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