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-v1.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 */
238 bool out_of_data_space:1;
240 struct dm_bio_prison *prison;
241 struct dm_kcopyd_client *copier;
243 struct work_struct worker;
244 struct workqueue_struct *wq;
245 struct throttle throttle;
246 struct delayed_work waker;
247 struct delayed_work no_space_timeout;
249 unsigned long last_commit_jiffies;
253 struct bio_list deferred_flush_bios;
254 struct list_head prepared_mappings;
255 struct list_head prepared_discards;
256 struct list_head prepared_discards_pt2;
257 struct list_head active_thins;
259 struct dm_deferred_set *shared_read_ds;
260 struct dm_deferred_set *all_io_ds;
262 struct dm_thin_new_mapping *next_mapping;
264 process_bio_fn process_bio;
265 process_bio_fn process_discard;
267 process_cell_fn process_cell;
268 process_cell_fn process_discard_cell;
270 process_mapping_fn process_prepared_mapping;
271 process_mapping_fn process_prepared_discard;
272 process_mapping_fn process_prepared_discard_pt2;
274 struct dm_bio_prison_cell **cell_sort_array;
276 mempool_t mapping_pool;
279 static enum pool_mode get_pool_mode(struct pool *pool);
280 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
283 * Target context for a pool.
286 struct dm_target *ti;
288 struct dm_dev *data_dev;
289 struct dm_dev *metadata_dev;
290 struct dm_target_callbacks callbacks;
292 dm_block_t low_water_blocks;
293 struct pool_features requested_pf; /* Features requested during table load */
294 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
298 * Target context for a thin.
301 struct list_head list;
302 struct dm_dev *pool_dev;
303 struct dm_dev *origin_dev;
304 sector_t origin_size;
308 struct dm_thin_device *td;
309 struct mapped_device *thin_md;
313 struct list_head deferred_cells;
314 struct bio_list deferred_bio_list;
315 struct bio_list retry_on_resume_list;
316 struct rb_root sort_bio_list; /* sorted list of deferred bios */
319 * Ensures the thin is not destroyed until the worker has finished
320 * iterating the active_thins list.
323 struct completion can_destroy;
326 /*----------------------------------------------------------------*/
328 static bool block_size_is_power_of_two(struct pool *pool)
330 return pool->sectors_per_block_shift >= 0;
333 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
335 return block_size_is_power_of_two(pool) ?
336 (b << pool->sectors_per_block_shift) :
337 (b * pool->sectors_per_block);
340 /*----------------------------------------------------------------*/
344 struct blk_plug plug;
345 struct bio *parent_bio;
349 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
354 blk_start_plug(&op->plug);
355 op->parent_bio = parent;
359 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
361 struct thin_c *tc = op->tc;
362 sector_t s = block_to_sectors(tc->pool, data_b);
363 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
365 return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
366 GFP_NOWAIT, 0, &op->bio);
369 static void end_discard(struct discard_op *op, int r)
373 * Even if one of the calls to issue_discard failed, we
374 * need to wait for the chain to complete.
376 bio_chain(op->bio, op->parent_bio);
377 bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
381 blk_finish_plug(&op->plug);
384 * Even if r is set, there could be sub discards in flight that we
387 if (r && !op->parent_bio->bi_status)
388 op->parent_bio->bi_status = errno_to_blk_status(r);
389 bio_endio(op->parent_bio);
392 /*----------------------------------------------------------------*/
395 * wake_worker() is used when new work is queued and when pool_resume is
396 * ready to continue deferred IO processing.
398 static void wake_worker(struct pool *pool)
400 queue_work(pool->wq, &pool->worker);
403 /*----------------------------------------------------------------*/
405 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
406 struct dm_bio_prison_cell **cell_result)
409 struct dm_bio_prison_cell *cell_prealloc;
412 * Allocate a cell from the prison's mempool.
413 * This might block but it can't fail.
415 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
417 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
420 * We reused an old cell; we can get rid of
423 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
428 static void cell_release(struct pool *pool,
429 struct dm_bio_prison_cell *cell,
430 struct bio_list *bios)
432 dm_cell_release(pool->prison, cell, bios);
433 dm_bio_prison_free_cell(pool->prison, cell);
436 static void cell_visit_release(struct pool *pool,
437 void (*fn)(void *, struct dm_bio_prison_cell *),
439 struct dm_bio_prison_cell *cell)
441 dm_cell_visit_release(pool->prison, fn, context, cell);
442 dm_bio_prison_free_cell(pool->prison, cell);
445 static void cell_release_no_holder(struct pool *pool,
446 struct dm_bio_prison_cell *cell,
447 struct bio_list *bios)
449 dm_cell_release_no_holder(pool->prison, cell, bios);
450 dm_bio_prison_free_cell(pool->prison, cell);
453 static void cell_error_with_code(struct pool *pool,
454 struct dm_bio_prison_cell *cell, blk_status_t error_code)
456 dm_cell_error(pool->prison, cell, error_code);
457 dm_bio_prison_free_cell(pool->prison, cell);
460 static blk_status_t get_pool_io_error_code(struct pool *pool)
462 return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
465 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
467 cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
470 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
472 cell_error_with_code(pool, cell, 0);
475 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
477 cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
480 /*----------------------------------------------------------------*/
483 * A global list of pools that uses a struct mapped_device as a key.
485 static struct dm_thin_pool_table {
487 struct list_head pools;
488 } dm_thin_pool_table;
490 static void pool_table_init(void)
492 mutex_init(&dm_thin_pool_table.mutex);
493 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
496 static void pool_table_exit(void)
498 mutex_destroy(&dm_thin_pool_table.mutex);
501 static void __pool_table_insert(struct pool *pool)
503 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
504 list_add(&pool->list, &dm_thin_pool_table.pools);
507 static void __pool_table_remove(struct pool *pool)
509 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
510 list_del(&pool->list);
513 static struct pool *__pool_table_lookup(struct mapped_device *md)
515 struct pool *pool = NULL, *tmp;
517 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
519 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
520 if (tmp->pool_md == md) {
529 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
531 struct pool *pool = NULL, *tmp;
533 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
535 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
536 if (tmp->md_dev == md_dev) {
545 /*----------------------------------------------------------------*/
547 struct dm_thin_endio_hook {
549 struct dm_deferred_entry *shared_read_entry;
550 struct dm_deferred_entry *all_io_entry;
551 struct dm_thin_new_mapping *overwrite_mapping;
552 struct rb_node rb_node;
553 struct dm_bio_prison_cell *cell;
556 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
558 bio_list_merge(bios, master);
559 bio_list_init(master);
562 static void error_bio_list(struct bio_list *bios, blk_status_t error)
566 while ((bio = bio_list_pop(bios))) {
567 bio->bi_status = error;
572 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
575 struct bio_list bios;
578 bio_list_init(&bios);
580 spin_lock_irqsave(&tc->lock, flags);
581 __merge_bio_list(&bios, master);
582 spin_unlock_irqrestore(&tc->lock, flags);
584 error_bio_list(&bios, error);
587 static void requeue_deferred_cells(struct thin_c *tc)
589 struct pool *pool = tc->pool;
591 struct list_head cells;
592 struct dm_bio_prison_cell *cell, *tmp;
594 INIT_LIST_HEAD(&cells);
596 spin_lock_irqsave(&tc->lock, flags);
597 list_splice_init(&tc->deferred_cells, &cells);
598 spin_unlock_irqrestore(&tc->lock, flags);
600 list_for_each_entry_safe(cell, tmp, &cells, user_list)
601 cell_requeue(pool, cell);
604 static void requeue_io(struct thin_c *tc)
606 struct bio_list bios;
609 bio_list_init(&bios);
611 spin_lock_irqsave(&tc->lock, flags);
612 __merge_bio_list(&bios, &tc->deferred_bio_list);
613 __merge_bio_list(&bios, &tc->retry_on_resume_list);
614 spin_unlock_irqrestore(&tc->lock, flags);
616 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
617 requeue_deferred_cells(tc);
620 static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
625 list_for_each_entry_rcu(tc, &pool->active_thins, list)
626 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
630 static void error_retry_list(struct pool *pool)
632 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
636 * This section of code contains the logic for processing a thin device's IO.
637 * Much of the code depends on pool object resources (lists, workqueues, etc)
638 * but most is exclusively called from the thin target rather than the thin-pool
642 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
644 struct pool *pool = tc->pool;
645 sector_t block_nr = bio->bi_iter.bi_sector;
647 if (block_size_is_power_of_two(pool))
648 block_nr >>= pool->sectors_per_block_shift;
650 (void) sector_div(block_nr, pool->sectors_per_block);
656 * Returns the _complete_ blocks that this bio covers.
658 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
659 dm_block_t *begin, dm_block_t *end)
661 struct pool *pool = tc->pool;
662 sector_t b = bio->bi_iter.bi_sector;
663 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
665 b += pool->sectors_per_block - 1ull; /* so we round up */
667 if (block_size_is_power_of_two(pool)) {
668 b >>= pool->sectors_per_block_shift;
669 e >>= pool->sectors_per_block_shift;
671 (void) sector_div(b, pool->sectors_per_block);
672 (void) sector_div(e, pool->sectors_per_block);
676 /* Can happen if the bio is within a single block. */
683 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
685 struct pool *pool = tc->pool;
686 sector_t bi_sector = bio->bi_iter.bi_sector;
688 bio_set_dev(bio, tc->pool_dev->bdev);
689 if (block_size_is_power_of_two(pool))
690 bio->bi_iter.bi_sector =
691 (block << pool->sectors_per_block_shift) |
692 (bi_sector & (pool->sectors_per_block - 1));
694 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
695 sector_div(bi_sector, pool->sectors_per_block);
698 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
700 bio_set_dev(bio, tc->origin_dev->bdev);
703 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
705 return op_is_flush(bio->bi_opf) &&
706 dm_thin_changed_this_transaction(tc->td);
709 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
711 struct dm_thin_endio_hook *h;
713 if (bio_op(bio) == REQ_OP_DISCARD)
716 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
717 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
720 static void issue(struct thin_c *tc, struct bio *bio)
722 struct pool *pool = tc->pool;
725 if (!bio_triggers_commit(tc, bio)) {
726 generic_make_request(bio);
731 * Complete bio with an error if earlier I/O caused changes to
732 * the metadata that can't be committed e.g, due to I/O errors
733 * on the metadata device.
735 if (dm_thin_aborted_changes(tc->td)) {
741 * Batch together any bios that trigger commits and then issue a
742 * single commit for them in process_deferred_bios().
744 spin_lock_irqsave(&pool->lock, flags);
745 bio_list_add(&pool->deferred_flush_bios, bio);
746 spin_unlock_irqrestore(&pool->lock, flags);
749 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
751 remap_to_origin(tc, bio);
755 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
758 remap(tc, bio, block);
762 /*----------------------------------------------------------------*/
765 * Bio endio functions.
767 struct dm_thin_new_mapping {
768 struct list_head list;
774 * Track quiescing, copying and zeroing preparation actions. When this
775 * counter hits zero the block is prepared and can be inserted into the
778 atomic_t prepare_actions;
782 dm_block_t virt_begin, virt_end;
783 dm_block_t data_block;
784 struct dm_bio_prison_cell *cell;
787 * If the bio covers the whole area of a block then we can avoid
788 * zeroing or copying. Instead this bio is hooked. The bio will
789 * still be in the cell, so care has to be taken to avoid issuing
793 bio_end_io_t *saved_bi_end_io;
796 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
798 struct pool *pool = m->tc->pool;
800 if (atomic_dec_and_test(&m->prepare_actions)) {
801 list_add_tail(&m->list, &pool->prepared_mappings);
806 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
809 struct pool *pool = m->tc->pool;
811 spin_lock_irqsave(&pool->lock, flags);
812 __complete_mapping_preparation(m);
813 spin_unlock_irqrestore(&pool->lock, flags);
816 static void copy_complete(int read_err, unsigned long write_err, void *context)
818 struct dm_thin_new_mapping *m = context;
820 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
821 complete_mapping_preparation(m);
824 static void overwrite_endio(struct bio *bio)
826 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
827 struct dm_thin_new_mapping *m = h->overwrite_mapping;
829 bio->bi_end_io = m->saved_bi_end_io;
831 m->status = bio->bi_status;
832 complete_mapping_preparation(m);
835 /*----------------------------------------------------------------*/
842 * Prepared mapping jobs.
846 * This sends the bios in the cell, except the original holder, back
847 * to the deferred_bios list.
849 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
851 struct pool *pool = tc->pool;
854 spin_lock_irqsave(&tc->lock, flags);
855 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
856 spin_unlock_irqrestore(&tc->lock, flags);
861 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
865 struct bio_list defer_bios;
866 struct bio_list issue_bios;
869 static void __inc_remap_and_issue_cell(void *context,
870 struct dm_bio_prison_cell *cell)
872 struct remap_info *info = context;
875 while ((bio = bio_list_pop(&cell->bios))) {
876 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
877 bio_list_add(&info->defer_bios, bio);
879 inc_all_io_entry(info->tc->pool, bio);
882 * We can't issue the bios with the bio prison lock
883 * held, so we add them to a list to issue on
884 * return from this function.
886 bio_list_add(&info->issue_bios, bio);
891 static void inc_remap_and_issue_cell(struct thin_c *tc,
892 struct dm_bio_prison_cell *cell,
896 struct remap_info info;
899 bio_list_init(&info.defer_bios);
900 bio_list_init(&info.issue_bios);
903 * We have to be careful to inc any bios we're about to issue
904 * before the cell is released, and avoid a race with new bios
905 * being added to the cell.
907 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
910 while ((bio = bio_list_pop(&info.defer_bios)))
911 thin_defer_bio(tc, bio);
913 while ((bio = bio_list_pop(&info.issue_bios)))
914 remap_and_issue(info.tc, bio, block);
917 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
919 cell_error(m->tc->pool, m->cell);
921 mempool_free(m, &m->tc->pool->mapping_pool);
924 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
926 struct thin_c *tc = m->tc;
927 struct pool *pool = tc->pool;
928 struct bio *bio = m->bio;
932 cell_error(pool, m->cell);
937 * Commit the prepared block into the mapping btree.
938 * Any I/O for this block arriving after this point will get
939 * remapped to it directly.
941 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
943 metadata_operation_failed(pool, "dm_thin_insert_block", r);
944 cell_error(pool, m->cell);
949 * Release any bios held while the block was being provisioned.
950 * If we are processing a write bio that completely covers the block,
951 * we already processed it so can ignore it now when processing
952 * the bios in the cell.
955 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
958 inc_all_io_entry(tc->pool, m->cell->holder);
959 remap_and_issue(tc, m->cell->holder, m->data_block);
960 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
965 mempool_free(m, &pool->mapping_pool);
968 /*----------------------------------------------------------------*/
970 static void free_discard_mapping(struct dm_thin_new_mapping *m)
972 struct thin_c *tc = m->tc;
974 cell_defer_no_holder(tc, m->cell);
975 mempool_free(m, &tc->pool->mapping_pool);
978 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
980 bio_io_error(m->bio);
981 free_discard_mapping(m);
984 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
987 free_discard_mapping(m);
990 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
993 struct thin_c *tc = m->tc;
995 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
997 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
998 bio_io_error(m->bio);
1002 cell_defer_no_holder(tc, m->cell);
1003 mempool_free(m, &tc->pool->mapping_pool);
1006 /*----------------------------------------------------------------*/
1008 static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1009 struct bio *discard_parent)
1012 * We've already unmapped this range of blocks, but before we
1013 * passdown we have to check that these blocks are now unused.
1017 struct thin_c *tc = m->tc;
1018 struct pool *pool = tc->pool;
1019 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1020 struct discard_op op;
1022 begin_discard(&op, tc, discard_parent);
1024 /* find start of unmapped run */
1025 for (; b < end; b++) {
1026 r = dm_pool_block_is_used(pool->pmd, b, &used);
1037 /* find end of run */
1038 for (e = b + 1; e != end; e++) {
1039 r = dm_pool_block_is_used(pool->pmd, e, &used);
1047 r = issue_discard(&op, b, e);
1054 end_discard(&op, r);
1057 static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1059 unsigned long flags;
1060 struct pool *pool = m->tc->pool;
1062 spin_lock_irqsave(&pool->lock, flags);
1063 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1064 spin_unlock_irqrestore(&pool->lock, flags);
1068 static void passdown_endio(struct bio *bio)
1071 * It doesn't matter if the passdown discard failed, we still want
1072 * to unmap (we ignore err).
1074 queue_passdown_pt2(bio->bi_private);
1078 static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1081 struct thin_c *tc = m->tc;
1082 struct pool *pool = tc->pool;
1083 struct bio *discard_parent;
1084 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1087 * Only this thread allocates blocks, so we can be sure that the
1088 * newly unmapped blocks will not be allocated before the end of
1091 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1093 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1094 bio_io_error(m->bio);
1095 cell_defer_no_holder(tc, m->cell);
1096 mempool_free(m, &pool->mapping_pool);
1101 * Increment the unmapped blocks. This prevents a race between the
1102 * passdown io and reallocation of freed blocks.
1104 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1106 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1107 bio_io_error(m->bio);
1108 cell_defer_no_holder(tc, m->cell);
1109 mempool_free(m, &pool->mapping_pool);
1113 discard_parent = bio_alloc(GFP_NOIO, 1);
1114 if (!discard_parent) {
1115 DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1116 dm_device_name(tc->pool->pool_md));
1117 queue_passdown_pt2(m);
1120 discard_parent->bi_end_io = passdown_endio;
1121 discard_parent->bi_private = m;
1123 if (m->maybe_shared)
1124 passdown_double_checking_shared_status(m, discard_parent);
1126 struct discard_op op;
1128 begin_discard(&op, tc, discard_parent);
1129 r = issue_discard(&op, m->data_block, data_end);
1130 end_discard(&op, r);
1135 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1138 struct thin_c *tc = m->tc;
1139 struct pool *pool = tc->pool;
1142 * The passdown has completed, so now we can decrement all those
1145 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1146 m->data_block + (m->virt_end - m->virt_begin));
1148 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1149 bio_io_error(m->bio);
1153 cell_defer_no_holder(tc, m->cell);
1154 mempool_free(m, &pool->mapping_pool);
1157 static void process_prepared(struct pool *pool, struct list_head *head,
1158 process_mapping_fn *fn)
1160 unsigned long flags;
1161 struct list_head maps;
1162 struct dm_thin_new_mapping *m, *tmp;
1164 INIT_LIST_HEAD(&maps);
1165 spin_lock_irqsave(&pool->lock, flags);
1166 list_splice_init(head, &maps);
1167 spin_unlock_irqrestore(&pool->lock, flags);
1169 list_for_each_entry_safe(m, tmp, &maps, list)
1174 * Deferred bio jobs.
1176 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1178 return bio->bi_iter.bi_size ==
1179 (pool->sectors_per_block << SECTOR_SHIFT);
1182 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1184 return (bio_data_dir(bio) == WRITE) &&
1185 io_overlaps_block(pool, bio);
1188 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1191 *save = bio->bi_end_io;
1192 bio->bi_end_io = fn;
1195 static int ensure_next_mapping(struct pool *pool)
1197 if (pool->next_mapping)
1200 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1202 return pool->next_mapping ? 0 : -ENOMEM;
1205 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1207 struct dm_thin_new_mapping *m = pool->next_mapping;
1209 BUG_ON(!pool->next_mapping);
1211 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1212 INIT_LIST_HEAD(&m->list);
1215 pool->next_mapping = NULL;
1220 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1221 sector_t begin, sector_t end)
1224 struct dm_io_region to;
1226 to.bdev = tc->pool_dev->bdev;
1228 to.count = end - begin;
1230 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1232 DMERR_LIMIT("dm_kcopyd_zero() failed");
1233 copy_complete(1, 1, m);
1237 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1238 dm_block_t data_begin,
1239 struct dm_thin_new_mapping *m)
1241 struct pool *pool = tc->pool;
1242 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1244 h->overwrite_mapping = m;
1246 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1247 inc_all_io_entry(pool, bio);
1248 remap_and_issue(tc, bio, data_begin);
1252 * A partial copy also needs to zero the uncopied region.
1254 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1255 struct dm_dev *origin, dm_block_t data_origin,
1256 dm_block_t data_dest,
1257 struct dm_bio_prison_cell *cell, struct bio *bio,
1261 struct pool *pool = tc->pool;
1262 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1265 m->virt_begin = virt_block;
1266 m->virt_end = virt_block + 1u;
1267 m->data_block = data_dest;
1271 * quiesce action + copy action + an extra reference held for the
1272 * duration of this function (we may need to inc later for a
1275 atomic_set(&m->prepare_actions, 3);
1277 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1278 complete_mapping_preparation(m); /* already quiesced */
1281 * IO to pool_dev remaps to the pool target's data_dev.
1283 * If the whole block of data is being overwritten, we can issue the
1284 * bio immediately. Otherwise we use kcopyd to clone the data first.
1286 if (io_overwrites_block(pool, bio))
1287 remap_and_issue_overwrite(tc, bio, data_dest, m);
1289 struct dm_io_region from, to;
1291 from.bdev = origin->bdev;
1292 from.sector = data_origin * pool->sectors_per_block;
1295 to.bdev = tc->pool_dev->bdev;
1296 to.sector = data_dest * pool->sectors_per_block;
1299 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1300 0, copy_complete, m);
1302 DMERR_LIMIT("dm_kcopyd_copy() failed");
1303 copy_complete(1, 1, m);
1306 * We allow the zero to be issued, to simplify the
1307 * error path. Otherwise we'd need to start
1308 * worrying about decrementing the prepare_actions
1314 * Do we need to zero a tail region?
1316 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1317 atomic_inc(&m->prepare_actions);
1319 data_dest * pool->sectors_per_block + len,
1320 (data_dest + 1) * pool->sectors_per_block);
1324 complete_mapping_preparation(m); /* drop our ref */
1327 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1328 dm_block_t data_origin, dm_block_t data_dest,
1329 struct dm_bio_prison_cell *cell, struct bio *bio)
1331 schedule_copy(tc, virt_block, tc->pool_dev,
1332 data_origin, data_dest, cell, bio,
1333 tc->pool->sectors_per_block);
1336 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1337 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1340 struct pool *pool = tc->pool;
1341 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1343 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1345 m->virt_begin = virt_block;
1346 m->virt_end = virt_block + 1u;
1347 m->data_block = data_block;
1351 * If the whole block of data is being overwritten or we are not
1352 * zeroing pre-existing data, we can issue the bio immediately.
1353 * Otherwise we use kcopyd to zero the data first.
1355 if (pool->pf.zero_new_blocks) {
1356 if (io_overwrites_block(pool, bio))
1357 remap_and_issue_overwrite(tc, bio, data_block, m);
1359 ll_zero(tc, m, data_block * pool->sectors_per_block,
1360 (data_block + 1) * pool->sectors_per_block);
1362 process_prepared_mapping(m);
1365 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1366 dm_block_t data_dest,
1367 struct dm_bio_prison_cell *cell, struct bio *bio)
1369 struct pool *pool = tc->pool;
1370 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1371 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1373 if (virt_block_end <= tc->origin_size)
1374 schedule_copy(tc, virt_block, tc->origin_dev,
1375 virt_block, data_dest, cell, bio,
1376 pool->sectors_per_block);
1378 else if (virt_block_begin < tc->origin_size)
1379 schedule_copy(tc, virt_block, tc->origin_dev,
1380 virt_block, data_dest, cell, bio,
1381 tc->origin_size - virt_block_begin);
1384 schedule_zero(tc, virt_block, data_dest, cell, bio);
1387 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1389 static void requeue_bios(struct pool *pool);
1391 static void check_for_space(struct pool *pool)
1396 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1399 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1404 set_pool_mode(pool, PM_WRITE);
1410 * A non-zero return indicates read_only or fail_io mode.
1411 * Many callers don't care about the return value.
1413 static int commit(struct pool *pool)
1417 if (get_pool_mode(pool) >= PM_READ_ONLY)
1420 r = dm_pool_commit_metadata(pool->pmd);
1422 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1424 check_for_space(pool);
1429 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1431 unsigned long flags;
1433 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1434 DMWARN("%s: reached low water mark for data device: sending event.",
1435 dm_device_name(pool->pool_md));
1436 spin_lock_irqsave(&pool->lock, flags);
1437 pool->low_water_triggered = true;
1438 spin_unlock_irqrestore(&pool->lock, flags);
1439 dm_table_event(pool->ti->table);
1443 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1446 dm_block_t free_blocks;
1447 struct pool *pool = tc->pool;
1449 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1452 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1454 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1458 check_low_water_mark(pool, free_blocks);
1462 * Try to commit to see if that will free up some
1469 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1471 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1476 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1481 r = dm_pool_alloc_data_block(pool->pmd, result);
1484 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1486 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1494 * If we have run out of space, queue bios until the device is
1495 * resumed, presumably after having been reloaded with more space.
1497 static void retry_on_resume(struct bio *bio)
1499 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1500 struct thin_c *tc = h->tc;
1501 unsigned long flags;
1503 spin_lock_irqsave(&tc->lock, flags);
1504 bio_list_add(&tc->retry_on_resume_list, bio);
1505 spin_unlock_irqrestore(&tc->lock, flags);
1508 static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1510 enum pool_mode m = get_pool_mode(pool);
1514 /* Shouldn't get here */
1515 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1516 return BLK_STS_IOERR;
1518 case PM_OUT_OF_DATA_SPACE:
1519 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1523 return BLK_STS_IOERR;
1525 /* Shouldn't get here */
1526 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1527 return BLK_STS_IOERR;
1531 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1533 blk_status_t error = should_error_unserviceable_bio(pool);
1536 bio->bi_status = error;
1539 retry_on_resume(bio);
1542 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1545 struct bio_list bios;
1548 error = should_error_unserviceable_bio(pool);
1550 cell_error_with_code(pool, cell, error);
1554 bio_list_init(&bios);
1555 cell_release(pool, cell, &bios);
1557 while ((bio = bio_list_pop(&bios)))
1558 retry_on_resume(bio);
1561 static void process_discard_cell_no_passdown(struct thin_c *tc,
1562 struct dm_bio_prison_cell *virt_cell)
1564 struct pool *pool = tc->pool;
1565 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1568 * We don't need to lock the data blocks, since there's no
1569 * passdown. We only lock data blocks for allocation and breaking sharing.
1572 m->virt_begin = virt_cell->key.block_begin;
1573 m->virt_end = virt_cell->key.block_end;
1574 m->cell = virt_cell;
1575 m->bio = virt_cell->holder;
1577 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1578 pool->process_prepared_discard(m);
1581 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1584 struct pool *pool = tc->pool;
1588 struct dm_cell_key data_key;
1589 struct dm_bio_prison_cell *data_cell;
1590 struct dm_thin_new_mapping *m;
1591 dm_block_t virt_begin, virt_end, data_begin;
1593 while (begin != end) {
1594 r = ensure_next_mapping(pool);
1596 /* we did our best */
1599 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1600 &data_begin, &maybe_shared);
1603 * Silently fail, letting any mappings we've
1608 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1609 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1610 /* contention, we'll give up with this range */
1616 * IO may still be going to the destination block. We must
1617 * quiesce before we can do the removal.
1619 m = get_next_mapping(pool);
1621 m->maybe_shared = maybe_shared;
1622 m->virt_begin = virt_begin;
1623 m->virt_end = virt_end;
1624 m->data_block = data_begin;
1625 m->cell = data_cell;
1629 * The parent bio must not complete before sub discard bios are
1630 * chained to it (see end_discard's bio_chain)!
1632 * This per-mapping bi_remaining increment is paired with
1633 * the implicit decrement that occurs via bio_endio() in
1636 bio_inc_remaining(bio);
1637 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1638 pool->process_prepared_discard(m);
1644 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1646 struct bio *bio = virt_cell->holder;
1647 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1650 * The virt_cell will only get freed once the origin bio completes.
1651 * This means it will remain locked while all the individual
1652 * passdown bios are in flight.
1654 h->cell = virt_cell;
1655 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1658 * We complete the bio now, knowing that the bi_remaining field
1659 * will prevent completion until the sub range discards have
1665 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1667 dm_block_t begin, end;
1668 struct dm_cell_key virt_key;
1669 struct dm_bio_prison_cell *virt_cell;
1671 get_bio_block_range(tc, bio, &begin, &end);
1674 * The discard covers less than a block.
1680 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1681 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1683 * Potential starvation issue: We're relying on the
1684 * fs/application being well behaved, and not trying to
1685 * send IO to a region at the same time as discarding it.
1686 * If they do this persistently then it's possible this
1687 * cell will never be granted.
1691 tc->pool->process_discard_cell(tc, virt_cell);
1694 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1695 struct dm_cell_key *key,
1696 struct dm_thin_lookup_result *lookup_result,
1697 struct dm_bio_prison_cell *cell)
1700 dm_block_t data_block;
1701 struct pool *pool = tc->pool;
1703 r = alloc_data_block(tc, &data_block);
1706 schedule_internal_copy(tc, block, lookup_result->block,
1707 data_block, cell, bio);
1711 retry_bios_on_resume(pool, cell);
1715 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1717 cell_error(pool, cell);
1722 static void __remap_and_issue_shared_cell(void *context,
1723 struct dm_bio_prison_cell *cell)
1725 struct remap_info *info = context;
1728 while ((bio = bio_list_pop(&cell->bios))) {
1729 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1730 bio_op(bio) == REQ_OP_DISCARD)
1731 bio_list_add(&info->defer_bios, bio);
1733 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1735 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1736 inc_all_io_entry(info->tc->pool, bio);
1737 bio_list_add(&info->issue_bios, bio);
1742 static void remap_and_issue_shared_cell(struct thin_c *tc,
1743 struct dm_bio_prison_cell *cell,
1747 struct remap_info info;
1750 bio_list_init(&info.defer_bios);
1751 bio_list_init(&info.issue_bios);
1753 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1756 while ((bio = bio_list_pop(&info.defer_bios)))
1757 thin_defer_bio(tc, bio);
1759 while ((bio = bio_list_pop(&info.issue_bios)))
1760 remap_and_issue(tc, bio, block);
1763 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1765 struct dm_thin_lookup_result *lookup_result,
1766 struct dm_bio_prison_cell *virt_cell)
1768 struct dm_bio_prison_cell *data_cell;
1769 struct pool *pool = tc->pool;
1770 struct dm_cell_key key;
1773 * If cell is already occupied, then sharing is already in the process
1774 * of being broken so we have nothing further to do here.
1776 build_data_key(tc->td, lookup_result->block, &key);
1777 if (bio_detain(pool, &key, bio, &data_cell)) {
1778 cell_defer_no_holder(tc, virt_cell);
1782 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1783 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1784 cell_defer_no_holder(tc, virt_cell);
1786 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1788 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1789 inc_all_io_entry(pool, bio);
1790 remap_and_issue(tc, bio, lookup_result->block);
1792 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1793 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1797 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1798 struct dm_bio_prison_cell *cell)
1801 dm_block_t data_block;
1802 struct pool *pool = tc->pool;
1805 * Remap empty bios (flushes) immediately, without provisioning.
1807 if (!bio->bi_iter.bi_size) {
1808 inc_all_io_entry(pool, bio);
1809 cell_defer_no_holder(tc, cell);
1811 remap_and_issue(tc, bio, 0);
1816 * Fill read bios with zeroes and complete them immediately.
1818 if (bio_data_dir(bio) == READ) {
1820 cell_defer_no_holder(tc, cell);
1825 r = alloc_data_block(tc, &data_block);
1829 schedule_external_copy(tc, block, data_block, cell, bio);
1831 schedule_zero(tc, block, data_block, cell, bio);
1835 retry_bios_on_resume(pool, cell);
1839 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1841 cell_error(pool, cell);
1846 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1849 struct pool *pool = tc->pool;
1850 struct bio *bio = cell->holder;
1851 dm_block_t block = get_bio_block(tc, bio);
1852 struct dm_thin_lookup_result lookup_result;
1854 if (tc->requeue_mode) {
1855 cell_requeue(pool, cell);
1859 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1862 if (lookup_result.shared)
1863 process_shared_bio(tc, bio, block, &lookup_result, cell);
1865 inc_all_io_entry(pool, bio);
1866 remap_and_issue(tc, bio, lookup_result.block);
1867 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1872 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1873 inc_all_io_entry(pool, bio);
1874 cell_defer_no_holder(tc, cell);
1876 if (bio_end_sector(bio) <= tc->origin_size)
1877 remap_to_origin_and_issue(tc, bio);
1879 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1881 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1882 remap_to_origin_and_issue(tc, bio);
1889 provision_block(tc, bio, block, cell);
1893 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1895 cell_defer_no_holder(tc, cell);
1901 static void process_bio(struct thin_c *tc, struct bio *bio)
1903 struct pool *pool = tc->pool;
1904 dm_block_t block = get_bio_block(tc, bio);
1905 struct dm_bio_prison_cell *cell;
1906 struct dm_cell_key key;
1909 * If cell is already occupied, then the block is already
1910 * being provisioned so we have nothing further to do here.
1912 build_virtual_key(tc->td, block, &key);
1913 if (bio_detain(pool, &key, bio, &cell))
1916 process_cell(tc, cell);
1919 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1920 struct dm_bio_prison_cell *cell)
1923 int rw = bio_data_dir(bio);
1924 dm_block_t block = get_bio_block(tc, bio);
1925 struct dm_thin_lookup_result lookup_result;
1927 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1930 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1931 handle_unserviceable_bio(tc->pool, bio);
1933 cell_defer_no_holder(tc, cell);
1935 inc_all_io_entry(tc->pool, bio);
1936 remap_and_issue(tc, bio, lookup_result.block);
1938 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1944 cell_defer_no_holder(tc, cell);
1946 handle_unserviceable_bio(tc->pool, bio);
1950 if (tc->origin_dev) {
1951 inc_all_io_entry(tc->pool, bio);
1952 remap_to_origin_and_issue(tc, bio);
1961 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1964 cell_defer_no_holder(tc, cell);
1970 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1972 __process_bio_read_only(tc, bio, NULL);
1975 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1977 __process_bio_read_only(tc, cell->holder, cell);
1980 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1985 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1990 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1992 cell_success(tc->pool, cell);
1995 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1997 cell_error(tc->pool, cell);
2001 * FIXME: should we also commit due to size of transaction, measured in
2004 static int need_commit_due_to_time(struct pool *pool)
2006 return !time_in_range(jiffies, pool->last_commit_jiffies,
2007 pool->last_commit_jiffies + COMMIT_PERIOD);
2010 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2011 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2013 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2015 struct rb_node **rbp, *parent;
2016 struct dm_thin_endio_hook *pbd;
2017 sector_t bi_sector = bio->bi_iter.bi_sector;
2019 rbp = &tc->sort_bio_list.rb_node;
2023 pbd = thin_pbd(parent);
2025 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2026 rbp = &(*rbp)->rb_left;
2028 rbp = &(*rbp)->rb_right;
2031 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2032 rb_link_node(&pbd->rb_node, parent, rbp);
2033 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2036 static void __extract_sorted_bios(struct thin_c *tc)
2038 struct rb_node *node;
2039 struct dm_thin_endio_hook *pbd;
2042 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2043 pbd = thin_pbd(node);
2044 bio = thin_bio(pbd);
2046 bio_list_add(&tc->deferred_bio_list, bio);
2047 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2050 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2053 static void __sort_thin_deferred_bios(struct thin_c *tc)
2056 struct bio_list bios;
2058 bio_list_init(&bios);
2059 bio_list_merge(&bios, &tc->deferred_bio_list);
2060 bio_list_init(&tc->deferred_bio_list);
2062 /* Sort deferred_bio_list using rb-tree */
2063 while ((bio = bio_list_pop(&bios)))
2064 __thin_bio_rb_add(tc, bio);
2067 * Transfer the sorted bios in sort_bio_list back to
2068 * deferred_bio_list to allow lockless submission of
2071 __extract_sorted_bios(tc);
2074 static void process_thin_deferred_bios(struct thin_c *tc)
2076 struct pool *pool = tc->pool;
2077 unsigned long flags;
2079 struct bio_list bios;
2080 struct blk_plug plug;
2083 if (tc->requeue_mode) {
2084 error_thin_bio_list(tc, &tc->deferred_bio_list,
2085 BLK_STS_DM_REQUEUE);
2089 bio_list_init(&bios);
2091 spin_lock_irqsave(&tc->lock, flags);
2093 if (bio_list_empty(&tc->deferred_bio_list)) {
2094 spin_unlock_irqrestore(&tc->lock, flags);
2098 __sort_thin_deferred_bios(tc);
2100 bio_list_merge(&bios, &tc->deferred_bio_list);
2101 bio_list_init(&tc->deferred_bio_list);
2103 spin_unlock_irqrestore(&tc->lock, flags);
2105 blk_start_plug(&plug);
2106 while ((bio = bio_list_pop(&bios))) {
2108 * If we've got no free new_mapping structs, and processing
2109 * this bio might require one, we pause until there are some
2110 * prepared mappings to process.
2112 if (ensure_next_mapping(pool)) {
2113 spin_lock_irqsave(&tc->lock, flags);
2114 bio_list_add(&tc->deferred_bio_list, bio);
2115 bio_list_merge(&tc->deferred_bio_list, &bios);
2116 spin_unlock_irqrestore(&tc->lock, flags);
2120 if (bio_op(bio) == REQ_OP_DISCARD)
2121 pool->process_discard(tc, bio);
2123 pool->process_bio(tc, bio);
2125 if ((count++ & 127) == 0) {
2126 throttle_work_update(&pool->throttle);
2127 dm_pool_issue_prefetches(pool->pmd);
2130 blk_finish_plug(&plug);
2133 static int cmp_cells(const void *lhs, const void *rhs)
2135 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2136 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2138 BUG_ON(!lhs_cell->holder);
2139 BUG_ON(!rhs_cell->holder);
2141 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2144 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2150 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2153 struct dm_bio_prison_cell *cell, *tmp;
2155 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2156 if (count >= CELL_SORT_ARRAY_SIZE)
2159 pool->cell_sort_array[count++] = cell;
2160 list_del(&cell->user_list);
2163 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2168 static void process_thin_deferred_cells(struct thin_c *tc)
2170 struct pool *pool = tc->pool;
2171 unsigned long flags;
2172 struct list_head cells;
2173 struct dm_bio_prison_cell *cell;
2174 unsigned i, j, count;
2176 INIT_LIST_HEAD(&cells);
2178 spin_lock_irqsave(&tc->lock, flags);
2179 list_splice_init(&tc->deferred_cells, &cells);
2180 spin_unlock_irqrestore(&tc->lock, flags);
2182 if (list_empty(&cells))
2186 count = sort_cells(tc->pool, &cells);
2188 for (i = 0; i < count; i++) {
2189 cell = pool->cell_sort_array[i];
2190 BUG_ON(!cell->holder);
2193 * If we've got no free new_mapping structs, and processing
2194 * this bio might require one, we pause until there are some
2195 * prepared mappings to process.
2197 if (ensure_next_mapping(pool)) {
2198 for (j = i; j < count; j++)
2199 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2201 spin_lock_irqsave(&tc->lock, flags);
2202 list_splice(&cells, &tc->deferred_cells);
2203 spin_unlock_irqrestore(&tc->lock, flags);
2207 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2208 pool->process_discard_cell(tc, cell);
2210 pool->process_cell(tc, cell);
2212 } while (!list_empty(&cells));
2215 static void thin_get(struct thin_c *tc);
2216 static void thin_put(struct thin_c *tc);
2219 * We can't hold rcu_read_lock() around code that can block. So we
2220 * find a thin with the rcu lock held; bump a refcount; then drop
2223 static struct thin_c *get_first_thin(struct pool *pool)
2225 struct thin_c *tc = NULL;
2228 if (!list_empty(&pool->active_thins)) {
2229 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2237 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2239 struct thin_c *old_tc = tc;
2242 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2254 static void process_deferred_bios(struct pool *pool)
2256 unsigned long flags;
2258 struct bio_list bios;
2261 tc = get_first_thin(pool);
2263 process_thin_deferred_cells(tc);
2264 process_thin_deferred_bios(tc);
2265 tc = get_next_thin(pool, tc);
2269 * If there are any deferred flush bios, we must commit
2270 * the metadata before issuing them.
2272 bio_list_init(&bios);
2273 spin_lock_irqsave(&pool->lock, flags);
2274 bio_list_merge(&bios, &pool->deferred_flush_bios);
2275 bio_list_init(&pool->deferred_flush_bios);
2276 spin_unlock_irqrestore(&pool->lock, flags);
2278 if (bio_list_empty(&bios) &&
2279 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2283 while ((bio = bio_list_pop(&bios)))
2287 pool->last_commit_jiffies = jiffies;
2289 while ((bio = bio_list_pop(&bios)))
2290 generic_make_request(bio);
2293 static void do_worker(struct work_struct *ws)
2295 struct pool *pool = container_of(ws, struct pool, worker);
2297 throttle_work_start(&pool->throttle);
2298 dm_pool_issue_prefetches(pool->pmd);
2299 throttle_work_update(&pool->throttle);
2300 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2301 throttle_work_update(&pool->throttle);
2302 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2303 throttle_work_update(&pool->throttle);
2304 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2305 throttle_work_update(&pool->throttle);
2306 process_deferred_bios(pool);
2307 throttle_work_complete(&pool->throttle);
2311 * We want to commit periodically so that not too much
2312 * unwritten data builds up.
2314 static void do_waker(struct work_struct *ws)
2316 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2318 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2321 static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2324 * We're holding onto IO to allow userland time to react. After the
2325 * timeout either the pool will have been resized (and thus back in
2326 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2328 static void do_no_space_timeout(struct work_struct *ws)
2330 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2333 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2334 pool->pf.error_if_no_space = true;
2335 notify_of_pool_mode_change_to_oods(pool);
2336 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2340 /*----------------------------------------------------------------*/
2343 struct work_struct worker;
2344 struct completion complete;
2347 static struct pool_work *to_pool_work(struct work_struct *ws)
2349 return container_of(ws, struct pool_work, worker);
2352 static void pool_work_complete(struct pool_work *pw)
2354 complete(&pw->complete);
2357 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2358 void (*fn)(struct work_struct *))
2360 INIT_WORK_ONSTACK(&pw->worker, fn);
2361 init_completion(&pw->complete);
2362 queue_work(pool->wq, &pw->worker);
2363 wait_for_completion(&pw->complete);
2366 /*----------------------------------------------------------------*/
2368 struct noflush_work {
2369 struct pool_work pw;
2373 static struct noflush_work *to_noflush(struct work_struct *ws)
2375 return container_of(to_pool_work(ws), struct noflush_work, pw);
2378 static void do_noflush_start(struct work_struct *ws)
2380 struct noflush_work *w = to_noflush(ws);
2381 w->tc->requeue_mode = true;
2383 pool_work_complete(&w->pw);
2386 static void do_noflush_stop(struct work_struct *ws)
2388 struct noflush_work *w = to_noflush(ws);
2389 w->tc->requeue_mode = false;
2390 pool_work_complete(&w->pw);
2393 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2395 struct noflush_work w;
2398 pool_work_wait(&w.pw, tc->pool, fn);
2401 /*----------------------------------------------------------------*/
2403 static enum pool_mode get_pool_mode(struct pool *pool)
2405 return pool->pf.mode;
2408 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2410 dm_table_event(pool->ti->table);
2411 DMINFO("%s: switching pool to %s mode",
2412 dm_device_name(pool->pool_md), new_mode);
2415 static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2417 if (!pool->pf.error_if_no_space)
2418 notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2420 notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2423 static bool passdown_enabled(struct pool_c *pt)
2425 return pt->adjusted_pf.discard_passdown;
2428 static void set_discard_callbacks(struct pool *pool)
2430 struct pool_c *pt = pool->ti->private;
2432 if (passdown_enabled(pt)) {
2433 pool->process_discard_cell = process_discard_cell_passdown;
2434 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2435 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2437 pool->process_discard_cell = process_discard_cell_no_passdown;
2438 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2442 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2444 struct pool_c *pt = pool->ti->private;
2445 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2446 enum pool_mode old_mode = get_pool_mode(pool);
2447 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2450 * Never allow the pool to transition to PM_WRITE mode if user
2451 * intervention is required to verify metadata and data consistency.
2453 if (new_mode == PM_WRITE && needs_check) {
2454 DMERR("%s: unable to switch pool to write mode until repaired.",
2455 dm_device_name(pool->pool_md));
2456 if (old_mode != new_mode)
2457 new_mode = old_mode;
2459 new_mode = PM_READ_ONLY;
2462 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2463 * not going to recover without a thin_repair. So we never let the
2464 * pool move out of the old mode.
2466 if (old_mode == PM_FAIL)
2467 new_mode = old_mode;
2471 if (old_mode != new_mode)
2472 notify_of_pool_mode_change(pool, "failure");
2473 dm_pool_metadata_read_only(pool->pmd);
2474 pool->process_bio = process_bio_fail;
2475 pool->process_discard = process_bio_fail;
2476 pool->process_cell = process_cell_fail;
2477 pool->process_discard_cell = process_cell_fail;
2478 pool->process_prepared_mapping = process_prepared_mapping_fail;
2479 pool->process_prepared_discard = process_prepared_discard_fail;
2481 error_retry_list(pool);
2485 if (old_mode != new_mode)
2486 notify_of_pool_mode_change(pool, "read-only");
2487 dm_pool_metadata_read_only(pool->pmd);
2488 pool->process_bio = process_bio_read_only;
2489 pool->process_discard = process_bio_success;
2490 pool->process_cell = process_cell_read_only;
2491 pool->process_discard_cell = process_cell_success;
2492 pool->process_prepared_mapping = process_prepared_mapping_fail;
2493 pool->process_prepared_discard = process_prepared_discard_success;
2495 error_retry_list(pool);
2498 case PM_OUT_OF_DATA_SPACE:
2500 * Ideally we'd never hit this state; the low water mark
2501 * would trigger userland to extend the pool before we
2502 * completely run out of data space. However, many small
2503 * IOs to unprovisioned space can consume data space at an
2504 * alarming rate. Adjust your low water mark if you're
2505 * frequently seeing this mode.
2507 if (old_mode != new_mode)
2508 notify_of_pool_mode_change_to_oods(pool);
2509 pool->out_of_data_space = true;
2510 pool->process_bio = process_bio_read_only;
2511 pool->process_discard = process_discard_bio;
2512 pool->process_cell = process_cell_read_only;
2513 pool->process_prepared_mapping = process_prepared_mapping;
2514 set_discard_callbacks(pool);
2516 if (!pool->pf.error_if_no_space && no_space_timeout)
2517 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2521 if (old_mode != new_mode)
2522 notify_of_pool_mode_change(pool, "write");
2523 pool->out_of_data_space = false;
2524 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2525 dm_pool_metadata_read_write(pool->pmd);
2526 pool->process_bio = process_bio;
2527 pool->process_discard = process_discard_bio;
2528 pool->process_cell = process_cell;
2529 pool->process_prepared_mapping = process_prepared_mapping;
2530 set_discard_callbacks(pool);
2534 pool->pf.mode = new_mode;
2536 * The pool mode may have changed, sync it so bind_control_target()
2537 * doesn't cause an unexpected mode transition on resume.
2539 pt->adjusted_pf.mode = new_mode;
2542 static void abort_transaction(struct pool *pool)
2544 const char *dev_name = dm_device_name(pool->pool_md);
2546 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2547 if (dm_pool_abort_metadata(pool->pmd)) {
2548 DMERR("%s: failed to abort metadata transaction", dev_name);
2549 set_pool_mode(pool, PM_FAIL);
2552 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2553 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2554 set_pool_mode(pool, PM_FAIL);
2558 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2560 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2561 dm_device_name(pool->pool_md), op, r);
2563 abort_transaction(pool);
2564 set_pool_mode(pool, PM_READ_ONLY);
2567 /*----------------------------------------------------------------*/
2570 * Mapping functions.
2574 * Called only while mapping a thin bio to hand it over to the workqueue.
2576 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2578 unsigned long flags;
2579 struct pool *pool = tc->pool;
2581 spin_lock_irqsave(&tc->lock, flags);
2582 bio_list_add(&tc->deferred_bio_list, bio);
2583 spin_unlock_irqrestore(&tc->lock, flags);
2588 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2590 struct pool *pool = tc->pool;
2592 throttle_lock(&pool->throttle);
2593 thin_defer_bio(tc, bio);
2594 throttle_unlock(&pool->throttle);
2597 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2599 unsigned long flags;
2600 struct pool *pool = tc->pool;
2602 throttle_lock(&pool->throttle);
2603 spin_lock_irqsave(&tc->lock, flags);
2604 list_add_tail(&cell->user_list, &tc->deferred_cells);
2605 spin_unlock_irqrestore(&tc->lock, flags);
2606 throttle_unlock(&pool->throttle);
2611 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2613 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2616 h->shared_read_entry = NULL;
2617 h->all_io_entry = NULL;
2618 h->overwrite_mapping = NULL;
2623 * Non-blocking function called from the thin target's map function.
2625 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2628 struct thin_c *tc = ti->private;
2629 dm_block_t block = get_bio_block(tc, bio);
2630 struct dm_thin_device *td = tc->td;
2631 struct dm_thin_lookup_result result;
2632 struct dm_bio_prison_cell *virt_cell, *data_cell;
2633 struct dm_cell_key key;
2635 thin_hook_bio(tc, bio);
2637 if (tc->requeue_mode) {
2638 bio->bi_status = BLK_STS_DM_REQUEUE;
2640 return DM_MAPIO_SUBMITTED;
2643 if (get_pool_mode(tc->pool) == PM_FAIL) {
2645 return DM_MAPIO_SUBMITTED;
2648 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2649 thin_defer_bio_with_throttle(tc, bio);
2650 return DM_MAPIO_SUBMITTED;
2654 * We must hold the virtual cell before doing the lookup, otherwise
2655 * there's a race with discard.
2657 build_virtual_key(tc->td, block, &key);
2658 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2659 return DM_MAPIO_SUBMITTED;
2661 r = dm_thin_find_block(td, block, 0, &result);
2664 * Note that we defer readahead too.
2668 if (unlikely(result.shared)) {
2670 * We have a race condition here between the
2671 * result.shared value returned by the lookup and
2672 * snapshot creation, which may cause new
2675 * To avoid this always quiesce the origin before
2676 * taking the snap. You want to do this anyway to
2677 * ensure a consistent application view
2680 * More distant ancestors are irrelevant. The
2681 * shared flag will be set in their case.
2683 thin_defer_cell(tc, virt_cell);
2684 return DM_MAPIO_SUBMITTED;
2687 build_data_key(tc->td, result.block, &key);
2688 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2689 cell_defer_no_holder(tc, virt_cell);
2690 return DM_MAPIO_SUBMITTED;
2693 inc_all_io_entry(tc->pool, bio);
2694 cell_defer_no_holder(tc, data_cell);
2695 cell_defer_no_holder(tc, virt_cell);
2697 remap(tc, bio, result.block);
2698 return DM_MAPIO_REMAPPED;
2702 thin_defer_cell(tc, virt_cell);
2703 return DM_MAPIO_SUBMITTED;
2707 * Must always call bio_io_error on failure.
2708 * dm_thin_find_block can fail with -EINVAL if the
2709 * pool is switched to fail-io mode.
2712 cell_defer_no_holder(tc, virt_cell);
2713 return DM_MAPIO_SUBMITTED;
2717 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2719 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2720 struct request_queue *q;
2722 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2725 q = bdev_get_queue(pt->data_dev->bdev);
2726 return bdi_congested(q->backing_dev_info, bdi_bits);
2729 static void requeue_bios(struct pool *pool)
2731 unsigned long flags;
2735 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2736 spin_lock_irqsave(&tc->lock, flags);
2737 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2738 bio_list_init(&tc->retry_on_resume_list);
2739 spin_unlock_irqrestore(&tc->lock, flags);
2744 /*----------------------------------------------------------------
2745 * Binding of control targets to a pool object
2746 *--------------------------------------------------------------*/
2747 static bool data_dev_supports_discard(struct pool_c *pt)
2749 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2751 return q && blk_queue_discard(q);
2754 static bool is_factor(sector_t block_size, uint32_t n)
2756 return !sector_div(block_size, n);
2760 * If discard_passdown was enabled verify that the data device
2761 * supports discards. Disable discard_passdown if not.
2763 static void disable_passdown_if_not_supported(struct pool_c *pt)
2765 struct pool *pool = pt->pool;
2766 struct block_device *data_bdev = pt->data_dev->bdev;
2767 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2768 const char *reason = NULL;
2769 char buf[BDEVNAME_SIZE];
2771 if (!pt->adjusted_pf.discard_passdown)
2774 if (!data_dev_supports_discard(pt))
2775 reason = "discard unsupported";
2777 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2778 reason = "max discard sectors smaller than a block";
2781 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2782 pt->adjusted_pf.discard_passdown = false;
2786 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2788 struct pool_c *pt = ti->private;
2791 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2793 enum pool_mode old_mode = get_pool_mode(pool);
2794 enum pool_mode new_mode = pt->adjusted_pf.mode;
2797 * Don't change the pool's mode until set_pool_mode() below.
2798 * Otherwise the pool's process_* function pointers may
2799 * not match the desired pool mode.
2801 pt->adjusted_pf.mode = old_mode;
2804 pool->pf = pt->adjusted_pf;
2805 pool->low_water_blocks = pt->low_water_blocks;
2807 set_pool_mode(pool, new_mode);
2812 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2818 /*----------------------------------------------------------------
2820 *--------------------------------------------------------------*/
2821 /* Initialize pool features. */
2822 static void pool_features_init(struct pool_features *pf)
2824 pf->mode = PM_WRITE;
2825 pf->zero_new_blocks = true;
2826 pf->discard_enabled = true;
2827 pf->discard_passdown = true;
2828 pf->error_if_no_space = false;
2831 static void __pool_destroy(struct pool *pool)
2833 __pool_table_remove(pool);
2835 vfree(pool->cell_sort_array);
2836 if (dm_pool_metadata_close(pool->pmd) < 0)
2837 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2839 dm_bio_prison_destroy(pool->prison);
2840 dm_kcopyd_client_destroy(pool->copier);
2843 destroy_workqueue(pool->wq);
2845 if (pool->next_mapping)
2846 mempool_free(pool->next_mapping, &pool->mapping_pool);
2847 mempool_exit(&pool->mapping_pool);
2848 dm_deferred_set_destroy(pool->shared_read_ds);
2849 dm_deferred_set_destroy(pool->all_io_ds);
2853 static struct kmem_cache *_new_mapping_cache;
2855 static struct pool *pool_create(struct mapped_device *pool_md,
2856 struct block_device *metadata_dev,
2857 unsigned long block_size,
2858 int read_only, char **error)
2863 struct dm_pool_metadata *pmd;
2864 bool format_device = read_only ? false : true;
2866 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2868 *error = "Error creating metadata object";
2869 return (struct pool *)pmd;
2872 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2874 *error = "Error allocating memory for pool";
2875 err_p = ERR_PTR(-ENOMEM);
2880 pool->sectors_per_block = block_size;
2881 if (block_size & (block_size - 1))
2882 pool->sectors_per_block_shift = -1;
2884 pool->sectors_per_block_shift = __ffs(block_size);
2885 pool->low_water_blocks = 0;
2886 pool_features_init(&pool->pf);
2887 pool->prison = dm_bio_prison_create();
2888 if (!pool->prison) {
2889 *error = "Error creating pool's bio prison";
2890 err_p = ERR_PTR(-ENOMEM);
2894 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2895 if (IS_ERR(pool->copier)) {
2896 r = PTR_ERR(pool->copier);
2897 *error = "Error creating pool's kcopyd client";
2899 goto bad_kcopyd_client;
2903 * Create singlethreaded workqueue that will service all devices
2904 * that use this metadata.
2906 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2908 *error = "Error creating pool's workqueue";
2909 err_p = ERR_PTR(-ENOMEM);
2913 throttle_init(&pool->throttle);
2914 INIT_WORK(&pool->worker, do_worker);
2915 INIT_DELAYED_WORK(&pool->waker, do_waker);
2916 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2917 spin_lock_init(&pool->lock);
2918 bio_list_init(&pool->deferred_flush_bios);
2919 INIT_LIST_HEAD(&pool->prepared_mappings);
2920 INIT_LIST_HEAD(&pool->prepared_discards);
2921 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2922 INIT_LIST_HEAD(&pool->active_thins);
2923 pool->low_water_triggered = false;
2924 pool->suspended = true;
2925 pool->out_of_data_space = false;
2927 pool->shared_read_ds = dm_deferred_set_create();
2928 if (!pool->shared_read_ds) {
2929 *error = "Error creating pool's shared read deferred set";
2930 err_p = ERR_PTR(-ENOMEM);
2931 goto bad_shared_read_ds;
2934 pool->all_io_ds = dm_deferred_set_create();
2935 if (!pool->all_io_ds) {
2936 *error = "Error creating pool's all io deferred set";
2937 err_p = ERR_PTR(-ENOMEM);
2941 pool->next_mapping = NULL;
2942 r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
2943 _new_mapping_cache);
2945 *error = "Error creating pool's mapping mempool";
2947 goto bad_mapping_pool;
2950 pool->cell_sort_array =
2951 vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
2952 sizeof(*pool->cell_sort_array)));
2953 if (!pool->cell_sort_array) {
2954 *error = "Error allocating cell sort array";
2955 err_p = ERR_PTR(-ENOMEM);
2956 goto bad_sort_array;
2959 pool->ref_count = 1;
2960 pool->last_commit_jiffies = jiffies;
2961 pool->pool_md = pool_md;
2962 pool->md_dev = metadata_dev;
2963 __pool_table_insert(pool);
2968 mempool_exit(&pool->mapping_pool);
2970 dm_deferred_set_destroy(pool->all_io_ds);
2972 dm_deferred_set_destroy(pool->shared_read_ds);
2974 destroy_workqueue(pool->wq);
2976 dm_kcopyd_client_destroy(pool->copier);
2978 dm_bio_prison_destroy(pool->prison);
2982 if (dm_pool_metadata_close(pmd))
2983 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2988 static void __pool_inc(struct pool *pool)
2990 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2994 static void __pool_dec(struct pool *pool)
2996 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2997 BUG_ON(!pool->ref_count);
2998 if (!--pool->ref_count)
2999 __pool_destroy(pool);
3002 static struct pool *__pool_find(struct mapped_device *pool_md,
3003 struct block_device *metadata_dev,
3004 unsigned long block_size, int read_only,
3005 char **error, int *created)
3007 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3010 if (pool->pool_md != pool_md) {
3011 *error = "metadata device already in use by a pool";
3012 return ERR_PTR(-EBUSY);
3017 pool = __pool_table_lookup(pool_md);
3019 if (pool->md_dev != metadata_dev) {
3020 *error = "different pool cannot replace a pool";
3021 return ERR_PTR(-EINVAL);
3026 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3034 /*----------------------------------------------------------------
3035 * Pool target methods
3036 *--------------------------------------------------------------*/
3037 static void pool_dtr(struct dm_target *ti)
3039 struct pool_c *pt = ti->private;
3041 mutex_lock(&dm_thin_pool_table.mutex);
3043 unbind_control_target(pt->pool, ti);
3044 __pool_dec(pt->pool);
3045 dm_put_device(ti, pt->metadata_dev);
3046 dm_put_device(ti, pt->data_dev);
3049 mutex_unlock(&dm_thin_pool_table.mutex);
3052 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3053 struct dm_target *ti)
3057 const char *arg_name;
3059 static const struct dm_arg _args[] = {
3060 {0, 4, "Invalid number of pool feature arguments"},
3064 * No feature arguments supplied.
3069 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3073 while (argc && !r) {
3074 arg_name = dm_shift_arg(as);
3077 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3078 pf->zero_new_blocks = false;
3080 else if (!strcasecmp(arg_name, "ignore_discard"))
3081 pf->discard_enabled = false;
3083 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3084 pf->discard_passdown = false;
3086 else if (!strcasecmp(arg_name, "read_only"))
3087 pf->mode = PM_READ_ONLY;
3089 else if (!strcasecmp(arg_name, "error_if_no_space"))
3090 pf->error_if_no_space = true;
3093 ti->error = "Unrecognised pool feature requested";
3102 static void metadata_low_callback(void *context)
3104 struct pool *pool = context;
3106 DMWARN("%s: reached low water mark for metadata device: sending event.",
3107 dm_device_name(pool->pool_md));
3109 dm_table_event(pool->ti->table);
3112 static sector_t get_dev_size(struct block_device *bdev)
3114 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3117 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3119 sector_t metadata_dev_size = get_dev_size(bdev);
3120 char buffer[BDEVNAME_SIZE];
3122 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3123 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3124 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3127 static sector_t get_metadata_dev_size(struct block_device *bdev)
3129 sector_t metadata_dev_size = get_dev_size(bdev);
3131 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3132 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3134 return metadata_dev_size;
3137 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3139 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3141 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3143 return metadata_dev_size;
3147 * When a metadata threshold is crossed a dm event is triggered, and
3148 * userland should respond by growing the metadata device. We could let
3149 * userland set the threshold, like we do with the data threshold, but I'm
3150 * not sure they know enough to do this well.
3152 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3155 * 4M is ample for all ops with the possible exception of thin
3156 * device deletion which is harmless if it fails (just retry the
3157 * delete after you've grown the device).
3159 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3160 return min((dm_block_t)1024ULL /* 4M */, quarter);
3164 * thin-pool <metadata dev> <data dev>
3165 * <data block size (sectors)>
3166 * <low water mark (blocks)>
3167 * [<#feature args> [<arg>]*]
3169 * Optional feature arguments are:
3170 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3171 * ignore_discard: disable discard
3172 * no_discard_passdown: don't pass discards down to the data device
3173 * read_only: Don't allow any changes to be made to the pool metadata.
3174 * error_if_no_space: error IOs, instead of queueing, if no space.
3176 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3178 int r, pool_created = 0;
3181 struct pool_features pf;
3182 struct dm_arg_set as;
3183 struct dm_dev *data_dev;
3184 unsigned long block_size;
3185 dm_block_t low_water_blocks;
3186 struct dm_dev *metadata_dev;
3187 fmode_t metadata_mode;
3190 * FIXME Remove validation from scope of lock.
3192 mutex_lock(&dm_thin_pool_table.mutex);
3195 ti->error = "Invalid argument count";
3204 * Set default pool features.
3206 pool_features_init(&pf);
3208 dm_consume_args(&as, 4);
3209 r = parse_pool_features(&as, &pf, ti);
3213 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3214 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3216 ti->error = "Error opening metadata block device";
3219 warn_if_metadata_device_too_big(metadata_dev->bdev);
3221 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3223 ti->error = "Error getting data device";
3227 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3228 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3229 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3230 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3231 ti->error = "Invalid block size";
3236 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3237 ti->error = "Invalid low water mark";
3242 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3248 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3249 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3256 * 'pool_created' reflects whether this is the first table load.
3257 * Top level discard support is not allowed to be changed after
3258 * initial load. This would require a pool reload to trigger thin
3261 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3262 ti->error = "Discard support cannot be disabled once enabled";
3264 goto out_flags_changed;
3269 pt->metadata_dev = metadata_dev;
3270 pt->data_dev = data_dev;
3271 pt->low_water_blocks = low_water_blocks;
3272 pt->adjusted_pf = pt->requested_pf = pf;
3273 ti->num_flush_bios = 1;
3276 * Only need to enable discards if the pool should pass
3277 * them down to the data device. The thin device's discard
3278 * processing will cause mappings to be removed from the btree.
3280 if (pf.discard_enabled && pf.discard_passdown) {
3281 ti->num_discard_bios = 1;
3284 * Setting 'discards_supported' circumvents the normal
3285 * stacking of discard limits (this keeps the pool and
3286 * thin devices' discard limits consistent).
3288 ti->discards_supported = true;
3292 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3293 calc_metadata_threshold(pt),
3294 metadata_low_callback,
3297 goto out_flags_changed;
3299 pt->callbacks.congested_fn = pool_is_congested;
3300 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3302 mutex_unlock(&dm_thin_pool_table.mutex);
3311 dm_put_device(ti, data_dev);
3313 dm_put_device(ti, metadata_dev);
3315 mutex_unlock(&dm_thin_pool_table.mutex);
3320 static int pool_map(struct dm_target *ti, struct bio *bio)
3323 struct pool_c *pt = ti->private;
3324 struct pool *pool = pt->pool;
3325 unsigned long flags;
3328 * As this is a singleton target, ti->begin is always zero.
3330 spin_lock_irqsave(&pool->lock, flags);
3331 bio_set_dev(bio, pt->data_dev->bdev);
3332 r = DM_MAPIO_REMAPPED;
3333 spin_unlock_irqrestore(&pool->lock, flags);
3338 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3341 struct pool_c *pt = ti->private;
3342 struct pool *pool = pt->pool;
3343 sector_t data_size = ti->len;
3344 dm_block_t sb_data_size;
3346 *need_commit = false;
3348 (void) sector_div(data_size, pool->sectors_per_block);
3350 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3352 DMERR("%s: failed to retrieve data device size",
3353 dm_device_name(pool->pool_md));
3357 if (data_size < sb_data_size) {
3358 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3359 dm_device_name(pool->pool_md),
3360 (unsigned long long)data_size, sb_data_size);
3363 } else if (data_size > sb_data_size) {
3364 if (dm_pool_metadata_needs_check(pool->pmd)) {
3365 DMERR("%s: unable to grow the data device until repaired.",
3366 dm_device_name(pool->pool_md));
3371 DMINFO("%s: growing the data device from %llu to %llu blocks",
3372 dm_device_name(pool->pool_md),
3373 sb_data_size, (unsigned long long)data_size);
3374 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3376 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3380 *need_commit = true;
3386 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3389 struct pool_c *pt = ti->private;
3390 struct pool *pool = pt->pool;
3391 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3393 *need_commit = false;
3395 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3397 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3399 DMERR("%s: failed to retrieve metadata device size",
3400 dm_device_name(pool->pool_md));
3404 if (metadata_dev_size < sb_metadata_dev_size) {
3405 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3406 dm_device_name(pool->pool_md),
3407 metadata_dev_size, sb_metadata_dev_size);
3410 } else if (metadata_dev_size > sb_metadata_dev_size) {
3411 if (dm_pool_metadata_needs_check(pool->pmd)) {
3412 DMERR("%s: unable to grow the metadata device until repaired.",
3413 dm_device_name(pool->pool_md));
3417 warn_if_metadata_device_too_big(pool->md_dev);
3418 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3419 dm_device_name(pool->pool_md),
3420 sb_metadata_dev_size, metadata_dev_size);
3421 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3423 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3427 *need_commit = true;
3434 * Retrieves the number of blocks of the data device from
3435 * the superblock and compares it to the actual device size,
3436 * thus resizing the data device in case it has grown.
3438 * This both copes with opening preallocated data devices in the ctr
3439 * being followed by a resume
3441 * calling the resume method individually after userspace has
3442 * grown the data device in reaction to a table event.
3444 static int pool_preresume(struct dm_target *ti)
3447 bool need_commit1, need_commit2;
3448 struct pool_c *pt = ti->private;
3449 struct pool *pool = pt->pool;
3452 * Take control of the pool object.
3454 r = bind_control_target(pool, ti);
3458 r = maybe_resize_data_dev(ti, &need_commit1);
3462 r = maybe_resize_metadata_dev(ti, &need_commit2);
3466 if (need_commit1 || need_commit2)
3467 (void) commit(pool);
3472 static void pool_suspend_active_thins(struct pool *pool)
3476 /* Suspend all active thin devices */
3477 tc = get_first_thin(pool);
3479 dm_internal_suspend_noflush(tc->thin_md);
3480 tc = get_next_thin(pool, tc);
3484 static void pool_resume_active_thins(struct pool *pool)
3488 /* Resume all active thin devices */
3489 tc = get_first_thin(pool);
3491 dm_internal_resume(tc->thin_md);
3492 tc = get_next_thin(pool, tc);
3496 static void pool_resume(struct dm_target *ti)
3498 struct pool_c *pt = ti->private;
3499 struct pool *pool = pt->pool;
3500 unsigned long flags;
3503 * Must requeue active_thins' bios and then resume
3504 * active_thins _before_ clearing 'suspend' flag.
3507 pool_resume_active_thins(pool);
3509 spin_lock_irqsave(&pool->lock, flags);
3510 pool->low_water_triggered = false;
3511 pool->suspended = false;
3512 spin_unlock_irqrestore(&pool->lock, flags);
3514 do_waker(&pool->waker.work);
3517 static void pool_presuspend(struct dm_target *ti)
3519 struct pool_c *pt = ti->private;
3520 struct pool *pool = pt->pool;
3521 unsigned long flags;
3523 spin_lock_irqsave(&pool->lock, flags);
3524 pool->suspended = true;
3525 spin_unlock_irqrestore(&pool->lock, flags);
3527 pool_suspend_active_thins(pool);
3530 static void pool_presuspend_undo(struct dm_target *ti)
3532 struct pool_c *pt = ti->private;
3533 struct pool *pool = pt->pool;
3534 unsigned long flags;
3536 pool_resume_active_thins(pool);
3538 spin_lock_irqsave(&pool->lock, flags);
3539 pool->suspended = false;
3540 spin_unlock_irqrestore(&pool->lock, flags);
3543 static void pool_postsuspend(struct dm_target *ti)
3545 struct pool_c *pt = ti->private;
3546 struct pool *pool = pt->pool;
3548 cancel_delayed_work_sync(&pool->waker);
3549 cancel_delayed_work_sync(&pool->no_space_timeout);
3550 flush_workqueue(pool->wq);
3551 (void) commit(pool);
3554 static int check_arg_count(unsigned argc, unsigned args_required)
3556 if (argc != args_required) {
3557 DMWARN("Message received with %u arguments instead of %u.",
3558 argc, args_required);
3565 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3567 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3568 *dev_id <= MAX_DEV_ID)
3572 DMWARN("Message received with invalid device id: %s", arg);
3577 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3582 r = check_arg_count(argc, 2);
3586 r = read_dev_id(argv[1], &dev_id, 1);
3590 r = dm_pool_create_thin(pool->pmd, dev_id);
3592 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3600 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3603 dm_thin_id origin_dev_id;
3606 r = check_arg_count(argc, 3);
3610 r = read_dev_id(argv[1], &dev_id, 1);
3614 r = read_dev_id(argv[2], &origin_dev_id, 1);
3618 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3620 DMWARN("Creation of new snapshot %s of device %s failed.",
3628 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3633 r = check_arg_count(argc, 2);
3637 r = read_dev_id(argv[1], &dev_id, 1);
3641 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3643 DMWARN("Deletion of thin device %s failed.", argv[1]);
3648 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3650 dm_thin_id old_id, new_id;
3653 r = check_arg_count(argc, 3);
3657 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3658 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3662 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3663 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3667 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3669 DMWARN("Failed to change transaction id from %s to %s.",
3677 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3681 r = check_arg_count(argc, 1);
3685 (void) commit(pool);
3687 r = dm_pool_reserve_metadata_snap(pool->pmd);
3689 DMWARN("reserve_metadata_snap message failed.");
3694 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3698 r = check_arg_count(argc, 1);
3702 r = dm_pool_release_metadata_snap(pool->pmd);
3704 DMWARN("release_metadata_snap message failed.");
3710 * Messages supported:
3711 * create_thin <dev_id>
3712 * create_snap <dev_id> <origin_id>
3714 * set_transaction_id <current_trans_id> <new_trans_id>
3715 * reserve_metadata_snap
3716 * release_metadata_snap
3718 static int pool_message(struct dm_target *ti, unsigned argc, char **argv,
3719 char *result, unsigned maxlen)
3722 struct pool_c *pt = ti->private;
3723 struct pool *pool = pt->pool;
3725 if (get_pool_mode(pool) >= PM_READ_ONLY) {
3726 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3727 dm_device_name(pool->pool_md));
3731 if (!strcasecmp(argv[0], "create_thin"))
3732 r = process_create_thin_mesg(argc, argv, pool);
3734 else if (!strcasecmp(argv[0], "create_snap"))
3735 r = process_create_snap_mesg(argc, argv, pool);
3737 else if (!strcasecmp(argv[0], "delete"))
3738 r = process_delete_mesg(argc, argv, pool);
3740 else if (!strcasecmp(argv[0], "set_transaction_id"))
3741 r = process_set_transaction_id_mesg(argc, argv, pool);
3743 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3744 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3746 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3747 r = process_release_metadata_snap_mesg(argc, argv, pool);
3750 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3753 (void) commit(pool);
3758 static void emit_flags(struct pool_features *pf, char *result,
3759 unsigned sz, unsigned maxlen)
3761 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3762 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3763 pf->error_if_no_space;
3764 DMEMIT("%u ", count);
3766 if (!pf->zero_new_blocks)
3767 DMEMIT("skip_block_zeroing ");
3769 if (!pf->discard_enabled)
3770 DMEMIT("ignore_discard ");
3772 if (!pf->discard_passdown)
3773 DMEMIT("no_discard_passdown ");
3775 if (pf->mode == PM_READ_ONLY)
3776 DMEMIT("read_only ");
3778 if (pf->error_if_no_space)
3779 DMEMIT("error_if_no_space ");
3784 * <transaction id> <used metadata sectors>/<total metadata sectors>
3785 * <used data sectors>/<total data sectors> <held metadata root>
3786 * <pool mode> <discard config> <no space config> <needs_check>
3788 static void pool_status(struct dm_target *ti, status_type_t type,
3789 unsigned status_flags, char *result, unsigned maxlen)
3793 uint64_t transaction_id;
3794 dm_block_t nr_free_blocks_data;
3795 dm_block_t nr_free_blocks_metadata;
3796 dm_block_t nr_blocks_data;
3797 dm_block_t nr_blocks_metadata;
3798 dm_block_t held_root;
3799 char buf[BDEVNAME_SIZE];
3800 char buf2[BDEVNAME_SIZE];
3801 struct pool_c *pt = ti->private;
3802 struct pool *pool = pt->pool;
3805 case STATUSTYPE_INFO:
3806 if (get_pool_mode(pool) == PM_FAIL) {
3811 /* Commit to ensure statistics aren't out-of-date */
3812 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3813 (void) commit(pool);
3815 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3817 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3818 dm_device_name(pool->pool_md), r);
3822 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3824 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3825 dm_device_name(pool->pool_md), r);
3829 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3831 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3832 dm_device_name(pool->pool_md), r);
3836 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3838 DMERR("%s: dm_pool_get_free_block_count returned %d",
3839 dm_device_name(pool->pool_md), r);
3843 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3845 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3846 dm_device_name(pool->pool_md), r);
3850 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3852 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3853 dm_device_name(pool->pool_md), r);
3857 DMEMIT("%llu %llu/%llu %llu/%llu ",
3858 (unsigned long long)transaction_id,
3859 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3860 (unsigned long long)nr_blocks_metadata,
3861 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3862 (unsigned long long)nr_blocks_data);
3865 DMEMIT("%llu ", held_root);
3869 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3870 DMEMIT("out_of_data_space ");
3871 else if (pool->pf.mode == PM_READ_ONLY)
3876 if (!pool->pf.discard_enabled)
3877 DMEMIT("ignore_discard ");
3878 else if (pool->pf.discard_passdown)
3879 DMEMIT("discard_passdown ");
3881 DMEMIT("no_discard_passdown ");
3883 if (pool->pf.error_if_no_space)
3884 DMEMIT("error_if_no_space ");
3886 DMEMIT("queue_if_no_space ");
3888 if (dm_pool_metadata_needs_check(pool->pmd))
3889 DMEMIT("needs_check ");
3895 case STATUSTYPE_TABLE:
3896 DMEMIT("%s %s %lu %llu ",
3897 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3898 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3899 (unsigned long)pool->sectors_per_block,
3900 (unsigned long long)pt->low_water_blocks);
3901 emit_flags(&pt->requested_pf, result, sz, maxlen);
3910 static int pool_iterate_devices(struct dm_target *ti,
3911 iterate_devices_callout_fn fn, void *data)
3913 struct pool_c *pt = ti->private;
3915 return fn(ti, pt->data_dev, 0, ti->len, data);
3918 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3920 struct pool_c *pt = ti->private;
3921 struct pool *pool = pt->pool;
3922 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3925 * If max_sectors is smaller than pool->sectors_per_block adjust it
3926 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3927 * This is especially beneficial when the pool's data device is a RAID
3928 * device that has a full stripe width that matches pool->sectors_per_block
3929 * -- because even though partial RAID stripe-sized IOs will be issued to a
3930 * single RAID stripe; when aggregated they will end on a full RAID stripe
3931 * boundary.. which avoids additional partial RAID stripe writes cascading
3933 if (limits->max_sectors < pool->sectors_per_block) {
3934 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3935 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3936 limits->max_sectors--;
3937 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3942 * If the system-determined stacked limits are compatible with the
3943 * pool's blocksize (io_opt is a factor) do not override them.
3945 if (io_opt_sectors < pool->sectors_per_block ||
3946 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3947 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3948 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3950 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3951 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3955 * pt->adjusted_pf is a staging area for the actual features to use.
3956 * They get transferred to the live pool in bind_control_target()
3957 * called from pool_preresume().
3959 if (!pt->adjusted_pf.discard_enabled) {
3961 * Must explicitly disallow stacking discard limits otherwise the
3962 * block layer will stack them if pool's data device has support.
3963 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3964 * user to see that, so make sure to set all discard limits to 0.
3966 limits->discard_granularity = 0;
3970 disable_passdown_if_not_supported(pt);
3973 * The pool uses the same discard limits as the underlying data
3974 * device. DM core has already set this up.
3978 static struct target_type pool_target = {
3979 .name = "thin-pool",
3980 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3981 DM_TARGET_IMMUTABLE,
3982 .version = {1, 19, 0},
3983 .module = THIS_MODULE,
3987 .presuspend = pool_presuspend,
3988 .presuspend_undo = pool_presuspend_undo,
3989 .postsuspend = pool_postsuspend,
3990 .preresume = pool_preresume,
3991 .resume = pool_resume,
3992 .message = pool_message,
3993 .status = pool_status,
3994 .iterate_devices = pool_iterate_devices,
3995 .io_hints = pool_io_hints,
3998 /*----------------------------------------------------------------
3999 * Thin target methods
4000 *--------------------------------------------------------------*/
4001 static void thin_get(struct thin_c *tc)
4003 atomic_inc(&tc->refcount);
4006 static void thin_put(struct thin_c *tc)
4008 if (atomic_dec_and_test(&tc->refcount))
4009 complete(&tc->can_destroy);
4012 static void thin_dtr(struct dm_target *ti)
4014 struct thin_c *tc = ti->private;
4015 unsigned long flags;
4017 spin_lock_irqsave(&tc->pool->lock, flags);
4018 list_del_rcu(&tc->list);
4019 spin_unlock_irqrestore(&tc->pool->lock, flags);
4023 wait_for_completion(&tc->can_destroy);
4025 mutex_lock(&dm_thin_pool_table.mutex);
4027 __pool_dec(tc->pool);
4028 dm_pool_close_thin_device(tc->td);
4029 dm_put_device(ti, tc->pool_dev);
4031 dm_put_device(ti, tc->origin_dev);
4034 mutex_unlock(&dm_thin_pool_table.mutex);
4038 * Thin target parameters:
4040 * <pool_dev> <dev_id> [origin_dev]
4042 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4043 * dev_id: the internal device identifier
4044 * origin_dev: a device external to the pool that should act as the origin
4046 * If the pool device has discards disabled, they get disabled for the thin
4049 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4053 struct dm_dev *pool_dev, *origin_dev;
4054 struct mapped_device *pool_md;
4055 unsigned long flags;
4057 mutex_lock(&dm_thin_pool_table.mutex);
4059 if (argc != 2 && argc != 3) {
4060 ti->error = "Invalid argument count";
4065 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4067 ti->error = "Out of memory";
4071 tc->thin_md = dm_table_get_md(ti->table);
4072 spin_lock_init(&tc->lock);
4073 INIT_LIST_HEAD(&tc->deferred_cells);
4074 bio_list_init(&tc->deferred_bio_list);
4075 bio_list_init(&tc->retry_on_resume_list);
4076 tc->sort_bio_list = RB_ROOT;
4079 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4081 ti->error = "Error opening origin device";
4082 goto bad_origin_dev;
4084 tc->origin_dev = origin_dev;
4087 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4089 ti->error = "Error opening pool device";
4092 tc->pool_dev = pool_dev;
4094 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4095 ti->error = "Invalid device id";
4100 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4102 ti->error = "Couldn't get pool mapped device";
4107 tc->pool = __pool_table_lookup(pool_md);
4109 ti->error = "Couldn't find pool object";
4111 goto bad_pool_lookup;
4113 __pool_inc(tc->pool);
4115 if (get_pool_mode(tc->pool) == PM_FAIL) {
4116 ti->error = "Couldn't open thin device, Pool is in fail mode";
4121 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4123 ti->error = "Couldn't open thin internal device";
4127 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4131 ti->num_flush_bios = 1;
4132 ti->flush_supported = true;
4133 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4135 /* In case the pool supports discards, pass them on. */
4136 if (tc->pool->pf.discard_enabled) {
4137 ti->discards_supported = true;
4138 ti->num_discard_bios = 1;
4139 ti->split_discard_bios = false;
4142 mutex_unlock(&dm_thin_pool_table.mutex);
4144 spin_lock_irqsave(&tc->pool->lock, flags);
4145 if (tc->pool->suspended) {
4146 spin_unlock_irqrestore(&tc->pool->lock, flags);
4147 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4148 ti->error = "Unable to activate thin device while pool is suspended";
4152 atomic_set(&tc->refcount, 1);
4153 init_completion(&tc->can_destroy);
4154 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4155 spin_unlock_irqrestore(&tc->pool->lock, flags);
4157 * This synchronize_rcu() call is needed here otherwise we risk a
4158 * wake_worker() call finding no bios to process (because the newly
4159 * added tc isn't yet visible). So this reduces latency since we
4160 * aren't then dependent on the periodic commit to wake_worker().
4169 dm_pool_close_thin_device(tc->td);
4171 __pool_dec(tc->pool);
4175 dm_put_device(ti, tc->pool_dev);
4178 dm_put_device(ti, tc->origin_dev);
4182 mutex_unlock(&dm_thin_pool_table.mutex);
4187 static int thin_map(struct dm_target *ti, struct bio *bio)
4189 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4191 return thin_bio_map(ti, bio);
4194 static int thin_endio(struct dm_target *ti, struct bio *bio,
4197 unsigned long flags;
4198 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4199 struct list_head work;
4200 struct dm_thin_new_mapping *m, *tmp;
4201 struct pool *pool = h->tc->pool;
4203 if (h->shared_read_entry) {
4204 INIT_LIST_HEAD(&work);
4205 dm_deferred_entry_dec(h->shared_read_entry, &work);
4207 spin_lock_irqsave(&pool->lock, flags);
4208 list_for_each_entry_safe(m, tmp, &work, list) {
4210 __complete_mapping_preparation(m);
4212 spin_unlock_irqrestore(&pool->lock, flags);
4215 if (h->all_io_entry) {
4216 INIT_LIST_HEAD(&work);
4217 dm_deferred_entry_dec(h->all_io_entry, &work);
4218 if (!list_empty(&work)) {
4219 spin_lock_irqsave(&pool->lock, flags);
4220 list_for_each_entry_safe(m, tmp, &work, list)
4221 list_add_tail(&m->list, &pool->prepared_discards);
4222 spin_unlock_irqrestore(&pool->lock, flags);
4228 cell_defer_no_holder(h->tc, h->cell);
4230 return DM_ENDIO_DONE;
4233 static void thin_presuspend(struct dm_target *ti)
4235 struct thin_c *tc = ti->private;
4237 if (dm_noflush_suspending(ti))
4238 noflush_work(tc, do_noflush_start);
4241 static void thin_postsuspend(struct dm_target *ti)
4243 struct thin_c *tc = ti->private;
4246 * The dm_noflush_suspending flag has been cleared by now, so
4247 * unfortunately we must always run this.
4249 noflush_work(tc, do_noflush_stop);
4252 static int thin_preresume(struct dm_target *ti)
4254 struct thin_c *tc = ti->private;
4257 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4263 * <nr mapped sectors> <highest mapped sector>
4265 static void thin_status(struct dm_target *ti, status_type_t type,
4266 unsigned status_flags, char *result, unsigned maxlen)
4270 dm_block_t mapped, highest;
4271 char buf[BDEVNAME_SIZE];
4272 struct thin_c *tc = ti->private;
4274 if (get_pool_mode(tc->pool) == PM_FAIL) {
4283 case STATUSTYPE_INFO:
4284 r = dm_thin_get_mapped_count(tc->td, &mapped);
4286 DMERR("dm_thin_get_mapped_count returned %d", r);
4290 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4292 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4296 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4298 DMEMIT("%llu", ((highest + 1) *
4299 tc->pool->sectors_per_block) - 1);
4304 case STATUSTYPE_TABLE:
4306 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4307 (unsigned long) tc->dev_id);
4309 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4320 static int thin_iterate_devices(struct dm_target *ti,
4321 iterate_devices_callout_fn fn, void *data)
4324 struct thin_c *tc = ti->private;
4325 struct pool *pool = tc->pool;
4328 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4329 * we follow a more convoluted path through to the pool's target.
4332 return 0; /* nothing is bound */
4334 blocks = pool->ti->len;
4335 (void) sector_div(blocks, pool->sectors_per_block);
4337 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4342 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4344 struct thin_c *tc = ti->private;
4345 struct pool *pool = tc->pool;
4347 if (!pool->pf.discard_enabled)
4350 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4351 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4354 static struct target_type thin_target = {
4356 .version = {1, 19, 0},
4357 .module = THIS_MODULE,
4361 .end_io = thin_endio,
4362 .preresume = thin_preresume,
4363 .presuspend = thin_presuspend,
4364 .postsuspend = thin_postsuspend,
4365 .status = thin_status,
4366 .iterate_devices = thin_iterate_devices,
4367 .io_hints = thin_io_hints,
4370 /*----------------------------------------------------------------*/
4372 static int __init dm_thin_init(void)
4378 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4379 if (!_new_mapping_cache)
4382 r = dm_register_target(&thin_target);
4384 goto bad_new_mapping_cache;
4386 r = dm_register_target(&pool_target);
4388 goto bad_thin_target;
4393 dm_unregister_target(&thin_target);
4394 bad_new_mapping_cache:
4395 kmem_cache_destroy(_new_mapping_cache);
4400 static void dm_thin_exit(void)
4402 dm_unregister_target(&thin_target);
4403 dm_unregister_target(&pool_target);
4405 kmem_cache_destroy(_new_mapping_cache);
4410 module_init(dm_thin_init);
4411 module_exit(dm_thin_exit);
4413 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4414 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4416 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4417 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4418 MODULE_LICENSE("GPL");