 *save = bio->bi_end_io;
938 static int ensure_next_mapping(struct pool *pool)
940 if (pool->next_mapping)
943 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
945 return pool->next_mapping ? 0 : -ENOMEM;
948 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
950 struct dm_thin_new_mapping *m = pool->next_mapping;
952 BUG_ON(!pool->next_mapping);
954 memset(m, 0, sizeof(struct dm_thin_new_mapping));
955 INIT_LIST_HEAD(&m->list);
958 pool->next_mapping = NULL;
963 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
964 sector_t begin, sector_t end)
967 struct dm_io_region to;
969 to.bdev = tc->pool_dev->bdev;
971 to.count = end - begin;
973 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
975 DMERR_LIMIT("dm_kcopyd_zero() failed");
976 copy_complete(1, 1, m);
980 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
981 dm_block_t data_block,
982 struct dm_thin_new_mapping *m)
984 struct pool *pool = tc->pool;
985 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
987 h->overwrite_mapping = m;
989 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
990 inc_all_io_entry(pool, bio);
991 remap_and_issue(tc, bio, data_block);
995 * A partial copy also needs to zero the uncopied region.
997 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
998 struct dm_dev *origin, dm_block_t data_origin,
999 dm_block_t data_dest,
1000 struct dm_bio_prison_cell *cell, struct bio *bio,
1004 struct pool *pool = tc->pool;
1005 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1008 m->virt_block = virt_block;
1009 m->data_block = data_dest;
1013 * quiesce action + copy action + an extra reference held for the
1014 * duration of this function (we may need to inc later for a
1017 atomic_set(&m->prepare_actions, 3);
1019 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1020 complete_mapping_preparation(m); /* already quiesced */
1023 * IO to pool_dev remaps to the pool target's data_dev.
1025 * If the whole block of data is being overwritten, we can issue the
1026 * bio immediately. Otherwise we use kcopyd to clone the data first.
1028 if (io_overwrites_block(pool, bio))
1029 remap_and_issue_overwrite(tc, bio, data_dest, m);
1031 struct dm_io_region from, to;
1033 from.bdev = origin->bdev;
1034 from.sector = data_origin * pool->sectors_per_block;
1037 to.bdev = tc->pool_dev->bdev;
1038 to.sector = data_dest * pool->sectors_per_block;
1041 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1042 0, copy_complete, m);
1044 DMERR_LIMIT("dm_kcopyd_copy() failed");
1045 copy_complete(1, 1, m);
1048 * We allow the zero to be issued, to simplify the
1049 * error path. Otherwise we'd need to start
1050 * worrying about decrementing the prepare_actions
1056 * Do we need to zero a tail region?
1058 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1059 atomic_inc(&m->prepare_actions);
1061 data_dest * pool->sectors_per_block + len,
1062 (data_dest + 1) * pool->sectors_per_block);
1066 complete_mapping_preparation(m); /* drop our ref */
1069 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1070 dm_block_t data_origin, dm_block_t data_dest,
1071 struct dm_bio_prison_cell *cell, struct bio *bio)
1073 schedule_copy(tc, virt_block, tc->pool_dev,
1074 data_origin, data_dest, cell, bio,
1075 tc->pool->sectors_per_block);
1078 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1079 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1082 struct pool *pool = tc->pool;
1083 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1085 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1087 m->virt_block = virt_block;
1088 m->data_block = data_block;
1092 * If the whole block of data is being overwritten or we are not
1093 * zeroing pre-existing data, we can issue the bio immediately.
1094 * Otherwise we use kcopyd to zero the data first.
1096 if (!pool->pf.zero_new_blocks)
1097 process_prepared_mapping(m);
1099 else if (io_overwrites_block(pool, bio))
1100 remap_and_issue_overwrite(tc, bio, data_block, m);
1104 data_block * pool->sectors_per_block,
1105 (data_block + 1) * pool->sectors_per_block);
1108 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1109 dm_block_t data_dest,
1110 struct dm_bio_prison_cell *cell, struct bio *bio)
1112 struct pool *pool = tc->pool;
1113 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1114 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1116 if (virt_block_end <= tc->origin_size)
1117 schedule_copy(tc, virt_block, tc->origin_dev,
1118 virt_block, data_dest, cell, bio,
1119 pool->sectors_per_block);
1121 else if (virt_block_begin < tc->origin_size)
1122 schedule_copy(tc, virt_block, tc->origin_dev,
1123 virt_block, data_dest, cell, bio,
1124 tc->origin_size - virt_block_begin);
1127 schedule_zero(tc, virt_block, data_dest, cell, bio);
1130 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1132 static void check_for_space(struct pool *pool)
1137 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1140 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1145 set_pool_mode(pool, PM_WRITE);
1149 * A non-zero return indicates read_only or fail_io mode.
1150 * Many callers don't care about the return value.
1152 static int commit(struct pool *pool)
1156 if (get_pool_mode(pool) >= PM_READ_ONLY)
1159 r = dm_pool_commit_metadata(pool->pmd);
1161 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1163 check_for_space(pool);
1168 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1170 unsigned long flags;
1172 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1173 DMWARN("%s: reached low water mark for data device: sending event.",
1174 dm_device_name(pool->pool_md));
1175 spin_lock_irqsave(&pool->lock, flags);
1176 pool->low_water_triggered = true;
1177 spin_unlock_irqrestore(&pool->lock, flags);
1178 dm_table_event(pool->ti->table);
1182 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1185 dm_block_t free_blocks;
1186 struct pool *pool = tc->pool;
1188 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1191 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1193 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1197 check_low_water_mark(pool, free_blocks);
1201 * Try to commit to see if that will free up some
1208 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1210 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1215 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1220 r = dm_pool_alloc_data_block(pool->pmd, result);
1222 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1230 * If we have run out of space, queue bios until the device is
1231 * resumed, presumably after having been reloaded with more space.
1233 static void retry_on_resume(struct bio *bio)
1235 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1236 struct thin_c *tc = h->tc;
1237 unsigned long flags;
1239 spin_lock_irqsave(&tc->lock, flags);
1240 bio_list_add(&tc->retry_on_resume_list, bio);
1241 spin_unlock_irqrestore(&tc->lock, flags);
1244 static int should_error_unserviceable_bio(struct pool *pool)
1246 enum pool_mode m = get_pool_mode(pool);
1250 /* Shouldn't get here */
1251 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1254 case PM_OUT_OF_DATA_SPACE:
1255 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1261 /* Shouldn't get here */
1262 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1267 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1269 int error = should_error_unserviceable_bio(pool);
1272 bio_endio(bio, error);
1274 retry_on_resume(bio);
1277 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1280 struct bio_list bios;
1283 error = should_error_unserviceable_bio(pool);
1285 cell_error_with_code(pool, cell, error);
1289 bio_list_init(&bios);
1290 cell_release(pool, cell, &bios);
1292 while ((bio = bio_list_pop(&bios)))
1293 retry_on_resume(bio);
1296 static void process_discard_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1299 struct bio *bio = cell->holder;
1300 struct pool *pool = tc->pool;
1301 struct dm_bio_prison_cell *cell2;
1302 struct dm_cell_key key2;
1303 dm_block_t block = get_bio_block(tc, bio);
1304 struct dm_thin_lookup_result lookup_result;
1305 struct dm_thin_new_mapping *m;
1307 if (tc->requeue_mode) {
1308 cell_requeue(pool, cell);
1312 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1316 * Check nobody is fiddling with this pool block. This can
1317 * happen if someone's in the process of breaking sharing
1320 build_data_key(tc->td, lookup_result.block, &key2);
1321 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1322 cell_defer_no_holder(tc, cell);
1326 if (io_overlaps_block(pool, bio)) {
1328 * IO may still be going to the destination block. We must
1329 * quiesce before we can do the removal.
1331 m = get_next_mapping(pool);
1333 m->pass_discard = pool->pf.discard_passdown;
1334 m->definitely_not_shared = !lookup_result.shared;
1335 m->virt_block = block;
1336 m->data_block = lookup_result.block;
1341 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1342 pool->process_prepared_discard(m);
1345 inc_all_io_entry(pool, bio);
1346 cell_defer_no_holder(tc, cell);
1347 cell_defer_no_holder(tc, cell2);
1350 * The DM core makes sure that the discard doesn't span
1351 * a block boundary. So we submit the discard of a
1352 * partial block appropriately.
1354 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1355 remap_and_issue(tc, bio, lookup_result.block);
1363 * It isn't provisioned, just forget it.
1365 cell_defer_no_holder(tc, cell);
1370 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1372 cell_defer_no_holder(tc, cell);
1378 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1380 struct dm_bio_prison_cell *cell;
1381 struct dm_cell_key key;
1382 dm_block_t block = get_bio_block(tc, bio);
1384 build_virtual_key(tc->td, block, &key);
1385 if (bio_detain(tc->pool, &key, bio, &cell))
1388 process_discard_cell(tc, cell);
1391 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1392 struct dm_cell_key *key,
1393 struct dm_thin_lookup_result *lookup_result,
1394 struct dm_bio_prison_cell *cell)
1397 dm_block_t data_block;
1398 struct pool *pool = tc->pool;
1400 r = alloc_data_block(tc, &data_block);
1403 schedule_internal_copy(tc, block, lookup_result->block,
1404 data_block, cell, bio);
1408 retry_bios_on_resume(pool, cell);
1412 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1414 cell_error(pool, cell);
1419 static void __remap_and_issue_shared_cell(void *context,
1420 struct dm_bio_prison_cell *cell)
1422 struct remap_info *info = context;
1425 while ((bio = bio_list_pop(&cell->bios))) {
1426 if ((bio_data_dir(bio) == WRITE) ||
1427 (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)))
1428 bio_list_add(&info->defer_bios, bio);
1430 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1432 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1433 inc_all_io_entry(info->tc->pool, bio);
1434 bio_list_add(&info->issue_bios, bio);
1439 static void remap_and_issue_shared_cell(struct thin_c *tc,
1440 struct dm_bio_prison_cell *cell,
1444 struct remap_info info;
1447 bio_list_init(&info.defer_bios);
1448 bio_list_init(&info.issue_bios);
1450 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1453 while ((bio = bio_list_pop(&info.defer_bios)))
1454 thin_defer_bio(tc, bio);
1456 while ((bio = bio_list_pop(&info.issue_bios)))
1457 remap_and_issue(tc, bio, block);
1460 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1462 struct dm_thin_lookup_result *lookup_result,
1463 struct dm_bio_prison_cell *virt_cell)
1465 struct dm_bio_prison_cell *data_cell;
1466 struct pool *pool = tc->pool;
1467 struct dm_cell_key key;
1470 * If cell is already occupied, then sharing is already in the process
1471 * of being broken so we have nothing further to do here.
1473 build_data_key(tc->td, lookup_result->block, &key);
1474 if (bio_detain(pool, &key, bio, &data_cell)) {
1475 cell_defer_no_holder(tc, virt_cell);
1479 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1480 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1481 cell_defer_no_holder(tc, virt_cell);
1483 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1485 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1486 inc_all_io_entry(pool, bio);
1487 remap_and_issue(tc, bio, lookup_result->block);
1489 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1490 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1494 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1495 struct dm_bio_prison_cell *cell)
1498 dm_block_t data_block;
1499 struct pool *pool = tc->pool;
1502 * Remap empty bios (flushes) immediately, without provisioning.
1504 if (!bio->bi_iter.bi_size) {
1505 inc_all_io_entry(pool, bio);
1506 cell_defer_no_holder(tc, cell);
1508 remap_and_issue(tc, bio, 0);
1513 * Fill read bios with zeroes and complete them immediately.
1515 if (bio_data_dir(bio) == READ) {
1517 cell_defer_no_holder(tc, cell);
1522 r = alloc_data_block(tc, &data_block);
1526 schedule_external_copy(tc, block, data_block, cell, bio);
1528 schedule_zero(tc, block, data_block, cell, bio);
1532 retry_bios_on_resume(pool, cell);
1536 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1538 cell_error(pool, cell);
1543 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1546 struct pool *pool = tc->pool;
1547 struct bio *bio = cell->holder;
1548 dm_block_t block = get_bio_block(tc, bio);
1549 struct dm_thin_lookup_result lookup_result;
1551 if (tc->requeue_mode) {
1552 cell_requeue(pool, cell);
1556 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1559 if (lookup_result.shared)
1560 process_shared_bio(tc, bio, block, &lookup_result, cell);
1562 inc_all_io_entry(pool, bio);
1563 remap_and_issue(tc, bio, lookup_result.block);
1564 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1569 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1570 inc_all_io_entry(pool, bio);
1571 cell_defer_no_holder(tc, cell);
1573 if (bio_end_sector(bio) <= tc->origin_size)
1574 remap_to_origin_and_issue(tc, bio);
1576 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1578 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1579 remap_to_origin_and_issue(tc, bio);
1586 provision_block(tc, bio, block, cell);
1590 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1592 cell_defer_no_holder(tc, cell);
1598 static void process_bio(struct thin_c *tc, struct bio *bio)
1600 struct pool *pool = tc->pool;
1601 dm_block_t block = get_bio_block(tc, bio);
1602 struct dm_bio_prison_cell *cell;
1603 struct dm_cell_key key;
1606 * If cell is already occupied, then the block is already
1607 * being provisioned so we have nothing further to do here.
1609 build_virtual_key(tc->td, block, &key);
1610 if (bio_detain(pool, &key, bio, &cell))
1613 process_cell(tc, cell);
1616 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1617 struct dm_bio_prison_cell *cell)
1620 int rw = bio_data_dir(bio);
1621 dm_block_t block = get_bio_block(tc, bio);
1622 struct dm_thin_lookup_result lookup_result;
1624 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1627 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1628 handle_unserviceable_bio(tc->pool, bio);
1630 cell_defer_no_holder(tc, cell);
1632 inc_all_io_entry(tc->pool, bio);
1633 remap_and_issue(tc, bio, lookup_result.block);
1635 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1641 cell_defer_no_holder(tc, cell);
1643 handle_unserviceable_bio(tc->pool, bio);
1647 if (tc->origin_dev) {
1648 inc_all_io_entry(tc->pool, bio);
1649 remap_to_origin_and_issue(tc, bio);
1658 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1661 cell_defer_no_holder(tc, cell);
1667 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1669 __process_bio_read_only(tc, bio, NULL);
1672 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1674 __process_bio_read_only(tc, cell->holder, cell);
1677 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1682 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1687 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1689 cell_success(tc->pool, cell);
1692 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1694 cell_error(tc->pool, cell);
1698 * FIXME: should we also commit due to size of transaction, measured in
1701 static int need_commit_due_to_time(struct pool *pool)
1703 return jiffies < pool->last_commit_jiffies ||
1704 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1707 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1708 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1710 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1712 struct rb_node **rbp, *parent;
1713 struct dm_thin_endio_hook *pbd;
1714 sector_t bi_sector = bio->bi_iter.bi_sector;
1716 rbp = &tc->sort_bio_list.rb_node;
1720 pbd = thin_pbd(parent);
1722 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1723 rbp = &(*rbp)->rb_left;
1725 rbp = &(*rbp)->rb_right;
1728 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1729 rb_link_node(&pbd->rb_node, parent, rbp);
1730 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1733 static void __extract_sorted_bios(struct thin_c *tc)
1735 struct rb_node *node;
1736 struct dm_thin_endio_hook *pbd;
1739 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1740 pbd = thin_pbd(node);
1741 bio = thin_bio(pbd);
1743 bio_list_add(&tc->deferred_bio_list, bio);
1744 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1747 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1750 static void __sort_thin_deferred_bios(struct thin_c *tc)
1753 struct bio_list bios;
1755 bio_list_init(&bios);
1756 bio_list_merge(&bios, &tc->deferred_bio_list);
1757 bio_list_init(&tc->deferred_bio_list);
1759 /* Sort deferred_bio_list using rb-tree */
1760 while ((bio = bio_list_pop(&bios)))
1761 __thin_bio_rb_add(tc, bio);
1764 * Transfer the sorted bios in sort_bio_list back to
1765 * deferred_bio_list to allow lockless submission of
1768 __extract_sorted_bios(tc);
1771 static void process_thin_deferred_bios(struct thin_c *tc)
1773 struct pool *pool = tc->pool;
1774 unsigned long flags;
1776 struct bio_list bios;
1777 struct blk_plug plug;
1780 if (tc->requeue_mode) {
1781 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
1785 bio_list_init(&bios);
1787 spin_lock_irqsave(&tc->lock, flags);
1789 if (bio_list_empty(&tc->deferred_bio_list)) {
1790 spin_unlock_irqrestore(&tc->lock, flags);
1794 __sort_thin_deferred_bios(tc);
1796 bio_list_merge(&bios, &tc->deferred_bio_list);
1797 bio_list_init(&tc->deferred_bio_list);
1799 spin_unlock_irqrestore(&tc->lock, flags);
1801 blk_start_plug(&plug);
1802 while ((bio = bio_list_pop(&bios))) {
1804 * If we've got no free new_mapping structs, and processing
1805 * this bio might require one, we pause until there are some
1806 * prepared mappings to process.
1808 if (ensure_next_mapping(pool)) {
1809 spin_lock_irqsave(&tc->lock, flags);
1810 bio_list_add(&tc->deferred_bio_list, bio);
1811 bio_list_merge(&tc->deferred_bio_list, &bios);
1812 spin_unlock_irqrestore(&tc->lock, flags);
1816 if (bio->bi_rw & REQ_DISCARD)
1817 pool->process_discard(tc, bio);
1819 pool->process_bio(tc, bio);
1821 if ((count++ & 127) == 0) {
1822 throttle_work_update(&pool->throttle);
1823 dm_pool_issue_prefetches(pool->pmd);
1826 blk_finish_plug(&plug);
1829 static int cmp_cells(const void *lhs, const void *rhs)
1831 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
1832 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
1834 BUG_ON(!lhs_cell->holder);
1835 BUG_ON(!rhs_cell->holder);
1837 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
1840 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
1846 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
1849 struct dm_bio_prison_cell *cell, *tmp;
1851 list_for_each_entry_safe(cell, tmp, cells, user_list) {
1852 if (count >= CELL_SORT_ARRAY_SIZE)
1855 pool->cell_sort_array[count++] = cell;
1856 list_del(&cell->user_list);
1859 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
1864 static void process_thin_deferred_cells(struct thin_c *tc)
1866 struct pool *pool = tc->pool;
1867 unsigned long flags;
1868 struct list_head cells;
1869 struct dm_bio_prison_cell *cell;
1870 unsigned i, j, count;
1872 INIT_LIST_HEAD(&cells);
1874 spin_lock_irqsave(&tc->lock, flags);
1875 list_splice_init(&tc->deferred_cells, &cells);
1876 spin_unlock_irqrestore(&tc->lock, flags);
1878 if (list_empty(&cells))
1882 count = sort_cells(tc->pool, &cells);
1884 for (i = 0; i < count; i++) {
1885 cell = pool->cell_sort_array[i];
1886 BUG_ON(!cell->holder);
1889 * If we've got no free new_mapping structs, and processing
1890 * this bio might require one, we pause until there are some
1891 * prepared mappings to process.
1893 if (ensure_next_mapping(pool)) {
1894 for (j = i; j < count; j++)
1895 list_add(&pool->cell_sort_array[j]->user_list, &cells);
1897 spin_lock_irqsave(&tc->lock, flags);
1898 list_splice(&cells, &tc->deferred_cells);
1899 spin_unlock_irqrestore(&tc->lock, flags);
1903 if (cell->holder->bi_rw & REQ_DISCARD)
1904 pool->process_discard_cell(tc, cell);
1906 pool->process_cell(tc, cell);
1908 } while (!list_empty(&cells));
1911 static void thin_get(struct thin_c *tc);
1912 static void thin_put(struct thin_c *tc);
1915 * We can't hold rcu_read_lock() around code that can block. So we
1916 * find a thin with the rcu lock held; bump a refcount; then drop
1919 static struct thin_c *get_first_thin(struct pool *pool)
1921 struct thin_c *tc = NULL;
1924 if (!list_empty(&pool->active_thins)) {
1925 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
1933 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
1935 struct thin_c *old_tc = tc;
1938 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
1950 static void process_deferred_bios(struct pool *pool)
1952 unsigned long flags;
1954 struct bio_list bios;
1957 tc = get_first_thin(pool);
1959 process_thin_deferred_cells(tc);
1960 process_thin_deferred_bios(tc);
1961 tc = get_next_thin(pool, tc);
1965 * If there are any deferred flush bios, we must commit
1966 * the metadata before issuing them.
1968 bio_list_init(&bios);
1969 spin_lock_irqsave(&pool->lock, flags);
1970 bio_list_merge(&bios, &pool->deferred_flush_bios);
1971 bio_list_init(&pool->deferred_flush_bios);
1972 spin_unlock_irqrestore(&pool->lock, flags);
1974 if (bio_list_empty(&bios) &&
1975 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
1979 while ((bio = bio_list_pop(&bios)))
1983 pool->last_commit_jiffies = jiffies;
1985 while ((bio = bio_list_pop(&bios)))
1986 generic_make_request(bio);
1989 static void do_worker(struct work_struct *ws)
1991 struct pool *pool = container_of(ws, struct pool, worker);
1993 throttle_work_start(&pool->throttle);
1994 dm_pool_issue_prefetches(pool->pmd);
1995 throttle_work_update(&pool->throttle);
1996 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1997 throttle_work_update(&pool->throttle);
1998 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1999 throttle_work_update(&pool->throttle);
2000 process_deferred_bios(pool);
2001 throttle_work_complete(&pool->throttle);
2005 * We want to commit periodically so that not too much
2006 * unwritten data builds up.
2008 static void do_waker(struct work_struct *ws)
2010 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2012 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2016 * We're holding onto IO to allow userland time to react. After the
2017 * timeout either the pool will have been resized (and thus back in
2018 * PM_WRITE mode), or we degrade to PM_READ_ONLY and start erroring IO.
2020 static void do_no_space_timeout(struct work_struct *ws)
2022 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2025 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space)
2026 set_pool_mode(pool, PM_READ_ONLY);
2029 /*----------------------------------------------------------------*/
2032 struct work_struct worker;
2033 struct completion complete;
2036 static struct pool_work *to_pool_work(struct work_struct *ws)
2038 return container_of(ws, struct pool_work, worker);
2041 static void pool_work_complete(struct pool_work *pw)
2043 complete(&pw->complete);
2046 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2047 void (*fn)(struct work_struct *))
2049 INIT_WORK_ONSTACK(&pw->worker, fn);
2050 init_completion(&pw->complete);
2051 queue_work(pool->wq, &pw->worker);
2052 wait_for_completion(&pw->complete);
2055 /*----------------------------------------------------------------*/
2057 struct noflush_work {
2058 struct pool_work pw;
2062 static struct noflush_work *to_noflush(struct work_struct *ws)
2064 return container_of(to_pool_work(ws), struct noflush_work, pw);
2067 static void do_noflush_start(struct work_struct *ws)
2069 struct noflush_work *w = to_noflush(ws);
2070 w->tc->requeue_mode = true;
2072 pool_work_complete(&w->pw);
2075 static void do_noflush_stop(struct work_struct *ws)
2077 struct noflush_work *w = to_noflush(ws);
2078 w->tc->requeue_mode = false;
2079 pool_work_complete(&w->pw);
2082 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2084 struct noflush_work w;
2087 pool_work_wait(&w.pw, tc->pool, fn);
2090 /*----------------------------------------------------------------*/
2092 static enum pool_mode get_pool_mode(struct pool *pool)
2094 return pool->pf.mode;
2097 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2099 dm_table_event(pool->ti->table);
2100 DMINFO("%s: switching pool to %s mode",
2101 dm_device_name(pool->pool_md), new_mode);
2104 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2106 struct pool_c *pt = pool->ti->private;
2107 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2108 enum pool_mode old_mode = get_pool_mode(pool);
2109 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2112 * Never allow the pool to transition to PM_WRITE mode if user
2113 * intervention is required to verify metadata and data consistency.
2115 if (new_mode == PM_WRITE && needs_check) {
2116 DMERR("%s: unable to switch pool to write mode until repaired.",
2117 dm_device_name(pool->pool_md));
2118 if (old_mode != new_mode)
2119 new_mode = old_mode;
2121 new_mode = PM_READ_ONLY;
2124 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2125 * not going to recover without a thin_repair. So we never let the
2126 * pool move out of the old mode.
2128 if (old_mode == PM_FAIL)
2129 new_mode = old_mode;
2133 if (old_mode != new_mode)
2134 notify_of_pool_mode_change(pool, "failure");
2135 dm_pool_metadata_read_only(pool->pmd);
2136 pool->process_bio = process_bio_fail;
2137 pool->process_discard = process_bio_fail;
2138 pool->process_cell = process_cell_fail;
2139 pool->process_discard_cell = process_cell_fail;
2140 pool->process_prepared_mapping = process_prepared_mapping_fail;
2141 pool->process_prepared_discard = process_prepared_discard_fail;
2143 error_retry_list(pool);
2147 if (old_mode != new_mode)
2148 notify_of_pool_mode_change(pool, "read-only");
2149 dm_pool_metadata_read_only(pool->pmd);
2150 pool->process_bio = process_bio_read_only;
2151 pool->process_discard = process_bio_success;
2152 pool->process_cell = process_cell_read_only;
2153 pool->process_discard_cell = process_cell_success;
2154 pool->process_prepared_mapping = process_prepared_mapping_fail;
2155 pool->process_prepared_discard = process_prepared_discard_passdown;
2157 error_retry_list(pool);
2160 case PM_OUT_OF_DATA_SPACE:
2162 * Ideally we'd never hit this state; the low water mark
2163 * would trigger userland to extend the pool before we
2164 * completely run out of data space. However, many small
2165 * IOs to unprovisioned space can consume data space at an
2166 * alarming rate. Adjust your low water mark if you're
2167 * frequently seeing this mode.
2169 if (old_mode != new_mode)
2170 notify_of_pool_mode_change(pool, "out-of-data-space");
2171 pool->process_bio = process_bio_read_only;
2172 pool->process_discard = process_discard_bio;
2173 pool->process_cell = process_cell_read_only;
2174 pool->process_discard_cell = process_discard_cell;
2175 pool->process_prepared_mapping = process_prepared_mapping;
2176 pool->process_prepared_discard = process_prepared_discard;
2178 if (!pool->pf.error_if_no_space && no_space_timeout)
2179 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2183 if (old_mode != new_mode)
2184 notify_of_pool_mode_change(pool, "write");
2185 dm_pool_metadata_read_write(pool->pmd);
2186 pool->process_bio = process_bio;
2187 pool->process_discard = process_discard_bio;
2188 pool->process_cell = process_cell;
2189 pool->process_discard_cell = process_discard_cell;
2190 pool->process_prepared_mapping = process_prepared_mapping;
2191 pool->process_prepared_discard = process_prepared_discard;
2195 pool->pf.mode = new_mode;
2197 * The pool mode may have changed, sync it so bind_control_target()
2198 * doesn't cause an unexpected mode transition on resume.
2200 pt->adjusted_pf.mode = new_mode;
2203 static void abort_transaction(struct pool *pool)
2205 const char *dev_name = dm_device_name(pool->pool_md);
2207 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2208 if (dm_pool_abort_metadata(pool->pmd)) {
2209 DMERR("%s: failed to abort metadata transaction", dev_name);
2210 set_pool_mode(pool, PM_FAIL);
2213 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2214 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2215 set_pool_mode(pool, PM_FAIL);
2219 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2221 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2222 dm_device_name(pool->pool_md), op, r);
2224 abort_transaction(pool);
2225 set_pool_mode(pool, PM_READ_ONLY);
2228 /*----------------------------------------------------------------*/
2231 * Mapping functions.
2235 * Called only while mapping a thin bio to hand it over to the workqueue.
2237 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2239 unsigned long flags;
2240 struct pool *pool = tc->pool;
2242 spin_lock_irqsave(&tc->lock, flags);
2243 bio_list_add(&tc->deferred_bio_list, bio);
2244 spin_unlock_irqrestore(&tc->lock, flags);
2249 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2251 struct pool *pool = tc->pool;
2253 throttle_lock(&pool->throttle);
2254 thin_defer_bio(tc, bio);
2255 throttle_unlock(&pool->throttle);
2258 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2260 unsigned long flags;
2261 struct pool *pool = tc->pool;
2263 throttle_lock(&pool->throttle);
2264 spin_lock_irqsave(&tc->lock, flags);
2265 list_add_tail(&cell->user_list, &tc->deferred_cells);
2266 spin_unlock_irqrestore(&tc->lock, flags);
2267 throttle_unlock(&pool->throttle);
2272 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2274 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2277 h->shared_read_entry = NULL;
2278 h->all_io_entry = NULL;
2279 h->overwrite_mapping = NULL;
2283 * Non-blocking function called from the thin target's map function.
2285 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2288 struct thin_c *tc = ti->private;
2289 dm_block_t block = get_bio_block(tc, bio);
2290 struct dm_thin_device *td = tc->td;
2291 struct dm_thin_lookup_result result;
2292 struct dm_bio_prison_cell *virt_cell, *data_cell;
2293 struct dm_cell_key key;
2295 thin_hook_bio(tc, bio);
2297 if (tc->requeue_mode) {
2298 bio_endio(bio, DM_ENDIO_REQUEUE);
2299 return DM_MAPIO_SUBMITTED;
2302 if (get_pool_mode(tc->pool) == PM_FAIL) {
2304 return DM_MAPIO_SUBMITTED;
2307 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
2308 thin_defer_bio_with_throttle(tc, bio);
2309 return DM_MAPIO_SUBMITTED;
2313 * We must hold the virtual cell before doing the lookup, otherwise
2314 * there's a race with discard.
2316 build_virtual_key(tc->td, block, &key);
2317 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2318 return DM_MAPIO_SUBMITTED;
2320 r = dm_thin_find_block(td, block, 0, &result);
2323 * Note that we defer readahead too.
2327 if (unlikely(result.shared)) {
2329 * We have a race condition here between the
2330 * result.shared value returned by the lookup and
2331 * snapshot creation, which may cause new
2334 * To avoid this always quiesce the origin before
2335 * taking the snap. You want to do this anyway to
2336 * ensure a consistent application view
2339 * More distant ancestors are irrelevant. The
2340 * shared flag will be set in their case.
2342 thin_defer_cell(tc, virt_cell);
2343 return DM_MAPIO_SUBMITTED;
2346 build_data_key(tc->td, result.block, &key);
2347 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2348 cell_defer_no_holder(tc, virt_cell);
2349 return DM_MAPIO_SUBMITTED;
2352 inc_all_io_entry(tc->pool, bio);
2353 cell_defer_no_holder(tc, data_cell);
2354 cell_defer_no_holder(tc, virt_cell);
2356 remap(tc, bio, result.block);
2357 return DM_MAPIO_REMAPPED;
2360 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
2362 * This block isn't provisioned, and we have no way
2365 handle_unserviceable_bio(tc->pool, bio);
2366 cell_defer_no_holder(tc, virt_cell);
2367 return DM_MAPIO_SUBMITTED;
2372 thin_defer_cell(tc, virt_cell);
2373 return DM_MAPIO_SUBMITTED;
2377 * Must always call bio_io_error on failure.
2378 * dm_thin_find_block can fail with -EINVAL if the
2379 * pool is switched to fail-io mode.
2382 cell_defer_no_holder(tc, virt_cell);
2383 return DM_MAPIO_SUBMITTED;
2387 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2389 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2390 struct request_queue *q;
2392 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2395 q = bdev_get_queue(pt->data_dev->bdev);
2396 return bdi_congested(&q->backing_dev_info, bdi_bits);
2399 static void requeue_bios(struct pool *pool)
2401 unsigned long flags;
2405 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2406 spin_lock_irqsave(&tc->lock, flags);
2407 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2408 bio_list_init(&tc->retry_on_resume_list);
2409 spin_unlock_irqrestore(&tc->lock, flags);
2414 /*----------------------------------------------------------------
2415 * Binding of control targets to a pool object
2416 *--------------------------------------------------------------*/
2417 static bool data_dev_supports_discard(struct pool_c *pt)
2419 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2421 return q && blk_queue_discard(q);
2424 static bool is_factor(sector_t block_size, uint32_t n)
2426 return !sector_div(block_size, n);
2430 * If discard_passdown was enabled verify that the data device
2431 * supports discards. Disable discard_passdown if not.
2433 static void disable_passdown_if_not_supported(struct pool_c *pt)
2435 struct pool *pool = pt->pool;
2436 struct block_device *data_bdev = pt->data_dev->bdev;
2437 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2438 sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
2439 const char *reason = NULL;
2440 char buf[BDEVNAME_SIZE];
2442 if (!pt->adjusted_pf.discard_passdown)
2445 if (!data_dev_supports_discard(pt))
2446 reason = "discard unsupported";
2448 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2449 reason = "max discard sectors smaller than a block";
2451 else if (data_limits->discard_granularity > block_size)
2452 reason = "discard granularity larger than a block";
2454 else if (!is_factor(block_size, data_limits->discard_granularity))
2455 reason = "discard granularity not a factor of block size";
2458 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2459 pt->adjusted_pf.discard_passdown = false;
2463 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2465 struct pool_c *pt = ti->private;
2468 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2470 enum pool_mode old_mode = get_pool_mode(pool);
2471 enum pool_mode new_mode = pt->adjusted_pf.mode;
2474 * Don't change the pool's mode until set_pool_mode() below.
2475 * Otherwise the pool's process_* function pointers may
2476 * not match the desired pool mode.
2478 pt->adjusted_pf.mode = old_mode;
2481 pool->pf = pt->adjusted_pf;
2482 pool->low_water_blocks = pt->low_water_blocks;
2484 set_pool_mode(pool, new_mode);
2489 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2495 /*----------------------------------------------------------------
2497 *--------------------------------------------------------------*/
2498 /* Initialize pool features. */
2499 static void pool_features_init(struct pool_features *pf)
2501 pf->mode = PM_WRITE;
2502 pf->zero_new_blocks = true;
2503 pf->discard_enabled = true;
2504 pf->discard_passdown = true;
2505 pf->error_if_no_space = false;
2508 static void __pool_destroy(struct pool *pool)
2510 __pool_table_remove(pool);
2512 if (dm_pool_metadata_close(pool->pmd) < 0)
2513 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2515 dm_bio_prison_destroy(pool->prison);
2516 dm_kcopyd_client_destroy(pool->copier);
2519 destroy_workqueue(pool->wq);
2521 if (pool->next_mapping)
2522 mempool_free(pool->next_mapping, pool->mapping_pool);
2523 mempool_destroy(pool->mapping_pool);
2524 dm_deferred_set_destroy(pool->shared_read_ds);
2525 dm_deferred_set_destroy(pool->all_io_ds);
2529 static struct kmem_cache *_new_mapping_cache;
2531 static struct pool *pool_create(struct mapped_device *pool_md,
2532 struct block_device *metadata_dev,
2533 unsigned long block_size,
2534 int read_only, char **error)
2539 struct dm_pool_metadata *pmd;
2540 bool format_device = read_only ? false : true;
2542 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2544 *error = "Error creating metadata object";
2545 return (struct pool *)pmd;
2548 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2550 *error = "Error allocating memory for pool";
2551 err_p = ERR_PTR(-ENOMEM);
2556 pool->sectors_per_block = block_size;
2557 if (block_size & (block_size - 1))
2558 pool->sectors_per_block_shift = -1;
2560 pool->sectors_per_block_shift = __ffs(block_size);
2561 pool->low_water_blocks = 0;
2562 pool_features_init(&pool->pf);
2563 pool->prison = dm_bio_prison_create();
2564 if (!pool->prison) {
2565 *error = "Error creating pool's bio prison";
2566 err_p = ERR_PTR(-ENOMEM);
2570 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2571 if (IS_ERR(pool->copier)) {
2572 r = PTR_ERR(pool->copier);
2573 *error = "Error creating pool's kcopyd client";
2575 goto bad_kcopyd_client;
2579 * Create singlethreaded workqueue that will service all devices
2580 * that use this metadata.
2582 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2584 *error = "Error creating pool's workqueue";
2585 err_p = ERR_PTR(-ENOMEM);
2589 throttle_init(&pool->throttle);
2590 INIT_WORK(&pool->worker, do_worker);
2591 INIT_DELAYED_WORK(&pool->waker, do_waker);
2592 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2593 spin_lock_init(&pool->lock);
2594 bio_list_init(&pool->deferred_flush_bios);
2595 INIT_LIST_HEAD(&pool->prepared_mappings);
2596 INIT_LIST_HEAD(&pool->prepared_discards);
2597 INIT_LIST_HEAD(&pool->active_thins);
2598 pool->low_water_triggered = false;
2599 pool->suspended = true;
2601 pool->shared_read_ds = dm_deferred_set_create();
2602 if (!pool->shared_read_ds) {
2603 *error = "Error creating pool's shared read deferred set";
2604 err_p = ERR_PTR(-ENOMEM);
2605 goto bad_shared_read_ds;
2608 pool->all_io_ds = dm_deferred_set_create();
2609 if (!pool->all_io_ds) {
2610 *error = "Error creating pool's all io deferred set";
2611 err_p = ERR_PTR(-ENOMEM);
2615 pool->next_mapping = NULL;
2616 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2617 _new_mapping_cache);
2618 if (!pool->mapping_pool) {
2619 *error = "Error creating pool's mapping mempool";
2620 err_p = ERR_PTR(-ENOMEM);
2621 goto bad_mapping_pool;
2624 pool->ref_count = 1;
2625 pool->last_commit_jiffies = jiffies;
2626 pool->pool_md = pool_md;
2627 pool->md_dev = metadata_dev;
2628 __pool_table_insert(pool);
2633 dm_deferred_set_destroy(pool->all_io_ds);
2635 dm_deferred_set_destroy(pool->shared_read_ds);
2637 destroy_workqueue(pool->wq);
2639 dm_kcopyd_client_destroy(pool->copier);
2641 dm_bio_prison_destroy(pool->prison);
2645 if (dm_pool_metadata_close(pmd))
2646 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2651 static void __pool_inc(struct pool *pool)
2653 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2657 static void __pool_dec(struct pool *pool)
2659 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2660 BUG_ON(!pool->ref_count);
2661 if (!--pool->ref_count)
2662 __pool_destroy(pool);
2665 static struct pool *__pool_find(struct mapped_device *pool_md,
2666 struct block_device *metadata_dev,
2667 unsigned long block_size, int read_only,
2668 char **error, int *created)
2670 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2673 if (pool->pool_md != pool_md) {
2674 *error = "metadata device already in use by a pool";
2675 return ERR_PTR(-EBUSY);
2680 pool = __pool_table_lookup(pool_md);
2682 if (pool->md_dev != metadata_dev) {
2683 *error = "different pool cannot replace a pool";
2684 return ERR_PTR(-EINVAL);
2689 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2697 /*----------------------------------------------------------------
2698 * Pool target methods
2699 *--------------------------------------------------------------*/
2700 static void pool_dtr(struct dm_target *ti)
2702 struct pool_c *pt = ti->private;
2704 mutex_lock(&dm_thin_pool_table.mutex);
2706 unbind_control_target(pt->pool, ti);
2707 __pool_dec(pt->pool);
2708 dm_put_device(ti, pt->metadata_dev);
2709 dm_put_device(ti, pt->data_dev);
2712 mutex_unlock(&dm_thin_pool_table.mutex);
2715 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2716 struct dm_target *ti)
2720 const char *arg_name;
2722 static struct dm_arg _args[] = {
2723 {0, 4, "Invalid number of pool feature arguments"},
2727 * No feature arguments supplied.
2732 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2736 while (argc && !r) {
2737 arg_name = dm_shift_arg(as);
2740 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2741 pf->zero_new_blocks = false;
2743 else if (!strcasecmp(arg_name, "ignore_discard"))
2744 pf->discard_enabled = false;
2746 else if (!strcasecmp(arg_name, "no_discard_passdown"))
2747 pf->discard_passdown = false;
2749 else if (!strcasecmp(arg_name, "read_only"))
2750 pf->mode = PM_READ_ONLY;
2752 else if (!strcasecmp(arg_name, "error_if_no_space"))
2753 pf->error_if_no_space = true;
2756 ti->error = "Unrecognised pool feature requested";
2765 static void metadata_low_callback(void *context)
2767 struct pool *pool = context;
2769 DMWARN("%s: reached low water mark for metadata device: sending event.",
2770 dm_device_name(pool->pool_md));
2772 dm_table_event(pool->ti->table);
2775 static sector_t get_dev_size(struct block_device *bdev)
2777 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
2780 static void warn_if_metadata_device_too_big(struct block_device *bdev)
2782 sector_t metadata_dev_size = get_dev_size(bdev);
2783 char buffer[BDEVNAME_SIZE];
2785 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
2786 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2787 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
2790 static sector_t get_metadata_dev_size(struct block_device *bdev)
2792 sector_t metadata_dev_size = get_dev_size(bdev);
2794 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
2795 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
2797 return metadata_dev_size;
2800 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
2802 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
2804 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
2806 return metadata_dev_size;
2810 * When a metadata threshold is crossed a dm event is triggered, and
2811 * userland should respond by growing the metadata device. We could let
2812 * userland set the threshold, like we do with the data threshold, but I'm
2813 * not sure they know enough to do this well.
2815 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
2818 * 4M is ample for all ops with the possible exception of thin
2819 * device deletion which is harmless if it fails (just retry the
2820 * delete after you've grown the device).
2822 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
2823 return min((dm_block_t)1024ULL /* 4M */, quarter);
2827 * thin-pool <metadata dev> <data dev>
2828 * <data block size (sectors)>
2829 * <low water mark (blocks)>
2830 * [<#feature args> [<arg>]*]
2832 * Optional feature arguments are:
2833 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
2834 * ignore_discard: disable discard
2835 * no_discard_passdown: don't pass discards down to the data device
2836 * read_only: Don't allow any changes to be made to the pool metadata.
2837 * error_if_no_space: error IOs, instead of queueing, if no space.
2839 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
2841 int r, pool_created = 0;
2844 struct pool_features pf;
2845 struct dm_arg_set as;
2846 struct dm_dev *data_dev;
2847 unsigned long block_size;
2848 dm_block_t low_water_blocks;
2849 struct dm_dev *metadata_dev;
2850 fmode_t metadata_mode;
2853 * FIXME Remove validation from scope of lock.
2855 mutex_lock(&dm_thin_pool_table.mutex);
2858 ti->error = "Invalid argument count";
2867 * Set default pool features.
2869 pool_features_init(&pf);
2871 dm_consume_args(&as, 4);
2872 r = parse_pool_features(&as, &pf, ti);
2876 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2877 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2879 ti->error = "Error opening metadata block device";
2882 warn_if_metadata_device_too_big(metadata_dev->bdev);
2884 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2886 ti->error = "Error getting data device";
2890 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2891 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2892 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2893 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2894 ti->error = "Invalid block size";
2899 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2900 ti->error = "Invalid low water mark";
2905 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2911 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2912 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2919 * 'pool_created' reflects whether this is the first table load.
2920 * Top level discard support is not allowed to be changed after
2921 * initial load. This would require a pool reload to trigger thin
2924 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2925 ti->error = "Discard support cannot be disabled once enabled";
2927 goto out_flags_changed;
2932 pt->metadata_dev = metadata_dev;
2933 pt->data_dev = data_dev;
2934 pt->low_water_blocks = low_water_blocks;
2935 pt->adjusted_pf = pt->requested_pf = pf;
2936 ti->num_flush_bios = 1;
2939 * Only need to enable discards if the pool should pass
2940 * them down to the data device. The thin device's discard
2941 * processing will cause mappings to be removed from the btree.
2943 ti->discard_zeroes_data_unsupported = true;
2944 if (pf.discard_enabled && pf.discard_passdown) {
2945 ti->num_discard_bios = 1;
2948 * Setting 'discards_supported' circumvents the normal
2949 * stacking of discard limits (this keeps the pool and
2950 * thin devices' discard limits consistent).
2952 ti->discards_supported = true;
2956 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2957 calc_metadata_threshold(pt),
2958 metadata_low_callback,
2963 pt->callbacks.congested_fn = pool_is_congested;
2964 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2966 mutex_unlock(&dm_thin_pool_table.mutex);
2975 dm_put_device(ti, data_dev);
2977 dm_put_device(ti, metadata_dev);
2979 mutex_unlock(&dm_thin_pool_table.mutex);
2984 static int pool_map(struct dm_target *ti, struct bio *bio)
2987 struct pool_c *pt = ti->private;
2988 struct pool *pool = pt->pool;
2989 unsigned long flags;
2992 * As this is a singleton target, ti->begin is always zero.
2994 spin_lock_irqsave(&pool->lock, flags);
2995 bio->bi_bdev = pt->data_dev->bdev;
2996 r = DM_MAPIO_REMAPPED;
2997 spin_unlock_irqrestore(&pool->lock, flags);
3002 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3005 struct pool_c *pt = ti->private;
3006 struct pool *pool = pt->pool;
3007 sector_t data_size = ti->len;
3008 dm_block_t sb_data_size;
3010 *need_commit = false;
3012 (void) sector_div(data_size, pool->sectors_per_block);
3014 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3016 DMERR("%s: failed to retrieve data device size",
3017 dm_device_name(pool->pool_md));
3021 if (data_size < sb_data_size) {
3022 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3023 dm_device_name(pool->pool_md),
3024 (unsigned long long)data_size, sb_data_size);
3027 } else if (data_size > sb_data_size) {
3028 if (dm_pool_metadata_needs_check(pool->pmd)) {
3029 DMERR("%s: unable to grow the data device until repaired.",
3030 dm_device_name(pool->pool_md));
3035 DMINFO("%s: growing the data device from %llu to %llu blocks",
3036 dm_device_name(pool->pool_md),
3037 sb_data_size, (unsigned long long)data_size);
3038 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3040 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3044 *need_commit = true;
3050 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3053 struct pool_c *pt = ti->private;
3054 struct pool *pool = pt->pool;
3055 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3057 *need_commit = false;
3059 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3061 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3063 DMERR("%s: failed to retrieve metadata device size",
3064 dm_device_name(pool->pool_md));
3068 if (metadata_dev_size < sb_metadata_dev_size) {
3069 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3070 dm_device_name(pool->pool_md),
3071 metadata_dev_size, sb_metadata_dev_size);
3074 } else if (metadata_dev_size > sb_metadata_dev_size) {
3075 if (dm_pool_metadata_needs_check(pool->pmd)) {
3076 DMERR("%s: unable to grow the metadata device until repaired.",
3077 dm_device_name(pool->pool_md));
3081 warn_if_metadata_device_too_big(pool->md_dev);
3082 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3083 dm_device_name(pool->pool_md),
3084 sb_metadata_dev_size, metadata_dev_size);
3085 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3087 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3091 *need_commit = true;
3098 * Retrieves the number of blocks of the data device from
3099 * the superblock and compares it to the actual device size,
3100 * thus resizing the data device in case it has grown.
3102 * This both copes with opening preallocated data devices in the ctr
3103 * being followed by a resume
3105 * calling the resume method individually after userspace has
3106 * grown the data device in reaction to a table event.
3108 static int pool_preresume(struct dm_target *ti)
3111 bool need_commit1, need_commit2;
3112 struct pool_c *pt = ti->private;
3113 struct pool *pool = pt->pool;
3116 * Take control of the pool object.
3118 r = bind_control_target(pool, ti);
3122 r = maybe_resize_data_dev(ti, &need_commit1);
3126 r = maybe_resize_metadata_dev(ti, &need_commit2);
3130 if (need_commit1 || need_commit2)
3131 (void) commit(pool);
3136 static void pool_suspend_active_thins(struct pool *pool)
3140 /* Suspend all active thin devices */
3141 tc = get_first_thin(pool);
3143 dm_internal_suspend_noflush(tc->thin_md);
3144 tc = get_next_thin(pool, tc);
3148 static void pool_resume_active_thins(struct pool *pool)
3152 /* Resume all active thin devices */
3153 tc = get_first_thin(pool);
3155 dm_internal_resume(tc->thin_md);
3156 tc = get_next_thin(pool, tc);
3160 static void pool_resume(struct dm_target *ti)
3162 struct pool_c *pt = ti->private;
3163 struct pool *pool = pt->pool;
3164 unsigned long flags;
3167 * Must requeue active_thins' bios and then resume
3168 * active_thins _before_ clearing 'suspend' flag.
3171 pool_resume_active_thins(pool);
3173 spin_lock_irqsave(&pool->lock, flags);
3174 pool->low_water_triggered = false;
3175 pool->suspended = false;
3176 spin_unlock_irqrestore(&pool->lock, flags);
3178 do_waker(&pool->waker.work);
3181 static void pool_presuspend(struct dm_target *ti)
3183 struct pool_c *pt = ti->private;
3184 struct pool *pool = pt->pool;
3185 unsigned long flags;
3187 spin_lock_irqsave(&pool->lock, flags);
3188 pool->suspended = true;
3189 spin_unlock_irqrestore(&pool->lock, flags);
3191 pool_suspend_active_thins(pool);
3194 static void pool_presuspend_undo(struct dm_target *ti)
3196 struct pool_c *pt = ti->private;
3197 struct pool *pool = pt->pool;
3198 unsigned long flags;
3200 pool_resume_active_thins(pool);
3202 spin_lock_irqsave(&pool->lock, flags);
3203 pool->suspended = false;
3204 spin_unlock_irqrestore(&pool->lock, flags);
3207 static void pool_postsuspend(struct dm_target *ti)
3209 struct pool_c *pt = ti->private;
3210 struct pool *pool = pt->pool;
3212 cancel_delayed_work(&pool->waker);
3213 cancel_delayed_work(&pool->no_space_timeout);
3214 flush_workqueue(pool->wq);
3215 (void) commit(pool);
3218 static int check_arg_count(unsigned argc, unsigned args_required)
3220 if (argc != args_required) {
3221 DMWARN("Message received with %u arguments instead of %u.",
3222 argc, args_required);
3229 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3231 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3232 *dev_id <= MAX_DEV_ID)
3236 DMWARN("Message received with invalid device id: %s", arg);
3241 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3246 r = check_arg_count(argc, 2);
3250 r = read_dev_id(argv[1], &dev_id, 1);
3254 r = dm_pool_create_thin(pool->pmd, dev_id);
3256 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3264 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3267 dm_thin_id origin_dev_id;
3270 r = check_arg_count(argc, 3);
3274 r = read_dev_id(argv[1], &dev_id, 1);
3278 r = read_dev_id(argv[2], &origin_dev_id, 1);
3282 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3284 DMWARN("Creation of new snapshot %s of device %s failed.",
3292 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3297 r = check_arg_count(argc, 2);
3301 r = read_dev_id(argv[1], &dev_id, 1);
3305 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3307 DMWARN("Deletion of thin device %s failed.", argv[1]);
3312 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3314 dm_thin_id old_id, new_id;
3317 r = check_arg_count(argc, 3);
3321 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3322 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3326 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3327 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3331 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3333 DMWARN("Failed to change transaction id from %s to %s.",
3341 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3345 r = check_arg_count(argc, 1);
3349 (void) commit(pool);
3351 r = dm_pool_reserve_metadata_snap(pool->pmd);
3353 DMWARN("reserve_metadata_snap message failed.");
3358 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3362 r = check_arg_count(argc, 1);
3366 r = dm_pool_release_metadata_snap(pool->pmd);
3368 DMWARN("release_metadata_snap message failed.");
3374 * Messages supported:
3375 * create_thin <dev_id>
3376 * create_snap <dev_id> <origin_id>
3378 * set_transaction_id <current_trans_id> <new_trans_id>
3379 * reserve_metadata_snap
3380 * release_metadata_snap
3382 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3385 struct pool_c *pt = ti->private;
3386 struct pool *pool = pt->pool;
3388 if (!strcasecmp(argv[0], "create_thin"))
3389 r = process_create_thin_mesg(argc, argv, pool);
3391 else if (!strcasecmp(argv[0], "create_snap"))
3392 r = process_create_snap_mesg(argc, argv, pool);
3394 else if (!strcasecmp(argv[0], "delete"))
3395 r = process_delete_mesg(argc, argv, pool);
3397 else if (!strcasecmp(argv[0], "set_transaction_id"))
3398 r = process_set_transaction_id_mesg(argc, argv, pool);
3400 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3401 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3403 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3404 r = process_release_metadata_snap_mesg(argc, argv, pool);
3407 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3410 (void) commit(pool);
3415 static void emit_flags(struct pool_features *pf, char *result,
3416 unsigned sz, unsigned maxlen)
3418 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3419 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3420 pf->error_if_no_space;
3421 DMEMIT("%u ", count);
3423 if (!pf->zero_new_blocks)
3424 DMEMIT("skip_block_zeroing ");
3426 if (!pf->discard_enabled)
3427 DMEMIT("ignore_discard ");
3429 if (!pf->discard_passdown)
3430 DMEMIT("no_discard_passdown ");
3432 if (pf->mode == PM_READ_ONLY)
3433 DMEMIT("read_only ");
3435 if (pf->error_if_no_space)
3436 DMEMIT("error_if_no_space ");
3441 * <transaction id> <used metadata sectors>/<total metadata sectors>
3442 * <used data sectors>/<total data sectors> <held metadata root>
3444 static void pool_status(struct dm_target *ti, status_type_t type,
3445 unsigned status_flags, char *result, unsigned maxlen)
3449 uint64_t transaction_id;
3450 dm_block_t nr_free_blocks_data;
3451 dm_block_t nr_free_blocks_metadata;
3452 dm_block_t nr_blocks_data;
3453 dm_block_t nr_blocks_metadata;
3454 dm_block_t held_root;
3455 char buf[BDEVNAME_SIZE];
3456 char buf2[BDEVNAME_SIZE];
3457 struct pool_c *pt = ti->private;
3458 struct pool *pool = pt->pool;
3461 case STATUSTYPE_INFO:
3462 if (get_pool_mode(pool) == PM_FAIL) {
3467 /* Commit to ensure statistics aren't out-of-date */
3468 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3469 (void) commit(pool);
3471 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3473 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3474 dm_device_name(pool->pool_md), r);
3478 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3480 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3481 dm_device_name(pool->pool_md), r);
3485 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3487 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3488 dm_device_name(pool->pool_md), r);
3492 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3494 DMERR("%s: dm_pool_get_free_block_count returned %d",
3495 dm_device_name(pool->pool_md), r);
3499 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3501 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3502 dm_device_name(pool->pool_md), r);
3506 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3508 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3509 dm_device_name(pool->pool_md), r);
3513 DMEMIT("%llu %llu/%llu %llu/%llu ",
3514 (unsigned long long)transaction_id,
3515 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3516 (unsigned long long)nr_blocks_metadata,
3517 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3518 (unsigned long long)nr_blocks_data);
3521 DMEMIT("%llu ", held_root);
3525 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3526 DMEMIT("out_of_data_space ");
3527 else if (pool->pf.mode == PM_READ_ONLY)
3532 if (!pool->pf.discard_enabled)
3533 DMEMIT("ignore_discard ");
3534 else if (pool->pf.discard_passdown)
3535 DMEMIT("discard_passdown ");
3537 DMEMIT("no_discard_passdown ");
3539 if (pool->pf.error_if_no_space)
3540 DMEMIT("error_if_no_space ");
3542 DMEMIT("queue_if_no_space ");
3546 case STATUSTYPE_TABLE:
3547 DMEMIT("%s %s %lu %llu ",
3548 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3549 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3550 (unsigned long)pool->sectors_per_block,
3551 (unsigned long long)pt->low_water_blocks);
3552 emit_flags(&pt->requested_pf, result, sz, maxlen);
3561 static int pool_iterate_devices(struct dm_target *ti,
3562 iterate_devices_callout_fn fn, void *data)
3564 struct pool_c *pt = ti->private;
3566 return fn(ti, pt->data_dev, 0, ti->len, data);
3569 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3570 struct bio_vec *biovec, int max_size)
3572 struct pool_c *pt = ti->private;
3573 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3575 if (!q->merge_bvec_fn)
3578 bvm->bi_bdev = pt->data_dev->bdev;
3580 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3583 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
3585 struct pool *pool = pt->pool;
3586 struct queue_limits *data_limits;
3588 limits->max_discard_sectors = pool->sectors_per_block;
3591 * discard_granularity is just a hint, and not enforced.
3593 if (pt->adjusted_pf.discard_passdown) {
3594 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
3595 limits->discard_granularity = max(data_limits->discard_granularity,
3596 pool->sectors_per_block << SECTOR_SHIFT);
3598 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
3601 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3603 struct pool_c *pt = ti->private;
3604 struct pool *pool = pt->pool;
3605 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3608 * If max_sectors is smaller than pool->sectors_per_block adjust it
3609 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3610 * This is especially beneficial when the pool's data device is a RAID
3611 * device that has a full stripe width that matches pool->sectors_per_block
3612 * -- because even though partial RAID stripe-sized IOs will be issued to a
3613 * single RAID stripe; when aggregated they will end on a full RAID stripe
3614 * boundary.. which avoids additional partial RAID stripe writes cascading
3616 if (limits->max_sectors < pool->sectors_per_block) {
3617 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3618 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3619 limits->max_sectors--;
3620 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3625 * If the system-determined stacked limits are compatible with the
3626 * pool's blocksize (io_opt is a factor) do not override them.
3628 if (io_opt_sectors < pool->sectors_per_block ||
3629 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3630 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3631 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3633 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3634 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3638 * pt->adjusted_pf is a staging area for the actual features to use.
3639 * They get transferred to the live pool in bind_control_target()
3640 * called from pool_preresume().
3642 if (!pt->adjusted_pf.discard_enabled) {
3644 * Must explicitly disallow stacking discard limits otherwise the
3645 * block layer will stack them if pool's data device has support.
3646 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3647 * user to see that, so make sure to set all discard limits to 0.
3649 limits->discard_granularity = 0;
3653 disable_passdown_if_not_supported(pt);
3655 set_discard_limits(pt, limits);
3658 static struct target_type pool_target = {
3659 .name = "thin-pool",
3660 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3661 DM_TARGET_IMMUTABLE,
3662 .version = {1, 14, 0},
3663 .module = THIS_MODULE,
3667 .presuspend = pool_presuspend,
3668 .presuspend_undo = pool_presuspend_undo,
3669 .postsuspend = pool_postsuspend,
3670 .preresume = pool_preresume,
3671 .resume = pool_resume,
3672 .message = pool_message,
3673 .status = pool_status,
3674 .merge = pool_merge,
3675 .iterate_devices = pool_iterate_devices,
3676 .io_hints = pool_io_hints,
3679 /*----------------------------------------------------------------
3680 * Thin target methods
3681 *--------------------------------------------------------------*/
3682 static void thin_get(struct thin_c *tc)
3684 atomic_inc(&tc->refcount);
3687 static void thin_put(struct thin_c *tc)
3689 if (atomic_dec_and_test(&tc->refcount))
3690 complete(&tc->can_destroy);
3693 static void thin_dtr(struct dm_target *ti)
3695 struct thin_c *tc = ti->private;
3696 unsigned long flags;
3698 spin_lock_irqsave(&tc->pool->lock, flags);
3699 list_del_rcu(&tc->list);
3700 spin_unlock_irqrestore(&tc->pool->lock, flags);
3704 wait_for_completion(&tc->can_destroy);
3706 mutex_lock(&dm_thin_pool_table.mutex);
3708 __pool_dec(tc->pool);
3709 dm_pool_close_thin_device(tc->td);
3710 dm_put_device(ti, tc->pool_dev);
3712 dm_put_device(ti, tc->origin_dev);
3715 mutex_unlock(&dm_thin_pool_table.mutex);
3719 * Thin target parameters:
3721 * <pool_dev> <dev_id> [origin_dev]
3723 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3724 * dev_id: the internal device identifier
3725 * origin_dev: a device external to the pool that should act as the origin
3727 * If the pool device has discards disabled, they get disabled for the thin
3730 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3734 struct dm_dev *pool_dev, *origin_dev;
3735 struct mapped_device *pool_md;
3736 unsigned long flags;
3738 mutex_lock(&dm_thin_pool_table.mutex);
3740 if (argc != 2 && argc != 3) {
3741 ti->error = "Invalid argument count";
3746 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3748 ti->error = "Out of memory";
3752 tc->thin_md = dm_table_get_md(ti->table);
3753 spin_lock_init(&tc->lock);
3754 INIT_LIST_HEAD(&tc->deferred_cells);
3755 bio_list_init(&tc->deferred_bio_list);
3756 bio_list_init(&tc->retry_on_resume_list);
3757 tc->sort_bio_list = RB_ROOT;
3760 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
3762 ti->error = "Error opening origin device";
3763 goto bad_origin_dev;
3765 tc->origin_dev = origin_dev;
3768 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
3770 ti->error = "Error opening pool device";
3773 tc->pool_dev = pool_dev;
3775 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
3776 ti->error = "Invalid device id";
3781 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
3783 ti->error = "Couldn't get pool mapped device";
3788 tc->pool = __pool_table_lookup(pool_md);
3790 ti->error = "Couldn't find pool object";
3792 goto bad_pool_lookup;
3794 __pool_inc(tc->pool);
3796 if (get_pool_mode(tc->pool) == PM_FAIL) {
3797 ti->error = "Couldn't open thin device, Pool is in fail mode";
3802 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
3804 ti->error = "Couldn't open thin internal device";
3808 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
3812 ti->num_flush_bios = 1;
3813 ti->flush_supported = true;
3814 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
3816 /* In case the pool supports discards, pass them on. */
3817 ti->discard_zeroes_data_unsupported = true;
3818 if (tc->pool->pf.discard_enabled) {
3819 ti->discards_supported = true;
3820 ti->num_discard_bios = 1;
3821 /* Discard bios must be split on a block boundary */
3822 ti->split_discard_bios = true;
3825 mutex_unlock(&dm_thin_pool_table.mutex);
3827 spin_lock_irqsave(&tc->pool->lock, flags);
3828 if (tc->pool->suspended) {
3829 spin_unlock_irqrestore(&tc->pool->lock, flags);
3830 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
3831 ti->error = "Unable to activate thin device while pool is suspended";
3835 atomic_set(&tc->refcount, 1);
3836 init_completion(&tc->can_destroy);
3837 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
3838 spin_unlock_irqrestore(&tc->pool->lock, flags);
3840 * This synchronize_rcu() call is needed here otherwise we risk a
3841 * wake_worker() call finding no bios to process (because the newly
3842 * added tc isn't yet visible). So this reduces latency since we
3843 * aren't then dependent on the periodic commit to wake_worker().
3852 dm_pool_close_thin_device(tc->td);
3854 __pool_dec(tc->pool);
3858 dm_put_device(ti, tc->pool_dev);
3861 dm_put_device(ti, tc->origin_dev);
3865 mutex_unlock(&dm_thin_pool_table.mutex);
3870 static int thin_map(struct dm_target *ti, struct bio *bio)
3872 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
3874 return thin_bio_map(ti, bio);
3877 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
3879 unsigned long flags;
3880 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
3881 struct list_head work;
3882 struct dm_thin_new_mapping *m, *tmp;
3883 struct pool *pool = h->tc->pool;
3885 if (h->shared_read_entry) {
3886 INIT_LIST_HEAD(&work);
3887 dm_deferred_entry_dec(h->shared_read_entry, &work);
3889 spin_lock_irqsave(&pool->lock, flags);
3890 list_for_each_entry_safe(m, tmp, &work, list) {
3892 __complete_mapping_preparation(m);
3894 spin_unlock_irqrestore(&pool->lock, flags);
3897 if (h->all_io_entry) {
3898 INIT_LIST_HEAD(&work);
3899 dm_deferred_entry_dec(h->all_io_entry, &work);
3900 if (!list_empty(&work)) {
3901 spin_lock_irqsave(&pool->lock, flags);
3902 list_for_each_entry_safe(m, tmp, &work, list)
3903 list_add_tail(&m->list, &pool->prepared_discards);
3904 spin_unlock_irqrestore(&pool->lock, flags);
3912 static void thin_presuspend(struct dm_target *ti)
3914 struct thin_c *tc = ti->private;
3916 if (dm_noflush_suspending(ti))
3917 noflush_work(tc, do_noflush_start);
3920 static void thin_postsuspend(struct dm_target *ti)
3922 struct thin_c *tc = ti->private;
3925 * The dm_noflush_suspending flag has been cleared by now, so
3926 * unfortunately we must always run this.
3928 noflush_work(tc, do_noflush_stop);
3931 static int thin_preresume(struct dm_target *ti)
3933 struct thin_c *tc = ti->private;
3936 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
3942 * <nr mapped sectors> <highest mapped sector>
3944 static void thin_status(struct dm_target *ti, status_type_t type,
3945 unsigned status_flags, char *result, unsigned maxlen)
3949 dm_block_t mapped, highest;
3950 char buf[BDEVNAME_SIZE];
3951 struct thin_c *tc = ti->private;
3953 if (get_pool_mode(tc->pool) == PM_FAIL) {
3962 case STATUSTYPE_INFO:
3963 r = dm_thin_get_mapped_count(tc->td, &mapped);
3965 DMERR("dm_thin_get_mapped_count returned %d", r);
3969 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
3971 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
3975 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
3977 DMEMIT("%llu", ((highest + 1) *
3978 tc->pool->sectors_per_block) - 1);
3983 case STATUSTYPE_TABLE:
3985 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
3986 (unsigned long) tc->dev_id);
3988 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
3999 static int thin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
4000 struct bio_vec *biovec, int max_size)
4002 struct thin_c *tc = ti->private;
4003 struct request_queue *q = bdev_get_queue(tc->pool_dev->bdev);
4005 if (!q->merge_bvec_fn)
4008 bvm->bi_bdev = tc->pool_dev->bdev;
4009 bvm->bi_sector = dm_target_offset(ti, bvm->bi_sector);
4011 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
4014 static int thin_iterate_devices(struct dm_target *ti,
4015 iterate_devices_callout_fn fn, void *data)
4018 struct thin_c *tc = ti->private;
4019 struct pool *pool = tc->pool;
4022 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4023 * we follow a more convoluted path through to the pool's target.
4026 return 0; /* nothing is bound */
4028 blocks = pool->ti->len;
4029 (void) sector_div(blocks, pool->sectors_per_block);
4031 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4036 static struct target_type thin_target = {
4038 .version = {1, 14, 0},
4039 .module = THIS_MODULE,
4043 .end_io = thin_endio,
4044 .preresume = thin_preresume,
4045 .presuspend = thin_presuspend,
4046 .postsuspend = thin_postsuspend,
4047 .status = thin_status,
4048 .merge = thin_merge,
4049 .iterate_devices = thin_iterate_devices,
4052 /*----------------------------------------------------------------*/
4054 static int __init dm_thin_init(void)
4060 r = dm_register_target(&thin_target);
4064 r = dm_register_target(&pool_target);
4066 goto bad_pool_target;
4070 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4071 if (!_new_mapping_cache)
4072 goto bad_new_mapping_cache;
4076 bad_new_mapping_cache:
4077 dm_unregister_target(&pool_target);
4079 dm_unregister_target(&thin_target);
4084 static void dm_thin_exit(void)
4086 dm_unregister_target(&thin_target);
4087 dm_unregister_target(&pool_target);
4089 kmem_cache_destroy(_new_mapping_cache);
4092 module_init(dm_thin_init);
4093 module_exit(dm_thin_exit);
4095 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4096 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4098 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4099 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4100 MODULE_LICENSE("GPL");