1/*
2 * Copyright (C) 2011-2012 Red Hat UK.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm-thin-metadata.h"
8#include "dm-bio-prison-v1.h"
9#include "dm.h"
10
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>
24
25#define DM_MSG_PREFIX "thin"
26
27/*
28 * Tunable constants
29 */
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
34
35static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36
37DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
38 "A percentage of time allocated for copy on write");
39
40/*
41 * The block size of the device holding pool data must be
42 * between 64KB and 1GB.
43 */
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)
46
47/*
48 * Device id is restricted to 24 bits.
49 */
50#define MAX_DEV_ID ((1 << 24) - 1)
51
52/*
53 * How do we handle breaking sharing of data blocks?
54 * =================================================
55 *
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.
62 *
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.
65 *
66 * Let's say we write to a shared block in what was the origin. The
67 * steps are:
68 *
69 * i) plug io further to this physical block. (see bio_prison code).
70 *
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)
73 *
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).
76 *
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.
84 *
85 * v) unplug io to this physical block, including the io that triggered
86 * the breaking of sharing.
87 *
88 * Steps (ii) and (iii) occur in parallel.
89 *
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:
93 *
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.
97 *
98 * - The snap mapping still points to the old block. As it would after
99 * the commit.
100 *
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.
108 */
109
110/*----------------------------------------------------------------*/
111
112/*
113 * Key building.
114 */
115enum lock_space {
116 VIRTUAL,
117 PHYSICAL
118};
119
120static 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)
122{
123 key->virtual = (ls == VIRTUAL);
124 key->dev = dm_thin_dev_id(td);
125 key->block_begin = b;
126 key->block_end = e;
127}
128
129static void build_data_key(struct dm_thin_device *td, dm_block_t b,
130 struct dm_cell_key *key)
131{
132 build_key(td, PHYSICAL, b, b + 1llu, key);
133}
134
135static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
136 struct dm_cell_key *key)
137{
138 build_key(td, VIRTUAL, b, b + 1llu, key);
139}
140
141/*----------------------------------------------------------------*/
142
143#define THROTTLE_THRESHOLD (1 * HZ)
144
145struct throttle {
146 struct rw_semaphore lock;
147 unsigned long threshold;
148 bool throttle_applied;
149};
150
151static void throttle_init(struct throttle *t)
152{
153 init_rwsem(&t->lock);
154 t->throttle_applied = false;
155}
156
157static void throttle_work_start(struct throttle *t)
158{
159 t->threshold = jiffies + THROTTLE_THRESHOLD;
160}
161
162static void throttle_work_update(struct throttle *t)
163{
164 if (!t->throttle_applied && jiffies > t->threshold) {
165 down_write(&t->lock);
166 t->throttle_applied = true;
167 }
168}
169
170static void throttle_work_complete(struct throttle *t)
171{
172 if (t->throttle_applied) {
173 t->throttle_applied = false;
174 up_write(&t->lock);
175 }
176}
177
178static void throttle_lock(struct throttle *t)
179{
180 down_read(&t->lock);
181}
182
183static void throttle_unlock(struct throttle *t)
184{
185 up_read(&t->lock);
186}
187
188/*----------------------------------------------------------------*/
189
190/*
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.
194 */
195struct dm_thin_new_mapping;
196
197/*
198 * The pool runs in various modes. Ordered in degraded order for comparisons.
199 */
200enum 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
204 /*
205 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
206 */
207 PM_OUT_OF_METADATA_SPACE,
208 PM_READ_ONLY, /* metadata may not be changed */
209
210 PM_FAIL, /* all I/O fails */
211};
212
213struct pool_features {
214 enum pool_mode mode;
215
216 bool zero_new_blocks:1;
217 bool discard_enabled:1;
218 bool discard_passdown:1;
219 bool error_if_no_space:1;
220};
221
222struct thin_c;
223typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
224typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
225typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
226
227#define CELL_SORT_ARRAY_SIZE 8192
228
229struct pool {
230 struct list_head list;
231 struct dm_target *ti; /* Only set if a pool target is bound */
232
233 struct mapped_device *pool_md;
234 struct block_device *md_dev;
235 struct dm_pool_metadata *pmd;
236
237 dm_block_t low_water_blocks;
238 uint32_t sectors_per_block;
239 int sectors_per_block_shift;
240
241 struct pool_features pf;
242 bool low_water_triggered:1; /* A dm event has been sent */
243 bool suspended:1;
244 bool out_of_data_space:1;
245
246 struct dm_bio_prison *prison;
247 struct dm_kcopyd_client *copier;
248
249 struct work_struct worker;
250 struct workqueue_struct *wq;
251 struct throttle throttle;
252 struct delayed_work waker;
253 struct delayed_work no_space_timeout;
254
255 unsigned long last_commit_jiffies;
256 unsigned ref_count;
257
258 spinlock_t lock;
259 struct bio_list deferred_flush_bios;
260 struct bio_list deferred_flush_completions;
261 struct list_head prepared_mappings;
262 struct list_head prepared_discards;
263 struct list_head prepared_discards_pt2;
264 struct list_head active_thins;
265
266 struct dm_deferred_set *shared_read_ds;
267 struct dm_deferred_set *all_io_ds;
268
269 struct dm_thin_new_mapping *next_mapping;
270
271 process_bio_fn process_bio;
272 process_bio_fn process_discard;
273
274 process_cell_fn process_cell;
275 process_cell_fn process_discard_cell;
276
277 process_mapping_fn process_prepared_mapping;
278 process_mapping_fn process_prepared_discard;
279 process_mapping_fn process_prepared_discard_pt2;
280
281 struct dm_bio_prison_cell **cell_sort_array;
282
283 mempool_t mapping_pool;
284};
285
286static void metadata_operation_failed(struct pool *pool, const char *op, int r);
287
288static enum pool_mode get_pool_mode(struct pool *pool)
289{
290 return pool->pf.mode;
291}
292
293static void notify_of_pool_mode_change(struct pool *pool)
294{
295 const char *descs[] = {
296 "write",
297 "out-of-data-space",
298 "read-only",
299 "read-only",
300 "fail"
301 };
302 const char *extra_desc = NULL;
303 enum pool_mode mode = get_pool_mode(pool);
304
305 if (mode == PM_OUT_OF_DATA_SPACE) {
306 if (!pool->pf.error_if_no_space)
307 extra_desc = " (queue IO)";
308 else
309 extra_desc = " (error IO)";
310 }
311
312 dm_table_event(pool->ti->table);
313 DMINFO("%s: switching pool to %s%s mode",
314 dm_device_name(pool->pool_md),
315 descs[(int)mode], extra_desc ? : "");
316}
317
318/*
319 * Target context for a pool.
320 */
321struct pool_c {
322 struct dm_target *ti;
323 struct pool *pool;
324 struct dm_dev *data_dev;
325 struct dm_dev *metadata_dev;
326 struct dm_target_callbacks callbacks;
327
328 dm_block_t low_water_blocks;
329 struct pool_features requested_pf; /* Features requested during table load */
330 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
331};
332
333/*
334 * Target context for a thin.
335 */
336struct thin_c {
337 struct list_head list;
338 struct dm_dev *pool_dev;
339 struct dm_dev *origin_dev;
340 sector_t origin_size;
341 dm_thin_id dev_id;
342
343 struct pool *pool;
344 struct dm_thin_device *td;
345 struct mapped_device *thin_md;
346
347 bool requeue_mode:1;
348 spinlock_t lock;
349 struct list_head deferred_cells;
350 struct bio_list deferred_bio_list;
351 struct bio_list retry_on_resume_list;
352 struct rb_root sort_bio_list; /* sorted list of deferred bios */
353
354 /*
355 * Ensures the thin is not destroyed until the worker has finished
356 * iterating the active_thins list.
357 */
358 refcount_t refcount;
359 struct completion can_destroy;
360};
361
362/*----------------------------------------------------------------*/
363
364static bool block_size_is_power_of_two(struct pool *pool)
365{
366 return pool->sectors_per_block_shift >= 0;
367}
368
369static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
370{
371 return block_size_is_power_of_two(pool) ?
372 (b << pool->sectors_per_block_shift) :
373 (b * pool->sectors_per_block);
374}
375
376/*----------------------------------------------------------------*/
377
378struct discard_op {
379 struct thin_c *tc;
380 struct blk_plug plug;
381 struct bio *parent_bio;
382 struct bio *bio;
383};
384
385static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
386{
387 BUG_ON(!parent);
388
389 op->tc = tc;
390 blk_start_plug(&op->plug);
391 op->parent_bio = parent;
392 op->bio = NULL;
393}
394
395static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
396{
397 struct thin_c *tc = op->tc;
398 sector_t s = block_to_sectors(tc->pool, data_b);
399 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
400
401 return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
402 GFP_NOWAIT, 0, &op->bio);
403}
404
405static void end_discard(struct discard_op *op, int r)
406{
407 if (op->bio) {
408 /*
409 * Even if one of the calls to issue_discard failed, we
410 * need to wait for the chain to complete.
411 */
412 bio_chain(op->bio, op->parent_bio);
413 bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
414 submit_bio(op->bio);
415 }
416
417 blk_finish_plug(&op->plug);
418
419 /*
420 * Even if r is set, there could be sub discards in flight that we
421 * need to wait for.
422 */
423 if (r && !op->parent_bio->bi_status)
424 op->parent_bio->bi_status = errno_to_blk_status(r);
425 bio_endio(op->parent_bio);
426}
427
428/*----------------------------------------------------------------*/
429
430/*
431 * wake_worker() is used when new work is queued and when pool_resume is
432 * ready to continue deferred IO processing.
433 */
434static void wake_worker(struct pool *pool)
435{
436 queue_work(pool->wq, &pool->worker);
437}
438
439/*----------------------------------------------------------------*/
440
441static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
442 struct dm_bio_prison_cell **cell_result)
443{
444 int r;
445 struct dm_bio_prison_cell *cell_prealloc;
446
447 /*
448 * Allocate a cell from the prison's mempool.
449 * This might block but it can't fail.
450 */
451 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
452
453 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
454 if (r)
455 /*
456 * We reused an old cell; we can get rid of
457 * the new one.
458 */
459 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
460
461 return r;
462}
463
464static void cell_release(struct pool *pool,
465 struct dm_bio_prison_cell *cell,
466 struct bio_list *bios)
467{
468 dm_cell_release(pool->prison, cell, bios);
469 dm_bio_prison_free_cell(pool->prison, cell);
470}
471
472static void cell_visit_release(struct pool *pool,
473 void (*fn)(void *, struct dm_bio_prison_cell *),
474 void *context,
475 struct dm_bio_prison_cell *cell)
476{
477 dm_cell_visit_release(pool->prison, fn, context, cell);
478 dm_bio_prison_free_cell(pool->prison, cell);
479}
480
481static void cell_release_no_holder(struct pool *pool,
482 struct dm_bio_prison_cell *cell,
483 struct bio_list *bios)
484{
485 dm_cell_release_no_holder(pool->prison, cell, bios);
486 dm_bio_prison_free_cell(pool->prison, cell);
487}
488
489static void cell_error_with_code(struct pool *pool,
490 struct dm_bio_prison_cell *cell, blk_status_t error_code)
491{
492 dm_cell_error(pool->prison, cell, error_code);
493 dm_bio_prison_free_cell(pool->prison, cell);
494}
495
496static blk_status_t get_pool_io_error_code(struct pool *pool)
497{
498 return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
499}
500
501static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
502{
503 cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
504}
505
506static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
507{
508 cell_error_with_code(pool, cell, 0);
509}
510
511static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
512{
513 cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
514}
515
516/*----------------------------------------------------------------*/
517
518/*
519 * A global list of pools that uses a struct mapped_device as a key.
520 */
521static struct dm_thin_pool_table {
522 struct mutex mutex;
523 struct list_head pools;
524} dm_thin_pool_table;
525
526static void pool_table_init(void)
527{
528 mutex_init(&dm_thin_pool_table.mutex);
529 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
530}
531
532static void pool_table_exit(void)
533{
534 mutex_destroy(&dm_thin_pool_table.mutex);
535}
536
537static void __pool_table_insert(struct pool *pool)
538{
539 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
540 list_add(&pool->list, &dm_thin_pool_table.pools);
541}
542
543static void __pool_table_remove(struct pool *pool)
544{
545 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
546 list_del(&pool->list);
547}
548
549static struct pool *__pool_table_lookup(struct mapped_device *md)
550{
551 struct pool *pool = NULL, *tmp;
552
553 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
554
555 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
556 if (tmp->pool_md == md) {
557 pool = tmp;
558 break;
559 }
560 }
561
562 return pool;
563}
564
565static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
566{
567 struct pool *pool = NULL, *tmp;
568
569 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
570
571 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
572 if (tmp->md_dev == md_dev) {
573 pool = tmp;
574 break;
575 }
576 }
577
578 return pool;
579}
580
581/*----------------------------------------------------------------*/
582
583struct dm_thin_endio_hook {
584 struct thin_c *tc;
585 struct dm_deferred_entry *shared_read_entry;
586 struct dm_deferred_entry *all_io_entry;
587 struct dm_thin_new_mapping *overwrite_mapping;
588 struct rb_node rb_node;
589 struct dm_bio_prison_cell *cell;
590};
591
592static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
593{
594 bio_list_merge(bios, master);
595 bio_list_init(master);
596}
597
598static void error_bio_list(struct bio_list *bios, blk_status_t error)
599{
600 struct bio *bio;
601
602 while ((bio = bio_list_pop(bios))) {
603 bio->bi_status = error;
604 bio_endio(bio);
605 }
606}
607
608static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
609 blk_status_t error)
610{
611 struct bio_list bios;
612 unsigned long flags;
613
614 bio_list_init(&bios);
615
616 spin_lock_irqsave(&tc->lock, flags);
617 __merge_bio_list(&bios, master);
618 spin_unlock_irqrestore(&tc->lock, flags);
619
620 error_bio_list(&bios, error);
621}
622
623static void requeue_deferred_cells(struct thin_c *tc)
624{
625 struct pool *pool = tc->pool;
626 unsigned long flags;
627 struct list_head cells;
628 struct dm_bio_prison_cell *cell, *tmp;
629
630 INIT_LIST_HEAD(&cells);
631
632 spin_lock_irqsave(&tc->lock, flags);
633 list_splice_init(&tc->deferred_cells, &cells);
634 spin_unlock_irqrestore(&tc->lock, flags);
635
636 list_for_each_entry_safe(cell, tmp, &cells, user_list)
637 cell_requeue(pool, cell);
638}
639
640static void requeue_io(struct thin_c *tc)
641{
642 struct bio_list bios;
643 unsigned long flags;
644
645 bio_list_init(&bios);
646
647 spin_lock_irqsave(&tc->lock, flags);
648 __merge_bio_list(&bios, &tc->deferred_bio_list);
649 __merge_bio_list(&bios, &tc->retry_on_resume_list);
650 spin_unlock_irqrestore(&tc->lock, flags);
651
652 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
653 requeue_deferred_cells(tc);
654}
655
656static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
657{
658 struct thin_c *tc;
659
660 rcu_read_lock();
661 list_for_each_entry_rcu(tc, &pool->active_thins, list)
662 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
663 rcu_read_unlock();
664}
665
666static void error_retry_list(struct pool *pool)
667{
668 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
669}
670
671/*
672 * This section of code contains the logic for processing a thin device's IO.
673 * Much of the code depends on pool object resources (lists, workqueues, etc)
674 * but most is exclusively called from the thin target rather than the thin-pool
675 * target.
676 */
677
678static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
679{
680 struct pool *pool = tc->pool;
681 sector_t block_nr = bio->bi_iter.bi_sector;
682
683 if (block_size_is_power_of_two(pool))
684 block_nr >>= pool->sectors_per_block_shift;
685 else
686 (void) sector_div(block_nr, pool->sectors_per_block);
687
688 return block_nr;
689}
690
691/*
692 * Returns the _complete_ blocks that this bio covers.
693 */
694static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
695 dm_block_t *begin, dm_block_t *end)
696{
697 struct pool *pool = tc->pool;
698 sector_t b = bio->bi_iter.bi_sector;
699 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
700
701 b += pool->sectors_per_block - 1ull; /* so we round up */
702
703 if (block_size_is_power_of_two(pool)) {
704 b >>= pool->sectors_per_block_shift;
705 e >>= pool->sectors_per_block_shift;
706 } else {
707 (void) sector_div(b, pool->sectors_per_block);
708 (void) sector_div(e, pool->sectors_per_block);
709 }
710
711 if (e < b)
712 /* Can happen if the bio is within a single block. */
713 e = b;
714
715 *begin = b;
716 *end = e;
717}
718
719static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
720{
721 struct pool *pool = tc->pool;
722 sector_t bi_sector = bio->bi_iter.bi_sector;
723
724 bio_set_dev(bio, tc->pool_dev->bdev);
725 if (block_size_is_power_of_two(pool))
726 bio->bi_iter.bi_sector =
727 (block << pool->sectors_per_block_shift) |
728 (bi_sector & (pool->sectors_per_block - 1));
729 else
730 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
731 sector_div(bi_sector, pool->sectors_per_block);
732}
733
734static void remap_to_origin(struct thin_c *tc, struct bio *bio)
735{
736 bio_set_dev(bio, tc->origin_dev->bdev);
737}
738
739static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
740{
741 return op_is_flush(bio->bi_opf) &&
742 dm_thin_changed_this_transaction(tc->td);
743}
744
745static void inc_all_io_entry(struct pool *pool, struct bio *bio)
746{
747 struct dm_thin_endio_hook *h;
748
749 if (bio_op(bio) == REQ_OP_DISCARD)
750 return;
751
752 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
753 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
754}
755
756static void issue(struct thin_c *tc, struct bio *bio)
757{
758 struct pool *pool = tc->pool;
759 unsigned long flags;
760
761 if (!bio_triggers_commit(tc, bio)) {
762 generic_make_request(bio);
763 return;
764 }
765
766 /*
767 * Complete bio with an error if earlier I/O caused changes to
768 * the metadata that can't be committed e.g, due to I/O errors
769 * on the metadata device.
770 */
771 if (dm_thin_aborted_changes(tc->td)) {
772 bio_io_error(bio);
773 return;
774 }
775
776 /*
777 * Batch together any bios that trigger commits and then issue a
778 * single commit for them in process_deferred_bios().
779 */
780 spin_lock_irqsave(&pool->lock, flags);
781 bio_list_add(&pool->deferred_flush_bios, bio);
782 spin_unlock_irqrestore(&pool->lock, flags);
783}
784
785static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
786{
787 remap_to_origin(tc, bio);
788 issue(tc, bio);
789}
790
791static void remap_and_issue(struct thin_c *tc, struct bio *bio,
792 dm_block_t block)
793{
794 remap(tc, bio, block);
795 issue(tc, bio);
796}
797
798/*----------------------------------------------------------------*/
799
800/*
801 * Bio endio functions.
802 */
803struct dm_thin_new_mapping {
804 struct list_head list;
805
806 bool pass_discard:1;
807 bool maybe_shared:1;
808
809 /*
810 * Track quiescing, copying and zeroing preparation actions. When this
811 * counter hits zero the block is prepared and can be inserted into the
812 * btree.
813 */
814 atomic_t prepare_actions;
815
816 blk_status_t status;
817 struct thin_c *tc;
818 dm_block_t virt_begin, virt_end;
819 dm_block_t data_block;
820 struct dm_bio_prison_cell *cell;
821
822 /*
823 * If the bio covers the whole area of a block then we can avoid
824 * zeroing or copying. Instead this bio is hooked. The bio will
825 * still be in the cell, so care has to be taken to avoid issuing
826 * the bio twice.
827 */
828 struct bio *bio;
829 bio_end_io_t *saved_bi_end_io;
830};
831
832static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
833{
834 struct pool *pool = m->tc->pool;
835
836 if (atomic_dec_and_test(&m->prepare_actions)) {
837 list_add_tail(&m->list, &pool->prepared_mappings);
838 wake_worker(pool);
839 }
840}
841
842static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
843{
844 unsigned long flags;
845 struct pool *pool = m->tc->pool;
846
847 spin_lock_irqsave(&pool->lock, flags);
848 __complete_mapping_preparation(m);
849 spin_unlock_irqrestore(&pool->lock, flags);
850}
851
852static void copy_complete(int read_err, unsigned long write_err, void *context)
853{
854 struct dm_thin_new_mapping *m = context;
855
856 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
857 complete_mapping_preparation(m);
858}
859
860static void overwrite_endio(struct bio *bio)
861{
862 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
863 struct dm_thin_new_mapping *m = h->overwrite_mapping;
864
865 bio->bi_end_io = m->saved_bi_end_io;
866
867 m->status = bio->bi_status;
868 complete_mapping_preparation(m);
869}
870
871/*----------------------------------------------------------------*/
872
873/*
874 * Workqueue.
875 */
876
877/*
878 * Prepared mapping jobs.
879 */
880
881/*
882 * This sends the bios in the cell, except the original holder, back
883 * to the deferred_bios list.
884 */
885static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
886{
887 struct pool *pool = tc->pool;
888 unsigned long flags;
889
890 spin_lock_irqsave(&tc->lock, flags);
891 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
892 spin_unlock_irqrestore(&tc->lock, flags);
893
894 wake_worker(pool);
895}
896
897static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
898
899struct remap_info {
900 struct thin_c *tc;
901 struct bio_list defer_bios;
902 struct bio_list issue_bios;
903};
904
905static void __inc_remap_and_issue_cell(void *context,
906 struct dm_bio_prison_cell *cell)
907{
908 struct remap_info *info = context;
909 struct bio *bio;
910
911 while ((bio = bio_list_pop(&cell->bios))) {
912 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
913 bio_list_add(&info->defer_bios, bio);
914 else {
915 inc_all_io_entry(info->tc->pool, bio);
916
917 /*
918 * We can't issue the bios with the bio prison lock
919 * held, so we add them to a list to issue on
920 * return from this function.
921 */
922 bio_list_add(&info->issue_bios, bio);
923 }
924 }
925}
926
927static void inc_remap_and_issue_cell(struct thin_c *tc,
928 struct dm_bio_prison_cell *cell,
929 dm_block_t block)
930{
931 struct bio *bio;
932 struct remap_info info;
933
934 info.tc = tc;
935 bio_list_init(&info.defer_bios);
936 bio_list_init(&info.issue_bios);
937
938 /*
939 * We have to be careful to inc any bios we're about to issue
940 * before the cell is released, and avoid a race with new bios
941 * being added to the cell.
942 */
943 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
944 &info, cell);
945
946 while ((bio = bio_list_pop(&info.defer_bios)))
947 thin_defer_bio(tc, bio);
948
949 while ((bio = bio_list_pop(&info.issue_bios)))
950 remap_and_issue(info.tc, bio, block);
951}
952
953static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
954{
955 cell_error(m->tc->pool, m->cell);
956 list_del(&m->list);
957 mempool_free(m, &m->tc->pool->mapping_pool);
958}
959
960static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
961{
962 struct pool *pool = tc->pool;
963 unsigned long flags;
964
965 /*
966 * If the bio has the REQ_FUA flag set we must commit the metadata
967 * before signaling its completion.
968 */
969 if (!bio_triggers_commit(tc, bio)) {
970 bio_endio(bio);
971 return;
972 }
973
974 /*
975 * Complete bio with an error if earlier I/O caused changes to the
976 * metadata that can't be committed, e.g, due to I/O errors on the
977 * metadata device.
978 */
979 if (dm_thin_aborted_changes(tc->td)) {
980 bio_io_error(bio);
981 return;
982 }
983
984 /*
985 * Batch together any bios that trigger commits and then issue a
986 * single commit for them in process_deferred_bios().
987 */
988 spin_lock_irqsave(&pool->lock, flags);
989 bio_list_add(&pool->deferred_flush_completions, bio);
990 spin_unlock_irqrestore(&pool->lock, flags);
991}
992
993static void process_prepared_mapping(struct dm_thin_new_mapping *m)
994{
995 struct thin_c *tc = m->tc;
996 struct pool *pool = tc->pool;
997 struct bio *bio = m->bio;
998 int r;
999
1000 if (m->status) {
1001 cell_error(pool, m->cell);
1002 goto out;
1003 }
1004
1005 /*
1006 * Commit the prepared block into the mapping btree.
1007 * Any I/O for this block arriving after this point will get
1008 * remapped to it directly.
1009 */
1010 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1011 if (r) {
1012 metadata_operation_failed(pool, "dm_thin_insert_block", r);
1013 cell_error(pool, m->cell);
1014 goto out;
1015 }
1016
1017 /*
1018 * Release any bios held while the block was being provisioned.
1019 * If we are processing a write bio that completely covers the block,
1020 * we already processed it so can ignore it now when processing
1021 * the bios in the cell.
1022 */
1023 if (bio) {
1024 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1025 complete_overwrite_bio(tc, bio);
1026 } else {
1027 inc_all_io_entry(tc->pool, m->cell->holder);
1028 remap_and_issue(tc, m->cell->holder, m->data_block);
1029 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1030 }
1031
1032out:
1033 list_del(&m->list);
1034 mempool_free(m, &pool->mapping_pool);
1035}
1036
1037/*----------------------------------------------------------------*/
1038
1039static void free_discard_mapping(struct dm_thin_new_mapping *m)
1040{
1041 struct thin_c *tc = m->tc;
1042 if (m->cell)
1043 cell_defer_no_holder(tc, m->cell);
1044 mempool_free(m, &tc->pool->mapping_pool);
1045}
1046
1047static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1048{
1049 bio_io_error(m->bio);
1050 free_discard_mapping(m);
1051}
1052
1053static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1054{
1055 bio_endio(m->bio);
1056 free_discard_mapping(m);
1057}
1058
1059static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1060{
1061 int r;
1062 struct thin_c *tc = m->tc;
1063
1064 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1065 if (r) {
1066 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1067 bio_io_error(m->bio);
1068 } else
1069 bio_endio(m->bio);
1070
1071 cell_defer_no_holder(tc, m->cell);
1072 mempool_free(m, &tc->pool->mapping_pool);
1073}
1074
1075/*----------------------------------------------------------------*/
1076
1077static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1078 struct bio *discard_parent)
1079{
1080 /*
1081 * We've already unmapped this range of blocks, but before we
1082 * passdown we have to check that these blocks are now unused.
1083 */
1084 int r = 0;
1085 bool shared = true;
1086 struct thin_c *tc = m->tc;
1087 struct pool *pool = tc->pool;
1088 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1089 struct discard_op op;
1090
1091 begin_discard(&op, tc, discard_parent);
1092 while (b != end) {
1093 /* find start of unmapped run */
1094 for (; b < end; b++) {
1095 r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1096 if (r)
1097 goto out;
1098
1099 if (!shared)
1100 break;
1101 }
1102
1103 if (b == end)
1104 break;
1105
1106 /* find end of run */
1107 for (e = b + 1; e != end; e++) {
1108 r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1109 if (r)
1110 goto out;
1111
1112 if (shared)
1113 break;
1114 }
1115
1116 r = issue_discard(&op, b, e);
1117 if (r)
1118 goto out;
1119
1120 b = e;
1121 }
1122out:
1123 end_discard(&op, r);
1124}
1125
1126static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1127{
1128 unsigned long flags;
1129 struct pool *pool = m->tc->pool;
1130
1131 spin_lock_irqsave(&pool->lock, flags);
1132 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1133 spin_unlock_irqrestore(&pool->lock, flags);
1134 wake_worker(pool);
1135}
1136
1137static void passdown_endio(struct bio *bio)
1138{
1139 /*
1140 * It doesn't matter if the passdown discard failed, we still want
1141 * to unmap (we ignore err).
1142 */
1143 queue_passdown_pt2(bio->bi_private);
1144 bio_put(bio);
1145}
1146
1147static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1148{
1149 int r;
1150 struct thin_c *tc = m->tc;
1151 struct pool *pool = tc->pool;
1152 struct bio *discard_parent;
1153 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1154
1155 /*
1156 * Only this thread allocates blocks, so we can be sure that the
1157 * newly unmapped blocks will not be allocated before the end of
1158 * the function.
1159 */
1160 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1161 if (r) {
1162 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1163 bio_io_error(m->bio);
1164 cell_defer_no_holder(tc, m->cell);
1165 mempool_free(m, &pool->mapping_pool);
1166 return;
1167 }
1168
1169 /*
1170 * Increment the unmapped blocks. This prevents a race between the
1171 * passdown io and reallocation of freed blocks.
1172 */
1173 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1174 if (r) {
1175 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1176 bio_io_error(m->bio);
1177 cell_defer_no_holder(tc, m->cell);
1178 mempool_free(m, &pool->mapping_pool);
1179 return;
1180 }
1181
1182 discard_parent = bio_alloc(GFP_NOIO, 1);
1183 if (!discard_parent) {
1184 DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1185 dm_device_name(tc->pool->pool_md));
1186 queue_passdown_pt2(m);
1187
1188 } else {
1189 discard_parent->bi_end_io = passdown_endio;
1190 discard_parent->bi_private = m;
1191
1192 if (m->maybe_shared)
1193 passdown_double_checking_shared_status(m, discard_parent);
1194 else {
1195 struct discard_op op;
1196
1197 begin_discard(&op, tc, discard_parent);
1198 r = issue_discard(&op, m->data_block, data_end);
1199 end_discard(&op, r);
1200 }
1201 }
1202}
1203
1204static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1205{
1206 int r;
1207 struct thin_c *tc = m->tc;
1208 struct pool *pool = tc->pool;
1209
1210 /*
1211 * The passdown has completed, so now we can decrement all those
1212 * unmapped blocks.
1213 */
1214 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1215 m->data_block + (m->virt_end - m->virt_begin));
1216 if (r) {
1217 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1218 bio_io_error(m->bio);
1219 } else
1220 bio_endio(m->bio);
1221
1222 cell_defer_no_holder(tc, m->cell);
1223 mempool_free(m, &pool->mapping_pool);
1224}
1225
1226static void process_prepared(struct pool *pool, struct list_head *head,
1227 process_mapping_fn *fn)
1228{
1229 unsigned long flags;
1230 struct list_head maps;
1231 struct dm_thin_new_mapping *m, *tmp;
1232
1233 INIT_LIST_HEAD(&maps);
1234 spin_lock_irqsave(&pool->lock, flags);
1235 list_splice_init(head, &maps);
1236 spin_unlock_irqrestore(&pool->lock, flags);
1237
1238 list_for_each_entry_safe(m, tmp, &maps, list)
1239 (*fn)(m);
1240}
1241
1242/*
1243 * Deferred bio jobs.
1244 */
1245static int io_overlaps_block(struct pool *pool, struct bio *bio)
1246{
1247 return bio->bi_iter.bi_size ==
1248 (pool->sectors_per_block << SECTOR_SHIFT);
1249}
1250
1251static int io_overwrites_block(struct pool *pool, struct bio *bio)
1252{
1253 return (bio_data_dir(bio) == WRITE) &&
1254 io_overlaps_block(pool, bio);
1255}
1256
1257static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1258 bio_end_io_t *fn)
1259{
1260 *save = bio->bi_end_io;
1261 bio->bi_end_io = fn;
1262}
1263
1264static int ensure_next_mapping(struct pool *pool)
1265{
1266 if (pool->next_mapping)
1267 return 0;
1268
1269 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1270
1271 return pool->next_mapping ? 0 : -ENOMEM;
1272}
1273
1274static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1275{
1276 struct dm_thin_new_mapping *m = pool->next_mapping;
1277
1278 BUG_ON(!pool->next_mapping);
1279
1280 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1281 INIT_LIST_HEAD(&m->list);
1282 m->bio = NULL;
1283
1284 pool->next_mapping = NULL;
1285
1286 return m;
1287}
1288
1289static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1290 sector_t begin, sector_t end)
1291{
1292 struct dm_io_region to;
1293
1294 to.bdev = tc->pool_dev->bdev;
1295 to.sector = begin;
1296 to.count = end - begin;
1297
1298 dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1299}
1300
1301static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1302 dm_block_t data_begin,
1303 struct dm_thin_new_mapping *m)
1304{
1305 struct pool *pool = tc->pool;
1306 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1307
1308 h->overwrite_mapping = m;
1309 m->bio = bio;
1310 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1311 inc_all_io_entry(pool, bio);
1312 remap_and_issue(tc, bio, data_begin);
1313}
1314
1315/*
1316 * A partial copy also needs to zero the uncopied region.
1317 */
1318static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1319 struct dm_dev *origin, dm_block_t data_origin,
1320 dm_block_t data_dest,
1321 struct dm_bio_prison_cell *cell, struct bio *bio,
1322 sector_t len)
1323{
1324 struct pool *pool = tc->pool;
1325 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1326
1327 m->tc = tc;
1328 m->virt_begin = virt_block;
1329 m->virt_end = virt_block + 1u;
1330 m->data_block = data_dest;
1331 m->cell = cell;
1332
1333 /*
1334 * quiesce action + copy action + an extra reference held for the
1335 * duration of this function (we may need to inc later for a
1336 * partial zero).
1337 */
1338 atomic_set(&m->prepare_actions, 3);
1339
1340 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1341 complete_mapping_preparation(m); /* already quiesced */
1342
1343 /*
1344 * IO to pool_dev remaps to the pool target's data_dev.
1345 *
1346 * If the whole block of data is being overwritten, we can issue the
1347 * bio immediately. Otherwise we use kcopyd to clone the data first.
1348 */
1349 if (io_overwrites_block(pool, bio))
1350 remap_and_issue_overwrite(tc, bio, data_dest, m);
1351 else {
1352 struct dm_io_region from, to;
1353
1354 from.bdev = origin->bdev;
1355 from.sector = data_origin * pool->sectors_per_block;
1356 from.count = len;
1357
1358 to.bdev = tc->pool_dev->bdev;
1359 to.sector = data_dest * pool->sectors_per_block;
1360 to.count = len;
1361
1362 dm_kcopyd_copy(pool->copier, &from, 1, &to,
1363 0, copy_complete, m);
1364
1365 /*
1366 * Do we need to zero a tail region?
1367 */
1368 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1369 atomic_inc(&m->prepare_actions);
1370 ll_zero(tc, m,
1371 data_dest * pool->sectors_per_block + len,
1372 (data_dest + 1) * pool->sectors_per_block);
1373 }
1374 }
1375
1376 complete_mapping_preparation(m); /* drop our ref */
1377}
1378
1379static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1380 dm_block_t data_origin, dm_block_t data_dest,
1381 struct dm_bio_prison_cell *cell, struct bio *bio)
1382{
1383 schedule_copy(tc, virt_block, tc->pool_dev,
1384 data_origin, data_dest, cell, bio,
1385 tc->pool->sectors_per_block);
1386}
1387
1388static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1389 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1390 struct bio *bio)
1391{
1392 struct pool *pool = tc->pool;
1393 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1394
1395 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1396 m->tc = tc;
1397 m->virt_begin = virt_block;
1398 m->virt_end = virt_block + 1u;
1399 m->data_block = data_block;
1400 m->cell = cell;
1401
1402 /*
1403 * If the whole block of data is being overwritten or we are not
1404 * zeroing pre-existing data, we can issue the bio immediately.
1405 * Otherwise we use kcopyd to zero the data first.
1406 */
1407 if (pool->pf.zero_new_blocks) {
1408 if (io_overwrites_block(pool, bio))
1409 remap_and_issue_overwrite(tc, bio, data_block, m);
1410 else
1411 ll_zero(tc, m, data_block * pool->sectors_per_block,
1412 (data_block + 1) * pool->sectors_per_block);
1413 } else
1414 process_prepared_mapping(m);
1415}
1416
1417static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1418 dm_block_t data_dest,
1419 struct dm_bio_prison_cell *cell, struct bio *bio)
1420{
1421 struct pool *pool = tc->pool;
1422 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1423 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1424
1425 if (virt_block_end <= tc->origin_size)
1426 schedule_copy(tc, virt_block, tc->origin_dev,
1427 virt_block, data_dest, cell, bio,
1428 pool->sectors_per_block);
1429
1430 else if (virt_block_begin < tc->origin_size)
1431 schedule_copy(tc, virt_block, tc->origin_dev,
1432 virt_block, data_dest, cell, bio,
1433 tc->origin_size - virt_block_begin);
1434
1435 else
1436 schedule_zero(tc, virt_block, data_dest, cell, bio);
1437}
1438
1439static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1440
1441static void requeue_bios(struct pool *pool);
1442
1443static bool is_read_only_pool_mode(enum pool_mode mode)
1444{
1445 return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1446}
1447
1448static bool is_read_only(struct pool *pool)
1449{
1450 return is_read_only_pool_mode(get_pool_mode(pool));
1451}
1452
1453static void check_for_metadata_space(struct pool *pool)
1454{
1455 int r;
1456 const char *ooms_reason = NULL;
1457 dm_block_t nr_free;
1458
1459 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1460 if (r)
1461 ooms_reason = "Could not get free metadata blocks";
1462 else if (!nr_free)
1463 ooms_reason = "No free metadata blocks";
1464
1465 if (ooms_reason && !is_read_only(pool)) {
1466 DMERR("%s", ooms_reason);
1467 set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1468 }
1469}
1470
1471static void check_for_data_space(struct pool *pool)
1472{
1473 int r;
1474 dm_block_t nr_free;
1475
1476 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1477 return;
1478
1479 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1480 if (r)
1481 return;
1482
1483 if (nr_free) {
1484 set_pool_mode(pool, PM_WRITE);
1485 requeue_bios(pool);
1486 }
1487}
1488
1489/*
1490 * A non-zero return indicates read_only or fail_io mode.
1491 * Many callers don't care about the return value.
1492 */
1493static int commit(struct pool *pool)
1494{
1495 int r;
1496
1497 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1498 return -EINVAL;
1499
1500 r = dm_pool_commit_metadata(pool->pmd);
1501 if (r)
1502 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1503 else {
1504 check_for_metadata_space(pool);
1505 check_for_data_space(pool);
1506 }
1507
1508 return r;
1509}
1510
1511static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1512{
1513 unsigned long flags;
1514
1515 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1516 DMWARN("%s: reached low water mark for data device: sending event.",
1517 dm_device_name(pool->pool_md));
1518 spin_lock_irqsave(&pool->lock, flags);
1519 pool->low_water_triggered = true;
1520 spin_unlock_irqrestore(&pool->lock, flags);
1521 dm_table_event(pool->ti->table);
1522 }
1523}
1524
1525static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1526{
1527 int r;
1528 dm_block_t free_blocks;
1529 struct pool *pool = tc->pool;
1530
1531 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1532 return -EINVAL;
1533
1534 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1535 if (r) {
1536 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1537 return r;
1538 }
1539
1540 check_low_water_mark(pool, free_blocks);
1541
1542 if (!free_blocks) {
1543 /*
1544 * Try to commit to see if that will free up some
1545 * more space.
1546 */
1547 r = commit(pool);
1548 if (r)
1549 return r;
1550
1551 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1552 if (r) {
1553 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1554 return r;
1555 }
1556
1557 if (!free_blocks) {
1558 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1559 return -ENOSPC;
1560 }
1561 }
1562
1563 r = dm_pool_alloc_data_block(pool->pmd, result);
1564 if (r) {
1565 if (r == -ENOSPC)
1566 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1567 else
1568 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1569 return r;
1570 }
1571
1572 r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1573 if (r) {
1574 metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1575 return r;
1576 }
1577
1578 if (!free_blocks) {
1579 /* Let's commit before we use up the metadata reserve. */
1580 r = commit(pool);
1581 if (r)
1582 return r;
1583 }
1584
1585 return 0;
1586}
1587
1588/*
1589 * If we have run out of space, queue bios until the device is
1590 * resumed, presumably after having been reloaded with more space.
1591 */
1592static void retry_on_resume(struct bio *bio)
1593{
1594 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1595 struct thin_c *tc = h->tc;
1596 unsigned long flags;
1597
1598 spin_lock_irqsave(&tc->lock, flags);
1599 bio_list_add(&tc->retry_on_resume_list, bio);
1600 spin_unlock_irqrestore(&tc->lock, flags);
1601}
1602
1603static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1604{
1605 enum pool_mode m = get_pool_mode(pool);
1606
1607 switch (m) {
1608 case PM_WRITE:
1609 /* Shouldn't get here */
1610 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1611 return BLK_STS_IOERR;
1612
1613 case PM_OUT_OF_DATA_SPACE:
1614 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1615
1616 case PM_OUT_OF_METADATA_SPACE:
1617 case PM_READ_ONLY:
1618 case PM_FAIL:
1619 return BLK_STS_IOERR;
1620 default:
1621 /* Shouldn't get here */
1622 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1623 return BLK_STS_IOERR;
1624 }
1625}
1626
1627static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1628{
1629 blk_status_t error = should_error_unserviceable_bio(pool);
1630
1631 if (error) {
1632 bio->bi_status = error;
1633 bio_endio(bio);
1634 } else
1635 retry_on_resume(bio);
1636}
1637
1638static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1639{
1640 struct bio *bio;
1641 struct bio_list bios;
1642 blk_status_t error;
1643
1644 error = should_error_unserviceable_bio(pool);
1645 if (error) {
1646 cell_error_with_code(pool, cell, error);
1647 return;
1648 }
1649
1650 bio_list_init(&bios);
1651 cell_release(pool, cell, &bios);
1652
1653 while ((bio = bio_list_pop(&bios)))
1654 retry_on_resume(bio);
1655}
1656
1657static void process_discard_cell_no_passdown(struct thin_c *tc,
1658 struct dm_bio_prison_cell *virt_cell)
1659{
1660 struct pool *pool = tc->pool;
1661 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1662
1663 /*
1664 * We don't need to lock the data blocks, since there's no
1665 * passdown. We only lock data blocks for allocation and breaking sharing.
1666 */
1667 m->tc = tc;
1668 m->virt_begin = virt_cell->key.block_begin;
1669 m->virt_end = virt_cell->key.block_end;
1670 m->cell = virt_cell;
1671 m->bio = virt_cell->holder;
1672
1673 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1674 pool->process_prepared_discard(m);
1675}
1676
1677static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1678 struct bio *bio)
1679{
1680 struct pool *pool = tc->pool;
1681
1682 int r;
1683 bool maybe_shared;
1684 struct dm_cell_key data_key;
1685 struct dm_bio_prison_cell *data_cell;
1686 struct dm_thin_new_mapping *m;
1687 dm_block_t virt_begin, virt_end, data_begin;
1688
1689 while (begin != end) {
1690 r = ensure_next_mapping(pool);
1691 if (r)
1692 /* we did our best */
1693 return;
1694
1695 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1696 &data_begin, &maybe_shared);
1697 if (r)
1698 /*
1699 * Silently fail, letting any mappings we've
1700 * created complete.
1701 */
1702 break;
1703
1704 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1705 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1706 /* contention, we'll give up with this range */
1707 begin = virt_end;
1708 continue;
1709 }
1710
1711 /*
1712 * IO may still be going to the destination block. We must
1713 * quiesce before we can do the removal.
1714 */
1715 m = get_next_mapping(pool);
1716 m->tc = tc;
1717 m->maybe_shared = maybe_shared;
1718 m->virt_begin = virt_begin;
1719 m->virt_end = virt_end;
1720 m->data_block = data_begin;
1721 m->cell = data_cell;
1722 m->bio = bio;
1723
1724 /*
1725 * The parent bio must not complete before sub discard bios are
1726 * chained to it (see end_discard's bio_chain)!
1727 *
1728 * This per-mapping bi_remaining increment is paired with
1729 * the implicit decrement that occurs via bio_endio() in
1730 * end_discard().
1731 */
1732 bio_inc_remaining(bio);
1733 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1734 pool->process_prepared_discard(m);
1735
1736 begin = virt_end;
1737 }
1738}
1739
1740static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1741{
1742 struct bio *bio = virt_cell->holder;
1743 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1744
1745 /*
1746 * The virt_cell will only get freed once the origin bio completes.
1747 * This means it will remain locked while all the individual
1748 * passdown bios are in flight.
1749 */
1750 h->cell = virt_cell;
1751 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1752
1753 /*
1754 * We complete the bio now, knowing that the bi_remaining field
1755 * will prevent completion until the sub range discards have
1756 * completed.
1757 */
1758 bio_endio(bio);
1759}
1760
1761static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1762{
1763 dm_block_t begin, end;
1764 struct dm_cell_key virt_key;
1765 struct dm_bio_prison_cell *virt_cell;
1766
1767 get_bio_block_range(tc, bio, &begin, &end);
1768 if (begin == end) {
1769 /*
1770 * The discard covers less than a block.
1771 */
1772 bio_endio(bio);
1773 return;
1774 }
1775
1776 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1777 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1778 /*
1779 * Potential starvation issue: We're relying on the
1780 * fs/application being well behaved, and not trying to
1781 * send IO to a region at the same time as discarding it.
1782 * If they do this persistently then it's possible this
1783 * cell will never be granted.
1784 */
1785 return;
1786
1787 tc->pool->process_discard_cell(tc, virt_cell);
1788}
1789
1790static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1791 struct dm_cell_key *key,
1792 struct dm_thin_lookup_result *lookup_result,
1793 struct dm_bio_prison_cell *cell)
1794{
1795 int r;
1796 dm_block_t data_block;
1797 struct pool *pool = tc->pool;
1798
1799 r = alloc_data_block(tc, &data_block);
1800 switch (r) {
1801 case 0:
1802 schedule_internal_copy(tc, block, lookup_result->block,
1803 data_block, cell, bio);
1804 break;
1805
1806 case -ENOSPC:
1807 retry_bios_on_resume(pool, cell);
1808 break;
1809
1810 default:
1811 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1812 __func__, r);
1813 cell_error(pool, cell);
1814 break;
1815 }
1816}
1817
1818static void __remap_and_issue_shared_cell(void *context,
1819 struct dm_bio_prison_cell *cell)
1820{
1821 struct remap_info *info = context;
1822 struct bio *bio;
1823
1824 while ((bio = bio_list_pop(&cell->bios))) {
1825 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1826 bio_op(bio) == REQ_OP_DISCARD)
1827 bio_list_add(&info->defer_bios, bio);
1828 else {
1829 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1830
1831 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1832 inc_all_io_entry(info->tc->pool, bio);
1833 bio_list_add(&info->issue_bios, bio);
1834 }
1835 }
1836}
1837
1838static void remap_and_issue_shared_cell(struct thin_c *tc,
1839 struct dm_bio_prison_cell *cell,
1840 dm_block_t block)
1841{
1842 struct bio *bio;
1843 struct remap_info info;
1844
1845 info.tc = tc;
1846 bio_list_init(&info.defer_bios);
1847 bio_list_init(&info.issue_bios);
1848
1849 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1850 &info, cell);
1851
1852 while ((bio = bio_list_pop(&info.defer_bios)))
1853 thin_defer_bio(tc, bio);
1854
1855 while ((bio = bio_list_pop(&info.issue_bios)))
1856 remap_and_issue(tc, bio, block);
1857}
1858
1859static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1860 dm_block_t block,
1861 struct dm_thin_lookup_result *lookup_result,
1862 struct dm_bio_prison_cell *virt_cell)
1863{
1864 struct dm_bio_prison_cell *data_cell;
1865 struct pool *pool = tc->pool;
1866 struct dm_cell_key key;
1867
1868 /*
1869 * If cell is already occupied, then sharing is already in the process
1870 * of being broken so we have nothing further to do here.
1871 */
1872 build_data_key(tc->td, lookup_result->block, &key);
1873 if (bio_detain(pool, &key, bio, &data_cell)) {
1874 cell_defer_no_holder(tc, virt_cell);
1875 return;
1876 }
1877
1878 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1879 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1880 cell_defer_no_holder(tc, virt_cell);
1881 } else {
1882 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1883
1884 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1885 inc_all_io_entry(pool, bio);
1886 remap_and_issue(tc, bio, lookup_result->block);
1887
1888 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1889 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1890 }
1891}
1892
1893static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1894 struct dm_bio_prison_cell *cell)
1895{
1896 int r;
1897 dm_block_t data_block;
1898 struct pool *pool = tc->pool;
1899
1900 /*
1901 * Remap empty bios (flushes) immediately, without provisioning.
1902 */
1903 if (!bio->bi_iter.bi_size) {
1904 inc_all_io_entry(pool, bio);
1905 cell_defer_no_holder(tc, cell);
1906
1907 remap_and_issue(tc, bio, 0);
1908 return;
1909 }
1910
1911 /*
1912 * Fill read bios with zeroes and complete them immediately.
1913 */
1914 if (bio_data_dir(bio) == READ) {
1915 zero_fill_bio(bio);
1916 cell_defer_no_holder(tc, cell);
1917 bio_endio(bio);
1918 return;
1919 }
1920
1921 r = alloc_data_block(tc, &data_block);
1922 switch (r) {
1923 case 0:
1924 if (tc->origin_dev)
1925 schedule_external_copy(tc, block, data_block, cell, bio);
1926 else
1927 schedule_zero(tc, block, data_block, cell, bio);
1928 break;
1929
1930 case -ENOSPC:
1931 retry_bios_on_resume(pool, cell);
1932 break;
1933
1934 default:
1935 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1936 __func__, r);
1937 cell_error(pool, cell);
1938 break;
1939 }
1940}
1941
1942static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1943{
1944 int r;
1945 struct pool *pool = tc->pool;
1946 struct bio *bio = cell->holder;
1947 dm_block_t block = get_bio_block(tc, bio);
1948 struct dm_thin_lookup_result lookup_result;
1949
1950 if (tc->requeue_mode) {
1951 cell_requeue(pool, cell);
1952 return;
1953 }
1954
1955 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1956 switch (r) {
1957 case 0:
1958 if (lookup_result.shared)
1959 process_shared_bio(tc, bio, block, &lookup_result, cell);
1960 else {
1961 inc_all_io_entry(pool, bio);
1962 remap_and_issue(tc, bio, lookup_result.block);
1963 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1964 }
1965 break;
1966
1967 case -ENODATA:
1968 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1969 inc_all_io_entry(pool, bio);
1970 cell_defer_no_holder(tc, cell);
1971
1972 if (bio_end_sector(bio) <= tc->origin_size)
1973 remap_to_origin_and_issue(tc, bio);
1974
1975 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1976 zero_fill_bio(bio);
1977 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1978 remap_to_origin_and_issue(tc, bio);
1979
1980 } else {
1981 zero_fill_bio(bio);
1982 bio_endio(bio);
1983 }
1984 } else
1985 provision_block(tc, bio, block, cell);
1986 break;
1987
1988 default:
1989 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1990 __func__, r);
1991 cell_defer_no_holder(tc, cell);
1992 bio_io_error(bio);
1993 break;
1994 }
1995}
1996
1997static void process_bio(struct thin_c *tc, struct bio *bio)
1998{
1999 struct pool *pool = tc->pool;
2000 dm_block_t block = get_bio_block(tc, bio);
2001 struct dm_bio_prison_cell *cell;
2002 struct dm_cell_key key;
2003
2004 /*
2005 * If cell is already occupied, then the block is already
2006 * being provisioned so we have nothing further to do here.
2007 */
2008 build_virtual_key(tc->td, block, &key);
2009 if (bio_detain(pool, &key, bio, &cell))
2010 return;
2011
2012 process_cell(tc, cell);
2013}
2014
2015static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2016 struct dm_bio_prison_cell *cell)
2017{
2018 int r;
2019 int rw = bio_data_dir(bio);
2020 dm_block_t block = get_bio_block(tc, bio);
2021 struct dm_thin_lookup_result lookup_result;
2022
2023 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2024 switch (r) {
2025 case 0:
2026 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2027 handle_unserviceable_bio(tc->pool, bio);
2028 if (cell)
2029 cell_defer_no_holder(tc, cell);
2030 } else {
2031 inc_all_io_entry(tc->pool, bio);
2032 remap_and_issue(tc, bio, lookup_result.block);
2033 if (cell)
2034 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2035 }
2036 break;
2037
2038 case -ENODATA:
2039 if (cell)
2040 cell_defer_no_holder(tc, cell);
2041 if (rw != READ) {
2042 handle_unserviceable_bio(tc->pool, bio);
2043 break;
2044 }
2045
2046 if (tc->origin_dev) {
2047 inc_all_io_entry(tc->pool, bio);
2048 remap_to_origin_and_issue(tc, bio);
2049 break;
2050 }
2051
2052 zero_fill_bio(bio);
2053 bio_endio(bio);
2054 break;
2055
2056 default:
2057 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2058 __func__, r);
2059 if (cell)
2060 cell_defer_no_holder(tc, cell);
2061 bio_io_error(bio);
2062 break;
2063 }
2064}
2065
2066static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2067{
2068 __process_bio_read_only(tc, bio, NULL);
2069}
2070
2071static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2072{
2073 __process_bio_read_only(tc, cell->holder, cell);
2074}
2075
2076static void process_bio_success(struct thin_c *tc, struct bio *bio)
2077{
2078 bio_endio(bio);
2079}
2080
2081static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2082{
2083 bio_io_error(bio);
2084}
2085
2086static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2087{
2088 cell_success(tc->pool, cell);
2089}
2090
2091static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2092{
2093 cell_error(tc->pool, cell);
2094}
2095
2096/*
2097 * FIXME: should we also commit due to size of transaction, measured in
2098 * metadata blocks?
2099 */
2100static int need_commit_due_to_time(struct pool *pool)
2101{
2102 return !time_in_range(jiffies, pool->last_commit_jiffies,
2103 pool->last_commit_jiffies + COMMIT_PERIOD);
2104}
2105
2106#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2107#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2108
2109static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2110{
2111 struct rb_node **rbp, *parent;
2112 struct dm_thin_endio_hook *pbd;
2113 sector_t bi_sector = bio->bi_iter.bi_sector;
2114
2115 rbp = &tc->sort_bio_list.rb_node;
2116 parent = NULL;
2117 while (*rbp) {
2118 parent = *rbp;
2119 pbd = thin_pbd(parent);
2120
2121 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2122 rbp = &(*rbp)->rb_left;
2123 else
2124 rbp = &(*rbp)->rb_right;
2125 }
2126
2127 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2128 rb_link_node(&pbd->rb_node, parent, rbp);
2129 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2130}
2131
2132static void __extract_sorted_bios(struct thin_c *tc)
2133{
2134 struct rb_node *node;
2135 struct dm_thin_endio_hook *pbd;
2136 struct bio *bio;
2137
2138 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2139 pbd = thin_pbd(node);
2140 bio = thin_bio(pbd);
2141
2142 bio_list_add(&tc->deferred_bio_list, bio);
2143 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2144 }
2145
2146 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2147}
2148
2149static void __sort_thin_deferred_bios(struct thin_c *tc)
2150{
2151 struct bio *bio;
2152 struct bio_list bios;
2153
2154 bio_list_init(&bios);
2155 bio_list_merge(&bios, &tc->deferred_bio_list);
2156 bio_list_init(&tc->deferred_bio_list);
2157
2158 /* Sort deferred_bio_list using rb-tree */
2159 while ((bio = bio_list_pop(&bios)))
2160 __thin_bio_rb_add(tc, bio);
2161
2162 /*
2163 * Transfer the sorted bios in sort_bio_list back to
2164 * deferred_bio_list to allow lockless submission of
2165 * all bios.
2166 */
2167 __extract_sorted_bios(tc);
2168}
2169
2170static void process_thin_deferred_bios(struct thin_c *tc)
2171{
2172 struct pool *pool = tc->pool;
2173 unsigned long flags;
2174 struct bio *bio;
2175 struct bio_list bios;
2176 struct blk_plug plug;
2177 unsigned count = 0;
2178
2179 if (tc->requeue_mode) {
2180 error_thin_bio_list(tc, &tc->deferred_bio_list,
2181 BLK_STS_DM_REQUEUE);
2182 return;
2183 }
2184
2185 bio_list_init(&bios);
2186
2187 spin_lock_irqsave(&tc->lock, flags);
2188
2189 if (bio_list_empty(&tc->deferred_bio_list)) {
2190 spin_unlock_irqrestore(&tc->lock, flags);
2191 return;
2192 }
2193
2194 __sort_thin_deferred_bios(tc);
2195
2196 bio_list_merge(&bios, &tc->deferred_bio_list);
2197 bio_list_init(&tc->deferred_bio_list);
2198
2199 spin_unlock_irqrestore(&tc->lock, flags);
2200
2201 blk_start_plug(&plug);
2202 while ((bio = bio_list_pop(&bios))) {
2203 /*
2204 * If we've got no free new_mapping structs, and processing
2205 * this bio might require one, we pause until there are some
2206 * prepared mappings to process.
2207 */
2208 if (ensure_next_mapping(pool)) {
2209 spin_lock_irqsave(&tc->lock, flags);
2210 bio_list_add(&tc->deferred_bio_list, bio);
2211 bio_list_merge(&tc->deferred_bio_list, &bios);
2212 spin_unlock_irqrestore(&tc->lock, flags);
2213 break;
2214 }
2215
2216 if (bio_op(bio) == REQ_OP_DISCARD)
2217 pool->process_discard(tc, bio);
2218 else
2219 pool->process_bio(tc, bio);
2220
2221 if ((count++ & 127) == 0) {
2222 throttle_work_update(&pool->throttle);
2223 dm_pool_issue_prefetches(pool->pmd);
2224 }
2225 }
2226 blk_finish_plug(&plug);
2227}
2228
2229static int cmp_cells(const void *lhs, const void *rhs)
2230{
2231 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2232 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2233
2234 BUG_ON(!lhs_cell->holder);
2235 BUG_ON(!rhs_cell->holder);
2236
2237 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2238 return -1;
2239
2240 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2241 return 1;
2242
2243 return 0;
2244}
2245
2246static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2247{
2248 unsigned count = 0;
2249 struct dm_bio_prison_cell *cell, *tmp;
2250
2251 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2252 if (count >= CELL_SORT_ARRAY_SIZE)
2253 break;
2254
2255 pool->cell_sort_array[count++] = cell;
2256 list_del(&cell->user_list);
2257 }
2258
2259 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2260
2261 return count;
2262}
2263
2264static void process_thin_deferred_cells(struct thin_c *tc)
2265{
2266 struct pool *pool = tc->pool;
2267 unsigned long flags;
2268 struct list_head cells;
2269 struct dm_bio_prison_cell *cell;
2270 unsigned i, j, count;
2271
2272 INIT_LIST_HEAD(&cells);
2273
2274 spin_lock_irqsave(&tc->lock, flags);
2275 list_splice_init(&tc->deferred_cells, &cells);
2276 spin_unlock_irqrestore(&tc->lock, flags);
2277
2278 if (list_empty(&cells))
2279 return;
2280
2281 do {
2282 count = sort_cells(tc->pool, &cells);
2283
2284 for (i = 0; i < count; i++) {
2285 cell = pool->cell_sort_array[i];
2286 BUG_ON(!cell->holder);
2287
2288 /*
2289 * If we've got no free new_mapping structs, and processing
2290 * this bio might require one, we pause until there are some
2291 * prepared mappings to process.
2292 */
2293 if (ensure_next_mapping(pool)) {
2294 for (j = i; j < count; j++)
2295 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2296
2297 spin_lock_irqsave(&tc->lock, flags);
2298 list_splice(&cells, &tc->deferred_cells);
2299 spin_unlock_irqrestore(&tc->lock, flags);
2300 return;
2301 }
2302
2303 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2304 pool->process_discard_cell(tc, cell);
2305 else
2306 pool->process_cell(tc, cell);
2307 }
2308 } while (!list_empty(&cells));
2309}
2310
2311static void thin_get(struct thin_c *tc);
2312static void thin_put(struct thin_c *tc);
2313
2314/*
2315 * We can't hold rcu_read_lock() around code that can block. So we
2316 * find a thin with the rcu lock held; bump a refcount; then drop
2317 * the lock.
2318 */
2319static struct thin_c *get_first_thin(struct pool *pool)
2320{
2321 struct thin_c *tc = NULL;
2322
2323 rcu_read_lock();
2324 if (!list_empty(&pool->active_thins)) {
2325 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2326 thin_get(tc);
2327 }
2328 rcu_read_unlock();
2329
2330 return tc;
2331}
2332
2333static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2334{
2335 struct thin_c *old_tc = tc;
2336
2337 rcu_read_lock();
2338 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2339 thin_get(tc);
2340 thin_put(old_tc);
2341 rcu_read_unlock();
2342 return tc;
2343 }
2344 thin_put(old_tc);
2345 rcu_read_unlock();
2346
2347 return NULL;
2348}
2349
2350static void process_deferred_bios(struct pool *pool)
2351{
2352 unsigned long flags;
2353 struct bio *bio;
2354 struct bio_list bios, bio_completions;
2355 struct thin_c *tc;
2356
2357 tc = get_first_thin(pool);
2358 while (tc) {
2359 process_thin_deferred_cells(tc);
2360 process_thin_deferred_bios(tc);
2361 tc = get_next_thin(pool, tc);
2362 }
2363
2364 /*
2365 * If there are any deferred flush bios, we must commit the metadata
2366 * before issuing them or signaling their completion.
2367 */
2368 bio_list_init(&bios);
2369 bio_list_init(&bio_completions);
2370
2371 spin_lock_irqsave(&pool->lock, flags);
2372 bio_list_merge(&bios, &pool->deferred_flush_bios);
2373 bio_list_init(&pool->deferred_flush_bios);
2374
2375 bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2376 bio_list_init(&pool->deferred_flush_completions);
2377 spin_unlock_irqrestore(&pool->lock, flags);
2378
2379 if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2380 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2381 return;
2382
2383 if (commit(pool)) {
2384 bio_list_merge(&bios, &bio_completions);
2385
2386 while ((bio = bio_list_pop(&bios)))
2387 bio_io_error(bio);
2388 return;
2389 }
2390 pool->last_commit_jiffies = jiffies;
2391
2392 while ((bio = bio_list_pop(&bio_completions)))
2393 bio_endio(bio);
2394
2395 while ((bio = bio_list_pop(&bios)))
2396 generic_make_request(bio);
2397}
2398
2399static void do_worker(struct work_struct *ws)
2400{
2401 struct pool *pool = container_of(ws, struct pool, worker);
2402
2403 throttle_work_start(&pool->throttle);
2404 dm_pool_issue_prefetches(pool->pmd);
2405 throttle_work_update(&pool->throttle);
2406 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2407 throttle_work_update(&pool->throttle);
2408 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2409 throttle_work_update(&pool->throttle);
2410 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2411 throttle_work_update(&pool->throttle);
2412 process_deferred_bios(pool);
2413 throttle_work_complete(&pool->throttle);
2414}
2415
2416/*
2417 * We want to commit periodically so that not too much
2418 * unwritten data builds up.
2419 */
2420static void do_waker(struct work_struct *ws)
2421{
2422 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2423 wake_worker(pool);
2424 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2425}
2426
2427/*
2428 * We're holding onto IO to allow userland time to react. After the
2429 * timeout either the pool will have been resized (and thus back in
2430 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2431 */
2432static void do_no_space_timeout(struct work_struct *ws)
2433{
2434 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2435 no_space_timeout);
2436
2437 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2438 pool->pf.error_if_no_space = true;
2439 notify_of_pool_mode_change(pool);
2440 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2441 }
2442}
2443
2444/*----------------------------------------------------------------*/
2445
2446struct pool_work {
2447 struct work_struct worker;
2448 struct completion complete;
2449};
2450
2451static struct pool_work *to_pool_work(struct work_struct *ws)
2452{
2453 return container_of(ws, struct pool_work, worker);
2454}
2455
2456static void pool_work_complete(struct pool_work *pw)
2457{
2458 complete(&pw->complete);
2459}
2460
2461static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2462 void (*fn)(struct work_struct *))
2463{
2464 INIT_WORK_ONSTACK(&pw->worker, fn);
2465 init_completion(&pw->complete);
2466 queue_work(pool->wq, &pw->worker);
2467 wait_for_completion(&pw->complete);
2468}
2469
2470/*----------------------------------------------------------------*/
2471
2472struct noflush_work {
2473 struct pool_work pw;
2474 struct thin_c *tc;
2475};
2476
2477static struct noflush_work *to_noflush(struct work_struct *ws)
2478{
2479 return container_of(to_pool_work(ws), struct noflush_work, pw);
2480}
2481
2482static void do_noflush_start(struct work_struct *ws)
2483{
2484 struct noflush_work *w = to_noflush(ws);
2485 w->tc->requeue_mode = true;
2486 requeue_io(w->tc);
2487 pool_work_complete(&w->pw);
2488}
2489
2490static void do_noflush_stop(struct work_struct *ws)
2491{
2492 struct noflush_work *w = to_noflush(ws);
2493 w->tc->requeue_mode = false;
2494 pool_work_complete(&w->pw);
2495}
2496
2497static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2498{
2499 struct noflush_work w;
2500
2501 w.tc = tc;
2502 pool_work_wait(&w.pw, tc->pool, fn);
2503}
2504
2505/*----------------------------------------------------------------*/
2506
2507static bool passdown_enabled(struct pool_c *pt)
2508{
2509 return pt->adjusted_pf.discard_passdown;
2510}
2511
2512static void set_discard_callbacks(struct pool *pool)
2513{
2514 struct pool_c *pt = pool->ti->private;
2515
2516 if (passdown_enabled(pt)) {
2517 pool->process_discard_cell = process_discard_cell_passdown;
2518 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2519 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2520 } else {
2521 pool->process_discard_cell = process_discard_cell_no_passdown;
2522 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2523 }
2524}
2525
2526static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2527{
2528 struct pool_c *pt = pool->ti->private;
2529 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2530 enum pool_mode old_mode = get_pool_mode(pool);
2531 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2532
2533 /*
2534 * Never allow the pool to transition to PM_WRITE mode if user
2535 * intervention is required to verify metadata and data consistency.
2536 */
2537 if (new_mode == PM_WRITE && needs_check) {
2538 DMERR("%s: unable to switch pool to write mode until repaired.",
2539 dm_device_name(pool->pool_md));
2540 if (old_mode != new_mode)
2541 new_mode = old_mode;
2542 else
2543 new_mode = PM_READ_ONLY;
2544 }
2545 /*
2546 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2547 * not going to recover without a thin_repair. So we never let the
2548 * pool move out of the old mode.
2549 */
2550 if (old_mode == PM_FAIL)
2551 new_mode = old_mode;
2552
2553 switch (new_mode) {
2554 case PM_FAIL:
2555 dm_pool_metadata_read_only(pool->pmd);
2556 pool->process_bio = process_bio_fail;
2557 pool->process_discard = process_bio_fail;
2558 pool->process_cell = process_cell_fail;
2559 pool->process_discard_cell = process_cell_fail;
2560 pool->process_prepared_mapping = process_prepared_mapping_fail;
2561 pool->process_prepared_discard = process_prepared_discard_fail;
2562
2563 error_retry_list(pool);
2564 break;
2565
2566 case PM_OUT_OF_METADATA_SPACE:
2567 case PM_READ_ONLY:
2568 dm_pool_metadata_read_only(pool->pmd);
2569 pool->process_bio = process_bio_read_only;
2570 pool->process_discard = process_bio_success;
2571 pool->process_cell = process_cell_read_only;
2572 pool->process_discard_cell = process_cell_success;
2573 pool->process_prepared_mapping = process_prepared_mapping_fail;
2574 pool->process_prepared_discard = process_prepared_discard_success;
2575
2576 error_retry_list(pool);
2577 break;
2578
2579 case PM_OUT_OF_DATA_SPACE:
2580 /*
2581 * Ideally we'd never hit this state; the low water mark
2582 * would trigger userland to extend the pool before we
2583 * completely run out of data space. However, many small
2584 * IOs to unprovisioned space can consume data space at an
2585 * alarming rate. Adjust your low water mark if you're
2586 * frequently seeing this mode.
2587 */
2588 pool->out_of_data_space = true;
2589 pool->process_bio = process_bio_read_only;
2590 pool->process_discard = process_discard_bio;
2591 pool->process_cell = process_cell_read_only;
2592 pool->process_prepared_mapping = process_prepared_mapping;
2593 set_discard_callbacks(pool);
2594
2595 if (!pool->pf.error_if_no_space && no_space_timeout)
2596 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2597 break;
2598
2599 case PM_WRITE:
2600 if (old_mode == PM_OUT_OF_DATA_SPACE)
2601 cancel_delayed_work_sync(&pool->no_space_timeout);
2602 pool->out_of_data_space = false;
2603 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2604 dm_pool_metadata_read_write(pool->pmd);
2605 pool->process_bio = process_bio;
2606 pool->process_discard = process_discard_bio;
2607 pool->process_cell = process_cell;
2608 pool->process_prepared_mapping = process_prepared_mapping;
2609 set_discard_callbacks(pool);
2610 break;
2611 }
2612
2613 pool->pf.mode = new_mode;
2614 /*
2615 * The pool mode may have changed, sync it so bind_control_target()
2616 * doesn't cause an unexpected mode transition on resume.
2617 */
2618 pt->adjusted_pf.mode = new_mode;
2619
2620 if (old_mode != new_mode)
2621 notify_of_pool_mode_change(pool);
2622}
2623
2624static void abort_transaction(struct pool *pool)
2625{
2626 const char *dev_name = dm_device_name(pool->pool_md);
2627
2628 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2629 if (dm_pool_abort_metadata(pool->pmd)) {
2630 DMERR("%s: failed to abort metadata transaction", dev_name);
2631 set_pool_mode(pool, PM_FAIL);
2632 }
2633
2634 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2635 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2636 set_pool_mode(pool, PM_FAIL);
2637 }
2638}
2639
2640static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2641{
2642 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2643 dm_device_name(pool->pool_md), op, r);
2644
2645 abort_transaction(pool);
2646 set_pool_mode(pool, PM_READ_ONLY);
2647}
2648
2649/*----------------------------------------------------------------*/
2650
2651/*
2652 * Mapping functions.
2653 */
2654
2655/*
2656 * Called only while mapping a thin bio to hand it over to the workqueue.
2657 */
2658static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2659{
2660 unsigned long flags;
2661 struct pool *pool = tc->pool;
2662
2663 spin_lock_irqsave(&tc->lock, flags);
2664 bio_list_add(&tc->deferred_bio_list, bio);
2665 spin_unlock_irqrestore(&tc->lock, flags);
2666
2667 wake_worker(pool);
2668}
2669
2670static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2671{
2672 struct pool *pool = tc->pool;
2673
2674 throttle_lock(&pool->throttle);
2675 thin_defer_bio(tc, bio);
2676 throttle_unlock(&pool->throttle);
2677}
2678
2679static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2680{
2681 unsigned long flags;
2682 struct pool *pool = tc->pool;
2683
2684 throttle_lock(&pool->throttle);
2685 spin_lock_irqsave(&tc->lock, flags);
2686 list_add_tail(&cell->user_list, &tc->deferred_cells);
2687 spin_unlock_irqrestore(&tc->lock, flags);
2688 throttle_unlock(&pool->throttle);
2689
2690 wake_worker(pool);
2691}
2692
2693static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2694{
2695 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2696
2697 h->tc = tc;
2698 h->shared_read_entry = NULL;
2699 h->all_io_entry = NULL;
2700 h->overwrite_mapping = NULL;
2701 h->cell = NULL;
2702}
2703
2704/*
2705 * Non-blocking function called from the thin target's map function.
2706 */
2707static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2708{
2709 int r;
2710 struct thin_c *tc = ti->private;
2711 dm_block_t block = get_bio_block(tc, bio);
2712 struct dm_thin_device *td = tc->td;
2713 struct dm_thin_lookup_result result;
2714 struct dm_bio_prison_cell *virt_cell, *data_cell;
2715 struct dm_cell_key key;
2716
2717 thin_hook_bio(tc, bio);
2718
2719 if (tc->requeue_mode) {
2720 bio->bi_status = BLK_STS_DM_REQUEUE;
2721 bio_endio(bio);
2722 return DM_MAPIO_SUBMITTED;
2723 }
2724
2725 if (get_pool_mode(tc->pool) == PM_FAIL) {
2726 bio_io_error(bio);
2727 return DM_MAPIO_SUBMITTED;
2728 }
2729
2730 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2731 thin_defer_bio_with_throttle(tc, bio);
2732 return DM_MAPIO_SUBMITTED;
2733 }
2734
2735 /*
2736 * We must hold the virtual cell before doing the lookup, otherwise
2737 * there's a race with discard.
2738 */
2739 build_virtual_key(tc->td, block, &key);
2740 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2741 return DM_MAPIO_SUBMITTED;
2742
2743 r = dm_thin_find_block(td, block, 0, &result);
2744
2745 /*
2746 * Note that we defer readahead too.
2747 */
2748 switch (r) {
2749 case 0:
2750 if (unlikely(result.shared)) {
2751 /*
2752 * We have a race condition here between the
2753 * result.shared value returned by the lookup and
2754 * snapshot creation, which may cause new
2755 * sharing.
2756 *
2757 * To avoid this always quiesce the origin before
2758 * taking the snap. You want to do this anyway to
2759 * ensure a consistent application view
2760 * (i.e. lockfs).
2761 *
2762 * More distant ancestors are irrelevant. The
2763 * shared flag will be set in their case.
2764 */
2765 thin_defer_cell(tc, virt_cell);
2766 return DM_MAPIO_SUBMITTED;
2767 }
2768
2769 build_data_key(tc->td, result.block, &key);
2770 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2771 cell_defer_no_holder(tc, virt_cell);
2772 return DM_MAPIO_SUBMITTED;
2773 }
2774
2775 inc_all_io_entry(tc->pool, bio);
2776 cell_defer_no_holder(tc, data_cell);
2777 cell_defer_no_holder(tc, virt_cell);
2778
2779 remap(tc, bio, result.block);
2780 return DM_MAPIO_REMAPPED;
2781
2782 case -ENODATA:
2783 case -EWOULDBLOCK:
2784 thin_defer_cell(tc, virt_cell);
2785 return DM_MAPIO_SUBMITTED;
2786
2787 default:
2788 /*
2789 * Must always call bio_io_error on failure.
2790 * dm_thin_find_block can fail with -EINVAL if the
2791 * pool is switched to fail-io mode.
2792 */
2793 bio_io_error(bio);
2794 cell_defer_no_holder(tc, virt_cell);
2795 return DM_MAPIO_SUBMITTED;
2796 }
2797}
2798
2799static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2800{
2801 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2802 struct request_queue *q;
2803
2804 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2805 return 1;
2806
2807 q = bdev_get_queue(pt->data_dev->bdev);
2808 return bdi_congested(q->backing_dev_info, bdi_bits);
2809}
2810
2811static void requeue_bios(struct pool *pool)
2812{
2813 unsigned long flags;
2814 struct thin_c *tc;
2815
2816 rcu_read_lock();
2817 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2818 spin_lock_irqsave(&tc->lock, flags);
2819 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2820 bio_list_init(&tc->retry_on_resume_list);
2821 spin_unlock_irqrestore(&tc->lock, flags);
2822 }
2823 rcu_read_unlock();
2824}
2825
2826/*----------------------------------------------------------------
2827 * Binding of control targets to a pool object
2828 *--------------------------------------------------------------*/
2829static bool data_dev_supports_discard(struct pool_c *pt)
2830{
2831 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2832
2833 return q && blk_queue_discard(q);
2834}
2835
2836static bool is_factor(sector_t block_size, uint32_t n)
2837{
2838 return !sector_div(block_size, n);
2839}
2840
2841/*
2842 * If discard_passdown was enabled verify that the data device
2843 * supports discards. Disable discard_passdown if not.
2844 */
2845static void disable_passdown_if_not_supported(struct pool_c *pt)
2846{
2847 struct pool *pool = pt->pool;
2848 struct block_device *data_bdev = pt->data_dev->bdev;
2849 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2850 const char *reason = NULL;
2851 char buf[BDEVNAME_SIZE];
2852
2853 if (!pt->adjusted_pf.discard_passdown)
2854 return;
2855
2856 if (!data_dev_supports_discard(pt))
2857 reason = "discard unsupported";
2858
2859 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2860 reason = "max discard sectors smaller than a block";
2861
2862 if (reason) {
2863 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2864 pt->adjusted_pf.discard_passdown = false;
2865 }
2866}
2867
2868static int bind_control_target(struct pool *pool, struct dm_target *ti)
2869{
2870 struct pool_c *pt = ti->private;
2871
2872 /*
2873 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2874 */
2875 enum pool_mode old_mode = get_pool_mode(pool);
2876 enum pool_mode new_mode = pt->adjusted_pf.mode;
2877
2878 /*
2879 * Don't change the pool's mode until set_pool_mode() below.
2880 * Otherwise the pool's process_* function pointers may
2881 * not match the desired pool mode.
2882 */
2883 pt->adjusted_pf.mode = old_mode;
2884
2885 pool->ti = ti;
2886 pool->pf = pt->adjusted_pf;
2887 pool->low_water_blocks = pt->low_water_blocks;
2888
2889 set_pool_mode(pool, new_mode);
2890
2891 return 0;
2892}
2893
2894static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2895{
2896 if (pool->ti == ti)
2897 pool->ti = NULL;
2898}
2899
2900/*----------------------------------------------------------------
2901 * Pool creation
2902 *--------------------------------------------------------------*/
2903/* Initialize pool features. */
2904static void pool_features_init(struct pool_features *pf)
2905{
2906 pf->mode = PM_WRITE;
2907 pf->zero_new_blocks = true;
2908 pf->discard_enabled = true;
2909 pf->discard_passdown = true;
2910 pf->error_if_no_space = false;
2911}
2912
2913static void __pool_destroy(struct pool *pool)
2914{
2915 __pool_table_remove(pool);
2916
2917 vfree(pool->cell_sort_array);
2918 if (dm_pool_metadata_close(pool->pmd) < 0)
2919 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2920
2921 dm_bio_prison_destroy(pool->prison);
2922 dm_kcopyd_client_destroy(pool->copier);
2923
2924 if (pool->wq)
2925 destroy_workqueue(pool->wq);
2926
2927 if (pool->next_mapping)
2928 mempool_free(pool->next_mapping, &pool->mapping_pool);
2929 mempool_exit(&pool->mapping_pool);
2930 dm_deferred_set_destroy(pool->shared_read_ds);
2931 dm_deferred_set_destroy(pool->all_io_ds);
2932 kfree(pool);
2933}
2934
2935static struct kmem_cache *_new_mapping_cache;
2936
2937static struct pool *pool_create(struct mapped_device *pool_md,
2938 struct block_device *metadata_dev,
2939 unsigned long block_size,
2940 int read_only, char **error)
2941{
2942 int r;
2943 void *err_p;
2944 struct pool *pool;
2945 struct dm_pool_metadata *pmd;
2946 bool format_device = read_only ? false : true;
2947
2948 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2949 if (IS_ERR(pmd)) {
2950 *error = "Error creating metadata object";
2951 return (struct pool *)pmd;
2952 }
2953
2954 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2955 if (!pool) {
2956 *error = "Error allocating memory for pool";
2957 err_p = ERR_PTR(-ENOMEM);
2958 goto bad_pool;
2959 }
2960
2961 pool->pmd = pmd;
2962 pool->sectors_per_block = block_size;
2963 if (block_size & (block_size - 1))
2964 pool->sectors_per_block_shift = -1;
2965 else
2966 pool->sectors_per_block_shift = __ffs(block_size);
2967 pool->low_water_blocks = 0;
2968 pool_features_init(&pool->pf);
2969 pool->prison = dm_bio_prison_create();
2970 if (!pool->prison) {
2971 *error = "Error creating pool's bio prison";
2972 err_p = ERR_PTR(-ENOMEM);
2973 goto bad_prison;
2974 }
2975
2976 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2977 if (IS_ERR(pool->copier)) {
2978 r = PTR_ERR(pool->copier);
2979 *error = "Error creating pool's kcopyd client";
2980 err_p = ERR_PTR(r);
2981 goto bad_kcopyd_client;
2982 }
2983
2984 /*
2985 * Create singlethreaded workqueue that will service all devices
2986 * that use this metadata.
2987 */
2988 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2989 if (!pool->wq) {
2990 *error = "Error creating pool's workqueue";
2991 err_p = ERR_PTR(-ENOMEM);
2992 goto bad_wq;
2993 }
2994
2995 throttle_init(&pool->throttle);
2996 INIT_WORK(&pool->worker, do_worker);
2997 INIT_DELAYED_WORK(&pool->waker, do_waker);
2998 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2999 spin_lock_init(&pool->lock);
3000 bio_list_init(&pool->deferred_flush_bios);
3001 bio_list_init(&pool->deferred_flush_completions);
3002 INIT_LIST_HEAD(&pool->prepared_mappings);
3003 INIT_LIST_HEAD(&pool->prepared_discards);
3004 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
3005 INIT_LIST_HEAD(&pool->active_thins);
3006 pool->low_water_triggered = false;
3007 pool->suspended = true;
3008 pool->out_of_data_space = false;
3009
3010 pool->shared_read_ds = dm_deferred_set_create();
3011 if (!pool->shared_read_ds) {
3012 *error = "Error creating pool's shared read deferred set";
3013 err_p = ERR_PTR(-ENOMEM);
3014 goto bad_shared_read_ds;
3015 }
3016
3017 pool->all_io_ds = dm_deferred_set_create();
3018 if (!pool->all_io_ds) {
3019 *error = "Error creating pool's all io deferred set";
3020 err_p = ERR_PTR(-ENOMEM);
3021 goto bad_all_io_ds;
3022 }
3023
3024 pool->next_mapping = NULL;
3025 r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
3026 _new_mapping_cache);
3027 if (r) {
3028 *error = "Error creating pool's mapping mempool";
3029 err_p = ERR_PTR(r);
3030 goto bad_mapping_pool;
3031 }
3032
3033 pool->cell_sort_array =
3034 vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3035 sizeof(*pool->cell_sort_array)));
3036 if (!pool->cell_sort_array) {
3037 *error = "Error allocating cell sort array";
3038 err_p = ERR_PTR(-ENOMEM);
3039 goto bad_sort_array;
3040 }
3041
3042 pool->ref_count = 1;
3043 pool->last_commit_jiffies = jiffies;
3044 pool->pool_md = pool_md;
3045 pool->md_dev = metadata_dev;
3046 __pool_table_insert(pool);
3047
3048 return pool;
3049
3050bad_sort_array:
3051 mempool_exit(&pool->mapping_pool);
3052bad_mapping_pool:
3053 dm_deferred_set_destroy(pool->all_io_ds);
3054bad_all_io_ds:
3055 dm_deferred_set_destroy(pool->shared_read_ds);
3056bad_shared_read_ds:
3057 destroy_workqueue(pool->wq);
3058bad_wq:
3059 dm_kcopyd_client_destroy(pool->copier);
3060bad_kcopyd_client:
3061 dm_bio_prison_destroy(pool->prison);
3062bad_prison:
3063 kfree(pool);
3064bad_pool:
3065 if (dm_pool_metadata_close(pmd))
3066 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3067
3068 return err_p;
3069}
3070
3071static void __pool_inc(struct pool *pool)
3072{
3073 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3074 pool->ref_count++;
3075}
3076
3077static void __pool_dec(struct pool *pool)
3078{
3079 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3080 BUG_ON(!pool->ref_count);
3081 if (!--pool->ref_count)
3082 __pool_destroy(pool);
3083}
3084
3085static struct pool *__pool_find(struct mapped_device *pool_md,
3086 struct block_device *metadata_dev,
3087 unsigned long block_size, int read_only,
3088 char **error, int *created)
3089{
3090 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3091
3092 if (pool) {
3093 if (pool->pool_md != pool_md) {
3094 *error = "metadata device already in use by a pool";
3095 return ERR_PTR(-EBUSY);
3096 }
3097 __pool_inc(pool);
3098
3099 } else {
3100 pool = __pool_table_lookup(pool_md);
3101 if (pool) {
3102 if (pool->md_dev != metadata_dev) {
3103 *error = "different pool cannot replace a pool";
3104 return ERR_PTR(-EINVAL);
3105 }
3106 __pool_inc(pool);
3107
3108 } else {
3109 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3110 *created = 1;
3111 }
3112 }
3113
3114 return pool;
3115}
3116
3117/*----------------------------------------------------------------
3118 * Pool target methods
3119 *--------------------------------------------------------------*/
3120static void pool_dtr(struct dm_target *ti)
3121{
3122 struct pool_c *pt = ti->private;
3123
3124 mutex_lock(&dm_thin_pool_table.mutex);
3125
3126 unbind_control_target(pt->pool, ti);
3127 __pool_dec(pt->pool);
3128 dm_put_device(ti, pt->metadata_dev);
3129 dm_put_device(ti, pt->data_dev);
3130 kfree(pt);
3131
3132 mutex_unlock(&dm_thin_pool_table.mutex);
3133}
3134
3135static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3136 struct dm_target *ti)
3137{
3138 int r;
3139 unsigned argc;
3140 const char *arg_name;
3141
3142 static const struct dm_arg _args[] = {
3143 {0, 4, "Invalid number of pool feature arguments"},
3144 };
3145
3146 /*
3147 * No feature arguments supplied.
3148 */
3149 if (!as->argc)
3150 return 0;
3151
3152 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3153 if (r)
3154 return -EINVAL;
3155
3156 while (argc && !r) {
3157 arg_name = dm_shift_arg(as);
3158 argc--;
3159
3160 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3161 pf->zero_new_blocks = false;
3162
3163 else if (!strcasecmp(arg_name, "ignore_discard"))
3164 pf->discard_enabled = false;
3165
3166 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3167 pf->discard_passdown = false;
3168
3169 else if (!strcasecmp(arg_name, "read_only"))
3170 pf->mode = PM_READ_ONLY;
3171
3172 else if (!strcasecmp(arg_name, "error_if_no_space"))
3173 pf->error_if_no_space = true;
3174
3175 else {
3176 ti->error = "Unrecognised pool feature requested";
3177 r = -EINVAL;
3178 break;
3179 }
3180 }
3181
3182 return r;
3183}
3184
3185static void metadata_low_callback(void *context)
3186{
3187 struct pool *pool = context;
3188
3189 DMWARN("%s: reached low water mark for metadata device: sending event.",
3190 dm_device_name(pool->pool_md));
3191
3192 dm_table_event(pool->ti->table);
3193}
3194
3195static sector_t get_dev_size(struct block_device *bdev)
3196{
3197 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3198}
3199
3200static void warn_if_metadata_device_too_big(struct block_device *bdev)
3201{
3202 sector_t metadata_dev_size = get_dev_size(bdev);
3203 char buffer[BDEVNAME_SIZE];
3204
3205 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3206 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3207 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3208}
3209
3210static sector_t get_metadata_dev_size(struct block_device *bdev)
3211{
3212 sector_t metadata_dev_size = get_dev_size(bdev);
3213
3214 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3215 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3216
3217 return metadata_dev_size;
3218}
3219
3220static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3221{
3222 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3223
3224 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3225
3226 return metadata_dev_size;
3227}
3228
3229/*
3230 * When a metadata threshold is crossed a dm event is triggered, and
3231 * userland should respond by growing the metadata device. We could let
3232 * userland set the threshold, like we do with the data threshold, but I'm
3233 * not sure they know enough to do this well.
3234 */
3235static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3236{
3237 /*
3238 * 4M is ample for all ops with the possible exception of thin
3239 * device deletion which is harmless if it fails (just retry the
3240 * delete after you've grown the device).
3241 */
3242 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3243 return min((dm_block_t)1024ULL /* 4M */, quarter);
3244}
3245
3246/*
3247 * thin-pool <metadata dev> <data dev>
3248 * <data block size (sectors)>
3249 * <low water mark (blocks)>
3250 * [<#feature args> [<arg>]*]
3251 *
3252 * Optional feature arguments are:
3253 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3254 * ignore_discard: disable discard
3255 * no_discard_passdown: don't pass discards down to the data device
3256 * read_only: Don't allow any changes to be made to the pool metadata.
3257 * error_if_no_space: error IOs, instead of queueing, if no space.
3258 */
3259static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3260{
3261 int r, pool_created = 0;
3262 struct pool_c *pt;
3263 struct pool *pool;
3264 struct pool_features pf;
3265 struct dm_arg_set as;
3266 struct dm_dev *data_dev;
3267 unsigned long block_size;
3268 dm_block_t low_water_blocks;
3269 struct dm_dev *metadata_dev;
3270 fmode_t metadata_mode;
3271
3272 /*
3273 * FIXME Remove validation from scope of lock.
3274 */
3275 mutex_lock(&dm_thin_pool_table.mutex);
3276
3277 if (argc < 4) {
3278 ti->error = "Invalid argument count";
3279 r = -EINVAL;
3280 goto out_unlock;
3281 }
3282
3283 as.argc = argc;
3284 as.argv = argv;
3285
3286 /* make sure metadata and data are different devices */
3287 if (!strcmp(argv[0], argv[1])) {
3288 ti->error = "Error setting metadata or data device";
3289 r = -EINVAL;
3290 goto out_unlock;
3291 }
3292
3293 /*
3294 * Set default pool features.
3295 */
3296 pool_features_init(&pf);
3297
3298 dm_consume_args(&as, 4);
3299 r = parse_pool_features(&as, &pf, ti);
3300 if (r)
3301 goto out_unlock;
3302
3303 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3304 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3305 if (r) {
3306 ti->error = "Error opening metadata block device";
3307 goto out_unlock;
3308 }
3309 warn_if_metadata_device_too_big(metadata_dev->bdev);
3310
3311 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3312 if (r) {
3313 ti->error = "Error getting data device";
3314 goto out_metadata;
3315 }
3316
3317 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3318 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3319 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3320 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3321 ti->error = "Invalid block size";
3322 r = -EINVAL;
3323 goto out;
3324 }
3325
3326 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3327 ti->error = "Invalid low water mark";
3328 r = -EINVAL;
3329 goto out;
3330 }
3331
3332 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3333 if (!pt) {
3334 r = -ENOMEM;
3335 goto out;
3336 }
3337
3338 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3339 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3340 if (IS_ERR(pool)) {
3341 r = PTR_ERR(pool);
3342 goto out_free_pt;
3343 }
3344
3345 /*
3346 * 'pool_created' reflects whether this is the first table load.
3347 * Top level discard support is not allowed to be changed after
3348 * initial load. This would require a pool reload to trigger thin
3349 * device changes.
3350 */
3351 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3352 ti->error = "Discard support cannot be disabled once enabled";
3353 r = -EINVAL;
3354 goto out_flags_changed;
3355 }
3356
3357 pt->pool = pool;
3358 pt->ti = ti;
3359 pt->metadata_dev = metadata_dev;
3360 pt->data_dev = data_dev;
3361 pt->low_water_blocks = low_water_blocks;
3362 pt->adjusted_pf = pt->requested_pf = pf;
3363 ti->num_flush_bios = 1;
3364
3365 /*
3366 * Only need to enable discards if the pool should pass
3367 * them down to the data device. The thin device's discard
3368 * processing will cause mappings to be removed from the btree.
3369 */
3370 if (pf.discard_enabled && pf.discard_passdown) {
3371 ti->num_discard_bios = 1;
3372
3373 /*
3374 * Setting 'discards_supported' circumvents the normal
3375 * stacking of discard limits (this keeps the pool and
3376 * thin devices' discard limits consistent).
3377 */
3378 ti->discards_supported = true;
3379 }
3380 ti->private = pt;
3381
3382 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3383 calc_metadata_threshold(pt),
3384 metadata_low_callback,
3385 pool);
3386 if (r)
3387 goto out_flags_changed;
3388
3389 pt->callbacks.congested_fn = pool_is_congested;
3390 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3391
3392 mutex_unlock(&dm_thin_pool_table.mutex);
3393
3394 return 0;
3395
3396out_flags_changed:
3397 __pool_dec(pool);
3398out_free_pt:
3399 kfree(pt);
3400out:
3401 dm_put_device(ti, data_dev);
3402out_metadata:
3403 dm_put_device(ti, metadata_dev);
3404out_unlock:
3405 mutex_unlock(&dm_thin_pool_table.mutex);
3406
3407 return r;
3408}
3409
3410static int pool_map(struct dm_target *ti, struct bio *bio)
3411{
3412 int r;
3413 struct pool_c *pt = ti->private;
3414 struct pool *pool = pt->pool;
3415 unsigned long flags;
3416
3417 /*
3418 * As this is a singleton target, ti->begin is always zero.
3419 */
3420 spin_lock_irqsave(&pool->lock, flags);
3421 bio_set_dev(bio, pt->data_dev->bdev);
3422 r = DM_MAPIO_REMAPPED;
3423 spin_unlock_irqrestore(&pool->lock, flags);
3424
3425 return r;
3426}
3427
3428static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3429{
3430 int r;
3431 struct pool_c *pt = ti->private;
3432 struct pool *pool = pt->pool;
3433 sector_t data_size = ti->len;
3434 dm_block_t sb_data_size;
3435
3436 *need_commit = false;
3437
3438 (void) sector_div(data_size, pool->sectors_per_block);
3439
3440 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3441 if (r) {
3442 DMERR("%s: failed to retrieve data device size",
3443 dm_device_name(pool->pool_md));
3444 return r;
3445 }
3446
3447 if (data_size < sb_data_size) {
3448 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3449 dm_device_name(pool->pool_md),
3450 (unsigned long long)data_size, sb_data_size);
3451 return -EINVAL;
3452
3453 } else if (data_size > sb_data_size) {
3454 if (dm_pool_metadata_needs_check(pool->pmd)) {
3455 DMERR("%s: unable to grow the data device until repaired.",
3456 dm_device_name(pool->pool_md));
3457 return 0;
3458 }
3459
3460 if (sb_data_size)
3461 DMINFO("%s: growing the data device from %llu to %llu blocks",
3462 dm_device_name(pool->pool_md),
3463 sb_data_size, (unsigned long long)data_size);
3464 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3465 if (r) {
3466 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3467 return r;
3468 }
3469
3470 *need_commit = true;
3471 }
3472
3473 return 0;
3474}
3475
3476static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3477{
3478 int r;
3479 struct pool_c *pt = ti->private;
3480 struct pool *pool = pt->pool;
3481 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3482
3483 *need_commit = false;
3484
3485 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3486
3487 r =