perf bench futex: Cache align the worker struct
[linux/fpc-iii.git] / drivers / md / dm-thin.c
blobd1c05c12a9db3851f0f8dfafee26b3853e25a666
1 /*
2 * Copyright (C) 2011-2012 Red Hat UK.
4 * This file is released under the GPL.
5 */
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison.h"
9 #include "dm.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"
28 * Tunable constants
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
61 * same data blocks.
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
67 * steps are:
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
99 * the commit.
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 /*----------------------------------------------------------------*/
113 * Key building.
115 enum lock_space {
116 VIRTUAL,
117 PHYSICAL
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;
126 key->block_end = e;
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)
145 struct throttle {
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;
174 up_write(&t->lock);
178 static void throttle_lock(struct throttle *t)
180 down_read(&t->lock);
183 static void throttle_unlock(struct throttle *t)
185 up_read(&t->lock);
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
193 * devices.
195 struct dm_thin_new_mapping;
198 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
200 enum pool_mode {
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 {
208 enum pool_mode mode;
210 bool zero_new_blocks:1;
211 bool discard_enabled:1;
212 bool discard_passdown:1;
213 bool error_if_no_space:1;
216 struct thin_c;
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
223 struct pool {
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 */
237 bool suspended:1;
238 bool out_of_data_space:1;
240 struct dm_bio_prison *prison;
241 struct dm_kcopyd_client *copier;
243 struct workqueue_struct *wq;
244 struct throttle throttle;
245 struct work_struct worker;
246 struct delayed_work waker;
247 struct delayed_work no_space_timeout;
249 unsigned long last_commit_jiffies;
250 unsigned ref_count;
252 spinlock_t lock;
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;
263 mempool_t *mapping_pool;
265 process_bio_fn process_bio;
266 process_bio_fn process_discard;
268 process_cell_fn process_cell;
269 process_cell_fn process_discard_cell;
271 process_mapping_fn process_prepared_mapping;
272 process_mapping_fn process_prepared_discard;
273 process_mapping_fn process_prepared_discard_pt2;
275 struct dm_bio_prison_cell **cell_sort_array;
278 static enum pool_mode get_pool_mode(struct pool *pool);
279 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
282 * Target context for a pool.
284 struct pool_c {
285 struct dm_target *ti;
286 struct pool *pool;
287 struct dm_dev *data_dev;
288 struct dm_dev *metadata_dev;
289 struct dm_target_callbacks callbacks;
291 dm_block_t low_water_blocks;
292 struct pool_features requested_pf; /* Features requested during table load */
293 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
297 * Target context for a thin.
299 struct thin_c {
300 struct list_head list;
301 struct dm_dev *pool_dev;
302 struct dm_dev *origin_dev;
303 sector_t origin_size;
304 dm_thin_id dev_id;
306 struct pool *pool;
307 struct dm_thin_device *td;
308 struct mapped_device *thin_md;
310 bool requeue_mode:1;
311 spinlock_t lock;
312 struct list_head deferred_cells;
313 struct bio_list deferred_bio_list;
314 struct bio_list retry_on_resume_list;
315 struct rb_root sort_bio_list; /* sorted list of deferred bios */
318 * Ensures the thin is not destroyed until the worker has finished
319 * iterating the active_thins list.
321 atomic_t refcount;
322 struct completion can_destroy;
325 /*----------------------------------------------------------------*/
327 static bool block_size_is_power_of_two(struct pool *pool)
329 return pool->sectors_per_block_shift >= 0;
332 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
334 return block_size_is_power_of_two(pool) ?
335 (b << pool->sectors_per_block_shift) :
336 (b * pool->sectors_per_block);
339 /*----------------------------------------------------------------*/
341 struct discard_op {
342 struct thin_c *tc;
343 struct blk_plug plug;
344 struct bio *parent_bio;
345 struct bio *bio;
348 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
350 BUG_ON(!parent);
352 op->tc = tc;
353 blk_start_plug(&op->plug);
354 op->parent_bio = parent;
355 op->bio = NULL;
358 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
360 struct thin_c *tc = op->tc;
361 sector_t s = block_to_sectors(tc->pool, data_b);
362 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
364 return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
365 GFP_NOWAIT, 0, &op->bio);
368 static void end_discard(struct discard_op *op, int r)
370 if (op->bio) {
372 * Even if one of the calls to issue_discard failed, we
373 * need to wait for the chain to complete.
375 bio_chain(op->bio, op->parent_bio);
376 bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
377 submit_bio(op->bio);
380 blk_finish_plug(&op->plug);
383 * Even if r is set, there could be sub discards in flight that we
384 * need to wait for.
386 if (r && !op->parent_bio->bi_error)
387 op->parent_bio->bi_error = r;
388 bio_endio(op->parent_bio);
391 /*----------------------------------------------------------------*/
394 * wake_worker() is used when new work is queued and when pool_resume is
395 * ready to continue deferred IO processing.
397 static void wake_worker(struct pool *pool)
399 queue_work(pool->wq, &pool->worker);
402 /*----------------------------------------------------------------*/
404 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
405 struct dm_bio_prison_cell **cell_result)
407 int r;
408 struct dm_bio_prison_cell *cell_prealloc;
411 * Allocate a cell from the prison's mempool.
412 * This might block but it can't fail.
414 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
416 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
417 if (r)
419 * We reused an old cell; we can get rid of
420 * the new one.
422 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
424 return r;
427 static void cell_release(struct pool *pool,
428 struct dm_bio_prison_cell *cell,
429 struct bio_list *bios)
431 dm_cell_release(pool->prison, cell, bios);
432 dm_bio_prison_free_cell(pool->prison, cell);
435 static void cell_visit_release(struct pool *pool,
436 void (*fn)(void *, struct dm_bio_prison_cell *),
437 void *context,
438 struct dm_bio_prison_cell *cell)
440 dm_cell_visit_release(pool->prison, fn, context, cell);
441 dm_bio_prison_free_cell(pool->prison, cell);
444 static void cell_release_no_holder(struct pool *pool,
445 struct dm_bio_prison_cell *cell,
446 struct bio_list *bios)
448 dm_cell_release_no_holder(pool->prison, cell, bios);
449 dm_bio_prison_free_cell(pool->prison, cell);
452 static void cell_error_with_code(struct pool *pool,
453 struct dm_bio_prison_cell *cell, int error_code)
455 dm_cell_error(pool->prison, cell, error_code);
456 dm_bio_prison_free_cell(pool->prison, cell);
459 static int get_pool_io_error_code(struct pool *pool)
461 return pool->out_of_data_space ? -ENOSPC : -EIO;
464 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
466 int error = get_pool_io_error_code(pool);
468 cell_error_with_code(pool, cell, error);
471 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
473 cell_error_with_code(pool, cell, 0);
476 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
478 cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
481 /*----------------------------------------------------------------*/
484 * A global list of pools that uses a struct mapped_device as a key.
486 static struct dm_thin_pool_table {
487 struct mutex mutex;
488 struct list_head pools;
489 } dm_thin_pool_table;
491 static void pool_table_init(void)
493 mutex_init(&dm_thin_pool_table.mutex);
494 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
497 static void __pool_table_insert(struct pool *pool)
499 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
500 list_add(&pool->list, &dm_thin_pool_table.pools);
503 static void __pool_table_remove(struct pool *pool)
505 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
506 list_del(&pool->list);
509 static struct pool *__pool_table_lookup(struct mapped_device *md)
511 struct pool *pool = NULL, *tmp;
513 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
515 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
516 if (tmp->pool_md == md) {
517 pool = tmp;
518 break;
522 return pool;
525 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
527 struct pool *pool = NULL, *tmp;
529 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
531 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
532 if (tmp->md_dev == md_dev) {
533 pool = tmp;
534 break;
538 return pool;
541 /*----------------------------------------------------------------*/
543 struct dm_thin_endio_hook {
544 struct thin_c *tc;
545 struct dm_deferred_entry *shared_read_entry;
546 struct dm_deferred_entry *all_io_entry;
547 struct dm_thin_new_mapping *overwrite_mapping;
548 struct rb_node rb_node;
549 struct dm_bio_prison_cell *cell;
552 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
554 bio_list_merge(bios, master);
555 bio_list_init(master);
558 static void error_bio_list(struct bio_list *bios, int error)
560 struct bio *bio;
562 while ((bio = bio_list_pop(bios))) {
563 bio->bi_error = error;
564 bio_endio(bio);
568 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
570 struct bio_list bios;
571 unsigned long flags;
573 bio_list_init(&bios);
575 spin_lock_irqsave(&tc->lock, flags);
576 __merge_bio_list(&bios, master);
577 spin_unlock_irqrestore(&tc->lock, flags);
579 error_bio_list(&bios, error);
582 static void requeue_deferred_cells(struct thin_c *tc)
584 struct pool *pool = tc->pool;
585 unsigned long flags;
586 struct list_head cells;
587 struct dm_bio_prison_cell *cell, *tmp;
589 INIT_LIST_HEAD(&cells);
591 spin_lock_irqsave(&tc->lock, flags);
592 list_splice_init(&tc->deferred_cells, &cells);
593 spin_unlock_irqrestore(&tc->lock, flags);
595 list_for_each_entry_safe(cell, tmp, &cells, user_list)
596 cell_requeue(pool, cell);
599 static void requeue_io(struct thin_c *tc)
601 struct bio_list bios;
602 unsigned long flags;
604 bio_list_init(&bios);
606 spin_lock_irqsave(&tc->lock, flags);
607 __merge_bio_list(&bios, &tc->deferred_bio_list);
608 __merge_bio_list(&bios, &tc->retry_on_resume_list);
609 spin_unlock_irqrestore(&tc->lock, flags);
611 error_bio_list(&bios, DM_ENDIO_REQUEUE);
612 requeue_deferred_cells(tc);
615 static void error_retry_list_with_code(struct pool *pool, int error)
617 struct thin_c *tc;
619 rcu_read_lock();
620 list_for_each_entry_rcu(tc, &pool->active_thins, list)
621 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
622 rcu_read_unlock();
625 static void error_retry_list(struct pool *pool)
627 int error = get_pool_io_error_code(pool);
629 error_retry_list_with_code(pool, error);
633 * This section of code contains the logic for processing a thin device's IO.
634 * Much of the code depends on pool object resources (lists, workqueues, etc)
635 * but most is exclusively called from the thin target rather than the thin-pool
636 * target.
639 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
641 struct pool *pool = tc->pool;
642 sector_t block_nr = bio->bi_iter.bi_sector;
644 if (block_size_is_power_of_two(pool))
645 block_nr >>= pool->sectors_per_block_shift;
646 else
647 (void) sector_div(block_nr, pool->sectors_per_block);
649 return block_nr;
653 * Returns the _complete_ blocks that this bio covers.
655 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
656 dm_block_t *begin, dm_block_t *end)
658 struct pool *pool = tc->pool;
659 sector_t b = bio->bi_iter.bi_sector;
660 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
662 b += pool->sectors_per_block - 1ull; /* so we round up */
664 if (block_size_is_power_of_two(pool)) {
665 b >>= pool->sectors_per_block_shift;
666 e >>= pool->sectors_per_block_shift;
667 } else {
668 (void) sector_div(b, pool->sectors_per_block);
669 (void) sector_div(e, pool->sectors_per_block);
672 if (e < b)
673 /* Can happen if the bio is within a single block. */
674 e = b;
676 *begin = b;
677 *end = e;
680 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
682 struct pool *pool = tc->pool;
683 sector_t bi_sector = bio->bi_iter.bi_sector;
685 bio->bi_bdev = tc->pool_dev->bdev;
686 if (block_size_is_power_of_two(pool))
687 bio->bi_iter.bi_sector =
688 (block << pool->sectors_per_block_shift) |
689 (bi_sector & (pool->sectors_per_block - 1));
690 else
691 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
692 sector_div(bi_sector, pool->sectors_per_block);
695 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
697 bio->bi_bdev = tc->origin_dev->bdev;
700 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
702 return (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA)) &&
703 dm_thin_changed_this_transaction(tc->td);
706 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
708 struct dm_thin_endio_hook *h;
710 if (bio_op(bio) == REQ_OP_DISCARD)
711 return;
713 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
714 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
717 static void issue(struct thin_c *tc, struct bio *bio)
719 struct pool *pool = tc->pool;
720 unsigned long flags;
722 if (!bio_triggers_commit(tc, bio)) {
723 generic_make_request(bio);
724 return;
728 * Complete bio with an error if earlier I/O caused changes to
729 * the metadata that can't be committed e.g, due to I/O errors
730 * on the metadata device.
732 if (dm_thin_aborted_changes(tc->td)) {
733 bio_io_error(bio);
734 return;
738 * Batch together any bios that trigger commits and then issue a
739 * single commit for them in process_deferred_bios().
741 spin_lock_irqsave(&pool->lock, flags);
742 bio_list_add(&pool->deferred_flush_bios, bio);
743 spin_unlock_irqrestore(&pool->lock, flags);
746 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
748 remap_to_origin(tc, bio);
749 issue(tc, bio);
752 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
753 dm_block_t block)
755 remap(tc, bio, block);
756 issue(tc, bio);
759 /*----------------------------------------------------------------*/
762 * Bio endio functions.
764 struct dm_thin_new_mapping {
765 struct list_head list;
767 bool pass_discard:1;
768 bool maybe_shared:1;
771 * Track quiescing, copying and zeroing preparation actions. When this
772 * counter hits zero the block is prepared and can be inserted into the
773 * btree.
775 atomic_t prepare_actions;
777 int err;
778 struct thin_c *tc;
779 dm_block_t virt_begin, virt_end;
780 dm_block_t data_block;
781 struct dm_bio_prison_cell *cell;
784 * If the bio covers the whole area of a block then we can avoid
785 * zeroing or copying. Instead this bio is hooked. The bio will
786 * still be in the cell, so care has to be taken to avoid issuing
787 * the bio twice.
789 struct bio *bio;
790 bio_end_io_t *saved_bi_end_io;
793 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
795 struct pool *pool = m->tc->pool;
797 if (atomic_dec_and_test(&m->prepare_actions)) {
798 list_add_tail(&m->list, &pool->prepared_mappings);
799 wake_worker(pool);
803 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
805 unsigned long flags;
806 struct pool *pool = m->tc->pool;
808 spin_lock_irqsave(&pool->lock, flags);
809 __complete_mapping_preparation(m);
810 spin_unlock_irqrestore(&pool->lock, flags);
813 static void copy_complete(int read_err, unsigned long write_err, void *context)
815 struct dm_thin_new_mapping *m = context;
817 m->err = read_err || write_err ? -EIO : 0;
818 complete_mapping_preparation(m);
821 static void overwrite_endio(struct bio *bio)
823 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
824 struct dm_thin_new_mapping *m = h->overwrite_mapping;
826 bio->bi_end_io = m->saved_bi_end_io;
828 m->err = bio->bi_error;
829 complete_mapping_preparation(m);
832 /*----------------------------------------------------------------*/
835 * Workqueue.
839 * Prepared mapping jobs.
843 * This sends the bios in the cell, except the original holder, back
844 * to the deferred_bios list.
846 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
848 struct pool *pool = tc->pool;
849 unsigned long flags;
851 spin_lock_irqsave(&tc->lock, flags);
852 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
853 spin_unlock_irqrestore(&tc->lock, flags);
855 wake_worker(pool);
858 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
860 struct remap_info {
861 struct thin_c *tc;
862 struct bio_list defer_bios;
863 struct bio_list issue_bios;
866 static void __inc_remap_and_issue_cell(void *context,
867 struct dm_bio_prison_cell *cell)
869 struct remap_info *info = context;
870 struct bio *bio;
872 while ((bio = bio_list_pop(&cell->bios))) {
873 if (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
874 bio_op(bio) == REQ_OP_DISCARD)
875 bio_list_add(&info->defer_bios, bio);
876 else {
877 inc_all_io_entry(info->tc->pool, bio);
880 * We can't issue the bios with the bio prison lock
881 * held, so we add them to a list to issue on
882 * return from this function.
884 bio_list_add(&info->issue_bios, bio);
889 static void inc_remap_and_issue_cell(struct thin_c *tc,
890 struct dm_bio_prison_cell *cell,
891 dm_block_t block)
893 struct bio *bio;
894 struct remap_info info;
896 info.tc = tc;
897 bio_list_init(&info.defer_bios);
898 bio_list_init(&info.issue_bios);
901 * We have to be careful to inc any bios we're about to issue
902 * before the cell is released, and avoid a race with new bios
903 * being added to the cell.
905 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
906 &info, cell);
908 while ((bio = bio_list_pop(&info.defer_bios)))
909 thin_defer_bio(tc, bio);
911 while ((bio = bio_list_pop(&info.issue_bios)))
912 remap_and_issue(info.tc, bio, block);
915 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
917 cell_error(m->tc->pool, m->cell);
918 list_del(&m->list);
919 mempool_free(m, m->tc->pool->mapping_pool);
922 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
924 struct thin_c *tc = m->tc;
925 struct pool *pool = tc->pool;
926 struct bio *bio = m->bio;
927 int r;
929 if (m->err) {
930 cell_error(pool, m->cell);
931 goto out;
935 * Commit the prepared block into the mapping btree.
936 * Any I/O for this block arriving after this point will get
937 * remapped to it directly.
939 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
940 if (r) {
941 metadata_operation_failed(pool, "dm_thin_insert_block", r);
942 cell_error(pool, m->cell);
943 goto out;
947 * Release any bios held while the block was being provisioned.
948 * If we are processing a write bio that completely covers the block,
949 * we already processed it so can ignore it now when processing
950 * the bios in the cell.
952 if (bio) {
953 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
954 bio_endio(bio);
955 } else {
956 inc_all_io_entry(tc->pool, m->cell->holder);
957 remap_and_issue(tc, m->cell->holder, m->data_block);
958 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
961 out:
962 list_del(&m->list);
963 mempool_free(m, pool->mapping_pool);
966 /*----------------------------------------------------------------*/
968 static void free_discard_mapping(struct dm_thin_new_mapping *m)
970 struct thin_c *tc = m->tc;
971 if (m->cell)
972 cell_defer_no_holder(tc, m->cell);
973 mempool_free(m, tc->pool->mapping_pool);
976 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
978 bio_io_error(m->bio);
979 free_discard_mapping(m);
982 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
984 bio_endio(m->bio);
985 free_discard_mapping(m);
988 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
990 int r;
991 struct thin_c *tc = m->tc;
993 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
994 if (r) {
995 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
996 bio_io_error(m->bio);
997 } else
998 bio_endio(m->bio);
1000 cell_defer_no_holder(tc, m->cell);
1001 mempool_free(m, tc->pool->mapping_pool);
1004 /*----------------------------------------------------------------*/
1006 static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1007 struct bio *discard_parent)
1010 * We've already unmapped this range of blocks, but before we
1011 * passdown we have to check that these blocks are now unused.
1013 int r = 0;
1014 bool used = true;
1015 struct thin_c *tc = m->tc;
1016 struct pool *pool = tc->pool;
1017 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1018 struct discard_op op;
1020 begin_discard(&op, tc, discard_parent);
1021 while (b != end) {
1022 /* find start of unmapped run */
1023 for (; b < end; b++) {
1024 r = dm_pool_block_is_used(pool->pmd, b, &used);
1025 if (r)
1026 goto out;
1028 if (!used)
1029 break;
1032 if (b == end)
1033 break;
1035 /* find end of run */
1036 for (e = b + 1; e != end; e++) {
1037 r = dm_pool_block_is_used(pool->pmd, e, &used);
1038 if (r)
1039 goto out;
1041 if (used)
1042 break;
1045 r = issue_discard(&op, b, e);
1046 if (r)
1047 goto out;
1049 b = e;
1051 out:
1052 end_discard(&op, r);
1055 static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1057 unsigned long flags;
1058 struct pool *pool = m->tc->pool;
1060 spin_lock_irqsave(&pool->lock, flags);
1061 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1062 spin_unlock_irqrestore(&pool->lock, flags);
1063 wake_worker(pool);
1066 static void passdown_endio(struct bio *bio)
1069 * It doesn't matter if the passdown discard failed, we still want
1070 * to unmap (we ignore err).
1072 queue_passdown_pt2(bio->bi_private);
1075 static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1077 int r;
1078 struct thin_c *tc = m->tc;
1079 struct pool *pool = tc->pool;
1080 struct bio *discard_parent;
1081 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1084 * Only this thread allocates blocks, so we can be sure that the
1085 * newly unmapped blocks will not be allocated before the end of
1086 * the function.
1088 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1089 if (r) {
1090 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1091 bio_io_error(m->bio);
1092 cell_defer_no_holder(tc, m->cell);
1093 mempool_free(m, pool->mapping_pool);
1094 return;
1097 discard_parent = bio_alloc(GFP_NOIO, 1);
1098 if (!discard_parent) {
1099 DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1100 dm_device_name(tc->pool->pool_md));
1101 queue_passdown_pt2(m);
1103 } else {
1104 discard_parent->bi_end_io = passdown_endio;
1105 discard_parent->bi_private = m;
1107 if (m->maybe_shared)
1108 passdown_double_checking_shared_status(m, discard_parent);
1109 else {
1110 struct discard_op op;
1112 begin_discard(&op, tc, discard_parent);
1113 r = issue_discard(&op, m->data_block, data_end);
1114 end_discard(&op, r);
1119 * Increment the unmapped blocks. This prevents a race between the
1120 * passdown io and reallocation of freed blocks.
1122 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1123 if (r) {
1124 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1125 bio_io_error(m->bio);
1126 cell_defer_no_holder(tc, m->cell);
1127 mempool_free(m, pool->mapping_pool);
1128 return;
1132 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1134 int r;
1135 struct thin_c *tc = m->tc;
1136 struct pool *pool = tc->pool;
1139 * The passdown has completed, so now we can decrement all those
1140 * unmapped blocks.
1142 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1143 m->data_block + (m->virt_end - m->virt_begin));
1144 if (r) {
1145 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1146 bio_io_error(m->bio);
1147 } else
1148 bio_endio(m->bio);
1150 cell_defer_no_holder(tc, m->cell);
1151 mempool_free(m, pool->mapping_pool);
1154 static void process_prepared(struct pool *pool, struct list_head *head,
1155 process_mapping_fn *fn)
1157 unsigned long flags;
1158 struct list_head maps;
1159 struct dm_thin_new_mapping *m, *tmp;
1161 INIT_LIST_HEAD(&maps);
1162 spin_lock_irqsave(&pool->lock, flags);
1163 list_splice_init(head, &maps);
1164 spin_unlock_irqrestore(&pool->lock, flags);
1166 list_for_each_entry_safe(m, tmp, &maps, list)
1167 (*fn)(m);
1171 * Deferred bio jobs.
1173 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1175 return bio->bi_iter.bi_size ==
1176 (pool->sectors_per_block << SECTOR_SHIFT);
1179 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1181 return (bio_data_dir(bio) == WRITE) &&
1182 io_overlaps_block(pool, bio);
1185 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1186 bio_end_io_t *fn)
1188 *save = bio->bi_end_io;
1189 bio->bi_end_io = fn;
1192 static int ensure_next_mapping(struct pool *pool)
1194 if (pool->next_mapping)
1195 return 0;
1197 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
1199 return pool->next_mapping ? 0 : -ENOMEM;
1202 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1204 struct dm_thin_new_mapping *m = pool->next_mapping;
1206 BUG_ON(!pool->next_mapping);
1208 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1209 INIT_LIST_HEAD(&m->list);
1210 m->bio = NULL;
1212 pool->next_mapping = NULL;
1214 return m;
1217 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1218 sector_t begin, sector_t end)
1220 int r;
1221 struct dm_io_region to;
1223 to.bdev = tc->pool_dev->bdev;
1224 to.sector = begin;
1225 to.count = end - begin;
1227 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1228 if (r < 0) {
1229 DMERR_LIMIT("dm_kcopyd_zero() failed");
1230 copy_complete(1, 1, m);
1234 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1235 dm_block_t data_begin,
1236 struct dm_thin_new_mapping *m)
1238 struct pool *pool = tc->pool;
1239 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1241 h->overwrite_mapping = m;
1242 m->bio = bio;
1243 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1244 inc_all_io_entry(pool, bio);
1245 remap_and_issue(tc, bio, data_begin);
1249 * A partial copy also needs to zero the uncopied region.
1251 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1252 struct dm_dev *origin, dm_block_t data_origin,
1253 dm_block_t data_dest,
1254 struct dm_bio_prison_cell *cell, struct bio *bio,
1255 sector_t len)
1257 int r;
1258 struct pool *pool = tc->pool;
1259 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1261 m->tc = tc;
1262 m->virt_begin = virt_block;
1263 m->virt_end = virt_block + 1u;
1264 m->data_block = data_dest;
1265 m->cell = cell;
1268 * quiesce action + copy action + an extra reference held for the
1269 * duration of this function (we may need to inc later for a
1270 * partial zero).
1272 atomic_set(&m->prepare_actions, 3);
1274 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1275 complete_mapping_preparation(m); /* already quiesced */
1278 * IO to pool_dev remaps to the pool target's data_dev.
1280 * If the whole block of data is being overwritten, we can issue the
1281 * bio immediately. Otherwise we use kcopyd to clone the data first.
1283 if (io_overwrites_block(pool, bio))
1284 remap_and_issue_overwrite(tc, bio, data_dest, m);
1285 else {
1286 struct dm_io_region from, to;
1288 from.bdev = origin->bdev;
1289 from.sector = data_origin * pool->sectors_per_block;
1290 from.count = len;
1292 to.bdev = tc->pool_dev->bdev;
1293 to.sector = data_dest * pool->sectors_per_block;
1294 to.count = len;
1296 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1297 0, copy_complete, m);
1298 if (r < 0) {
1299 DMERR_LIMIT("dm_kcopyd_copy() failed");
1300 copy_complete(1, 1, m);
1303 * We allow the zero to be issued, to simplify the
1304 * error path. Otherwise we'd need to start
1305 * worrying about decrementing the prepare_actions
1306 * counter.
1311 * Do we need to zero a tail region?
1313 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1314 atomic_inc(&m->prepare_actions);
1315 ll_zero(tc, m,
1316 data_dest * pool->sectors_per_block + len,
1317 (data_dest + 1) * pool->sectors_per_block);
1321 complete_mapping_preparation(m); /* drop our ref */
1324 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1325 dm_block_t data_origin, dm_block_t data_dest,
1326 struct dm_bio_prison_cell *cell, struct bio *bio)
1328 schedule_copy(tc, virt_block, tc->pool_dev,
1329 data_origin, data_dest, cell, bio,
1330 tc->pool->sectors_per_block);
1333 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1334 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1335 struct bio *bio)
1337 struct pool *pool = tc->pool;
1338 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1340 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1341 m->tc = tc;
1342 m->virt_begin = virt_block;
1343 m->virt_end = virt_block + 1u;
1344 m->data_block = data_block;
1345 m->cell = cell;
1348 * If the whole block of data is being overwritten or we are not
1349 * zeroing pre-existing data, we can issue the bio immediately.
1350 * Otherwise we use kcopyd to zero the data first.
1352 if (pool->pf.zero_new_blocks) {
1353 if (io_overwrites_block(pool, bio))
1354 remap_and_issue_overwrite(tc, bio, data_block, m);
1355 else
1356 ll_zero(tc, m, data_block * pool->sectors_per_block,
1357 (data_block + 1) * pool->sectors_per_block);
1358 } else
1359 process_prepared_mapping(m);
1362 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1363 dm_block_t data_dest,
1364 struct dm_bio_prison_cell *cell, struct bio *bio)
1366 struct pool *pool = tc->pool;
1367 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1368 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1370 if (virt_block_end <= tc->origin_size)
1371 schedule_copy(tc, virt_block, tc->origin_dev,
1372 virt_block, data_dest, cell, bio,
1373 pool->sectors_per_block);
1375 else if (virt_block_begin < tc->origin_size)
1376 schedule_copy(tc, virt_block, tc->origin_dev,
1377 virt_block, data_dest, cell, bio,
1378 tc->origin_size - virt_block_begin);
1380 else
1381 schedule_zero(tc, virt_block, data_dest, cell, bio);
1384 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1386 static void check_for_space(struct pool *pool)
1388 int r;
1389 dm_block_t nr_free;
1391 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1392 return;
1394 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1395 if (r)
1396 return;
1398 if (nr_free)
1399 set_pool_mode(pool, PM_WRITE);
1403 * A non-zero return indicates read_only or fail_io mode.
1404 * Many callers don't care about the return value.
1406 static int commit(struct pool *pool)
1408 int r;
1410 if (get_pool_mode(pool) >= PM_READ_ONLY)
1411 return -EINVAL;
1413 r = dm_pool_commit_metadata(pool->pmd);
1414 if (r)
1415 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1416 else
1417 check_for_space(pool);
1419 return r;
1422 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1424 unsigned long flags;
1426 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1427 DMWARN("%s: reached low water mark for data device: sending event.",
1428 dm_device_name(pool->pool_md));
1429 spin_lock_irqsave(&pool->lock, flags);
1430 pool->low_water_triggered = true;
1431 spin_unlock_irqrestore(&pool->lock, flags);
1432 dm_table_event(pool->ti->table);
1436 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1438 int r;
1439 dm_block_t free_blocks;
1440 struct pool *pool = tc->pool;
1442 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1443 return -EINVAL;
1445 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1446 if (r) {
1447 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1448 return r;
1451 check_low_water_mark(pool, free_blocks);
1453 if (!free_blocks) {
1455 * Try to commit to see if that will free up some
1456 * more space.
1458 r = commit(pool);
1459 if (r)
1460 return r;
1462 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1463 if (r) {
1464 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1465 return r;
1468 if (!free_blocks) {
1469 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1470 return -ENOSPC;
1474 r = dm_pool_alloc_data_block(pool->pmd, result);
1475 if (r) {
1476 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1477 return r;
1480 return 0;
1484 * If we have run out of space, queue bios until the device is
1485 * resumed, presumably after having been reloaded with more space.
1487 static void retry_on_resume(struct bio *bio)
1489 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1490 struct thin_c *tc = h->tc;
1491 unsigned long flags;
1493 spin_lock_irqsave(&tc->lock, flags);
1494 bio_list_add(&tc->retry_on_resume_list, bio);
1495 spin_unlock_irqrestore(&tc->lock, flags);
1498 static int should_error_unserviceable_bio(struct pool *pool)
1500 enum pool_mode m = get_pool_mode(pool);
1502 switch (m) {
1503 case PM_WRITE:
1504 /* Shouldn't get here */
1505 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1506 return -EIO;
1508 case PM_OUT_OF_DATA_SPACE:
1509 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1511 case PM_READ_ONLY:
1512 case PM_FAIL:
1513 return -EIO;
1514 default:
1515 /* Shouldn't get here */
1516 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1517 return -EIO;
1521 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1523 int error = should_error_unserviceable_bio(pool);
1525 if (error) {
1526 bio->bi_error = error;
1527 bio_endio(bio);
1528 } else
1529 retry_on_resume(bio);
1532 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1534 struct bio *bio;
1535 struct bio_list bios;
1536 int error;
1538 error = should_error_unserviceable_bio(pool);
1539 if (error) {
1540 cell_error_with_code(pool, cell, error);
1541 return;
1544 bio_list_init(&bios);
1545 cell_release(pool, cell, &bios);
1547 while ((bio = bio_list_pop(&bios)))
1548 retry_on_resume(bio);
1551 static void process_discard_cell_no_passdown(struct thin_c *tc,
1552 struct dm_bio_prison_cell *virt_cell)
1554 struct pool *pool = tc->pool;
1555 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1558 * We don't need to lock the data blocks, since there's no
1559 * passdown. We only lock data blocks for allocation and breaking sharing.
1561 m->tc = tc;
1562 m->virt_begin = virt_cell->key.block_begin;
1563 m->virt_end = virt_cell->key.block_end;
1564 m->cell = virt_cell;
1565 m->bio = virt_cell->holder;
1567 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1568 pool->process_prepared_discard(m);
1571 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1572 struct bio *bio)
1574 struct pool *pool = tc->pool;
1576 int r;
1577 bool maybe_shared;
1578 struct dm_cell_key data_key;
1579 struct dm_bio_prison_cell *data_cell;
1580 struct dm_thin_new_mapping *m;
1581 dm_block_t virt_begin, virt_end, data_begin;
1583 while (begin != end) {
1584 r = ensure_next_mapping(pool);
1585 if (r)
1586 /* we did our best */
1587 return;
1589 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1590 &data_begin, &maybe_shared);
1591 if (r)
1593 * Silently fail, letting any mappings we've
1594 * created complete.
1596 break;
1598 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1599 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1600 /* contention, we'll give up with this range */
1601 begin = virt_end;
1602 continue;
1606 * IO may still be going to the destination block. We must
1607 * quiesce before we can do the removal.
1609 m = get_next_mapping(pool);
1610 m->tc = tc;
1611 m->maybe_shared = maybe_shared;
1612 m->virt_begin = virt_begin;
1613 m->virt_end = virt_end;
1614 m->data_block = data_begin;
1615 m->cell = data_cell;
1616 m->bio = bio;
1619 * The parent bio must not complete before sub discard bios are
1620 * chained to it (see end_discard's bio_chain)!
1622 * This per-mapping bi_remaining increment is paired with
1623 * the implicit decrement that occurs via bio_endio() in
1624 * end_discard().
1626 bio_inc_remaining(bio);
1627 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1628 pool->process_prepared_discard(m);
1630 begin = virt_end;
1634 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1636 struct bio *bio = virt_cell->holder;
1637 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1640 * The virt_cell will only get freed once the origin bio completes.
1641 * This means it will remain locked while all the individual
1642 * passdown bios are in flight.
1644 h->cell = virt_cell;
1645 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1648 * We complete the bio now, knowing that the bi_remaining field
1649 * will prevent completion until the sub range discards have
1650 * completed.
1652 bio_endio(bio);
1655 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1657 dm_block_t begin, end;
1658 struct dm_cell_key virt_key;
1659 struct dm_bio_prison_cell *virt_cell;
1661 get_bio_block_range(tc, bio, &begin, &end);
1662 if (begin == end) {
1664 * The discard covers less than a block.
1666 bio_endio(bio);
1667 return;
1670 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1671 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1673 * Potential starvation issue: We're relying on the
1674 * fs/application being well behaved, and not trying to
1675 * send IO to a region at the same time as discarding it.
1676 * If they do this persistently then it's possible this
1677 * cell will never be granted.
1679 return;
1681 tc->pool->process_discard_cell(tc, virt_cell);
1684 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1685 struct dm_cell_key *key,
1686 struct dm_thin_lookup_result *lookup_result,
1687 struct dm_bio_prison_cell *cell)
1689 int r;
1690 dm_block_t data_block;
1691 struct pool *pool = tc->pool;
1693 r = alloc_data_block(tc, &data_block);
1694 switch (r) {
1695 case 0:
1696 schedule_internal_copy(tc, block, lookup_result->block,
1697 data_block, cell, bio);
1698 break;
1700 case -ENOSPC:
1701 retry_bios_on_resume(pool, cell);
1702 break;
1704 default:
1705 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1706 __func__, r);
1707 cell_error(pool, cell);
1708 break;
1712 static void __remap_and_issue_shared_cell(void *context,
1713 struct dm_bio_prison_cell *cell)
1715 struct remap_info *info = context;
1716 struct bio *bio;
1718 while ((bio = bio_list_pop(&cell->bios))) {
1719 if ((bio_data_dir(bio) == WRITE) ||
1720 (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
1721 bio_op(bio) == REQ_OP_DISCARD))
1722 bio_list_add(&info->defer_bios, bio);
1723 else {
1724 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1726 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1727 inc_all_io_entry(info->tc->pool, bio);
1728 bio_list_add(&info->issue_bios, bio);
1733 static void remap_and_issue_shared_cell(struct thin_c *tc,
1734 struct dm_bio_prison_cell *cell,
1735 dm_block_t block)
1737 struct bio *bio;
1738 struct remap_info info;
1740 info.tc = tc;
1741 bio_list_init(&info.defer_bios);
1742 bio_list_init(&info.issue_bios);
1744 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1745 &info, cell);
1747 while ((bio = bio_list_pop(&info.defer_bios)))
1748 thin_defer_bio(tc, bio);
1750 while ((bio = bio_list_pop(&info.issue_bios)))
1751 remap_and_issue(tc, bio, block);
1754 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1755 dm_block_t block,
1756 struct dm_thin_lookup_result *lookup_result,
1757 struct dm_bio_prison_cell *virt_cell)
1759 struct dm_bio_prison_cell *data_cell;
1760 struct pool *pool = tc->pool;
1761 struct dm_cell_key key;
1764 * If cell is already occupied, then sharing is already in the process
1765 * of being broken so we have nothing further to do here.
1767 build_data_key(tc->td, lookup_result->block, &key);
1768 if (bio_detain(pool, &key, bio, &data_cell)) {
1769 cell_defer_no_holder(tc, virt_cell);
1770 return;
1773 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1774 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1775 cell_defer_no_holder(tc, virt_cell);
1776 } else {
1777 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1779 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1780 inc_all_io_entry(pool, bio);
1781 remap_and_issue(tc, bio, lookup_result->block);
1783 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1784 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1788 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1789 struct dm_bio_prison_cell *cell)
1791 int r;
1792 dm_block_t data_block;
1793 struct pool *pool = tc->pool;
1796 * Remap empty bios (flushes) immediately, without provisioning.
1798 if (!bio->bi_iter.bi_size) {
1799 inc_all_io_entry(pool, bio);
1800 cell_defer_no_holder(tc, cell);
1802 remap_and_issue(tc, bio, 0);
1803 return;
1807 * Fill read bios with zeroes and complete them immediately.
1809 if (bio_data_dir(bio) == READ) {
1810 zero_fill_bio(bio);
1811 cell_defer_no_holder(tc, cell);
1812 bio_endio(bio);
1813 return;
1816 r = alloc_data_block(tc, &data_block);
1817 switch (r) {
1818 case 0:
1819 if (tc->origin_dev)
1820 schedule_external_copy(tc, block, data_block, cell, bio);
1821 else
1822 schedule_zero(tc, block, data_block, cell, bio);
1823 break;
1825 case -ENOSPC:
1826 retry_bios_on_resume(pool, cell);
1827 break;
1829 default:
1830 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1831 __func__, r);
1832 cell_error(pool, cell);
1833 break;
1837 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1839 int r;
1840 struct pool *pool = tc->pool;
1841 struct bio *bio = cell->holder;
1842 dm_block_t block = get_bio_block(tc, bio);
1843 struct dm_thin_lookup_result lookup_result;
1845 if (tc->requeue_mode) {
1846 cell_requeue(pool, cell);
1847 return;
1850 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1851 switch (r) {
1852 case 0:
1853 if (lookup_result.shared)
1854 process_shared_bio(tc, bio, block, &lookup_result, cell);
1855 else {
1856 inc_all_io_entry(pool, bio);
1857 remap_and_issue(tc, bio, lookup_result.block);
1858 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1860 break;
1862 case -ENODATA:
1863 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1864 inc_all_io_entry(pool, bio);
1865 cell_defer_no_holder(tc, cell);
1867 if (bio_end_sector(bio) <= tc->origin_size)
1868 remap_to_origin_and_issue(tc, bio);
1870 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1871 zero_fill_bio(bio);
1872 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1873 remap_to_origin_and_issue(tc, bio);
1875 } else {
1876 zero_fill_bio(bio);
1877 bio_endio(bio);
1879 } else
1880 provision_block(tc, bio, block, cell);
1881 break;
1883 default:
1884 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1885 __func__, r);
1886 cell_defer_no_holder(tc, cell);
1887 bio_io_error(bio);
1888 break;
1892 static void process_bio(struct thin_c *tc, struct bio *bio)
1894 struct pool *pool = tc->pool;
1895 dm_block_t block = get_bio_block(tc, bio);
1896 struct dm_bio_prison_cell *cell;
1897 struct dm_cell_key key;
1900 * If cell is already occupied, then the block is already
1901 * being provisioned so we have nothing further to do here.
1903 build_virtual_key(tc->td, block, &key);
1904 if (bio_detain(pool, &key, bio, &cell))
1905 return;
1907 process_cell(tc, cell);
1910 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1911 struct dm_bio_prison_cell *cell)
1913 int r;
1914 int rw = bio_data_dir(bio);
1915 dm_block_t block = get_bio_block(tc, bio);
1916 struct dm_thin_lookup_result lookup_result;
1918 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1919 switch (r) {
1920 case 0:
1921 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1922 handle_unserviceable_bio(tc->pool, bio);
1923 if (cell)
1924 cell_defer_no_holder(tc, cell);
1925 } else {
1926 inc_all_io_entry(tc->pool, bio);
1927 remap_and_issue(tc, bio, lookup_result.block);
1928 if (cell)
1929 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1931 break;
1933 case -ENODATA:
1934 if (cell)
1935 cell_defer_no_holder(tc, cell);
1936 if (rw != READ) {
1937 handle_unserviceable_bio(tc->pool, bio);
1938 break;
1941 if (tc->origin_dev) {
1942 inc_all_io_entry(tc->pool, bio);
1943 remap_to_origin_and_issue(tc, bio);
1944 break;
1947 zero_fill_bio(bio);
1948 bio_endio(bio);
1949 break;
1951 default:
1952 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1953 __func__, r);
1954 if (cell)
1955 cell_defer_no_holder(tc, cell);
1956 bio_io_error(bio);
1957 break;
1961 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1963 __process_bio_read_only(tc, bio, NULL);
1966 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1968 __process_bio_read_only(tc, cell->holder, cell);
1971 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1973 bio_endio(bio);
1976 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1978 bio_io_error(bio);
1981 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1983 cell_success(tc->pool, cell);
1986 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1988 cell_error(tc->pool, cell);
1992 * FIXME: should we also commit due to size of transaction, measured in
1993 * metadata blocks?
1995 static int need_commit_due_to_time(struct pool *pool)
1997 return !time_in_range(jiffies, pool->last_commit_jiffies,
1998 pool->last_commit_jiffies + COMMIT_PERIOD);
2001 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2002 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2004 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2006 struct rb_node **rbp, *parent;
2007 struct dm_thin_endio_hook *pbd;
2008 sector_t bi_sector = bio->bi_iter.bi_sector;
2010 rbp = &tc->sort_bio_list.rb_node;
2011 parent = NULL;
2012 while (*rbp) {
2013 parent = *rbp;
2014 pbd = thin_pbd(parent);
2016 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2017 rbp = &(*rbp)->rb_left;
2018 else
2019 rbp = &(*rbp)->rb_right;
2022 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2023 rb_link_node(&pbd->rb_node, parent, rbp);
2024 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2027 static void __extract_sorted_bios(struct thin_c *tc)
2029 struct rb_node *node;
2030 struct dm_thin_endio_hook *pbd;
2031 struct bio *bio;
2033 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2034 pbd = thin_pbd(node);
2035 bio = thin_bio(pbd);
2037 bio_list_add(&tc->deferred_bio_list, bio);
2038 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2041 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2044 static void __sort_thin_deferred_bios(struct thin_c *tc)
2046 struct bio *bio;
2047 struct bio_list bios;
2049 bio_list_init(&bios);
2050 bio_list_merge(&bios, &tc->deferred_bio_list);
2051 bio_list_init(&tc->deferred_bio_list);
2053 /* Sort deferred_bio_list using rb-tree */
2054 while ((bio = bio_list_pop(&bios)))
2055 __thin_bio_rb_add(tc, bio);
2058 * Transfer the sorted bios in sort_bio_list back to
2059 * deferred_bio_list to allow lockless submission of
2060 * all bios.
2062 __extract_sorted_bios(tc);
2065 static void process_thin_deferred_bios(struct thin_c *tc)
2067 struct pool *pool = tc->pool;
2068 unsigned long flags;
2069 struct bio *bio;
2070 struct bio_list bios;
2071 struct blk_plug plug;
2072 unsigned count = 0;
2074 if (tc->requeue_mode) {
2075 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
2076 return;
2079 bio_list_init(&bios);
2081 spin_lock_irqsave(&tc->lock, flags);
2083 if (bio_list_empty(&tc->deferred_bio_list)) {
2084 spin_unlock_irqrestore(&tc->lock, flags);
2085 return;
2088 __sort_thin_deferred_bios(tc);
2090 bio_list_merge(&bios, &tc->deferred_bio_list);
2091 bio_list_init(&tc->deferred_bio_list);
2093 spin_unlock_irqrestore(&tc->lock, flags);
2095 blk_start_plug(&plug);
2096 while ((bio = bio_list_pop(&bios))) {
2098 * If we've got no free new_mapping structs, and processing
2099 * this bio might require one, we pause until there are some
2100 * prepared mappings to process.
2102 if (ensure_next_mapping(pool)) {
2103 spin_lock_irqsave(&tc->lock, flags);
2104 bio_list_add(&tc->deferred_bio_list, bio);
2105 bio_list_merge(&tc->deferred_bio_list, &bios);
2106 spin_unlock_irqrestore(&tc->lock, flags);
2107 break;
2110 if (bio_op(bio) == REQ_OP_DISCARD)
2111 pool->process_discard(tc, bio);
2112 else
2113 pool->process_bio(tc, bio);
2115 if ((count++ & 127) == 0) {
2116 throttle_work_update(&pool->throttle);
2117 dm_pool_issue_prefetches(pool->pmd);
2120 blk_finish_plug(&plug);
2123 static int cmp_cells(const void *lhs, const void *rhs)
2125 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2126 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2128 BUG_ON(!lhs_cell->holder);
2129 BUG_ON(!rhs_cell->holder);
2131 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2132 return -1;
2134 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2135 return 1;
2137 return 0;
2140 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2142 unsigned count = 0;
2143 struct dm_bio_prison_cell *cell, *tmp;
2145 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2146 if (count >= CELL_SORT_ARRAY_SIZE)
2147 break;
2149 pool->cell_sort_array[count++] = cell;
2150 list_del(&cell->user_list);
2153 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2155 return count;
2158 static void process_thin_deferred_cells(struct thin_c *tc)
2160 struct pool *pool = tc->pool;
2161 unsigned long flags;
2162 struct list_head cells;
2163 struct dm_bio_prison_cell *cell;
2164 unsigned i, j, count;
2166 INIT_LIST_HEAD(&cells);
2168 spin_lock_irqsave(&tc->lock, flags);
2169 list_splice_init(&tc->deferred_cells, &cells);
2170 spin_unlock_irqrestore(&tc->lock, flags);
2172 if (list_empty(&cells))
2173 return;
2175 do {
2176 count = sort_cells(tc->pool, &cells);
2178 for (i = 0; i < count; i++) {
2179 cell = pool->cell_sort_array[i];
2180 BUG_ON(!cell->holder);
2183 * If we've got no free new_mapping structs, and processing
2184 * this bio might require one, we pause until there are some
2185 * prepared mappings to process.
2187 if (ensure_next_mapping(pool)) {
2188 for (j = i; j < count; j++)
2189 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2191 spin_lock_irqsave(&tc->lock, flags);
2192 list_splice(&cells, &tc->deferred_cells);
2193 spin_unlock_irqrestore(&tc->lock, flags);
2194 return;
2197 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2198 pool->process_discard_cell(tc, cell);
2199 else
2200 pool->process_cell(tc, cell);
2202 } while (!list_empty(&cells));
2205 static void thin_get(struct thin_c *tc);
2206 static void thin_put(struct thin_c *tc);
2209 * We can't hold rcu_read_lock() around code that can block. So we
2210 * find a thin with the rcu lock held; bump a refcount; then drop
2211 * the lock.
2213 static struct thin_c *get_first_thin(struct pool *pool)
2215 struct thin_c *tc = NULL;
2217 rcu_read_lock();
2218 if (!list_empty(&pool->active_thins)) {
2219 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2220 thin_get(tc);
2222 rcu_read_unlock();
2224 return tc;
2227 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2229 struct thin_c *old_tc = tc;
2231 rcu_read_lock();
2232 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2233 thin_get(tc);
2234 thin_put(old_tc);
2235 rcu_read_unlock();
2236 return tc;
2238 thin_put(old_tc);
2239 rcu_read_unlock();
2241 return NULL;
2244 static void process_deferred_bios(struct pool *pool)
2246 unsigned long flags;
2247 struct bio *bio;
2248 struct bio_list bios;
2249 struct thin_c *tc;
2251 tc = get_first_thin(pool);
2252 while (tc) {
2253 process_thin_deferred_cells(tc);
2254 process_thin_deferred_bios(tc);
2255 tc = get_next_thin(pool, tc);
2259 * If there are any deferred flush bios, we must commit
2260 * the metadata before issuing them.
2262 bio_list_init(&bios);
2263 spin_lock_irqsave(&pool->lock, flags);
2264 bio_list_merge(&bios, &pool->deferred_flush_bios);
2265 bio_list_init(&pool->deferred_flush_bios);
2266 spin_unlock_irqrestore(&pool->lock, flags);
2268 if (bio_list_empty(&bios) &&
2269 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2270 return;
2272 if (commit(pool)) {
2273 while ((bio = bio_list_pop(&bios)))
2274 bio_io_error(bio);
2275 return;
2277 pool->last_commit_jiffies = jiffies;
2279 while ((bio = bio_list_pop(&bios)))
2280 generic_make_request(bio);
2283 static void do_worker(struct work_struct *ws)
2285 struct pool *pool = container_of(ws, struct pool, worker);
2287 throttle_work_start(&pool->throttle);
2288 dm_pool_issue_prefetches(pool->pmd);
2289 throttle_work_update(&pool->throttle);
2290 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2291 throttle_work_update(&pool->throttle);
2292 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2293 throttle_work_update(&pool->throttle);
2294 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2295 throttle_work_update(&pool->throttle);
2296 process_deferred_bios(pool);
2297 throttle_work_complete(&pool->throttle);
2301 * We want to commit periodically so that not too much
2302 * unwritten data builds up.
2304 static void do_waker(struct work_struct *ws)
2306 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2307 wake_worker(pool);
2308 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2311 static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2314 * We're holding onto IO to allow userland time to react. After the
2315 * timeout either the pool will have been resized (and thus back in
2316 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2318 static void do_no_space_timeout(struct work_struct *ws)
2320 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2321 no_space_timeout);
2323 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2324 pool->pf.error_if_no_space = true;
2325 notify_of_pool_mode_change_to_oods(pool);
2326 error_retry_list_with_code(pool, -ENOSPC);
2330 /*----------------------------------------------------------------*/
2332 struct pool_work {
2333 struct work_struct worker;
2334 struct completion complete;
2337 static struct pool_work *to_pool_work(struct work_struct *ws)
2339 return container_of(ws, struct pool_work, worker);
2342 static void pool_work_complete(struct pool_work *pw)
2344 complete(&pw->complete);
2347 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2348 void (*fn)(struct work_struct *))
2350 INIT_WORK_ONSTACK(&pw->worker, fn);
2351 init_completion(&pw->complete);
2352 queue_work(pool->wq, &pw->worker);
2353 wait_for_completion(&pw->complete);
2356 /*----------------------------------------------------------------*/
2358 struct noflush_work {
2359 struct pool_work pw;
2360 struct thin_c *tc;
2363 static struct noflush_work *to_noflush(struct work_struct *ws)
2365 return container_of(to_pool_work(ws), struct noflush_work, pw);
2368 static void do_noflush_start(struct work_struct *ws)
2370 struct noflush_work *w = to_noflush(ws);
2371 w->tc->requeue_mode = true;
2372 requeue_io(w->tc);
2373 pool_work_complete(&w->pw);
2376 static void do_noflush_stop(struct work_struct *ws)
2378 struct noflush_work *w = to_noflush(ws);
2379 w->tc->requeue_mode = false;
2380 pool_work_complete(&w->pw);
2383 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2385 struct noflush_work w;
2387 w.tc = tc;
2388 pool_work_wait(&w.pw, tc->pool, fn);
2391 /*----------------------------------------------------------------*/
2393 static enum pool_mode get_pool_mode(struct pool *pool)
2395 return pool->pf.mode;
2398 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2400 dm_table_event(pool->ti->table);
2401 DMINFO("%s: switching pool to %s mode",
2402 dm_device_name(pool->pool_md), new_mode);
2405 static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2407 if (!pool->pf.error_if_no_space)
2408 notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2409 else
2410 notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2413 static bool passdown_enabled(struct pool_c *pt)
2415 return pt->adjusted_pf.discard_passdown;
2418 static void set_discard_callbacks(struct pool *pool)
2420 struct pool_c *pt = pool->ti->private;
2422 if (passdown_enabled(pt)) {
2423 pool->process_discard_cell = process_discard_cell_passdown;
2424 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2425 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2426 } else {
2427 pool->process_discard_cell = process_discard_cell_no_passdown;
2428 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2432 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2434 struct pool_c *pt = pool->ti->private;
2435 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2436 enum pool_mode old_mode = get_pool_mode(pool);
2437 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2440 * Never allow the pool to transition to PM_WRITE mode if user
2441 * intervention is required to verify metadata and data consistency.
2443 if (new_mode == PM_WRITE && needs_check) {
2444 DMERR("%s: unable to switch pool to write mode until repaired.",
2445 dm_device_name(pool->pool_md));
2446 if (old_mode != new_mode)
2447 new_mode = old_mode;
2448 else
2449 new_mode = PM_READ_ONLY;
2452 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2453 * not going to recover without a thin_repair. So we never let the
2454 * pool move out of the old mode.
2456 if (old_mode == PM_FAIL)
2457 new_mode = old_mode;
2459 switch (new_mode) {
2460 case PM_FAIL:
2461 if (old_mode != new_mode)
2462 notify_of_pool_mode_change(pool, "failure");
2463 dm_pool_metadata_read_only(pool->pmd);
2464 pool->process_bio = process_bio_fail;
2465 pool->process_discard = process_bio_fail;
2466 pool->process_cell = process_cell_fail;
2467 pool->process_discard_cell = process_cell_fail;
2468 pool->process_prepared_mapping = process_prepared_mapping_fail;
2469 pool->process_prepared_discard = process_prepared_discard_fail;
2471 error_retry_list(pool);
2472 break;
2474 case PM_READ_ONLY:
2475 if (old_mode != new_mode)
2476 notify_of_pool_mode_change(pool, "read-only");
2477 dm_pool_metadata_read_only(pool->pmd);
2478 pool->process_bio = process_bio_read_only;
2479 pool->process_discard = process_bio_success;
2480 pool->process_cell = process_cell_read_only;
2481 pool->process_discard_cell = process_cell_success;
2482 pool->process_prepared_mapping = process_prepared_mapping_fail;
2483 pool->process_prepared_discard = process_prepared_discard_success;
2485 error_retry_list(pool);
2486 break;
2488 case PM_OUT_OF_DATA_SPACE:
2490 * Ideally we'd never hit this state; the low water mark
2491 * would trigger userland to extend the pool before we
2492 * completely run out of data space. However, many small
2493 * IOs to unprovisioned space can consume data space at an
2494 * alarming rate. Adjust your low water mark if you're
2495 * frequently seeing this mode.
2497 if (old_mode != new_mode)
2498 notify_of_pool_mode_change_to_oods(pool);
2499 pool->out_of_data_space = true;
2500 pool->process_bio = process_bio_read_only;
2501 pool->process_discard = process_discard_bio;
2502 pool->process_cell = process_cell_read_only;
2503 pool->process_prepared_mapping = process_prepared_mapping;
2504 set_discard_callbacks(pool);
2506 if (!pool->pf.error_if_no_space && no_space_timeout)
2507 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2508 break;
2510 case PM_WRITE:
2511 if (old_mode != new_mode)
2512 notify_of_pool_mode_change(pool, "write");
2513 pool->out_of_data_space = false;
2514 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2515 dm_pool_metadata_read_write(pool->pmd);
2516 pool->process_bio = process_bio;
2517 pool->process_discard = process_discard_bio;
2518 pool->process_cell = process_cell;
2519 pool->process_prepared_mapping = process_prepared_mapping;
2520 set_discard_callbacks(pool);
2521 break;
2524 pool->pf.mode = new_mode;
2526 * The pool mode may have changed, sync it so bind_control_target()
2527 * doesn't cause an unexpected mode transition on resume.
2529 pt->adjusted_pf.mode = new_mode;
2532 static void abort_transaction(struct pool *pool)
2534 const char *dev_name = dm_device_name(pool->pool_md);
2536 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2537 if (dm_pool_abort_metadata(pool->pmd)) {
2538 DMERR("%s: failed to abort metadata transaction", dev_name);
2539 set_pool_mode(pool, PM_FAIL);
2542 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2543 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2544 set_pool_mode(pool, PM_FAIL);
2548 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2550 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2551 dm_device_name(pool->pool_md), op, r);
2553 abort_transaction(pool);
2554 set_pool_mode(pool, PM_READ_ONLY);
2557 /*----------------------------------------------------------------*/
2560 * Mapping functions.
2564 * Called only while mapping a thin bio to hand it over to the workqueue.
2566 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2568 unsigned long flags;
2569 struct pool *pool = tc->pool;
2571 spin_lock_irqsave(&tc->lock, flags);
2572 bio_list_add(&tc->deferred_bio_list, bio);
2573 spin_unlock_irqrestore(&tc->lock, flags);
2575 wake_worker(pool);
2578 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2580 struct pool *pool = tc->pool;
2582 throttle_lock(&pool->throttle);
2583 thin_defer_bio(tc, bio);
2584 throttle_unlock(&pool->throttle);
2587 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2589 unsigned long flags;
2590 struct pool *pool = tc->pool;
2592 throttle_lock(&pool->throttle);
2593 spin_lock_irqsave(&tc->lock, flags);
2594 list_add_tail(&cell->user_list, &tc->deferred_cells);
2595 spin_unlock_irqrestore(&tc->lock, flags);
2596 throttle_unlock(&pool->throttle);
2598 wake_worker(pool);
2601 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2603 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2605 h->tc = tc;
2606 h->shared_read_entry = NULL;
2607 h->all_io_entry = NULL;
2608 h->overwrite_mapping = NULL;
2609 h->cell = NULL;
2613 * Non-blocking function called from the thin target's map function.
2615 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2617 int r;
2618 struct thin_c *tc = ti->private;
2619 dm_block_t block = get_bio_block(tc, bio);
2620 struct dm_thin_device *td = tc->td;
2621 struct dm_thin_lookup_result result;
2622 struct dm_bio_prison_cell *virt_cell, *data_cell;
2623 struct dm_cell_key key;
2625 thin_hook_bio(tc, bio);
2627 if (tc->requeue_mode) {
2628 bio->bi_error = DM_ENDIO_REQUEUE;
2629 bio_endio(bio);
2630 return DM_MAPIO_SUBMITTED;
2633 if (get_pool_mode(tc->pool) == PM_FAIL) {
2634 bio_io_error(bio);
2635 return DM_MAPIO_SUBMITTED;
2638 if (bio->bi_opf & (REQ_PREFLUSH | REQ_FUA) ||
2639 bio_op(bio) == REQ_OP_DISCARD) {
2640 thin_defer_bio_with_throttle(tc, bio);
2641 return DM_MAPIO_SUBMITTED;
2645 * We must hold the virtual cell before doing the lookup, otherwise
2646 * there's a race with discard.
2648 build_virtual_key(tc->td, block, &key);
2649 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2650 return DM_MAPIO_SUBMITTED;
2652 r = dm_thin_find_block(td, block, 0, &result);
2655 * Note that we defer readahead too.
2657 switch (r) {
2658 case 0:
2659 if (unlikely(result.shared)) {
2661 * We have a race condition here between the
2662 * result.shared value returned by the lookup and
2663 * snapshot creation, which may cause new
2664 * sharing.
2666 * To avoid this always quiesce the origin before
2667 * taking the snap. You want to do this anyway to
2668 * ensure a consistent application view
2669 * (i.e. lockfs).
2671 * More distant ancestors are irrelevant. The
2672 * shared flag will be set in their case.
2674 thin_defer_cell(tc, virt_cell);
2675 return DM_MAPIO_SUBMITTED;
2678 build_data_key(tc->td, result.block, &key);
2679 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2680 cell_defer_no_holder(tc, virt_cell);
2681 return DM_MAPIO_SUBMITTED;
2684 inc_all_io_entry(tc->pool, bio);
2685 cell_defer_no_holder(tc, data_cell);
2686 cell_defer_no_holder(tc, virt_cell);
2688 remap(tc, bio, result.block);
2689 return DM_MAPIO_REMAPPED;
2691 case -ENODATA:
2692 case -EWOULDBLOCK:
2693 thin_defer_cell(tc, virt_cell);
2694 return DM_MAPIO_SUBMITTED;
2696 default:
2698 * Must always call bio_io_error on failure.
2699 * dm_thin_find_block can fail with -EINVAL if the
2700 * pool is switched to fail-io mode.
2702 bio_io_error(bio);
2703 cell_defer_no_holder(tc, virt_cell);
2704 return DM_MAPIO_SUBMITTED;
2708 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2710 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2711 struct request_queue *q;
2713 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2714 return 1;
2716 q = bdev_get_queue(pt->data_dev->bdev);
2717 return bdi_congested(&q->backing_dev_info, bdi_bits);
2720 static void requeue_bios(struct pool *pool)
2722 unsigned long flags;
2723 struct thin_c *tc;
2725 rcu_read_lock();
2726 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2727 spin_lock_irqsave(&tc->lock, flags);
2728 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2729 bio_list_init(&tc->retry_on_resume_list);
2730 spin_unlock_irqrestore(&tc->lock, flags);
2732 rcu_read_unlock();
2735 /*----------------------------------------------------------------
2736 * Binding of control targets to a pool object
2737 *--------------------------------------------------------------*/
2738 static bool data_dev_supports_discard(struct pool_c *pt)
2740 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2742 return q && blk_queue_discard(q);
2745 static bool is_factor(sector_t block_size, uint32_t n)
2747 return !sector_div(block_size, n);
2751 * If discard_passdown was enabled verify that the data device
2752 * supports discards. Disable discard_passdown if not.
2754 static void disable_passdown_if_not_supported(struct pool_c *pt)
2756 struct pool *pool = pt->pool;
2757 struct block_device *data_bdev = pt->data_dev->bdev;
2758 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2759 const char *reason = NULL;
2760 char buf[BDEVNAME_SIZE];
2762 if (!pt->adjusted_pf.discard_passdown)
2763 return;
2765 if (!data_dev_supports_discard(pt))
2766 reason = "discard unsupported";
2768 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2769 reason = "max discard sectors smaller than a block";
2771 if (reason) {
2772 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2773 pt->adjusted_pf.discard_passdown = false;
2777 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2779 struct pool_c *pt = ti->private;
2782 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2784 enum pool_mode old_mode = get_pool_mode(pool);
2785 enum pool_mode new_mode = pt->adjusted_pf.mode;
2788 * Don't change the pool's mode until set_pool_mode() below.
2789 * Otherwise the pool's process_* function pointers may
2790 * not match the desired pool mode.
2792 pt->adjusted_pf.mode = old_mode;
2794 pool->ti = ti;
2795 pool->pf = pt->adjusted_pf;
2796 pool->low_water_blocks = pt->low_water_blocks;
2798 set_pool_mode(pool, new_mode);
2800 return 0;
2803 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2805 if (pool->ti == ti)
2806 pool->ti = NULL;
2809 /*----------------------------------------------------------------
2810 * Pool creation
2811 *--------------------------------------------------------------*/
2812 /* Initialize pool features. */
2813 static void pool_features_init(struct pool_features *pf)
2815 pf->mode = PM_WRITE;
2816 pf->zero_new_blocks = true;
2817 pf->discard_enabled = true;
2818 pf->discard_passdown = true;
2819 pf->error_if_no_space = false;
2822 static void __pool_destroy(struct pool *pool)
2824 __pool_table_remove(pool);
2826 vfree(pool->cell_sort_array);
2827 if (dm_pool_metadata_close(pool->pmd) < 0)
2828 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2830 dm_bio_prison_destroy(pool->prison);
2831 dm_kcopyd_client_destroy(pool->copier);
2833 if (pool->wq)
2834 destroy_workqueue(pool->wq);
2836 if (pool->next_mapping)
2837 mempool_free(pool->next_mapping, pool->mapping_pool);
2838 mempool_destroy(pool->mapping_pool);
2839 dm_deferred_set_destroy(pool->shared_read_ds);
2840 dm_deferred_set_destroy(pool->all_io_ds);
2841 kfree(pool);
2844 static struct kmem_cache *_new_mapping_cache;
2846 static struct pool *pool_create(struct mapped_device *pool_md,
2847 struct block_device *metadata_dev,
2848 unsigned long block_size,
2849 int read_only, char **error)
2851 int r;
2852 void *err_p;
2853 struct pool *pool;
2854 struct dm_pool_metadata *pmd;
2855 bool format_device = read_only ? false : true;
2857 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2858 if (IS_ERR(pmd)) {
2859 *error = "Error creating metadata object";
2860 return (struct pool *)pmd;
2863 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2864 if (!pool) {
2865 *error = "Error allocating memory for pool";
2866 err_p = ERR_PTR(-ENOMEM);
2867 goto bad_pool;
2870 pool->pmd = pmd;
2871 pool->sectors_per_block = block_size;
2872 if (block_size & (block_size - 1))
2873 pool->sectors_per_block_shift = -1;
2874 else
2875 pool->sectors_per_block_shift = __ffs(block_size);
2876 pool->low_water_blocks = 0;
2877 pool_features_init(&pool->pf);
2878 pool->prison = dm_bio_prison_create();
2879 if (!pool->prison) {
2880 *error = "Error creating pool's bio prison";
2881 err_p = ERR_PTR(-ENOMEM);
2882 goto bad_prison;
2885 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2886 if (IS_ERR(pool->copier)) {
2887 r = PTR_ERR(pool->copier);
2888 *error = "Error creating pool's kcopyd client";
2889 err_p = ERR_PTR(r);
2890 goto bad_kcopyd_client;
2894 * Create singlethreaded workqueue that will service all devices
2895 * that use this metadata.
2897 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2898 if (!pool->wq) {
2899 *error = "Error creating pool's workqueue";
2900 err_p = ERR_PTR(-ENOMEM);
2901 goto bad_wq;
2904 throttle_init(&pool->throttle);
2905 INIT_WORK(&pool->worker, do_worker);
2906 INIT_DELAYED_WORK(&pool->waker, do_waker);
2907 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2908 spin_lock_init(&pool->lock);
2909 bio_list_init(&pool->deferred_flush_bios);
2910 INIT_LIST_HEAD(&pool->prepared_mappings);
2911 INIT_LIST_HEAD(&pool->prepared_discards);
2912 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2913 INIT_LIST_HEAD(&pool->active_thins);
2914 pool->low_water_triggered = false;
2915 pool->suspended = true;
2916 pool->out_of_data_space = false;
2918 pool->shared_read_ds = dm_deferred_set_create();
2919 if (!pool->shared_read_ds) {
2920 *error = "Error creating pool's shared read deferred set";
2921 err_p = ERR_PTR(-ENOMEM);
2922 goto bad_shared_read_ds;
2925 pool->all_io_ds = dm_deferred_set_create();
2926 if (!pool->all_io_ds) {
2927 *error = "Error creating pool's all io deferred set";
2928 err_p = ERR_PTR(-ENOMEM);
2929 goto bad_all_io_ds;
2932 pool->next_mapping = NULL;
2933 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2934 _new_mapping_cache);
2935 if (!pool->mapping_pool) {
2936 *error = "Error creating pool's mapping mempool";
2937 err_p = ERR_PTR(-ENOMEM);
2938 goto bad_mapping_pool;
2941 pool->cell_sort_array = vmalloc(sizeof(*pool->cell_sort_array) * CELL_SORT_ARRAY_SIZE);
2942 if (!pool->cell_sort_array) {
2943 *error = "Error allocating cell sort array";
2944 err_p = ERR_PTR(-ENOMEM);
2945 goto bad_sort_array;
2948 pool->ref_count = 1;
2949 pool->last_commit_jiffies = jiffies;
2950 pool->pool_md = pool_md;
2951 pool->md_dev = metadata_dev;
2952 __pool_table_insert(pool);
2954 return pool;
2956 bad_sort_array:
2957 mempool_destroy(pool->mapping_pool);
2958 bad_mapping_pool:
2959 dm_deferred_set_destroy(pool->all_io_ds);
2960 bad_all_io_ds:
2961 dm_deferred_set_destroy(pool->shared_read_ds);
2962 bad_shared_read_ds:
2963 destroy_workqueue(pool->wq);
2964 bad_wq:
2965 dm_kcopyd_client_destroy(pool->copier);
2966 bad_kcopyd_client:
2967 dm_bio_prison_destroy(pool->prison);
2968 bad_prison:
2969 kfree(pool);
2970 bad_pool:
2971 if (dm_pool_metadata_close(pmd))
2972 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2974 return err_p;
2977 static void __pool_inc(struct pool *pool)
2979 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2980 pool->ref_count++;
2983 static void __pool_dec(struct pool *pool)
2985 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2986 BUG_ON(!pool->ref_count);
2987 if (!--pool->ref_count)
2988 __pool_destroy(pool);
2991 static struct pool *__pool_find(struct mapped_device *pool_md,
2992 struct block_device *metadata_dev,
2993 unsigned long block_size, int read_only,
2994 char **error, int *created)
2996 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2998 if (pool) {
2999 if (pool->pool_md != pool_md) {
3000 *error = "metadata device already in use by a pool";
3001 return ERR_PTR(-EBUSY);
3003 __pool_inc(pool);
3005 } else {
3006 pool = __pool_table_lookup(pool_md);
3007 if (pool) {
3008 if (pool->md_dev != metadata_dev) {
3009 *error = "different pool cannot replace a pool";
3010 return ERR_PTR(-EINVAL);
3012 __pool_inc(pool);
3014 } else {
3015 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3016 *created = 1;
3020 return pool;
3023 /*----------------------------------------------------------------
3024 * Pool target methods
3025 *--------------------------------------------------------------*/
3026 static void pool_dtr(struct dm_target *ti)
3028 struct pool_c *pt = ti->private;
3030 mutex_lock(&dm_thin_pool_table.mutex);
3032 unbind_control_target(pt->pool, ti);
3033 __pool_dec(pt->pool);
3034 dm_put_device(ti, pt->metadata_dev);
3035 dm_put_device(ti, pt->data_dev);
3036 kfree(pt);
3038 mutex_unlock(&dm_thin_pool_table.mutex);
3041 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3042 struct dm_target *ti)
3044 int r;
3045 unsigned argc;
3046 const char *arg_name;
3048 static struct dm_arg _args[] = {
3049 {0, 4, "Invalid number of pool feature arguments"},
3053 * No feature arguments supplied.
3055 if (!as->argc)
3056 return 0;
3058 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3059 if (r)
3060 return -EINVAL;
3062 while (argc && !r) {
3063 arg_name = dm_shift_arg(as);
3064 argc--;
3066 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3067 pf->zero_new_blocks = false;
3069 else if (!strcasecmp(arg_name, "ignore_discard"))
3070 pf->discard_enabled = false;
3072 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3073 pf->discard_passdown = false;
3075 else if (!strcasecmp(arg_name, "read_only"))
3076 pf->mode = PM_READ_ONLY;
3078 else if (!strcasecmp(arg_name, "error_if_no_space"))
3079 pf->error_if_no_space = true;
3081 else {
3082 ti->error = "Unrecognised pool feature requested";
3083 r = -EINVAL;
3084 break;
3088 return r;
3091 static void metadata_low_callback(void *context)
3093 struct pool *pool = context;
3095 DMWARN("%s: reached low water mark for metadata device: sending event.",
3096 dm_device_name(pool->pool_md));
3098 dm_table_event(pool->ti->table);
3101 static sector_t get_dev_size(struct block_device *bdev)
3103 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3106 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3108 sector_t metadata_dev_size = get_dev_size(bdev);
3109 char buffer[BDEVNAME_SIZE];
3111 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3112 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3113 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3116 static sector_t get_metadata_dev_size(struct block_device *bdev)
3118 sector_t metadata_dev_size = get_dev_size(bdev);
3120 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3121 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3123 return metadata_dev_size;
3126 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3128 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3130 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3132 return metadata_dev_size;
3136 * When a metadata threshold is crossed a dm event is triggered, and
3137 * userland should respond by growing the metadata device. We could let
3138 * userland set the threshold, like we do with the data threshold, but I'm
3139 * not sure they know enough to do this well.
3141 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3144 * 4M is ample for all ops with the possible exception of thin
3145 * device deletion which is harmless if it fails (just retry the
3146 * delete after you've grown the device).
3148 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3149 return min((dm_block_t)1024ULL /* 4M */, quarter);
3153 * thin-pool <metadata dev> <data dev>
3154 * <data block size (sectors)>
3155 * <low water mark (blocks)>
3156 * [<#feature args> [<arg>]*]
3158 * Optional feature arguments are:
3159 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3160 * ignore_discard: disable discard
3161 * no_discard_passdown: don't pass discards down to the data device
3162 * read_only: Don't allow any changes to be made to the pool metadata.
3163 * error_if_no_space: error IOs, instead of queueing, if no space.
3165 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3167 int r, pool_created = 0;
3168 struct pool_c *pt;
3169 struct pool *pool;
3170 struct pool_features pf;
3171 struct dm_arg_set as;
3172 struct dm_dev *data_dev;
3173 unsigned long block_size;
3174 dm_block_t low_water_blocks;
3175 struct dm_dev *metadata_dev;
3176 fmode_t metadata_mode;
3179 * FIXME Remove validation from scope of lock.
3181 mutex_lock(&dm_thin_pool_table.mutex);
3183 if (argc < 4) {
3184 ti->error = "Invalid argument count";
3185 r = -EINVAL;
3186 goto out_unlock;
3189 as.argc = argc;
3190 as.argv = argv;
3193 * Set default pool features.
3195 pool_features_init(&pf);
3197 dm_consume_args(&as, 4);
3198 r = parse_pool_features(&as, &pf, ti);
3199 if (r)
3200 goto out_unlock;
3202 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3203 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3204 if (r) {
3205 ti->error = "Error opening metadata block device";
3206 goto out_unlock;
3208 warn_if_metadata_device_too_big(metadata_dev->bdev);
3210 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3211 if (r) {
3212 ti->error = "Error getting data device";
3213 goto out_metadata;
3216 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3217 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3218 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3219 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3220 ti->error = "Invalid block size";
3221 r = -EINVAL;
3222 goto out;
3225 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3226 ti->error = "Invalid low water mark";
3227 r = -EINVAL;
3228 goto out;
3231 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3232 if (!pt) {
3233 r = -ENOMEM;
3234 goto out;
3237 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3238 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3239 if (IS_ERR(pool)) {
3240 r = PTR_ERR(pool);
3241 goto out_free_pt;
3245 * 'pool_created' reflects whether this is the first table load.
3246 * Top level discard support is not allowed to be changed after
3247 * initial load. This would require a pool reload to trigger thin
3248 * device changes.
3250 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3251 ti->error = "Discard support cannot be disabled once enabled";
3252 r = -EINVAL;
3253 goto out_flags_changed;
3256 pt->pool = pool;
3257 pt->ti = ti;
3258 pt->metadata_dev = metadata_dev;
3259 pt->data_dev = data_dev;
3260 pt->low_water_blocks = low_water_blocks;
3261 pt->adjusted_pf = pt->requested_pf = pf;
3262 ti->num_flush_bios = 1;
3265 * Only need to enable discards if the pool should pass
3266 * them down to the data device. The thin device's discard
3267 * processing will cause mappings to be removed from the btree.
3269 ti->discard_zeroes_data_unsupported = true;
3270 if (pf.discard_enabled && pf.discard_passdown) {
3271 ti->num_discard_bios = 1;
3274 * Setting 'discards_supported' circumvents the normal
3275 * stacking of discard limits (this keeps the pool and
3276 * thin devices' discard limits consistent).
3278 ti->discards_supported = true;
3280 ti->private = pt;
3282 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3283 calc_metadata_threshold(pt),
3284 metadata_low_callback,
3285 pool);
3286 if (r)
3287 goto out_flags_changed;
3289 pt->callbacks.congested_fn = pool_is_congested;
3290 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3292 mutex_unlock(&dm_thin_pool_table.mutex);
3294 return 0;
3296 out_flags_changed:
3297 __pool_dec(pool);
3298 out_free_pt:
3299 kfree(pt);
3300 out:
3301 dm_put_device(ti, data_dev);
3302 out_metadata:
3303 dm_put_device(ti, metadata_dev);
3304 out_unlock:
3305 mutex_unlock(&dm_thin_pool_table.mutex);
3307 return r;
3310 static int pool_map(struct dm_target *ti, struct bio *bio)
3312 int r;
3313 struct pool_c *pt = ti->private;
3314 struct pool *pool = pt->pool;
3315 unsigned long flags;
3318 * As this is a singleton target, ti->begin is always zero.
3320 spin_lock_irqsave(&pool->lock, flags);
3321 bio->bi_bdev = pt->data_dev->bdev;
3322 r = DM_MAPIO_REMAPPED;
3323 spin_unlock_irqrestore(&pool->lock, flags);
3325 return r;
3328 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3330 int r;
3331 struct pool_c *pt = ti->private;
3332 struct pool *pool = pt->pool;
3333 sector_t data_size = ti->len;
3334 dm_block_t sb_data_size;
3336 *need_commit = false;
3338 (void) sector_div(data_size, pool->sectors_per_block);
3340 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3341 if (r) {
3342 DMERR("%s: failed to retrieve data device size",
3343 dm_device_name(pool->pool_md));
3344 return r;
3347 if (data_size < sb_data_size) {
3348 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3349 dm_device_name(pool->pool_md),
3350 (unsigned long long)data_size, sb_data_size);
3351 return -EINVAL;
3353 } else if (data_size > sb_data_size) {
3354 if (dm_pool_metadata_needs_check(pool->pmd)) {
3355 DMERR("%s: unable to grow the data device until repaired.",
3356 dm_device_name(pool->pool_md));
3357 return 0;
3360 if (sb_data_size)
3361 DMINFO("%s: growing the data device from %llu to %llu blocks",
3362 dm_device_name(pool->pool_md),
3363 sb_data_size, (unsigned long long)data_size);
3364 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3365 if (r) {
3366 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3367 return r;
3370 *need_commit = true;
3373 return 0;
3376 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3378 int r;
3379 struct pool_c *pt = ti->private;
3380 struct pool *pool = pt->pool;
3381 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3383 *need_commit = false;
3385 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3387 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3388 if (r) {
3389 DMERR("%s: failed to retrieve metadata device size",
3390 dm_device_name(pool->pool_md));
3391 return r;
3394 if (metadata_dev_size < sb_metadata_dev_size) {
3395 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3396 dm_device_name(pool->pool_md),
3397 metadata_dev_size, sb_metadata_dev_size);
3398 return -EINVAL;
3400 } else if (metadata_dev_size > sb_metadata_dev_size) {
3401 if (dm_pool_metadata_needs_check(pool->pmd)) {
3402 DMERR("%s: unable to grow the metadata device until repaired.",
3403 dm_device_name(pool->pool_md));
3404 return 0;
3407 warn_if_metadata_device_too_big(pool->md_dev);
3408 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3409 dm_device_name(pool->pool_md),
3410 sb_metadata_dev_size, metadata_dev_size);
3411 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3412 if (r) {
3413 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3414 return r;
3417 *need_commit = true;
3420 return 0;
3424 * Retrieves the number of blocks of the data device from
3425 * the superblock and compares it to the actual device size,
3426 * thus resizing the data device in case it has grown.
3428 * This both copes with opening preallocated data devices in the ctr
3429 * being followed by a resume
3430 * -and-
3431 * calling the resume method individually after userspace has
3432 * grown the data device in reaction to a table event.
3434 static int pool_preresume(struct dm_target *ti)
3436 int r;
3437 bool need_commit1, need_commit2;
3438 struct pool_c *pt = ti->private;
3439 struct pool *pool = pt->pool;
3442 * Take control of the pool object.
3444 r = bind_control_target(pool, ti);
3445 if (r)
3446 return r;
3448 r = maybe_resize_data_dev(ti, &need_commit1);
3449 if (r)
3450 return r;
3452 r = maybe_resize_metadata_dev(ti, &need_commit2);
3453 if (r)
3454 return r;
3456 if (need_commit1 || need_commit2)
3457 (void) commit(pool);
3459 return 0;
3462 static void pool_suspend_active_thins(struct pool *pool)
3464 struct thin_c *tc;
3466 /* Suspend all active thin devices */
3467 tc = get_first_thin(pool);
3468 while (tc) {
3469 dm_internal_suspend_noflush(tc->thin_md);
3470 tc = get_next_thin(pool, tc);
3474 static void pool_resume_active_thins(struct pool *pool)
3476 struct thin_c *tc;
3478 /* Resume all active thin devices */
3479 tc = get_first_thin(pool);
3480 while (tc) {
3481 dm_internal_resume(tc->thin_md);
3482 tc = get_next_thin(pool, tc);
3486 static void pool_resume(struct dm_target *ti)
3488 struct pool_c *pt = ti->private;
3489 struct pool *pool = pt->pool;
3490 unsigned long flags;
3493 * Must requeue active_thins' bios and then resume
3494 * active_thins _before_ clearing 'suspend' flag.
3496 requeue_bios(pool);
3497 pool_resume_active_thins(pool);
3499 spin_lock_irqsave(&pool->lock, flags);
3500 pool->low_water_triggered = false;
3501 pool->suspended = false;
3502 spin_unlock_irqrestore(&pool->lock, flags);
3504 do_waker(&pool->waker.work);
3507 static void pool_presuspend(struct dm_target *ti)
3509 struct pool_c *pt = ti->private;
3510 struct pool *pool = pt->pool;
3511 unsigned long flags;
3513 spin_lock_irqsave(&pool->lock, flags);
3514 pool->suspended = true;
3515 spin_unlock_irqrestore(&pool->lock, flags);
3517 pool_suspend_active_thins(pool);
3520 static void pool_presuspend_undo(struct dm_target *ti)
3522 struct pool_c *pt = ti->private;
3523 struct pool *pool = pt->pool;
3524 unsigned long flags;
3526 pool_resume_active_thins(pool);
3528 spin_lock_irqsave(&pool->lock, flags);
3529 pool->suspended = false;
3530 spin_unlock_irqrestore(&pool->lock, flags);
3533 static void pool_postsuspend(struct dm_target *ti)
3535 struct pool_c *pt = ti->private;
3536 struct pool *pool = pt->pool;
3538 cancel_delayed_work_sync(&pool->waker);
3539 cancel_delayed_work_sync(&pool->no_space_timeout);
3540 flush_workqueue(pool->wq);
3541 (void) commit(pool);
3544 static int check_arg_count(unsigned argc, unsigned args_required)
3546 if (argc != args_required) {
3547 DMWARN("Message received with %u arguments instead of %u.",
3548 argc, args_required);
3549 return -EINVAL;
3552 return 0;
3555 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3557 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3558 *dev_id <= MAX_DEV_ID)
3559 return 0;
3561 if (warning)
3562 DMWARN("Message received with invalid device id: %s", arg);
3564 return -EINVAL;
3567 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3569 dm_thin_id dev_id;
3570 int r;
3572 r = check_arg_count(argc, 2);
3573 if (r)
3574 return r;
3576 r = read_dev_id(argv[1], &dev_id, 1);
3577 if (r)
3578 return r;
3580 r = dm_pool_create_thin(pool->pmd, dev_id);
3581 if (r) {
3582 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3583 argv[1]);
3584 return r;
3587 return 0;
3590 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3592 dm_thin_id dev_id;
3593 dm_thin_id origin_dev_id;
3594 int r;
3596 r = check_arg_count(argc, 3);
3597 if (r)
3598 return r;
3600 r = read_dev_id(argv[1], &dev_id, 1);
3601 if (r)
3602 return r;
3604 r = read_dev_id(argv[2], &origin_dev_id, 1);
3605 if (r)
3606 return r;
3608 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3609 if (r) {
3610 DMWARN("Creation of new snapshot %s of device %s failed.",
3611 argv[1], argv[2]);
3612 return r;
3615 return 0;
3618 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3620 dm_thin_id dev_id;
3621 int r;
3623 r = check_arg_count(argc, 2);
3624 if (r)
3625 return r;
3627 r = read_dev_id(argv[1], &dev_id, 1);
3628 if (r)
3629 return r;
3631 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3632 if (r)
3633 DMWARN("Deletion of thin device %s failed.", argv[1]);
3635 return r;
3638 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3640 dm_thin_id old_id, new_id;
3641 int r;
3643 r = check_arg_count(argc, 3);
3644 if (r)
3645 return r;
3647 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3648 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3649 return -EINVAL;
3652 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3653 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3654 return -EINVAL;
3657 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3658 if (r) {
3659 DMWARN("Failed to change transaction id from %s to %s.",
3660 argv[1], argv[2]);
3661 return r;
3664 return 0;
3667 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3669 int r;
3671 r = check_arg_count(argc, 1);
3672 if (r)
3673 return r;
3675 (void) commit(pool);
3677 r = dm_pool_reserve_metadata_snap(pool->pmd);
3678 if (r)
3679 DMWARN("reserve_metadata_snap message failed.");
3681 return r;
3684 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3686 int r;
3688 r = check_arg_count(argc, 1);
3689 if (r)
3690 return r;
3692 r = dm_pool_release_metadata_snap(pool->pmd);
3693 if (r)
3694 DMWARN("release_metadata_snap message failed.");
3696 return r;
3700 * Messages supported:
3701 * create_thin <dev_id>
3702 * create_snap <dev_id> <origin_id>
3703 * delete <dev_id>
3704 * set_transaction_id <current_trans_id> <new_trans_id>
3705 * reserve_metadata_snap
3706 * release_metadata_snap
3708 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3710 int r = -EINVAL;
3711 struct pool_c *pt = ti->private;
3712 struct pool *pool = pt->pool;
3714 if (get_pool_mode(pool) >= PM_READ_ONLY) {
3715 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3716 dm_device_name(pool->pool_md));
3717 return -EOPNOTSUPP;
3720 if (!strcasecmp(argv[0], "create_thin"))
3721 r = process_create_thin_mesg(argc, argv, pool);
3723 else if (!strcasecmp(argv[0], "create_snap"))
3724 r = process_create_snap_mesg(argc, argv, pool);
3726 else if (!strcasecmp(argv[0], "delete"))
3727 r = process_delete_mesg(argc, argv, pool);
3729 else if (!strcasecmp(argv[0], "set_transaction_id"))
3730 r = process_set_transaction_id_mesg(argc, argv, pool);
3732 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3733 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3735 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3736 r = process_release_metadata_snap_mesg(argc, argv, pool);
3738 else
3739 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3741 if (!r)
3742 (void) commit(pool);
3744 return r;
3747 static void emit_flags(struct pool_features *pf, char *result,
3748 unsigned sz, unsigned maxlen)
3750 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3751 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3752 pf->error_if_no_space;
3753 DMEMIT("%u ", count);
3755 if (!pf->zero_new_blocks)
3756 DMEMIT("skip_block_zeroing ");
3758 if (!pf->discard_enabled)
3759 DMEMIT("ignore_discard ");
3761 if (!pf->discard_passdown)
3762 DMEMIT("no_discard_passdown ");
3764 if (pf->mode == PM_READ_ONLY)
3765 DMEMIT("read_only ");
3767 if (pf->error_if_no_space)
3768 DMEMIT("error_if_no_space ");
3772 * Status line is:
3773 * <transaction id> <used metadata sectors>/<total metadata sectors>
3774 * <used data sectors>/<total data sectors> <held metadata root>
3775 * <pool mode> <discard config> <no space config> <needs_check>
3777 static void pool_status(struct dm_target *ti, status_type_t type,
3778 unsigned status_flags, char *result, unsigned maxlen)
3780 int r;
3781 unsigned sz = 0;
3782 uint64_t transaction_id;
3783 dm_block_t nr_free_blocks_data;
3784 dm_block_t nr_free_blocks_metadata;
3785 dm_block_t nr_blocks_data;
3786 dm_block_t nr_blocks_metadata;
3787 dm_block_t held_root;
3788 char buf[BDEVNAME_SIZE];
3789 char buf2[BDEVNAME_SIZE];
3790 struct pool_c *pt = ti->private;
3791 struct pool *pool = pt->pool;
3793 switch (type) {
3794 case STATUSTYPE_INFO:
3795 if (get_pool_mode(pool) == PM_FAIL) {
3796 DMEMIT("Fail");
3797 break;
3800 /* Commit to ensure statistics aren't out-of-date */
3801 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3802 (void) commit(pool);
3804 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3805 if (r) {
3806 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3807 dm_device_name(pool->pool_md), r);
3808 goto err;
3811 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3812 if (r) {
3813 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3814 dm_device_name(pool->pool_md), r);
3815 goto err;
3818 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3819 if (r) {
3820 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3821 dm_device_name(pool->pool_md), r);
3822 goto err;
3825 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3826 if (r) {
3827 DMERR("%s: dm_pool_get_free_block_count returned %d",
3828 dm_device_name(pool->pool_md), r);
3829 goto err;
3832 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3833 if (r) {
3834 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3835 dm_device_name(pool->pool_md), r);
3836 goto err;
3839 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3840 if (r) {
3841 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3842 dm_device_name(pool->pool_md), r);
3843 goto err;
3846 DMEMIT("%llu %llu/%llu %llu/%llu ",
3847 (unsigned long long)transaction_id,
3848 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3849 (unsigned long long)nr_blocks_metadata,
3850 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3851 (unsigned long long)nr_blocks_data);
3853 if (held_root)
3854 DMEMIT("%llu ", held_root);
3855 else
3856 DMEMIT("- ");
3858 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3859 DMEMIT("out_of_data_space ");
3860 else if (pool->pf.mode == PM_READ_ONLY)
3861 DMEMIT("ro ");
3862 else
3863 DMEMIT("rw ");
3865 if (!pool->pf.discard_enabled)
3866 DMEMIT("ignore_discard ");
3867 else if (pool->pf.discard_passdown)
3868 DMEMIT("discard_passdown ");
3869 else
3870 DMEMIT("no_discard_passdown ");
3872 if (pool->pf.error_if_no_space)
3873 DMEMIT("error_if_no_space ");
3874 else
3875 DMEMIT("queue_if_no_space ");
3877 if (dm_pool_metadata_needs_check(pool->pmd))
3878 DMEMIT("needs_check ");
3879 else
3880 DMEMIT("- ");
3882 break;
3884 case STATUSTYPE_TABLE:
3885 DMEMIT("%s %s %lu %llu ",
3886 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3887 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3888 (unsigned long)pool->sectors_per_block,
3889 (unsigned long long)pt->low_water_blocks);
3890 emit_flags(&pt->requested_pf, result, sz, maxlen);
3891 break;
3893 return;
3895 err:
3896 DMEMIT("Error");
3899 static int pool_iterate_devices(struct dm_target *ti,
3900 iterate_devices_callout_fn fn, void *data)
3902 struct pool_c *pt = ti->private;
3904 return fn(ti, pt->data_dev, 0, ti->len, data);
3907 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3909 struct pool_c *pt = ti->private;
3910 struct pool *pool = pt->pool;
3911 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3914 * If max_sectors is smaller than pool->sectors_per_block adjust it
3915 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3916 * This is especially beneficial when the pool's data device is a RAID
3917 * device that has a full stripe width that matches pool->sectors_per_block
3918 * -- because even though partial RAID stripe-sized IOs will be issued to a
3919 * single RAID stripe; when aggregated they will end on a full RAID stripe
3920 * boundary.. which avoids additional partial RAID stripe writes cascading
3922 if (limits->max_sectors < pool->sectors_per_block) {
3923 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3924 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3925 limits->max_sectors--;
3926 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3931 * If the system-determined stacked limits are compatible with the
3932 * pool's blocksize (io_opt is a factor) do not override them.
3934 if (io_opt_sectors < pool->sectors_per_block ||
3935 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3936 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3937 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3938 else
3939 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3940 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3944 * pt->adjusted_pf is a staging area for the actual features to use.
3945 * They get transferred to the live pool in bind_control_target()
3946 * called from pool_preresume().
3948 if (!pt->adjusted_pf.discard_enabled) {
3950 * Must explicitly disallow stacking discard limits otherwise the
3951 * block layer will stack them if pool's data device has support.
3952 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3953 * user to see that, so make sure to set all discard limits to 0.
3955 limits->discard_granularity = 0;
3956 return;
3959 disable_passdown_if_not_supported(pt);
3962 * The pool uses the same discard limits as the underlying data
3963 * device. DM core has already set this up.
3967 static struct target_type pool_target = {
3968 .name = "thin-pool",
3969 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3970 DM_TARGET_IMMUTABLE,
3971 .version = {1, 19, 0},
3972 .module = THIS_MODULE,
3973 .ctr = pool_ctr,
3974 .dtr = pool_dtr,
3975 .map = pool_map,
3976 .presuspend = pool_presuspend,
3977 .presuspend_undo = pool_presuspend_undo,
3978 .postsuspend = pool_postsuspend,
3979 .preresume = pool_preresume,
3980 .resume = pool_resume,
3981 .message = pool_message,
3982 .status = pool_status,
3983 .iterate_devices = pool_iterate_devices,
3984 .io_hints = pool_io_hints,
3987 /*----------------------------------------------------------------
3988 * Thin target methods
3989 *--------------------------------------------------------------*/
3990 static void thin_get(struct thin_c *tc)
3992 atomic_inc(&tc->refcount);
3995 static void thin_put(struct thin_c *tc)
3997 if (atomic_dec_and_test(&tc->refcount))
3998 complete(&tc->can_destroy);
4001 static void thin_dtr(struct dm_target *ti)
4003 struct thin_c *tc = ti->private;
4004 unsigned long flags;
4006 spin_lock_irqsave(&tc->pool->lock, flags);
4007 list_del_rcu(&tc->list);
4008 spin_unlock_irqrestore(&tc->pool->lock, flags);
4009 synchronize_rcu();
4011 thin_put(tc);
4012 wait_for_completion(&tc->can_destroy);
4014 mutex_lock(&dm_thin_pool_table.mutex);
4016 __pool_dec(tc->pool);
4017 dm_pool_close_thin_device(tc->td);
4018 dm_put_device(ti, tc->pool_dev);
4019 if (tc->origin_dev)
4020 dm_put_device(ti, tc->origin_dev);
4021 kfree(tc);
4023 mutex_unlock(&dm_thin_pool_table.mutex);
4027 * Thin target parameters:
4029 * <pool_dev> <dev_id> [origin_dev]
4031 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4032 * dev_id: the internal device identifier
4033 * origin_dev: a device external to the pool that should act as the origin
4035 * If the pool device has discards disabled, they get disabled for the thin
4036 * device as well.
4038 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4040 int r;
4041 struct thin_c *tc;
4042 struct dm_dev *pool_dev, *origin_dev;
4043 struct mapped_device *pool_md;
4044 unsigned long flags;
4046 mutex_lock(&dm_thin_pool_table.mutex);
4048 if (argc != 2 && argc != 3) {
4049 ti->error = "Invalid argument count";
4050 r = -EINVAL;
4051 goto out_unlock;
4054 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4055 if (!tc) {
4056 ti->error = "Out of memory";
4057 r = -ENOMEM;
4058 goto out_unlock;
4060 tc->thin_md = dm_table_get_md(ti->table);
4061 spin_lock_init(&tc->lock);
4062 INIT_LIST_HEAD(&tc->deferred_cells);
4063 bio_list_init(&tc->deferred_bio_list);
4064 bio_list_init(&tc->retry_on_resume_list);
4065 tc->sort_bio_list = RB_ROOT;
4067 if (argc == 3) {
4068 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4069 if (r) {
4070 ti->error = "Error opening origin device";
4071 goto bad_origin_dev;
4073 tc->origin_dev = origin_dev;
4076 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4077 if (r) {
4078 ti->error = "Error opening pool device";
4079 goto bad_pool_dev;
4081 tc->pool_dev = pool_dev;
4083 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4084 ti->error = "Invalid device id";
4085 r = -EINVAL;
4086 goto bad_common;
4089 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4090 if (!pool_md) {
4091 ti->error = "Couldn't get pool mapped device";
4092 r = -EINVAL;
4093 goto bad_common;
4096 tc->pool = __pool_table_lookup(pool_md);
4097 if (!tc->pool) {
4098 ti->error = "Couldn't find pool object";
4099 r = -EINVAL;
4100 goto bad_pool_lookup;
4102 __pool_inc(tc->pool);
4104 if (get_pool_mode(tc->pool) == PM_FAIL) {
4105 ti->error = "Couldn't open thin device, Pool is in fail mode";
4106 r = -EINVAL;
4107 goto bad_pool;
4110 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4111 if (r) {
4112 ti->error = "Couldn't open thin internal device";
4113 goto bad_pool;
4116 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4117 if (r)
4118 goto bad;
4120 ti->num_flush_bios = 1;
4121 ti->flush_supported = true;
4122 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4124 /* In case the pool supports discards, pass them on. */
4125 ti->discard_zeroes_data_unsupported = true;
4126 if (tc->pool->pf.discard_enabled) {
4127 ti->discards_supported = true;
4128 ti->num_discard_bios = 1;
4129 ti->split_discard_bios = false;
4132 mutex_unlock(&dm_thin_pool_table.mutex);
4134 spin_lock_irqsave(&tc->pool->lock, flags);
4135 if (tc->pool->suspended) {
4136 spin_unlock_irqrestore(&tc->pool->lock, flags);
4137 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4138 ti->error = "Unable to activate thin device while pool is suspended";
4139 r = -EINVAL;
4140 goto bad;
4142 atomic_set(&tc->refcount, 1);
4143 init_completion(&tc->can_destroy);
4144 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4145 spin_unlock_irqrestore(&tc->pool->lock, flags);
4147 * This synchronize_rcu() call is needed here otherwise we risk a
4148 * wake_worker() call finding no bios to process (because the newly
4149 * added tc isn't yet visible). So this reduces latency since we
4150 * aren't then dependent on the periodic commit to wake_worker().
4152 synchronize_rcu();
4154 dm_put(pool_md);
4156 return 0;
4158 bad:
4159 dm_pool_close_thin_device(tc->td);
4160 bad_pool:
4161 __pool_dec(tc->pool);
4162 bad_pool_lookup:
4163 dm_put(pool_md);
4164 bad_common:
4165 dm_put_device(ti, tc->pool_dev);
4166 bad_pool_dev:
4167 if (tc->origin_dev)
4168 dm_put_device(ti, tc->origin_dev);
4169 bad_origin_dev:
4170 kfree(tc);
4171 out_unlock:
4172 mutex_unlock(&dm_thin_pool_table.mutex);
4174 return r;
4177 static int thin_map(struct dm_target *ti, struct bio *bio)
4179 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4181 return thin_bio_map(ti, bio);
4184 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
4186 unsigned long flags;
4187 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4188 struct list_head work;
4189 struct dm_thin_new_mapping *m, *tmp;
4190 struct pool *pool = h->tc->pool;
4192 if (h->shared_read_entry) {
4193 INIT_LIST_HEAD(&work);
4194 dm_deferred_entry_dec(h->shared_read_entry, &work);
4196 spin_lock_irqsave(&pool->lock, flags);
4197 list_for_each_entry_safe(m, tmp, &work, list) {
4198 list_del(&m->list);
4199 __complete_mapping_preparation(m);
4201 spin_unlock_irqrestore(&pool->lock, flags);
4204 if (h->all_io_entry) {
4205 INIT_LIST_HEAD(&work);
4206 dm_deferred_entry_dec(h->all_io_entry, &work);
4207 if (!list_empty(&work)) {
4208 spin_lock_irqsave(&pool->lock, flags);
4209 list_for_each_entry_safe(m, tmp, &work, list)
4210 list_add_tail(&m->list, &pool->prepared_discards);
4211 spin_unlock_irqrestore(&pool->lock, flags);
4212 wake_worker(pool);
4216 if (h->cell)
4217 cell_defer_no_holder(h->tc, h->cell);
4219 return 0;
4222 static void thin_presuspend(struct dm_target *ti)
4224 struct thin_c *tc = ti->private;
4226 if (dm_noflush_suspending(ti))
4227 noflush_work(tc, do_noflush_start);
4230 static void thin_postsuspend(struct dm_target *ti)
4232 struct thin_c *tc = ti->private;
4235 * The dm_noflush_suspending flag has been cleared by now, so
4236 * unfortunately we must always run this.
4238 noflush_work(tc, do_noflush_stop);
4241 static int thin_preresume(struct dm_target *ti)
4243 struct thin_c *tc = ti->private;
4245 if (tc->origin_dev)
4246 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4248 return 0;
4252 * <nr mapped sectors> <highest mapped sector>
4254 static void thin_status(struct dm_target *ti, status_type_t type,
4255 unsigned status_flags, char *result, unsigned maxlen)
4257 int r;
4258 ssize_t sz = 0;
4259 dm_block_t mapped, highest;
4260 char buf[BDEVNAME_SIZE];
4261 struct thin_c *tc = ti->private;
4263 if (get_pool_mode(tc->pool) == PM_FAIL) {
4264 DMEMIT("Fail");
4265 return;
4268 if (!tc->td)
4269 DMEMIT("-");
4270 else {
4271 switch (type) {
4272 case STATUSTYPE_INFO:
4273 r = dm_thin_get_mapped_count(tc->td, &mapped);
4274 if (r) {
4275 DMERR("dm_thin_get_mapped_count returned %d", r);
4276 goto err;
4279 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4280 if (r < 0) {
4281 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4282 goto err;
4285 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4286 if (r)
4287 DMEMIT("%llu", ((highest + 1) *
4288 tc->pool->sectors_per_block) - 1);
4289 else
4290 DMEMIT("-");
4291 break;
4293 case STATUSTYPE_TABLE:
4294 DMEMIT("%s %lu",
4295 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4296 (unsigned long) tc->dev_id);
4297 if (tc->origin_dev)
4298 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4299 break;
4303 return;
4305 err:
4306 DMEMIT("Error");
4309 static int thin_iterate_devices(struct dm_target *ti,
4310 iterate_devices_callout_fn fn, void *data)
4312 sector_t blocks;
4313 struct thin_c *tc = ti->private;
4314 struct pool *pool = tc->pool;
4317 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4318 * we follow a more convoluted path through to the pool's target.
4320 if (!pool->ti)
4321 return 0; /* nothing is bound */
4323 blocks = pool->ti->len;
4324 (void) sector_div(blocks, pool->sectors_per_block);
4325 if (blocks)
4326 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4328 return 0;
4331 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4333 struct thin_c *tc = ti->private;
4334 struct pool *pool = tc->pool;
4336 if (!pool->pf.discard_enabled)
4337 return;
4339 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4340 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4343 static struct target_type thin_target = {
4344 .name = "thin",
4345 .version = {1, 19, 0},
4346 .module = THIS_MODULE,
4347 .ctr = thin_ctr,
4348 .dtr = thin_dtr,
4349 .map = thin_map,
4350 .end_io = thin_endio,
4351 .preresume = thin_preresume,
4352 .presuspend = thin_presuspend,
4353 .postsuspend = thin_postsuspend,
4354 .status = thin_status,
4355 .iterate_devices = thin_iterate_devices,
4356 .io_hints = thin_io_hints,
4359 /*----------------------------------------------------------------*/
4361 static int __init dm_thin_init(void)
4363 int r;
4365 pool_table_init();
4367 r = dm_register_target(&thin_target);
4368 if (r)
4369 return r;
4371 r = dm_register_target(&pool_target);
4372 if (r)
4373 goto bad_pool_target;
4375 r = -ENOMEM;
4377 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4378 if (!_new_mapping_cache)
4379 goto bad_new_mapping_cache;
4381 return 0;
4383 bad_new_mapping_cache:
4384 dm_unregister_target(&pool_target);
4385 bad_pool_target:
4386 dm_unregister_target(&thin_target);
4388 return r;
4391 static void dm_thin_exit(void)
4393 dm_unregister_target(&thin_target);
4394 dm_unregister_target(&pool_target);
4396 kmem_cache_destroy(_new_mapping_cache);
4399 module_init(dm_thin_init);
4400 module_exit(dm_thin_exit);
4402 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4403 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4405 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4406 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4407 MODULE_LICENSE("GPL");