Loading...
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 && time_is_before_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 *data_dev;
235 struct block_device *md_dev;
236 struct dm_pool_metadata *pmd;
237
238 dm_block_t low_water_blocks;
239 uint32_t sectors_per_block;
240 int sectors_per_block_shift;
241
242 struct pool_features pf;
243 bool low_water_triggered:1; /* A dm event has been sent */
244 bool suspended:1;
245 bool out_of_data_space:1;
246
247 struct dm_bio_prison *prison;
248 struct dm_kcopyd_client *copier;
249
250 struct work_struct worker;
251 struct workqueue_struct *wq;
252 struct throttle throttle;
253 struct delayed_work waker;
254 struct delayed_work no_space_timeout;
255
256 unsigned long last_commit_jiffies;
257 unsigned ref_count;
258
259 spinlock_t lock;
260 struct bio_list deferred_flush_bios;
261 struct bio_list deferred_flush_completions;
262 struct list_head prepared_mappings;
263 struct list_head prepared_discards;
264 struct list_head prepared_discards_pt2;
265 struct list_head active_thins;
266
267 struct dm_deferred_set *shared_read_ds;
268 struct dm_deferred_set *all_io_ds;
269
270 struct dm_thin_new_mapping *next_mapping;
271
272 process_bio_fn process_bio;
273 process_bio_fn process_discard;
274
275 process_cell_fn process_cell;
276 process_cell_fn process_discard_cell;
277
278 process_mapping_fn process_prepared_mapping;
279 process_mapping_fn process_prepared_discard;
280 process_mapping_fn process_prepared_discard_pt2;
281
282 struct dm_bio_prison_cell **cell_sort_array;
283
284 mempool_t mapping_pool;
285};
286
287static void metadata_operation_failed(struct pool *pool, const char *op, int r);
288
289static enum pool_mode get_pool_mode(struct pool *pool)
290{
291 return pool->pf.mode;
292}
293
294static void notify_of_pool_mode_change(struct pool *pool)
295{
296 const char *descs[] = {
297 "write",
298 "out-of-data-space",
299 "read-only",
300 "read-only",
301 "fail"
302 };
303 const char *extra_desc = NULL;
304 enum pool_mode mode = get_pool_mode(pool);
305
306 if (mode == PM_OUT_OF_DATA_SPACE) {
307 if (!pool->pf.error_if_no_space)
308 extra_desc = " (queue IO)";
309 else
310 extra_desc = " (error IO)";
311 }
312
313 dm_table_event(pool->ti->table);
314 DMINFO("%s: switching pool to %s%s mode",
315 dm_device_name(pool->pool_md),
316 descs[(int)mode], extra_desc ? : "");
317}
318
319/*
320 * Target context for a pool.
321 */
322struct pool_c {
323 struct dm_target *ti;
324 struct pool *pool;
325 struct dm_dev *data_dev;
326 struct dm_dev *metadata_dev;
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, GFP_NOWAIT,
402 &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 op->bio->bi_opf = REQ_OP_DISCARD;
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
613 bio_list_init(&bios);
614
615 spin_lock_irq(&tc->lock);
616 __merge_bio_list(&bios, master);
617 spin_unlock_irq(&tc->lock);
618
619 error_bio_list(&bios, error);
620}
621
622static void requeue_deferred_cells(struct thin_c *tc)
623{
624 struct pool *pool = tc->pool;
625 struct list_head cells;
626 struct dm_bio_prison_cell *cell, *tmp;
627
628 INIT_LIST_HEAD(&cells);
629
630 spin_lock_irq(&tc->lock);
631 list_splice_init(&tc->deferred_cells, &cells);
632 spin_unlock_irq(&tc->lock);
633
634 list_for_each_entry_safe(cell, tmp, &cells, user_list)
635 cell_requeue(pool, cell);
636}
637
638static void requeue_io(struct thin_c *tc)
639{
640 struct bio_list bios;
641
642 bio_list_init(&bios);
643
644 spin_lock_irq(&tc->lock);
645 __merge_bio_list(&bios, &tc->deferred_bio_list);
646 __merge_bio_list(&bios, &tc->retry_on_resume_list);
647 spin_unlock_irq(&tc->lock);
648
649 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
650 requeue_deferred_cells(tc);
651}
652
653static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
654{
655 struct thin_c *tc;
656
657 rcu_read_lock();
658 list_for_each_entry_rcu(tc, &pool->active_thins, list)
659 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
660 rcu_read_unlock();
661}
662
663static void error_retry_list(struct pool *pool)
664{
665 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
666}
667
668/*
669 * This section of code contains the logic for processing a thin device's IO.
670 * Much of the code depends on pool object resources (lists, workqueues, etc)
671 * but most is exclusively called from the thin target rather than the thin-pool
672 * target.
673 */
674
675static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
676{
677 struct pool *pool = tc->pool;
678 sector_t block_nr = bio->bi_iter.bi_sector;
679
680 if (block_size_is_power_of_two(pool))
681 block_nr >>= pool->sectors_per_block_shift;
682 else
683 (void) sector_div(block_nr, pool->sectors_per_block);
684
685 return block_nr;
686}
687
688/*
689 * Returns the _complete_ blocks that this bio covers.
690 */
691static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
692 dm_block_t *begin, dm_block_t *end)
693{
694 struct pool *pool = tc->pool;
695 sector_t b = bio->bi_iter.bi_sector;
696 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
697
698 b += pool->sectors_per_block - 1ull; /* so we round up */
699
700 if (block_size_is_power_of_two(pool)) {
701 b >>= pool->sectors_per_block_shift;
702 e >>= pool->sectors_per_block_shift;
703 } else {
704 (void) sector_div(b, pool->sectors_per_block);
705 (void) sector_div(e, pool->sectors_per_block);
706 }
707
708 if (e < b)
709 /* Can happen if the bio is within a single block. */
710 e = b;
711
712 *begin = b;
713 *end = e;
714}
715
716static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
717{
718 struct pool *pool = tc->pool;
719 sector_t bi_sector = bio->bi_iter.bi_sector;
720
721 bio_set_dev(bio, tc->pool_dev->bdev);
722 if (block_size_is_power_of_two(pool))
723 bio->bi_iter.bi_sector =
724 (block << pool->sectors_per_block_shift) |
725 (bi_sector & (pool->sectors_per_block - 1));
726 else
727 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
728 sector_div(bi_sector, pool->sectors_per_block);
729}
730
731static void remap_to_origin(struct thin_c *tc, struct bio *bio)
732{
733 bio_set_dev(bio, tc->origin_dev->bdev);
734}
735
736static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
737{
738 return op_is_flush(bio->bi_opf) &&
739 dm_thin_changed_this_transaction(tc->td);
740}
741
742static void inc_all_io_entry(struct pool *pool, struct bio *bio)
743{
744 struct dm_thin_endio_hook *h;
745
746 if (bio_op(bio) == REQ_OP_DISCARD)
747 return;
748
749 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
750 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
751}
752
753static void issue(struct thin_c *tc, struct bio *bio)
754{
755 struct pool *pool = tc->pool;
756
757 if (!bio_triggers_commit(tc, bio)) {
758 dm_submit_bio_remap(bio, NULL);
759 return;
760 }
761
762 /*
763 * Complete bio with an error if earlier I/O caused changes to
764 * the metadata that can't be committed e.g, due to I/O errors
765 * on the metadata device.
766 */
767 if (dm_thin_aborted_changes(tc->td)) {
768 bio_io_error(bio);
769 return;
770 }
771
772 /*
773 * Batch together any bios that trigger commits and then issue a
774 * single commit for them in process_deferred_bios().
775 */
776 spin_lock_irq(&pool->lock);
777 bio_list_add(&pool->deferred_flush_bios, bio);
778 spin_unlock_irq(&pool->lock);
779}
780
781static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
782{
783 remap_to_origin(tc, bio);
784 issue(tc, bio);
785}
786
787static void remap_and_issue(struct thin_c *tc, struct bio *bio,
788 dm_block_t block)
789{
790 remap(tc, bio, block);
791 issue(tc, bio);
792}
793
794/*----------------------------------------------------------------*/
795
796/*
797 * Bio endio functions.
798 */
799struct dm_thin_new_mapping {
800 struct list_head list;
801
802 bool pass_discard:1;
803 bool maybe_shared:1;
804
805 /*
806 * Track quiescing, copying and zeroing preparation actions. When this
807 * counter hits zero the block is prepared and can be inserted into the
808 * btree.
809 */
810 atomic_t prepare_actions;
811
812 blk_status_t status;
813 struct thin_c *tc;
814 dm_block_t virt_begin, virt_end;
815 dm_block_t data_block;
816 struct dm_bio_prison_cell *cell;
817
818 /*
819 * If the bio covers the whole area of a block then we can avoid
820 * zeroing or copying. Instead this bio is hooked. The bio will
821 * still be in the cell, so care has to be taken to avoid issuing
822 * the bio twice.
823 */
824 struct bio *bio;
825 bio_end_io_t *saved_bi_end_io;
826};
827
828static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
829{
830 struct pool *pool = m->tc->pool;
831
832 if (atomic_dec_and_test(&m->prepare_actions)) {
833 list_add_tail(&m->list, &pool->prepared_mappings);
834 wake_worker(pool);
835 }
836}
837
838static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
839{
840 unsigned long flags;
841 struct pool *pool = m->tc->pool;
842
843 spin_lock_irqsave(&pool->lock, flags);
844 __complete_mapping_preparation(m);
845 spin_unlock_irqrestore(&pool->lock, flags);
846}
847
848static void copy_complete(int read_err, unsigned long write_err, void *context)
849{
850 struct dm_thin_new_mapping *m = context;
851
852 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
853 complete_mapping_preparation(m);
854}
855
856static void overwrite_endio(struct bio *bio)
857{
858 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
859 struct dm_thin_new_mapping *m = h->overwrite_mapping;
860
861 bio->bi_end_io = m->saved_bi_end_io;
862
863 m->status = bio->bi_status;
864 complete_mapping_preparation(m);
865}
866
867/*----------------------------------------------------------------*/
868
869/*
870 * Workqueue.
871 */
872
873/*
874 * Prepared mapping jobs.
875 */
876
877/*
878 * This sends the bios in the cell, except the original holder, back
879 * to the deferred_bios list.
880 */
881static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
882{
883 struct pool *pool = tc->pool;
884 unsigned long flags;
885 int has_work;
886
887 spin_lock_irqsave(&tc->lock, flags);
888 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
889 has_work = !bio_list_empty(&tc->deferred_bio_list);
890 spin_unlock_irqrestore(&tc->lock, flags);
891
892 if (has_work)
893 wake_worker(pool);
894}
895
896static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
897
898struct remap_info {
899 struct thin_c *tc;
900 struct bio_list defer_bios;
901 struct bio_list issue_bios;
902};
903
904static void __inc_remap_and_issue_cell(void *context,
905 struct dm_bio_prison_cell *cell)
906{
907 struct remap_info *info = context;
908 struct bio *bio;
909
910 while ((bio = bio_list_pop(&cell->bios))) {
911 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
912 bio_list_add(&info->defer_bios, bio);
913 else {
914 inc_all_io_entry(info->tc->pool, bio);
915
916 /*
917 * We can't issue the bios with the bio prison lock
918 * held, so we add them to a list to issue on
919 * return from this function.
920 */
921 bio_list_add(&info->issue_bios, bio);
922 }
923 }
924}
925
926static void inc_remap_and_issue_cell(struct thin_c *tc,
927 struct dm_bio_prison_cell *cell,
928 dm_block_t block)
929{
930 struct bio *bio;
931 struct remap_info info;
932
933 info.tc = tc;
934 bio_list_init(&info.defer_bios);
935 bio_list_init(&info.issue_bios);
936
937 /*
938 * We have to be careful to inc any bios we're about to issue
939 * before the cell is released, and avoid a race with new bios
940 * being added to the cell.
941 */
942 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
943 &info, cell);
944
945 while ((bio = bio_list_pop(&info.defer_bios)))
946 thin_defer_bio(tc, bio);
947
948 while ((bio = bio_list_pop(&info.issue_bios)))
949 remap_and_issue(info.tc, bio, block);
950}
951
952static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
953{
954 cell_error(m->tc->pool, m->cell);
955 list_del(&m->list);
956 mempool_free(m, &m->tc->pool->mapping_pool);
957}
958
959static void complete_overwrite_bio(struct thin_c *tc, struct bio *bio)
960{
961 struct pool *pool = tc->pool;
962
963 /*
964 * If the bio has the REQ_FUA flag set we must commit the metadata
965 * before signaling its completion.
966 */
967 if (!bio_triggers_commit(tc, bio)) {
968 bio_endio(bio);
969 return;
970 }
971
972 /*
973 * Complete bio with an error if earlier I/O caused changes to the
974 * metadata that can't be committed, e.g, due to I/O errors on the
975 * metadata device.
976 */
977 if (dm_thin_aborted_changes(tc->td)) {
978 bio_io_error(bio);
979 return;
980 }
981
982 /*
983 * Batch together any bios that trigger commits and then issue a
984 * single commit for them in process_deferred_bios().
985 */
986 spin_lock_irq(&pool->lock);
987 bio_list_add(&pool->deferred_flush_completions, bio);
988 spin_unlock_irq(&pool->lock);
989}
990
991static void process_prepared_mapping(struct dm_thin_new_mapping *m)
992{
993 struct thin_c *tc = m->tc;
994 struct pool *pool = tc->pool;
995 struct bio *bio = m->bio;
996 int r;
997
998 if (m->status) {
999 cell_error(pool, m->cell);
1000 goto out;
1001 }
1002
1003 /*
1004 * Commit the prepared block into the mapping btree.
1005 * Any I/O for this block arriving after this point will get
1006 * remapped to it directly.
1007 */
1008 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1009 if (r) {
1010 metadata_operation_failed(pool, "dm_thin_insert_block", r);
1011 cell_error(pool, m->cell);
1012 goto out;
1013 }
1014
1015 /*
1016 * Release any bios held while the block was being provisioned.
1017 * If we are processing a write bio that completely covers the block,
1018 * we already processed it so can ignore it now when processing
1019 * the bios in the cell.
1020 */
1021 if (bio) {
1022 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1023 complete_overwrite_bio(tc, bio);
1024 } else {
1025 inc_all_io_entry(tc->pool, m->cell->holder);
1026 remap_and_issue(tc, m->cell->holder, m->data_block);
1027 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1028 }
1029
1030out:
1031 list_del(&m->list);
1032 mempool_free(m, &pool->mapping_pool);
1033}
1034
1035/*----------------------------------------------------------------*/
1036
1037static void free_discard_mapping(struct dm_thin_new_mapping *m)
1038{
1039 struct thin_c *tc = m->tc;
1040 if (m->cell)
1041 cell_defer_no_holder(tc, m->cell);
1042 mempool_free(m, &tc->pool->mapping_pool);
1043}
1044
1045static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
1046{
1047 bio_io_error(m->bio);
1048 free_discard_mapping(m);
1049}
1050
1051static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
1052{
1053 bio_endio(m->bio);
1054 free_discard_mapping(m);
1055}
1056
1057static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
1058{
1059 int r;
1060 struct thin_c *tc = m->tc;
1061
1062 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
1063 if (r) {
1064 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1065 bio_io_error(m->bio);
1066 } else
1067 bio_endio(m->bio);
1068
1069 cell_defer_no_holder(tc, m->cell);
1070 mempool_free(m, &tc->pool->mapping_pool);
1071}
1072
1073/*----------------------------------------------------------------*/
1074
1075static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1076 struct bio *discard_parent)
1077{
1078 /*
1079 * We've already unmapped this range of blocks, but before we
1080 * passdown we have to check that these blocks are now unused.
1081 */
1082 int r = 0;
1083 bool shared = true;
1084 struct thin_c *tc = m->tc;
1085 struct pool *pool = tc->pool;
1086 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1087 struct discard_op op;
1088
1089 begin_discard(&op, tc, discard_parent);
1090 while (b != end) {
1091 /* find start of unmapped run */
1092 for (; b < end; b++) {
1093 r = dm_pool_block_is_shared(pool->pmd, b, &shared);
1094 if (r)
1095 goto out;
1096
1097 if (!shared)
1098 break;
1099 }
1100
1101 if (b == end)
1102 break;
1103
1104 /* find end of run */
1105 for (e = b + 1; e != end; e++) {
1106 r = dm_pool_block_is_shared(pool->pmd, e, &shared);
1107 if (r)
1108 goto out;
1109
1110 if (shared)
1111 break;
1112 }
1113
1114 r = issue_discard(&op, b, e);
1115 if (r)
1116 goto out;
1117
1118 b = e;
1119 }
1120out:
1121 end_discard(&op, r);
1122}
1123
1124static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1125{
1126 unsigned long flags;
1127 struct pool *pool = m->tc->pool;
1128
1129 spin_lock_irqsave(&pool->lock, flags);
1130 list_add_tail(&m->list, &pool->prepared_discards_pt2);
1131 spin_unlock_irqrestore(&pool->lock, flags);
1132 wake_worker(pool);
1133}
1134
1135static void passdown_endio(struct bio *bio)
1136{
1137 /*
1138 * It doesn't matter if the passdown discard failed, we still want
1139 * to unmap (we ignore err).
1140 */
1141 queue_passdown_pt2(bio->bi_private);
1142 bio_put(bio);
1143}
1144
1145static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1146{
1147 int r;
1148 struct thin_c *tc = m->tc;
1149 struct pool *pool = tc->pool;
1150 struct bio *discard_parent;
1151 dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1152
1153 /*
1154 * Only this thread allocates blocks, so we can be sure that the
1155 * newly unmapped blocks will not be allocated before the end of
1156 * the function.
1157 */
1158 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1159 if (r) {
1160 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1161 bio_io_error(m->bio);
1162 cell_defer_no_holder(tc, m->cell);
1163 mempool_free(m, &pool->mapping_pool);
1164 return;
1165 }
1166
1167 /*
1168 * Increment the unmapped blocks. This prevents a race between the
1169 * passdown io and reallocation of freed blocks.
1170 */
1171 r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1172 if (r) {
1173 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1174 bio_io_error(m->bio);
1175 cell_defer_no_holder(tc, m->cell);
1176 mempool_free(m, &pool->mapping_pool);
1177 return;
1178 }
1179
1180 discard_parent = bio_alloc(NULL, 1, 0, GFP_NOIO);
1181 discard_parent->bi_end_io = passdown_endio;
1182 discard_parent->bi_private = m;
1183 if (m->maybe_shared)
1184 passdown_double_checking_shared_status(m, discard_parent);
1185 else {
1186 struct discard_op op;
1187
1188 begin_discard(&op, tc, discard_parent);
1189 r = issue_discard(&op, m->data_block, data_end);
1190 end_discard(&op, r);
1191 }
1192}
1193
1194static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1195{
1196 int r;
1197 struct thin_c *tc = m->tc;
1198 struct pool *pool = tc->pool;
1199
1200 /*
1201 * The passdown has completed, so now we can decrement all those
1202 * unmapped blocks.
1203 */
1204 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1205 m->data_block + (m->virt_end - m->virt_begin));
1206 if (r) {
1207 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1208 bio_io_error(m->bio);
1209 } else
1210 bio_endio(m->bio);
1211
1212 cell_defer_no_holder(tc, m->cell);
1213 mempool_free(m, &pool->mapping_pool);
1214}
1215
1216static void process_prepared(struct pool *pool, struct list_head *head,
1217 process_mapping_fn *fn)
1218{
1219 struct list_head maps;
1220 struct dm_thin_new_mapping *m, *tmp;
1221
1222 INIT_LIST_HEAD(&maps);
1223 spin_lock_irq(&pool->lock);
1224 list_splice_init(head, &maps);
1225 spin_unlock_irq(&pool->lock);
1226
1227 list_for_each_entry_safe(m, tmp, &maps, list)
1228 (*fn)(m);
1229}
1230
1231/*
1232 * Deferred bio jobs.
1233 */
1234static int io_overlaps_block(struct pool *pool, struct bio *bio)
1235{
1236 return bio->bi_iter.bi_size ==
1237 (pool->sectors_per_block << SECTOR_SHIFT);
1238}
1239
1240static int io_overwrites_block(struct pool *pool, struct bio *bio)
1241{
1242 return (bio_data_dir(bio) == WRITE) &&
1243 io_overlaps_block(pool, bio);
1244}
1245
1246static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1247 bio_end_io_t *fn)
1248{
1249 *save = bio->bi_end_io;
1250 bio->bi_end_io = fn;
1251}
1252
1253static int ensure_next_mapping(struct pool *pool)
1254{
1255 if (pool->next_mapping)
1256 return 0;
1257
1258 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1259
1260 return pool->next_mapping ? 0 : -ENOMEM;
1261}
1262
1263static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1264{
1265 struct dm_thin_new_mapping *m = pool->next_mapping;
1266
1267 BUG_ON(!pool->next_mapping);
1268
1269 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1270 INIT_LIST_HEAD(&m->list);
1271 m->bio = NULL;
1272
1273 pool->next_mapping = NULL;
1274
1275 return m;
1276}
1277
1278static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1279 sector_t begin, sector_t end)
1280{
1281 struct dm_io_region to;
1282
1283 to.bdev = tc->pool_dev->bdev;
1284 to.sector = begin;
1285 to.count = end - begin;
1286
1287 dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1288}
1289
1290static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1291 dm_block_t data_begin,
1292 struct dm_thin_new_mapping *m)
1293{
1294 struct pool *pool = tc->pool;
1295 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1296
1297 h->overwrite_mapping = m;
1298 m->bio = bio;
1299 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1300 inc_all_io_entry(pool, bio);
1301 remap_and_issue(tc, bio, data_begin);
1302}
1303
1304/*
1305 * A partial copy also needs to zero the uncopied region.
1306 */
1307static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1308 struct dm_dev *origin, dm_block_t data_origin,
1309 dm_block_t data_dest,
1310 struct dm_bio_prison_cell *cell, struct bio *bio,
1311 sector_t len)
1312{
1313 struct pool *pool = tc->pool;
1314 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1315
1316 m->tc = tc;
1317 m->virt_begin = virt_block;
1318 m->virt_end = virt_block + 1u;
1319 m->data_block = data_dest;
1320 m->cell = cell;
1321
1322 /*
1323 * quiesce action + copy action + an extra reference held for the
1324 * duration of this function (we may need to inc later for a
1325 * partial zero).
1326 */
1327 atomic_set(&m->prepare_actions, 3);
1328
1329 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1330 complete_mapping_preparation(m); /* already quiesced */
1331
1332 /*
1333 * IO to pool_dev remaps to the pool target's data_dev.
1334 *
1335 * If the whole block of data is being overwritten, we can issue the
1336 * bio immediately. Otherwise we use kcopyd to clone the data first.
1337 */
1338 if (io_overwrites_block(pool, bio))
1339 remap_and_issue_overwrite(tc, bio, data_dest, m);
1340 else {
1341 struct dm_io_region from, to;
1342
1343 from.bdev = origin->bdev;
1344 from.sector = data_origin * pool->sectors_per_block;
1345 from.count = len;
1346
1347 to.bdev = tc->pool_dev->bdev;
1348 to.sector = data_dest * pool->sectors_per_block;
1349 to.count = len;
1350
1351 dm_kcopyd_copy(pool->copier, &from, 1, &to,
1352 0, copy_complete, m);
1353
1354 /*
1355 * Do we need to zero a tail region?
1356 */
1357 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1358 atomic_inc(&m->prepare_actions);
1359 ll_zero(tc, m,
1360 data_dest * pool->sectors_per_block + len,
1361 (data_dest + 1) * pool->sectors_per_block);
1362 }
1363 }
1364
1365 complete_mapping_preparation(m); /* drop our ref */
1366}
1367
1368static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1369 dm_block_t data_origin, dm_block_t data_dest,
1370 struct dm_bio_prison_cell *cell, struct bio *bio)
1371{
1372 schedule_copy(tc, virt_block, tc->pool_dev,
1373 data_origin, data_dest, cell, bio,
1374 tc->pool->sectors_per_block);
1375}
1376
1377static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1378 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1379 struct bio *bio)
1380{
1381 struct pool *pool = tc->pool;
1382 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1383
1384 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1385 m->tc = tc;
1386 m->virt_begin = virt_block;
1387 m->virt_end = virt_block + 1u;
1388 m->data_block = data_block;
1389 m->cell = cell;
1390
1391 /*
1392 * If the whole block of data is being overwritten or we are not
1393 * zeroing pre-existing data, we can issue the bio immediately.
1394 * Otherwise we use kcopyd to zero the data first.
1395 */
1396 if (pool->pf.zero_new_blocks) {
1397 if (io_overwrites_block(pool, bio))
1398 remap_and_issue_overwrite(tc, bio, data_block, m);
1399 else
1400 ll_zero(tc, m, data_block * pool->sectors_per_block,
1401 (data_block + 1) * pool->sectors_per_block);
1402 } else
1403 process_prepared_mapping(m);
1404}
1405
1406static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1407 dm_block_t data_dest,
1408 struct dm_bio_prison_cell *cell, struct bio *bio)
1409{
1410 struct pool *pool = tc->pool;
1411 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1412 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1413
1414 if (virt_block_end <= tc->origin_size)
1415 schedule_copy(tc, virt_block, tc->origin_dev,
1416 virt_block, data_dest, cell, bio,
1417 pool->sectors_per_block);
1418
1419 else if (virt_block_begin < tc->origin_size)
1420 schedule_copy(tc, virt_block, tc->origin_dev,
1421 virt_block, data_dest, cell, bio,
1422 tc->origin_size - virt_block_begin);
1423
1424 else
1425 schedule_zero(tc, virt_block, data_dest, cell, bio);
1426}
1427
1428static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1429
1430static void requeue_bios(struct pool *pool);
1431
1432static bool is_read_only_pool_mode(enum pool_mode mode)
1433{
1434 return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1435}
1436
1437static bool is_read_only(struct pool *pool)
1438{
1439 return is_read_only_pool_mode(get_pool_mode(pool));
1440}
1441
1442static void check_for_metadata_space(struct pool *pool)
1443{
1444 int r;
1445 const char *ooms_reason = NULL;
1446 dm_block_t nr_free;
1447
1448 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1449 if (r)
1450 ooms_reason = "Could not get free metadata blocks";
1451 else if (!nr_free)
1452 ooms_reason = "No free metadata blocks";
1453
1454 if (ooms_reason && !is_read_only(pool)) {
1455 DMERR("%s", ooms_reason);
1456 set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1457 }
1458}
1459
1460static void check_for_data_space(struct pool *pool)
1461{
1462 int r;
1463 dm_block_t nr_free;
1464
1465 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1466 return;
1467
1468 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1469 if (r)
1470 return;
1471
1472 if (nr_free) {
1473 set_pool_mode(pool, PM_WRITE);
1474 requeue_bios(pool);
1475 }
1476}
1477
1478/*
1479 * A non-zero return indicates read_only or fail_io mode.
1480 * Many callers don't care about the return value.
1481 */
1482static int commit(struct pool *pool)
1483{
1484 int r;
1485
1486 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1487 return -EINVAL;
1488
1489 r = dm_pool_commit_metadata(pool->pmd);
1490 if (r)
1491 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1492 else {
1493 check_for_metadata_space(pool);
1494 check_for_data_space(pool);
1495 }
1496
1497 return r;
1498}
1499
1500static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1501{
1502 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1503 DMWARN("%s: reached low water mark for data device: sending event.",
1504 dm_device_name(pool->pool_md));
1505 spin_lock_irq(&pool->lock);
1506 pool->low_water_triggered = true;
1507 spin_unlock_irq(&pool->lock);
1508 dm_table_event(pool->ti->table);
1509 }
1510}
1511
1512static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1513{
1514 int r;
1515 dm_block_t free_blocks;
1516 struct pool *pool = tc->pool;
1517
1518 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1519 return -EINVAL;
1520
1521 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1522 if (r) {
1523 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1524 return r;
1525 }
1526
1527 check_low_water_mark(pool, free_blocks);
1528
1529 if (!free_blocks) {
1530 /*
1531 * Try to commit to see if that will free up some
1532 * more space.
1533 */
1534 r = commit(pool);
1535 if (r)
1536 return r;
1537
1538 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1539 if (r) {
1540 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1541 return r;
1542 }
1543
1544 if (!free_blocks) {
1545 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1546 return -ENOSPC;
1547 }
1548 }
1549
1550 r = dm_pool_alloc_data_block(pool->pmd, result);
1551 if (r) {
1552 if (r == -ENOSPC)
1553 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1554 else
1555 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1556 return r;
1557 }
1558
1559 r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1560 if (r) {
1561 metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1562 return r;
1563 }
1564
1565 if (!free_blocks) {
1566 /* Let's commit before we use up the metadata reserve. */
1567 r = commit(pool);
1568 if (r)
1569 return r;
1570 }
1571
1572 return 0;
1573}
1574
1575/*
1576 * If we have run out of space, queue bios until the device is
1577 * resumed, presumably after having been reloaded with more space.
1578 */
1579static void retry_on_resume(struct bio *bio)
1580{
1581 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1582 struct thin_c *tc = h->tc;
1583
1584 spin_lock_irq(&tc->lock);
1585 bio_list_add(&tc->retry_on_resume_list, bio);
1586 spin_unlock_irq(&tc->lock);
1587}
1588
1589static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1590{
1591 enum pool_mode m = get_pool_mode(pool);
1592
1593 switch (m) {
1594 case PM_WRITE:
1595 /* Shouldn't get here */
1596 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1597 return BLK_STS_IOERR;
1598
1599 case PM_OUT_OF_DATA_SPACE:
1600 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1601
1602 case PM_OUT_OF_METADATA_SPACE:
1603 case PM_READ_ONLY:
1604 case PM_FAIL:
1605 return BLK_STS_IOERR;
1606 default:
1607 /* Shouldn't get here */
1608 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1609 return BLK_STS_IOERR;
1610 }
1611}
1612
1613static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1614{
1615 blk_status_t error = should_error_unserviceable_bio(pool);
1616
1617 if (error) {
1618 bio->bi_status = error;
1619 bio_endio(bio);
1620 } else
1621 retry_on_resume(bio);
1622}
1623
1624static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1625{
1626 struct bio *bio;
1627 struct bio_list bios;
1628 blk_status_t error;
1629
1630 error = should_error_unserviceable_bio(pool);
1631 if (error) {
1632 cell_error_with_code(pool, cell, error);
1633 return;
1634 }
1635
1636 bio_list_init(&bios);
1637 cell_release(pool, cell, &bios);
1638
1639 while ((bio = bio_list_pop(&bios)))
1640 retry_on_resume(bio);
1641}
1642
1643static void process_discard_cell_no_passdown(struct thin_c *tc,
1644 struct dm_bio_prison_cell *virt_cell)
1645{
1646 struct pool *pool = tc->pool;
1647 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1648
1649 /*
1650 * We don't need to lock the data blocks, since there's no
1651 * passdown. We only lock data blocks for allocation and breaking sharing.
1652 */
1653 m->tc = tc;
1654 m->virt_begin = virt_cell->key.block_begin;
1655 m->virt_end = virt_cell->key.block_end;
1656 m->cell = virt_cell;
1657 m->bio = virt_cell->holder;
1658
1659 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1660 pool->process_prepared_discard(m);
1661}
1662
1663static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1664 struct bio *bio)
1665{
1666 struct pool *pool = tc->pool;
1667
1668 int r;
1669 bool maybe_shared;
1670 struct dm_cell_key data_key;
1671 struct dm_bio_prison_cell *data_cell;
1672 struct dm_thin_new_mapping *m;
1673 dm_block_t virt_begin, virt_end, data_begin;
1674
1675 while (begin != end) {
1676 r = ensure_next_mapping(pool);
1677 if (r)
1678 /* we did our best */
1679 return;
1680
1681 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1682 &data_begin, &maybe_shared);
1683 if (r)
1684 /*
1685 * Silently fail, letting any mappings we've
1686 * created complete.
1687 */
1688 break;
1689
1690 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1691 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1692 /* contention, we'll give up with this range */
1693 begin = virt_end;
1694 continue;
1695 }
1696
1697 /*
1698 * IO may still be going to the destination block. We must
1699 * quiesce before we can do the removal.
1700 */
1701 m = get_next_mapping(pool);
1702 m->tc = tc;
1703 m->maybe_shared = maybe_shared;
1704 m->virt_begin = virt_begin;
1705 m->virt_end = virt_end;
1706 m->data_block = data_begin;
1707 m->cell = data_cell;
1708 m->bio = bio;
1709
1710 /*
1711 * The parent bio must not complete before sub discard bios are
1712 * chained to it (see end_discard's bio_chain)!
1713 *
1714 * This per-mapping bi_remaining increment is paired with
1715 * the implicit decrement that occurs via bio_endio() in
1716 * end_discard().
1717 */
1718 bio_inc_remaining(bio);
1719 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1720 pool->process_prepared_discard(m);
1721
1722 begin = virt_end;
1723 }
1724}
1725
1726static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1727{
1728 struct bio *bio = virt_cell->holder;
1729 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1730
1731 /*
1732 * The virt_cell will only get freed once the origin bio completes.
1733 * This means it will remain locked while all the individual
1734 * passdown bios are in flight.
1735 */
1736 h->cell = virt_cell;
1737 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1738
1739 /*
1740 * We complete the bio now, knowing that the bi_remaining field
1741 * will prevent completion until the sub range discards have
1742 * completed.
1743 */
1744 bio_endio(bio);
1745}
1746
1747static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1748{
1749 dm_block_t begin, end;
1750 struct dm_cell_key virt_key;
1751 struct dm_bio_prison_cell *virt_cell;
1752
1753 get_bio_block_range(tc, bio, &begin, &end);
1754 if (begin == end) {
1755 /*
1756 * The discard covers less than a block.
1757 */
1758 bio_endio(bio);
1759 return;
1760 }
1761
1762 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1763 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1764 /*
1765 * Potential starvation issue: We're relying on the
1766 * fs/application being well behaved, and not trying to
1767 * send IO to a region at the same time as discarding it.
1768 * If they do this persistently then it's possible this
1769 * cell will never be granted.
1770 */
1771 return;
1772
1773 tc->pool->process_discard_cell(tc, virt_cell);
1774}
1775
1776static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1777 struct dm_cell_key *key,
1778 struct dm_thin_lookup_result *lookup_result,
1779 struct dm_bio_prison_cell *cell)
1780{
1781 int r;
1782 dm_block_t data_block;
1783 struct pool *pool = tc->pool;
1784
1785 r = alloc_data_block(tc, &data_block);
1786 switch (r) {
1787 case 0:
1788 schedule_internal_copy(tc, block, lookup_result->block,
1789 data_block, cell, bio);
1790 break;
1791
1792 case -ENOSPC:
1793 retry_bios_on_resume(pool, cell);
1794 break;
1795
1796 default:
1797 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1798 __func__, r);
1799 cell_error(pool, cell);
1800 break;
1801 }
1802}
1803
1804static void __remap_and_issue_shared_cell(void *context,
1805 struct dm_bio_prison_cell *cell)
1806{
1807 struct remap_info *info = context;
1808 struct bio *bio;
1809
1810 while ((bio = bio_list_pop(&cell->bios))) {
1811 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1812 bio_op(bio) == REQ_OP_DISCARD)
1813 bio_list_add(&info->defer_bios, bio);
1814 else {
1815 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1816
1817 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1818 inc_all_io_entry(info->tc->pool, bio);
1819 bio_list_add(&info->issue_bios, bio);
1820 }
1821 }
1822}
1823
1824static void remap_and_issue_shared_cell(struct thin_c *tc,
1825 struct dm_bio_prison_cell *cell,
1826 dm_block_t block)
1827{
1828 struct bio *bio;
1829 struct remap_info info;
1830
1831 info.tc = tc;
1832 bio_list_init(&info.defer_bios);
1833 bio_list_init(&info.issue_bios);
1834
1835 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1836 &info, cell);
1837
1838 while ((bio = bio_list_pop(&info.defer_bios)))
1839 thin_defer_bio(tc, bio);
1840
1841 while ((bio = bio_list_pop(&info.issue_bios)))
1842 remap_and_issue(tc, bio, block);
1843}
1844
1845static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1846 dm_block_t block,
1847 struct dm_thin_lookup_result *lookup_result,
1848 struct dm_bio_prison_cell *virt_cell)
1849{
1850 struct dm_bio_prison_cell *data_cell;
1851 struct pool *pool = tc->pool;
1852 struct dm_cell_key key;
1853
1854 /*
1855 * If cell is already occupied, then sharing is already in the process
1856 * of being broken so we have nothing further to do here.
1857 */
1858 build_data_key(tc->td, lookup_result->block, &key);
1859 if (bio_detain(pool, &key, bio, &data_cell)) {
1860 cell_defer_no_holder(tc, virt_cell);
1861 return;
1862 }
1863
1864 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1865 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1866 cell_defer_no_holder(tc, virt_cell);
1867 } else {
1868 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1869
1870 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1871 inc_all_io_entry(pool, bio);
1872 remap_and_issue(tc, bio, lookup_result->block);
1873
1874 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1875 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1876 }
1877}
1878
1879static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1880 struct dm_bio_prison_cell *cell)
1881{
1882 int r;
1883 dm_block_t data_block;
1884 struct pool *pool = tc->pool;
1885
1886 /*
1887 * Remap empty bios (flushes) immediately, without provisioning.
1888 */
1889 if (!bio->bi_iter.bi_size) {
1890 inc_all_io_entry(pool, bio);
1891 cell_defer_no_holder(tc, cell);
1892
1893 remap_and_issue(tc, bio, 0);
1894 return;
1895 }
1896
1897 /*
1898 * Fill read bios with zeroes and complete them immediately.
1899 */
1900 if (bio_data_dir(bio) == READ) {
1901 zero_fill_bio(bio);
1902 cell_defer_no_holder(tc, cell);
1903 bio_endio(bio);
1904 return;
1905 }
1906
1907 r = alloc_data_block(tc, &data_block);
1908 switch (r) {
1909 case 0:
1910 if (tc->origin_dev)
1911 schedule_external_copy(tc, block, data_block, cell, bio);
1912 else
1913 schedule_zero(tc, block, data_block, cell, bio);
1914 break;
1915
1916 case -ENOSPC:
1917 retry_bios_on_resume(pool, cell);
1918 break;
1919
1920 default:
1921 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1922 __func__, r);
1923 cell_error(pool, cell);
1924 break;
1925 }
1926}
1927
1928static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1929{
1930 int r;
1931 struct pool *pool = tc->pool;
1932 struct bio *bio = cell->holder;
1933 dm_block_t block = get_bio_block(tc, bio);
1934 struct dm_thin_lookup_result lookup_result;
1935
1936 if (tc->requeue_mode) {
1937 cell_requeue(pool, cell);
1938 return;
1939 }
1940
1941 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1942 switch (r) {
1943 case 0:
1944 if (lookup_result.shared)
1945 process_shared_bio(tc, bio, block, &lookup_result, cell);
1946 else {
1947 inc_all_io_entry(pool, bio);
1948 remap_and_issue(tc, bio, lookup_result.block);
1949 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1950 }
1951 break;
1952
1953 case -ENODATA:
1954 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1955 inc_all_io_entry(pool, bio);
1956 cell_defer_no_holder(tc, cell);
1957
1958 if (bio_end_sector(bio) <= tc->origin_size)
1959 remap_to_origin_and_issue(tc, bio);
1960
1961 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1962 zero_fill_bio(bio);
1963 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1964 remap_to_origin_and_issue(tc, bio);
1965
1966 } else {
1967 zero_fill_bio(bio);
1968 bio_endio(bio);
1969 }
1970 } else
1971 provision_block(tc, bio, block, cell);
1972 break;
1973
1974 default:
1975 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1976 __func__, r);
1977 cell_defer_no_holder(tc, cell);
1978 bio_io_error(bio);
1979 break;
1980 }
1981}
1982
1983static void process_bio(struct thin_c *tc, struct bio *bio)
1984{
1985 struct pool *pool = tc->pool;
1986 dm_block_t block = get_bio_block(tc, bio);
1987 struct dm_bio_prison_cell *cell;
1988 struct dm_cell_key key;
1989
1990 /*
1991 * If cell is already occupied, then the block is already
1992 * being provisioned so we have nothing further to do here.
1993 */
1994 build_virtual_key(tc->td, block, &key);
1995 if (bio_detain(pool, &key, bio, &cell))
1996 return;
1997
1998 process_cell(tc, cell);
1999}
2000
2001static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2002 struct dm_bio_prison_cell *cell)
2003{
2004 int r;
2005 int rw = bio_data_dir(bio);
2006 dm_block_t block = get_bio_block(tc, bio);
2007 struct dm_thin_lookup_result lookup_result;
2008
2009 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2010 switch (r) {
2011 case 0:
2012 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2013 handle_unserviceable_bio(tc->pool, bio);
2014 if (cell)
2015 cell_defer_no_holder(tc, cell);
2016 } else {
2017 inc_all_io_entry(tc->pool, bio);
2018 remap_and_issue(tc, bio, lookup_result.block);
2019 if (cell)
2020 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2021 }
2022 break;
2023
2024 case -ENODATA:
2025 if (cell)
2026 cell_defer_no_holder(tc, cell);
2027 if (rw != READ) {
2028 handle_unserviceable_bio(tc->pool, bio);
2029 break;
2030 }
2031
2032 if (tc->origin_dev) {
2033 inc_all_io_entry(tc->pool, bio);
2034 remap_to_origin_and_issue(tc, bio);
2035 break;
2036 }
2037
2038 zero_fill_bio(bio);
2039 bio_endio(bio);
2040 break;
2041
2042 default:
2043 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2044 __func__, r);
2045 if (cell)
2046 cell_defer_no_holder(tc, cell);
2047 bio_io_error(bio);
2048 break;
2049 }
2050}
2051
2052static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2053{
2054 __process_bio_read_only(tc, bio, NULL);
2055}
2056
2057static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2058{
2059 __process_bio_read_only(tc, cell->holder, cell);
2060}
2061
2062static void process_bio_success(struct thin_c *tc, struct bio *bio)
2063{
2064 bio_endio(bio);
2065}
2066
2067static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2068{
2069 bio_io_error(bio);
2070}
2071
2072static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2073{
2074 cell_success(tc->pool, cell);
2075}
2076
2077static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2078{
2079 cell_error(tc->pool, cell);
2080}
2081
2082/*
2083 * FIXME: should we also commit due to size of transaction, measured in
2084 * metadata blocks?
2085 */
2086static int need_commit_due_to_time(struct pool *pool)
2087{
2088 return !time_in_range(jiffies, pool->last_commit_jiffies,
2089 pool->last_commit_jiffies + COMMIT_PERIOD);
2090}
2091
2092#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2093#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2094
2095static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2096{
2097 struct rb_node **rbp, *parent;
2098 struct dm_thin_endio_hook *pbd;
2099 sector_t bi_sector = bio->bi_iter.bi_sector;
2100
2101 rbp = &tc->sort_bio_list.rb_node;
2102 parent = NULL;
2103 while (*rbp) {
2104 parent = *rbp;
2105 pbd = thin_pbd(parent);
2106
2107 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2108 rbp = &(*rbp)->rb_left;
2109 else
2110 rbp = &(*rbp)->rb_right;
2111 }
2112
2113 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2114 rb_link_node(&pbd->rb_node, parent, rbp);
2115 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2116}
2117
2118static void __extract_sorted_bios(struct thin_c *tc)
2119{
2120 struct rb_node *node;
2121 struct dm_thin_endio_hook *pbd;
2122 struct bio *bio;
2123
2124 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2125 pbd = thin_pbd(node);
2126 bio = thin_bio(pbd);
2127
2128 bio_list_add(&tc->deferred_bio_list, bio);
2129 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2130 }
2131
2132 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2133}
2134
2135static void __sort_thin_deferred_bios(struct thin_c *tc)
2136{
2137 struct bio *bio;
2138 struct bio_list bios;
2139
2140 bio_list_init(&bios);
2141 bio_list_merge(&bios, &tc->deferred_bio_list);
2142 bio_list_init(&tc->deferred_bio_list);
2143
2144 /* Sort deferred_bio_list using rb-tree */
2145 while ((bio = bio_list_pop(&bios)))
2146 __thin_bio_rb_add(tc, bio);
2147
2148 /*
2149 * Transfer the sorted bios in sort_bio_list back to
2150 * deferred_bio_list to allow lockless submission of
2151 * all bios.
2152 */
2153 __extract_sorted_bios(tc);
2154}
2155
2156static void process_thin_deferred_bios(struct thin_c *tc)
2157{
2158 struct pool *pool = tc->pool;
2159 struct bio *bio;
2160 struct bio_list bios;
2161 struct blk_plug plug;
2162 unsigned count = 0;
2163
2164 if (tc->requeue_mode) {
2165 error_thin_bio_list(tc, &tc->deferred_bio_list,
2166 BLK_STS_DM_REQUEUE);
2167 return;
2168 }
2169
2170 bio_list_init(&bios);
2171
2172 spin_lock_irq(&tc->lock);
2173
2174 if (bio_list_empty(&tc->deferred_bio_list)) {
2175 spin_unlock_irq(&tc->lock);
2176 return;
2177 }
2178
2179 __sort_thin_deferred_bios(tc);
2180
2181 bio_list_merge(&bios, &tc->deferred_bio_list);
2182 bio_list_init(&tc->deferred_bio_list);
2183
2184 spin_unlock_irq(&tc->lock);
2185
2186 blk_start_plug(&plug);
2187 while ((bio = bio_list_pop(&bios))) {
2188 /*
2189 * If we've got no free new_mapping structs, and processing
2190 * this bio might require one, we pause until there are some
2191 * prepared mappings to process.
2192 */
2193 if (ensure_next_mapping(pool)) {
2194 spin_lock_irq(&tc->lock);
2195 bio_list_add(&tc->deferred_bio_list, bio);
2196 bio_list_merge(&tc->deferred_bio_list, &bios);
2197 spin_unlock_irq(&tc->lock);
2198 break;
2199 }
2200
2201 if (bio_op(bio) == REQ_OP_DISCARD)
2202 pool->process_discard(tc, bio);
2203 else
2204 pool->process_bio(tc, bio);
2205
2206 if ((count++ & 127) == 0) {
2207 throttle_work_update(&pool->throttle);
2208 dm_pool_issue_prefetches(pool->pmd);
2209 }
2210 }
2211 blk_finish_plug(&plug);
2212}
2213
2214static int cmp_cells(const void *lhs, const void *rhs)
2215{
2216 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2217 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2218
2219 BUG_ON(!lhs_cell->holder);
2220 BUG_ON(!rhs_cell->holder);
2221
2222 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2223 return -1;
2224
2225 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2226 return 1;
2227
2228 return 0;
2229}
2230
2231static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2232{
2233 unsigned count = 0;
2234 struct dm_bio_prison_cell *cell, *tmp;
2235
2236 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2237 if (count >= CELL_SORT_ARRAY_SIZE)
2238 break;
2239
2240 pool->cell_sort_array[count++] = cell;
2241 list_del(&cell->user_list);
2242 }
2243
2244 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2245
2246 return count;
2247}
2248
2249static void process_thin_deferred_cells(struct thin_c *tc)
2250{
2251 struct pool *pool = tc->pool;
2252 struct list_head cells;
2253 struct dm_bio_prison_cell *cell;
2254 unsigned i, j, count;
2255
2256 INIT_LIST_HEAD(&cells);
2257
2258 spin_lock_irq(&tc->lock);
2259 list_splice_init(&tc->deferred_cells, &cells);
2260 spin_unlock_irq(&tc->lock);
2261
2262 if (list_empty(&cells))
2263 return;
2264
2265 do {
2266 count = sort_cells(tc->pool, &cells);
2267
2268 for (i = 0; i < count; i++) {
2269 cell = pool->cell_sort_array[i];
2270 BUG_ON(!cell->holder);
2271
2272 /*
2273 * If we've got no free new_mapping structs, and processing
2274 * this bio might require one, we pause until there are some
2275 * prepared mappings to process.
2276 */
2277 if (ensure_next_mapping(pool)) {
2278 for (j = i; j < count; j++)
2279 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2280
2281 spin_lock_irq(&tc->lock);
2282 list_splice(&cells, &tc->deferred_cells);
2283 spin_unlock_irq(&tc->lock);
2284 return;
2285 }
2286
2287 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2288 pool->process_discard_cell(tc, cell);
2289 else
2290 pool->process_cell(tc, cell);
2291 }
2292 } while (!list_empty(&cells));
2293}
2294
2295static void thin_get(struct thin_c *tc);
2296static void thin_put(struct thin_c *tc);
2297
2298/*
2299 * We can't hold rcu_read_lock() around code that can block. So we
2300 * find a thin with the rcu lock held; bump a refcount; then drop
2301 * the lock.
2302 */
2303static struct thin_c *get_first_thin(struct pool *pool)
2304{
2305 struct thin_c *tc = NULL;
2306
2307 rcu_read_lock();
2308 if (!list_empty(&pool->active_thins)) {
2309 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2310 thin_get(tc);
2311 }
2312 rcu_read_unlock();
2313
2314 return tc;
2315}
2316
2317static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2318{
2319 struct thin_c *old_tc = tc;
2320
2321 rcu_read_lock();
2322 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2323 thin_get(tc);
2324 thin_put(old_tc);
2325 rcu_read_unlock();
2326 return tc;
2327 }
2328 thin_put(old_tc);
2329 rcu_read_unlock();
2330
2331 return NULL;
2332}
2333
2334static void process_deferred_bios(struct pool *pool)
2335{
2336 struct bio *bio;
2337 struct bio_list bios, bio_completions;
2338 struct thin_c *tc;
2339
2340 tc = get_first_thin(pool);
2341 while (tc) {
2342 process_thin_deferred_cells(tc);
2343 process_thin_deferred_bios(tc);
2344 tc = get_next_thin(pool, tc);
2345 }
2346
2347 /*
2348 * If there are any deferred flush bios, we must commit the metadata
2349 * before issuing them or signaling their completion.
2350 */
2351 bio_list_init(&bios);
2352 bio_list_init(&bio_completions);
2353
2354 spin_lock_irq(&pool->lock);
2355 bio_list_merge(&bios, &pool->deferred_flush_bios);
2356 bio_list_init(&pool->deferred_flush_bios);
2357
2358 bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2359 bio_list_init(&pool->deferred_flush_completions);
2360 spin_unlock_irq(&pool->lock);
2361
2362 if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2363 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2364 return;
2365
2366 if (commit(pool)) {
2367 bio_list_merge(&bios, &bio_completions);
2368
2369 while ((bio = bio_list_pop(&bios)))
2370 bio_io_error(bio);
2371 return;
2372 }
2373 pool->last_commit_jiffies = jiffies;
2374
2375 while ((bio = bio_list_pop(&bio_completions)))
2376 bio_endio(bio);
2377
2378 while ((bio = bio_list_pop(&bios))) {
2379 /*
2380 * The data device was flushed as part of metadata commit,
2381 * so complete redundant flushes immediately.
2382 */
2383 if (bio->bi_opf & REQ_PREFLUSH)
2384 bio_endio(bio);
2385 else
2386 dm_submit_bio_remap(bio, NULL);
2387 }
2388}
2389
2390static void do_worker(struct work_struct *ws)
2391{
2392 struct pool *pool = container_of(ws, struct pool, worker);
2393
2394 throttle_work_start(&pool->throttle);
2395 dm_pool_issue_prefetches(pool->pmd);
2396 throttle_work_update(&pool->throttle);
2397 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2398 throttle_work_update(&pool->throttle);
2399 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2400 throttle_work_update(&pool->throttle);
2401 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2402 throttle_work_update(&pool->throttle);
2403 process_deferred_bios(pool);
2404 throttle_work_complete(&pool->throttle);
2405}
2406
2407/*
2408 * We want to commit periodically so that not too much
2409 * unwritten data builds up.
2410 */
2411static void do_waker(struct work_struct *ws)
2412{
2413 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2414 wake_worker(pool);
2415 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2416}
2417
2418/*
2419 * We're holding onto IO to allow userland time to react. After the
2420 * timeout either the pool will have been resized (and thus back in
2421 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2422 */
2423static void do_no_space_timeout(struct work_struct *ws)
2424{
2425 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2426 no_space_timeout);
2427
2428 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2429 pool->pf.error_if_no_space = true;
2430 notify_of_pool_mode_change(pool);
2431 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2432 }
2433}
2434
2435/*----------------------------------------------------------------*/
2436
2437struct pool_work {
2438 struct work_struct worker;
2439 struct completion complete;
2440};
2441
2442static struct pool_work *to_pool_work(struct work_struct *ws)
2443{
2444 return container_of(ws, struct pool_work, worker);
2445}
2446
2447static void pool_work_complete(struct pool_work *pw)
2448{
2449 complete(&pw->complete);
2450}
2451
2452static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2453 void (*fn)(struct work_struct *))
2454{
2455 INIT_WORK_ONSTACK(&pw->worker, fn);
2456 init_completion(&pw->complete);
2457 queue_work(pool->wq, &pw->worker);
2458 wait_for_completion(&pw->complete);
2459}
2460
2461/*----------------------------------------------------------------*/
2462
2463struct noflush_work {
2464 struct pool_work pw;
2465 struct thin_c *tc;
2466};
2467
2468static struct noflush_work *to_noflush(struct work_struct *ws)
2469{
2470 return container_of(to_pool_work(ws), struct noflush_work, pw);
2471}
2472
2473static void do_noflush_start(struct work_struct *ws)
2474{
2475 struct noflush_work *w = to_noflush(ws);
2476 w->tc->requeue_mode = true;
2477 requeue_io(w->tc);
2478 pool_work_complete(&w->pw);
2479}
2480
2481static void do_noflush_stop(struct work_struct *ws)
2482{
2483 struct noflush_work *w = to_noflush(ws);
2484 w->tc->requeue_mode = false;
2485 pool_work_complete(&w->pw);
2486}
2487
2488static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2489{
2490 struct noflush_work w;
2491
2492 w.tc = tc;
2493 pool_work_wait(&w.pw, tc->pool, fn);
2494}
2495
2496/*----------------------------------------------------------------*/
2497
2498static bool passdown_enabled(struct pool_c *pt)
2499{
2500 return pt->adjusted_pf.discard_passdown;
2501}
2502
2503static void set_discard_callbacks(struct pool *pool)
2504{
2505 struct pool_c *pt = pool->ti->private;
2506
2507 if (passdown_enabled(pt)) {
2508 pool->process_discard_cell = process_discard_cell_passdown;
2509 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2510 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2511 } else {
2512 pool->process_discard_cell = process_discard_cell_no_passdown;
2513 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2514 }
2515}
2516
2517static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2518{
2519 struct pool_c *pt = pool->ti->private;
2520 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2521 enum pool_mode old_mode = get_pool_mode(pool);
2522 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2523
2524 /*
2525 * Never allow the pool to transition to PM_WRITE mode if user
2526 * intervention is required to verify metadata and data consistency.
2527 */
2528 if (new_mode == PM_WRITE && needs_check) {
2529 DMERR("%s: unable to switch pool to write mode until repaired.",
2530 dm_device_name(pool->pool_md));
2531 if (old_mode != new_mode)
2532 new_mode = old_mode;
2533 else
2534 new_mode = PM_READ_ONLY;
2535 }
2536 /*
2537 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2538 * not going to recover without a thin_repair. So we never let the
2539 * pool move out of the old mode.
2540 */
2541 if (old_mode == PM_FAIL)
2542 new_mode = old_mode;
2543
2544 switch (new_mode) {
2545 case PM_FAIL:
2546 dm_pool_metadata_read_only(pool->pmd);
2547 pool->process_bio = process_bio_fail;
2548 pool->process_discard = process_bio_fail;
2549 pool->process_cell = process_cell_fail;
2550 pool->process_discard_cell = process_cell_fail;
2551 pool->process_prepared_mapping = process_prepared_mapping_fail;
2552 pool->process_prepared_discard = process_prepared_discard_fail;
2553
2554 error_retry_list(pool);
2555 break;
2556
2557 case PM_OUT_OF_METADATA_SPACE:
2558 case PM_READ_ONLY:
2559 dm_pool_metadata_read_only(pool->pmd);
2560 pool->process_bio = process_bio_read_only;
2561 pool->process_discard = process_bio_success;
2562 pool->process_cell = process_cell_read_only;
2563 pool->process_discard_cell = process_cell_success;
2564 pool->process_prepared_mapping = process_prepared_mapping_fail;
2565 pool->process_prepared_discard = process_prepared_discard_success;
2566
2567 error_retry_list(pool);
2568 break;
2569
2570 case PM_OUT_OF_DATA_SPACE:
2571 /*
2572 * Ideally we'd never hit this state; the low water mark
2573 * would trigger userland to extend the pool before we
2574 * completely run out of data space. However, many small
2575 * IOs to unprovisioned space can consume data space at an
2576 * alarming rate. Adjust your low water mark if you're
2577 * frequently seeing this mode.
2578 */
2579 pool->out_of_data_space = true;
2580 pool->process_bio = process_bio_read_only;
2581 pool->process_discard = process_discard_bio;
2582 pool->process_cell = process_cell_read_only;
2583 pool->process_prepared_mapping = process_prepared_mapping;
2584 set_discard_callbacks(pool);
2585
2586 if (!pool->pf.error_if_no_space && no_space_timeout)
2587 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2588 break;
2589
2590 case PM_WRITE:
2591 if (old_mode == PM_OUT_OF_DATA_SPACE)
2592 cancel_delayed_work_sync(&pool->no_space_timeout);
2593 pool->out_of_data_space = false;
2594 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2595 dm_pool_metadata_read_write(pool->pmd);
2596 pool->process_bio = process_bio;
2597 pool->process_discard = process_discard_bio;
2598 pool->process_cell = process_cell;
2599 pool->process_prepared_mapping = process_prepared_mapping;
2600 set_discard_callbacks(pool);
2601 break;
2602 }
2603
2604 pool->pf.mode = new_mode;
2605 /*
2606 * The pool mode may have changed, sync it so bind_control_target()
2607 * doesn't cause an unexpected mode transition on resume.
2608 */
2609 pt->adjusted_pf.mode = new_mode;
2610
2611 if (old_mode != new_mode)
2612 notify_of_pool_mode_change(pool);
2613}
2614
2615static void abort_transaction(struct pool *pool)
2616{
2617 const char *dev_name = dm_device_name(pool->pool_md);
2618
2619 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2620 if (dm_pool_abort_metadata(pool->pmd)) {
2621 DMERR("%s: failed to abort metadata transaction", dev_name);
2622 set_pool_mode(pool, PM_FAIL);
2623 }
2624
2625 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2626 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2627 set_pool_mode(pool, PM_FAIL);
2628 }
2629}
2630
2631static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2632{
2633 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2634 dm_device_name(pool->pool_md), op, r);
2635
2636 abort_transaction(pool);
2637 set_pool_mode(pool, PM_READ_ONLY);
2638}
2639
2640/*----------------------------------------------------------------*/
2641
2642/*
2643 * Mapping functions.
2644 */
2645
2646/*
2647 * Called only while mapping a thin bio to hand it over to the workqueue.
2648 */
2649static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2650{
2651 struct pool *pool = tc->pool;
2652
2653 spin_lock_irq(&tc->lock);
2654 bio_list_add(&tc->deferred_bio_list, bio);
2655 spin_unlock_irq(&tc->lock);
2656
2657 wake_worker(pool);
2658}
2659
2660static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2661{
2662 struct pool *pool = tc->pool;
2663
2664 throttle_lock(&pool->throttle);
2665 thin_defer_bio(tc, bio);
2666 throttle_unlock(&pool->throttle);
2667}
2668
2669static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2670{
2671 struct pool *pool = tc->pool;
2672
2673 throttle_lock(&pool->throttle);
2674 spin_lock_irq(&tc->lock);
2675 list_add_tail(&cell->user_list, &tc->deferred_cells);
2676 spin_unlock_irq(&tc->lock);
2677 throttle_unlock(&pool->throttle);
2678
2679 wake_worker(pool);
2680}
2681
2682static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2683{
2684 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2685
2686 h->tc = tc;
2687 h->shared_read_entry = NULL;
2688 h->all_io_entry = NULL;
2689 h->overwrite_mapping = NULL;
2690 h->cell = NULL;
2691}
2692
2693/*
2694 * Non-blocking function called from the thin target's map function.
2695 */
2696static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2697{
2698 int r;
2699 struct thin_c *tc = ti->private;
2700 dm_block_t block = get_bio_block(tc, bio);
2701 struct dm_thin_device *td = tc->td;
2702 struct dm_thin_lookup_result result;
2703 struct dm_bio_prison_cell *virt_cell, *data_cell;
2704 struct dm_cell_key key;
2705
2706 thin_hook_bio(tc, bio);
2707
2708 if (tc->requeue_mode) {
2709 bio->bi_status = BLK_STS_DM_REQUEUE;
2710 bio_endio(bio);
2711 return DM_MAPIO_SUBMITTED;
2712 }
2713
2714 if (get_pool_mode(tc->pool) == PM_FAIL) {
2715 bio_io_error(bio);
2716 return DM_MAPIO_SUBMITTED;
2717 }
2718
2719 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2720 thin_defer_bio_with_throttle(tc, bio);
2721 return DM_MAPIO_SUBMITTED;
2722 }
2723
2724 /*
2725 * We must hold the virtual cell before doing the lookup, otherwise
2726 * there's a race with discard.
2727 */
2728 build_virtual_key(tc->td, block, &key);
2729 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2730 return DM_MAPIO_SUBMITTED;
2731
2732 r = dm_thin_find_block(td, block, 0, &result);
2733
2734 /*
2735 * Note that we defer readahead too.
2736 */
2737 switch (r) {
2738 case 0:
2739 if (unlikely(result.shared)) {
2740 /*
2741 * We have a race condition here between the
2742 * result.shared value returned by the lookup and
2743 * snapshot creation, which may cause new
2744 * sharing.
2745 *
2746 * To avoid this always quiesce the origin before
2747 * taking the snap. You want to do this anyway to
2748 * ensure a consistent application view
2749 * (i.e. lockfs).
2750 *
2751 * More distant ancestors are irrelevant. The
2752 * shared flag will be set in their case.
2753 */
2754 thin_defer_cell(tc, virt_cell);
2755 return DM_MAPIO_SUBMITTED;
2756 }
2757
2758 build_data_key(tc->td, result.block, &key);
2759 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2760 cell_defer_no_holder(tc, virt_cell);
2761 return DM_MAPIO_SUBMITTED;
2762 }
2763
2764 inc_all_io_entry(tc->pool, bio);
2765 cell_defer_no_holder(tc, data_cell);
2766 cell_defer_no_holder(tc, virt_cell);
2767
2768 remap(tc, bio, result.block);
2769 return DM_MAPIO_REMAPPED;
2770
2771 case -ENODATA:
2772 case -EWOULDBLOCK:
2773 thin_defer_cell(tc, virt_cell);
2774 return DM_MAPIO_SUBMITTED;
2775
2776 default:
2777 /*
2778 * Must always call bio_io_error on failure.
2779 * dm_thin_find_block can fail with -EINVAL if the
2780 * pool is switched to fail-io mode.
2781 */
2782 bio_io_error(bio);
2783 cell_defer_no_holder(tc, virt_cell);
2784 return DM_MAPIO_SUBMITTED;
2785 }
2786}
2787
2788static void requeue_bios(struct pool *pool)
2789{
2790 struct thin_c *tc;
2791
2792 rcu_read_lock();
2793 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2794 spin_lock_irq(&tc->lock);
2795 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2796 bio_list_init(&tc->retry_on_resume_list);
2797 spin_unlock_irq(&tc->lock);
2798 }
2799 rcu_read_unlock();
2800}
2801
2802/*----------------------------------------------------------------
2803 * Binding of control targets to a pool object
2804 *--------------------------------------------------------------*/
2805static bool is_factor(sector_t block_size, uint32_t n)
2806{
2807 return !sector_div(block_size, n);
2808}
2809
2810/*
2811 * If discard_passdown was enabled verify that the data device
2812 * supports discards. Disable discard_passdown if not.
2813 */
2814static void disable_passdown_if_not_supported(struct pool_c *pt)
2815{
2816 struct pool *pool = pt->pool;
2817 struct block_device *data_bdev = pt->data_dev->bdev;
2818 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2819 const char *reason = NULL;
2820
2821 if (!pt->adjusted_pf.discard_passdown)
2822 return;
2823
2824 if (!bdev_max_discard_sectors(pt->data_dev->bdev))
2825 reason = "discard unsupported";
2826
2827 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2828 reason = "max discard sectors smaller than a block";
2829
2830 if (reason) {
2831 DMWARN("Data device (%pg) %s: Disabling discard passdown.", data_bdev, reason);
2832 pt->adjusted_pf.discard_passdown = false;
2833 }
2834}
2835
2836static int bind_control_target(struct pool *pool, struct dm_target *ti)
2837{
2838 struct pool_c *pt = ti->private;
2839
2840 /*
2841 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2842 */
2843 enum pool_mode old_mode = get_pool_mode(pool);
2844 enum pool_mode new_mode = pt->adjusted_pf.mode;
2845
2846 /*
2847 * Don't change the pool's mode until set_pool_mode() below.
2848 * Otherwise the pool's process_* function pointers may
2849 * not match the desired pool mode.
2850 */
2851 pt->adjusted_pf.mode = old_mode;
2852
2853 pool->ti = ti;
2854 pool->pf = pt->adjusted_pf;
2855 pool->low_water_blocks = pt->low_water_blocks;
2856
2857 set_pool_mode(pool, new_mode);
2858
2859 return 0;
2860}
2861
2862static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2863{
2864 if (pool->ti == ti)
2865 pool->ti = NULL;
2866}
2867
2868/*----------------------------------------------------------------
2869 * Pool creation
2870 *--------------------------------------------------------------*/
2871/* Initialize pool features. */
2872static void pool_features_init(struct pool_features *pf)
2873{
2874 pf->mode = PM_WRITE;
2875 pf->zero_new_blocks = true;
2876 pf->discard_enabled = true;
2877 pf->discard_passdown = true;
2878 pf->error_if_no_space = false;
2879}
2880
2881static void __pool_destroy(struct pool *pool)
2882{
2883 __pool_table_remove(pool);
2884
2885 vfree(pool->cell_sort_array);
2886 if (dm_pool_metadata_close(pool->pmd) < 0)
2887 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2888
2889 dm_bio_prison_destroy(pool->prison);
2890 dm_kcopyd_client_destroy(pool->copier);
2891
2892 cancel_delayed_work_sync(&pool->waker);
2893 cancel_delayed_work_sync(&pool->no_space_timeout);
2894 if (pool->wq)
2895 destroy_workqueue(pool->wq);
2896
2897 if (pool->next_mapping)
2898 mempool_free(pool->next_mapping, &pool->mapping_pool);
2899 mempool_exit(&pool->mapping_pool);
2900 dm_deferred_set_destroy(pool->shared_read_ds);
2901 dm_deferred_set_destroy(pool->all_io_ds);
2902 kfree(pool);
2903}
2904
2905static struct kmem_cache *_new_mapping_cache;
2906
2907static struct pool *pool_create(struct mapped_device *pool_md,
2908 struct block_device *metadata_dev,
2909 struct block_device *data_dev,
2910 unsigned long block_size,
2911 int read_only, char **error)
2912{
2913 int r;
2914 void *err_p;
2915 struct pool *pool;
2916 struct dm_pool_metadata *pmd;
2917 bool format_device = read_only ? false : true;
2918
2919 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2920 if (IS_ERR(pmd)) {
2921 *error = "Error creating metadata object";
2922 return (struct pool *)pmd;
2923 }
2924
2925 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2926 if (!pool) {
2927 *error = "Error allocating memory for pool";
2928 err_p = ERR_PTR(-ENOMEM);
2929 goto bad_pool;
2930 }
2931
2932 pool->pmd = pmd;
2933 pool->sectors_per_block = block_size;
2934 if (block_size & (block_size - 1))
2935 pool->sectors_per_block_shift = -1;
2936 else
2937 pool->sectors_per_block_shift = __ffs(block_size);
2938 pool->low_water_blocks = 0;
2939 pool_features_init(&pool->pf);
2940 pool->prison = dm_bio_prison_create();
2941 if (!pool->prison) {
2942 *error = "Error creating pool's bio prison";
2943 err_p = ERR_PTR(-ENOMEM);
2944 goto bad_prison;
2945 }
2946
2947 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2948 if (IS_ERR(pool->copier)) {
2949 r = PTR_ERR(pool->copier);
2950 *error = "Error creating pool's kcopyd client";
2951 err_p = ERR_PTR(r);
2952 goto bad_kcopyd_client;
2953 }
2954
2955 /*
2956 * Create singlethreaded workqueue that will service all devices
2957 * that use this metadata.
2958 */
2959 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2960 if (!pool->wq) {
2961 *error = "Error creating pool's workqueue";
2962 err_p = ERR_PTR(-ENOMEM);
2963 goto bad_wq;
2964 }
2965
2966 throttle_init(&pool->throttle);
2967 INIT_WORK(&pool->worker, do_worker);
2968 INIT_DELAYED_WORK(&pool->waker, do_waker);
2969 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2970 spin_lock_init(&pool->lock);
2971 bio_list_init(&pool->deferred_flush_bios);
2972 bio_list_init(&pool->deferred_flush_completions);
2973 INIT_LIST_HEAD(&pool->prepared_mappings);
2974 INIT_LIST_HEAD(&pool->prepared_discards);
2975 INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2976 INIT_LIST_HEAD(&pool->active_thins);
2977 pool->low_water_triggered = false;
2978 pool->suspended = true;
2979 pool->out_of_data_space = false;
2980
2981 pool->shared_read_ds = dm_deferred_set_create();
2982 if (!pool->shared_read_ds) {
2983 *error = "Error creating pool's shared read deferred set";
2984 err_p = ERR_PTR(-ENOMEM);
2985 goto bad_shared_read_ds;
2986 }
2987
2988 pool->all_io_ds = dm_deferred_set_create();
2989 if (!pool->all_io_ds) {
2990 *error = "Error creating pool's all io deferred set";
2991 err_p = ERR_PTR(-ENOMEM);
2992 goto bad_all_io_ds;
2993 }
2994
2995 pool->next_mapping = NULL;
2996 r = mempool_init_slab_pool(&pool->mapping_pool, MAPPING_POOL_SIZE,
2997 _new_mapping_cache);
2998 if (r) {
2999 *error = "Error creating pool's mapping mempool";
3000 err_p = ERR_PTR(r);
3001 goto bad_mapping_pool;
3002 }
3003
3004 pool->cell_sort_array =
3005 vmalloc(array_size(CELL_SORT_ARRAY_SIZE,
3006 sizeof(*pool->cell_sort_array)));
3007 if (!pool->cell_sort_array) {
3008 *error = "Error allocating cell sort array";
3009 err_p = ERR_PTR(-ENOMEM);
3010 goto bad_sort_array;
3011 }
3012
3013 pool->ref_count = 1;
3014 pool->last_commit_jiffies = jiffies;
3015 pool->pool_md = pool_md;
3016 pool->md_dev = metadata_dev;
3017 pool->data_dev = data_dev;
3018 __pool_table_insert(pool);
3019
3020 return pool;
3021
3022bad_sort_array:
3023 mempool_exit(&pool->mapping_pool);
3024bad_mapping_pool:
3025 dm_deferred_set_destroy(pool->all_io_ds);
3026bad_all_io_ds:
3027 dm_deferred_set_destroy(pool->shared_read_ds);
3028bad_shared_read_ds:
3029 destroy_workqueue(pool->wq);
3030bad_wq:
3031 dm_kcopyd_client_destroy(pool->copier);
3032bad_kcopyd_client:
3033 dm_bio_prison_destroy(pool->prison);
3034bad_prison:
3035 kfree(pool);
3036bad_pool:
3037 if (dm_pool_metadata_close(pmd))
3038 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3039
3040 return err_p;
3041}
3042
3043static void __pool_inc(struct pool *pool)
3044{
3045 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3046 pool->ref_count++;
3047}
3048
3049static void __pool_dec(struct pool *pool)
3050{
3051 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3052 BUG_ON(!pool->ref_count);
3053 if (!--pool->ref_count)
3054 __pool_destroy(pool);
3055}
3056
3057static struct pool *__pool_find(struct mapped_device *pool_md,
3058 struct block_device *metadata_dev,
3059 struct block_device *data_dev,
3060 unsigned long block_size, int read_only,
3061 char **error, int *created)
3062{
3063 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3064
3065 if (pool) {
3066 if (pool->pool_md != pool_md) {
3067 *error = "metadata device already in use by a pool";
3068 return ERR_PTR(-EBUSY);
3069 }
3070 if (pool->data_dev != data_dev) {
3071 *error = "data device already in use by a pool";
3072 return ERR_PTR(-EBUSY);
3073 }
3074 __pool_inc(pool);
3075
3076 } else {
3077 pool = __pool_table_lookup(pool_md);
3078 if (pool) {
3079 if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3080 *error = "different pool cannot replace a pool";
3081 return ERR_PTR(-EINVAL);
3082 }
3083 __pool_inc(pool);
3084
3085 } else {
3086 pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3087 *created = 1;
3088 }
3089 }
3090
3091 return pool;
3092}
3093
3094/*----------------------------------------------------------------
3095 * Pool target methods
3096 *--------------------------------------------------------------*/
3097static void pool_dtr(struct dm_target *ti)
3098{
3099 struct pool_c *pt = ti->private;
3100
3101 mutex_lock(&dm_thin_pool_table.mutex);
3102
3103 unbind_control_target(pt->pool, ti);
3104 __pool_dec(pt->pool);
3105 dm_put_device(ti, pt->metadata_dev);
3106 dm_put_device(ti, pt->data_dev);
3107 kfree(pt);
3108
3109 mutex_unlock(&dm_thin_pool_table.mutex);
3110}
3111
3112static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3113 struct dm_target *ti)
3114{
3115 int r;
3116 unsigned argc;
3117 const char *arg_name;
3118
3119 static const struct dm_arg _args[] = {
3120 {0, 4, "Invalid number of pool feature arguments"},
3121 };
3122
3123 /*
3124 * No feature arguments supplied.
3125 */
3126 if (!as->argc)
3127 return 0;
3128
3129 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3130 if (r)
3131 return -EINVAL;
3132
3133 while (argc && !r) {
3134 arg_name = dm_shift_arg(as);
3135 argc--;
3136
3137 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3138 pf->zero_new_blocks = false;
3139
3140 else if (!strcasecmp(arg_name, "ignore_discard"))
3141 pf->discard_enabled = false;
3142
3143 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3144 pf->discard_passdown = false;
3145
3146 else if (!strcasecmp(arg_name, "read_only"))
3147 pf->mode = PM_READ_ONLY;
3148
3149 else if (!strcasecmp(arg_name, "error_if_no_space"))
3150 pf->error_if_no_space = true;
3151
3152 else {
3153 ti->error = "Unrecognised pool feature requested";
3154 r = -EINVAL;
3155 break;
3156 }
3157 }
3158
3159 return r;
3160}
3161
3162static void metadata_low_callback(void *context)
3163{
3164 struct pool *pool = context;
3165
3166 DMWARN("%s: reached low water mark for metadata device: sending event.",
3167 dm_device_name(pool->pool_md));
3168
3169 dm_table_event(pool->ti->table);
3170}
3171
3172/*
3173 * We need to flush the data device **before** committing the metadata.
3174 *
3175 * This ensures that the data blocks of any newly inserted mappings are
3176 * properly written to non-volatile storage and won't be lost in case of a
3177 * crash.
3178 *
3179 * Failure to do so can result in data corruption in the case of internal or
3180 * external snapshots and in the case of newly provisioned blocks, when block
3181 * zeroing is enabled.
3182 */
3183static int metadata_pre_commit_callback(void *context)
3184{
3185 struct pool *pool = context;
3186
3187 return blkdev_issue_flush(pool->data_dev);
3188}
3189
3190static sector_t get_dev_size(struct block_device *bdev)
3191{
3192 return bdev_nr_sectors(bdev);
3193}
3194
3195static void warn_if_metadata_device_too_big(struct block_device *bdev)
3196{
3197 sector_t metadata_dev_size = get_dev_size(bdev);
3198
3199 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3200 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
3201 bdev, THIN_METADATA_MAX_SECTORS);
3202}
3203
3204static sector_t get_metadata_dev_size(struct block_device *bdev)
3205{
3206 sector_t metadata_dev_size = get_dev_size(bdev);
3207
3208 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3209 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3210
3211 return metadata_dev_size;
3212}
3213
3214static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3215{
3216 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3217
3218 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3219
3220 return metadata_dev_size;
3221}
3222
3223/*
3224 * When a metadata threshold is crossed a dm event is triggered, and
3225 * userland should respond by growing the metadata device. We could let
3226 * userland set the threshold, like we do with the data threshold, but I'm
3227 * not sure they know enough to do this well.
3228 */
3229static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3230{
3231 /*
3232 * 4M is ample for all ops with the possible exception of thin
3233 * device deletion which is harmless if it fails (just retry the
3234 * delete after you've grown the device).
3235 */
3236 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3237 return min((dm_block_t)1024ULL /* 4M */, quarter);
3238}
3239
3240/*
3241 * thin-pool <metadata dev> <data dev>
3242 * <data block size (sectors)>
3243 * <low water mark (blocks)>
3244 * [<#feature args> [<arg>]*]
3245 *
3246 * Optional feature arguments are:
3247 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3248 * ignore_discard: disable discard
3249 * no_discard_passdown: don't pass discards down to the data device
3250 * read_only: Don't allow any changes to be made to the pool metadata.
3251 * error_if_no_space: error IOs, instead of queueing, if no space.
3252 */
3253static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3254{
3255 int r, pool_created = 0;
3256 struct pool_c *pt;
3257 struct pool *pool;
3258 struct pool_features pf;
3259 struct dm_arg_set as;
3260 struct dm_dev *data_dev;
3261 unsigned long block_size;
3262 dm_block_t low_water_blocks;
3263 struct dm_dev *metadata_dev;
3264 fmode_t metadata_mode;
3265
3266 /*
3267 * FIXME Remove validation from scope of lock.
3268 */
3269 mutex_lock(&dm_thin_pool_table.mutex);
3270
3271 if (argc < 4) {
3272 ti->error = "Invalid argument count";
3273 r = -EINVAL;
3274 goto out_unlock;
3275 }
3276
3277 as.argc = argc;
3278 as.argv = argv;
3279
3280 /* make sure metadata and data are different devices */
3281 if (!strcmp(argv[0], argv[1])) {
3282 ti->error = "Error setting metadata or data device";
3283 r = -EINVAL;
3284 goto out_unlock;
3285 }
3286
3287 /*
3288 * Set default pool features.
3289 */
3290 pool_features_init(&pf);
3291
3292 dm_consume_args(&as, 4);
3293 r = parse_pool_features(&as, &pf, ti);
3294 if (r)
3295 goto out_unlock;
3296
3297 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3298 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3299 if (r) {
3300 ti->error = "Error opening metadata block device";
3301 goto out_unlock;
3302 }
3303 warn_if_metadata_device_too_big(metadata_dev->bdev);
3304
3305 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3306 if (r) {
3307 ti->error = "Error getting data device";
3308 goto out_metadata;
3309 }
3310
3311 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3312 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3313 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3314 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3315 ti->error = "Invalid block size";
3316 r = -EINVAL;
3317 goto out;
3318 }
3319
3320 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3321 ti->error = "Invalid low water mark";
3322 r = -EINVAL;
3323 goto out;
3324 }
3325
3326 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3327 if (!pt) {
3328 r = -ENOMEM;
3329 goto out;
3330 }
3331
3332 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3333 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3334 if (IS_ERR(pool)) {
3335 r = PTR_ERR(pool);
3336 goto out_free_pt;
3337 }
3338
3339 /*
3340 * 'pool_created' reflects whether this is the first table load.
3341 * Top level discard support is not allowed to be changed after
3342 * initial load. This would require a pool reload to trigger thin
3343 * device changes.
3344 */
3345 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3346 ti->error = "Discard support cannot be disabled once enabled";
3347 r = -EINVAL;
3348 goto out_flags_changed;
3349 }
3350
3351 pt->pool = pool;
3352 pt->ti = ti;
3353 pt->metadata_dev = metadata_dev;
3354 pt->data_dev = data_dev;
3355 pt->low_water_blocks = low_water_blocks;
3356 pt->adjusted_pf = pt->requested_pf = pf;
3357 ti->num_flush_bios = 1;
3358
3359 /*
3360 * Only need to enable discards if the pool should pass
3361 * them down to the data device. The thin device's discard
3362 * processing will cause mappings to be removed from the btree.
3363 */
3364 if (pf.discard_enabled && pf.discard_passdown) {
3365 ti->num_discard_bios = 1;
3366
3367 /*
3368 * Setting 'discards_supported' circumvents the normal
3369 * stacking of discard limits (this keeps the pool and
3370 * thin devices' discard limits consistent).
3371 */
3372 ti->discards_supported = true;
3373 }
3374 ti->private = pt;
3375
3376 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3377 calc_metadata_threshold(pt),
3378 metadata_low_callback,
3379 pool);
3380 if (r) {
3381 ti->error = "Error registering metadata threshold";
3382 goto out_flags_changed;
3383 }
3384
3385 dm_pool_register_pre_commit_callback(pool->pmd,
3386 metadata_pre_commit_callback, pool);
3387
3388 mutex_unlock(&dm_thin_pool_table.mutex);
3389
3390 return 0;
3391
3392out_flags_changed:
3393 __pool_dec(pool);
3394out_free_pt:
3395 kfree(pt);
3396out:
3397 dm_put_device(ti, data_dev);
3398out_metadata:
3399 dm_put_device(ti, metadata_dev);
3400out_unlock:
3401 mutex_unlock(&dm_thin_pool_table.mutex);
3402
3403 return r;
3404}
3405
3406static int pool_map(struct dm_target *ti, struct bio *bio)
3407{
3408 int r;
3409 struct pool_c *pt = ti->private;
3410 struct pool *pool = pt->pool;
3411
3412 /*
3413 * As this is a singleton target, ti->begin is always zero.
3414 */
3415 spin_lock_irq(&pool->lock);
3416 bio_set_dev(bio, pt->data_dev->bdev);
3417 r = DM_MAPIO_REMAPPED;
3418 spin_unlock_irq(&pool->lock);
3419
3420 return r;
3421}
3422
3423static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3424{
3425 int r;
3426 struct pool_c *pt = ti->private;
3427 struct pool *pool = pt->pool;
3428 sector_t data_size = ti->len;
3429 dm_block_t sb_data_size;
3430
3431 *need_commit = false;
3432
3433 (void) sector_div(data_size, pool->sectors_per_block);
3434
3435 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3436 if (r) {
3437 DMERR("%s: failed to retrieve data device size",
3438 dm_device_name(pool->pool_md));
3439 return r;
3440 }
3441
3442 if (data_size < sb_data_size) {
3443 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3444 dm_device_name(pool->pool_md),
3445 (unsigned long long)data_size, sb_data_size);
3446 return -EINVAL;
3447
3448 } else if (data_size > sb_data_size) {
3449 if (dm_pool_metadata_needs_check(pool->pmd)) {
3450 DMERR("%s: unable to grow the data device until repaired.",
3451 dm_device_name(pool->pool_md));
3452 return 0;
3453 }
3454
3455 if (sb_data_size)
3456 DMINFO("%s: growing the data device from %llu to %llu blocks",
3457 dm_device_name(pool->pool_md),
3458 sb_data_size, (unsigned long long)data_size);
3459 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3460 if (r) {
3461 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3462 return r;
3463 }
3464
3465 *need_commit = true;
3466 }
3467
3468 return 0;
3469}
3470
3471static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3472{
3473 int r;
3474 struct pool_c *pt = ti->private;
3475 struct pool *pool = pt->pool;
3476 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3477
3478 *need_commit = false;
3479
3480 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3481
3482 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3483 if (r) {
3484 DMERR("%s: failed to retrieve metadata device size",
3485 dm_device_name(pool->pool_md));
3486 return r;
3487 }
3488
3489 if (metadata_dev_size < sb_metadata_dev_size) {
3490 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3491 dm_device_name(pool->pool_md),
3492 metadata_dev_size, sb_metadata_dev_size);
3493 return -EINVAL;
3494
3495 } else if (metadata_dev_size > sb_metadata_dev_size) {
3496 if (dm_pool_metadata_needs_check(pool->pmd)) {
3497 DMERR("%s: unable to grow the metadata device until repaired.",
3498 dm_device_name(pool->pool_md));
3499 return 0;
3500 }
3501
3502 warn_if_metadata_device_too_big(pool->md_dev);
3503 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3504 dm_device_name(pool->pool_md),
3505 sb_metadata_dev_size, metadata_dev_size);
3506
3507 if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3508 set_pool_mode(pool, PM_WRITE);
3509
3510 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3511 if (r) {
3512 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3513 return r;
3514 }
3515
3516 *need_commit = true;
3517 }
3518
3519 return 0;
3520}
3521
3522/*
3523 * Retrieves the number of blocks of the data device from
3524 * the superblock and compares it to the actual device size,
3525 * thus resizing the data device in case it has grown.
3526 *
3527 * This both copes with opening preallocated data devices in the ctr
3528 * being followed by a resume
3529 * -and-
3530 * calling the resume method individually after userspace has
3531 * grown the data device in reaction to a table event.
3532 */
3533static int pool_preresume(struct dm_target *ti)
3534{
3535 int r;
3536 bool need_commit1, need_commit2;
3537 struct pool_c *pt = ti->private;
3538 struct pool *pool = pt->pool;
3539
3540 /*
3541 * Take control of the pool object.
3542 */
3543 r = bind_control_target(pool, ti);
3544 if (r)
3545 goto out;
3546
3547 r = maybe_resize_data_dev(ti, &need_commit1);
3548 if (r)
3549 goto out;
3550
3551 r = maybe_resize_metadata_dev(ti, &need_commit2);
3552 if (r)
3553 goto out;
3554
3555 if (need_commit1 || need_commit2)
3556 (void) commit(pool);
3557out:
3558 /*
3559 * When a thin-pool is PM_FAIL, it cannot be rebuilt if
3560 * bio is in deferred list. Therefore need to return 0
3561 * to allow pool_resume() to flush IO.
3562 */
3563 if (r && get_pool_mode(pool) == PM_FAIL)
3564 r = 0;
3565
3566 return r;
3567}
3568
3569static void pool_suspend_active_thins(struct pool *pool)
3570{
3571 struct thin_c *tc;
3572
3573 /* Suspend all active thin devices */
3574 tc = get_first_thin(pool);
3575 while (tc) {
3576 dm_internal_suspend_noflush(tc->thin_md);
3577 tc = get_next_thin(pool, tc);
3578 }
3579}
3580
3581static void pool_resume_active_thins(struct pool *pool)
3582{
3583 struct thin_c *tc;
3584
3585 /* Resume all active thin devices */
3586 tc = get_first_thin(pool);
3587 while (tc) {
3588 dm_internal_resume(tc->thin_md);
3589 tc = get_next_thin(pool, tc);
3590 }
3591}
3592
3593static void pool_resume(struct dm_target *ti)
3594{
3595 struct pool_c *pt = ti->private;
3596 struct pool *pool = pt->pool;
3597
3598 /*
3599 * Must requeue active_thins' bios and then resume
3600 * active_thins _before_ clearing 'suspend' flag.
3601 */
3602 requeue_bios(pool);
3603 pool_resume_active_thins(pool);
3604
3605 spin_lock_irq(&pool->lock);
3606 pool->low_water_triggered = false;
3607 pool->suspended = false;
3608 spin_unlock_irq(&pool->lock);
3609
3610 do_waker(&pool->waker.work);
3611}
3612
3613static void pool_presuspend(struct dm_target *ti)
3614{
3615 struct pool_c *pt = ti->private;
3616 struct pool *pool = pt->pool;
3617
3618 spin_lock_irq(&pool->lock);
3619 pool->suspended = true;
3620 spin_unlock_irq(&pool->lock);
3621
3622 pool_suspend_active_thins(pool);
3623}
3624
3625static void pool_presuspend_undo(struct dm_target *ti)
3626{
3627 struct pool_c *pt = ti->private;
3628 struct pool *pool = pt->pool;
3629
3630 pool_resume_active_thins(pool);
3631
3632 spin_lock_irq(&pool->lock);
3633 pool->suspended = false;
3634 spin_unlock_irq(&pool->lock);
3635}
3636
3637static void pool_postsuspend(struct dm_target *ti)
3638{
3639 struct pool_c *pt = ti->private;
3640 struct pool *pool = pt->pool;
3641
3642 cancel_delayed_work_sync(&pool->waker);
3643 cancel_delayed_work_sync(&pool->no_space_timeout);
3644 flush_workqueue(pool->wq);
3645 (void) commit(pool);
3646}
3647
3648static int check_arg_count(unsigned argc, unsigned args_required)
3649{
3650 if (argc != args_required) {
3651 DMWARN("Message received with %u arguments instead of %u.",
3652 argc, args_required);
3653 return -EINVAL;
3654 }
3655
3656 return 0;
3657}
3658
3659static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3660{
3661 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3662 *dev_id <= MAX_DEV_ID)
3663 return 0;
3664
3665 if (warning)
3666 DMWARN("Message received with invalid device id: %s", arg);
3667
3668 return -EINVAL;
3669}
3670
3671static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3672{
3673 dm_thin_id dev_id;
3674 int r;
3675
3676 r = check_arg_count(argc, 2);
3677 if (r)
3678 return r;
3679
3680 r = read_dev_id(argv[1], &dev_id, 1);
3681 if (r)
3682 return r;
3683
3684 r = dm_pool_create_thin(pool->pmd, dev_id);
3685 if (r) {
3686 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3687 argv[1]);
3688 return r;
3689 }
3690
3691 return 0;
3692}
3693
3694static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3695{
3696 dm_thin_id dev_id;
3697 dm_thin_id origin_dev_id;
3698 int r;
3699
3700 r = check_arg_count(argc, 3);
3701 if (r)
3702 return r;
3703
3704 r = read_dev_id(argv[1], &dev_id, 1);
3705 if (r)
3706 return r;
3707
3708 r = read_dev_id(argv[2], &origin_dev_id, 1);
3709 if (r)
3710 return r;
3711
3712 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3713 if (r) {
3714 DMWARN("Creation of new snapshot %s of device %s failed.",
3715 argv[1], argv[2]);
3716 return r;
3717 }
3718
3719 return 0;
3720}
3721
3722static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3723{
3724 dm_thin_id dev_id;
3725 int r;
3726
3727 r = check_arg_count(argc, 2);
3728 if (r)
3729 return r;
3730
3731 r = read_dev_id(argv[1], &dev_id, 1);
3732 if (r)
3733 return r;
3734
3735 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3736 if (r)
3737 DMWARN("Deletion of thin device %s failed.", argv[1]);
3738
3739 return r;
3740}
3741
3742static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3743{
3744 dm_thin_id old_id, new_id;
3745 int r;
3746
3747 r = check_arg_count(argc, 3);
3748 if (r)
3749 return r;
3750
3751 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3752 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3753 return -EINVAL;
3754 }
3755
3756 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3757 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3758 return -EINVAL;
3759 }
3760
3761 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3762 if (r) {
3763 DMWARN("Failed to change transaction id from %s to %s.",
3764 argv[1], argv[2]);
3765 return r;
3766 }
3767
3768 return 0;
3769}
3770
3771static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3772{
3773 int r;
3774
3775 r = check_arg_count(argc, 1);
3776 if (r)
3777 return r;
3778
3779 (void) commit(pool);
3780
3781 r = dm_pool_reserve_metadata_snap(pool->pmd);
3782 if (r)
3783 DMWARN("reserve_metadata_snap message failed.");
3784
3785 return r;
3786}
3787
3788static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3789{
3790 int r;
3791
3792 r = check_arg_count(argc, 1);
3793 if (r)
3794 return r;
3795
3796 r = dm_pool_release_metadata_snap(pool->pmd);
3797 if (r)
3798 DMWARN("release_metadata_snap message failed.");
3799
3800 return r;
3801}
3802
3803/*
3804 * Messages supported:
3805 * create_thin <dev_id>
3806 * create_snap <dev_id> <origin_id>
3807 * delete <dev_id>
3808 * set_transaction_id <current_trans_id> <new_trans_id>
3809 * reserve_metadata_snap
3810 * release_metadata_snap
3811 */
3812static int pool_message(struct dm_target *ti, unsigned argc, char **argv,
3813 char *result, unsigned maxlen)
3814{
3815 int r = -EINVAL;
3816 struct pool_c *pt = ti->private;
3817 struct pool *pool = pt->pool;
3818
3819 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3820 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3821 dm_device_name(pool->pool_md));
3822 return -EOPNOTSUPP;
3823 }
3824
3825 if (!strcasecmp(argv[0], "create_thin"))
3826 r = process_create_thin_mesg(argc, argv, pool);
3827
3828 else if (!strcasecmp(argv[0], "create_snap"))
3829 r = process_create_snap_mesg(argc, argv, pool);
3830
3831 else if (!strcasecmp(argv[0], "delete"))
3832 r = process_delete_mesg(argc, argv, pool);
3833
3834 else if (!strcasecmp(argv[0], "set_transaction_id"))
3835 r = process_set_transaction_id_mesg(argc, argv, pool);
3836
3837 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3838 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3839
3840 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3841 r = process_release_metadata_snap_mesg(argc, argv, pool);
3842
3843 else
3844 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3845
3846 if (!r)
3847 (void) commit(pool);
3848
3849 return r;
3850}
3851
3852static void emit_flags(struct pool_features *pf, char *result,
3853 unsigned sz, unsigned maxlen)
3854{
3855 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3856 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3857 pf->error_if_no_space;
3858 DMEMIT("%u ", count);
3859
3860 if (!pf->zero_new_blocks)
3861 DMEMIT("skip_block_zeroing ");
3862
3863 if (!pf->discard_enabled)
3864 DMEMIT("ignore_discard ");
3865
3866 if (!pf->discard_passdown)
3867 DMEMIT("no_discard_passdown ");
3868
3869 if (pf->mode == PM_READ_ONLY)
3870 DMEMIT("read_only ");
3871
3872 if (pf->error_if_no_space)
3873 DMEMIT("error_if_no_space ");
3874}
3875
3876/*
3877 * Status line is:
3878 * <transaction id> <used metadata sectors>/<total metadata sectors>
3879 * <used data sectors>/<total data sectors> <held metadata root>
3880 * <pool mode> <discard config> <no space config> <needs_check>
3881 */
3882static void pool_status(struct dm_target *ti, status_type_t type,
3883 unsigned status_flags, char *result, unsigned maxlen)
3884{
3885 int r;
3886 unsigned sz = 0;
3887 uint64_t transaction_id;
3888 dm_block_t nr_free_blocks_data;
3889 dm_block_t nr_free_blocks_metadata;
3890 dm_block_t nr_blocks_data;
3891 dm_block_t nr_blocks_metadata;
3892 dm_block_t held_root;
3893 enum pool_mode mode;
3894 char buf[BDEVNAME_SIZE];
3895 char buf2[BDEVNAME_SIZE];
3896 struct pool_c *pt = ti->private;
3897 struct pool *pool = pt->pool;
3898
3899 switch (type) {
3900 case STATUSTYPE_INFO:
3901 if (get_pool_mode(pool) == PM_FAIL) {
3902 DMEMIT("Fail");
3903 break;
3904 }
3905
3906 /* Commit to ensure statistics aren't out-of-date */
3907 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3908 (void) commit(pool);
3909
3910 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3911 if (r) {
3912 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3913 dm_device_name(pool->pool_md), r);
3914 goto err;
3915 }
3916
3917 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3918 if (r) {
3919 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3920 dm_device_name(pool->pool_md), r);
3921 goto err;
3922 }
3923
3924 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3925 if (r) {
3926 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3927 dm_device_name(pool->pool_md), r);
3928 goto err;
3929 }
3930
3931 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3932 if (r) {
3933 DMERR("%s: dm_pool_get_free_block_count returned %d",
3934 dm_device_name(pool->pool_md), r);
3935 goto err;
3936 }
3937
3938 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3939 if (r) {
3940 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3941 dm_device_name(pool->pool_md), r);
3942 goto err;
3943 }
3944
3945 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3946 if (r) {
3947 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3948 dm_device_name(pool->pool_md), r);
3949 goto err;
3950 }
3951
3952 DMEMIT("%llu %llu/%llu %llu/%llu ",
3953 (unsigned long long)transaction_id,
3954 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3955 (unsigned long long)nr_blocks_metadata,
3956 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3957 (unsigned long long)nr_blocks_data);
3958
3959 if (held_root)
3960 DMEMIT("%llu ", held_root);
3961 else
3962 DMEMIT("- ");
3963
3964 mode = get_pool_mode(pool);
3965 if (mode == PM_OUT_OF_DATA_SPACE)
3966 DMEMIT("out_of_data_space ");
3967 else if (is_read_only_pool_mode(mode))
3968 DMEMIT("ro ");
3969 else
3970 DMEMIT("rw ");
3971
3972 if (!pool->pf.discard_enabled)
3973 DMEMIT("ignore_discard ");
3974 else if (pool->pf.discard_passdown)
3975 DMEMIT("discard_passdown ");
3976 else
3977 DMEMIT("no_discard_passdown ");
3978
3979 if (pool->pf.error_if_no_space)
3980 DMEMIT("error_if_no_space ");
3981 else
3982 DMEMIT("queue_if_no_space ");
3983
3984 if (dm_pool_metadata_needs_check(pool->pmd))
3985 DMEMIT("needs_check ");
3986 else
3987 DMEMIT("- ");
3988
3989 DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
3990
3991 break;
3992
3993 case STATUSTYPE_TABLE:
3994 DMEMIT("%s %s %lu %llu ",
3995 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3996 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3997 (unsigned long)pool->sectors_per_block,
3998 (unsigned long long)pt->low_water_blocks);
3999 emit_flags(&pt->requested_pf, result, sz, maxlen);
4000 break;
4001
4002 case STATUSTYPE_IMA:
4003 *result = '\0';
4004 break;
4005 }
4006 return;
4007
4008err:
4009 DMEMIT("Error");
4010}
4011
4012static int pool_iterate_devices(struct dm_target *ti,
4013 iterate_devices_callout_fn fn, void *data)
4014{
4015 struct pool_c *pt = ti->private;
4016
4017 return fn(ti, pt->data_dev, 0, ti->len, data);
4018}
4019
4020static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4021{
4022 struct pool_c *pt = ti->private;
4023 struct pool *pool = pt->pool;
4024 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4025
4026 /*
4027 * If max_sectors is smaller than pool->sectors_per_block adjust it
4028 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4029 * This is especially beneficial when the pool's data device is a RAID
4030 * device that has a full stripe width that matches pool->sectors_per_block
4031 * -- because even though partial RAID stripe-sized IOs will be issued to a
4032 * single RAID stripe; when aggregated they will end on a full RAID stripe
4033 * boundary.. which avoids additional partial RAID stripe writes cascading
4034 */
4035 if (limits->max_sectors < pool->sectors_per_block) {
4036 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4037 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4038 limits->max_sectors--;
4039 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4040 }
4041 }
4042
4043 /*
4044 * If the system-determined stacked limits are compatible with the
4045 * pool's blocksize (io_opt is a factor) do not override them.
4046 */
4047 if (io_opt_sectors < pool->sectors_per_block ||
4048 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4049 if (is_factor(pool->sectors_per_block, limits->max_sectors))
4050 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
4051 else
4052 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
4053 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
4054 }
4055
4056 /*
4057 * pt->adjusted_pf is a staging area for the actual features to use.
4058 * They get transferred to the live pool in bind_control_target()
4059 * called from pool_preresume().
4060 */
4061 if (!pt->adjusted_pf.discard_enabled) {
4062 /*
4063 * Must explicitly disallow stacking discard limits otherwise the
4064 * block layer will stack them if pool's data device has support.
4065 */
4066 limits->discard_granularity = 0;
4067 return;
4068 }
4069
4070 disable_passdown_if_not_supported(pt);
4071
4072 /*
4073 * The pool uses the same discard limits as the underlying data
4074 * device. DM core has already set this up.
4075 */
4076}
4077
4078static struct target_type pool_target = {
4079 .name = "thin-pool",
4080 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4081 DM_TARGET_IMMUTABLE,
4082 .version = {1, 22, 0},
4083 .module = THIS_MODULE,
4084 .ctr = pool_ctr,
4085 .dtr = pool_dtr,
4086 .map = pool_map,
4087 .presuspend = pool_presuspend,
4088 .presuspend_undo = pool_presuspend_undo,
4089 .postsuspend = pool_postsuspend,
4090 .preresume = pool_preresume,
4091 .resume = pool_resume,
4092 .message = pool_message,
4093 .status = pool_status,
4094 .iterate_devices = pool_iterate_devices,
4095 .io_hints = pool_io_hints,
4096};
4097
4098/*----------------------------------------------------------------
4099 * Thin target methods
4100 *--------------------------------------------------------------*/
4101static void thin_get(struct thin_c *tc)
4102{
4103 refcount_inc(&tc->refcount);
4104}
4105
4106static void thin_put(struct thin_c *tc)
4107{
4108 if (refcount_dec_and_test(&tc->refcount))
4109 complete(&tc->can_destroy);
4110}
4111
4112static void thin_dtr(struct dm_target *ti)
4113{
4114 struct thin_c *tc = ti->private;
4115
4116 spin_lock_irq(&tc->pool->lock);
4117 list_del_rcu(&tc->list);
4118 spin_unlock_irq(&tc->pool->lock);
4119 synchronize_rcu();
4120
4121 thin_put(tc);
4122 wait_for_completion(&tc->can_destroy);
4123
4124 mutex_lock(&dm_thin_pool_table.mutex);
4125
4126 __pool_dec(tc->pool);
4127 dm_pool_close_thin_device(tc->td);
4128 dm_put_device(ti, tc->pool_dev);
4129 if (tc->origin_dev)
4130 dm_put_device(ti, tc->origin_dev);
4131 kfree(tc);
4132
4133 mutex_unlock(&dm_thin_pool_table.mutex);
4134}
4135
4136/*
4137 * Thin target parameters:
4138 *
4139 * <pool_dev> <dev_id> [origin_dev]
4140 *
4141 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4142 * dev_id: the internal device identifier
4143 * origin_dev: a device external to the pool that should act as the origin
4144 *
4145 * If the pool device has discards disabled, they get disabled for the thin
4146 * device as well.
4147 */
4148static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4149{
4150 int r;
4151 struct thin_c *tc;
4152 struct dm_dev *pool_dev, *origin_dev;
4153 struct mapped_device *pool_md;
4154
4155 mutex_lock(&dm_thin_pool_table.mutex);
4156
4157 if (argc != 2 && argc != 3) {
4158 ti->error = "Invalid argument count";
4159 r = -EINVAL;
4160 goto out_unlock;
4161 }
4162
4163 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4164 if (!tc) {
4165 ti->error = "Out of memory";
4166 r = -ENOMEM;
4167 goto out_unlock;
4168 }
4169 tc->thin_md = dm_table_get_md(ti->table);
4170 spin_lock_init(&tc->lock);
4171 INIT_LIST_HEAD(&tc->deferred_cells);
4172 bio_list_init(&tc->deferred_bio_list);
4173 bio_list_init(&tc->retry_on_resume_list);
4174 tc->sort_bio_list = RB_ROOT;
4175
4176 if (argc == 3) {
4177 if (!strcmp(argv[0], argv[2])) {
4178 ti->error = "Error setting origin device";
4179 r = -EINVAL;
4180 goto bad_origin_dev;
4181 }
4182
4183 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4184 if (r) {
4185 ti->error = "Error opening origin device";
4186 goto bad_origin_dev;
4187 }
4188 tc->origin_dev = origin_dev;
4189 }
4190
4191 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4192 if (r) {
4193 ti->error = "Error opening pool device";
4194 goto bad_pool_dev;
4195 }
4196 tc->pool_dev = pool_dev;
4197
4198 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4199 ti->error = "Invalid device id";
4200 r = -EINVAL;
4201 goto bad_common;
4202 }
4203
4204 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4205 if (!pool_md) {
4206 ti->error = "Couldn't get pool mapped device";
4207 r = -EINVAL;
4208 goto bad_common;
4209 }
4210
4211 tc->pool = __pool_table_lookup(pool_md);
4212 if (!tc->pool) {
4213 ti->error = "Couldn't find pool object";
4214 r = -EINVAL;
4215 goto bad_pool_lookup;
4216 }
4217 __pool_inc(tc->pool);
4218
4219 if (get_pool_mode(tc->pool) == PM_FAIL) {
4220 ti->error = "Couldn't open thin device, Pool is in fail mode";
4221 r = -EINVAL;
4222 goto bad_pool;
4223 }
4224
4225 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4226 if (r) {
4227 ti->error = "Couldn't open thin internal device";
4228 goto bad_pool;
4229 }
4230
4231 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4232 if (r)
4233 goto bad;
4234
4235 ti->num_flush_bios = 1;
4236 ti->flush_supported = true;
4237 ti->accounts_remapped_io = true;
4238 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4239
4240 /* In case the pool supports discards, pass them on. */
4241 if (tc->pool->pf.discard_enabled) {
4242 ti->discards_supported = true;
4243 ti->num_discard_bios = 1;
4244 }
4245
4246 mutex_unlock(&dm_thin_pool_table.mutex);
4247
4248 spin_lock_irq(&tc->pool->lock);
4249 if (tc->pool->suspended) {
4250 spin_unlock_irq(&tc->pool->lock);
4251 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4252 ti->error = "Unable to activate thin device while pool is suspended";
4253 r = -EINVAL;
4254 goto bad;
4255 }
4256 refcount_set(&tc->refcount, 1);
4257 init_completion(&tc->can_destroy);
4258 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4259 spin_unlock_irq(&tc->pool->lock);
4260 /*
4261 * This synchronize_rcu() call is needed here otherwise we risk a
4262 * wake_worker() call finding no bios to process (because the newly
4263 * added tc isn't yet visible). So this reduces latency since we
4264 * aren't then dependent on the periodic commit to wake_worker().
4265 */
4266 synchronize_rcu();
4267
4268 dm_put(pool_md);
4269
4270 return 0;
4271
4272bad:
4273 dm_pool_close_thin_device(tc->td);
4274bad_pool:
4275 __pool_dec(tc->pool);
4276bad_pool_lookup:
4277 dm_put(pool_md);
4278bad_common:
4279 dm_put_device(ti, tc->pool_dev);
4280bad_pool_dev:
4281 if (tc->origin_dev)
4282 dm_put_device(ti, tc->origin_dev);
4283bad_origin_dev:
4284 kfree(tc);
4285out_unlock:
4286 mutex_unlock(&dm_thin_pool_table.mutex);
4287
4288 return r;
4289}
4290
4291static int thin_map(struct dm_target *ti, struct bio *bio)
4292{
4293 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4294
4295 return thin_bio_map(ti, bio);
4296}
4297
4298static int thin_endio(struct dm_target *ti, struct bio *bio,
4299 blk_status_t *err)
4300{
4301 unsigned long flags;
4302 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4303 struct list_head work;
4304 struct dm_thin_new_mapping *m, *tmp;
4305 struct pool *pool = h->tc->pool;
4306
4307 if (h->shared_read_entry) {
4308 INIT_LIST_HEAD(&work);
4309 dm_deferred_entry_dec(h->shared_read_entry, &work);
4310
4311 spin_lock_irqsave(&pool->lock, flags);
4312 list_for_each_entry_safe(m, tmp, &work, list) {
4313 list_del(&m->list);
4314 __complete_mapping_preparation(m);
4315 }
4316 spin_unlock_irqrestore(&pool->lock, flags);
4317 }
4318
4319 if (h->all_io_entry) {
4320 INIT_LIST_HEAD(&work);
4321 dm_deferred_entry_dec(h->all_io_entry, &work);
4322 if (!list_empty(&work)) {
4323 spin_lock_irqsave(&pool->lock, flags);
4324 list_for_each_entry_safe(m, tmp, &work, list)
4325 list_add_tail(&m->list, &pool->prepared_discards);
4326 spin_unlock_irqrestore(&pool->lock, flags);
4327 wake_worker(pool);
4328 }
4329 }
4330
4331 if (h->cell)
4332 cell_defer_no_holder(h->tc, h->cell);
4333
4334 return DM_ENDIO_DONE;
4335}
4336
4337static void thin_presuspend(struct dm_target *ti)
4338{
4339 struct thin_c *tc = ti->private;
4340
4341 if (dm_noflush_suspending(ti))
4342 noflush_work(tc, do_noflush_start);
4343}
4344
4345static void thin_postsuspend(struct dm_target *ti)
4346{
4347 struct thin_c *tc = ti->private;
4348
4349 /*
4350 * The dm_noflush_suspending flag has been cleared by now, so
4351 * unfortunately we must always run this.
4352 */
4353 noflush_work(tc, do_noflush_stop);
4354}
4355
4356static int thin_preresume(struct dm_target *ti)
4357{
4358 struct thin_c *tc = ti->private;
4359
4360 if (tc->origin_dev)
4361 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4362
4363 return 0;
4364}
4365
4366/*
4367 * <nr mapped sectors> <highest mapped sector>
4368 */
4369static void thin_status(struct dm_target *ti, status_type_t type,
4370 unsigned status_flags, char *result, unsigned maxlen)
4371{
4372 int r;
4373 ssize_t sz = 0;
4374 dm_block_t mapped, highest;
4375 char buf[BDEVNAME_SIZE];
4376 struct thin_c *tc = ti->private;
4377
4378 if (get_pool_mode(tc->pool) == PM_FAIL) {
4379 DMEMIT("Fail");
4380 return;
4381 }
4382
4383 if (!tc->td)
4384 DMEMIT("-");
4385 else {
4386 switch (type) {
4387 case STATUSTYPE_INFO:
4388 r = dm_thin_get_mapped_count(tc->td, &mapped);
4389 if (r) {
4390 DMERR("dm_thin_get_mapped_count returned %d", r);
4391 goto err;
4392 }
4393
4394 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4395 if (r < 0) {
4396 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4397 goto err;
4398 }
4399
4400 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4401 if (r)
4402 DMEMIT("%llu", ((highest + 1) *
4403 tc->pool->sectors_per_block) - 1);
4404 else
4405 DMEMIT("-");
4406 break;
4407
4408 case STATUSTYPE_TABLE:
4409 DMEMIT("%s %lu",
4410 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4411 (unsigned long) tc->dev_id);
4412 if (tc->origin_dev)
4413 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4414 break;
4415
4416 case STATUSTYPE_IMA:
4417 *result = '\0';
4418 break;
4419 }
4420 }
4421
4422 return;
4423
4424err:
4425 DMEMIT("Error");
4426}
4427
4428static int thin_iterate_devices(struct dm_target *ti,
4429 iterate_devices_callout_fn fn, void *data)
4430{
4431 sector_t blocks;
4432 struct thin_c *tc = ti->private;
4433 struct pool *pool = tc->pool;
4434
4435 /*
4436 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4437 * we follow a more convoluted path through to the pool's target.
4438 */
4439 if (!pool->ti)
4440 return 0; /* nothing is bound */
4441
4442 blocks = pool->ti->len;
4443 (void) sector_div(blocks, pool->sectors_per_block);
4444 if (blocks)
4445 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4446
4447 return 0;
4448}
4449
4450static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4451{
4452 struct thin_c *tc = ti->private;
4453 struct pool *pool = tc->pool;
4454
4455 if (!pool->pf.discard_enabled)
4456 return;
4457
4458 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4459 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4460}
4461
4462static struct target_type thin_target = {
4463 .name = "thin",
4464 .version = {1, 22, 0},
4465 .module = THIS_MODULE,
4466 .ctr = thin_ctr,
4467 .dtr = thin_dtr,
4468 .map = thin_map,
4469 .end_io = thin_endio,
4470 .preresume = thin_preresume,
4471 .presuspend = thin_presuspend,
4472 .postsuspend = thin_postsuspend,
4473 .status = thin_status,
4474 .iterate_devices = thin_iterate_devices,
4475 .io_hints = thin_io_hints,
4476};
4477
4478/*----------------------------------------------------------------*/
4479
4480static int __init dm_thin_init(void)
4481{
4482 int r = -ENOMEM;
4483
4484 pool_table_init();
4485
4486 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4487 if (!_new_mapping_cache)
4488 return r;
4489
4490 r = dm_register_target(&thin_target);
4491 if (r)
4492 goto bad_new_mapping_cache;
4493
4494 r = dm_register_target(&pool_target);
4495 if (r)
4496 goto bad_thin_target;
4497
4498 return 0;
4499
4500bad_thin_target:
4501 dm_unregister_target(&thin_target);
4502bad_new_mapping_cache:
4503 kmem_cache_destroy(_new_mapping_cache);
4504
4505 return r;
4506}
4507
4508static void dm_thin_exit(void)
4509{
4510 dm_unregister_target(&thin_target);
4511 dm_unregister_target(&pool_target);
4512
4513 kmem_cache_destroy(_new_mapping_cache);
4514
4515 pool_table_exit();
4516}
4517
4518module_init(dm_thin_init);
4519module_exit(dm_thin_exit);
4520
4521module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4522MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4523
4524MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4525MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4526MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011-2012 Red Hat UK.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-thin-metadata.h"
9#include "dm-bio-prison-v1.h"
10#include "dm.h"
11
12#include <linux/device-mapper.h>
13#include <linux/dm-io.h>
14#include <linux/dm-kcopyd.h>
15#include <linux/jiffies.h>
16#include <linux/log2.h>
17#include <linux/list.h>
18#include <linux/rculist.h>
19#include <linux/init.h>
20#include <linux/module.h>
21#include <linux/slab.h>
22#include <linux/vmalloc.h>
23#include <linux/sort.h>
24#include <linux/rbtree.h>
25
26#define DM_MSG_PREFIX "thin"
27
28/*
29 * Tunable constants
30 */
31#define ENDIO_HOOK_POOL_SIZE 1024
32#define MAPPING_POOL_SIZE 1024
33#define COMMIT_PERIOD HZ
34#define NO_SPACE_TIMEOUT_SECS 60
35
36static unsigned int no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
37
38DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
39 "A percentage of time allocated for copy on write");
40
41/*
42 * The block size of the device holding pool data must be
43 * between 64KB and 1GB.
44 */
45#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
46#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
47
48/*
49 * Device id is restricted to 24 bits.
50 */
51#define MAX_DEV_ID ((1 << 24) - 1)
52
53/*
54 * How do we handle breaking sharing of data blocks?
55 * =================================================
56 *
57 * We use a standard copy-on-write btree to store the mappings for the
58 * devices (note I'm talking about copy-on-write of the metadata here, not
59 * the data). When you take an internal snapshot you clone the root node
60 * of the origin btree. After this there is no concept of an origin or a
61 * snapshot. They are just two device trees that happen to point to the
62 * same data blocks.
63 *
64 * When we get a write in we decide if it's to a shared data block using
65 * some timestamp magic. If it is, we have to break sharing.
66 *
67 * Let's say we write to a shared block in what was the origin. The
68 * steps are:
69 *
70 * i) plug io further to this physical block. (see bio_prison code).
71 *
72 * ii) quiesce any read io to that shared data block. Obviously
73 * including all devices that share this block. (see dm_deferred_set code)
74 *
75 * iii) copy the data block to a newly allocate block. This step can be
76 * missed out if the io covers the block. (schedule_copy).
77 *
78 * iv) insert the new mapping into the origin's btree
79 * (process_prepared_mapping). This act of inserting breaks some
80 * sharing of btree nodes between the two devices. Breaking sharing only
81 * effects the btree of that specific device. Btrees for the other
82 * devices that share the block never change. The btree for the origin
83 * device as it was after the last commit is untouched, ie. we're using
84 * persistent data structures in the functional programming sense.
85 *
86 * v) unplug io to this physical block, including the io that triggered
87 * the breaking of sharing.
88 *
89 * Steps (ii) and (iii) occur in parallel.
90 *
91 * The metadata _doesn't_ need to be committed before the io continues. We
92 * get away with this because the io is always written to a _new_ block.
93 * If there's a crash, then:
94 *
95 * - The origin mapping will point to the old origin block (the shared
96 * one). This will contain the data as it was before the io that triggered
97 * the breaking of sharing came in.
98 *
99 * - The snap mapping still points to the old block. As it would after
100 * the commit.
101 *
102 * The downside of this scheme is the timestamp magic isn't perfect, and
103 * will continue to think that data block in the snapshot device is shared
104 * even after the write to the origin has broken sharing. I suspect data
105 * blocks will typically be shared by many different devices, so we're
106 * breaking sharing n + 1 times, rather than n, where n is the number of
107 * devices that reference this data block. At the moment I think the
108 * benefits far, far outweigh the disadvantages.
109 */
110
111/*----------------------------------------------------------------*/
112
113/*
114 * Key building.
115 */
116enum lock_space {
117 VIRTUAL,
118 PHYSICAL
119};
120
121static bool build_key(struct dm_thin_device *td, enum lock_space ls,
122 dm_block_t b, dm_block_t e, struct dm_cell_key *key)
123{
124 key->virtual = (ls == VIRTUAL);
125 key->dev = dm_thin_dev_id(td);
126 key->block_begin = b;
127 key->block_end = e;
128
129 return dm_cell_key_has_valid_range(key);
130}
131
132static void build_data_key(struct dm_thin_device *td, dm_block_t b,
133 struct dm_cell_key *key)
134{
135 (void) build_key(td, PHYSICAL, b, b + 1llu, key);
136}
137
138static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
139 struct dm_cell_key *key)
140{
141 (void) build_key(td, VIRTUAL, b, b + 1llu, key);
142}
143
144/*----------------------------------------------------------------*/
145
146#define THROTTLE_THRESHOLD (1 * HZ)
147
148struct throttle {
149 struct rw_semaphore lock;
150 unsigned long threshold;
151 bool throttle_applied;
152};
153
154static void throttle_init(struct throttle *t)
155{
156 init_rwsem(&t->lock);
157 t->throttle_applied = false;
158}
159
160static void throttle_work_start(struct throttle *t)
161{
162 t->threshold = jiffies + THROTTLE_THRESHOLD;
163}
164
165static void throttle_work_update(struct throttle *t)
166{
167 if (!t->throttle_applied && time_is_before_jiffies(t->threshold)) {
168 down_write(&t->lock);
169 t->throttle_applied = true;
170 }
171}
172
173static void throttle_work_complete(struct throttle *t)
174{
175 if (t->throttle_applied) {
176 t->throttle_applied = false;
177 up_write(&t->lock);
178 }
179}
180
181static void throttle_lock(struct throttle *t)
182{
183 down_read(&t->lock);
184}
185
186static void throttle_unlock(struct throttle *t)
187{
188 up_read(&t->lock);
189}
190
191/*----------------------------------------------------------------*/
192
193/*
194 * A pool device ties together a metadata device and a data device. It
195 * also provides the interface for creating and destroying internal
196 * devices.
197 */
198struct dm_thin_new_mapping;
199
200/*
201 * The pool runs in various modes. Ordered in degraded order for comparisons.
202 */
203enum pool_mode {
204 PM_WRITE, /* metadata may be changed */
205 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
206
207 /*
208 * Like READ_ONLY, except may switch back to WRITE on metadata resize. Reported as READ_ONLY.
209 */
210 PM_OUT_OF_METADATA_SPACE,
211 PM_READ_ONLY, /* metadata may not be changed */
212
213 PM_FAIL, /* all I/O fails */
214};
215
216struct pool_features {
217 enum pool_mode mode;
218
219 bool zero_new_blocks:1;
220 bool discard_enabled:1;
221 bool discard_passdown:1;
222 bool error_if_no_space:1;
223};
224
225struct thin_c;
226typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
227typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
228typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
229
230#define CELL_SORT_ARRAY_SIZE 8192
231
232struct pool {
233 struct list_head list;
234 struct dm_target *ti; /* Only set if a pool target is bound */
235
236 struct mapped_device *pool_md;
237 struct block_device *data_dev;
238 struct block_device *md_dev;
239 struct dm_pool_metadata *pmd;
240
241 dm_block_t low_water_blocks;
242 uint32_t sectors_per_block;
243 int sectors_per_block_shift;
244
245 struct pool_features pf;
246 bool low_water_triggered:1; /* A dm event has been sent */
247 bool suspended:1;
248 bool out_of_data_space:1;
249
250 struct dm_bio_prison *prison;
251 struct dm_kcopyd_client *copier;
252
253 struct work_struct worker;
254 struct workqueue_struct *wq;
255 struct throttle throttle;
256 struct delayed_work waker;
257 struct delayed_work no_space_timeout;
258
259 unsigned long last_commit_jiffies;
260 unsigned int ref_count;
261
262 spinlock_t lock;
263 struct bio_list deferred_flush_bios;
264 struct bio_list deferred_flush_completions;
265 struct list_head prepared_mappings;
266 struct list_head prepared_discards;
267 struct list_head prepared_discards_pt2;
268 struct list_head active_thins;
269
270 struct dm_deferred_set *shared_read_ds;
271 struct dm_deferred_set *all_io_ds;
272
273 struct dm_thin_new_mapping *next_mapping;
274
275 process_bio_fn process_bio;
276 process_bio_fn process_discard;
277
278 process_cell_fn process_cell;
279 process_cell_fn process_discard_cell;
280
281 process_mapping_fn process_prepared_mapping;
282 process_mapping_fn process_prepared_discard;
283 process_mapping_fn process_prepared_discard_pt2;
284
285 struct dm_bio_prison_cell **cell_sort_array;
286
287 mempool_t mapping_pool;
288};
289
290static void metadata_operation_failed(struct pool *pool, const char *op, int r);
291
292static enum pool_mode get_pool_mode(struct pool *pool)
293{
294 return pool->pf.mode;
295}
296
297static void notify_of_pool_mode_change(struct pool *pool)
298{
299 static const char *descs[] = {
300 "write",
301 "out-of-data-space",
302 "read-only",
303 "read-only",
304 "fail"
305 };
306 const char *extra_desc = NULL;
307 enum pool_mode mode = get_pool_mode(pool);
308
309 if (mode == PM_OUT_OF_DATA_SPACE) {
310 if (!pool->pf.error_if_no_space)
311 extra_desc = " (queue IO)";
312 else
313 extra_desc = " (error IO)";
314 }
315
316 dm_table_event(pool->ti->table);
317 DMINFO("%s: switching pool to %s%s mode",
318 dm_device_name(pool->pool_md),
319 descs[(int)mode], extra_desc ? : "");
320}
321
322/*
323 * Target context for a pool.
324 */
325struct pool_c {
326 struct dm_target *ti;
327 struct pool *pool;
328 struct dm_dev *data_dev;
329 struct dm_dev *metadata_dev;
330
331 dm_block_t low_water_blocks;
332 struct pool_features requested_pf; /* Features requested during table load */
333 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
334};
335
336/*
337 * Target context for a thin.
338 */
339struct thin_c {
340 struct list_head list;
341 struct dm_dev *pool_dev;
342 struct dm_dev *origin_dev;
343 sector_t origin_size;
344 dm_thin_id dev_id;
345
346 struct pool *pool;
347 struct dm_thin_device *td;
348 struct mapped_device *thin_md;
349
350 bool requeue_mode:1;
351 spinlock_t lock;
352 struct list_head deferred_cells;
353 struct bio_list deferred_bio_list;
354 struct bio_list retry_on_resume_list;
355 struct rb_root sort_bio_list; /* sorted list of deferred bios */
356
357 /*
358 * Ensures the thin is not destroyed until the worker has finished
359 * iterating the active_thins list.
360 */
361 refcount_t refcount;
362 struct completion can_destroy;
363};
364
365/*----------------------------------------------------------------*/
366
367static bool block_size_is_power_of_two(struct pool *pool)
368{
369 return pool->sectors_per_block_shift >= 0;
370}
371
372static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
373{
374 return block_size_is_power_of_two(pool) ?
375 (b << pool->sectors_per_block_shift) :
376 (b * pool->sectors_per_block);
377}
378
379/*----------------------------------------------------------------*/
380
381struct discard_op {
382 struct thin_c *tc;
383 struct blk_plug plug;
384 struct bio *parent_bio;
385 struct bio *bio;
386};
387
388static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
389{
390 BUG_ON(!parent);
391
392 op->tc = tc;
393 blk_start_plug(&op->plug);
394 op->parent_bio = parent;
395 op->bio = NULL;
396}
397
398static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
399{
400 struct thin_c *tc = op->tc;
401 sector_t s = block_to_sectors(tc->pool, data_b);
402 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
403
404 return __blkdev_issue_discard(tc->pool_dev->bdev, s, len, GFP_NOIO, &op->bio);
405}
406
407static void end_discard(struct discard_op *op, int r)
408{
409 if (op->bio) {
410 /*
411 * Even if one of the calls to issue_discard failed, we
412 * need to wait for the chain to complete.
413 */
414 bio_chain(op->bio, op->parent_bio);
415 op->bio->bi_opf = REQ_OP_DISCARD;
416 submit_bio(op->bio);
417 }
418
419 blk_finish_plug(&op->plug);
420
421 /*
422 * Even if r is set, there could be sub discards in flight that we
423 * need to wait for.
424 */
425 if (r && !op->parent_bio->bi_status)
426 op->parent_bio->bi_status = errno_to_blk_status(r);
427 bio_endio(op->parent_bio);
428}
429
430/*----------------------------------------------------------------*/
431
432/*
433 * wake_worker() is used when new work is queued and when pool_resume is
434 * ready to continue deferred IO processing.
435 */
436static void wake_worker(struct pool *pool)
437{
438 queue_work(pool->wq, &pool->worker);
439}
440
441/*----------------------------------------------------------------*/
442
443static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
444 struct dm_bio_prison_cell **cell_result)
445{
446 int r;
447 struct dm_bio_prison_cell *cell_prealloc;
448
449 /*
450 * Allocate a cell from the prison's mempool.
451 * This might block but it can't fail.
452 */
453 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
454
455 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
456 if (r) {
457 /*
458 * We reused an old cell; we can get rid of
459 * the new one.
460 */
461 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
462 }
463
464 return r;
465}
466
467static void cell_release(struct pool *pool,
468 struct dm_bio_prison_cell *cell,
469 struct bio_list *bios)
470{
471 dm_cell_release(pool->prison, cell, bios);
472 dm_bio_prison_free_cell(pool->prison, cell);
473}
474
475static void cell_visit_release(struct pool *pool,
476 void (*fn)(void *, struct dm_bio_prison_cell *),
477 void *context,
478 struct dm_bio_prison_cell *cell)
479{
480 dm_cell_visit_release(pool->prison, fn, context, cell);
481 dm_bio_prison_free_cell(pool->prison, cell);
482}
483
484static void cell_release_no_holder(struct pool *pool,
485 struct dm_bio_prison_cell *cell,
486 struct bio_list *bios)
487{
488 dm_cell_release_no_holder(pool->prison, cell, bios);
489 dm_bio_prison_free_cell(pool->prison, cell);
490}
491
492static void cell_error_with_code(struct pool *pool,
493 struct dm_bio_prison_cell *cell, blk_status_t error_code)
494{
495 dm_cell_error(pool->prison, cell, error_code);
496 dm_bio_prison_free_cell(pool->prison, cell);
497}
498
499static blk_status_t get_pool_io_error_code(struct pool *pool)
500{
501 return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
502}
503
504static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
505{
506 cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
507}
508
509static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
510{
511 cell_error_with_code(pool, cell, 0);
512}
513
514static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
515{
516 cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
517}
518
519/*----------------------------------------------------------------*/
520
521/*
522 * A global list of pools that uses a struct mapped_device as a key.
523 */
524static struct dm_thin_pool_table {
525 struct mutex mutex;
526 struct list_head pools;
527} dm_thin_pool_table;
528
529static void pool_table_init(void)
530{
531 mutex_init(&dm_thin_pool_table.mutex);
532 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
533}
534
535static void pool_table_exit(void)
536{
537 mutex_destroy(&dm_thin_pool_table.mutex);
538}
539
540static void __pool_table_insert(struct pool *pool)
541{
542 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
543 list_add(&pool->list, &dm_thin_pool_table.pools);
544}
545
546static void __pool_table_remove(struct pool *pool)
547{
548 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
549 list_del(&pool->list);
550}
551
552static struct pool *__pool_table_lookup(struct mapped_device *md)
553{
554 struct pool *pool = NULL, *tmp;
555
556 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
557
558 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
559 if (tmp->pool_md == md) {
560 pool = tmp;
561 break;
562 }
563 }
564
565 return pool;
566}
567
568static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
569{
570 struct pool *pool = NULL, *tmp;
571
572 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
573
574 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
575 if (tmp->md_dev == md_dev) {
576 pool = tmp;
577 break;
578 }
579 }
580
581 return pool;
582}
583
584/*----------------------------------------------------------------*/
585
586struct dm_thin_endio_hook {
587 struct thin_c *tc;
588 struct dm_deferred_entry *shared_read_entry;
589 struct dm_deferred_entry *all_io_entry;
590 struct dm_thin_new_mapping *overwrite_mapping;
591 struct rb_node rb_node;
592 struct dm_bio_prison_cell *cell;
593};
594
595static void error_bio_list(struct bio_list *bios, blk_status_t error)
596{
597 struct bio *bio;
598
599 while ((bio = bio_list_pop(bios))) {
600 bio->bi_status = error;
601 bio_endio(bio);
602 }
603}
604
605static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
606 blk_status_t error)
607{
608 struct bio_list bios;
609
610 bio_list_init(&bios);
611
612 spin_lock_irq(&tc->lock);
613 bio_list_merge_init(&bios, master);
614 spin_unlock_irq(&tc->lock);
615
616 error_bio_list(&bios, error);
617}
618
619static void requeue_deferred_cells(struct thin_c *tc)
620{
621 struct pool *pool = tc->pool;
622 struct list_head cells;
623 struct dm_bio_prison_cell *cell, *tmp;
624
625 INIT_LIST_HEAD(&cells);
626
627 spin_lock_irq(&tc->lock);
628 list_splice_init(&tc->deferred_cells, &cells);
629 spin_unlock_irq(&tc->lock);
630
631 list_for_each_entry_safe(cell, tmp, &cells, user_list)
632 cell_requeue(pool, cell);
633}
634
635static void requeue_io(struct thin_c *tc)
636{
637 struct bio_list bios;
638
639 bio_list_init(&bios);
640
641 spin_lock_irq(&tc->lock);
642 bio_list_merge_init(&bios, &tc->deferred_bio_list);
643 bio_list_merge_init(&bios, &tc->retry_on_resume_list);
644 spin_unlock_irq(&tc->lock);
645
646 error_bio_list(&bios, BLK_STS_DM_REQUEUE);
647 requeue_deferred_cells(tc);
648}
649
650static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
651{
652 struct thin_c *tc;
653
654 rcu_read_lock();
655 list_for_each_entry_rcu(tc, &pool->active_thins, list)
656 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
657 rcu_read_unlock();
658}
659
660static void error_retry_list(struct pool *pool)
661{
662 error_retry_list_with_code(pool, get_pool_io_error_code(pool));
663}
664
665/*
666 * This section of code contains the logic for processing a thin device's IO.
667 * Much of the code depends on pool object resources (lists, workqueues, etc)
668 * but most is exclusively called from the thin target rather than the thin-pool
669 * target.
670 */
671
672static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
673{
674 struct pool *pool = tc->pool;
675 sector_t block_nr = bio->bi_iter.bi_sector;
676
677 if (block_size_is_power_of_two(pool))
678 block_nr >>= pool->sectors_per_block_shift;
679 else
680 (void) sector_div(block_nr, pool->sectors_per_block);
681
682 return block_nr;
683}
684
685/*
686 * Returns the _complete_ blocks that this bio covers.
687 */
688static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
689 dm_block_t *begin, dm_block_t *end)
690{
691 struct pool *pool = tc->pool;
692 sector_t b = bio->bi_iter.bi_sector;
693 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
694
695 b += pool->sectors_per_block - 1ull; /* so we round up */
696
697 if (block_size_is_power_of_two(pool)) {
698 b >>= pool->sectors_per_block_shift;
699 e >>= pool->sectors_per_block_shift;
700 } else {
701 (void) sector_div(b, pool->sectors_per_block);
702 (void) sector_div(e, pool->sectors_per_block);
703 }
704
705 if (e < b) {
706 /* Can happen if the bio is within a single block. */
707 e = b;
708 }
709
710 *begin = b;
711 *end = e;
712}
713
714static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
715{
716 struct pool *pool = tc->pool;
717 sector_t bi_sector = bio->bi_iter.bi_sector;
718
719 bio_set_dev(bio, tc->pool_dev->bdev);
720 if (block_size_is_power_of_two(pool)) {
721 bio->bi_iter.bi_sector =
722 (block << pool->sectors_per_block_shift) |
723 (bi_sector & (pool->sectors_per_block - 1));
724 } else {
725 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
726 sector_div(bi_sector, pool->sectors_per_block);
727 }
728}
729
730static void remap_to_origin(struct thin_c *tc, struct bio *bio)
731{
732 bio_set_dev(bio, tc->origin_dev->bdev);
733}
734
735static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
736{
737 return op_is_flush(bio->bi_opf) &&
738 dm_thin_changed_this_transaction(tc->td);
739}
740
741static void inc_all_io_entry(struct pool *pool, struct bio *bio)
742{
743 struct dm_thin_endio_hook *h;
744
745 if (bio_op(bio) == REQ_OP_DISCARD)
746 return;
747
748 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
749 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
750}
751
752static void issue(struct thin_c *tc, struct bio *bio)
753{
754 struct pool *pool = tc->pool;
755
756 if (!bio_triggers_commit(tc, bio)) {
757 dm_submit_bio_remap(bio, NULL);
758 return;
759 }
760
761 /*
762 * Complete bio with an error if earlier I/O caused changes to
763 * the metadata that can't be committed e.g, due to I/O errors
764 * on the metadata device.
765 */
766 if (dm_thin_aborted_changes(tc->td)) {
767 bio_io_error(bio);
768 return;
769 }
770
771 /*
772 * Batch together any bios that trigger commits and then issue a
773 * single commit for them in process_deferred_bios().
774 */
775 spin_lock_irq(&pool->lock);
776 bio_list_add(&pool->deferred_flush_bios, bio);
777 spin_unlock_irq(&pool->lock);
778}
779
780static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
781{
782 remap_to_origin(tc, bio);
783 issue(tc, bio);
784}
785
786static void remap_and_issue(struct thin_c *tc, struct bio *bio,
787 dm_block_t block)
788{
789 remap(tc, bio, block);
790 issue(tc, bio);
791}
792
793/*----------------------------------------------------------------*/
794
795/*
796 * Bio endio functions.
797 */
798struct dm_thin_new_mapping {
799 struct list_head list;
800
801 bool pass_discard:1;
802 bool maybe_shared:1;
803
804 /*
805 * Track quiescing, copying and zeroing preparation actions. When this
806 * counter hits zero the block is prepared and can be inserted into the
807 * btree.
808 */
809 atomic_t prepare_actions;
810
811 blk_status_t status;
812 struct thin_c *tc;
813 dm_block_t virt_begin, virt_end;
814 dm_block_t data_block;
815 struct dm_bio_prison_cell *cell;
816
817 /*
818 * If the bio covers the whole area of a block then we can avoid
819 * zeroing or copying. Instead this bio is hooked. The bio will
820 * still be in the cell, so care has to be taken to avoid issuing
821 * the bio twice.
822 */
823 struct bio *bio;
824 bio_end_io_t *saved_bi_end_io;
825};
826
827static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
828{
829 struct pool *pool = m->tc->pool;
830
831 if (atomic_dec_and_test(&m->prepare_actions)) {
832 list_add_tail(&m->list, &pool->prepared_mappings);
833 wake_worker(pool);
834 }
835}
836
837static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
838{
839 unsigned long flags;
840 struct pool *pool = m->tc->pool;
841
842 spin_lock_irqsave(&pool->lock, flags);
843 __complete_mapping_preparation(m);
844 spin_unlock_irqrestore(&pool->lock, flags);
845}
846
847static void copy_complete(int read_err, unsigned long write_err, void *context)
848{
849 struct dm_thin_new_mapping *m = context;
850
851 m->status = read_err || write_err ? BLK_STS_IOERR : 0;
852 complete_mapping_preparation(m);
853}
854
855static void overwrite_endio(struct bio *bio)
856{
857 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
858 struct dm_thin_new_mapping *m = h->overwrite_mapping;
859
860 bio->bi_end_io = m->saved_bi_end_io;
861
862 m->status = bio->bi_status;
863 complete_mapping_preparation(m);
864}
865
866/*----------------------------------------------------------------*/
867
868/*
869 * Workqueue.
870 */
871
872/*
873 * Prepared mapping jobs.
874 */
875
876/*
877 * This sends the bios in the cell, except the original holder, back
878 * to the deferred_bios list.
879 */
880static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
881{
882 struct pool *pool = tc->pool;
883 unsigned long flags;
884 struct bio_list bios;
885
886 bio_list_init(&bios);
887 cell_release_no_holder(pool, cell, &bios);
888
889 if (!bio_list_empty(&bios)) {
890 spin_lock_irqsave(&tc->lock, flags);
891 bio_list_merge(&tc->deferred_bio_list, &bios);
892 spin_unlock_irqrestore(&tc->lock, flags);
893 wake_worker(pool);
894 }
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
964 /*
965 * If the bio has the REQ_FUA flag set we must commit the metadata
966 * before signaling its completion.
967 */
968 if (!bio_triggers_commit(tc, bio)) {
969 bio_endio(bio);
970 return;
971 }
972
973 /*
974 * Complete bio with an error if earlier I/O caused changes to the
975 * metadata that can't be committed, e.g, due to I/O errors on the
976 * metadata device.
977 */
978 if (dm_thin_aborted_changes(tc->td)) {
979 bio_io_error(bio);
980 return;
981 }
982
983 /*
984 * Batch together any bios that trigger commits and then issue a
985 * single commit for them in process_deferred_bios().
986 */
987 spin_lock_irq(&pool->lock);
988 bio_list_add(&pool->deferred_flush_completions, bio);
989 spin_unlock_irq(&pool->lock);
990}
991
992static void process_prepared_mapping(struct dm_thin_new_mapping *m)
993{
994 struct thin_c *tc = m->tc;
995 struct pool *pool = tc->pool;
996 struct bio *bio = m->bio;
997 int r;
998
999 if (m->status) {
1000 cell_error(pool, m->cell);
1001 goto out;
1002 }
1003
1004 /*
1005 * Commit the prepared block into the mapping btree.
1006 * Any I/O for this block arriving after this point will get
1007 * remapped to it directly.
1008 */
1009 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
1010 if (r) {
1011 metadata_operation_failed(pool, "dm_thin_insert_block", r);
1012 cell_error(pool, m->cell);
1013 goto out;
1014 }
1015
1016 /*
1017 * Release any bios held while the block was being provisioned.
1018 * If we are processing a write bio that completely covers the block,
1019 * we already processed it so can ignore it now when processing
1020 * the bios in the cell.
1021 */
1022 if (bio) {
1023 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1024 complete_overwrite_bio(tc, bio);
1025 } else {
1026 inc_all_io_entry(tc->pool, m->cell->holder);
1027 remap_and_issue(tc, m->cell->holder, m->data_block);
1028 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
1029 }
1030
1031out:
1032 list_del(&m->list);
1033 mempool_free(m, &pool->mapping_pool);
1034}
1035
1036/*----------------------------------------------------------------*/
1037
1038static void free_discard_mapping(struct dm_thin_new_mapping *m)
1039{
1040 struct thin_c *tc = m->tc;
1041
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(NULL, 1, 0, GFP_NOIO);
1183 discard_parent->bi_end_io = passdown_endio;
1184 discard_parent->bi_private = m;
1185 if (m->maybe_shared)
1186 passdown_double_checking_shared_status(m, discard_parent);
1187 else {
1188 struct discard_op op;
1189
1190 begin_discard(&op, tc, discard_parent);
1191 r = issue_discard(&op, m->data_block, data_end);
1192 end_discard(&op, r);
1193 }
1194}
1195
1196static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1197{
1198 int r;
1199 struct thin_c *tc = m->tc;
1200 struct pool *pool = tc->pool;
1201
1202 /*
1203 * The passdown has completed, so now we can decrement all those
1204 * unmapped blocks.
1205 */
1206 r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1207 m->data_block + (m->virt_end - m->virt_begin));
1208 if (r) {
1209 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1210 bio_io_error(m->bio);
1211 } else
1212 bio_endio(m->bio);
1213
1214 cell_defer_no_holder(tc, m->cell);
1215 mempool_free(m, &pool->mapping_pool);
1216}
1217
1218static void process_prepared(struct pool *pool, struct list_head *head,
1219 process_mapping_fn *fn)
1220{
1221 struct list_head maps;
1222 struct dm_thin_new_mapping *m, *tmp;
1223
1224 INIT_LIST_HEAD(&maps);
1225 spin_lock_irq(&pool->lock);
1226 list_splice_init(head, &maps);
1227 spin_unlock_irq(&pool->lock);
1228
1229 list_for_each_entry_safe(m, tmp, &maps, list)
1230 (*fn)(m);
1231}
1232
1233/*
1234 * Deferred bio jobs.
1235 */
1236static int io_overlaps_block(struct pool *pool, struct bio *bio)
1237{
1238 return bio->bi_iter.bi_size ==
1239 (pool->sectors_per_block << SECTOR_SHIFT);
1240}
1241
1242static int io_overwrites_block(struct pool *pool, struct bio *bio)
1243{
1244 return (bio_data_dir(bio) == WRITE) &&
1245 io_overlaps_block(pool, bio);
1246}
1247
1248static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1249 bio_end_io_t *fn)
1250{
1251 *save = bio->bi_end_io;
1252 bio->bi_end_io = fn;
1253}
1254
1255static int ensure_next_mapping(struct pool *pool)
1256{
1257 if (pool->next_mapping)
1258 return 0;
1259
1260 pool->next_mapping = mempool_alloc(&pool->mapping_pool, GFP_ATOMIC);
1261
1262 return pool->next_mapping ? 0 : -ENOMEM;
1263}
1264
1265static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1266{
1267 struct dm_thin_new_mapping *m = pool->next_mapping;
1268
1269 BUG_ON(!pool->next_mapping);
1270
1271 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1272 INIT_LIST_HEAD(&m->list);
1273 m->bio = NULL;
1274
1275 pool->next_mapping = NULL;
1276
1277 return m;
1278}
1279
1280static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1281 sector_t begin, sector_t end)
1282{
1283 struct dm_io_region to;
1284
1285 to.bdev = tc->pool_dev->bdev;
1286 to.sector = begin;
1287 to.count = end - begin;
1288
1289 dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1290}
1291
1292static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1293 dm_block_t data_begin,
1294 struct dm_thin_new_mapping *m)
1295{
1296 struct pool *pool = tc->pool;
1297 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1298
1299 h->overwrite_mapping = m;
1300 m->bio = bio;
1301 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1302 inc_all_io_entry(pool, bio);
1303 remap_and_issue(tc, bio, data_begin);
1304}
1305
1306/*
1307 * A partial copy also needs to zero the uncopied region.
1308 */
1309static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1310 struct dm_dev *origin, dm_block_t data_origin,
1311 dm_block_t data_dest,
1312 struct dm_bio_prison_cell *cell, struct bio *bio,
1313 sector_t len)
1314{
1315 struct pool *pool = tc->pool;
1316 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1317
1318 m->tc = tc;
1319 m->virt_begin = virt_block;
1320 m->virt_end = virt_block + 1u;
1321 m->data_block = data_dest;
1322 m->cell = cell;
1323
1324 /*
1325 * quiesce action + copy action + an extra reference held for the
1326 * duration of this function (we may need to inc later for a
1327 * partial zero).
1328 */
1329 atomic_set(&m->prepare_actions, 3);
1330
1331 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1332 complete_mapping_preparation(m); /* already quiesced */
1333
1334 /*
1335 * IO to pool_dev remaps to the pool target's data_dev.
1336 *
1337 * If the whole block of data is being overwritten, we can issue the
1338 * bio immediately. Otherwise we use kcopyd to clone the data first.
1339 */
1340 if (io_overwrites_block(pool, bio))
1341 remap_and_issue_overwrite(tc, bio, data_dest, m);
1342 else {
1343 struct dm_io_region from, to;
1344
1345 from.bdev = origin->bdev;
1346 from.sector = data_origin * pool->sectors_per_block;
1347 from.count = len;
1348
1349 to.bdev = tc->pool_dev->bdev;
1350 to.sector = data_dest * pool->sectors_per_block;
1351 to.count = len;
1352
1353 dm_kcopyd_copy(pool->copier, &from, 1, &to,
1354 0, copy_complete, m);
1355
1356 /*
1357 * Do we need to zero a tail region?
1358 */
1359 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1360 atomic_inc(&m->prepare_actions);
1361 ll_zero(tc, m,
1362 data_dest * pool->sectors_per_block + len,
1363 (data_dest + 1) * pool->sectors_per_block);
1364 }
1365 }
1366
1367 complete_mapping_preparation(m); /* drop our ref */
1368}
1369
1370static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1371 dm_block_t data_origin, dm_block_t data_dest,
1372 struct dm_bio_prison_cell *cell, struct bio *bio)
1373{
1374 schedule_copy(tc, virt_block, tc->pool_dev,
1375 data_origin, data_dest, cell, bio,
1376 tc->pool->sectors_per_block);
1377}
1378
1379static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1380 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1381 struct bio *bio)
1382{
1383 struct pool *pool = tc->pool;
1384 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1385
1386 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1387 m->tc = tc;
1388 m->virt_begin = virt_block;
1389 m->virt_end = virt_block + 1u;
1390 m->data_block = data_block;
1391 m->cell = cell;
1392
1393 /*
1394 * If the whole block of data is being overwritten or we are not
1395 * zeroing pre-existing data, we can issue the bio immediately.
1396 * Otherwise we use kcopyd to zero the data first.
1397 */
1398 if (pool->pf.zero_new_blocks) {
1399 if (io_overwrites_block(pool, bio))
1400 remap_and_issue_overwrite(tc, bio, data_block, m);
1401 else {
1402 ll_zero(tc, m, data_block * pool->sectors_per_block,
1403 (data_block + 1) * pool->sectors_per_block);
1404 }
1405 } else
1406 process_prepared_mapping(m);
1407}
1408
1409static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1410 dm_block_t data_dest,
1411 struct dm_bio_prison_cell *cell, struct bio *bio)
1412{
1413 struct pool *pool = tc->pool;
1414 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1415 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1416
1417 if (virt_block_end <= tc->origin_size) {
1418 schedule_copy(tc, virt_block, tc->origin_dev,
1419 virt_block, data_dest, cell, bio,
1420 pool->sectors_per_block);
1421
1422 } else if (virt_block_begin < tc->origin_size) {
1423 schedule_copy(tc, virt_block, tc->origin_dev,
1424 virt_block, data_dest, cell, bio,
1425 tc->origin_size - virt_block_begin);
1426
1427 } else
1428 schedule_zero(tc, virt_block, data_dest, cell, bio);
1429}
1430
1431static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1432
1433static void requeue_bios(struct pool *pool);
1434
1435static bool is_read_only_pool_mode(enum pool_mode mode)
1436{
1437 return (mode == PM_OUT_OF_METADATA_SPACE || mode == PM_READ_ONLY);
1438}
1439
1440static bool is_read_only(struct pool *pool)
1441{
1442 return is_read_only_pool_mode(get_pool_mode(pool));
1443}
1444
1445static void check_for_metadata_space(struct pool *pool)
1446{
1447 int r;
1448 const char *ooms_reason = NULL;
1449 dm_block_t nr_free;
1450
1451 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free);
1452 if (r)
1453 ooms_reason = "Could not get free metadata blocks";
1454 else if (!nr_free)
1455 ooms_reason = "No free metadata blocks";
1456
1457 if (ooms_reason && !is_read_only(pool)) {
1458 DMERR("%s", ooms_reason);
1459 set_pool_mode(pool, PM_OUT_OF_METADATA_SPACE);
1460 }
1461}
1462
1463static void check_for_data_space(struct pool *pool)
1464{
1465 int r;
1466 dm_block_t nr_free;
1467
1468 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1469 return;
1470
1471 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1472 if (r)
1473 return;
1474
1475 if (nr_free) {
1476 set_pool_mode(pool, PM_WRITE);
1477 requeue_bios(pool);
1478 }
1479}
1480
1481/*
1482 * A non-zero return indicates read_only or fail_io mode.
1483 * Many callers don't care about the return value.
1484 */
1485static int commit(struct pool *pool)
1486{
1487 int r;
1488
1489 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE)
1490 return -EINVAL;
1491
1492 r = dm_pool_commit_metadata(pool->pmd);
1493 if (r)
1494 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1495 else {
1496 check_for_metadata_space(pool);
1497 check_for_data_space(pool);
1498 }
1499
1500 return r;
1501}
1502
1503static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1504{
1505 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1506 DMWARN("%s: reached low water mark for data device: sending event.",
1507 dm_device_name(pool->pool_md));
1508 spin_lock_irq(&pool->lock);
1509 pool->low_water_triggered = true;
1510 spin_unlock_irq(&pool->lock);
1511 dm_table_event(pool->ti->table);
1512 }
1513}
1514
1515static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1516{
1517 int r;
1518 dm_block_t free_blocks;
1519 struct pool *pool = tc->pool;
1520
1521 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1522 return -EINVAL;
1523
1524 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1525 if (r) {
1526 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1527 return r;
1528 }
1529
1530 check_low_water_mark(pool, free_blocks);
1531
1532 if (!free_blocks) {
1533 /*
1534 * Try to commit to see if that will free up some
1535 * more space.
1536 */
1537 r = commit(pool);
1538 if (r)
1539 return r;
1540
1541 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1542 if (r) {
1543 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1544 return r;
1545 }
1546
1547 if (!free_blocks) {
1548 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1549 return -ENOSPC;
1550 }
1551 }
1552
1553 r = dm_pool_alloc_data_block(pool->pmd, result);
1554 if (r) {
1555 if (r == -ENOSPC)
1556 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1557 else
1558 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1559 return r;
1560 }
1561
1562 r = dm_pool_get_free_metadata_block_count(pool->pmd, &free_blocks);
1563 if (r) {
1564 metadata_operation_failed(pool, "dm_pool_get_free_metadata_block_count", r);
1565 return r;
1566 }
1567
1568 if (!free_blocks) {
1569 /* Let's commit before we use up the metadata reserve. */
1570 r = commit(pool);
1571 if (r)
1572 return r;
1573 }
1574
1575 return 0;
1576}
1577
1578/*
1579 * If we have run out of space, queue bios until the device is
1580 * resumed, presumably after having been reloaded with more space.
1581 */
1582static void retry_on_resume(struct bio *bio)
1583{
1584 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1585 struct thin_c *tc = h->tc;
1586
1587 spin_lock_irq(&tc->lock);
1588 bio_list_add(&tc->retry_on_resume_list, bio);
1589 spin_unlock_irq(&tc->lock);
1590}
1591
1592static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1593{
1594 enum pool_mode m = get_pool_mode(pool);
1595
1596 switch (m) {
1597 case PM_WRITE:
1598 /* Shouldn't get here */
1599 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1600 return BLK_STS_IOERR;
1601
1602 case PM_OUT_OF_DATA_SPACE:
1603 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1604
1605 case PM_OUT_OF_METADATA_SPACE:
1606 case PM_READ_ONLY:
1607 case PM_FAIL:
1608 return BLK_STS_IOERR;
1609 default:
1610 /* Shouldn't get here */
1611 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1612 return BLK_STS_IOERR;
1613 }
1614}
1615
1616static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1617{
1618 blk_status_t error = should_error_unserviceable_bio(pool);
1619
1620 if (error) {
1621 bio->bi_status = error;
1622 bio_endio(bio);
1623 } else
1624 retry_on_resume(bio);
1625}
1626
1627static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1628{
1629 struct bio *bio;
1630 struct bio_list bios;
1631 blk_status_t error;
1632
1633 error = should_error_unserviceable_bio(pool);
1634 if (error) {
1635 cell_error_with_code(pool, cell, error);
1636 return;
1637 }
1638
1639 bio_list_init(&bios);
1640 cell_release(pool, cell, &bios);
1641
1642 while ((bio = bio_list_pop(&bios)))
1643 retry_on_resume(bio);
1644}
1645
1646static void process_discard_cell_no_passdown(struct thin_c *tc,
1647 struct dm_bio_prison_cell *virt_cell)
1648{
1649 struct pool *pool = tc->pool;
1650 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1651
1652 /*
1653 * We don't need to lock the data blocks, since there's no
1654 * passdown. We only lock data blocks for allocation and breaking sharing.
1655 */
1656 m->tc = tc;
1657 m->virt_begin = virt_cell->key.block_begin;
1658 m->virt_end = virt_cell->key.block_end;
1659 m->cell = virt_cell;
1660 m->bio = virt_cell->holder;
1661
1662 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1663 pool->process_prepared_discard(m);
1664}
1665
1666static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1667 struct bio *bio)
1668{
1669 struct pool *pool = tc->pool;
1670
1671 int r;
1672 bool maybe_shared;
1673 struct dm_cell_key data_key;
1674 struct dm_bio_prison_cell *data_cell;
1675 struct dm_thin_new_mapping *m;
1676 dm_block_t virt_begin, virt_end, data_begin, data_end;
1677 dm_block_t len, next_boundary;
1678
1679 while (begin != end) {
1680 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1681 &data_begin, &maybe_shared);
1682 if (r) {
1683 /*
1684 * Silently fail, letting any mappings we've
1685 * created complete.
1686 */
1687 break;
1688 }
1689
1690 data_end = data_begin + (virt_end - virt_begin);
1691
1692 /*
1693 * Make sure the data region obeys the bio prison restrictions.
1694 */
1695 while (data_begin < data_end) {
1696 r = ensure_next_mapping(pool);
1697 if (r)
1698 return; /* we did our best */
1699
1700 next_boundary = ((data_begin >> BIO_PRISON_MAX_RANGE_SHIFT) + 1)
1701 << BIO_PRISON_MAX_RANGE_SHIFT;
1702 len = min_t(sector_t, data_end - data_begin, next_boundary - data_begin);
1703
1704 /* This key is certainly within range given the above splitting */
1705 (void) build_key(tc->td, PHYSICAL, data_begin, data_begin + len, &data_key);
1706 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1707 /* contention, we'll give up with this range */
1708 data_begin += len;
1709 continue;
1710 }
1711
1712 /*
1713 * IO may still be going to the destination block. We must
1714 * quiesce before we can do the removal.
1715 */
1716 m = get_next_mapping(pool);
1717 m->tc = tc;
1718 m->maybe_shared = maybe_shared;
1719 m->virt_begin = virt_begin;
1720 m->virt_end = virt_begin + len;
1721 m->data_block = data_begin;
1722 m->cell = data_cell;
1723 m->bio = bio;
1724
1725 /*
1726 * The parent bio must not complete before sub discard bios are
1727 * chained to it (see end_discard's bio_chain)!
1728 *
1729 * This per-mapping bi_remaining increment is paired with
1730 * the implicit decrement that occurs via bio_endio() in
1731 * end_discard().
1732 */
1733 bio_inc_remaining(bio);
1734 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1735 pool->process_prepared_discard(m);
1736
1737 virt_begin += len;
1738 data_begin += len;
1739 }
1740
1741 begin = virt_end;
1742 }
1743}
1744
1745static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1746{
1747 struct bio *bio = virt_cell->holder;
1748 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1749
1750 /*
1751 * The virt_cell will only get freed once the origin bio completes.
1752 * This means it will remain locked while all the individual
1753 * passdown bios are in flight.
1754 */
1755 h->cell = virt_cell;
1756 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1757
1758 /*
1759 * We complete the bio now, knowing that the bi_remaining field
1760 * will prevent completion until the sub range discards have
1761 * completed.
1762 */
1763 bio_endio(bio);
1764}
1765
1766static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1767{
1768 dm_block_t begin, end;
1769 struct dm_cell_key virt_key;
1770 struct dm_bio_prison_cell *virt_cell;
1771
1772 get_bio_block_range(tc, bio, &begin, &end);
1773 if (begin == end) {
1774 /*
1775 * The discard covers less than a block.
1776 */
1777 bio_endio(bio);
1778 return;
1779 }
1780
1781 if (unlikely(!build_key(tc->td, VIRTUAL, begin, end, &virt_key))) {
1782 DMERR_LIMIT("Discard doesn't respect bio prison limits");
1783 bio_endio(bio);
1784 return;
1785 }
1786
1787 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell)) {
1788 /*
1789 * Potential starvation issue: We're relying on the
1790 * fs/application being well behaved, and not trying to
1791 * send IO to a region at the same time as discarding it.
1792 * If they do this persistently then it's possible this
1793 * cell will never be granted.
1794 */
1795 return;
1796 }
1797
1798 tc->pool->process_discard_cell(tc, virt_cell);
1799}
1800
1801static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1802 struct dm_cell_key *key,
1803 struct dm_thin_lookup_result *lookup_result,
1804 struct dm_bio_prison_cell *cell)
1805{
1806 int r;
1807 dm_block_t data_block;
1808 struct pool *pool = tc->pool;
1809
1810 r = alloc_data_block(tc, &data_block);
1811 switch (r) {
1812 case 0:
1813 schedule_internal_copy(tc, block, lookup_result->block,
1814 data_block, cell, bio);
1815 break;
1816
1817 case -ENOSPC:
1818 retry_bios_on_resume(pool, cell);
1819 break;
1820
1821 default:
1822 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1823 __func__, r);
1824 cell_error(pool, cell);
1825 break;
1826 }
1827}
1828
1829static void __remap_and_issue_shared_cell(void *context,
1830 struct dm_bio_prison_cell *cell)
1831{
1832 struct remap_info *info = context;
1833 struct bio *bio;
1834
1835 while ((bio = bio_list_pop(&cell->bios))) {
1836 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1837 bio_op(bio) == REQ_OP_DISCARD)
1838 bio_list_add(&info->defer_bios, bio);
1839 else {
1840 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1841
1842 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1843 inc_all_io_entry(info->tc->pool, bio);
1844 bio_list_add(&info->issue_bios, bio);
1845 }
1846 }
1847}
1848
1849static void remap_and_issue_shared_cell(struct thin_c *tc,
1850 struct dm_bio_prison_cell *cell,
1851 dm_block_t block)
1852{
1853 struct bio *bio;
1854 struct remap_info info;
1855
1856 info.tc = tc;
1857 bio_list_init(&info.defer_bios);
1858 bio_list_init(&info.issue_bios);
1859
1860 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1861 &info, cell);
1862
1863 while ((bio = bio_list_pop(&info.defer_bios)))
1864 thin_defer_bio(tc, bio);
1865
1866 while ((bio = bio_list_pop(&info.issue_bios)))
1867 remap_and_issue(tc, bio, block);
1868}
1869
1870static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1871 dm_block_t block,
1872 struct dm_thin_lookup_result *lookup_result,
1873 struct dm_bio_prison_cell *virt_cell)
1874{
1875 struct dm_bio_prison_cell *data_cell;
1876 struct pool *pool = tc->pool;
1877 struct dm_cell_key key;
1878
1879 /*
1880 * If cell is already occupied, then sharing is already in the process
1881 * of being broken so we have nothing further to do here.
1882 */
1883 build_data_key(tc->td, lookup_result->block, &key);
1884 if (bio_detain(pool, &key, bio, &data_cell)) {
1885 cell_defer_no_holder(tc, virt_cell);
1886 return;
1887 }
1888
1889 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1890 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1891 cell_defer_no_holder(tc, virt_cell);
1892 } else {
1893 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1894
1895 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1896 inc_all_io_entry(pool, bio);
1897 remap_and_issue(tc, bio, lookup_result->block);
1898
1899 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1900 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1901 }
1902}
1903
1904static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1905 struct dm_bio_prison_cell *cell)
1906{
1907 int r;
1908 dm_block_t data_block;
1909 struct pool *pool = tc->pool;
1910
1911 /*
1912 * Remap empty bios (flushes) immediately, without provisioning.
1913 */
1914 if (!bio->bi_iter.bi_size) {
1915 inc_all_io_entry(pool, bio);
1916 cell_defer_no_holder(tc, cell);
1917
1918 remap_and_issue(tc, bio, 0);
1919 return;
1920 }
1921
1922 /*
1923 * Fill read bios with zeroes and complete them immediately.
1924 */
1925 if (bio_data_dir(bio) == READ) {
1926 zero_fill_bio(bio);
1927 cell_defer_no_holder(tc, cell);
1928 bio_endio(bio);
1929 return;
1930 }
1931
1932 r = alloc_data_block(tc, &data_block);
1933 switch (r) {
1934 case 0:
1935 if (tc->origin_dev)
1936 schedule_external_copy(tc, block, data_block, cell, bio);
1937 else
1938 schedule_zero(tc, block, data_block, cell, bio);
1939 break;
1940
1941 case -ENOSPC:
1942 retry_bios_on_resume(pool, cell);
1943 break;
1944
1945 default:
1946 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1947 __func__, r);
1948 cell_error(pool, cell);
1949 break;
1950 }
1951}
1952
1953static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1954{
1955 int r;
1956 struct pool *pool = tc->pool;
1957 struct bio *bio = cell->holder;
1958 dm_block_t block = get_bio_block(tc, bio);
1959 struct dm_thin_lookup_result lookup_result;
1960
1961 if (tc->requeue_mode) {
1962 cell_requeue(pool, cell);
1963 return;
1964 }
1965
1966 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1967 switch (r) {
1968 case 0:
1969 if (lookup_result.shared)
1970 process_shared_bio(tc, bio, block, &lookup_result, cell);
1971 else {
1972 inc_all_io_entry(pool, bio);
1973 remap_and_issue(tc, bio, lookup_result.block);
1974 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1975 }
1976 break;
1977
1978 case -ENODATA:
1979 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1980 inc_all_io_entry(pool, bio);
1981 cell_defer_no_holder(tc, cell);
1982
1983 if (bio_end_sector(bio) <= tc->origin_size)
1984 remap_to_origin_and_issue(tc, bio);
1985
1986 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1987 zero_fill_bio(bio);
1988 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1989 remap_to_origin_and_issue(tc, bio);
1990
1991 } else {
1992 zero_fill_bio(bio);
1993 bio_endio(bio);
1994 }
1995 } else
1996 provision_block(tc, bio, block, cell);
1997 break;
1998
1999 default:
2000 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2001 __func__, r);
2002 cell_defer_no_holder(tc, cell);
2003 bio_io_error(bio);
2004 break;
2005 }
2006}
2007
2008static void process_bio(struct thin_c *tc, struct bio *bio)
2009{
2010 struct pool *pool = tc->pool;
2011 dm_block_t block = get_bio_block(tc, bio);
2012 struct dm_bio_prison_cell *cell;
2013 struct dm_cell_key key;
2014
2015 /*
2016 * If cell is already occupied, then the block is already
2017 * being provisioned so we have nothing further to do here.
2018 */
2019 build_virtual_key(tc->td, block, &key);
2020 if (bio_detain(pool, &key, bio, &cell))
2021 return;
2022
2023 process_cell(tc, cell);
2024}
2025
2026static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
2027 struct dm_bio_prison_cell *cell)
2028{
2029 int r;
2030 int rw = bio_data_dir(bio);
2031 dm_block_t block = get_bio_block(tc, bio);
2032 struct dm_thin_lookup_result lookup_result;
2033
2034 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
2035 switch (r) {
2036 case 0:
2037 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
2038 handle_unserviceable_bio(tc->pool, bio);
2039 if (cell)
2040 cell_defer_no_holder(tc, cell);
2041 } else {
2042 inc_all_io_entry(tc->pool, bio);
2043 remap_and_issue(tc, bio, lookup_result.block);
2044 if (cell)
2045 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
2046 }
2047 break;
2048
2049 case -ENODATA:
2050 if (cell)
2051 cell_defer_no_holder(tc, cell);
2052 if (rw != READ) {
2053 handle_unserviceable_bio(tc->pool, bio);
2054 break;
2055 }
2056
2057 if (tc->origin_dev) {
2058 inc_all_io_entry(tc->pool, bio);
2059 remap_to_origin_and_issue(tc, bio);
2060 break;
2061 }
2062
2063 zero_fill_bio(bio);
2064 bio_endio(bio);
2065 break;
2066
2067 default:
2068 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
2069 __func__, r);
2070 if (cell)
2071 cell_defer_no_holder(tc, cell);
2072 bio_io_error(bio);
2073 break;
2074 }
2075}
2076
2077static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
2078{
2079 __process_bio_read_only(tc, bio, NULL);
2080}
2081
2082static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2083{
2084 __process_bio_read_only(tc, cell->holder, cell);
2085}
2086
2087static void process_bio_success(struct thin_c *tc, struct bio *bio)
2088{
2089 bio_endio(bio);
2090}
2091
2092static void process_bio_fail(struct thin_c *tc, struct bio *bio)
2093{
2094 bio_io_error(bio);
2095}
2096
2097static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2098{
2099 cell_success(tc->pool, cell);
2100}
2101
2102static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2103{
2104 cell_error(tc->pool, cell);
2105}
2106
2107/*
2108 * FIXME: should we also commit due to size of transaction, measured in
2109 * metadata blocks?
2110 */
2111static int need_commit_due_to_time(struct pool *pool)
2112{
2113 return !time_in_range(jiffies, pool->last_commit_jiffies,
2114 pool->last_commit_jiffies + COMMIT_PERIOD);
2115}
2116
2117#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
2118#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
2119
2120static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2121{
2122 struct rb_node **rbp, *parent;
2123 struct dm_thin_endio_hook *pbd;
2124 sector_t bi_sector = bio->bi_iter.bi_sector;
2125
2126 rbp = &tc->sort_bio_list.rb_node;
2127 parent = NULL;
2128 while (*rbp) {
2129 parent = *rbp;
2130 pbd = thin_pbd(parent);
2131
2132 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2133 rbp = &(*rbp)->rb_left;
2134 else
2135 rbp = &(*rbp)->rb_right;
2136 }
2137
2138 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2139 rb_link_node(&pbd->rb_node, parent, rbp);
2140 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2141}
2142
2143static void __extract_sorted_bios(struct thin_c *tc)
2144{
2145 struct rb_node *node;
2146 struct dm_thin_endio_hook *pbd;
2147 struct bio *bio;
2148
2149 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2150 pbd = thin_pbd(node);
2151 bio = thin_bio(pbd);
2152
2153 bio_list_add(&tc->deferred_bio_list, bio);
2154 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2155 }
2156
2157 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2158}
2159
2160static void __sort_thin_deferred_bios(struct thin_c *tc)
2161{
2162 struct bio *bio;
2163 struct bio_list bios;
2164
2165 bio_list_init(&bios);
2166 bio_list_merge(&bios, &tc->deferred_bio_list);
2167 bio_list_init(&tc->deferred_bio_list);
2168
2169 /* Sort deferred_bio_list using rb-tree */
2170 while ((bio = bio_list_pop(&bios)))
2171 __thin_bio_rb_add(tc, bio);
2172
2173 /*
2174 * Transfer the sorted bios in sort_bio_list back to
2175 * deferred_bio_list to allow lockless submission of
2176 * all bios.
2177 */
2178 __extract_sorted_bios(tc);
2179}
2180
2181static void process_thin_deferred_bios(struct thin_c *tc)
2182{
2183 struct pool *pool = tc->pool;
2184 struct bio *bio;
2185 struct bio_list bios;
2186 struct blk_plug plug;
2187 unsigned int count = 0;
2188
2189 if (tc->requeue_mode) {
2190 error_thin_bio_list(tc, &tc->deferred_bio_list,
2191 BLK_STS_DM_REQUEUE);
2192 return;
2193 }
2194
2195 bio_list_init(&bios);
2196
2197 spin_lock_irq(&tc->lock);
2198
2199 if (bio_list_empty(&tc->deferred_bio_list)) {
2200 spin_unlock_irq(&tc->lock);
2201 return;
2202 }
2203
2204 __sort_thin_deferred_bios(tc);
2205
2206 bio_list_merge(&bios, &tc->deferred_bio_list);
2207 bio_list_init(&tc->deferred_bio_list);
2208
2209 spin_unlock_irq(&tc->lock);
2210
2211 blk_start_plug(&plug);
2212 while ((bio = bio_list_pop(&bios))) {
2213 /*
2214 * If we've got no free new_mapping structs, and processing
2215 * this bio might require one, we pause until there are some
2216 * prepared mappings to process.
2217 */
2218 if (ensure_next_mapping(pool)) {
2219 spin_lock_irq(&tc->lock);
2220 bio_list_add(&tc->deferred_bio_list, bio);
2221 bio_list_merge(&tc->deferred_bio_list, &bios);
2222 spin_unlock_irq(&tc->lock);
2223 break;
2224 }
2225
2226 if (bio_op(bio) == REQ_OP_DISCARD)
2227 pool->process_discard(tc, bio);
2228 else
2229 pool->process_bio(tc, bio);
2230
2231 if ((count++ & 127) == 0) {
2232 throttle_work_update(&pool->throttle);
2233 dm_pool_issue_prefetches(pool->pmd);
2234 }
2235 cond_resched();
2236 }
2237 blk_finish_plug(&plug);
2238}
2239
2240static int cmp_cells(const void *lhs, const void *rhs)
2241{
2242 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2243 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2244
2245 BUG_ON(!lhs_cell->holder);
2246 BUG_ON(!rhs_cell->holder);
2247
2248 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2249 return -1;
2250
2251 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2252 return 1;
2253
2254 return 0;
2255}
2256
2257static unsigned int sort_cells(struct pool *pool, struct list_head *cells)
2258{
2259 unsigned int count = 0;
2260 struct dm_bio_prison_cell *cell, *tmp;
2261
2262 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2263 if (count >= CELL_SORT_ARRAY_SIZE)
2264 break;
2265
2266 pool->cell_sort_array[count++] = cell;
2267 list_del(&cell->user_list);
2268 }
2269
2270 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2271
2272 return count;
2273}
2274
2275static void process_thin_deferred_cells(struct thin_c *tc)
2276{
2277 struct pool *pool = tc->pool;
2278 struct list_head cells;
2279 struct dm_bio_prison_cell *cell;
2280 unsigned int i, j, count;
2281
2282 INIT_LIST_HEAD(&cells);
2283
2284 spin_lock_irq(&tc->lock);
2285 list_splice_init(&tc->deferred_cells, &cells);
2286 spin_unlock_irq(&tc->lock);
2287
2288 if (list_empty(&cells))
2289 return;
2290
2291 do {
2292 count = sort_cells(tc->pool, &cells);
2293
2294 for (i = 0; i < count; i++) {
2295 cell = pool->cell_sort_array[i];
2296 BUG_ON(!cell->holder);
2297
2298 /*
2299 * If we've got no free new_mapping structs, and processing
2300 * this bio might require one, we pause until there are some
2301 * prepared mappings to process.
2302 */
2303 if (ensure_next_mapping(pool)) {
2304 for (j = i; j < count; j++)
2305 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2306
2307 spin_lock_irq(&tc->lock);
2308 list_splice(&cells, &tc->deferred_cells);
2309 spin_unlock_irq(&tc->lock);
2310 return;
2311 }
2312
2313 if (bio_op(cell->holder) == REQ_OP_DISCARD)
2314 pool->process_discard_cell(tc, cell);
2315 else
2316 pool->process_cell(tc, cell);
2317 }
2318 cond_resched();
2319 } while (!list_empty(&cells));
2320}
2321
2322static void thin_get(struct thin_c *tc);
2323static void thin_put(struct thin_c *tc);
2324
2325/*
2326 * We can't hold rcu_read_lock() around code that can block. So we
2327 * find a thin with the rcu lock held; bump a refcount; then drop
2328 * the lock.
2329 */
2330static struct thin_c *get_first_thin(struct pool *pool)
2331{
2332 struct thin_c *tc = NULL;
2333
2334 rcu_read_lock();
2335 tc = list_first_or_null_rcu(&pool->active_thins, struct thin_c, list);
2336 if (tc)
2337 thin_get(tc);
2338 rcu_read_unlock();
2339
2340 return tc;
2341}
2342
2343static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2344{
2345 struct thin_c *old_tc = tc;
2346
2347 rcu_read_lock();
2348 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2349 thin_get(tc);
2350 thin_put(old_tc);
2351 rcu_read_unlock();
2352 return tc;
2353 }
2354 thin_put(old_tc);
2355 rcu_read_unlock();
2356
2357 return NULL;
2358}
2359
2360static void process_deferred_bios(struct pool *pool)
2361{
2362 struct bio *bio;
2363 struct bio_list bios, bio_completions;
2364 struct thin_c *tc;
2365
2366 tc = get_first_thin(pool);
2367 while (tc) {
2368 process_thin_deferred_cells(tc);
2369 process_thin_deferred_bios(tc);
2370 tc = get_next_thin(pool, tc);
2371 }
2372
2373 /*
2374 * If there are any deferred flush bios, we must commit the metadata
2375 * before issuing them or signaling their completion.
2376 */
2377 bio_list_init(&bios);
2378 bio_list_init(&bio_completions);
2379
2380 spin_lock_irq(&pool->lock);
2381 bio_list_merge(&bios, &pool->deferred_flush_bios);
2382 bio_list_init(&pool->deferred_flush_bios);
2383
2384 bio_list_merge(&bio_completions, &pool->deferred_flush_completions);
2385 bio_list_init(&pool->deferred_flush_completions);
2386 spin_unlock_irq(&pool->lock);
2387
2388 if (bio_list_empty(&bios) && bio_list_empty(&bio_completions) &&
2389 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2390 return;
2391
2392 if (commit(pool)) {
2393 bio_list_merge(&bios, &bio_completions);
2394
2395 while ((bio = bio_list_pop(&bios)))
2396 bio_io_error(bio);
2397 return;
2398 }
2399 pool->last_commit_jiffies = jiffies;
2400
2401 while ((bio = bio_list_pop(&bio_completions)))
2402 bio_endio(bio);
2403
2404 while ((bio = bio_list_pop(&bios))) {
2405 /*
2406 * The data device was flushed as part of metadata commit,
2407 * so complete redundant flushes immediately.
2408 */
2409 if (bio->bi_opf & REQ_PREFLUSH)
2410 bio_endio(bio);
2411 else
2412 dm_submit_bio_remap(bio, NULL);
2413 }
2414}
2415
2416static void do_worker(struct work_struct *ws)
2417{
2418 struct pool *pool = container_of(ws, struct pool, worker);
2419
2420 throttle_work_start(&pool->throttle);
2421 dm_pool_issue_prefetches(pool->pmd);
2422 throttle_work_update(&pool->throttle);
2423 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2424 throttle_work_update(&pool->throttle);
2425 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2426 throttle_work_update(&pool->throttle);
2427 process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2428 throttle_work_update(&pool->throttle);
2429 process_deferred_bios(pool);
2430 throttle_work_complete(&pool->throttle);
2431}
2432
2433/*
2434 * We want to commit periodically so that not too much
2435 * unwritten data builds up.
2436 */
2437static void do_waker(struct work_struct *ws)
2438{
2439 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2440
2441 wake_worker(pool);
2442 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2443}
2444
2445/*
2446 * We're holding onto IO to allow userland time to react. After the
2447 * timeout either the pool will have been resized (and thus back in
2448 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2449 */
2450static void do_no_space_timeout(struct work_struct *ws)
2451{
2452 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2453 no_space_timeout);
2454
2455 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2456 pool->pf.error_if_no_space = true;
2457 notify_of_pool_mode_change(pool);
2458 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2459 }
2460}
2461
2462/*----------------------------------------------------------------*/
2463
2464struct pool_work {
2465 struct work_struct worker;
2466 struct completion complete;
2467};
2468
2469static struct pool_work *to_pool_work(struct work_struct *ws)
2470{
2471 return container_of(ws, struct pool_work, worker);
2472}
2473
2474static void pool_work_complete(struct pool_work *pw)
2475{
2476 complete(&pw->complete);
2477}
2478
2479static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2480 void (*fn)(struct work_struct *))
2481{
2482 INIT_WORK_ONSTACK(&pw->worker, fn);
2483 init_completion(&pw->complete);
2484 queue_work(pool->wq, &pw->worker);
2485 wait_for_completion(&pw->complete);
2486 destroy_work_on_stack(&pw->worker);
2487}
2488
2489/*----------------------------------------------------------------*/
2490
2491struct noflush_work {
2492 struct pool_work pw;
2493 struct thin_c *tc;
2494};
2495
2496static struct noflush_work *to_noflush(struct work_struct *ws)
2497{
2498 return container_of(to_pool_work(ws), struct noflush_work, pw);
2499}
2500
2501static void do_noflush_start(struct work_struct *ws)
2502{
2503 struct noflush_work *w = to_noflush(ws);
2504
2505 w->tc->requeue_mode = true;
2506 requeue_io(w->tc);
2507 pool_work_complete(&w->pw);
2508}
2509
2510static void do_noflush_stop(struct work_struct *ws)
2511{
2512 struct noflush_work *w = to_noflush(ws);
2513
2514 w->tc->requeue_mode = false;
2515 pool_work_complete(&w->pw);
2516}
2517
2518static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2519{
2520 struct noflush_work w;
2521
2522 w.tc = tc;
2523 pool_work_wait(&w.pw, tc->pool, fn);
2524}
2525
2526/*----------------------------------------------------------------*/
2527
2528static void set_discard_callbacks(struct pool *pool)
2529{
2530 struct pool_c *pt = pool->ti->private;
2531
2532 if (pt->adjusted_pf.discard_passdown) {
2533 pool->process_discard_cell = process_discard_cell_passdown;
2534 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2535 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2536 } else {
2537 pool->process_discard_cell = process_discard_cell_no_passdown;
2538 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2539 }
2540}
2541
2542static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2543{
2544 struct pool_c *pt = pool->ti->private;
2545 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2546 enum pool_mode old_mode = get_pool_mode(pool);
2547 unsigned long no_space_timeout = READ_ONCE(no_space_timeout_secs) * HZ;
2548
2549 /*
2550 * Never allow the pool to transition to PM_WRITE mode if user
2551 * intervention is required to verify metadata and data consistency.
2552 */
2553 if (new_mode == PM_WRITE && needs_check) {
2554 DMERR("%s: unable to switch pool to write mode until repaired.",
2555 dm_device_name(pool->pool_md));
2556 if (old_mode != new_mode)
2557 new_mode = old_mode;
2558 else
2559 new_mode = PM_READ_ONLY;
2560 }
2561 /*
2562 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2563 * not going to recover without a thin_repair. So we never let the
2564 * pool move out of the old mode.
2565 */
2566 if (old_mode == PM_FAIL)
2567 new_mode = old_mode;
2568
2569 switch (new_mode) {
2570 case PM_FAIL:
2571 dm_pool_metadata_read_only(pool->pmd);
2572 pool->process_bio = process_bio_fail;
2573 pool->process_discard = process_bio_fail;
2574 pool->process_cell = process_cell_fail;
2575 pool->process_discard_cell = process_cell_fail;
2576 pool->process_prepared_mapping = process_prepared_mapping_fail;
2577 pool->process_prepared_discard = process_prepared_discard_fail;
2578
2579 error_retry_list(pool);
2580 break;
2581
2582 case PM_OUT_OF_METADATA_SPACE:
2583 case PM_READ_ONLY:
2584 dm_pool_metadata_read_only(pool->pmd);
2585 pool->process_bio = process_bio_read_only;
2586 pool->process_discard = process_bio_success;
2587 pool->process_cell = process_cell_read_only;
2588 pool->process_discard_cell = process_cell_success;
2589 pool->process_prepared_mapping = process_prepared_mapping_fail;
2590 pool->process_prepared_discard = process_prepared_discard_success;
2591
2592 error_retry_list(pool);
2593 break;
2594
2595 case PM_OUT_OF_DATA_SPACE:
2596 /*
2597 * Ideally we'd never hit this state; the low water mark
2598 * would trigger userland to extend the pool before we
2599 * completely run out of data space. However, many small
2600 * IOs to unprovisioned space can consume data space at an
2601 * alarming rate. Adjust your low water mark if you're
2602 * frequently seeing this mode.
2603 */
2604 pool->out_of_data_space = true;
2605 pool->process_bio = process_bio_read_only;
2606 pool->process_discard = process_discard_bio;
2607 pool->process_cell = process_cell_read_only;
2608 pool->process_prepared_mapping = process_prepared_mapping;
2609 set_discard_callbacks(pool);
2610
2611 if (!pool->pf.error_if_no_space && no_space_timeout)
2612 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2613 break;
2614
2615 case PM_WRITE:
2616 if (old_mode == PM_OUT_OF_DATA_SPACE)
2617 cancel_delayed_work_sync(&pool->no_space_timeout);
2618 pool->out_of_data_space = false;
2619 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2620 dm_pool_metadata_read_write(pool->pmd);
2621 pool->process_bio = process_bio;
2622 pool->process_discard = process_discard_bio;
2623 pool->process_cell = process_cell;
2624 pool->process_prepared_mapping = process_prepared_mapping;
2625 set_discard_callbacks(pool);
2626 break;
2627 }
2628
2629 pool->pf.mode = new_mode;
2630 /*
2631 * The pool mode may have changed, sync it so bind_control_target()
2632 * doesn't cause an unexpected mode transition on resume.
2633 */
2634 pt->adjusted_pf.mode = new_mode;
2635
2636 if (old_mode != new_mode)
2637 notify_of_pool_mode_change(pool);
2638}
2639
2640static void abort_transaction(struct pool *pool)
2641{
2642 const char *dev_name = dm_device_name(pool->pool_md);
2643
2644 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2645 if (dm_pool_abort_metadata(pool->pmd)) {
2646 DMERR("%s: failed to abort metadata transaction", dev_name);
2647 set_pool_mode(pool, PM_FAIL);
2648 }
2649
2650 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2651 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2652 set_pool_mode(pool, PM_FAIL);
2653 }
2654}
2655
2656static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2657{
2658 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2659 dm_device_name(pool->pool_md), op, r);
2660
2661 abort_transaction(pool);
2662 set_pool_mode(pool, PM_READ_ONLY);
2663}
2664
2665/*----------------------------------------------------------------*/
2666
2667/*
2668 * Mapping functions.
2669 */
2670
2671/*
2672 * Called only while mapping a thin bio to hand it over to the workqueue.
2673 */
2674static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2675{
2676 struct pool *pool = tc->pool;
2677
2678 spin_lock_irq(&tc->lock);
2679 bio_list_add(&tc->deferred_bio_list, bio);
2680 spin_unlock_irq(&tc->lock);
2681
2682 wake_worker(pool);
2683}
2684
2685static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2686{
2687 struct pool *pool = tc->pool;
2688
2689 throttle_lock(&pool->throttle);
2690 thin_defer_bio(tc, bio);
2691 throttle_unlock(&pool->throttle);
2692}
2693
2694static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2695{
2696 struct pool *pool = tc->pool;
2697
2698 throttle_lock(&pool->throttle);
2699 spin_lock_irq(&tc->lock);
2700 list_add_tail(&cell->user_list, &tc->deferred_cells);
2701 spin_unlock_irq(&tc->lock);
2702 throttle_unlock(&pool->throttle);
2703
2704 wake_worker(pool);
2705}
2706
2707static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2708{
2709 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2710
2711 h->tc = tc;
2712 h->shared_read_entry = NULL;
2713 h->all_io_entry = NULL;
2714 h->overwrite_mapping = NULL;
2715 h->cell = NULL;
2716}
2717
2718/*
2719 * Non-blocking function called from the thin target's map function.
2720 */
2721static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2722{
2723 int r;
2724 struct thin_c *tc = ti->private;
2725 dm_block_t block = get_bio_block(tc, bio);
2726 struct dm_thin_device *td = tc->td;
2727 struct dm_thin_lookup_result result;
2728 struct dm_bio_prison_cell *virt_cell, *data_cell;
2729 struct dm_cell_key key;
2730
2731 thin_hook_bio(tc, bio);
2732
2733 if (tc->requeue_mode) {
2734 bio->bi_status = BLK_STS_DM_REQUEUE;
2735 bio_endio(bio);
2736 return DM_MAPIO_SUBMITTED;
2737 }
2738
2739 if (get_pool_mode(tc->pool) == PM_FAIL) {
2740 bio_io_error(bio);
2741 return DM_MAPIO_SUBMITTED;
2742 }
2743
2744 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2745 thin_defer_bio_with_throttle(tc, bio);
2746 return DM_MAPIO_SUBMITTED;
2747 }
2748
2749 /*
2750 * We must hold the virtual cell before doing the lookup, otherwise
2751 * there's a race with discard.
2752 */
2753 build_virtual_key(tc->td, block, &key);
2754 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2755 return DM_MAPIO_SUBMITTED;
2756
2757 r = dm_thin_find_block(td, block, 0, &result);
2758
2759 /*
2760 * Note that we defer readahead too.
2761 */
2762 switch (r) {
2763 case 0:
2764 if (unlikely(result.shared)) {
2765 /*
2766 * We have a race condition here between the
2767 * result.shared value returned by the lookup and
2768 * snapshot creation, which may cause new
2769 * sharing.
2770 *
2771 * To avoid this always quiesce the origin before
2772 * taking the snap. You want to do this anyway to
2773 * ensure a consistent application view
2774 * (i.e. lockfs).
2775 *
2776 * More distant ancestors are irrelevant. The
2777 * shared flag will be set in their case.
2778 */
2779 thin_defer_cell(tc, virt_cell);
2780 return DM_MAPIO_SUBMITTED;
2781 }
2782
2783 build_data_key(tc->td, result.block, &key);
2784 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2785 cell_defer_no_holder(tc, virt_cell);
2786 return DM_MAPIO_SUBMITTED;
2787 }
2788
2789 inc_all_io_entry(tc->pool, bio);
2790 cell_defer_no_holder(tc, data_cell);
2791 cell_defer_no_holder(tc, virt_cell);
2792
2793 remap(tc, bio, result.block);
2794 return DM_MAPIO_REMAPPED;
2795
2796 case -ENODATA:
2797 case -EWOULDBLOCK:
2798 thin_defer_cell(tc, virt_cell);
2799 return DM_MAPIO_SUBMITTED;
2800
2801 default:
2802 /*
2803 * Must always call bio_io_error on failure.
2804 * dm_thin_find_block can fail with -EINVAL if the
2805 * pool is switched to fail-io mode.
2806 */
2807 bio_io_error(bio);
2808 cell_defer_no_holder(tc, virt_cell);
2809 return DM_MAPIO_SUBMITTED;
2810 }
2811}
2812
2813static void requeue_bios(struct pool *pool)
2814{
2815 struct thin_c *tc;
2816
2817 rcu_read_lock();
2818 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2819 spin_lock_irq(&tc->lock);
2820 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2821 bio_list_init(&tc->retry_on_resume_list);
2822 spin_unlock_irq(&tc->lock);
2823 }
2824 rcu_read_unlock();
2825}
2826
2827/*
2828 *--------------------------------------------------------------
2829 * Binding of control targets to a pool object
2830 *--------------------------------------------------------------
2831 */
2832static bool is_factor(sector_t block_size, uint32_t n)
2833{
2834 return !sector_div(block_size, n);
2835}
2836
2837/*
2838 * If discard_passdown was enabled verify that the data device
2839 * supports discards. Disable discard_passdown if not.
2840 */
2841static void disable_discard_passdown_if_not_supported(struct pool_c *pt)
2842{
2843 struct pool *pool = pt->pool;
2844 struct block_device *data_bdev = pt->data_dev->bdev;
2845 struct queue_limits *data_limits = bdev_limits(data_bdev);
2846 const char *reason = NULL;
2847
2848 if (!pt->adjusted_pf.discard_passdown)
2849 return;
2850
2851 if (!bdev_max_discard_sectors(pt->data_dev->bdev))
2852 reason = "discard unsupported";
2853
2854 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2855 reason = "max discard sectors smaller than a block";
2856
2857 if (reason) {
2858 DMWARN("Data device (%pg) %s: Disabling discard passdown.", data_bdev, reason);
2859 pt->adjusted_pf.discard_passdown = false;
2860 }
2861}
2862
2863static int bind_control_target(struct pool *pool, struct dm_target *ti)
2864{
2865 struct pool_c *pt = ti->private;
2866
2867 /*
2868 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2869 */
2870 enum pool_mode old_mode = get_pool_mode(pool);
2871 enum pool_mode new_mode = pt->adjusted_pf.mode;
2872
2873 /*
2874 * Don't change the pool's mode until set_pool_mode() below.
2875 * Otherwise the pool's process_* function pointers may
2876 * not match the desired pool mode.
2877 */
2878 pt->adjusted_pf.mode = old_mode;
2879
2880 pool->ti = ti;
2881 pool->pf = pt->adjusted_pf;
2882 pool->low_water_blocks = pt->low_water_blocks;
2883
2884 set_pool_mode(pool, new_mode);
2885
2886 return 0;
2887}
2888
2889static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2890{
2891 if (pool->ti == ti)
2892 pool->ti = NULL;
2893}
2894
2895/*
2896 *--------------------------------------------------------------
2897 * Pool creation
2898 *--------------------------------------------------------------
2899 */
2900/* Initialize pool features. */
2901static void pool_features_init(struct pool_features *pf)
2902{
2903 pf->mode = PM_WRITE;
2904 pf->zero_new_blocks = true;
2905 pf->discard_enabled = true;
2906 pf->discard_passdown = true;
2907 pf->error_if_no_space = false;
2908}
2909
2910static void __pool_destroy(struct pool *pool)
2911{
2912 __pool_table_remove(pool);
2913
2914 vfree(pool->cell_sort_array);
2915 if (dm_pool_metadata_close(pool->pmd) < 0)
2916 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2917
2918 dm_bio_prison_destroy(pool->prison);
2919 dm_kcopyd_client_destroy(pool->copier);
2920
2921 cancel_delayed_work_sync(&pool->waker);
2922 cancel_delayed_work_sync(&pool->no_space_timeout);
2923 if (pool->wq)
2924 destroy_workqueue(pool->wq);
2925
2926 if (pool->next_mapping)
2927 mempool_free(pool->next_mapping, &pool->mapping_pool);
2928 mempool_exit(&pool->mapping_pool);
2929 dm_deferred_set_destroy(pool->shared_read_ds);
2930 dm_deferred_set_destroy(pool->all_io_ds);
2931 kfree(pool);
2932}
2933
2934static struct kmem_cache *_new_mapping_cache;
2935
2936static struct pool *pool_create(struct mapped_device *pool_md,
2937 struct block_device *metadata_dev,
2938 struct block_device *data_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 ERR_CAST(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->data_dev = data_dev;
3047 __pool_table_insert(pool);
3048
3049 return pool;
3050
3051bad_sort_array:
3052 mempool_exit(&pool->mapping_pool);
3053bad_mapping_pool:
3054 dm_deferred_set_destroy(pool->all_io_ds);
3055bad_all_io_ds:
3056 dm_deferred_set_destroy(pool->shared_read_ds);
3057bad_shared_read_ds:
3058 destroy_workqueue(pool->wq);
3059bad_wq:
3060 dm_kcopyd_client_destroy(pool->copier);
3061bad_kcopyd_client:
3062 dm_bio_prison_destroy(pool->prison);
3063bad_prison:
3064 kfree(pool);
3065bad_pool:
3066 if (dm_pool_metadata_close(pmd))
3067 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
3068
3069 return err_p;
3070}
3071
3072static void __pool_inc(struct pool *pool)
3073{
3074 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3075 pool->ref_count++;
3076}
3077
3078static void __pool_dec(struct pool *pool)
3079{
3080 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
3081 BUG_ON(!pool->ref_count);
3082 if (!--pool->ref_count)
3083 __pool_destroy(pool);
3084}
3085
3086static struct pool *__pool_find(struct mapped_device *pool_md,
3087 struct block_device *metadata_dev,
3088 struct block_device *data_dev,
3089 unsigned long block_size, int read_only,
3090 char **error, int *created)
3091{
3092 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
3093
3094 if (pool) {
3095 if (pool->pool_md != pool_md) {
3096 *error = "metadata device already in use by a pool";
3097 return ERR_PTR(-EBUSY);
3098 }
3099 if (pool->data_dev != data_dev) {
3100 *error = "data device already in use by a pool";
3101 return ERR_PTR(-EBUSY);
3102 }
3103 __pool_inc(pool);
3104
3105 } else {
3106 pool = __pool_table_lookup(pool_md);
3107 if (pool) {
3108 if (pool->md_dev != metadata_dev || pool->data_dev != data_dev) {
3109 *error = "different pool cannot replace a pool";
3110 return ERR_PTR(-EINVAL);
3111 }
3112 __pool_inc(pool);
3113
3114 } else {
3115 pool = pool_create(pool_md, metadata_dev, data_dev, block_size, read_only, error);
3116 *created = 1;
3117 }
3118 }
3119
3120 return pool;
3121}
3122
3123/*
3124 *--------------------------------------------------------------
3125 * Pool target methods
3126 *--------------------------------------------------------------
3127 */
3128static void pool_dtr(struct dm_target *ti)
3129{
3130 struct pool_c *pt = ti->private;
3131
3132 mutex_lock(&dm_thin_pool_table.mutex);
3133
3134 unbind_control_target(pt->pool, ti);
3135 __pool_dec(pt->pool);
3136 dm_put_device(ti, pt->metadata_dev);
3137 dm_put_device(ti, pt->data_dev);
3138 kfree(pt);
3139
3140 mutex_unlock(&dm_thin_pool_table.mutex);
3141}
3142
3143static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3144 struct dm_target *ti)
3145{
3146 int r;
3147 unsigned int argc;
3148 const char *arg_name;
3149
3150 static const struct dm_arg _args[] = {
3151 {0, 4, "Invalid number of pool feature arguments"},
3152 };
3153
3154 /*
3155 * No feature arguments supplied.
3156 */
3157 if (!as->argc)
3158 return 0;
3159
3160 r = dm_read_arg_group(_args, as, &argc, &ti->error);
3161 if (r)
3162 return -EINVAL;
3163
3164 while (argc && !r) {
3165 arg_name = dm_shift_arg(as);
3166 argc--;
3167
3168 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3169 pf->zero_new_blocks = false;
3170
3171 else if (!strcasecmp(arg_name, "ignore_discard"))
3172 pf->discard_enabled = false;
3173
3174 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3175 pf->discard_passdown = false;
3176
3177 else if (!strcasecmp(arg_name, "read_only"))
3178 pf->mode = PM_READ_ONLY;
3179
3180 else if (!strcasecmp(arg_name, "error_if_no_space"))
3181 pf->error_if_no_space = true;
3182
3183 else {
3184 ti->error = "Unrecognised pool feature requested";
3185 r = -EINVAL;
3186 break;
3187 }
3188 }
3189
3190 return r;
3191}
3192
3193static void metadata_low_callback(void *context)
3194{
3195 struct pool *pool = context;
3196
3197 DMWARN("%s: reached low water mark for metadata device: sending event.",
3198 dm_device_name(pool->pool_md));
3199
3200 dm_table_event(pool->ti->table);
3201}
3202
3203/*
3204 * We need to flush the data device **before** committing the metadata.
3205 *
3206 * This ensures that the data blocks of any newly inserted mappings are
3207 * properly written to non-volatile storage and won't be lost in case of a
3208 * crash.
3209 *
3210 * Failure to do so can result in data corruption in the case of internal or
3211 * external snapshots and in the case of newly provisioned blocks, when block
3212 * zeroing is enabled.
3213 */
3214static int metadata_pre_commit_callback(void *context)
3215{
3216 struct pool *pool = context;
3217
3218 return blkdev_issue_flush(pool->data_dev);
3219}
3220
3221static sector_t get_dev_size(struct block_device *bdev)
3222{
3223 return bdev_nr_sectors(bdev);
3224}
3225
3226static void warn_if_metadata_device_too_big(struct block_device *bdev)
3227{
3228 sector_t metadata_dev_size = get_dev_size(bdev);
3229
3230 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3231 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
3232 bdev, THIN_METADATA_MAX_SECTORS);
3233}
3234
3235static sector_t get_metadata_dev_size(struct block_device *bdev)
3236{
3237 sector_t metadata_dev_size = get_dev_size(bdev);
3238
3239 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3240 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3241
3242 return metadata_dev_size;
3243}
3244
3245static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3246{
3247 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3248
3249 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3250
3251 return metadata_dev_size;
3252}
3253
3254/*
3255 * When a metadata threshold is crossed a dm event is triggered, and
3256 * userland should respond by growing the metadata device. We could let
3257 * userland set the threshold, like we do with the data threshold, but I'm
3258 * not sure they know enough to do this well.
3259 */
3260static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3261{
3262 /*
3263 * 4M is ample for all ops with the possible exception of thin
3264 * device deletion which is harmless if it fails (just retry the
3265 * delete after you've grown the device).
3266 */
3267 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3268
3269 return min((dm_block_t)1024ULL /* 4M */, quarter);
3270}
3271
3272/*
3273 * thin-pool <metadata dev> <data dev>
3274 * <data block size (sectors)>
3275 * <low water mark (blocks)>
3276 * [<#feature args> [<arg>]*]
3277 *
3278 * Optional feature arguments are:
3279 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3280 * ignore_discard: disable discard
3281 * no_discard_passdown: don't pass discards down to the data device
3282 * read_only: Don't allow any changes to be made to the pool metadata.
3283 * error_if_no_space: error IOs, instead of queueing, if no space.
3284 */
3285static int pool_ctr(struct dm_target *ti, unsigned int argc, char **argv)
3286{
3287 int r, pool_created = 0;
3288 struct pool_c *pt;
3289 struct pool *pool;
3290 struct pool_features pf;
3291 struct dm_arg_set as;
3292 struct dm_dev *data_dev;
3293 unsigned long block_size;
3294 dm_block_t low_water_blocks;
3295 struct dm_dev *metadata_dev;
3296 blk_mode_t metadata_mode;
3297
3298 /*
3299 * FIXME Remove validation from scope of lock.
3300 */
3301 mutex_lock(&dm_thin_pool_table.mutex);
3302
3303 if (argc < 4) {
3304 ti->error = "Invalid argument count";
3305 r = -EINVAL;
3306 goto out_unlock;
3307 }
3308
3309 as.argc = argc;
3310 as.argv = argv;
3311
3312 /* make sure metadata and data are different devices */
3313 if (!strcmp(argv[0], argv[1])) {
3314 ti->error = "Error setting metadata or data device";
3315 r = -EINVAL;
3316 goto out_unlock;
3317 }
3318
3319 /*
3320 * Set default pool features.
3321 */
3322 pool_features_init(&pf);
3323
3324 dm_consume_args(&as, 4);
3325 r = parse_pool_features(&as, &pf, ti);
3326 if (r)
3327 goto out_unlock;
3328
3329 metadata_mode = BLK_OPEN_READ |
3330 ((pf.mode == PM_READ_ONLY) ? 0 : BLK_OPEN_WRITE);
3331 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3332 if (r) {
3333 ti->error = "Error opening metadata block device";
3334 goto out_unlock;
3335 }
3336 warn_if_metadata_device_too_big(metadata_dev->bdev);
3337
3338 r = dm_get_device(ti, argv[1], BLK_OPEN_READ | BLK_OPEN_WRITE, &data_dev);
3339 if (r) {
3340 ti->error = "Error getting data device";
3341 goto out_metadata;
3342 }
3343
3344 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3345 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3346 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3347 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3348 ti->error = "Invalid block size";
3349 r = -EINVAL;
3350 goto out;
3351 }
3352
3353 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3354 ti->error = "Invalid low water mark";
3355 r = -EINVAL;
3356 goto out;
3357 }
3358
3359 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3360 if (!pt) {
3361 r = -ENOMEM;
3362 goto out;
3363 }
3364
3365 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev, data_dev->bdev,
3366 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3367 if (IS_ERR(pool)) {
3368 r = PTR_ERR(pool);
3369 goto out_free_pt;
3370 }
3371
3372 /*
3373 * 'pool_created' reflects whether this is the first table load.
3374 * Top level discard support is not allowed to be changed after
3375 * initial load. This would require a pool reload to trigger thin
3376 * device changes.
3377 */
3378 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3379 ti->error = "Discard support cannot be disabled once enabled";
3380 r = -EINVAL;
3381 goto out_flags_changed;
3382 }
3383
3384 pt->pool = pool;
3385 pt->ti = ti;
3386 pt->metadata_dev = metadata_dev;
3387 pt->data_dev = data_dev;
3388 pt->low_water_blocks = low_water_blocks;
3389 pt->adjusted_pf = pt->requested_pf = pf;
3390 ti->num_flush_bios = 1;
3391 ti->limit_swap_bios = true;
3392
3393 /*
3394 * Only need to enable discards if the pool should pass
3395 * them down to the data device. The thin device's discard
3396 * processing will cause mappings to be removed from the btree.
3397 */
3398 if (pf.discard_enabled && pf.discard_passdown) {
3399 ti->num_discard_bios = 1;
3400 /*
3401 * Setting 'discards_supported' circumvents the normal
3402 * stacking of discard limits (this keeps the pool and
3403 * thin devices' discard limits consistent).
3404 */
3405 ti->discards_supported = true;
3406 ti->max_discard_granularity = true;
3407 }
3408 ti->private = pt;
3409
3410 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3411 calc_metadata_threshold(pt),
3412 metadata_low_callback,
3413 pool);
3414 if (r) {
3415 ti->error = "Error registering metadata threshold";
3416 goto out_flags_changed;
3417 }
3418
3419 dm_pool_register_pre_commit_callback(pool->pmd,
3420 metadata_pre_commit_callback, pool);
3421
3422 mutex_unlock(&dm_thin_pool_table.mutex);
3423
3424 return 0;
3425
3426out_flags_changed:
3427 __pool_dec(pool);
3428out_free_pt:
3429 kfree(pt);
3430out:
3431 dm_put_device(ti, data_dev);
3432out_metadata:
3433 dm_put_device(ti, metadata_dev);
3434out_unlock:
3435 mutex_unlock(&dm_thin_pool_table.mutex);
3436
3437 return r;
3438}
3439
3440static int pool_map(struct dm_target *ti, struct bio *bio)
3441{
3442 struct pool_c *pt = ti->private;
3443 struct pool *pool = pt->pool;
3444
3445 /*
3446 * As this is a singleton target, ti->begin is always zero.
3447 */
3448 spin_lock_irq(&pool->lock);
3449 bio_set_dev(bio, pt->data_dev->bdev);
3450 spin_unlock_irq(&pool->lock);
3451
3452 return DM_MAPIO_REMAPPED;
3453}
3454
3455static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3456{
3457 int r;
3458 struct pool_c *pt = ti->private;
3459 struct pool *pool = pt->pool;
3460 sector_t data_size = ti->len;
3461 dm_block_t sb_data_size;
3462
3463 *need_commit = false;
3464
3465 (void) sector_div(data_size, pool->sectors_per_block);
3466
3467 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3468 if (r) {
3469 DMERR("%s: failed to retrieve data device size",
3470 dm_device_name(pool->pool_md));
3471 return r;
3472 }
3473
3474 if (data_size < sb_data_size) {
3475 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3476 dm_device_name(pool->pool_md),
3477 (unsigned long long)data_size, sb_data_size);
3478 return -EINVAL;
3479
3480 } else if (data_size > sb_data_size) {
3481 if (dm_pool_metadata_needs_check(pool->pmd)) {
3482 DMERR("%s: unable to grow the data device until repaired.",
3483 dm_device_name(pool->pool_md));
3484 return 0;
3485 }
3486
3487 if (sb_data_size)
3488 DMINFO("%s: growing the data device from %llu to %llu blocks",
3489 dm_device_name(pool->pool_md),
3490 sb_data_size, (unsigned long long)data_size);
3491 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3492 if (r) {
3493 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3494 return r;
3495 }
3496
3497 *need_commit = true;
3498 }
3499
3500 return 0;
3501}
3502
3503static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3504{
3505 int r;
3506 struct pool_c *pt = ti->private;
3507 struct pool *pool = pt->pool;
3508 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3509
3510 *need_commit = false;
3511
3512 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3513
3514 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3515 if (r) {
3516 DMERR("%s: failed to retrieve metadata device size",
3517 dm_device_name(pool->pool_md));
3518 return r;
3519 }
3520
3521 if (metadata_dev_size < sb_metadata_dev_size) {
3522 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3523 dm_device_name(pool->pool_md),
3524 metadata_dev_size, sb_metadata_dev_size);
3525 return -EINVAL;
3526
3527 } else if (metadata_dev_size > sb_metadata_dev_size) {
3528 if (dm_pool_metadata_needs_check(pool->pmd)) {
3529 DMERR("%s: unable to grow the metadata device until repaired.",
3530 dm_device_name(pool->pool_md));
3531 return 0;
3532 }
3533
3534 warn_if_metadata_device_too_big(pool->md_dev);
3535 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3536 dm_device_name(pool->pool_md),
3537 sb_metadata_dev_size, metadata_dev_size);
3538
3539 if (get_pool_mode(pool) == PM_OUT_OF_METADATA_SPACE)
3540 set_pool_mode(pool, PM_WRITE);
3541
3542 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3543 if (r) {
3544 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3545 return r;
3546 }
3547
3548 *need_commit = true;
3549 }
3550
3551 return 0;
3552}
3553
3554/*
3555 * Retrieves the number of blocks of the data device from
3556 * the superblock and compares it to the actual device size,
3557 * thus resizing the data device in case it has grown.
3558 *
3559 * This both copes with opening preallocated data devices in the ctr
3560 * being followed by a resume
3561 * -and-
3562 * calling the resume method individually after userspace has
3563 * grown the data device in reaction to a table event.
3564 */
3565static int pool_preresume(struct dm_target *ti)
3566{
3567 int r;
3568 bool need_commit1, need_commit2;
3569 struct pool_c *pt = ti->private;
3570 struct pool *pool = pt->pool;
3571
3572 /*
3573 * Take control of the pool object.
3574 */
3575 r = bind_control_target(pool, ti);
3576 if (r)
3577 goto out;
3578
3579 r = maybe_resize_data_dev(ti, &need_commit1);
3580 if (r)
3581 goto out;
3582
3583 r = maybe_resize_metadata_dev(ti, &need_commit2);
3584 if (r)
3585 goto out;
3586
3587 if (need_commit1 || need_commit2)
3588 (void) commit(pool);
3589out:
3590 /*
3591 * When a thin-pool is PM_FAIL, it cannot be rebuilt if
3592 * bio is in deferred list. Therefore need to return 0
3593 * to allow pool_resume() to flush IO.
3594 */
3595 if (r && get_pool_mode(pool) == PM_FAIL)
3596 r = 0;
3597
3598 return r;
3599}
3600
3601static void pool_suspend_active_thins(struct pool *pool)
3602{
3603 struct thin_c *tc;
3604
3605 /* Suspend all active thin devices */
3606 tc = get_first_thin(pool);
3607 while (tc) {
3608 dm_internal_suspend_noflush(tc->thin_md);
3609 tc = get_next_thin(pool, tc);
3610 }
3611}
3612
3613static void pool_resume_active_thins(struct pool *pool)
3614{
3615 struct thin_c *tc;
3616
3617 /* Resume all active thin devices */
3618 tc = get_first_thin(pool);
3619 while (tc) {
3620 dm_internal_resume(tc->thin_md);
3621 tc = get_next_thin(pool, tc);
3622 }
3623}
3624
3625static void pool_resume(struct dm_target *ti)
3626{
3627 struct pool_c *pt = ti->private;
3628 struct pool *pool = pt->pool;
3629
3630 /*
3631 * Must requeue active_thins' bios and then resume
3632 * active_thins _before_ clearing 'suspend' flag.
3633 */
3634 requeue_bios(pool);
3635 pool_resume_active_thins(pool);
3636
3637 spin_lock_irq(&pool->lock);
3638 pool->low_water_triggered = false;
3639 pool->suspended = false;
3640 spin_unlock_irq(&pool->lock);
3641
3642 do_waker(&pool->waker.work);
3643}
3644
3645static void pool_presuspend(struct dm_target *ti)
3646{
3647 struct pool_c *pt = ti->private;
3648 struct pool *pool = pt->pool;
3649
3650 spin_lock_irq(&pool->lock);
3651 pool->suspended = true;
3652 spin_unlock_irq(&pool->lock);
3653
3654 pool_suspend_active_thins(pool);
3655}
3656
3657static void pool_presuspend_undo(struct dm_target *ti)
3658{
3659 struct pool_c *pt = ti->private;
3660 struct pool *pool = pt->pool;
3661
3662 pool_resume_active_thins(pool);
3663
3664 spin_lock_irq(&pool->lock);
3665 pool->suspended = false;
3666 spin_unlock_irq(&pool->lock);
3667}
3668
3669static void pool_postsuspend(struct dm_target *ti)
3670{
3671 struct pool_c *pt = ti->private;
3672 struct pool *pool = pt->pool;
3673
3674 cancel_delayed_work_sync(&pool->waker);
3675 cancel_delayed_work_sync(&pool->no_space_timeout);
3676 flush_workqueue(pool->wq);
3677 (void) commit(pool);
3678}
3679
3680static int check_arg_count(unsigned int argc, unsigned int args_required)
3681{
3682 if (argc != args_required) {
3683 DMWARN("Message received with %u arguments instead of %u.",
3684 argc, args_required);
3685 return -EINVAL;
3686 }
3687
3688 return 0;
3689}
3690
3691static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3692{
3693 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3694 *dev_id <= MAX_DEV_ID)
3695 return 0;
3696
3697 if (warning)
3698 DMWARN("Message received with invalid device id: %s", arg);
3699
3700 return -EINVAL;
3701}
3702
3703static int process_create_thin_mesg(unsigned int argc, char **argv, struct pool *pool)
3704{
3705 dm_thin_id dev_id;
3706 int r;
3707
3708 r = check_arg_count(argc, 2);
3709 if (r)
3710 return r;
3711
3712 r = read_dev_id(argv[1], &dev_id, 1);
3713 if (r)
3714 return r;
3715
3716 r = dm_pool_create_thin(pool->pmd, dev_id);
3717 if (r) {
3718 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3719 argv[1]);
3720 return r;
3721 }
3722
3723 return 0;
3724}
3725
3726static int process_create_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3727{
3728 dm_thin_id dev_id;
3729 dm_thin_id origin_dev_id;
3730 int r;
3731
3732 r = check_arg_count(argc, 3);
3733 if (r)
3734 return r;
3735
3736 r = read_dev_id(argv[1], &dev_id, 1);
3737 if (r)
3738 return r;
3739
3740 r = read_dev_id(argv[2], &origin_dev_id, 1);
3741 if (r)
3742 return r;
3743
3744 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3745 if (r) {
3746 DMWARN("Creation of new snapshot %s of device %s failed.",
3747 argv[1], argv[2]);
3748 return r;
3749 }
3750
3751 return 0;
3752}
3753
3754static int process_delete_mesg(unsigned int argc, char **argv, struct pool *pool)
3755{
3756 dm_thin_id dev_id;
3757 int r;
3758
3759 r = check_arg_count(argc, 2);
3760 if (r)
3761 return r;
3762
3763 r = read_dev_id(argv[1], &dev_id, 1);
3764 if (r)
3765 return r;
3766
3767 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3768 if (r)
3769 DMWARN("Deletion of thin device %s failed.", argv[1]);
3770
3771 return r;
3772}
3773
3774static int process_set_transaction_id_mesg(unsigned int argc, char **argv, struct pool *pool)
3775{
3776 dm_thin_id old_id, new_id;
3777 int r;
3778
3779 r = check_arg_count(argc, 3);
3780 if (r)
3781 return r;
3782
3783 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3784 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3785 return -EINVAL;
3786 }
3787
3788 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3789 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3790 return -EINVAL;
3791 }
3792
3793 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3794 if (r) {
3795 DMWARN("Failed to change transaction id from %s to %s.",
3796 argv[1], argv[2]);
3797 return r;
3798 }
3799
3800 return 0;
3801}
3802
3803static int process_reserve_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3804{
3805 int r;
3806
3807 r = check_arg_count(argc, 1);
3808 if (r)
3809 return r;
3810
3811 (void) commit(pool);
3812
3813 r = dm_pool_reserve_metadata_snap(pool->pmd);
3814 if (r)
3815 DMWARN("reserve_metadata_snap message failed.");
3816
3817 return r;
3818}
3819
3820static int process_release_metadata_snap_mesg(unsigned int argc, char **argv, struct pool *pool)
3821{
3822 int r;
3823
3824 r = check_arg_count(argc, 1);
3825 if (r)
3826 return r;
3827
3828 r = dm_pool_release_metadata_snap(pool->pmd);
3829 if (r)
3830 DMWARN("release_metadata_snap message failed.");
3831
3832 return r;
3833}
3834
3835/*
3836 * Messages supported:
3837 * create_thin <dev_id>
3838 * create_snap <dev_id> <origin_id>
3839 * delete <dev_id>
3840 * set_transaction_id <current_trans_id> <new_trans_id>
3841 * reserve_metadata_snap
3842 * release_metadata_snap
3843 */
3844static int pool_message(struct dm_target *ti, unsigned int argc, char **argv,
3845 char *result, unsigned int maxlen)
3846{
3847 int r = -EINVAL;
3848 struct pool_c *pt = ti->private;
3849 struct pool *pool = pt->pool;
3850
3851 if (get_pool_mode(pool) >= PM_OUT_OF_METADATA_SPACE) {
3852 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3853 dm_device_name(pool->pool_md));
3854 return -EOPNOTSUPP;
3855 }
3856
3857 if (!strcasecmp(argv[0], "create_thin"))
3858 r = process_create_thin_mesg(argc, argv, pool);
3859
3860 else if (!strcasecmp(argv[0], "create_snap"))
3861 r = process_create_snap_mesg(argc, argv, pool);
3862
3863 else if (!strcasecmp(argv[0], "delete"))
3864 r = process_delete_mesg(argc, argv, pool);
3865
3866 else if (!strcasecmp(argv[0], "set_transaction_id"))
3867 r = process_set_transaction_id_mesg(argc, argv, pool);
3868
3869 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3870 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3871
3872 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3873 r = process_release_metadata_snap_mesg(argc, argv, pool);
3874
3875 else
3876 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3877
3878 if (!r)
3879 (void) commit(pool);
3880
3881 return r;
3882}
3883
3884static void emit_flags(struct pool_features *pf, char *result,
3885 unsigned int sz, unsigned int maxlen)
3886{
3887 unsigned int count = !pf->zero_new_blocks + !pf->discard_enabled +
3888 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3889 pf->error_if_no_space;
3890 DMEMIT("%u ", count);
3891
3892 if (!pf->zero_new_blocks)
3893 DMEMIT("skip_block_zeroing ");
3894
3895 if (!pf->discard_enabled)
3896 DMEMIT("ignore_discard ");
3897
3898 if (!pf->discard_passdown)
3899 DMEMIT("no_discard_passdown ");
3900
3901 if (pf->mode == PM_READ_ONLY)
3902 DMEMIT("read_only ");
3903
3904 if (pf->error_if_no_space)
3905 DMEMIT("error_if_no_space ");
3906}
3907
3908/*
3909 * Status line is:
3910 * <transaction id> <used metadata sectors>/<total metadata sectors>
3911 * <used data sectors>/<total data sectors> <held metadata root>
3912 * <pool mode> <discard config> <no space config> <needs_check>
3913 */
3914static void pool_status(struct dm_target *ti, status_type_t type,
3915 unsigned int status_flags, char *result, unsigned int maxlen)
3916{
3917 int r;
3918 unsigned int sz = 0;
3919 uint64_t transaction_id;
3920 dm_block_t nr_free_blocks_data;
3921 dm_block_t nr_free_blocks_metadata;
3922 dm_block_t nr_blocks_data;
3923 dm_block_t nr_blocks_metadata;
3924 dm_block_t held_root;
3925 enum pool_mode mode;
3926 char buf[BDEVNAME_SIZE];
3927 char buf2[BDEVNAME_SIZE];
3928 struct pool_c *pt = ti->private;
3929 struct pool *pool = pt->pool;
3930
3931 switch (type) {
3932 case STATUSTYPE_INFO:
3933 if (get_pool_mode(pool) == PM_FAIL) {
3934 DMEMIT("Fail");
3935 break;
3936 }
3937
3938 /* Commit to ensure statistics aren't out-of-date */
3939 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3940 (void) commit(pool);
3941
3942 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3943 if (r) {
3944 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3945 dm_device_name(pool->pool_md), r);
3946 goto err;
3947 }
3948
3949 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3950 if (r) {
3951 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3952 dm_device_name(pool->pool_md), r);
3953 goto err;
3954 }
3955
3956 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3957 if (r) {
3958 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3959 dm_device_name(pool->pool_md), r);
3960 goto err;
3961 }
3962
3963 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3964 if (r) {
3965 DMERR("%s: dm_pool_get_free_block_count returned %d",
3966 dm_device_name(pool->pool_md), r);
3967 goto err;
3968 }
3969
3970 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3971 if (r) {
3972 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3973 dm_device_name(pool->pool_md), r);
3974 goto err;
3975 }
3976
3977 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3978 if (r) {
3979 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3980 dm_device_name(pool->pool_md), r);
3981 goto err;
3982 }
3983
3984 DMEMIT("%llu %llu/%llu %llu/%llu ",
3985 (unsigned long long)transaction_id,
3986 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3987 (unsigned long long)nr_blocks_metadata,
3988 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3989 (unsigned long long)nr_blocks_data);
3990
3991 if (held_root)
3992 DMEMIT("%llu ", held_root);
3993 else
3994 DMEMIT("- ");
3995
3996 mode = get_pool_mode(pool);
3997 if (mode == PM_OUT_OF_DATA_SPACE)
3998 DMEMIT("out_of_data_space ");
3999 else if (is_read_only_pool_mode(mode))
4000 DMEMIT("ro ");
4001 else
4002 DMEMIT("rw ");
4003
4004 if (!pool->pf.discard_enabled)
4005 DMEMIT("ignore_discard ");
4006 else if (pool->pf.discard_passdown)
4007 DMEMIT("discard_passdown ");
4008 else
4009 DMEMIT("no_discard_passdown ");
4010
4011 if (pool->pf.error_if_no_space)
4012 DMEMIT("error_if_no_space ");
4013 else
4014 DMEMIT("queue_if_no_space ");
4015
4016 if (dm_pool_metadata_needs_check(pool->pmd))
4017 DMEMIT("needs_check ");
4018 else
4019 DMEMIT("- ");
4020
4021 DMEMIT("%llu ", (unsigned long long)calc_metadata_threshold(pt));
4022
4023 break;
4024
4025 case STATUSTYPE_TABLE:
4026 DMEMIT("%s %s %lu %llu ",
4027 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
4028 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
4029 (unsigned long)pool->sectors_per_block,
4030 (unsigned long long)pt->low_water_blocks);
4031 emit_flags(&pt->requested_pf, result, sz, maxlen);
4032 break;
4033
4034 case STATUSTYPE_IMA:
4035 *result = '\0';
4036 break;
4037 }
4038 return;
4039
4040err:
4041 DMEMIT("Error");
4042}
4043
4044static int pool_iterate_devices(struct dm_target *ti,
4045 iterate_devices_callout_fn fn, void *data)
4046{
4047 struct pool_c *pt = ti->private;
4048
4049 return fn(ti, pt->data_dev, 0, ti->len, data);
4050}
4051
4052static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
4053{
4054 struct pool_c *pt = ti->private;
4055 struct pool *pool = pt->pool;
4056 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
4057
4058 /*
4059 * If max_sectors is smaller than pool->sectors_per_block adjust it
4060 * to the highest possible power-of-2 factor of pool->sectors_per_block.
4061 * This is especially beneficial when the pool's data device is a RAID
4062 * device that has a full stripe width that matches pool->sectors_per_block
4063 * -- because even though partial RAID stripe-sized IOs will be issued to a
4064 * single RAID stripe; when aggregated they will end on a full RAID stripe
4065 * boundary.. which avoids additional partial RAID stripe writes cascading
4066 */
4067 if (limits->max_sectors < pool->sectors_per_block) {
4068 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
4069 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
4070 limits->max_sectors--;
4071 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
4072 }
4073 }
4074
4075 /*
4076 * If the system-determined stacked limits are compatible with the
4077 * pool's blocksize (io_opt is a factor) do not override them.
4078 */
4079 if (io_opt_sectors < pool->sectors_per_block ||
4080 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
4081 if (is_factor(pool->sectors_per_block, limits->max_sectors))
4082 limits->io_min = limits->max_sectors << SECTOR_SHIFT;
4083 else
4084 limits->io_min = pool->sectors_per_block << SECTOR_SHIFT;
4085 limits->io_opt = pool->sectors_per_block << SECTOR_SHIFT;
4086 }
4087
4088 /*
4089 * pt->adjusted_pf is a staging area for the actual features to use.
4090 * They get transferred to the live pool in bind_control_target()
4091 * called from pool_preresume().
4092 */
4093
4094 if (pt->adjusted_pf.discard_enabled) {
4095 disable_discard_passdown_if_not_supported(pt);
4096 if (!pt->adjusted_pf.discard_passdown)
4097 limits->max_hw_discard_sectors = 0;
4098 /*
4099 * The pool uses the same discard limits as the underlying data
4100 * device. DM core has already set this up.
4101 */
4102 } else {
4103 /*
4104 * Must explicitly disallow stacking discard limits otherwise the
4105 * block layer will stack them if pool's data device has support.
4106 */
4107 limits->discard_granularity = 0;
4108 }
4109}
4110
4111static struct target_type pool_target = {
4112 .name = "thin-pool",
4113 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
4114 DM_TARGET_IMMUTABLE,
4115 .version = {1, 23, 0},
4116 .module = THIS_MODULE,
4117 .ctr = pool_ctr,
4118 .dtr = pool_dtr,
4119 .map = pool_map,
4120 .presuspend = pool_presuspend,
4121 .presuspend_undo = pool_presuspend_undo,
4122 .postsuspend = pool_postsuspend,
4123 .preresume = pool_preresume,
4124 .resume = pool_resume,
4125 .message = pool_message,
4126 .status = pool_status,
4127 .iterate_devices = pool_iterate_devices,
4128 .io_hints = pool_io_hints,
4129};
4130
4131/*
4132 *--------------------------------------------------------------
4133 * Thin target methods
4134 *--------------------------------------------------------------
4135 */
4136static void thin_get(struct thin_c *tc)
4137{
4138 refcount_inc(&tc->refcount);
4139}
4140
4141static void thin_put(struct thin_c *tc)
4142{
4143 if (refcount_dec_and_test(&tc->refcount))
4144 complete(&tc->can_destroy);
4145}
4146
4147static void thin_dtr(struct dm_target *ti)
4148{
4149 struct thin_c *tc = ti->private;
4150
4151 spin_lock_irq(&tc->pool->lock);
4152 list_del_rcu(&tc->list);
4153 spin_unlock_irq(&tc->pool->lock);
4154 synchronize_rcu();
4155
4156 thin_put(tc);
4157 wait_for_completion(&tc->can_destroy);
4158
4159 mutex_lock(&dm_thin_pool_table.mutex);
4160
4161 __pool_dec(tc->pool);
4162 dm_pool_close_thin_device(tc->td);
4163 dm_put_device(ti, tc->pool_dev);
4164 if (tc->origin_dev)
4165 dm_put_device(ti, tc->origin_dev);
4166 kfree(tc);
4167
4168 mutex_unlock(&dm_thin_pool_table.mutex);
4169}
4170
4171/*
4172 * Thin target parameters:
4173 *
4174 * <pool_dev> <dev_id> [origin_dev]
4175 *
4176 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4177 * dev_id: the internal device identifier
4178 * origin_dev: a device external to the pool that should act as the origin
4179 *
4180 * If the pool device has discards disabled, they get disabled for the thin
4181 * device as well.
4182 */
4183static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
4184{
4185 int r;
4186 struct thin_c *tc;
4187 struct dm_dev *pool_dev, *origin_dev;
4188 struct mapped_device *pool_md;
4189
4190 mutex_lock(&dm_thin_pool_table.mutex);
4191
4192 if (argc != 2 && argc != 3) {
4193 ti->error = "Invalid argument count";
4194 r = -EINVAL;
4195 goto out_unlock;
4196 }
4197
4198 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4199 if (!tc) {
4200 ti->error = "Out of memory";
4201 r = -ENOMEM;
4202 goto out_unlock;
4203 }
4204 tc->thin_md = dm_table_get_md(ti->table);
4205 spin_lock_init(&tc->lock);
4206 INIT_LIST_HEAD(&tc->deferred_cells);
4207 bio_list_init(&tc->deferred_bio_list);
4208 bio_list_init(&tc->retry_on_resume_list);
4209 tc->sort_bio_list = RB_ROOT;
4210
4211 if (argc == 3) {
4212 if (!strcmp(argv[0], argv[2])) {
4213 ti->error = "Error setting origin device";
4214 r = -EINVAL;
4215 goto bad_origin_dev;
4216 }
4217
4218 r = dm_get_device(ti, argv[2], BLK_OPEN_READ, &origin_dev);
4219 if (r) {
4220 ti->error = "Error opening origin device";
4221 goto bad_origin_dev;
4222 }
4223 tc->origin_dev = origin_dev;
4224 }
4225
4226 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4227 if (r) {
4228 ti->error = "Error opening pool device";
4229 goto bad_pool_dev;
4230 }
4231 tc->pool_dev = pool_dev;
4232
4233 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4234 ti->error = "Invalid device id";
4235 r = -EINVAL;
4236 goto bad_common;
4237 }
4238
4239 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4240 if (!pool_md) {
4241 ti->error = "Couldn't get pool mapped device";
4242 r = -EINVAL;
4243 goto bad_common;
4244 }
4245
4246 tc->pool = __pool_table_lookup(pool_md);
4247 if (!tc->pool) {
4248 ti->error = "Couldn't find pool object";
4249 r = -EINVAL;
4250 goto bad_pool_lookup;
4251 }
4252 __pool_inc(tc->pool);
4253
4254 if (get_pool_mode(tc->pool) == PM_FAIL) {
4255 ti->error = "Couldn't open thin device, Pool is in fail mode";
4256 r = -EINVAL;
4257 goto bad_pool;
4258 }
4259
4260 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4261 if (r) {
4262 ti->error = "Couldn't open thin internal device";
4263 goto bad_pool;
4264 }
4265
4266 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4267 if (r)
4268 goto bad;
4269
4270 ti->num_flush_bios = 1;
4271 ti->limit_swap_bios = true;
4272 ti->flush_supported = true;
4273 ti->accounts_remapped_io = true;
4274 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4275
4276 /* In case the pool supports discards, pass them on. */
4277 if (tc->pool->pf.discard_enabled) {
4278 ti->discards_supported = true;
4279 ti->num_discard_bios = 1;
4280 ti->max_discard_granularity = true;
4281 }
4282
4283 mutex_unlock(&dm_thin_pool_table.mutex);
4284
4285 spin_lock_irq(&tc->pool->lock);
4286 if (tc->pool->suspended) {
4287 spin_unlock_irq(&tc->pool->lock);
4288 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4289 ti->error = "Unable to activate thin device while pool is suspended";
4290 r = -EINVAL;
4291 goto bad;
4292 }
4293 refcount_set(&tc->refcount, 1);
4294 init_completion(&tc->can_destroy);
4295 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4296 spin_unlock_irq(&tc->pool->lock);
4297 /*
4298 * This synchronize_rcu() call is needed here otherwise we risk a
4299 * wake_worker() call finding no bios to process (because the newly
4300 * added tc isn't yet visible). So this reduces latency since we
4301 * aren't then dependent on the periodic commit to wake_worker().
4302 */
4303 synchronize_rcu();
4304
4305 dm_put(pool_md);
4306
4307 return 0;
4308
4309bad:
4310 dm_pool_close_thin_device(tc->td);
4311bad_pool:
4312 __pool_dec(tc->pool);
4313bad_pool_lookup:
4314 dm_put(pool_md);
4315bad_common:
4316 dm_put_device(ti, tc->pool_dev);
4317bad_pool_dev:
4318 if (tc->origin_dev)
4319 dm_put_device(ti, tc->origin_dev);
4320bad_origin_dev:
4321 kfree(tc);
4322out_unlock:
4323 mutex_unlock(&dm_thin_pool_table.mutex);
4324
4325 return r;
4326}
4327
4328static int thin_map(struct dm_target *ti, struct bio *bio)
4329{
4330 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4331
4332 return thin_bio_map(ti, bio);
4333}
4334
4335static int thin_endio(struct dm_target *ti, struct bio *bio,
4336 blk_status_t *err)
4337{
4338 unsigned long flags;
4339 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4340 struct list_head work;
4341 struct dm_thin_new_mapping *m, *tmp;
4342 struct pool *pool = h->tc->pool;
4343
4344 if (h->shared_read_entry) {
4345 INIT_LIST_HEAD(&work);
4346 dm_deferred_entry_dec(h->shared_read_entry, &work);
4347
4348 spin_lock_irqsave(&pool->lock, flags);
4349 list_for_each_entry_safe(m, tmp, &work, list) {
4350 list_del(&m->list);
4351 __complete_mapping_preparation(m);
4352 }
4353 spin_unlock_irqrestore(&pool->lock, flags);
4354 }
4355
4356 if (h->all_io_entry) {
4357 INIT_LIST_HEAD(&work);
4358 dm_deferred_entry_dec(h->all_io_entry, &work);
4359 if (!list_empty(&work)) {
4360 spin_lock_irqsave(&pool->lock, flags);
4361 list_for_each_entry_safe(m, tmp, &work, list)
4362 list_add_tail(&m->list, &pool->prepared_discards);
4363 spin_unlock_irqrestore(&pool->lock, flags);
4364 wake_worker(pool);
4365 }
4366 }
4367
4368 if (h->cell)
4369 cell_defer_no_holder(h->tc, h->cell);
4370
4371 return DM_ENDIO_DONE;
4372}
4373
4374static void thin_presuspend(struct dm_target *ti)
4375{
4376 struct thin_c *tc = ti->private;
4377
4378 if (dm_noflush_suspending(ti))
4379 noflush_work(tc, do_noflush_start);
4380}
4381
4382static void thin_postsuspend(struct dm_target *ti)
4383{
4384 struct thin_c *tc = ti->private;
4385
4386 /*
4387 * The dm_noflush_suspending flag has been cleared by now, so
4388 * unfortunately we must always run this.
4389 */
4390 noflush_work(tc, do_noflush_stop);
4391}
4392
4393static int thin_preresume(struct dm_target *ti)
4394{
4395 struct thin_c *tc = ti->private;
4396
4397 if (tc->origin_dev)
4398 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4399
4400 return 0;
4401}
4402
4403/*
4404 * <nr mapped sectors> <highest mapped sector>
4405 */
4406static void thin_status(struct dm_target *ti, status_type_t type,
4407 unsigned int status_flags, char *result, unsigned int maxlen)
4408{
4409 int r;
4410 ssize_t sz = 0;
4411 dm_block_t mapped, highest;
4412 char buf[BDEVNAME_SIZE];
4413 struct thin_c *tc = ti->private;
4414
4415 if (get_pool_mode(tc->pool) == PM_FAIL) {
4416 DMEMIT("Fail");
4417 return;
4418 }
4419
4420 if (!tc->td)
4421 DMEMIT("-");
4422 else {
4423 switch (type) {
4424 case STATUSTYPE_INFO:
4425 r = dm_thin_get_mapped_count(tc->td, &mapped);
4426 if (r) {
4427 DMERR("dm_thin_get_mapped_count returned %d", r);
4428 goto err;
4429 }
4430
4431 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4432 if (r < 0) {
4433 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4434 goto err;
4435 }
4436
4437 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4438 if (r)
4439 DMEMIT("%llu", ((highest + 1) *
4440 tc->pool->sectors_per_block) - 1);
4441 else
4442 DMEMIT("-");
4443 break;
4444
4445 case STATUSTYPE_TABLE:
4446 DMEMIT("%s %lu",
4447 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4448 (unsigned long) tc->dev_id);
4449 if (tc->origin_dev)
4450 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4451 break;
4452
4453 case STATUSTYPE_IMA:
4454 *result = '\0';
4455 break;
4456 }
4457 }
4458
4459 return;
4460
4461err:
4462 DMEMIT("Error");
4463}
4464
4465static int thin_iterate_devices(struct dm_target *ti,
4466 iterate_devices_callout_fn fn, void *data)
4467{
4468 sector_t blocks;
4469 struct thin_c *tc = ti->private;
4470 struct pool *pool = tc->pool;
4471
4472 /*
4473 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4474 * we follow a more convoluted path through to the pool's target.
4475 */
4476 if (!pool->ti)
4477 return 0; /* nothing is bound */
4478
4479 blocks = pool->ti->len;
4480 (void) sector_div(blocks, pool->sectors_per_block);
4481 if (blocks)
4482 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4483
4484 return 0;
4485}
4486
4487static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4488{
4489 struct thin_c *tc = ti->private;
4490 struct pool *pool = tc->pool;
4491
4492 if (pool->pf.discard_enabled) {
4493 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4494 limits->max_hw_discard_sectors = pool->sectors_per_block * BIO_PRISON_MAX_RANGE;
4495 }
4496}
4497
4498static struct target_type thin_target = {
4499 .name = "thin",
4500 .version = {1, 23, 0},
4501 .module = THIS_MODULE,
4502 .ctr = thin_ctr,
4503 .dtr = thin_dtr,
4504 .map = thin_map,
4505 .end_io = thin_endio,
4506 .preresume = thin_preresume,
4507 .presuspend = thin_presuspend,
4508 .postsuspend = thin_postsuspend,
4509 .status = thin_status,
4510 .iterate_devices = thin_iterate_devices,
4511 .io_hints = thin_io_hints,
4512};
4513
4514/*----------------------------------------------------------------*/
4515
4516static int __init dm_thin_init(void)
4517{
4518 int r = -ENOMEM;
4519
4520 pool_table_init();
4521
4522 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4523 if (!_new_mapping_cache)
4524 return r;
4525
4526 r = dm_register_target(&thin_target);
4527 if (r)
4528 goto bad_new_mapping_cache;
4529
4530 r = dm_register_target(&pool_target);
4531 if (r)
4532 goto bad_thin_target;
4533
4534 return 0;
4535
4536bad_thin_target:
4537 dm_unregister_target(&thin_target);
4538bad_new_mapping_cache:
4539 kmem_cache_destroy(_new_mapping_cache);
4540
4541 return r;
4542}
4543
4544static void dm_thin_exit(void)
4545{
4546 dm_unregister_target(&thin_target);
4547 dm_unregister_target(&pool_target);
4548
4549 kmem_cache_destroy(_new_mapping_cache);
4550
4551 pool_table_exit();
4552}
4553
4554module_init(dm_thin_init);
4555module_exit(dm_thin_exit);
4556
4557module_param_named(no_space_timeout, no_space_timeout_secs, uint, 0644);
4558MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4559
4560MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4561MODULE_AUTHOR("Joe Thornber <dm-devel@lists.linux.dev>");
4562MODULE_LICENSE("GPL");