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