2 * Copyright (C) 2011-2012 Red Hat UK.
4 * This file is released under the GPL.
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison.h"
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/log2.h>
16 #include <linux/list.h>
17 #include <linux/rculist.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
22 #include <linux/sort.h>
23 #include <linux/rbtree.h>
25 #define DM_MSG_PREFIX "thin"
30 #define ENDIO_HOOK_POOL_SIZE 1024
31 #define MAPPING_POOL_SIZE 1024
32 #define COMMIT_PERIOD HZ
33 #define NO_SPACE_TIMEOUT_SECS 60
35 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
37 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
38 "A percentage of time allocated for copy on write");
41 * The block size of the device holding pool data must be
42 * between 64KB and 1GB.
44 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
45 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
48 * Device id is restricted to 24 bits.
50 #define MAX_DEV_ID ((1 << 24) - 1)
53 * How do we handle breaking sharing of data blocks?
54 * =================================================
56 * We use a standard copy-on-write btree to store the mappings for the
57 * devices (note I'm talking about copy-on-write of the metadata here, not
58 * the data). When you take an internal snapshot you clone the root node
59 * of the origin btree. After this there is no concept of an origin or a
60 * snapshot. They are just two device trees that happen to point to the
63 * When we get a write in we decide if it's to a shared data block using
64 * some timestamp magic. If it is, we have to break sharing.
66 * Let's say we write to a shared block in what was the origin. The
69 * i) plug io further to this physical block. (see bio_prison code).
71 * ii) quiesce any read io to that shared data block. Obviously
72 * including all devices that share this block. (see dm_deferred_set code)
74 * iii) copy the data block to a newly allocate block. This step can be
75 * missed out if the io covers the block. (schedule_copy).
77 * iv) insert the new mapping into the origin's btree
78 * (process_prepared_mapping). This act of inserting breaks some
79 * sharing of btree nodes between the two devices. Breaking sharing only
80 * effects the btree of that specific device. Btrees for the other
81 * devices that share the block never change. The btree for the origin
82 * device as it was after the last commit is untouched, ie. we're using
83 * persistent data structures in the functional programming sense.
85 * v) unplug io to this physical block, including the io that triggered
86 * the breaking of sharing.
88 * Steps (ii) and (iii) occur in parallel.
90 * The metadata _doesn't_ need to be committed before the io continues. We
91 * get away with this because the io is always written to a _new_ block.
92 * If there's a crash, then:
94 * - The origin mapping will point to the old origin block (the shared
95 * one). This will contain the data as it was before the io that triggered
96 * the breaking of sharing came in.
98 * - The snap mapping still points to the old block. As it would after
101 * The downside of this scheme is the timestamp magic isn't perfect, and
102 * will continue to think that data block in the snapshot device is shared
103 * even after the write to the origin has broken sharing. I suspect data
104 * blocks will typically be shared by many different devices, so we're
105 * breaking sharing n + 1 times, rather than n, where n is the number of
106 * devices that reference this data block. At the moment I think the
107 * benefits far, far outweigh the disadvantages.
110 /*----------------------------------------------------------------*/
120 static void build_key(struct dm_thin_device *td, enum lock_space ls,
121 dm_block_t b, dm_block_t e, struct dm_cell_key *key)
123 key->virtual = (ls == VIRTUAL);
124 key->dev = dm_thin_dev_id(td);
125 key->block_begin = b;
129 static void build_data_key(struct dm_thin_device *td, dm_block_t b,
130 struct dm_cell_key *key)
132 build_key(td, PHYSICAL, b, b + 1llu, key);
135 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
136 struct dm_cell_key *key)
138 build_key(td, VIRTUAL, b, b + 1llu, key);
141 /*----------------------------------------------------------------*/
143 #define THROTTLE_THRESHOLD (1 * HZ)
146 struct rw_semaphore lock;
147 unsigned long threshold;
148 bool throttle_applied;
151 static void throttle_init(struct throttle *t)
153 init_rwsem(&t->lock);
154 t->throttle_applied = false;
157 static void throttle_work_start(struct throttle *t)
159 t->threshold = jiffies + THROTTLE_THRESHOLD;
162 static void throttle_work_update(struct throttle *t)
164 if (!t->throttle_applied && jiffies > t->threshold) {
165 down_write(&t->lock);
166 t->throttle_applied = true;
170 static void throttle_work_complete(struct throttle *t)
172 if (t->throttle_applied) {
173 t->throttle_applied = false;
178 static void throttle_lock(struct throttle *t)
183 static void throttle_unlock(struct throttle *t)
188 /*----------------------------------------------------------------*/
191 * A pool device ties together a metadata device and a data device. It
192 * also provides the interface for creating and destroying internal
195 struct dm_thin_new_mapping;
198 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
201 PM_WRITE, /* metadata may be changed */
202 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
203 PM_READ_ONLY, /* metadata may not be changed */
204 PM_FAIL, /* all I/O fails */
207 struct pool_features {
210 bool zero_new_blocks:1;
211 bool discard_enabled:1;
212 bool discard_passdown:1;
213 bool error_if_no_space:1;
217 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
218 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
219 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
221 #define CELL_SORT_ARRAY_SIZE 8192
224 struct list_head list;
225 struct dm_target *ti; /* Only set if a pool target is bound */
227 struct mapped_device *pool_md;
228 struct block_device *md_dev;
229 struct dm_pool_metadata *pmd;
231 dm_block_t low_water_blocks;
232 uint32_t sectors_per_block;
233 int sectors_per_block_shift;
235 struct pool_features pf;
236 bool low_water_triggered:1; /* A dm event has been sent */
239 struct dm_bio_prison *prison;
240 struct dm_kcopyd_client *copier;
242 struct workqueue_struct *wq;
243 struct throttle throttle;
244 struct work_struct worker;
245 struct delayed_work waker;
246 struct delayed_work no_space_timeout;
248 unsigned long last_commit_jiffies;
252 struct bio_list deferred_flush_bios;
253 struct list_head prepared_mappings;
254 struct list_head prepared_discards;
255 struct list_head active_thins;
257 struct dm_deferred_set *shared_read_ds;
258 struct dm_deferred_set *all_io_ds;
260 struct dm_thin_new_mapping *next_mapping;
261 mempool_t *mapping_pool;
263 process_bio_fn process_bio;
264 process_bio_fn process_discard;
266 process_cell_fn process_cell;
267 process_cell_fn process_discard_cell;
269 process_mapping_fn process_prepared_mapping;
270 process_mapping_fn process_prepared_discard;
272 struct dm_bio_prison_cell **cell_sort_array;
275 static enum pool_mode get_pool_mode(struct pool *pool);
276 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
279 * Target context for a pool.
282 struct dm_target *ti;
284 struct dm_dev *data_dev;
285 struct dm_dev *metadata_dev;
286 struct dm_target_callbacks callbacks;
288 dm_block_t low_water_blocks;
289 struct pool_features requested_pf; /* Features requested during table load */
290 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
294 * Target context for a thin.
297 struct list_head list;
298 struct dm_dev *pool_dev;
299 struct dm_dev *origin_dev;
300 sector_t origin_size;
304 struct dm_thin_device *td;
305 struct mapped_device *thin_md;
309 struct list_head deferred_cells;
310 struct bio_list deferred_bio_list;
311 struct bio_list retry_on_resume_list;
312 struct rb_root sort_bio_list; /* sorted list of deferred bios */
315 * Ensures the thin is not destroyed until the worker has finished
316 * iterating the active_thins list.
319 struct completion can_destroy;
322 /*----------------------------------------------------------------*/
325 * __blkdev_issue_discard_async - queue a discard with async completion
326 * @bdev: blockdev to issue discard for
327 * @sector: start sector
328 * @nr_sects: number of sectors to discard
329 * @gfp_mask: memory allocation flags (for bio_alloc)
330 * @flags: BLKDEV_IFL_* flags to control behaviour
331 * @parent_bio: parent discard bio that all sub discards get chained to
334 * Asynchronously issue a discard request for the sectors in question.
335 * NOTE: this variant of blk-core's blkdev_issue_discard() is a stop-gap
336 * that is being kept local to DM thinp until the block changes to allow
337 * late bio splitting land upstream.
339 static int __blkdev_issue_discard_async(struct block_device *bdev, sector_t sector,
340 sector_t nr_sects, gfp_t gfp_mask, unsigned long flags,
341 struct bio *parent_bio)
343 struct request_queue *q = bdev_get_queue(bdev);
344 int type = REQ_WRITE | REQ_DISCARD;
345 unsigned int max_discard_sectors, granularity;
349 struct blk_plug plug;
354 if (!blk_queue_discard(q))
357 /* Zero-sector (unknown) and one-sector granularities are the same. */
358 granularity = max(q->limits.discard_granularity >> 9, 1U);
359 alignment = (bdev_discard_alignment(bdev) >> 9) % granularity;
362 * Ensure that max_discard_sectors is of the proper
363 * granularity, so that requests stay aligned after a split.
365 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
366 max_discard_sectors -= max_discard_sectors % granularity;
367 if (unlikely(!max_discard_sectors)) {
368 /* Avoid infinite loop below. Being cautious never hurts. */
372 if (flags & BLKDEV_DISCARD_SECURE) {
373 if (!blk_queue_secdiscard(q))
378 blk_start_plug(&plug);
380 unsigned int req_sects;
381 sector_t end_sect, tmp;
384 * Required bio_put occurs in bio_endio thanks to bio_chain below
386 bio = bio_alloc(gfp_mask, 1);
392 req_sects = min_t(sector_t, nr_sects, max_discard_sectors);
395 * If splitting a request, and the next starting sector would be
396 * misaligned, stop the discard at the previous aligned sector.
398 end_sect = sector + req_sects;
400 if (req_sects < nr_sects &&
401 sector_div(tmp, granularity) != alignment) {
402 end_sect = end_sect - alignment;
403 sector_div(end_sect, granularity);
404 end_sect = end_sect * granularity + alignment;
405 req_sects = end_sect - sector;
408 bio_chain(bio, parent_bio);
410 bio->bi_iter.bi_sector = sector;
413 bio->bi_iter.bi_size = req_sects << 9;
414 nr_sects -= req_sects;
417 submit_bio(type, bio);
420 * We can loop for a long time in here, if someone does
421 * full device discards (like mkfs). Be nice and allow
422 * us to schedule out to avoid softlocking if preempt
427 blk_finish_plug(&plug);
432 static bool block_size_is_power_of_two(struct pool *pool)
434 return pool->sectors_per_block_shift >= 0;
437 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
439 return block_size_is_power_of_two(pool) ?
440 (b << pool->sectors_per_block_shift) :
441 (b * pool->sectors_per_block);
444 static int issue_discard(struct thin_c *tc, dm_block_t data_b, dm_block_t data_e,
445 struct bio *parent_bio)
447 sector_t s = block_to_sectors(tc->pool, data_b);
448 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
450 return __blkdev_issue_discard_async(tc->pool_dev->bdev, s, len,
451 GFP_NOWAIT, 0, parent_bio);
454 /*----------------------------------------------------------------*/
457 * wake_worker() is used when new work is queued and when pool_resume is
458 * ready to continue deferred IO processing.
460 static void wake_worker(struct pool *pool)
462 queue_work(pool->wq, &pool->worker);
465 /*----------------------------------------------------------------*/
467 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
468 struct dm_bio_prison_cell **cell_result)
471 struct dm_bio_prison_cell *cell_prealloc;
474 * Allocate a cell from the prison's mempool.
475 * This might block but it can't fail.
477 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
479 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
482 * We reused an old cell; we can get rid of
485 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
490 static void cell_release(struct pool *pool,
491 struct dm_bio_prison_cell *cell,
492 struct bio_list *bios)
494 dm_cell_release(pool->prison, cell, bios);
495 dm_bio_prison_free_cell(pool->prison, cell);
498 static void cell_visit_release(struct pool *pool,
499 void (*fn)(void *, struct dm_bio_prison_cell *),
501 struct dm_bio_prison_cell *cell)
503 dm_cell_visit_release(pool->prison, fn, context, cell);
504 dm_bio_prison_free_cell(pool->prison, cell);
507 static void cell_release_no_holder(struct pool *pool,
508 struct dm_bio_prison_cell *cell,
509 struct bio_list *bios)
511 dm_cell_release_no_holder(pool->prison, cell, bios);
512 dm_bio_prison_free_cell(pool->prison, cell);
515 static void cell_error_with_code(struct pool *pool,
516 struct dm_bio_prison_cell *cell, int error_code)
518 dm_cell_error(pool->prison, cell, error_code);
519 dm_bio_prison_free_cell(pool->prison, cell);
522 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
524 cell_error_with_code(pool, cell, -EIO);
527 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
529 cell_error_with_code(pool, cell, 0);
532 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
534 cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
537 /*----------------------------------------------------------------*/
540 * A global list of pools that uses a struct mapped_device as a key.
542 static struct dm_thin_pool_table {
544 struct list_head pools;
545 } dm_thin_pool_table;
547 static void pool_table_init(void)
549 mutex_init(&dm_thin_pool_table.mutex);
550 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
553 static void __pool_table_insert(struct pool *pool)
555 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
556 list_add(&pool->list, &dm_thin_pool_table.pools);
559 static void __pool_table_remove(struct pool *pool)
561 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
562 list_del(&pool->list);
565 static struct pool *__pool_table_lookup(struct mapped_device *md)
567 struct pool *pool = NULL, *tmp;
569 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
571 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
572 if (tmp->pool_md == md) {
581 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
583 struct pool *pool = NULL, *tmp;
585 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
587 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
588 if (tmp->md_dev == md_dev) {
597 /*----------------------------------------------------------------*/
599 struct dm_thin_endio_hook {
601 struct dm_deferred_entry *shared_read_entry;
602 struct dm_deferred_entry *all_io_entry;
603 struct dm_thin_new_mapping *overwrite_mapping;
604 struct rb_node rb_node;
605 struct dm_bio_prison_cell *cell;
608 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
610 bio_list_merge(bios, master);
611 bio_list_init(master);
614 static void error_bio_list(struct bio_list *bios, int error)
618 while ((bio = bio_list_pop(bios)))
619 bio_endio(bio, error);
622 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
624 struct bio_list bios;
627 bio_list_init(&bios);
629 spin_lock_irqsave(&tc->lock, flags);
630 __merge_bio_list(&bios, master);
631 spin_unlock_irqrestore(&tc->lock, flags);
633 error_bio_list(&bios, error);
636 static void requeue_deferred_cells(struct thin_c *tc)
638 struct pool *pool = tc->pool;
640 struct list_head cells;
641 struct dm_bio_prison_cell *cell, *tmp;
643 INIT_LIST_HEAD(&cells);
645 spin_lock_irqsave(&tc->lock, flags);
646 list_splice_init(&tc->deferred_cells, &cells);
647 spin_unlock_irqrestore(&tc->lock, flags);
649 list_for_each_entry_safe(cell, tmp, &cells, user_list)
650 cell_requeue(pool, cell);
653 static void requeue_io(struct thin_c *tc)
655 struct bio_list bios;
658 bio_list_init(&bios);
660 spin_lock_irqsave(&tc->lock, flags);
661 __merge_bio_list(&bios, &tc->deferred_bio_list);
662 __merge_bio_list(&bios, &tc->retry_on_resume_list);
663 spin_unlock_irqrestore(&tc->lock, flags);
665 error_bio_list(&bios, DM_ENDIO_REQUEUE);
666 requeue_deferred_cells(tc);
669 static void error_retry_list(struct pool *pool)
674 list_for_each_entry_rcu(tc, &pool->active_thins, list)
675 error_thin_bio_list(tc, &tc->retry_on_resume_list, -EIO);
680 * This section of code contains the logic for processing a thin device's IO.
681 * Much of the code depends on pool object resources (lists, workqueues, etc)
682 * but most is exclusively called from the thin target rather than the thin-pool
686 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
688 struct pool *pool = tc->pool;
689 sector_t block_nr = bio->bi_iter.bi_sector;
691 if (block_size_is_power_of_two(pool))
692 block_nr >>= pool->sectors_per_block_shift;
694 (void) sector_div(block_nr, pool->sectors_per_block);
700 * Returns the _complete_ blocks that this bio covers.
702 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
703 dm_block_t *begin, dm_block_t *end)
705 struct pool *pool = tc->pool;
706 sector_t b = bio->bi_iter.bi_sector;
707 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
709 b += pool->sectors_per_block - 1ull; /* so we round up */
711 if (block_size_is_power_of_two(pool)) {
712 b >>= pool->sectors_per_block_shift;
713 e >>= pool->sectors_per_block_shift;
715 (void) sector_div(b, pool->sectors_per_block);
716 (void) sector_div(e, pool->sectors_per_block);
720 /* Can happen if the bio is within a single block. */
727 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
729 struct pool *pool = tc->pool;
730 sector_t bi_sector = bio->bi_iter.bi_sector;
732 bio->bi_bdev = tc->pool_dev->bdev;
733 if (block_size_is_power_of_two(pool))
734 bio->bi_iter.bi_sector =
735 (block << pool->sectors_per_block_shift) |
736 (bi_sector & (pool->sectors_per_block - 1));
738 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
739 sector_div(bi_sector, pool->sectors_per_block);
742 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
744 bio->bi_bdev = tc->origin_dev->bdev;
747 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
749 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
750 dm_thin_changed_this_transaction(tc->td);
753 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
755 struct dm_thin_endio_hook *h;
757 if (bio->bi_rw & REQ_DISCARD)
760 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
761 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
764 static void issue(struct thin_c *tc, struct bio *bio)
766 struct pool *pool = tc->pool;
769 if (!bio_triggers_commit(tc, bio)) {
770 generic_make_request(bio);
775 * Complete bio with an error if earlier I/O caused changes to
776 * the metadata that can't be committed e.g, due to I/O errors
777 * on the metadata device.
779 if (dm_thin_aborted_changes(tc->td)) {
785 * Batch together any bios that trigger commits and then issue a
786 * single commit for them in process_deferred_bios().
788 spin_lock_irqsave(&pool->lock, flags);
789 bio_list_add(&pool->deferred_flush_bios, bio);
790 spin_unlock_irqrestore(&pool->lock, flags);
793 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
795 remap_to_origin(tc, bio);
799 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
802 remap(tc, bio, block);
806 /*----------------------------------------------------------------*/
809 * Bio endio functions.
811 struct dm_thin_new_mapping {
812 struct list_head list;
818 * Track quiescing, copying and zeroing preparation actions. When this
819 * counter hits zero the block is prepared and can be inserted into the
822 atomic_t prepare_actions;
826 dm_block_t virt_begin, virt_end;
827 dm_block_t data_block;
828 struct dm_bio_prison_cell *cell;
831 * If the bio covers the whole area of a block then we can avoid
832 * zeroing or copying. Instead this bio is hooked. The bio will
833 * still be in the cell, so care has to be taken to avoid issuing
837 bio_end_io_t *saved_bi_end_io;
840 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
842 struct pool *pool = m->tc->pool;
844 if (atomic_dec_and_test(&m->prepare_actions)) {
845 list_add_tail(&m->list, &pool->prepared_mappings);
850 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
853 struct pool *pool = m->tc->pool;
855 spin_lock_irqsave(&pool->lock, flags);
856 __complete_mapping_preparation(m);
857 spin_unlock_irqrestore(&pool->lock, flags);
860 static void copy_complete(int read_err, unsigned long write_err, void *context)
862 struct dm_thin_new_mapping *m = context;
864 m->err = read_err || write_err ? -EIO : 0;
865 complete_mapping_preparation(m);
868 static void overwrite_endio(struct bio *bio, int err)
870 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
871 struct dm_thin_new_mapping *m = h->overwrite_mapping;
873 bio->bi_end_io = m->saved_bi_end_io;
876 complete_mapping_preparation(m);
879 /*----------------------------------------------------------------*/
886 * Prepared mapping jobs.
890 * This sends the bios in the cell, except the original holder, back
891 * to the deferred_bios list.
893 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
895 struct pool *pool = tc->pool;
898 spin_lock_irqsave(&tc->lock, flags);
899 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
900 spin_unlock_irqrestore(&tc->lock, flags);
905 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
909 struct bio_list defer_bios;
910 struct bio_list issue_bios;
913 static void __inc_remap_and_issue_cell(void *context,
914 struct dm_bio_prison_cell *cell)
916 struct remap_info *info = context;
919 while ((bio = bio_list_pop(&cell->bios))) {
920 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA))
921 bio_list_add(&info->defer_bios, bio);
923 inc_all_io_entry(info->tc->pool, bio);
926 * We can't issue the bios with the bio prison lock
927 * held, so we add them to a list to issue on
928 * return from this function.
930 bio_list_add(&info->issue_bios, bio);
935 static void inc_remap_and_issue_cell(struct thin_c *tc,
936 struct dm_bio_prison_cell *cell,
940 struct remap_info info;
943 bio_list_init(&info.defer_bios);
944 bio_list_init(&info.issue_bios);
947 * We have to be careful to inc any bios we're about to issue
948 * before the cell is released, and avoid a race with new bios
949 * being added to the cell.
951 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
954 while ((bio = bio_list_pop(&info.defer_bios)))
955 thin_defer_bio(tc, bio);
957 while ((bio = bio_list_pop(&info.issue_bios)))
958 remap_and_issue(info.tc, bio, block);
961 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
963 cell_error(m->tc->pool, m->cell);
965 mempool_free(m, m->tc->pool->mapping_pool);
968 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
970 struct thin_c *tc = m->tc;
971 struct pool *pool = tc->pool;
972 struct bio *bio = m->bio;
976 cell_error(pool, m->cell);
981 * Commit the prepared block into the mapping btree.
982 * Any I/O for this block arriving after this point will get
983 * remapped to it directly.
985 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
987 metadata_operation_failed(pool, "dm_thin_insert_block", r);
988 cell_error(pool, m->cell);
993 * Release any bios held while the block was being provisioned.
994 * If we are processing a write bio that completely covers the block,
995 * we already processed it so can ignore it now when processing
996 * the bios in the cell.
999 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1002 inc_all_io_entry(tc->pool, m->cell->holder);
1003 remap_and_issue(tc, m->cell->holder, m->data_block);
1004 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1009 mempool_free(m, pool->mapping_pool);
1012 /*----------------------------------------------------------------*/
1014 static void free_discard_mapping(struct dm_thin_new_mapping *m)
1016 struct thin_c *tc = m->tc;
1018 cell_defer_no_holder(tc, m->cell);
1019 mempool_free(m, tc->pool->mapping_pool);
1022 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1024 bio_io_error(m->bio);
1025 free_discard_mapping(m);
1028 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1030 bio_endio(m->bio, 0);
1031 free_discard_mapping(m);
1034 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1037 struct thin_c *tc = m->tc;
1039 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1041 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1042 bio_io_error(m->bio);
1044 bio_endio(m->bio, 0);
1046 cell_defer_no_holder(tc, m->cell);
1047 mempool_free(m, tc->pool->mapping_pool);
1050 static int passdown_double_checking_shared_status(struct dm_thin_new_mapping *m)
1053 * We've already unmapped this range of blocks, but before we
1054 * passdown we have to check that these blocks are now unused.
1058 struct thin_c *tc = m->tc;
1059 struct pool *pool = tc->pool;
1060 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1063 /* find start of unmapped run */
1064 for (; b < end; b++) {
1065 r = dm_pool_block_is_used(pool->pmd, b, &used);
1076 /* find end of run */
1077 for (e = b + 1; e != end; e++) {
1078 r = dm_pool_block_is_used(pool->pmd, e, &used);
1086 r = issue_discard(tc, b, e, m->bio);
1096 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
1099 struct thin_c *tc = m->tc;
1100 struct pool *pool = tc->pool;
1102 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1104 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1106 else if (m->maybe_shared)
1107 r = passdown_double_checking_shared_status(m);
1109 r = issue_discard(tc, m->data_block, m->data_block + (m->virt_end - m->virt_begin), m->bio);
1112 * Even if r is set, there could be sub discards in flight that we
1115 bio_endio(m->bio, r);
1116 cell_defer_no_holder(tc, m->cell);
1117 mempool_free(m, pool->mapping_pool);
1120 static void process_prepared(struct pool *pool, struct list_head *head,
1121 process_mapping_fn *fn)
1123 unsigned long flags;
1124 struct list_head maps;
1125 struct dm_thin_new_mapping *m, *tmp;
1127 INIT_LIST_HEAD(&maps);
1128 spin_lock_irqsave(&pool->lock, flags);
1129 list_splice_init(head, &maps);
1130 spin_unlock_irqrestore(&pool->lock, flags);
1132 list_for_each_entry_safe(m, tmp, &maps, list)
1137 * Deferred bio jobs.
1139 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1141 return bio->bi_iter.bi_size ==
1142 (pool->sectors_per_block << SECTOR_SHIFT);
1145 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1147 return (bio_data_dir(bio) == WRITE) &&
1148 io_overlaps_block(pool, bio);
1151 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1154 *save = bio->bi_end_io;
1155 bio->bi_end_io = fn;
1158 static int ensure_next_mapping(struct pool *pool)
1160 if (pool->next_mapping)
1163 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
1165 return pool->next_mapping ? 0 : -ENOMEM;
1168 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1170 struct dm_thin_new_mapping *m = pool->next_mapping;
1172 BUG_ON(!pool->next_mapping);
1174 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1175 INIT_LIST_HEAD(&m->list);
1178 pool->next_mapping = NULL;
1183 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1184 sector_t begin, sector_t end)
1187 struct dm_io_region to;
1189 to.bdev = tc->pool_dev->bdev;
1191 to.count = end - begin;
1193 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1195 DMERR_LIMIT("dm_kcopyd_zero() failed");
1196 copy_complete(1, 1, m);
1200 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1201 dm_block_t data_begin,
1202 struct dm_thin_new_mapping *m)
1204 struct pool *pool = tc->pool;
1205 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1207 h->overwrite_mapping = m;
1209 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1210 inc_all_io_entry(pool, bio);
1211 remap_and_issue(tc, bio, data_begin);
1215 * A partial copy also needs to zero the uncopied region.
1217 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1218 struct dm_dev *origin, dm_block_t data_origin,
1219 dm_block_t data_dest,
1220 struct dm_bio_prison_cell *cell, struct bio *bio,
1224 struct pool *pool = tc->pool;
1225 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1228 m->virt_begin = virt_block;
1229 m->virt_end = virt_block + 1u;
1230 m->data_block = data_dest;
1234 * quiesce action + copy action + an extra reference held for the
1235 * duration of this function (we may need to inc later for a
1238 atomic_set(&m->prepare_actions, 3);
1240 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1241 complete_mapping_preparation(m); /* already quiesced */
1244 * IO to pool_dev remaps to the pool target's data_dev.
1246 * If the whole block of data is being overwritten, we can issue the
1247 * bio immediately. Otherwise we use kcopyd to clone the data first.
1249 if (io_overwrites_block(pool, bio))
1250 remap_and_issue_overwrite(tc, bio, data_dest, m);
1252 struct dm_io_region from, to;
1254 from.bdev = origin->bdev;
1255 from.sector = data_origin * pool->sectors_per_block;
1258 to.bdev = tc->pool_dev->bdev;
1259 to.sector = data_dest * pool->sectors_per_block;
1262 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1263 0, copy_complete, m);
1265 DMERR_LIMIT("dm_kcopyd_copy() failed");
1266 copy_complete(1, 1, m);
1269 * We allow the zero to be issued, to simplify the
1270 * error path. Otherwise we'd need to start
1271 * worrying about decrementing the prepare_actions
1277 * Do we need to zero a tail region?
1279 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1280 atomic_inc(&m->prepare_actions);
1282 data_dest * pool->sectors_per_block + len,
1283 (data_dest + 1) * pool->sectors_per_block);
1287 complete_mapping_preparation(m); /* drop our ref */
1290 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1291 dm_block_t data_origin, dm_block_t data_dest,
1292 struct dm_bio_prison_cell *cell, struct bio *bio)
1294 schedule_copy(tc, virt_block, tc->pool_dev,
1295 data_origin, data_dest, cell, bio,
1296 tc->pool->sectors_per_block);
1299 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1300 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1303 struct pool *pool = tc->pool;
1304 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1306 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1308 m->virt_begin = virt_block;
1309 m->virt_end = virt_block + 1u;
1310 m->data_block = data_block;
1314 * If the whole block of data is being overwritten or we are not
1315 * zeroing pre-existing data, we can issue the bio immediately.
1316 * Otherwise we use kcopyd to zero the data first.
1318 if (pool->pf.zero_new_blocks) {
1319 if (io_overwrites_block(pool, bio))
1320 remap_and_issue_overwrite(tc, bio, data_block, m);
1322 ll_zero(tc, m, data_block * pool->sectors_per_block,
1323 (data_block + 1) * pool->sectors_per_block);
1325 process_prepared_mapping(m);
1328 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1329 dm_block_t data_dest,
1330 struct dm_bio_prison_cell *cell, struct bio *bio)
1332 struct pool *pool = tc->pool;
1333 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1334 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1336 if (virt_block_end <= tc->origin_size)
1337 schedule_copy(tc, virt_block, tc->origin_dev,
1338 virt_block, data_dest, cell, bio,
1339 pool->sectors_per_block);
1341 else if (virt_block_begin < tc->origin_size)
1342 schedule_copy(tc, virt_block, tc->origin_dev,
1343 virt_block, data_dest, cell, bio,
1344 tc->origin_size - virt_block_begin);
1347 schedule_zero(tc, virt_block, data_dest, cell, bio);
1350 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1352 static void check_for_space(struct pool *pool)
1357 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1360 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1365 set_pool_mode(pool, PM_WRITE);
1369 * A non-zero return indicates read_only or fail_io mode.
1370 * Many callers don't care about the return value.
1372 static int commit(struct pool *pool)
1376 if (get_pool_mode(pool) >= PM_READ_ONLY)
1379 r = dm_pool_commit_metadata(pool->pmd);
1381 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1383 check_for_space(pool);
1388 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1390 unsigned long flags;
1392 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1393 DMWARN("%s: reached low water mark for data device: sending event.",
1394 dm_device_name(pool->pool_md));
1395 spin_lock_irqsave(&pool->lock, flags);
1396 pool->low_water_triggered = true;
1397 spin_unlock_irqrestore(&pool->lock, flags);
1398 dm_table_event(pool->ti->table);
1402 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1405 dm_block_t free_blocks;
1406 struct pool *pool = tc->pool;
1408 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1411 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1413 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1417 check_low_water_mark(pool, free_blocks);
1421 * Try to commit to see if that will free up some
1428 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1430 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1435 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1440 r = dm_pool_alloc_data_block(pool->pmd, result);
1442 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1450 * If we have run out of space, queue bios until the device is
1451 * resumed, presumably after having been reloaded with more space.
1453 static void retry_on_resume(struct bio *bio)
1455 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1456 struct thin_c *tc = h->tc;
1457 unsigned long flags;
1459 spin_lock_irqsave(&tc->lock, flags);
1460 bio_list_add(&tc->retry_on_resume_list, bio);
1461 spin_unlock_irqrestore(&tc->lock, flags);
1464 static int should_error_unserviceable_bio(struct pool *pool)
1466 enum pool_mode m = get_pool_mode(pool);
1470 /* Shouldn't get here */
1471 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1474 case PM_OUT_OF_DATA_SPACE:
1475 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1481 /* Shouldn't get here */
1482 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1487 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1489 int error = should_error_unserviceable_bio(pool);
1492 bio_endio(bio, error);
1494 retry_on_resume(bio);
1497 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1500 struct bio_list bios;
1503 error = should_error_unserviceable_bio(pool);
1505 cell_error_with_code(pool, cell, error);
1509 bio_list_init(&bios);
1510 cell_release(pool, cell, &bios);
1512 while ((bio = bio_list_pop(&bios)))
1513 retry_on_resume(bio);
1516 static void process_discard_cell_no_passdown(struct thin_c *tc,
1517 struct dm_bio_prison_cell *virt_cell)
1519 struct pool *pool = tc->pool;
1520 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1523 * We don't need to lock the data blocks, since there's no
1524 * passdown. We only lock data blocks for allocation and breaking sharing.
1527 m->virt_begin = virt_cell->key.block_begin;
1528 m->virt_end = virt_cell->key.block_end;
1529 m->cell = virt_cell;
1530 m->bio = virt_cell->holder;
1532 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1533 pool->process_prepared_discard(m);
1537 * FIXME: DM local hack to defer parent bios's end_io until we
1538 * _know_ all chained sub range discard bios have completed.
1539 * Will go away once late bio splitting lands upstream!
1541 static inline void __bio_inc_remaining(struct bio *bio)
1543 bio->bi_flags |= (1 << BIO_CHAIN);
1544 smp_mb__before_atomic();
1545 atomic_inc(&bio->__bi_remaining);
1548 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1551 struct pool *pool = tc->pool;
1555 struct dm_cell_key data_key;
1556 struct dm_bio_prison_cell *data_cell;
1557 struct dm_thin_new_mapping *m;
1558 dm_block_t virt_begin, virt_end, data_begin;
1560 while (begin != end) {
1561 r = ensure_next_mapping(pool);
1563 /* we did our best */
1566 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1567 &data_begin, &maybe_shared);
1570 * Silently fail, letting any mappings we've
1575 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1576 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1577 /* contention, we'll give up with this range */
1583 * IO may still be going to the destination block. We must
1584 * quiesce before we can do the removal.
1586 m = get_next_mapping(pool);
1588 m->maybe_shared = maybe_shared;
1589 m->virt_begin = virt_begin;
1590 m->virt_end = virt_end;
1591 m->data_block = data_begin;
1592 m->cell = data_cell;
1596 * The parent bio must not complete before sub discard bios are
1597 * chained to it (see __blkdev_issue_discard_async's bio_chain)!
1599 * This per-mapping bi_remaining increment is paired with
1600 * the implicit decrement that occurs via bio_endio() in
1601 * process_prepared_discard_{passdown,no_passdown}.
1603 __bio_inc_remaining(bio);
1604 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1605 pool->process_prepared_discard(m);
1611 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1613 struct bio *bio = virt_cell->holder;
1614 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1617 * The virt_cell will only get freed once the origin bio completes.
1618 * This means it will remain locked while all the individual
1619 * passdown bios are in flight.
1621 h->cell = virt_cell;
1622 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1625 * We complete the bio now, knowing that the bi_remaining field
1626 * will prevent completion until the sub range discards have
1632 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1634 dm_block_t begin, end;
1635 struct dm_cell_key virt_key;
1636 struct dm_bio_prison_cell *virt_cell;
1638 get_bio_block_range(tc, bio, &begin, &end);
1641 * The discard covers less than a block.
1647 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1648 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1650 * Potential starvation issue: We're relying on the
1651 * fs/application being well behaved, and not trying to
1652 * send IO to a region at the same time as discarding it.
1653 * If they do this persistently then it's possible this
1654 * cell will never be granted.
1658 tc->pool->process_discard_cell(tc, virt_cell);
1661 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1662 struct dm_cell_key *key,
1663 struct dm_thin_lookup_result *lookup_result,
1664 struct dm_bio_prison_cell *cell)
1667 dm_block_t data_block;
1668 struct pool *pool = tc->pool;
1670 r = alloc_data_block(tc, &data_block);
1673 schedule_internal_copy(tc, block, lookup_result->block,
1674 data_block, cell, bio);
1678 retry_bios_on_resume(pool, cell);
1682 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1684 cell_error(pool, cell);
1689 static void __remap_and_issue_shared_cell(void *context,
1690 struct dm_bio_prison_cell *cell)
1692 struct remap_info *info = context;
1695 while ((bio = bio_list_pop(&cell->bios))) {
1696 if ((bio_data_dir(bio) == WRITE) ||
1697 (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)))
1698 bio_list_add(&info->defer_bios, bio);
1700 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1702 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1703 inc_all_io_entry(info->tc->pool, bio);
1704 bio_list_add(&info->issue_bios, bio);
1709 static void remap_and_issue_shared_cell(struct thin_c *tc,
1710 struct dm_bio_prison_cell *cell,
1714 struct remap_info info;
1717 bio_list_init(&info.defer_bios);
1718 bio_list_init(&info.issue_bios);
1720 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1723 while ((bio = bio_list_pop(&info.defer_bios)))
1724 thin_defer_bio(tc, bio);
1726 while ((bio = bio_list_pop(&info.issue_bios)))
1727 remap_and_issue(tc, bio, block);
1730 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1732 struct dm_thin_lookup_result *lookup_result,
1733 struct dm_bio_prison_cell *virt_cell)
1735 struct dm_bio_prison_cell *data_cell;
1736 struct pool *pool = tc->pool;
1737 struct dm_cell_key key;
1740 * If cell is already occupied, then sharing is already in the process
1741 * of being broken so we have nothing further to do here.
1743 build_data_key(tc->td, lookup_result->block, &key);
1744 if (bio_detain(pool, &key, bio, &data_cell)) {
1745 cell_defer_no_holder(tc, virt_cell);
1749 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1750 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1751 cell_defer_no_holder(tc, virt_cell);
1753 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1755 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1756 inc_all_io_entry(pool, bio);
1757 remap_and_issue(tc, bio, lookup_result->block);
1759 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1760 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1764 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1765 struct dm_bio_prison_cell *cell)
1768 dm_block_t data_block;
1769 struct pool *pool = tc->pool;
1772 * Remap empty bios (flushes) immediately, without provisioning.
1774 if (!bio->bi_iter.bi_size) {
1775 inc_all_io_entry(pool, bio);
1776 cell_defer_no_holder(tc, cell);
1778 remap_and_issue(tc, bio, 0);
1783 * Fill read bios with zeroes and complete them immediately.
1785 if (bio_data_dir(bio) == READ) {
1787 cell_defer_no_holder(tc, cell);
1792 r = alloc_data_block(tc, &data_block);
1796 schedule_external_copy(tc, block, data_block, cell, bio);
1798 schedule_zero(tc, block, data_block, cell, bio);
1802 retry_bios_on_resume(pool, cell);
1806 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1808 cell_error(pool, cell);
1813 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1816 struct pool *pool = tc->pool;
1817 struct bio *bio = cell->holder;
1818 dm_block_t block = get_bio_block(tc, bio);
1819 struct dm_thin_lookup_result lookup_result;
1821 if (tc->requeue_mode) {
1822 cell_requeue(pool, cell);
1826 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1829 if (lookup_result.shared)
1830 process_shared_bio(tc, bio, block, &lookup_result, cell);
1832 inc_all_io_entry(pool, bio);
1833 remap_and_issue(tc, bio, lookup_result.block);
1834 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1839 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1840 inc_all_io_entry(pool, bio);
1841 cell_defer_no_holder(tc, cell);
1843 if (bio_end_sector(bio) <= tc->origin_size)
1844 remap_to_origin_and_issue(tc, bio);
1846 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1848 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1849 remap_to_origin_and_issue(tc, bio);
1856 provision_block(tc, bio, block, cell);
1860 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1862 cell_defer_no_holder(tc, cell);
1868 static void process_bio(struct thin_c *tc, struct bio *bio)
1870 struct pool *pool = tc->pool;
1871 dm_block_t block = get_bio_block(tc, bio);
1872 struct dm_bio_prison_cell *cell;
1873 struct dm_cell_key key;
1876 * If cell is already occupied, then the block is already
1877 * being provisioned so we have nothing further to do here.
1879 build_virtual_key(tc->td, block, &key);
1880 if (bio_detain(pool, &key, bio, &cell))
1883 process_cell(tc, cell);
1886 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1887 struct dm_bio_prison_cell *cell)
1890 int rw = bio_data_dir(bio);
1891 dm_block_t block = get_bio_block(tc, bio);
1892 struct dm_thin_lookup_result lookup_result;
1894 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1897 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1898 handle_unserviceable_bio(tc->pool, bio);
1900 cell_defer_no_holder(tc, cell);
1902 inc_all_io_entry(tc->pool, bio);
1903 remap_and_issue(tc, bio, lookup_result.block);
1905 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1911 cell_defer_no_holder(tc, cell);
1913 handle_unserviceable_bio(tc->pool, bio);
1917 if (tc->origin_dev) {
1918 inc_all_io_entry(tc->pool, bio);
1919 remap_to_origin_and_issue(tc, bio);
1928 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1931 cell_defer_no_holder(tc, cell);
1937 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1939 __process_bio_read_only(tc, bio, NULL);
1942 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1944 __process_bio_read_only(tc, cell->holder, cell);
1947 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1952 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1957 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1959 cell_success(tc->pool, cell);
1962 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1964 cell_error(tc->pool, cell);
1968 * FIXME: should we also commit due to size of transaction, measured in
1971 static int need_commit_due_to_time(struct pool *pool)
1973 return !time_in_range(jiffies, pool->last_commit_jiffies,
1974 pool->last_commit_jiffies + COMMIT_PERIOD);
1977 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1978 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1980 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1982 struct rb_node **rbp, *parent;
1983 struct dm_thin_endio_hook *pbd;
1984 sector_t bi_sector = bio->bi_iter.bi_sector;
1986 rbp = &tc->sort_bio_list.rb_node;
1990 pbd = thin_pbd(parent);
1992 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1993 rbp = &(*rbp)->rb_left;
1995 rbp = &(*rbp)->rb_right;
1998 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1999 rb_link_node(&pbd->rb_node, parent, rbp);
2000 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2003 static void __extract_sorted_bios(struct thin_c *tc)
2005 struct rb_node *node;
2006 struct dm_thin_endio_hook *pbd;
2009 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2010 pbd = thin_pbd(node);
2011 bio = thin_bio(pbd);
2013 bio_list_add(&tc->deferred_bio_list, bio);
2014 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2017 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2020 static void __sort_thin_deferred_bios(struct thin_c *tc)
2023 struct bio_list bios;
2025 bio_list_init(&bios);
2026 bio_list_merge(&bios, &tc->deferred_bio_list);
2027 bio_list_init(&tc->deferred_bio_list);
2029 /* Sort deferred_bio_list using rb-tree */
2030 while ((bio = bio_list_pop(&bios)))
2031 __thin_bio_rb_add(tc, bio);
2034 * Transfer the sorted bios in sort_bio_list back to
2035 * deferred_bio_list to allow lockless submission of
2038 __extract_sorted_bios(tc);
2041 static void process_thin_deferred_bios(struct thin_c *tc)
2043 struct pool *pool = tc->pool;
2044 unsigned long flags;
2046 struct bio_list bios;
2047 struct blk_plug plug;
2050 if (tc->requeue_mode) {
2051 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
2055 bio_list_init(&bios);
2057 spin_lock_irqsave(&tc->lock, flags);
2059 if (bio_list_empty(&tc->deferred_bio_list)) {
2060 spin_unlock_irqrestore(&tc->lock, flags);
2064 __sort_thin_deferred_bios(tc);
2066 bio_list_merge(&bios, &tc->deferred_bio_list);
2067 bio_list_init(&tc->deferred_bio_list);
2069 spin_unlock_irqrestore(&tc->lock, flags);
2071 blk_start_plug(&plug);
2072 while ((bio = bio_list_pop(&bios))) {
2074 * If we've got no free new_mapping structs, and processing
2075 * this bio might require one, we pause until there are some
2076 * prepared mappings to process.
2078 if (ensure_next_mapping(pool)) {
2079 spin_lock_irqsave(&tc->lock, flags);
2080 bio_list_add(&tc->deferred_bio_list, bio);
2081 bio_list_merge(&tc->deferred_bio_list, &bios);
2082 spin_unlock_irqrestore(&tc->lock, flags);
2086 if (bio->bi_rw & REQ_DISCARD)
2087 pool->process_discard(tc, bio);
2089 pool->process_bio(tc, bio);
2091 if ((count++ & 127) == 0) {
2092 throttle_work_update(&pool->throttle);
2093 dm_pool_issue_prefetches(pool->pmd);
2096 blk_finish_plug(&plug);
2099 static int cmp_cells(const void *lhs, const void *rhs)
2101 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2102 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2104 BUG_ON(!lhs_cell->holder);
2105 BUG_ON(!rhs_cell->holder);
2107 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2110 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2116 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2119 struct dm_bio_prison_cell *cell, *tmp;
2121 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2122 if (count >= CELL_SORT_ARRAY_SIZE)
2125 pool->cell_sort_array[count++] = cell;
2126 list_del(&cell->user_list);
2129 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2134 static void process_thin_deferred_cells(struct thin_c *tc)
2136 struct pool *pool = tc->pool;
2137 unsigned long flags;
2138 struct list_head cells;
2139 struct dm_bio_prison_cell *cell;
2140 unsigned i, j, count;
2142 INIT_LIST_HEAD(&cells);
2144 spin_lock_irqsave(&tc->lock, flags);
2145 list_splice_init(&tc->deferred_cells, &cells);
2146 spin_unlock_irqrestore(&tc->lock, flags);
2148 if (list_empty(&cells))
2152 count = sort_cells(tc->pool, &cells);
2154 for (i = 0; i < count; i++) {
2155 cell = pool->cell_sort_array[i];
2156 BUG_ON(!cell->holder);
2159 * If we've got no free new_mapping structs, and processing
2160 * this bio might require one, we pause until there are some
2161 * prepared mappings to process.
2163 if (ensure_next_mapping(pool)) {
2164 for (j = i; j < count; j++)
2165 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2167 spin_lock_irqsave(&tc->lock, flags);
2168 list_splice(&cells, &tc->deferred_cells);
2169 spin_unlock_irqrestore(&tc->lock, flags);
2173 if (cell->holder->bi_rw & REQ_DISCARD)
2174 pool->process_discard_cell(tc, cell);
2176 pool->process_cell(tc, cell);
2178 } while (!list_empty(&cells));
2181 static void thin_get(struct thin_c *tc);
2182 static void thin_put(struct thin_c *tc);
2185 * We can't hold rcu_read_lock() around code that can block. So we
2186 * find a thin with the rcu lock held; bump a refcount; then drop
2189 static struct thin_c *get_first_thin(struct pool *pool)
2191 struct thin_c *tc = NULL;
2194 if (!list_empty(&pool->active_thins)) {
2195 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2203 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2205 struct thin_c *old_tc = tc;
2208 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2220 static void process_deferred_bios(struct pool *pool)
2222 unsigned long flags;
2224 struct bio_list bios;
2227 tc = get_first_thin(pool);
2229 process_thin_deferred_cells(tc);
2230 process_thin_deferred_bios(tc);
2231 tc = get_next_thin(pool, tc);
2235 * If there are any deferred flush bios, we must commit
2236 * the metadata before issuing them.
2238 bio_list_init(&bios);
2239 spin_lock_irqsave(&pool->lock, flags);
2240 bio_list_merge(&bios, &pool->deferred_flush_bios);
2241 bio_list_init(&pool->deferred_flush_bios);
2242 spin_unlock_irqrestore(&pool->lock, flags);
2244 if (bio_list_empty(&bios) &&
2245 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2249 while ((bio = bio_list_pop(&bios)))
2253 pool->last_commit_jiffies = jiffies;
2255 while ((bio = bio_list_pop(&bios)))
2256 generic_make_request(bio);
2259 static void do_worker(struct work_struct *ws)
2261 struct pool *pool = container_of(ws, struct pool, worker);
2263 throttle_work_start(&pool->throttle);
2264 dm_pool_issue_prefetches(pool->pmd);
2265 throttle_work_update(&pool->throttle);
2266 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2267 throttle_work_update(&pool->throttle);
2268 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2269 throttle_work_update(&pool->throttle);
2270 process_deferred_bios(pool);
2271 throttle_work_complete(&pool->throttle);
2275 * We want to commit periodically so that not too much
2276 * unwritten data builds up.
2278 static void do_waker(struct work_struct *ws)
2280 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2282 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2285 static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2288 * We're holding onto IO to allow userland time to react. After the
2289 * timeout either the pool will have been resized (and thus back in
2290 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2292 static void do_no_space_timeout(struct work_struct *ws)
2294 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2297 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2298 pool->pf.error_if_no_space = true;
2299 notify_of_pool_mode_change_to_oods(pool);
2300 error_retry_list(pool);
2304 /*----------------------------------------------------------------*/
2307 struct work_struct worker;
2308 struct completion complete;
2311 static struct pool_work *to_pool_work(struct work_struct *ws)
2313 return container_of(ws, struct pool_work, worker);
2316 static void pool_work_complete(struct pool_work *pw)
2318 complete(&pw->complete);
2321 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2322 void (*fn)(struct work_struct *))
2324 INIT_WORK_ONSTACK(&pw->worker, fn);
2325 init_completion(&pw->complete);
2326 queue_work(pool->wq, &pw->worker);
2327 wait_for_completion(&pw->complete);
2330 /*----------------------------------------------------------------*/
2332 struct noflush_work {
2333 struct pool_work pw;
2337 static struct noflush_work *to_noflush(struct work_struct *ws)
2339 return container_of(to_pool_work(ws), struct noflush_work, pw);
2342 static void do_noflush_start(struct work_struct *ws)
2344 struct noflush_work *w = to_noflush(ws);
2345 w->tc->requeue_mode = true;
2347 pool_work_complete(&w->pw);
2350 static void do_noflush_stop(struct work_struct *ws)
2352 struct noflush_work *w = to_noflush(ws);
2353 w->tc->requeue_mode = false;
2354 pool_work_complete(&w->pw);
2357 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2359 struct noflush_work w;
2362 pool_work_wait(&w.pw, tc->pool, fn);
2365 /*----------------------------------------------------------------*/
2367 static enum pool_mode get_pool_mode(struct pool *pool)
2369 return pool->pf.mode;
2372 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2374 dm_table_event(pool->ti->table);
2375 DMINFO("%s: switching pool to %s mode",
2376 dm_device_name(pool->pool_md), new_mode);
2379 static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2381 if (!pool->pf.error_if_no_space)
2382 notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2384 notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2387 static bool passdown_enabled(struct pool_c *pt)
2389 return pt->adjusted_pf.discard_passdown;
2392 static void set_discard_callbacks(struct pool *pool)
2394 struct pool_c *pt = pool->ti->private;
2396 if (passdown_enabled(pt)) {
2397 pool->process_discard_cell = process_discard_cell_passdown;
2398 pool->process_prepared_discard = process_prepared_discard_passdown;
2400 pool->process_discard_cell = process_discard_cell_no_passdown;
2401 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2405 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2407 struct pool_c *pt = pool->ti->private;
2408 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2409 enum pool_mode old_mode = get_pool_mode(pool);
2410 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2413 * Never allow the pool to transition to PM_WRITE mode if user
2414 * intervention is required to verify metadata and data consistency.
2416 if (new_mode == PM_WRITE && needs_check) {
2417 DMERR("%s: unable to switch pool to write mode until repaired.",
2418 dm_device_name(pool->pool_md));
2419 if (old_mode != new_mode)
2420 new_mode = old_mode;
2422 new_mode = PM_READ_ONLY;
2425 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2426 * not going to recover without a thin_repair. So we never let the
2427 * pool move out of the old mode.
2429 if (old_mode == PM_FAIL)
2430 new_mode = old_mode;
2434 if (old_mode != new_mode)
2435 notify_of_pool_mode_change(pool, "failure");
2436 dm_pool_metadata_read_only(pool->pmd);
2437 pool->process_bio = process_bio_fail;
2438 pool->process_discard = process_bio_fail;
2439 pool->process_cell = process_cell_fail;
2440 pool->process_discard_cell = process_cell_fail;
2441 pool->process_prepared_mapping = process_prepared_mapping_fail;
2442 pool->process_prepared_discard = process_prepared_discard_fail;
2444 error_retry_list(pool);
2448 if (old_mode != new_mode)
2449 notify_of_pool_mode_change(pool, "read-only");
2450 dm_pool_metadata_read_only(pool->pmd);
2451 pool->process_bio = process_bio_read_only;
2452 pool->process_discard = process_bio_success;
2453 pool->process_cell = process_cell_read_only;
2454 pool->process_discard_cell = process_cell_success;
2455 pool->process_prepared_mapping = process_prepared_mapping_fail;
2456 pool->process_prepared_discard = process_prepared_discard_success;
2458 error_retry_list(pool);
2461 case PM_OUT_OF_DATA_SPACE:
2463 * Ideally we'd never hit this state; the low water mark
2464 * would trigger userland to extend the pool before we
2465 * completely run out of data space. However, many small
2466 * IOs to unprovisioned space can consume data space at an
2467 * alarming rate. Adjust your low water mark if you're
2468 * frequently seeing this mode.
2470 if (old_mode != new_mode)
2471 notify_of_pool_mode_change_to_oods(pool);
2472 pool->process_bio = process_bio_read_only;
2473 pool->process_discard = process_discard_bio;
2474 pool->process_cell = process_cell_read_only;
2475 pool->process_prepared_mapping = process_prepared_mapping;
2476 set_discard_callbacks(pool);
2478 if (!pool->pf.error_if_no_space && no_space_timeout)
2479 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2483 if (old_mode != new_mode)
2484 notify_of_pool_mode_change(pool, "write");
2485 dm_pool_metadata_read_write(pool->pmd);
2486 pool->process_bio = process_bio;
2487 pool->process_discard = process_discard_bio;
2488 pool->process_cell = process_cell;
2489 pool->process_prepared_mapping = process_prepared_mapping;
2490 set_discard_callbacks(pool);
2494 pool->pf.mode = new_mode;
2496 * The pool mode may have changed, sync it so bind_control_target()
2497 * doesn't cause an unexpected mode transition on resume.
2499 pt->adjusted_pf.mode = new_mode;
2502 static void abort_transaction(struct pool *pool)
2504 const char *dev_name = dm_device_name(pool->pool_md);
2506 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2507 if (dm_pool_abort_metadata(pool->pmd)) {
2508 DMERR("%s: failed to abort metadata transaction", dev_name);
2509 set_pool_mode(pool, PM_FAIL);
2512 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2513 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2514 set_pool_mode(pool, PM_FAIL);
2518 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2520 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2521 dm_device_name(pool->pool_md), op, r);
2523 abort_transaction(pool);
2524 set_pool_mode(pool, PM_READ_ONLY);
2527 /*----------------------------------------------------------------*/
2530 * Mapping functions.
2534 * Called only while mapping a thin bio to hand it over to the workqueue.
2536 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2538 unsigned long flags;
2539 struct pool *pool = tc->pool;
2541 spin_lock_irqsave(&tc->lock, flags);
2542 bio_list_add(&tc->deferred_bio_list, bio);
2543 spin_unlock_irqrestore(&tc->lock, flags);
2548 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2550 struct pool *pool = tc->pool;
2552 throttle_lock(&pool->throttle);
2553 thin_defer_bio(tc, bio);
2554 throttle_unlock(&pool->throttle);
2557 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2559 unsigned long flags;
2560 struct pool *pool = tc->pool;
2562 throttle_lock(&pool->throttle);
2563 spin_lock_irqsave(&tc->lock, flags);
2564 list_add_tail(&cell->user_list, &tc->deferred_cells);
2565 spin_unlock_irqrestore(&tc->lock, flags);
2566 throttle_unlock(&pool->throttle);
2571 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2573 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2576 h->shared_read_entry = NULL;
2577 h->all_io_entry = NULL;
2578 h->overwrite_mapping = NULL;
2583 * Non-blocking function called from the thin target's map function.
2585 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2588 struct thin_c *tc = ti->private;
2589 dm_block_t block = get_bio_block(tc, bio);
2590 struct dm_thin_device *td = tc->td;
2591 struct dm_thin_lookup_result result;
2592 struct dm_bio_prison_cell *virt_cell, *data_cell;
2593 struct dm_cell_key key;
2595 thin_hook_bio(tc, bio);
2597 if (tc->requeue_mode) {
2598 bio_endio(bio, DM_ENDIO_REQUEUE);
2599 return DM_MAPIO_SUBMITTED;
2602 if (get_pool_mode(tc->pool) == PM_FAIL) {
2604 return DM_MAPIO_SUBMITTED;
2607 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
2608 thin_defer_bio_with_throttle(tc, bio);
2609 return DM_MAPIO_SUBMITTED;
2613 * We must hold the virtual cell before doing the lookup, otherwise
2614 * there's a race with discard.
2616 build_virtual_key(tc->td, block, &key);
2617 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2618 return DM_MAPIO_SUBMITTED;
2620 r = dm_thin_find_block(td, block, 0, &result);
2623 * Note that we defer readahead too.
2627 if (unlikely(result.shared)) {
2629 * We have a race condition here between the
2630 * result.shared value returned by the lookup and
2631 * snapshot creation, which may cause new
2634 * To avoid this always quiesce the origin before
2635 * taking the snap. You want to do this anyway to
2636 * ensure a consistent application view
2639 * More distant ancestors are irrelevant. The
2640 * shared flag will be set in their case.
2642 thin_defer_cell(tc, virt_cell);
2643 return DM_MAPIO_SUBMITTED;
2646 build_data_key(tc->td, result.block, &key);
2647 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2648 cell_defer_no_holder(tc, virt_cell);
2649 return DM_MAPIO_SUBMITTED;
2652 inc_all_io_entry(tc->pool, bio);
2653 cell_defer_no_holder(tc, data_cell);
2654 cell_defer_no_holder(tc, virt_cell);
2656 remap(tc, bio, result.block);
2657 return DM_MAPIO_REMAPPED;
2661 thin_defer_cell(tc, virt_cell);
2662 return DM_MAPIO_SUBMITTED;
2666 * Must always call bio_io_error on failure.
2667 * dm_thin_find_block can fail with -EINVAL if the
2668 * pool is switched to fail-io mode.
2671 cell_defer_no_holder(tc, virt_cell);
2672 return DM_MAPIO_SUBMITTED;
2676 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2678 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2679 struct request_queue *q;
2681 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2684 q = bdev_get_queue(pt->data_dev->bdev);
2685 return bdi_congested(&q->backing_dev_info, bdi_bits);
2688 static void requeue_bios(struct pool *pool)
2690 unsigned long flags;
2694 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2695 spin_lock_irqsave(&tc->lock, flags);
2696 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2697 bio_list_init(&tc->retry_on_resume_list);
2698 spin_unlock_irqrestore(&tc->lock, flags);
2703 /*----------------------------------------------------------------
2704 * Binding of control targets to a pool object
2705 *--------------------------------------------------------------*/
2706 static bool data_dev_supports_discard(struct pool_c *pt)
2708 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2710 return q && blk_queue_discard(q);
2713 static bool is_factor(sector_t block_size, uint32_t n)
2715 return !sector_div(block_size, n);
2719 * If discard_passdown was enabled verify that the data device
2720 * supports discards. Disable discard_passdown if not.
2722 static void disable_passdown_if_not_supported(struct pool_c *pt)
2724 struct pool *pool = pt->pool;
2725 struct block_device *data_bdev = pt->data_dev->bdev;
2726 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2727 const char *reason = NULL;
2728 char buf[BDEVNAME_SIZE];
2730 if (!pt->adjusted_pf.discard_passdown)
2733 if (!data_dev_supports_discard(pt))
2734 reason = "discard unsupported";
2736 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2737 reason = "max discard sectors smaller than a block";
2740 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2741 pt->adjusted_pf.discard_passdown = false;
2745 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2747 struct pool_c *pt = ti->private;
2750 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2752 enum pool_mode old_mode = get_pool_mode(pool);
2753 enum pool_mode new_mode = pt->adjusted_pf.mode;
2756 * Don't change the pool's mode until set_pool_mode() below.
2757 * Otherwise the pool's process_* function pointers may
2758 * not match the desired pool mode.
2760 pt->adjusted_pf.mode = old_mode;
2763 pool->pf = pt->adjusted_pf;
2764 pool->low_water_blocks = pt->low_water_blocks;
2766 set_pool_mode(pool, new_mode);
2771 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2777 /*----------------------------------------------------------------
2779 *--------------------------------------------------------------*/
2780 /* Initialize pool features. */
2781 static void pool_features_init(struct pool_features *pf)
2783 pf->mode = PM_WRITE;
2784 pf->zero_new_blocks = true;
2785 pf->discard_enabled = true;
2786 pf->discard_passdown = true;
2787 pf->error_if_no_space = false;
2790 static void __pool_destroy(struct pool *pool)
2792 __pool_table_remove(pool);
2794 vfree(pool->cell_sort_array);
2795 if (dm_pool_metadata_close(pool->pmd) < 0)
2796 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2798 dm_bio_prison_destroy(pool->prison);
2799 dm_kcopyd_client_destroy(pool->copier);
2802 destroy_workqueue(pool->wq);
2804 if (pool->next_mapping)
2805 mempool_free(pool->next_mapping, pool->mapping_pool);
2806 mempool_destroy(pool->mapping_pool);
2807 dm_deferred_set_destroy(pool->shared_read_ds);
2808 dm_deferred_set_destroy(pool->all_io_ds);
2812 static struct kmem_cache *_new_mapping_cache;
2814 static struct pool *pool_create(struct mapped_device *pool_md,
2815 struct block_device *metadata_dev,
2816 unsigned long block_size,
2817 int read_only, char **error)
2822 struct dm_pool_metadata *pmd;
2823 bool format_device = read_only ? false : true;
2825 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2827 *error = "Error creating metadata object";
2828 return (struct pool *)pmd;
2831 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2833 *error = "Error allocating memory for pool";
2834 err_p = ERR_PTR(-ENOMEM);
2839 pool->sectors_per_block = block_size;
2840 if (block_size & (block_size - 1))
2841 pool->sectors_per_block_shift = -1;
2843 pool->sectors_per_block_shift = __ffs(block_size);
2844 pool->low_water_blocks = 0;
2845 pool_features_init(&pool->pf);
2846 pool->prison = dm_bio_prison_create();
2847 if (!pool->prison) {
2848 *error = "Error creating pool's bio prison";
2849 err_p = ERR_PTR(-ENOMEM);
2853 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2854 if (IS_ERR(pool->copier)) {
2855 r = PTR_ERR(pool->copier);
2856 *error = "Error creating pool's kcopyd client";
2858 goto bad_kcopyd_client;
2862 * Create singlethreaded workqueue that will service all devices
2863 * that use this metadata.
2865 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2867 *error = "Error creating pool's workqueue";
2868 err_p = ERR_PTR(-ENOMEM);
2872 throttle_init(&pool->throttle);
2873 INIT_WORK(&pool->worker, do_worker);
2874 INIT_DELAYED_WORK(&pool->waker, do_waker);
2875 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2876 spin_lock_init(&pool->lock);
2877 bio_list_init(&pool->deferred_flush_bios);
2878 INIT_LIST_HEAD(&pool->prepared_mappings);
2879 INIT_LIST_HEAD(&pool->prepared_discards);
2880 INIT_LIST_HEAD(&pool->active_thins);
2881 pool->low_water_triggered = false;
2882 pool->suspended = true;
2884 pool->shared_read_ds = dm_deferred_set_create();
2885 if (!pool->shared_read_ds) {
2886 *error = "Error creating pool's shared read deferred set";
2887 err_p = ERR_PTR(-ENOMEM);
2888 goto bad_shared_read_ds;
2891 pool->all_io_ds = dm_deferred_set_create();
2892 if (!pool->all_io_ds) {
2893 *error = "Error creating pool's all io deferred set";
2894 err_p = ERR_PTR(-ENOMEM);
2898 pool->next_mapping = NULL;
2899 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2900 _new_mapping_cache);
2901 if (!pool->mapping_pool) {
2902 *error = "Error creating pool's mapping mempool";
2903 err_p = ERR_PTR(-ENOMEM);
2904 goto bad_mapping_pool;
2907 pool->cell_sort_array = vmalloc(sizeof(*pool->cell_sort_array) * CELL_SORT_ARRAY_SIZE);
2908 if (!pool->cell_sort_array) {
2909 *error = "Error allocating cell sort array";
2910 err_p = ERR_PTR(-ENOMEM);
2911 goto bad_sort_array;
2914 pool->ref_count = 1;
2915 pool->last_commit_jiffies = jiffies;
2916 pool->pool_md = pool_md;
2917 pool->md_dev = metadata_dev;
2918 __pool_table_insert(pool);
2923 mempool_destroy(pool->mapping_pool);
2925 dm_deferred_set_destroy(pool->all_io_ds);
2927 dm_deferred_set_destroy(pool->shared_read_ds);
2929 destroy_workqueue(pool->wq);
2931 dm_kcopyd_client_destroy(pool->copier);
2933 dm_bio_prison_destroy(pool->prison);
2937 if (dm_pool_metadata_close(pmd))
2938 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2943 static void __pool_inc(struct pool *pool)
2945 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2949 static void __pool_dec(struct pool *pool)
2951 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2952 BUG_ON(!pool->ref_count);
2953 if (!--pool->ref_count)
2954 __pool_destroy(pool);
2957 static struct pool *__pool_find(struct mapped_device *pool_md,
2958 struct block_device *metadata_dev,
2959 unsigned long block_size, int read_only,
2960 char **error, int *created)
2962 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2965 if (pool->pool_md != pool_md) {
2966 *error = "metadata device already in use by a pool";
2967 return ERR_PTR(-EBUSY);
2972 pool = __pool_table_lookup(pool_md);
2974 if (pool->md_dev != metadata_dev) {
2975 *error = "different pool cannot replace a pool";
2976 return ERR_PTR(-EINVAL);
2981 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2989 /*----------------------------------------------------------------
2990 * Pool target methods
2991 *--------------------------------------------------------------*/
2992 static void pool_dtr(struct dm_target *ti)
2994 struct pool_c *pt = ti->private;
2996 mutex_lock(&dm_thin_pool_table.mutex);
2998 unbind_control_target(pt->pool, ti);
2999 __pool_dec(pt->pool);
3000 dm_put_device(ti, pt->metadata_dev);
3001 dm_put_device(ti, pt->data_dev);
3004 mutex_unlock(&dm_thin_pool_table.mutex);
3007 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3008 struct dm_target *ti)
3012 const char *arg_name;
3014 static struct dm_arg _args[] = {
3015 {0, 4, "Invalid number of pool feature arguments"},
3019 * No feature arguments supplied.
3024 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3028 while (argc && !r) {
3029 arg_name = dm_shift_arg(as);
3032 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3033 pf->zero_new_blocks = false;
3035 else if (!strcasecmp(arg_name, "ignore_discard"))
3036 pf->discard_enabled = false;
3038 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3039 pf->discard_passdown = false;
3041 else if (!strcasecmp(arg_name, "read_only"))
3042 pf->mode = PM_READ_ONLY;
3044 else if (!strcasecmp(arg_name, "error_if_no_space"))
3045 pf->error_if_no_space = true;
3048 ti->error = "Unrecognised pool feature requested";
3057 static void metadata_low_callback(void *context)
3059 struct pool *pool = context;
3061 DMWARN("%s: reached low water mark for metadata device: sending event.",
3062 dm_device_name(pool->pool_md));
3064 dm_table_event(pool->ti->table);
3067 static sector_t get_dev_size(struct block_device *bdev)
3069 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3072 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3074 sector_t metadata_dev_size = get_dev_size(bdev);
3075 char buffer[BDEVNAME_SIZE];
3077 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3078 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3079 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3082 static sector_t get_metadata_dev_size(struct block_device *bdev)
3084 sector_t metadata_dev_size = get_dev_size(bdev);
3086 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3087 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3089 return metadata_dev_size;
3092 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3094 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3096 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3098 return metadata_dev_size;
3102 * When a metadata threshold is crossed a dm event is triggered, and
3103 * userland should respond by growing the metadata device. We could let
3104 * userland set the threshold, like we do with the data threshold, but I'm
3105 * not sure they know enough to do this well.
3107 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3110 * 4M is ample for all ops with the possible exception of thin
3111 * device deletion which is harmless if it fails (just retry the
3112 * delete after you've grown the device).
3114 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3115 return min((dm_block_t)1024ULL /* 4M */, quarter);
3119 * thin-pool <metadata dev> <data dev>
3120 * <data block size (sectors)>
3121 * <low water mark (blocks)>
3122 * [<#feature args> [<arg>]*]
3124 * Optional feature arguments are:
3125 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3126 * ignore_discard: disable discard
3127 * no_discard_passdown: don't pass discards down to the data device
3128 * read_only: Don't allow any changes to be made to the pool metadata.
3129 * error_if_no_space: error IOs, instead of queueing, if no space.
3131 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3133 int r, pool_created = 0;
3136 struct pool_features pf;
3137 struct dm_arg_set as;
3138 struct dm_dev *data_dev;
3139 unsigned long block_size;
3140 dm_block_t low_water_blocks;
3141 struct dm_dev *metadata_dev;
3142 fmode_t metadata_mode;
3145 * FIXME Remove validation from scope of lock.
3147 mutex_lock(&dm_thin_pool_table.mutex);
3150 ti->error = "Invalid argument count";
3159 * Set default pool features.
3161 pool_features_init(&pf);
3163 dm_consume_args(&as, 4);
3164 r = parse_pool_features(&as, &pf, ti);
3168 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3169 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3171 ti->error = "Error opening metadata block device";
3174 warn_if_metadata_device_too_big(metadata_dev->bdev);
3176 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3178 ti->error = "Error getting data device";
3182 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3183 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3184 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3185 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3186 ti->error = "Invalid block size";
3191 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3192 ti->error = "Invalid low water mark";
3197 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3203 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3204 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3211 * 'pool_created' reflects whether this is the first table load.
3212 * Top level discard support is not allowed to be changed after
3213 * initial load. This would require a pool reload to trigger thin
3216 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3217 ti->error = "Discard support cannot be disabled once enabled";
3219 goto out_flags_changed;
3224 pt->metadata_dev = metadata_dev;
3225 pt->data_dev = data_dev;
3226 pt->low_water_blocks = low_water_blocks;
3227 pt->adjusted_pf = pt->requested_pf = pf;
3228 ti->num_flush_bios = 1;
3231 * Only need to enable discards if the pool should pass
3232 * them down to the data device. The thin device's discard
3233 * processing will cause mappings to be removed from the btree.
3235 ti->discard_zeroes_data_unsupported = true;
3236 if (pf.discard_enabled && pf.discard_passdown) {
3237 ti->num_discard_bios = 1;
3240 * Setting 'discards_supported' circumvents the normal
3241 * stacking of discard limits (this keeps the pool and
3242 * thin devices' discard limits consistent).
3244 ti->discards_supported = true;
3248 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3249 calc_metadata_threshold(pt),
3250 metadata_low_callback,
3255 pt->callbacks.congested_fn = pool_is_congested;
3256 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3258 mutex_unlock(&dm_thin_pool_table.mutex);
3267 dm_put_device(ti, data_dev);
3269 dm_put_device(ti, metadata_dev);
3271 mutex_unlock(&dm_thin_pool_table.mutex);
3276 static int pool_map(struct dm_target *ti, struct bio *bio)
3279 struct pool_c *pt = ti->private;
3280 struct pool *pool = pt->pool;
3281 unsigned long flags;
3284 * As this is a singleton target, ti->begin is always zero.
3286 spin_lock_irqsave(&pool->lock, flags);
3287 bio->bi_bdev = pt->data_dev->bdev;
3288 r = DM_MAPIO_REMAPPED;
3289 spin_unlock_irqrestore(&pool->lock, flags);
3294 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3297 struct pool_c *pt = ti->private;
3298 struct pool *pool = pt->pool;
3299 sector_t data_size = ti->len;
3300 dm_block_t sb_data_size;
3302 *need_commit = false;
3304 (void) sector_div(data_size, pool->sectors_per_block);
3306 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3308 DMERR("%s: failed to retrieve data device size",
3309 dm_device_name(pool->pool_md));
3313 if (data_size < sb_data_size) {
3314 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3315 dm_device_name(pool->pool_md),
3316 (unsigned long long)data_size, sb_data_size);
3319 } else if (data_size > sb_data_size) {
3320 if (dm_pool_metadata_needs_check(pool->pmd)) {
3321 DMERR("%s: unable to grow the data device until repaired.",
3322 dm_device_name(pool->pool_md));
3327 DMINFO("%s: growing the data device from %llu to %llu blocks",
3328 dm_device_name(pool->pool_md),
3329 sb_data_size, (unsigned long long)data_size);
3330 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3332 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3336 *need_commit = true;
3342 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3345 struct pool_c *pt = ti->private;
3346 struct pool *pool = pt->pool;
3347 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3349 *need_commit = false;
3351 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3353 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3355 DMERR("%s: failed to retrieve metadata device size",
3356 dm_device_name(pool->pool_md));
3360 if (metadata_dev_size < sb_metadata_dev_size) {
3361 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3362 dm_device_name(pool->pool_md),
3363 metadata_dev_size, sb_metadata_dev_size);
3366 } else if (metadata_dev_size > sb_metadata_dev_size) {
3367 if (dm_pool_metadata_needs_check(pool->pmd)) {
3368 DMERR("%s: unable to grow the metadata device until repaired.",
3369 dm_device_name(pool->pool_md));
3373 warn_if_metadata_device_too_big(pool->md_dev);
3374 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3375 dm_device_name(pool->pool_md),
3376 sb_metadata_dev_size, metadata_dev_size);
3377 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3379 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3383 *need_commit = true;
3390 * Retrieves the number of blocks of the data device from
3391 * the superblock and compares it to the actual device size,
3392 * thus resizing the data device in case it has grown.
3394 * This both copes with opening preallocated data devices in the ctr
3395 * being followed by a resume
3397 * calling the resume method individually after userspace has
3398 * grown the data device in reaction to a table event.
3400 static int pool_preresume(struct dm_target *ti)
3403 bool need_commit1, need_commit2;
3404 struct pool_c *pt = ti->private;
3405 struct pool *pool = pt->pool;
3408 * Take control of the pool object.
3410 r = bind_control_target(pool, ti);
3414 r = maybe_resize_data_dev(ti, &need_commit1);
3418 r = maybe_resize_metadata_dev(ti, &need_commit2);
3422 if (need_commit1 || need_commit2)
3423 (void) commit(pool);
3428 static void pool_suspend_active_thins(struct pool *pool)
3432 /* Suspend all active thin devices */
3433 tc = get_first_thin(pool);
3435 dm_internal_suspend_noflush(tc->thin_md);
3436 tc = get_next_thin(pool, tc);
3440 static void pool_resume_active_thins(struct pool *pool)
3444 /* Resume all active thin devices */
3445 tc = get_first_thin(pool);
3447 dm_internal_resume(tc->thin_md);
3448 tc = get_next_thin(pool, tc);
3452 static void pool_resume(struct dm_target *ti)
3454 struct pool_c *pt = ti->private;
3455 struct pool *pool = pt->pool;
3456 unsigned long flags;
3459 * Must requeue active_thins' bios and then resume
3460 * active_thins _before_ clearing 'suspend' flag.
3463 pool_resume_active_thins(pool);
3465 spin_lock_irqsave(&pool->lock, flags);
3466 pool->low_water_triggered = false;
3467 pool->suspended = false;
3468 spin_unlock_irqrestore(&pool->lock, flags);
3470 do_waker(&pool->waker.work);
3473 static void pool_presuspend(struct dm_target *ti)
3475 struct pool_c *pt = ti->private;
3476 struct pool *pool = pt->pool;
3477 unsigned long flags;
3479 spin_lock_irqsave(&pool->lock, flags);
3480 pool->suspended = true;
3481 spin_unlock_irqrestore(&pool->lock, flags);
3483 pool_suspend_active_thins(pool);
3486 static void pool_presuspend_undo(struct dm_target *ti)
3488 struct pool_c *pt = ti->private;
3489 struct pool *pool = pt->pool;
3490 unsigned long flags;
3492 pool_resume_active_thins(pool);
3494 spin_lock_irqsave(&pool->lock, flags);
3495 pool->suspended = false;
3496 spin_unlock_irqrestore(&pool->lock, flags);
3499 static void pool_postsuspend(struct dm_target *ti)
3501 struct pool_c *pt = ti->private;
3502 struct pool *pool = pt->pool;
3504 cancel_delayed_work(&pool->waker);
3505 cancel_delayed_work(&pool->no_space_timeout);
3506 flush_workqueue(pool->wq);
3507 (void) commit(pool);
3510 static int check_arg_count(unsigned argc, unsigned args_required)
3512 if (argc != args_required) {
3513 DMWARN("Message received with %u arguments instead of %u.",
3514 argc, args_required);
3521 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3523 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3524 *dev_id <= MAX_DEV_ID)
3528 DMWARN("Message received with invalid device id: %s", arg);
3533 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3538 r = check_arg_count(argc, 2);
3542 r = read_dev_id(argv[1], &dev_id, 1);
3546 r = dm_pool_create_thin(pool->pmd, dev_id);
3548 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3556 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3559 dm_thin_id origin_dev_id;
3562 r = check_arg_count(argc, 3);
3566 r = read_dev_id(argv[1], &dev_id, 1);
3570 r = read_dev_id(argv[2], &origin_dev_id, 1);
3574 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3576 DMWARN("Creation of new snapshot %s of device %s failed.",
3584 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3589 r = check_arg_count(argc, 2);
3593 r = read_dev_id(argv[1], &dev_id, 1);
3597 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3599 DMWARN("Deletion of thin device %s failed.", argv[1]);
3604 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3606 dm_thin_id old_id, new_id;
3609 r = check_arg_count(argc, 3);
3613 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3614 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3618 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3619 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3623 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3625 DMWARN("Failed to change transaction id from %s to %s.",
3633 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3637 r = check_arg_count(argc, 1);
3641 (void) commit(pool);
3643 r = dm_pool_reserve_metadata_snap(pool->pmd);
3645 DMWARN("reserve_metadata_snap message failed.");
3650 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3654 r = check_arg_count(argc, 1);
3658 r = dm_pool_release_metadata_snap(pool->pmd);
3660 DMWARN("release_metadata_snap message failed.");
3666 * Messages supported:
3667 * create_thin <dev_id>
3668 * create_snap <dev_id> <origin_id>
3670 * set_transaction_id <current_trans_id> <new_trans_id>
3671 * reserve_metadata_snap
3672 * release_metadata_snap
3674 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3677 struct pool_c *pt = ti->private;
3678 struct pool *pool = pt->pool;
3680 if (get_pool_mode(pool) >= PM_READ_ONLY) {
3681 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3682 dm_device_name(pool->pool_md));
3686 if (!strcasecmp(argv[0], "create_thin"))
3687 r = process_create_thin_mesg(argc, argv, pool);
3689 else if (!strcasecmp(argv[0], "create_snap"))
3690 r = process_create_snap_mesg(argc, argv, pool);
3692 else if (!strcasecmp(argv[0], "delete"))
3693 r = process_delete_mesg(argc, argv, pool);
3695 else if (!strcasecmp(argv[0], "set_transaction_id"))
3696 r = process_set_transaction_id_mesg(argc, argv, pool);
3698 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3699 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3701 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3702 r = process_release_metadata_snap_mesg(argc, argv, pool);
3705 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3708 (void) commit(pool);
3713 static void emit_flags(struct pool_features *pf, char *result,
3714 unsigned sz, unsigned maxlen)
3716 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3717 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3718 pf->error_if_no_space;
3719 DMEMIT("%u ", count);
3721 if (!pf->zero_new_blocks)
3722 DMEMIT("skip_block_zeroing ");
3724 if (!pf->discard_enabled)
3725 DMEMIT("ignore_discard ");
3727 if (!pf->discard_passdown)
3728 DMEMIT("no_discard_passdown ");
3730 if (pf->mode == PM_READ_ONLY)
3731 DMEMIT("read_only ");
3733 if (pf->error_if_no_space)
3734 DMEMIT("error_if_no_space ");
3739 * <transaction id> <used metadata sectors>/<total metadata sectors>
3740 * <used data sectors>/<total data sectors> <held metadata root>
3741 * <pool mode> <discard config> <no space config> <needs_check>
3743 static void pool_status(struct dm_target *ti, status_type_t type,
3744 unsigned status_flags, char *result, unsigned maxlen)
3748 uint64_t transaction_id;
3749 dm_block_t nr_free_blocks_data;
3750 dm_block_t nr_free_blocks_metadata;
3751 dm_block_t nr_blocks_data;
3752 dm_block_t nr_blocks_metadata;
3753 dm_block_t held_root;
3754 char buf[BDEVNAME_SIZE];
3755 char buf2[BDEVNAME_SIZE];
3756 struct pool_c *pt = ti->private;
3757 struct pool *pool = pt->pool;
3760 case STATUSTYPE_INFO:
3761 if (get_pool_mode(pool) == PM_FAIL) {
3766 /* Commit to ensure statistics aren't out-of-date */
3767 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3768 (void) commit(pool);
3770 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3772 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3773 dm_device_name(pool->pool_md), r);
3777 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3779 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3780 dm_device_name(pool->pool_md), r);
3784 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3786 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3787 dm_device_name(pool->pool_md), r);
3791 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3793 DMERR("%s: dm_pool_get_free_block_count returned %d",
3794 dm_device_name(pool->pool_md), r);
3798 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3800 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3801 dm_device_name(pool->pool_md), r);
3805 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3807 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3808 dm_device_name(pool->pool_md), r);
3812 DMEMIT("%llu %llu/%llu %llu/%llu ",
3813 (unsigned long long)transaction_id,
3814 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3815 (unsigned long long)nr_blocks_metadata,
3816 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3817 (unsigned long long)nr_blocks_data);
3820 DMEMIT("%llu ", held_root);
3824 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3825 DMEMIT("out_of_data_space ");
3826 else if (pool->pf.mode == PM_READ_ONLY)
3831 if (!pool->pf.discard_enabled)
3832 DMEMIT("ignore_discard ");
3833 else if (pool->pf.discard_passdown)
3834 DMEMIT("discard_passdown ");
3836 DMEMIT("no_discard_passdown ");
3838 if (pool->pf.error_if_no_space)
3839 DMEMIT("error_if_no_space ");
3841 DMEMIT("queue_if_no_space ");
3843 if (dm_pool_metadata_needs_check(pool->pmd))
3844 DMEMIT("needs_check ");
3850 case STATUSTYPE_TABLE:
3851 DMEMIT("%s %s %lu %llu ",
3852 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3853 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3854 (unsigned long)pool->sectors_per_block,
3855 (unsigned long long)pt->low_water_blocks);
3856 emit_flags(&pt->requested_pf, result, sz, maxlen);
3865 static int pool_iterate_devices(struct dm_target *ti,
3866 iterate_devices_callout_fn fn, void *data)
3868 struct pool_c *pt = ti->private;
3870 return fn(ti, pt->data_dev, 0, ti->len, data);
3873 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
3874 struct bio_vec *biovec, int max_size)
3876 struct pool_c *pt = ti->private;
3877 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
3879 if (!q->merge_bvec_fn)
3882 bvm->bi_bdev = pt->data_dev->bdev;
3884 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3887 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3889 struct pool_c *pt = ti->private;
3890 struct pool *pool = pt->pool;
3891 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3894 * If max_sectors is smaller than pool->sectors_per_block adjust it
3895 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3896 * This is especially beneficial when the pool's data device is a RAID
3897 * device that has a full stripe width that matches pool->sectors_per_block
3898 * -- because even though partial RAID stripe-sized IOs will be issued to a
3899 * single RAID stripe; when aggregated they will end on a full RAID stripe
3900 * boundary.. which avoids additional partial RAID stripe writes cascading
3902 if (limits->max_sectors < pool->sectors_per_block) {
3903 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3904 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3905 limits->max_sectors--;
3906 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3911 * If the system-determined stacked limits are compatible with the
3912 * pool's blocksize (io_opt is a factor) do not override them.
3914 if (io_opt_sectors < pool->sectors_per_block ||
3915 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3916 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3917 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3919 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3920 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3924 * pt->adjusted_pf is a staging area for the actual features to use.
3925 * They get transferred to the live pool in bind_control_target()
3926 * called from pool_preresume().
3928 if (!pt->adjusted_pf.discard_enabled) {
3930 * Must explicitly disallow stacking discard limits otherwise the
3931 * block layer will stack them if pool's data device has support.
3932 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3933 * user to see that, so make sure to set all discard limits to 0.
3935 limits->discard_granularity = 0;
3939 disable_passdown_if_not_supported(pt);
3942 * The pool uses the same discard limits as the underlying data
3943 * device. DM core has already set this up.
3947 static struct target_type pool_target = {
3948 .name = "thin-pool",
3949 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3950 DM_TARGET_IMMUTABLE,
3951 .version = {1, 16, 0},
3952 .module = THIS_MODULE,
3956 .presuspend = pool_presuspend,
3957 .presuspend_undo = pool_presuspend_undo,
3958 .postsuspend = pool_postsuspend,
3959 .preresume = pool_preresume,
3960 .resume = pool_resume,
3961 .message = pool_message,
3962 .status = pool_status,
3963 .merge = pool_merge,
3964 .iterate_devices = pool_iterate_devices,
3965 .io_hints = pool_io_hints,
3968 /*----------------------------------------------------------------
3969 * Thin target methods
3970 *--------------------------------------------------------------*/
3971 static void thin_get(struct thin_c *tc)
3973 atomic_inc(&tc->refcount);
3976 static void thin_put(struct thin_c *tc)
3978 if (atomic_dec_and_test(&tc->refcount))
3979 complete(&tc->can_destroy);
3982 static void thin_dtr(struct dm_target *ti)
3984 struct thin_c *tc = ti->private;
3985 unsigned long flags;
3987 spin_lock_irqsave(&tc->pool->lock, flags);
3988 list_del_rcu(&tc->list);
3989 spin_unlock_irqrestore(&tc->pool->lock, flags);
3993 wait_for_completion(&tc->can_destroy);
3995 mutex_lock(&dm_thin_pool_table.mutex);
3997 __pool_dec(tc->pool);
3998 dm_pool_close_thin_device(tc->td);
3999 dm_put_device(ti, tc->pool_dev);
4001 dm_put_device(ti, tc->origin_dev);
4004 mutex_unlock(&dm_thin_pool_table.mutex);
4008 * Thin target parameters:
4010 * <pool_dev> <dev_id> [origin_dev]
4012 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4013 * dev_id: the internal device identifier
4014 * origin_dev: a device external to the pool that should act as the origin
4016 * If the pool device has discards disabled, they get disabled for the thin
4019 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4023 struct dm_dev *pool_dev, *origin_dev;
4024 struct mapped_device *pool_md;
4025 unsigned long flags;
4027 mutex_lock(&dm_thin_pool_table.mutex);
4029 if (argc != 2 && argc != 3) {
4030 ti->error = "Invalid argument count";
4035 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4037 ti->error = "Out of memory";
4041 tc->thin_md = dm_table_get_md(ti->table);
4042 spin_lock_init(&tc->lock);
4043 INIT_LIST_HEAD(&tc->deferred_cells);
4044 bio_list_init(&tc->deferred_bio_list);
4045 bio_list_init(&tc->retry_on_resume_list);
4046 tc->sort_bio_list = RB_ROOT;
4049 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4051 ti->error = "Error opening origin device";
4052 goto bad_origin_dev;
4054 tc->origin_dev = origin_dev;
4057 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4059 ti->error = "Error opening pool device";
4062 tc->pool_dev = pool_dev;
4064 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4065 ti->error = "Invalid device id";
4070 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4072 ti->error = "Couldn't get pool mapped device";
4077 tc->pool = __pool_table_lookup(pool_md);
4079 ti->error = "Couldn't find pool object";
4081 goto bad_pool_lookup;
4083 __pool_inc(tc->pool);
4085 if (get_pool_mode(tc->pool) == PM_FAIL) {
4086 ti->error = "Couldn't open thin device, Pool is in fail mode";
4091 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4093 ti->error = "Couldn't open thin internal device";
4097 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4101 ti->num_flush_bios = 1;
4102 ti->flush_supported = true;
4103 ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
4105 /* In case the pool supports discards, pass them on. */
4106 ti->discard_zeroes_data_unsupported = true;
4107 if (tc->pool->pf.discard_enabled) {
4108 ti->discards_supported = true;
4109 ti->num_discard_bios = 1;
4110 ti->split_discard_bios = false;
4113 mutex_unlock(&dm_thin_pool_table.mutex);
4115 spin_lock_irqsave(&tc->pool->lock, flags);
4116 if (tc->pool->suspended) {
4117 spin_unlock_irqrestore(&tc->pool->lock, flags);
4118 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4119 ti->error = "Unable to activate thin device while pool is suspended";
4123 atomic_set(&tc->refcount, 1);
4124 init_completion(&tc->can_destroy);
4125 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4126 spin_unlock_irqrestore(&tc->pool->lock, flags);
4128 * This synchronize_rcu() call is needed here otherwise we risk a
4129 * wake_worker() call finding no bios to process (because the newly
4130 * added tc isn't yet visible). So this reduces latency since we
4131 * aren't then dependent on the periodic commit to wake_worker().
4140 dm_pool_close_thin_device(tc->td);
4142 __pool_dec(tc->pool);
4146 dm_put_device(ti, tc->pool_dev);
4149 dm_put_device(ti, tc->origin_dev);
4153 mutex_unlock(&dm_thin_pool_table.mutex);
4158 static int thin_map(struct dm_target *ti, struct bio *bio)
4160 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4162 return thin_bio_map(ti, bio);
4165 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
4167 unsigned long flags;
4168 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4169 struct list_head work;
4170 struct dm_thin_new_mapping *m, *tmp;
4171 struct pool *pool = h->tc->pool;
4173 if (h->shared_read_entry) {
4174 INIT_LIST_HEAD(&work);
4175 dm_deferred_entry_dec(h->shared_read_entry, &work);
4177 spin_lock_irqsave(&pool->lock, flags);
4178 list_for_each_entry_safe(m, tmp, &work, list) {
4180 __complete_mapping_preparation(m);
4182 spin_unlock_irqrestore(&pool->lock, flags);
4185 if (h->all_io_entry) {
4186 INIT_LIST_HEAD(&work);
4187 dm_deferred_entry_dec(h->all_io_entry, &work);
4188 if (!list_empty(&work)) {
4189 spin_lock_irqsave(&pool->lock, flags);
4190 list_for_each_entry_safe(m, tmp, &work, list)
4191 list_add_tail(&m->list, &pool->prepared_discards);
4192 spin_unlock_irqrestore(&pool->lock, flags);
4198 cell_defer_no_holder(h->tc, h->cell);
4203 static void thin_presuspend(struct dm_target *ti)
4205 struct thin_c *tc = ti->private;
4207 if (dm_noflush_suspending(ti))
4208 noflush_work(tc, do_noflush_start);
4211 static void thin_postsuspend(struct dm_target *ti)
4213 struct thin_c *tc = ti->private;
4216 * The dm_noflush_suspending flag has been cleared by now, so
4217 * unfortunately we must always run this.
4219 noflush_work(tc, do_noflush_stop);
4222 static int thin_preresume(struct dm_target *ti)
4224 struct thin_c *tc = ti->private;
4227 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4233 * <nr mapped sectors> <highest mapped sector>
4235 static void thin_status(struct dm_target *ti, status_type_t type,
4236 unsigned status_flags, char *result, unsigned maxlen)
4240 dm_block_t mapped, highest;
4241 char buf[BDEVNAME_SIZE];
4242 struct thin_c *tc = ti->private;
4244 if (get_pool_mode(tc->pool) == PM_FAIL) {
4253 case STATUSTYPE_INFO:
4254 r = dm_thin_get_mapped_count(tc->td, &mapped);
4256 DMERR("dm_thin_get_mapped_count returned %d", r);
4260 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4262 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4266 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4268 DMEMIT("%llu", ((highest + 1) *
4269 tc->pool->sectors_per_block) - 1);
4274 case STATUSTYPE_TABLE:
4276 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4277 (unsigned long) tc->dev_id);
4279 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4290 static int thin_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
4291 struct bio_vec *biovec, int max_size)
4293 struct thin_c *tc = ti->private;
4294 struct request_queue *q = bdev_get_queue(tc->pool_dev->bdev);
4296 if (!q->merge_bvec_fn)
4299 bvm->bi_bdev = tc->pool_dev->bdev;
4300 bvm->bi_sector = dm_target_offset(ti, bvm->bi_sector);
4302 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
4305 static int thin_iterate_devices(struct dm_target *ti,
4306 iterate_devices_callout_fn fn, void *data)
4309 struct thin_c *tc = ti->private;
4310 struct pool *pool = tc->pool;
4313 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4314 * we follow a more convoluted path through to the pool's target.
4317 return 0; /* nothing is bound */
4319 blocks = pool->ti->len;
4320 (void) sector_div(blocks, pool->sectors_per_block);
4322 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4327 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4329 struct thin_c *tc = ti->private;
4330 struct pool *pool = tc->pool;
4332 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4333 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4336 static struct target_type thin_target = {
4338 .version = {1, 16, 0},
4339 .module = THIS_MODULE,
4343 .end_io = thin_endio,
4344 .preresume = thin_preresume,
4345 .presuspend = thin_presuspend,
4346 .postsuspend = thin_postsuspend,
4347 .status = thin_status,
4348 .merge = thin_merge,
4349 .iterate_devices = thin_iterate_devices,
4350 .io_hints = thin_io_hints,
4353 /*----------------------------------------------------------------*/
4355 static int __init dm_thin_init(void)
4361 r = dm_register_target(&thin_target);
4365 r = dm_register_target(&pool_target);
4367 goto bad_pool_target;
4371 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4372 if (!_new_mapping_cache)
4373 goto bad_new_mapping_cache;
4377 bad_new_mapping_cache:
4378 dm_unregister_target(&pool_target);
4380 dm_unregister_target(&thin_target);
4385 static void dm_thin_exit(void)
4387 dm_unregister_target(&thin_target);
4388 dm_unregister_target(&pool_target);
4390 kmem_cache_destroy(_new_mapping_cache);
4393 module_init(dm_thin_init);
4394 module_exit(dm_thin_exit);
4396 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4397 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4399 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4400 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4401 MODULE_LICENSE("GPL");