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);
1131 * A non-zero return indicates read_only or fail_io mode.
1132 * Many callers don't care about the return value.
1134 static int commit(struct pool *pool)
1138 if (get_pool_mode(pool) >= PM_READ_ONLY)
1141 r = dm_pool_commit_metadata(pool->pmd);
1143 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1148 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1150 unsigned long flags;
1152 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1153 DMWARN("%s: reached low water mark for data device: sending event.",
1154 dm_device_name(pool->pool_md));
1155 spin_lock_irqsave(&pool->lock, flags);
1156 pool->low_water_triggered = true;
1157 spin_unlock_irqrestore(&pool->lock, flags);
1158 dm_table_event(pool->ti->table);
1162 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1164 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1167 dm_block_t free_blocks;
1168 struct pool *pool = tc->pool;
1170 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1173 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1175 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1179 check_low_water_mark(pool, free_blocks);
1183 * Try to commit to see if that will free up some
1190 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1192 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1197 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1202 r = dm_pool_alloc_data_block(pool->pmd, result);
1204 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1212 * If we have run out of space, queue bios until the device is
1213 * resumed, presumably after having been reloaded with more space.
1215 static void retry_on_resume(struct bio *bio)
1217 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1218 struct thin_c *tc = h->tc;
1219 unsigned long flags;
1221 spin_lock_irqsave(&tc->lock, flags);
1222 bio_list_add(&tc->retry_on_resume_list, bio);
1223 spin_unlock_irqrestore(&tc->lock, flags);
1226 static int should_error_unserviceable_bio(struct pool *pool)
1228 enum pool_mode m = get_pool_mode(pool);
1232 /* Shouldn't get here */
1233 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1236 case PM_OUT_OF_DATA_SPACE:
1237 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1243 /* Shouldn't get here */
1244 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1249 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1251 int error = should_error_unserviceable_bio(pool);
1254 bio_endio(bio, error);
1256 retry_on_resume(bio);
1259 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1262 struct bio_list bios;
1265 error = should_error_unserviceable_bio(pool);
1267 cell_error_with_code(pool, cell, error);
1271 bio_list_init(&bios);
1272 cell_release(pool, cell, &bios);
1274 while ((bio = bio_list_pop(&bios)))
1275 retry_on_resume(bio);
1278 static void process_discard_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1281 struct bio *bio = cell->holder;
1282 struct pool *pool = tc->pool;
1283 struct dm_bio_prison_cell *cell2;
1284 struct dm_cell_key key2;
1285 dm_block_t block = get_bio_block(tc, bio);
1286 struct dm_thin_lookup_result lookup_result;
1287 struct dm_thin_new_mapping *m;
1289 if (tc->requeue_mode) {
1290 cell_requeue(pool, cell);
1294 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1298 * Check nobody is fiddling with this pool block. This can
1299 * happen if someone's in the process of breaking sharing
1302 build_data_key(tc->td, lookup_result.block, &key2);
1303 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1304 cell_defer_no_holder(tc, cell);
1308 if (io_overlaps_block(pool, bio)) {
1310 * IO may still be going to the destination block. We must
1311 * quiesce before we can do the removal.
1313 m = get_next_mapping(pool);
1315 m->pass_discard = pool->pf.discard_passdown;
1316 m->definitely_not_shared = !lookup_result.shared;
1317 m->virt_block = block;
1318 m->data_block = lookup_result.block;
1323 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1324 pool->process_prepared_discard(m);
1327 inc_all_io_entry(pool, bio);
1328 cell_defer_no_holder(tc, cell);
1329 cell_defer_no_holder(tc, cell2);
1332 * The DM core makes sure that the discard doesn't span
1333 * a block boundary. So we submit the discard of a
1334 * partial block appropriately.
1336 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1337 remap_and_issue(tc, bio, lookup_result.block);
1345 * It isn't provisioned, just forget it.
1347 cell_defer_no_holder(tc, cell);
1352 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1354 cell_defer_no_holder(tc, cell);
1360 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1362 struct dm_bio_prison_cell *cell;
1363 struct dm_cell_key key;
1364 dm_block_t block = get_bio_block(tc, bio);
1366 build_virtual_key(tc->td, block, &key);
1367 if (bio_detain(tc->pool, &key, bio, &cell))
1370 process_discard_cell(tc, cell);
1373 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1374 struct dm_cell_key *key,
1375 struct dm_thin_lookup_result *lookup_result,
1376 struct dm_bio_prison_cell *cell)
1379 dm_block_t data_block;
1380 struct pool *pool = tc->pool;
1382 r = alloc_data_block(tc, &data_block);
1385 schedule_internal_copy(tc, block, lookup_result->block,
1386 data_block, cell, bio);
1390 retry_bios_on_resume(pool, cell);
1394 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1396 cell_error(pool, cell);
1401 static void __remap_and_issue_shared_cell(void *context,
1402 struct dm_bio_prison_cell *cell)
1404 struct remap_info *info = context;
1407 while ((bio = bio_list_pop(&cell->bios))) {
1408 if ((bio_data_dir(bio) == WRITE) ||
1409 (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)))
1410 bio_list_add(&info->defer_bios, bio);
1412 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1414 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1415 inc_all_io_entry(info->tc->pool, bio);
1416 bio_list_add(&info->issue_bios, bio);
1421 static void remap_and_issue_shared_cell(struct thin_c *tc,
1422 struct dm_bio_prison_cell *cell,
1426 struct remap_info info;
1429 bio_list_init(&info.defer_bios);
1430 bio_list_init(&info.issue_bios);
1432 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1435 while ((bio = bio_list_pop(&info.defer_bios)))
1436 thin_defer_bio(tc, bio);
1438 while ((bio = bio_list_pop(&info.issue_bios)))
1439 remap_and_issue(tc, bio, block);
1442 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1444 struct dm_thin_lookup_result *lookup_result,
1445 struct dm_bio_prison_cell *virt_cell)
1447 struct dm_bio_prison_cell *data_cell;
1448 struct pool *pool = tc->pool;
1449 struct dm_cell_key key;
1452 * If cell is already occupied, then sharing is already in the process
1453 * of being broken so we have nothing further to do here.
1455 build_data_key(tc->td, lookup_result->block, &key);
1456 if (bio_detain(pool, &key, bio, &data_cell)) {
1457 cell_defer_no_holder(tc, virt_cell);
1461 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1462 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1463 cell_defer_no_holder(tc, virt_cell);
1465 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1467 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1468 inc_all_io_entry(pool, bio);
1469 remap_and_issue(tc, bio, lookup_result->block);
1471 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1472 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1476 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1477 struct dm_bio_prison_cell *cell)
1480 dm_block_t data_block;
1481 struct pool *pool = tc->pool;
1484 * Remap empty bios (flushes) immediately, without provisioning.
1486 if (!bio->bi_iter.bi_size) {
1487 inc_all_io_entry(pool, bio);
1488 cell_defer_no_holder(tc, cell);
1490 remap_and_issue(tc, bio, 0);
1495 * Fill read bios with zeroes and complete them immediately.
1497 if (bio_data_dir(bio) == READ) {
1499 cell_defer_no_holder(tc, cell);
1504 r = alloc_data_block(tc, &data_block);
1508 schedule_external_copy(tc, block, data_block, cell, bio);
1510 schedule_zero(tc, block, data_block, cell, bio);
1514 retry_bios_on_resume(pool, cell);
1518 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1520 cell_error(pool, cell);
1525 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1528 struct pool *pool = tc->pool;
1529 struct bio *bio = cell->holder;
1530 dm_block_t block = get_bio_block(tc, bio);
1531 struct dm_thin_lookup_result lookup_result;
1533 if (tc->requeue_mode) {
1534 cell_requeue(pool, cell);
1538 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1541 if (lookup_result.shared)
1542 process_shared_bio(tc, bio, block, &lookup_result, cell);
1544 inc_all_io_entry(pool, bio);
1545 remap_and_issue(tc, bio, lookup_result.block);
1546 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1551 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1552 inc_all_io_entry(pool, bio);
1553 cell_defer_no_holder(tc, cell);
1555 if (bio_end_sector(bio) <= tc->origin_size)
1556 remap_to_origin_and_issue(tc, bio);
1558 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1560 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1561 remap_to_origin_and_issue(tc, bio);
1568 provision_block(tc, bio, block, cell);
1572 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1574 cell_defer_no_holder(tc, cell);
1580 static void process_bio(struct thin_c *tc, struct bio *bio)
1582 struct pool *pool = tc->pool;
1583 dm_block_t block = get_bio_block(tc, bio);
1584 struct dm_bio_prison_cell *cell;
1585 struct dm_cell_key key;
1588 * If cell is already occupied, then the block is already
1589 * being provisioned so we have nothing further to do here.
1591 build_virtual_key(tc->td, block, &key);
1592 if (bio_detain(pool, &key, bio, &cell))
1595 process_cell(tc, cell);
1598 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1599 struct dm_bio_prison_cell *cell)
1602 int rw = bio_data_dir(bio);
1603 dm_block_t block = get_bio_block(tc, bio);
1604 struct dm_thin_lookup_result lookup_result;
1606 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1609 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1610 handle_unserviceable_bio(tc->pool, bio);
1612 cell_defer_no_holder(tc, cell);
1614 inc_all_io_entry(tc->pool, bio);
1615 remap_and_issue(tc, bio, lookup_result.block);
1617 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1623 cell_defer_no_holder(tc, cell);
1625 handle_unserviceable_bio(tc->pool, bio);
1629 if (tc->origin_dev) {
1630 inc_all_io_entry(tc->pool, bio);
1631 remap_to_origin_and_issue(tc, bio);
1640 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1643 cell_defer_no_holder(tc, cell);
1649 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1651 __process_bio_read_only(tc, bio, NULL);
1654 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1656 __process_bio_read_only(tc, cell->holder, cell);
1659 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1664 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1669 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1671 cell_success(tc->pool, cell);
1674 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1676 cell_error(tc->pool, cell);
1680 * FIXME: should we also commit due to size of transaction, measured in
1683 static int need_commit_due_to_time(struct pool *pool)
1685 return jiffies < pool->last_commit_jiffies ||
1686 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1689 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1690 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1692 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1694 struct rb_node **rbp, *parent;
1695 struct dm_thin_endio_hook *pbd;
1696 sector_t bi_sector = bio->bi_iter.bi_sector;
1698 rbp = &tc->sort_bio_list.rb_node;
1702 pbd = thin_pbd(parent);
1704 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1705 rbp = &(*rbp)->rb_left;
1707 rbp = &(*rbp)->rb_right;
1710 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1711 rb_link_node(&pbd->rb_node, parent, rbp);
1712 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1715 static void __extract_sorted_bios(struct thin_c *tc)
1717 struct rb_node *node;
1718 struct dm_thin_endio_hook *pbd;
1721 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1722 pbd = thin_pbd(node);
1723 bio = thin_bio(pbd);
1725 bio_list_add(&tc->deferred_bio_list, bio);
1726 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1729 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1732 static void __sort_thin_deferred_bios(struct thin_c *tc)
1735 struct bio_list bios;
1737 bio_list_init(&bios);
1738 bio_list_merge(&bios, &tc->deferred_bio_list);
1739 bio_list_init(&tc->deferred_bio_list);
1741 /* Sort deferred_bio_list using rb-tree */
1742 while ((bio = bio_list_pop(&bios)))
1743 __thin_bio_rb_add(tc, bio);
1746 * Transfer the sorted bios in sort_bio_list back to
1747 * deferred_bio_list to allow lockless submission of
1750 __extract_sorted_bios(tc);
1753 static void process_thin_deferred_bios(struct thin_c *tc)
1755 struct pool *pool = tc->pool;
1756 unsigned long flags;
1758 struct bio_list bios;
1759 struct blk_plug plug;
1762 if (tc->requeue_mode) {
1763 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
1767 bio_list_init(&bios);
1769 spin_lock_irqsave(&tc->lock, flags);
1771 if (bio_list_empty(&tc->deferred_bio_list)) {
1772 spin_unlock_irqrestore(&tc->lock, flags);
1776 __sort_thin_deferred_bios(tc);
1778 bio_list_merge(&bios, &tc->deferred_bio_list);
1779 bio_list_init(&tc->deferred_bio_list);
1781 spin_unlock_irqrestore(&tc->lock, flags);
1783 blk_start_plug(&plug);
1784 while ((bio = bio_list_pop(&bios))) {
1786 * If we've got no free new_mapping structs, and processing
1787 * this bio might require one, we pause until there are some
1788 * prepared mappings to process.
1790 if (ensure_next_mapping(pool)) {
1791 spin_lock_irqsave(&tc->lock, flags);
1792 bio_list_add(&tc->deferred_bio_list, bio);
1793 bio_list_merge(&tc->deferred_bio_list, &bios);
1794 spin_unlock_irqrestore(&tc->lock, flags);
1798 if (bio->bi_rw & REQ_DISCARD)
1799 pool->process_discard(tc, bio);
1801 pool->process_bio(tc, bio);
1803 if ((count++ & 127) == 0) {
1804 throttle_work_update(&pool->throttle);
1805 dm_pool_issue_prefetches(pool->pmd);
1808 blk_finish_plug(&plug);
1811 static int cmp_cells(const void *lhs, const void *rhs)
1813 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
1814 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
1816 BUG_ON(!lhs_cell->holder);
1817 BUG_ON(!rhs_cell->holder);
1819 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
1822 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
1828 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
1831 struct dm_bio_prison_cell *cell, *tmp;
1833 list_for_each_entry_safe(cell, tmp, cells, user_list) {
1834 if (count >= CELL_SORT_ARRAY_SIZE)
1837 pool->cell_sort_array[count++] = cell;
1838 list_del(&cell->user_list);
1841 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
1846 static void process_thin_deferred_cells(struct thin_c *tc)
1848 struct pool *pool = tc->pool;
1849 unsigned long flags;
1850 struct list_head cells;
1851 struct dm_bio_prison_cell *cell;
1852 unsigned i, j, count;
1854 INIT_LIST_HEAD(&cells);
1856 spin_lock_irqsave(&tc->lock, flags);
1857 list_splice_init(&tc->deferred_cells, &cells);
1858 spin_unlock_irqrestore(&tc->lock, flags);
1860 if (list_empty(&cells))
1864 count = sort_cells(tc->pool, &cells);
1866 for (i = 0; i < count; i++) {
1867 cell = pool->cell_sort_array[i];
1868 BUG_ON(!cell->holder);
1871 * If we've got no free new_mapping structs, and processing
1872 * this bio might require one, we pause until there are some
1873 * prepared mappings to process.
1875 if (ensure_next_mapping(pool)) {
1876 for (j = i; j < count; j++)
1877 list_add(&pool->cell_sort_array[j]->user_list, &cells);
1879 spin_lock_irqsave(&tc->lock, flags);
1880 list_splice(&cells, &tc->deferred_cells);
1881 spin_unlock_irqrestore(&tc->lock, flags);
1885 if (cell->holder->bi_rw & REQ_DISCARD)
1886 pool->process_discard_cell(tc, cell);
1888 pool->process_cell(tc, cell);
1890 } while (!list_empty(&cells));
1893 static void thin_get(struct thin_c *tc);
1894 static void thin_put(struct thin_c *tc);
1897 * We can't hold rcu_read_lock() around code that can block. So we
1898 * find a thin with the rcu lock held; bump a refcount; then drop
1901 static struct thin_c *get_first_thin(struct pool *pool)
1903 struct thin_c *tc = NULL;
1906 if (!list_empty(&pool->active_thins)) {
1907 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
1915 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
1917 struct thin_c *old_tc = tc;
1920 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
1932 static void process_deferred_bios(struct pool *pool)
1934 unsigned long flags;
1936 struct bio_list bios;
1939 tc = get_first_thin(pool);
1941 process_thin_deferred_cells(tc);
1942 process_thin_deferred_bios(tc);
1943 tc = get_next_thin(pool, tc);
1947 * If there are any deferred flush bios, we must commit
1948 * the metadata before issuing them.
1950 bio_list_init(&bios);
1951 spin_lock_irqsave(&pool->lock, flags);
1952 bio_list_merge(&bios, &pool->deferred_flush_bios);
1953 bio_list_init(&pool->deferred_flush_bios);
1954 spin_unlock_irqrestore(&pool->lock, flags);
1956 if (bio_list_empty(&bios) &&
1957 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
1961 while ((bio = bio_list_pop(&bios)))
1965 pool->last_commit_jiffies = jiffies;
1967 while ((bio = bio_list_pop(&bios)))
1968 generic_make_request(bio);
1971 static void do_worker(struct work_struct *ws)
1973 struct pool *pool = container_of(ws, struct pool, worker);
1975 throttle_work_start(&pool->throttle);
1976 dm_pool_issue_prefetches(pool->pmd);
1977 throttle_work_update(&pool->throttle);
1978 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1979 throttle_work_update(&pool->throttle);
1980 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1981 throttle_work_update(&pool->throttle);
1982 process_deferred_bios(pool);
1983 throttle_work_complete(&pool->throttle);
1987 * We want to commit periodically so that not too much
1988 * unwritten data builds up.
1990 static void do_waker(struct work_struct *ws)
1992 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1994 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1998 * We're holding onto IO to allow userland time to react. After the
1999 * timeout either the pool will have been resized (and thus back in
2000 * PM_WRITE mode), or we degrade to PM_READ_ONLY and start erroring IO.
2002 static void do_no_space_timeout(struct work_struct *ws)
2004 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2007 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space)
2008 set_pool_mode(pool, PM_READ_ONLY);
2011 /*----------------------------------------------------------------*/
2014 struct work_struct worker;
2015 struct completion complete;
2018 static struct pool_work *to_pool_work(struct work_struct *ws)
2020 return container_of(ws, struct pool_work, worker);
2023 static void pool_work_complete(struct pool_work *pw)
2025 complete(&pw->complete);
2028 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2029 void (*fn)(struct work_struct *))
2031 INIT_WORK_ONSTACK(&pw->worker, fn);
2032 init_completion(&pw->complete);
2033 queue_work(pool->wq, &pw->worker);
2034 wait_for_completion(&pw->complete);
2037 /*----------------------------------------------------------------*/
2039 struct noflush_work {
2040 struct pool_work pw;
2044 static struct noflush_work *to_noflush(struct work_struct *ws)
2046 return container_of(to_pool_work(ws), struct noflush_work, pw);
2049 static void do_noflush_start(struct work_struct *ws)
2051 struct noflush_work *w = to_noflush(ws);
2052 w->tc->requeue_mode = true;
2054 pool_work_complete(&w->pw);
2057 static void do_noflush_stop(struct work_struct *ws)
2059 struct noflush_work *w = to_noflush(ws);
2060 w->tc->requeue_mode = false;
2061 pool_work_complete(&w->pw);
2064 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2066 struct noflush_work w;
2069 pool_work_wait(&w.pw, tc->pool, fn);
2072 /*----------------------------------------------------------------*/
2074 static enum pool_mode get_pool_mode(struct pool *pool)
2076 return pool->pf.mode;
2079 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2081 dm_table_event(pool->ti->table);
2082 DMINFO("%s: switching pool to %s mode",
2083 dm_device_name(pool->pool_md), new_mode);
2086 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2088 struct pool_c *pt = pool->ti->private;
2089 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2090 enum pool_mode old_mode = get_pool_mode(pool);
2091 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2094 * Never allow the pool to transition to PM_WRITE mode if user
2095 * intervention is required to verify metadata and data consistency.
2097 if (new_mode == PM_WRITE && needs_check) {
2098 DMERR("%s: unable to switch pool to write mode until repaired.",
2099 dm_device_name(pool->pool_md));
2100 if (old_mode != new_mode)
2101 new_mode = old_mode;
2103 new_mode = PM_READ_ONLY;
2106 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2107 * not going to recover without a thin_repair. So we never let the
2108 * pool move out of the old mode.
2110 if (old_mode == PM_FAIL)
2111 new_mode = old_mode;
2115 if (old_mode != new_mode)
2116 notify_of_pool_mode_change(pool, "failure");
2117 dm_pool_metadata_read_only(pool->pmd);
2118 pool->process_bio = process_bio_fail;
2119 pool->process_discard = process_bio_fail;
2120 pool->process_cell = process_cell_fail;
2121 pool->process_discard_cell = process_cell_fail;
2122 pool->process_prepared_mapping = process_prepared_mapping_fail;
2123 pool->process_prepared_discard = process_prepared_discard_fail;
2125 error_retry_list(pool);
2129 if (old_mode != new_mode)
2130 notify_of_pool_mode_change(pool, "read-only");
2131 dm_pool_metadata_read_only(pool->pmd);
2132 pool->process_bio = process_bio_read_only;
2133 pool->process_discard = process_bio_success;
2134 pool->process_cell = process_cell_read_only;
2135 pool->process_discard_cell = process_cell_success;
2136 pool->process_prepared_mapping = process_prepared_mapping_fail;
2137 pool->process_prepared_discard = process_prepared_discard_passdown;
2139 error_retry_list(pool);
2142 case PM_OUT_OF_DATA_SPACE:
2144 * Ideally we'd never hit this state; the low water mark
2145 * would trigger userland to extend the pool before we
2146 * completely run out of data space. However, many small
2147 * IOs to unprovisioned space can consume data space at an
2148 * alarming rate. Adjust your low water mark if you're
2149 * frequently seeing this mode.
2151 if (old_mode != new_mode)
2152 notify_of_pool_mode_change(pool, "out-of-data-space");
2153 pool->process_bio = process_bio_read_only;
2154 pool->process_discard = process_discard_bio;
2155 pool->process_cell = process_cell_read_only;
2156 pool->process_discard_cell = process_discard_cell;
2157 pool->process_prepared_mapping = process_prepared_mapping;
2158 pool->process_prepared_discard = process_prepared_discard_passdown;
2160 if (!pool->pf.error_if_no_space && no_space_timeout)
2161 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2165 if (old_mode != new_mode)
2166 notify_of_pool_mode_change(pool, "write");
2167 dm_pool_metadata_read_write(pool->pmd);
2168 pool->process_bio = process_bio;
2169 pool->process_discard = process_discard_bio;
2170 pool->process_cell = process_cell;
2171 pool->process_discard_cell = process_discard_cell;
2172 pool->process_prepared_mapping = process_prepared_mapping;
2173 pool->process_prepared_discard = process_prepared_discard;
2177 pool->pf.mode = new_mode;
2179 * The pool mode may have changed, sync it so bind_control_target()
2180 * doesn't cause an unexpected mode transition on resume.
2182 pt->adjusted_pf.mode = new_mode;
2185 static void abort_transaction(struct pool *pool)
2187 const char *dev_name = dm_device_name(pool->pool_md);
2189 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2190 if (dm_pool_abort_metadata(pool->pmd)) {
2191 DMERR("%s: failed to abort metadata transaction", dev_name);
2192 set_pool_mode(pool, PM_FAIL);
2195 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2196 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2197 set_pool_mode(pool, PM_FAIL);
2201 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2203 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2204 dm_device_name(pool->pool_md), op, r);
2206 abort_transaction(pool);
2207 set_pool_mode(pool, PM_READ_ONLY);
2210 /*----------------------------------------------------------------*/
2213 * Mapping functions.
2217 * Called only while mapping a thin bio to hand it over to the workqueue.
2219 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2221 unsigned long flags;
2222 struct pool *pool = tc->pool;
2224 spin_lock_irqsave(&tc->lock, flags);
2225 bio_list_add(&tc->deferred_bio_list, bio);
2226 spin_unlock_irqrestore(&tc->lock, flags);
2231 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2233 struct pool *pool = tc->pool;
2235 throttle_lock(&pool->throttle);
2236 thin_defer_bio(tc, bio);
2237 throttle_unlock(&pool->throttle);
2240 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2242 unsigned long flags;
2243 struct pool *pool = tc->pool;
2245 throttle_lock(&pool->throttle);
2246 spin_lock_irqsave(&tc->lock, flags);
2247 list_add_tail(&cell->user_list, &tc->deferred_cells);
2248 spin_unlock_irqrestore(&tc->lock, flags);
2249 throttle_unlock(&pool->throttle);
2254 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2256 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2259 h->shared_read_entry = NULL;
2260 h->all_io_entry = NULL;
2261 h->overwrite_mapping = NULL;
2265 * Non-blocking function called from the thin target's map function.
2267 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2270 struct thin_c *tc = ti->private;
2271 dm_block_t block = get_bio_block(tc, bio);
2272 struct dm_thin_device *td = tc->td;
2273 struct dm_thin_lookup_result result;
2274 struct dm_bio_prison_cell *virt_cell, *data_cell;
2275 struct dm_cell_key key;
2277 thin_hook_bio(tc, bio);
2279 if (tc->requeue_mode) {
2280 bio_endio(bio, DM_ENDIO_REQUEUE);
2281 return DM_MAPIO_SUBMITTED;
2284 if (get_pool_mode(tc->pool) == PM_FAIL) {
2286 return DM_MAPIO_SUBMITTED;
2289 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
2290 thin_defer_bio_with_throttle(tc, bio);
2291 return DM_MAPIO_SUBMITTED;
2295 * We must hold the virtual cell before doing the lookup, otherwise
2296 * there's a race with discard.
2298 build_virtual_key(tc->td, block, &key);
2299 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2300 return DM_MAPIO_SUBMITTED;
2302 r = dm_thin_find_block(td, block, 0, &result);
2305 * Note that we defer readahead too.
2309 if (unlikely(result.shared)) {
2311 * We have a race condition here between the
2312 * result.shared value returned by the lookup and
2313 * snapshot creation, which may cause new
2316 * To avoid this always quiesce the origin before
2317 * taking the snap. You want to do this anyway to
2318 * ensure a consistent application view
2321 * More distant ancestors are irrelevant. The
2322 * shared flag will be set in their case.
2324 thin_defer_cell(tc, virt_cell);
2325 return DM_MAPIO_SUBMITTED;
2328 build_data_key(tc->td, result.block, &key);
2329 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2330 cell_defer_no_holder(tc, virt_cell);
2331 return DM_MAPIO_SUBMITTED;
2334 inc_all_io_entry(tc->pool, bio);
2335 cell_defer_no_holder(tc, data_cell);
2336 cell_defer_no_holder(tc, virt_cell);
2338 remap(tc, bio, result.block);
2339 return DM_MAPIO_REMAPPED;
2342 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
2344 * This block isn't provisioned, and we have no way
2347 handle_unserviceable_bio(tc->pool, bio);
2348 cell_defer_no_holder(tc, virt_cell);
2349 return DM_MAPIO_SUBMITTED;
2354 thin_defer_cell(tc, virt_cell);
2355 return DM_MAPIO_SUBMITTED;
2359 * Must always call bio_io_error on failure.
2360 * dm_thin_find_block can fail with -EINVAL if the
2361 * pool is switched to fail-io mode.
2364 cell_defer_no_holder(tc, virt_cell);
2365 return DM_MAPIO_SUBMITTED;
2369 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2371 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2372 struct request_queue *q;
2374 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2377 q = bdev_get_queue(pt->data_dev->bdev);
2378 return bdi_congested(&q->backing_dev_info, bdi_bits);
2381 static void requeue_bios(struct pool *pool)
2383 unsigned long flags;
2387 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2388 spin_lock_irqsave(&tc->lock, flags);
2389 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2390 bio_list_init(&tc->retry_on_resume_list);
2391 spin_unlock_irqrestore(&tc->lock, flags);
2396 /*----------------------------------------------------------------
2397 * Binding of control targets to a pool object
2398 *--------------------------------------------------------------*/
2399 static bool data_dev_supports_discard(struct pool_c *pt)
2401 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2403 return q && blk_queue_discard(q);
2406 static bool is_factor(sector_t block_size, uint32_t n)
2408 return !sector_div(block_size, n);
2412 * If discard_passdown was enabled verify that the data device
2413 * supports discards. Disable discard_passdown if not.
2415 static void disable_passdown_if_not_supported(struct pool_c *pt)
2417 struct pool *pool = pt->pool;
2418 struct block_device *data_bdev = pt->data_dev->bdev;
2419 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2420 sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
2421 const char *reason = NULL;
2422 char buf[BDEVNAME_SIZE];
2424 if (!pt->adjusted_pf.discard_passdown)
2427 if (!data_dev_supports_discard(pt))
2428 reason = "discard unsupported";
2430 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2431 reason = "max discard sectors smaller than a block";
2433 else if (data_limits->discard_granularity > block_size)
2434 reason = "discard granularity larger than a block";
2436 else if (!is_factor(block_size, data_limits->discard_granularity))
2437 reason = "discard granularity not a factor of block size";
2440 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2441 pt->adjusted_pf.discard_passdown = false;
2445 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2447 struct pool_c *pt = ti->private;
2450 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2452 enum pool_mode old_mode = get_pool_mode(pool);
2453 enum pool_mode new_mode = pt->adjusted_pf.mode;
2456 * Don't change the pool's mode until set_pool_mode() below.
2457 * Otherwise the pool's process_* function pointers may
2458 * not match the desired pool mode.
2460 pt->adjusted_pf.mode = old_mode;
2463 pool->pf = pt->adjusted_pf;
2464 pool->low_water_blocks = pt->low_water_blocks;
2466 set_pool_mode(pool, new_mode);
2471 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2477 /*----------------------------------------------------------------
2479 *--------------------------------------------------------------*/
2480 /* Initialize pool features. */
2481 static void pool_features_init(struct pool_features *pf)
2483 pf->mode = PM_WRITE;
2484 pf->zero_new_blocks = true;
2485 pf->discard_enabled = true;
2486 pf->discard_passdown = true;
2487 pf->error_if_no_space = false;
2490 static void __pool_destroy(struct pool *pool)
2492 __pool_table_remove(pool);
2494 if (dm_pool_metadata_close(pool->pmd) < 0)
2495 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2497 dm_bio_prison_destroy(pool->prison);
2498 dm_kcopyd_client_destroy(pool->copier);
2501 destroy_workqueue(pool->wq);
2503 if (pool->next_mapping)
2504 mempool_free(pool->next_mapping, pool->mapping_pool);
2505 mempool_destroy(pool->mapping_pool);
2506 dm_deferred_set_destroy(pool->shared_read_ds);
2507 dm_deferred_set_destroy(pool->all_io_ds);
2511 static struct kmem_cache *_new_mapping_cache;
2513 static struct pool *pool_create(struct mapped_device *pool_md,
2514 struct block_device *metadata_dev,
2515 unsigned long block_size,
2516 int read_only, char **error)
2521 struct dm_pool_metadata *pmd;
2522 bool format_device = read_only ? false : true;
2524 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2526 *error = "Error creating metadata object";
2527 return (struct pool *)pmd;
2530 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2532 *error = "Error allocating memory for pool";
2533 err_p = ERR_PTR(-ENOMEM);
2538 pool->sectors_per_block = block_size;
2539 if (block_size & (block_size - 1))
2540 pool->sectors_per_block_shift = -1;
2542 pool->sectors_per_block_shift = __ffs(block_size);
2543 pool->low_water_blocks = 0;
2544 pool_features_init(&pool->pf);
2545 pool->prison = dm_bio_prison_create();
2546 if (!pool->prison) {
2547 *error = "Error creating pool's bio prison";
2548 err_p = ERR_PTR(-ENOMEM);
2552 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2553 if (IS_ERR(pool->copier)) {
2554 r = PTR_ERR(pool->copier);
2555 *error = "Error creating pool's kcopyd client";
2557 goto bad_kcopyd_client;
2561 * Create singlethreaded workqueue that will service all devices
2562 * that use this metadata.
2564 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2566 *error = "Error creating pool's workqueue";
2567 err_p = ERR_PTR(-ENOMEM);
2571 throttle_init(&pool->throttle);
2572 INIT_WORK(&pool->worker, do_worker);
2573 INIT_DELAYED_WORK(&pool->waker, do_waker);
2574 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2575 spin_lock_init(&pool->lock);
2576 bio_list_init(&pool->deferred_flush_bios);
2577 INIT_LIST_HEAD(&pool->prepared_mappings);
2578 INIT_LIST_HEAD(&pool->prepared_discards);
2579 INIT_LIST_HEAD(&pool->active_thins);
2580 pool->low_water_triggered = false;
2581 pool->suspended = true;
2583 pool->shared_read_ds = dm_deferred_set_create();
2584 if (!pool->shared_read_ds) {
2585 *error = "Error creating pool's shared read deferred set";
2586 err_p = ERR_PTR(-ENOMEM);
2587 goto bad_shared_read_ds;
2590 pool->all_io_ds = dm_deferred_set_create();
2591 if (!pool->all_io_ds) {
2592 *error = "Error creating pool's all io deferred set";
2593 err_p = ERR_PTR(-ENOMEM);
2597 pool->next_mapping = NULL;
2598 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2599 _new_mapping_cache);
2600 if (!pool->mapping_pool) {
2601 *error = "Error creating pool's mapping mempool";
2602 err_p = ERR_PTR(-ENOMEM);
2603 goto bad_mapping_pool;
2606 pool->ref_count = 1;
2607 pool->last_commit_jiffies = jiffies;
2608 pool->pool_md = pool_md;
2609 pool->md_dev = metadata_dev;
2610 __pool_table_insert(pool);
2615 dm_deferred_set_destroy(pool->all_io_ds);
2617 dm_deferred_set_destroy(pool->shared_read_ds);
2619 destroy_workqueue(pool->wq);
2621 dm_kcopyd_client_destroy(pool->copier);
2623 dm_bio_prison_destroy(pool->prison);
2627 if (dm_pool_metadata_close(pmd))
2628 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2633 static void __pool_inc(struct pool *pool)
2635 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2639 static void __pool_dec(struct pool *pool)
2641 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2642 BUG_ON(!pool->ref_count);
2643 if (!--pool->ref_count)
2644 __pool_destroy(pool);
2647 static struct pool *__pool_find(struct mapped_device *pool_md,
2648 struct block_device *metadata_dev,
2649 unsigned long block_size, int read_only,
2650 char **error, int *created)
2652 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2655 if (pool->pool_md != pool_md) {
2656 *error = "metadata device already in use by a pool";
2657 return ERR_PTR(-EBUSY);
2662 pool = __pool_table_lookup(pool_md);
2664 if (pool->md_dev != metadata_dev) {
2665 *error = "different pool cannot replace a pool";
2666 return ERR_PTR(-EINVAL);
2671 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2679 /*----------------------------------------------------------------
2680 * Pool target methods
2681 *--------------------------------------------------------------*/
2682 static void pool_dtr(struct dm_target *ti)
2684 struct pool_c *pt = ti->private;
2686 mutex_lock(&dm_thin_pool_table.mutex);
2688 unbind_control_target(pt->pool, ti);
2689 __pool_dec(pt->pool);
2690 dm_put_device(ti, pt->metadata_dev);
2691 dm_put_device(ti, pt->data_dev);
2694 mutex_unlock(&dm_thin_pool_table.mutex);
2697 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2698 struct dm_target *ti)
2702 const char *arg_name;
2704 static struct dm_arg _args[] = {
2705 {0, 4, "Invalid number of pool feature arguments"},
2709 * No feature arguments supplied.
2714 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2718 while (argc && !r) {
2719 arg_name = dm_shift_arg(as);
2722 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2723 pf->zero_new_blocks = false;
2725 else if (!strcasecmp(arg_name, "ignore_discard"))
2726 pf->discard_enabled = false;
2728 else if (!strcasecmp(arg_name, "no_discard_passdown"))
2729 pf->discard_passdown = false;
2731 else if (!strcasecmp(arg_name, "read_only"))
2732 pf->mode = PM_READ_ONLY;
2734 else if (!strcasecmp(arg_name, "error_if_no_space"))
2735 pf->error_if_no_space = true;
2738 ti->error = "Unrecognised pool feature requested";
2747 static void metadata_low_callback(void *context)
2749 struct pool *pool = context;
2751 DMWARN("%s: reached low water mark for metadata device: sending event.",
2752 dm_device_name(pool->pool_md));
2754 dm_table_event(pool->ti->table);
2757 static sector_t get_dev_size(struct block_device *bdev)
2759 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
2762 static void warn_if_metadata_device_too_big(struct block_device *bdev)
2764 sector_t metadata_dev_size = get_dev_size(bdev);
2765 char buffer[BDEVNAME_SIZE];
2767 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
2768 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2769 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
2772 static sector_t get_metadata_dev_size(struct block_device *bdev)
2774 sector_t metadata_dev_size = get_dev_size(bdev);
2776 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
2777 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
2779 return metadata_dev_size;
2782 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
2784 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
2786 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
2788 return metadata_dev_size;
2792 * When a metadata threshold is crossed a dm event is triggered, and
2793 * userland should respond by growing the metadata device. We could let
2794 * userland set the threshold, like we do with the data threshold, but I'm
2795 * not sure they know enough to do this well.
2797 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
2800 * 4M is ample for all ops with the possible exception of thin
2801 * device deletion which is harmless if it fails (just retry the
2802 * delete after you've grown the device).
2804 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
2805 return min((dm_block_t)1024ULL /* 4M */, quarter);
2809 * thin-pool <metadata dev> <data dev>
2810 * <data block size (sectors)>
2811 * <low water mark (blocks)>
2812 * [<#feature args> [<arg>]*]
2814 * Optional feature arguments are:
2815 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
2816 * ignore_discard: disable discard
2817 * no_discard_passdown: don't pass discards down to the data device
2818 * read_only: Don't allow any changes to be made to the pool metadata.
2819 * error_if_no_space: error IOs, instead of queueing, if no space.
2821 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
2823 int r, pool_created = 0;
2826 struct pool_features pf;
2827 struct dm_arg_set as;
2828 struct dm_dev *data_dev;
2829 unsigned long block_size;
2830 dm_block_t low_water_blocks;
2831 struct dm_dev *metadata_dev;
2832 fmode_t metadata_mode;
2835 * FIXME Remove validation from scope of lock.
2837 mutex_lock(&dm_thin_pool_table.mutex);
2840 ti->error = "Invalid argument count";
2849 * Set default pool features.
2851 pool_features_init(&pf);
2853 dm_consume_args(&as, 4);
2854 r = parse_pool_features(&as, &pf, ti);
2858 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2859 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2861 ti->error = "Error opening metadata block device";
2864 warn_if_metadata_device_too_big(metadata_dev->bdev);
2866 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2868 ti->error = "Error getting data device";
2872 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2873 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2874 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2875 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2876 ti->error = "Invalid block size";
2881 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2882 ti->error = "Invalid low water mark";
2887 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2893 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2894 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2901 * 'pool_created' reflects whether this is the first table load.
2902 * Top level discard support is not allowed to be changed after
2903 * initial load. This would require a pool reload to trigger thin
2906 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2907 ti->error = "Discard support cannot be disabled once enabled";
2909 goto out_flags_changed;
2914 pt->metadata_dev = metadata_dev;
2915 pt->data_dev = data_dev;
2916 pt->low_water_blocks = low_water_blocks;
2917 pt->adjusted_pf = pt->requested_pf = pf;
2918 ti->num_flush_bios = 1;
2921 * Only need to enable discards if the pool should pass
2922 * them down to the data device. The thin device's discard
2923 * processing will cause mappings to be removed from the btree.
2925 ti->discard_zeroes_data_unsupported = true;
2926 if (pf.discard_enabled && pf.discard_passdown) {
2927 ti->num_discard_bios = 1;
2930 * Setting 'discards_supported' circumvents the normal
2931 * stacking of discard limits (this keeps the pool and
2932 * thin devices' discard limits consistent).
2934 ti->discards_supported = true;
2938 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2939 calc_metadata_threshold(pt),
2940 metadata_low_callback,
2945 pt->callbacks.congested_fn = pool_is_congested;
2946 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2948 mutex_unlock(&dm_thin_pool_table.mutex);
2957 dm_put_device(ti, data_dev);
2959 dm_put_device(ti, metadata_dev);
2961 mutex_unlock(&dm_thin_pool_table.mutex);
2966 static int pool_map(struct dm_target *ti, struct bio *bio)
2969 struct pool_c *pt = ti->private;
2970 struct pool *pool = pt->pool;
2971 unsigned long flags;
2974 * As this is a singleton target, ti->begin is always zero.
2976 spin_lock_irqsave(&pool->lock, flags);
2977 bio->bi_bdev = pt->data_dev->bdev;
2978 r = DM_MAPIO_REMAPPED;
2979 spin_unlock_irqrestore(&pool->lock, flags);
2984 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
2987 struct pool_c *pt = ti->private;
2988 struct pool *pool = pt->pool;
2989 sector_t data_size = ti->len;
2990 dm_block_t sb_data_size;
2992 *need_commit = false;
2994 (void) sector_div(data_size, pool->sectors_per_block);
2996 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2998 DMERR("%s: failed to retrieve data device size",
2999 dm_device_name(pool->pool_md));
3003 if (data_size < sb_data_size) {
3004 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3005 dm_device_name(pool->pool_md),
3006 (unsigned long long)data_size, sb_data_size);
3009 } else if (data_size > sb_data_size) {
3010 if (dm_pool_metadata_needs_check(pool->pmd)) {
3011 DMERR("%s: unable to grow the data device until repaired.",
3012 dm_device_name(pool->pool_md));
3017 DMINFO("%s: growing the data device from %llu to %llu blocks",
3018 dm_device_name(pool->pool_md),
3019 sb_data_size, (unsigned long long)data_size);
3020 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3022 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3026 *need_commit = true;
3032 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3035 struct pool_c *pt = ti->private;
3036 struct pool *pool = pt->pool;
3037 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3039 *need_commit = false;
3041 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3043 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3045 DMERR("%s: failed to retrieve metadata device size",
3046 dm_device_name(pool->pool_md));
3050 if (metadata_dev_size < sb_metadata_dev_size) {
3051 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3052 dm_device_name(pool->pool_md),
3053 metadata_dev_size, sb_metadata_dev_size);
3056 } else if (metadata_dev_size > sb_metadata_dev_size) {
3057 if (dm_pool_metadata_needs_check(pool->pmd)) {
3058 DMERR("%s: unable to grow the metadata device until repaired.",
3059 dm_device_name(pool->pool_md));
3063 warn_if_metadata_device_too_big(pool->md_dev);
3064 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3065 dm_device_name(pool->pool_md),
3066 sb_metadata_dev_size, metadata_dev_size);
3067 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3069 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3073 *need_commit = true;
3080 * Retrieves the number of blocks of the data device from
3081 * the superblock and compares it to the actual device size,
3082 * thus resizing the data device in case it has grown.
3084 * This both copes with opening preallocated data devices in the ctr
3085 * being followed by a resume
3087 * calling the resume method individually after userspace has
3088 * grown the data device in reaction to a table event.
3090 static int pool_preresume(struct dm_target *ti)
3093 bool need_commit1, need_commit2;
3094 struct pool_c *pt = ti->private;
3095 struct pool *pool = pt->pool;
3098 * Take control of the pool object.
3100 r = bind_control_target(pool, ti);
3104 r = maybe_resize_data_dev(ti, &need_commit1);
3108 r = maybe_resize_metadata_dev(ti, &need_commit2);
3112 if (need_commit1 || need_commit2)
3113 (void) commit(pool);
3118 static void pool_suspend_active_thins(struct pool *pool)
3122 /* Suspend all active thin devices */
3123 tc = get_first_thin(pool);
3125 dm_internal_suspend_noflush(tc->thin_md);
3126 tc = get_next_thin(pool, tc);
3130 static void pool_resume_active_thins(struct pool *pool)
3134 /* Resume all active thin devices */
3135 tc = get_first_thin(pool);
3137 dm_internal_resume(tc->thin_md);
3138 tc = get_next_thin(pool, tc);
3142 static void pool_resume(struct dm_target *ti)
3144 struct pool_c *pt = ti->private;
3145 struct pool *pool = pt->pool;
3146 unsigned long flags;
3149 * Must requeue active_thins' bios and then resume
3150 * active_thins _before_ clearing 'suspend' flag.
3153 pool_resume_active_thins(pool);
3155 spin_lock_irqsave(&pool->lock, flags);
3156 pool->low_water_triggered = false;
3157 pool->suspended = false;
3158 spin_unlock_irqrestore(&pool->lock, flags);
3160 do_waker(&pool->waker.work);
3163 static void pool_presuspend(struct dm_target *ti)
3165 struct pool_c *pt = ti->private;
3166 struct pool *pool = pt->pool;
3167 unsigned long flags;
3169 spin_lock_irqsave(&pool->lock, flags);
3170 pool->suspended = true;
3171 spin_unlock_irqrestore(&pool->lock, flags);
3173 pool_suspend_active_thins(pool);
3176 static void pool_presuspend_undo(struct dm_target *ti)
3178 struct pool_c *pt = ti->private;
3179 struct pool *pool = pt->pool;
3180 unsigned long flags;
3182 pool_resume_active_thins(pool);
3184 spin_lock_irqsave(&pool->lock, flags);
3185 pool->suspended = false;
3186 spin_unlock_irqrestore(&pool->lock, flags);
3189 static void pool_postsuspend(struct dm_target *ti)
3191 struct pool_c *pt = ti->private;
3192 struct pool *pool = pt->pool;
3194 cancel_delayed_work(&pool->waker);
3195 cancel_delayed_work(&pool->no_space_timeout);
3196 flush_workqueue(pool->wq);
3197 (void) commit(pool);
3200 static int check_arg_count(unsigned argc, unsigned args_required)
3202 if (argc != args_required) {
3203 DMWARN("Message received with %u arguments instead of %u.",
3204 argc, args_required);
3211 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3213 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3214 *dev_id <= MAX_DEV_ID)
3218 DMWARN("Message received with invalid device id: %s", arg);
3223 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3228 r = check_arg_count(argc, 2);
3232 r = read_dev_id(argv[1], &dev_id, 1);
3236 r = dm_pool_create_thin(pool->pmd, dev_id);
3238 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3246 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3249 dm_thin_id origin_dev_id;
3252 r = check_arg_count(argc, 3);
3256 r = read_dev_id(argv[1], &dev_id, 1);
3260 r = read_dev_id(argv[2], &origin_dev_id, 1);
3264 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3266 DMWARN("Creation of new snapshot %s of device %s failed.",
3274 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3279 r = check_arg_count(argc, 2);
3283 r = read_dev_id(argv[1], &dev_id, 1);
3287 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3289 DMWARN("Deletion of thin device %s failed.", argv[1]);
3294 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3296 dm_thin_id old_id, new_id;
3299 r = check_arg_count(argc, 3);
3303 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3304 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3308 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3309 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3313 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3315 DMWARN("Failed to change transaction id from %s to %s.",
3323 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3327 r = check_arg_count(argc, 1);
3331 (void) commit(pool);
3333 r = dm_pool_reserve_metadata_snap(pool->pmd);
3335 DMWARN("reserve_metadata_snap message failed.");
3340 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3344 r = check_arg_count(argc, 1);
3348 r = dm_pool_release_metadata_snap(pool->pmd);
3350 DMWARN("release_metadata_snap message failed.");
3356 * Messages supported:
3357 * create_thin <dev_id>
3358 * create_snap <dev_id> <origin_id>
3360 * set_transaction_id <current_trans_id> <new_trans_id>
3361 * reserve_metadata_snap
3362 * release_metadata_snap
3364 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3367 struct pool_c *pt = ti->private;
3368 struct pool *pool = pt->pool;
3370 if (!strcasecmp(argv[0], "create_thin"))
3371 r = process_create_thin_mesg(argc, argv, pool);
3373 else if (!strcasecmp(argv[0], "create_snap"))
3374 r = process_create_snap_mesg(argc, argv, pool);
3376 else if (!strcasecmp(argv[0], "delete"))
3377 r = process_delete_mesg(argc, argv, pool);
3379 else if (!strcasecmp(argv[0], "set_transaction_id"))
3380 r = process_set_transaction_id_mesg(argc, argv, pool);
3382 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3383 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3385 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3386 r = process_release_metadata_snap_mesg(argc, argv, pool);
3389 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3392 (void) commit(pool);
3397 static void emit_flags(struct pool_features *pf, char *result,
3398 unsigned sz, unsigned maxlen)
3400 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3401 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3402 pf->error_if_no_space;
3403 DMEMIT("%u ", count);
3405 if (!pf->zero_new_blocks)
3406 DMEMIT("skip_block_zeroing ");
3408 if (!pf->discard_enabled)
3409 DMEMIT("ignore_discard ");
3411 if (!pf->discard_passdown)
3412 DMEMIT("no_discard_passdown ");
3414 if (pf->mode == PM_READ_ONLY)
3415 DMEMIT("read_only ");
3417 if (pf->error_if_no_space)
3418 DMEMIT("error_if_no_space ");
3423 * <transaction id> <used metadata sectors>/<total metadata sectors>
3424 * <used data sectors>/<total data sectors> <held metadata root>
3426 static void pool_status(struct dm_target *ti, status_type_t type,
3427 unsigned status_flags, char *result, unsigned maxlen)
3431 uint64_t transaction_id;
3432 dm_block_t nr_free_blocks_data;
3433 dm_block_t nr_free_blocks_metadata;
3434 dm_block_t nr_blocks_data;
3435 dm_block_t nr_blocks_metadata;
3436 dm_block_t held_root;
3437 char buf[BDEVNAME_SIZE];
3438 char buf2[BDEVNAME_SIZE];
3439 struct pool_c *pt = ti->private;
3440 struct pool *pool = pt->pool;
3443 case STATUSTYPE_INFO:
3444 if (get_pool_mode(pool) == PM_FAIL) {
3449 /* Commit to ensure statistics aren't out-of-date */
3450 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3451 (void) commit(pool);
3453 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3455 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3456 dm_device_name(pool->pool_md), r);
3460 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3462 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3463 dm_device_name(pool->pool_md), r);
3467 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3469 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3470 dm_device_name(pool->pool_md), r);
3474 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3476 DMERR("%s: dm_pool_get_free_block_count returned %d",
3477 dm_device_name(pool->pool_md), r);
3481 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3483 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3484 dm_device_name(pool->pool_md), r);
3488 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3490 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3491 dm_device_name(pool->pool_md), r);
3495 DMEMIT("%llu %llu/%llu %llu/%llu ",
3496 (unsigned long long)transaction_id,
3497 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3498 (unsigned long long)nr_blocks_metadata,
3499 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3500 (unsigned long long)nr_blocks_data);
3503 DMEMIT("%llu ", held_root);
3507 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3508 DMEMIT("out_of_data_space ");
3509 else if (pool->pf.mode == PM_READ_ONLY)
3514 if (!pool->pf.discard_enabled)
3515 DMEMIT("ignore_discard ");
3516 else if (pool->pf.discard_passdown)
3517 DMEMIT("discard_passdown ");
3519 DMEMIT("no_discard_passdown ");
3521 if (pool->pf.error_if_no_space)
3522 DMEMIT("error_if_no_space ");
3524 DMEMIT("queue_if_no_space ");
3528 case STATUSTYPE_TABLE:
3529 DMEMIT("%s %s %lu %llu ",
3530 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3531 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3532 (unsigned long)pool->sectors_per_block,
3533 (unsigned long long)pt->low_water_blocks);
3534 emit_flags(&pt->requested_pf, result, sz, maxlen);
3543 static int pool_iterate_devices(struct dm_target *ti,
3544 iterate_devices_callout_fn fn, void *data)
3546 struct pool_c *pt = ti->private;
3548 return fn(ti, pt->data_dev, 0, ti->len, data);
3551 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3552 struct bio_vec *biovec, int max_size)
3554 struct pool_c *pt = ti->private;
3555 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3557 if (!q->merge_bvec_fn)
3560 bvm->bi_bdev = pt->data_dev->bdev;
3562 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3565 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
3567 struct pool *pool = pt->pool;
3568 struct queue_limits *data_limits;
3570 limits->max_discard_sectors = pool->sectors_per_block;
3573 * discard_granularity is just a hint, and not enforced.
3575 if (pt->adjusted_pf.discard_passdown) {
3576 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
3577 limits->discard_granularity = max(data_limits->discard_granularity,
3578 pool->sectors_per_block << SECTOR_SHIFT);
3580 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
3583 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3585 struct pool_c *pt = ti->private;
3586 struct pool *pool = pt->pool;
3587 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3590 * If max_sectors is smaller than pool->sectors_per_block adjust it
3591 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3592 * This is especially beneficial when the pool's data device is a RAID
3593 * device that has a full stripe width that matches pool->sectors_per_block
3594 * -- because even though partial RAID stripe-sized IOs will be issued to a
3595 * single RAID stripe; when aggregated they will end on a full RAID stripe
3596 * boundary.. which avoids additional partial RAID stripe writes cascading
3598 if (limits->max_sectors < pool->sectors_per_block) {
3599 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3600 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3601 limits->max_sectors--;
3602 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3607 * If the system-determined stacked limits are compatible with the
3608 * pool's blocksize (io_opt is a factor) do not override them.
3610 if (io_opt_sectors < pool->sectors_per_block ||
3611 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3612 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3613 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3615 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3616 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3620 * pt->adjusted_pf is a staging area for the actual features to use.
3621 * They get transferred to the live pool in bind_control_target()
3622 * called from pool_preresume().
3624 if (!pt->adjusted_pf.discard_enabled) {
3626 * Must explicitly disallow stacking discard limits otherwise the
3627 * block layer will stack them if pool's data device has support.
3628 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3629 * user to see that, so make sure to set all discard limits to 0.
3631 limits->discard_granularity = 0;
3635 disable_passdown_if_not_supported(pt);
3637 set_discard_limits(pt, limits);
3640 static struct target_type pool_target = {
3641 .name = "thin-pool",
3642 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3643 DM_TARGET_IMMUTABLE,
3644 .version = {1, 14, 0},
3645 .module = THIS_MODULE,
3649 .presuspend = pool_presuspend,
3650 .presuspend_undo = pool_presuspend_undo,
3651 .postsuspend = pool_postsuspend,
3652 .preresume = pool_preresume,
3653 .resume = pool_resume,
3654 .message = pool_message,
3655 .status = pool_status,
3656 .merge = pool_merge,
3657 .iterate_devices = pool_iterate_devices,
3658 .io_hints = pool_io_hints,
3661 /*----------------------------------------------------------------
3662 * Thin target methods
3663 *--------------------------------------------------------------*/
3664 static void thin_get(struct thin_c *tc)
3666 atomic_inc(&tc->refcount);
3669 static void thin_put(struct thin_c *tc)
3671 if (atomic_dec_and_test(&tc->refcount))
3672 complete(&tc->can_destroy);
3675 static void thin_dtr(struct dm_target *ti)
3677 struct thin_c *tc = ti->private;
3678 unsigned long flags;
3680 spin_lock_irqsave(&tc->pool->lock, flags);
3681 list_del_rcu(&tc->list);
3682 spin_unlock_irqrestore(&tc->pool->lock, flags);
3686 wait_for_completion(&tc->can_destroy);
3688 mutex_lock(&dm_thin_pool_table.mutex);
3690 __pool_dec(tc->pool);
3691 dm_pool_close_thin_device(tc->td);
3692 dm_put_device(ti, tc->pool_dev);
3694 dm_put_device(ti, tc->origin_dev);
3697 mutex_unlock(&dm_thin_pool_table.mutex);
3701 * Thin target parameters:
3703 * <pool_dev> <dev_id> [origin_dev]
3705 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3706 * dev_id: the internal device identifier
3707 * origin_dev: a device external to the pool that should act as the origin
3709 * If the pool device has discards disabled, they get disabled for the thin
3712 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3716 struct dm_dev *pool_dev, *origin_dev;
3717 struct mapped_device *pool_md;
3718 unsigned long flags;
3720 mutex_lock(&dm_thin_pool_table.mutex);
3722 if (argc != 2 && argc != 3) {
3723 ti->error = "Invalid argument count";
3728 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3730 ti->error = "Out of memory";
3734 tc->thin_md = dm_table_get_md(ti->table);
3735 spin_lock_init(&tc->lock);
3736 INIT_LIST_HEAD(&tc->deferred_cells);
3737 bio_list_init(&tc->deferred_bio_list);
3738 bio_list_init(&tc->retry_on_resume_list);
3739 tc->sort_bio_list = RB_ROOT;
3742 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
3744 ti->error = "Error opening origin device";
3745 goto bad_origin_dev;
3747 tc->origin_dev = origin_dev;
3750 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
3752 ti->error = "Error opening pool device";
3755 tc->pool_dev = pool_dev;
3757 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
3758 ti->error = "Invalid device id";
3763 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
3765 ti->error = "Couldn't get pool mapped device";
3770 tc->pool = __pool_table_lookup(pool_md);
3772 ti->error = "Couldn't find pool object";
3774 goto bad_pool_lookup;
3776 __pool_inc(tc->pool);
3778 if (get_pool_mode(tc->pool) == PM_FAIL) {
3779 ti->error = "Couldn't open thin device, Pool is in fail mode";
3784 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
3786 ti->error = "Couldn't open thin internal device";
3790 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
3794 ti->num_flush_bios = 1;
3795 ti->flush_supported = true;
3796 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
3798 /* In case the pool supports discards, pass them on. */
3799 ti->discard_zeroes_data_unsupported = true;
3800 if (tc->pool->pf.discard_enabled) {
3801 ti->discards_supported = true;
3802 ti->num_discard_bios = 1;
3803 /* Discard bios must be split on a block boundary */
3804 ti->split_discard_bios = true;
3807 mutex_unlock(&dm_thin_pool_table.mutex);
3809 spin_lock_irqsave(&tc->pool->lock, flags);
3810 if (tc->pool->suspended) {
3811 spin_unlock_irqrestore(&tc->pool->lock, flags);
3812 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
3813 ti->error = "Unable to activate thin device while pool is suspended";
3817 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
3818 spin_unlock_irqrestore(&tc->pool->lock, flags);
3820 * This synchronize_rcu() call is needed here otherwise we risk a
3821 * wake_worker() call finding no bios to process (because the newly
3822 * added tc isn't yet visible). So this reduces latency since we
3823 * aren't then dependent on the periodic commit to wake_worker().
3829 atomic_set(&tc->refcount, 1);
3830 init_completion(&tc->can_destroy);
3835 dm_pool_close_thin_device(tc->td);
3837 __pool_dec(tc->pool);
3841 dm_put_device(ti, tc->pool_dev);
3844 dm_put_device(ti, tc->origin_dev);
3848 mutex_unlock(&dm_thin_pool_table.mutex);
3853 static int thin_map(struct dm_target *ti, struct bio *bio)
3855 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
3857 return thin_bio_map(ti, bio);
3860 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
3862 unsigned long flags;
3863 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
3864 struct list_head work;
3865 struct dm_thin_new_mapping *m, *tmp;
3866 struct pool *pool = h->tc->pool;
3868 if (h->shared_read_entry) {
3869 INIT_LIST_HEAD(&work);
3870 dm_deferred_entry_dec(h->shared_read_entry, &work);
3872 spin_lock_irqsave(&pool->lock, flags);
3873 list_for_each_entry_safe(m, tmp, &work, list) {
3875 __complete_mapping_preparation(m);
3877 spin_unlock_irqrestore(&pool->lock, flags);
3880 if (h->all_io_entry) {
3881 INIT_LIST_HEAD(&work);
3882 dm_deferred_entry_dec(h->all_io_entry, &work);
3883 if (!list_empty(&work)) {
3884 spin_lock_irqsave(&pool->lock, flags);
3885 list_for_each_entry_safe(m, tmp, &work, list)
3886 list_add_tail(&m->list, &pool->prepared_discards);
3887 spin_unlock_irqrestore(&pool->lock, flags);
3895 static void thin_presuspend(struct dm_target *ti)
3897 struct thin_c *tc = ti->private;
3899 if (dm_noflush_suspending(ti))
3900 noflush_work(tc, do_noflush_start);
3903 static void thin_postsuspend(struct dm_target *ti)
3905 struct thin_c *tc = ti->private;
3908 * The dm_noflush_suspending flag has been cleared by now, so
3909 * unfortunately we must always run this.
3911 noflush_work(tc, do_noflush_stop);
3914 static int thin_preresume(struct dm_target *ti)
3916 struct thin_c *tc = ti->private;
3919 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
3925 * <nr mapped sectors> <highest mapped sector>
3927 static void thin_status(struct dm_target *ti, status_type_t type,
3928 unsigned status_flags, char *result, unsigned maxlen)
3932 dm_block_t mapped, highest;
3933 char buf[BDEVNAME_SIZE];
3934 struct thin_c *tc = ti->private;
3936 if (get_pool_mode(tc->pool) == PM_FAIL) {
3945 case STATUSTYPE_INFO:
3946 r = dm_thin_get_mapped_count(tc->td, &mapped);
3948 DMERR("dm_thin_get_mapped_count returned %d", r);
3952 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
3954 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
3958 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
3960 DMEMIT("%llu", ((highest + 1) *
3961 tc->pool->sectors_per_block) - 1);
3966 case STATUSTYPE_TABLE:
3968 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
3969 (unsigned long) tc->dev_id);
3971 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
3982 static int thin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3983 struct bio_vec *biovec, int max_size)
3985 struct thin_c *tc = ti->private;
3986 struct request_queue *q = bdev_get_queue(tc->pool_dev->bdev);
3988 if (!q->merge_bvec_fn)
3991 bvm->bi_bdev = tc->pool_dev->bdev;
3992 bvm->bi_sector = dm_target_offset(ti, bvm->bi_sector);
3994 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3997 static int thin_iterate_devices(struct dm_target *ti,
3998 iterate_devices_callout_fn fn, void *data)
4001 struct thin_c *tc = ti->private;
4002 struct pool *pool = tc->pool;
4005 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4006 * we follow a more convoluted path through to the pool's target.
4009 return 0; /* nothing is bound */
4011 blocks = pool->ti->len;
4012 (void) sector_div(blocks, pool->sectors_per_block);
4014 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4019 static struct target_type thin_target = {
4021 .version = {1, 14, 0},
4022 .module = THIS_MODULE,
4026 .end_io = thin_endio,
4027 .preresume = thin_preresume,
4028 .presuspend = thin_presuspend,
4029 .postsuspend = thin_postsuspend,
4030 .status = thin_status,
4031 .merge = thin_merge,
4032 .iterate_devices = thin_iterate_devices,
4035 /*----------------------------------------------------------------*/
4037 static int __init dm_thin_init(void)
4043 r = dm_register_target(&thin_target);
4047 r = dm_register_target(&pool_target);
4049 goto bad_pool_target;
4053 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4054 if (!_new_mapping_cache)
4055 goto bad_new_mapping_cache;
4059 bad_new_mapping_cache:
4060 dm_unregister_target(&pool_target);
4062 dm_unregister_target(&thin_target);
4067 static void dm_thin_exit(void)
4069 dm_unregister_target(&thin_target);
4070 dm_unregister_target(&pool_target);
4072 kmem_cache_destroy(_new_mapping_cache);
4075 module_init(dm_thin_init);
4076 module_exit(dm_thin_exit);
4078 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4079 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4081 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4082 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4083 MODULE_LICENSE("GPL");