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.h"
9#include "dm.h"
10
11#include <linux/device-mapper.h>
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
14#include <linux/jiffies.h>
15#include <linux/log2.h>
16#include <linux/list.h>
17#include <linux/rculist.h>
18#include <linux/init.h>
19#include <linux/module.h>
20#include <linux/slab.h>
21#include <linux/vmalloc.h>
22#include <linux/sort.h>
23#include <linux/rbtree.h>
24
25#define DM_MSG_PREFIX "thin"
26
27/*
28 * Tunable constants
29 */
30#define ENDIO_HOOK_POOL_SIZE 1024
31#define MAPPING_POOL_SIZE 1024
32#define COMMIT_PERIOD HZ
33#define NO_SPACE_TIMEOUT_SECS 60
34
35static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36
37DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
38 "A percentage of time allocated for copy on write");
39
40/*
41 * The block size of the device holding pool data must be
42 * between 64KB and 1GB.
43 */
44#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
45#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
46
47/*
48 * Device id is restricted to 24 bits.
49 */
50#define MAX_DEV_ID ((1 << 24) - 1)
51
52/*
53 * How do we handle breaking sharing of data blocks?
54 * =================================================
55 *
56 * We use a standard copy-on-write btree to store the mappings for the
57 * devices (note I'm talking about copy-on-write of the metadata here, not
58 * the data). When you take an internal snapshot you clone the root node
59 * of the origin btree. After this there is no concept of an origin or a
60 * snapshot. They are just two device trees that happen to point to the
61 * same data blocks.
62 *
63 * When we get a write in we decide if it's to a shared data block using
64 * some timestamp magic. If it is, we have to break sharing.
65 *
66 * Let's say we write to a shared block in what was the origin. The
67 * steps are:
68 *
69 * i) plug io further to this physical block. (see bio_prison code).
70 *
71 * ii) quiesce any read io to that shared data block. Obviously
72 * including all devices that share this block. (see dm_deferred_set code)
73 *
74 * iii) copy the data block to a newly allocate block. This step can be
75 * missed out if the io covers the block. (schedule_copy).
76 *
77 * iv) insert the new mapping into the origin's btree
78 * (process_prepared_mapping). This act of inserting breaks some
79 * sharing of btree nodes between the two devices. Breaking sharing only
80 * effects the btree of that specific device. Btrees for the other
81 * devices that share the block never change. The btree for the origin
82 * device as it was after the last commit is untouched, ie. we're using
83 * persistent data structures in the functional programming sense.
84 *
85 * v) unplug io to this physical block, including the io that triggered
86 * the breaking of sharing.
87 *
88 * Steps (ii) and (iii) occur in parallel.
89 *
90 * The metadata _doesn't_ need to be committed before the io continues. We
91 * get away with this because the io is always written to a _new_ block.
92 * If there's a crash, then:
93 *
94 * - The origin mapping will point to the old origin block (the shared
95 * one). This will contain the data as it was before the io that triggered
96 * the breaking of sharing came in.
97 *
98 * - The snap mapping still points to the old block. As it would after
99 * the commit.
100 *
101 * The downside of this scheme is the timestamp magic isn't perfect, and
102 * will continue to think that data block in the snapshot device is shared
103 * even after the write to the origin has broken sharing. I suspect data
104 * blocks will typically be shared by many different devices, so we're
105 * breaking sharing n + 1 times, rather than n, where n is the number of
106 * devices that reference this data block. At the moment I think the
107 * benefits far, far outweigh the disadvantages.
108 */
109
110/*----------------------------------------------------------------*/
111
112/*
113 * Key building.
114 */
115enum lock_space {
116 VIRTUAL,
117 PHYSICAL
118};
119
120static void build_key(struct dm_thin_device *td, enum lock_space ls,
121 dm_block_t b, dm_block_t e, struct dm_cell_key *key)
122{
123 key->virtual = (ls == VIRTUAL);
124 key->dev = dm_thin_dev_id(td);
125 key->block_begin = b;
126 key->block_end = e;
127}
128
129static void build_data_key(struct dm_thin_device *td, dm_block_t b,
130 struct dm_cell_key *key)
131{
132 build_key(td, PHYSICAL, b, b + 1llu, key);
133}
134
135static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
136 struct dm_cell_key *key)
137{
138 build_key(td, VIRTUAL, b, b + 1llu, key);
139}
140
141/*----------------------------------------------------------------*/
142
143#define THROTTLE_THRESHOLD (1 * HZ)
144
145struct throttle {
146 struct rw_semaphore lock;
147 unsigned long threshold;
148 bool throttle_applied;
149};
150
151static void throttle_init(struct throttle *t)
152{
153 init_rwsem(&t->lock);
154 t->throttle_applied = false;
155}
156
157static void throttle_work_start(struct throttle *t)
158{
159 t->threshold = jiffies + THROTTLE_THRESHOLD;
160}
161
162static void throttle_work_update(struct throttle *t)
163{
164 if (!t->throttle_applied && jiffies > t->threshold) {
165 down_write(&t->lock);
166 t->throttle_applied = true;
167 }
168}
169
170static void throttle_work_complete(struct throttle *t)
171{
172 if (t->throttle_applied) {
173 t->throttle_applied = false;
174 up_write(&t->lock);
175 }
176}
177
178static void throttle_lock(struct throttle *t)
179{
180 down_read(&t->lock);
181}
182
183static void throttle_unlock(struct throttle *t)
184{
185 up_read(&t->lock);
186}
187
188/*----------------------------------------------------------------*/
189
190/*
191 * A pool device ties together a metadata device and a data device. It
192 * also provides the interface for creating and destroying internal
193 * devices.
194 */
195struct dm_thin_new_mapping;
196
197/*
198 * The pool runs in 4 modes. Ordered in degraded order for comparisons.
199 */
200enum pool_mode {
201 PM_WRITE, /* metadata may be changed */
202 PM_OUT_OF_DATA_SPACE, /* metadata may be changed, though data may not be allocated */
203 PM_READ_ONLY, /* metadata may not be changed */
204 PM_FAIL, /* all I/O fails */
205};
206
207struct pool_features {
208 enum pool_mode mode;
209
210 bool zero_new_blocks:1;
211 bool discard_enabled:1;
212 bool discard_passdown:1;
213 bool error_if_no_space:1;
214};
215
216struct thin_c;
217typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
218typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
219typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
220
221#define CELL_SORT_ARRAY_SIZE 8192
222
223struct pool {
224 struct list_head list;
225 struct dm_target *ti; /* Only set if a pool target is bound */
226
227 struct mapped_device *pool_md;
228 struct block_device *md_dev;
229 struct dm_pool_metadata *pmd;
230
231 dm_block_t low_water_blocks;
232 uint32_t sectors_per_block;
233 int sectors_per_block_shift;
234
235 struct pool_features pf;
236 bool low_water_triggered:1; /* A dm event has been sent */
237 bool suspended:1;
238 bool out_of_data_space:1;
239
240 struct dm_bio_prison *prison;
241 struct dm_kcopyd_client *copier;
242
243 struct workqueue_struct *wq;
244 struct throttle throttle;
245 struct work_struct worker;
246 struct delayed_work waker;
247 struct delayed_work no_space_timeout;
248
249 unsigned long last_commit_jiffies;
250 unsigned ref_count;
251
252 spinlock_t lock;
253 struct bio_list deferred_flush_bios;
254 struct list_head prepared_mappings;
255 struct list_head prepared_discards;
256 struct list_head active_thins;
257
258 struct dm_deferred_set *shared_read_ds;
259 struct dm_deferred_set *all_io_ds;
260
261 struct dm_thin_new_mapping *next_mapping;
262 mempool_t *mapping_pool;
263
264 process_bio_fn process_bio;
265 process_bio_fn process_discard;
266
267 process_cell_fn process_cell;
268 process_cell_fn process_discard_cell;
269
270 process_mapping_fn process_prepared_mapping;
271 process_mapping_fn process_prepared_discard;
272
273 struct dm_bio_prison_cell **cell_sort_array;
274};
275
276static enum pool_mode get_pool_mode(struct pool *pool);
277static void metadata_operation_failed(struct pool *pool, const char *op, int r);
278
279/*
280 * Target context for a pool.
281 */
282struct pool_c {
283 struct dm_target *ti;
284 struct pool *pool;
285 struct dm_dev *data_dev;
286 struct dm_dev *metadata_dev;
287 struct dm_target_callbacks callbacks;
288
289 dm_block_t low_water_blocks;
290 struct pool_features requested_pf; /* Features requested during table load */
291 struct pool_features adjusted_pf; /* Features used after adjusting for constituent devices */
292};
293
294/*
295 * Target context for a thin.
296 */
297struct thin_c {
298 struct list_head list;
299 struct dm_dev *pool_dev;
300 struct dm_dev *origin_dev;
301 sector_t origin_size;
302 dm_thin_id dev_id;
303
304 struct pool *pool;
305 struct dm_thin_device *td;
306 struct mapped_device *thin_md;
307
308 bool requeue_mode:1;
309 spinlock_t lock;
310 struct list_head deferred_cells;
311 struct bio_list deferred_bio_list;
312 struct bio_list retry_on_resume_list;
313 struct rb_root sort_bio_list; /* sorted list of deferred bios */
314
315 /*
316 * Ensures the thin is not destroyed until the worker has finished
317 * iterating the active_thins list.
318 */
319 atomic_t refcount;
320 struct completion can_destroy;
321};
322
323/*----------------------------------------------------------------*/
324
325/**
326 * __blkdev_issue_discard_async - queue a discard with async completion
327 * @bdev: blockdev to issue discard for
328 * @sector: start sector
329 * @nr_sects: number of sectors to discard
330 * @gfp_mask: memory allocation flags (for bio_alloc)
331 * @flags: BLKDEV_IFL_* flags to control behaviour
332 * @parent_bio: parent discard bio that all sub discards get chained to
333 *
334 * Description:
335 * Asynchronously issue a discard request for the sectors in question.
336 */
337static int __blkdev_issue_discard_async(struct block_device *bdev, sector_t sector,
338 sector_t nr_sects, gfp_t gfp_mask, unsigned long flags,
339 struct bio *parent_bio)
340{
341 struct request_queue *q = bdev_get_queue(bdev);
342 int type = REQ_WRITE | REQ_DISCARD;
343 struct bio *bio;
344
345 if (!q || !nr_sects)
346 return -ENXIO;
347
348 if (!blk_queue_discard(q))
349 return -EOPNOTSUPP;
350
351 if (flags & BLKDEV_DISCARD_SECURE) {
352 if (!blk_queue_secdiscard(q))
353 return -EOPNOTSUPP;
354 type |= REQ_SECURE;
355 }
356
357 /*
358 * Required bio_put occurs in bio_endio thanks to bio_chain below
359 */
360 bio = bio_alloc(gfp_mask, 1);
361 if (!bio)
362 return -ENOMEM;
363
364 bio_chain(bio, parent_bio);
365
366 bio->bi_iter.bi_sector = sector;
367 bio->bi_bdev = bdev;
368 bio->bi_iter.bi_size = nr_sects << 9;
369
370 submit_bio(type, bio);
371
372 return 0;
373}
374
375static bool block_size_is_power_of_two(struct pool *pool)
376{
377 return pool->sectors_per_block_shift >= 0;
378}
379
380static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
381{
382 return block_size_is_power_of_two(pool) ?
383 (b << pool->sectors_per_block_shift) :
384 (b * pool->sectors_per_block);
385}
386
387static int issue_discard(struct thin_c *tc, dm_block_t data_b, dm_block_t data_e,
388 struct bio *parent_bio)
389{
390 sector_t s = block_to_sectors(tc->pool, data_b);
391 sector_t len = block_to_sectors(tc->pool, data_e - data_b);
392
393 return __blkdev_issue_discard_async(tc->pool_dev->bdev, s, len,
394 GFP_NOWAIT, 0, parent_bio);
395}
396
397/*----------------------------------------------------------------*/
398
399/*
400 * wake_worker() is used when new work is queued and when pool_resume is
401 * ready to continue deferred IO processing.
402 */
403static void wake_worker(struct pool *pool)
404{
405 queue_work(pool->wq, &pool->worker);
406}
407
408/*----------------------------------------------------------------*/
409
410static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
411 struct dm_bio_prison_cell **cell_result)
412{
413 int r;
414 struct dm_bio_prison_cell *cell_prealloc;
415
416 /*
417 * Allocate a cell from the prison's mempool.
418 * This might block but it can't fail.
419 */
420 cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
421
422 r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
423 if (r)
424 /*
425 * We reused an old cell; we can get rid of
426 * the new one.
427 */
428 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
429
430 return r;
431}
432
433static void cell_release(struct pool *pool,
434 struct dm_bio_prison_cell *cell,
435 struct bio_list *bios)
436{
437 dm_cell_release(pool->prison, cell, bios);
438 dm_bio_prison_free_cell(pool->prison, cell);
439}
440
441static void cell_visit_release(struct pool *pool,
442 void (*fn)(void *, struct dm_bio_prison_cell *),
443 void *context,
444 struct dm_bio_prison_cell *cell)
445{
446 dm_cell_visit_release(pool->prison, fn, context, cell);
447 dm_bio_prison_free_cell(pool->prison, cell);
448}
449
450static void cell_release_no_holder(struct pool *pool,
451 struct dm_bio_prison_cell *cell,
452 struct bio_list *bios)
453{
454 dm_cell_release_no_holder(pool->prison, cell, bios);
455 dm_bio_prison_free_cell(pool->prison, cell);
456}
457
458static void cell_error_with_code(struct pool *pool,
459 struct dm_bio_prison_cell *cell, int error_code)
460{
461 dm_cell_error(pool->prison, cell, error_code);
462 dm_bio_prison_free_cell(pool->prison, cell);
463}
464
465static int get_pool_io_error_code(struct pool *pool)
466{
467 return pool->out_of_data_space ? -ENOSPC : -EIO;
468}
469
470static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
471{
472 int error = get_pool_io_error_code(pool);
473
474 cell_error_with_code(pool, cell, error);
475}
476
477static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
478{
479 cell_error_with_code(pool, cell, 0);
480}
481
482static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
483{
484 cell_error_with_code(pool, cell, DM_ENDIO_REQUEUE);
485}
486
487/*----------------------------------------------------------------*/
488
489/*
490 * A global list of pools that uses a struct mapped_device as a key.
491 */
492static struct dm_thin_pool_table {
493 struct mutex mutex;
494 struct list_head pools;
495} dm_thin_pool_table;
496
497static void pool_table_init(void)
498{
499 mutex_init(&dm_thin_pool_table.mutex);
500 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
501}
502
503static void __pool_table_insert(struct pool *pool)
504{
505 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
506 list_add(&pool->list, &dm_thin_pool_table.pools);
507}
508
509static void __pool_table_remove(struct pool *pool)
510{
511 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
512 list_del(&pool->list);
513}
514
515static struct pool *__pool_table_lookup(struct mapped_device *md)
516{
517 struct pool *pool = NULL, *tmp;
518
519 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
520
521 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
522 if (tmp->pool_md == md) {
523 pool = tmp;
524 break;
525 }
526 }
527
528 return pool;
529}
530
531static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
532{
533 struct pool *pool = NULL, *tmp;
534
535 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
536
537 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
538 if (tmp->md_dev == md_dev) {
539 pool = tmp;
540 break;
541 }
542 }
543
544 return pool;
545}
546
547/*----------------------------------------------------------------*/
548
549struct dm_thin_endio_hook {
550 struct thin_c *tc;
551 struct dm_deferred_entry *shared_read_entry;
552 struct dm_deferred_entry *all_io_entry;
553 struct dm_thin_new_mapping *overwrite_mapping;
554 struct rb_node rb_node;
555 struct dm_bio_prison_cell *cell;
556};
557
558static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
559{
560 bio_list_merge(bios, master);
561 bio_list_init(master);
562}
563
564static void error_bio_list(struct bio_list *bios, int error)
565{
566 struct bio *bio;
567
568 while ((bio = bio_list_pop(bios))) {
569 bio->bi_error = error;
570 bio_endio(bio);
571 }
572}
573
574static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master, int error)
575{
576 struct bio_list bios;
577 unsigned long flags;
578
579 bio_list_init(&bios);
580
581 spin_lock_irqsave(&tc->lock, flags);
582 __merge_bio_list(&bios, master);
583 spin_unlock_irqrestore(&tc->lock, flags);
584
585 error_bio_list(&bios, error);
586}
587
588static void requeue_deferred_cells(struct thin_c *tc)
589{
590 struct pool *pool = tc->pool;
591 unsigned long flags;
592 struct list_head cells;
593 struct dm_bio_prison_cell *cell, *tmp;
594
595 INIT_LIST_HEAD(&cells);
596
597 spin_lock_irqsave(&tc->lock, flags);
598 list_splice_init(&tc->deferred_cells, &cells);
599 spin_unlock_irqrestore(&tc->lock, flags);
600
601 list_for_each_entry_safe(cell, tmp, &cells, user_list)
602 cell_requeue(pool, cell);
603}
604
605static void requeue_io(struct thin_c *tc)
606{
607 struct bio_list bios;
608 unsigned long flags;
609
610 bio_list_init(&bios);
611
612 spin_lock_irqsave(&tc->lock, flags);
613 __merge_bio_list(&bios, &tc->deferred_bio_list);
614 __merge_bio_list(&bios, &tc->retry_on_resume_list);
615 spin_unlock_irqrestore(&tc->lock, flags);
616
617 error_bio_list(&bios, DM_ENDIO_REQUEUE);
618 requeue_deferred_cells(tc);
619}
620
621static void error_retry_list_with_code(struct pool *pool, int error)
622{
623 struct thin_c *tc;
624
625 rcu_read_lock();
626 list_for_each_entry_rcu(tc, &pool->active_thins, list)
627 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
628 rcu_read_unlock();
629}
630
631static void error_retry_list(struct pool *pool)
632{
633 int error = get_pool_io_error_code(pool);
634
635 return error_retry_list_with_code(pool, error);
636}
637
638/*
639 * This section of code contains the logic for processing a thin device's IO.
640 * Much of the code depends on pool object resources (lists, workqueues, etc)
641 * but most is exclusively called from the thin target rather than the thin-pool
642 * target.
643 */
644
645static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
646{
647 struct pool *pool = tc->pool;
648 sector_t block_nr = bio->bi_iter.bi_sector;
649
650 if (block_size_is_power_of_two(pool))
651 block_nr >>= pool->sectors_per_block_shift;
652 else
653 (void) sector_div(block_nr, pool->sectors_per_block);
654
655 return block_nr;
656}
657
658/*
659 * Returns the _complete_ blocks that this bio covers.
660 */
661static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
662 dm_block_t *begin, dm_block_t *end)
663{
664 struct pool *pool = tc->pool;
665 sector_t b = bio->bi_iter.bi_sector;
666 sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
667
668 b += pool->sectors_per_block - 1ull; /* so we round up */
669
670 if (block_size_is_power_of_two(pool)) {
671 b >>= pool->sectors_per_block_shift;
672 e >>= pool->sectors_per_block_shift;
673 } else {
674 (void) sector_div(b, pool->sectors_per_block);
675 (void) sector_div(e, pool->sectors_per_block);
676 }
677
678 if (e < b)
679 /* Can happen if the bio is within a single block. */
680 e = b;
681
682 *begin = b;
683 *end = e;
684}
685
686static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
687{
688 struct pool *pool = tc->pool;
689 sector_t bi_sector = bio->bi_iter.bi_sector;
690
691 bio->bi_bdev = tc->pool_dev->bdev;
692 if (block_size_is_power_of_two(pool))
693 bio->bi_iter.bi_sector =
694 (block << pool->sectors_per_block_shift) |
695 (bi_sector & (pool->sectors_per_block - 1));
696 else
697 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
698 sector_div(bi_sector, pool->sectors_per_block);
699}
700
701static void remap_to_origin(struct thin_c *tc, struct bio *bio)
702{
703 bio->bi_bdev = tc->origin_dev->bdev;
704}
705
706static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
707{
708 return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
709 dm_thin_changed_this_transaction(tc->td);
710}
711
712static void inc_all_io_entry(struct pool *pool, struct bio *bio)
713{
714 struct dm_thin_endio_hook *h;
715
716 if (bio->bi_rw & REQ_DISCARD)
717 return;
718
719 h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
720 h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
721}
722
723static void issue(struct thin_c *tc, struct bio *bio)
724{
725 struct pool *pool = tc->pool;
726 unsigned long flags;
727
728 if (!bio_triggers_commit(tc, bio)) {
729 generic_make_request(bio);
730 return;
731 }
732
733 /*
734 * Complete bio with an error if earlier I/O caused changes to
735 * the metadata that can't be committed e.g, due to I/O errors
736 * on the metadata device.
737 */
738 if (dm_thin_aborted_changes(tc->td)) {
739 bio_io_error(bio);
740 return;
741 }
742
743 /*
744 * Batch together any bios that trigger commits and then issue a
745 * single commit for them in process_deferred_bios().
746 */
747 spin_lock_irqsave(&pool->lock, flags);
748 bio_list_add(&pool->deferred_flush_bios, bio);
749 spin_unlock_irqrestore(&pool->lock, flags);
750}
751
752static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
753{
754 remap_to_origin(tc, bio);
755 issue(tc, bio);
756}
757
758static void remap_and_issue(struct thin_c *tc, struct bio *bio,
759 dm_block_t block)
760{
761 remap(tc, bio, block);
762 issue(tc, bio);
763}
764
765/*----------------------------------------------------------------*/
766
767/*
768 * Bio endio functions.
769 */
770struct dm_thin_new_mapping {
771 struct list_head list;
772
773 bool pass_discard:1;
774 bool maybe_shared:1;
775
776 /*
777 * Track quiescing, copying and zeroing preparation actions. When this
778 * counter hits zero the block is prepared and can be inserted into the
779 * btree.
780 */
781 atomic_t prepare_actions;
782
783 int err;
784 struct thin_c *tc;
785 dm_block_t virt_begin, virt_end;
786 dm_block_t data_block;
787 struct dm_bio_prison_cell *cell;
788
789 /*
790 * If the bio covers the whole area of a block then we can avoid
791 * zeroing or copying. Instead this bio is hooked. The bio will
792 * still be in the cell, so care has to be taken to avoid issuing
793 * the bio twice.
794 */
795 struct bio *bio;
796 bio_end_io_t *saved_bi_end_io;
797};
798
799static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
800{
801 struct pool *pool = m->tc->pool;
802
803 if (atomic_dec_and_test(&m->prepare_actions)) {
804 list_add_tail(&m->list, &pool->prepared_mappings);
805 wake_worker(pool);
806 }
807}
808
809static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
810{
811 unsigned long flags;
812 struct pool *pool = m->tc->pool;
813
814 spin_lock_irqsave(&pool->lock, flags);
815 __complete_mapping_preparation(m);
816 spin_unlock_irqrestore(&pool->lock, flags);
817}
818
819static void copy_complete(int read_err, unsigned long write_err, void *context)
820{
821 struct dm_thin_new_mapping *m = context;
822
823 m->err = read_err || write_err ? -EIO : 0;
824 complete_mapping_preparation(m);
825}
826
827static void overwrite_endio(struct bio *bio)
828{
829 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
830 struct dm_thin_new_mapping *m = h->overwrite_mapping;
831
832 bio->bi_end_io = m->saved_bi_end_io;
833
834 m->err = bio->bi_error;
835 complete_mapping_preparation(m);
836}
837
838/*----------------------------------------------------------------*/
839
840/*
841 * Workqueue.
842 */
843
844/*
845 * Prepared mapping jobs.
846 */
847
848/*
849 * This sends the bios in the cell, except the original holder, back
850 * to the deferred_bios list.
851 */
852static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
853{
854 struct pool *pool = tc->pool;
855 unsigned long flags;
856
857 spin_lock_irqsave(&tc->lock, flags);
858 cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
859 spin_unlock_irqrestore(&tc->lock, flags);
860
861 wake_worker(pool);
862}
863
864static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
865
866struct remap_info {
867 struct thin_c *tc;
868 struct bio_list defer_bios;
869 struct bio_list issue_bios;
870};
871
872static void __inc_remap_and_issue_cell(void *context,
873 struct dm_bio_prison_cell *cell)
874{
875 struct remap_info *info = context;
876 struct bio *bio;
877
878 while ((bio = bio_list_pop(&cell->bios))) {
879 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA))
880 bio_list_add(&info->defer_bios, bio);
881 else {
882 inc_all_io_entry(info->tc->pool, bio);
883
884 /*
885 * We can't issue the bios with the bio prison lock
886 * held, so we add them to a list to issue on
887 * return from this function.
888 */
889 bio_list_add(&info->issue_bios, bio);
890 }
891 }
892}
893
894static void inc_remap_and_issue_cell(struct thin_c *tc,
895 struct dm_bio_prison_cell *cell,
896 dm_block_t block)
897{
898 struct bio *bio;
899 struct remap_info info;
900
901 info.tc = tc;
902 bio_list_init(&info.defer_bios);
903 bio_list_init(&info.issue_bios);
904
905 /*
906 * We have to be careful to inc any bios we're about to issue
907 * before the cell is released, and avoid a race with new bios
908 * being added to the cell.
909 */
910 cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
911 &info, cell);
912
913 while ((bio = bio_list_pop(&info.defer_bios)))
914 thin_defer_bio(tc, bio);
915
916 while ((bio = bio_list_pop(&info.issue_bios)))
917 remap_and_issue(info.tc, bio, block);
918}
919
920static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
921{
922 cell_error(m->tc->pool, m->cell);
923 list_del(&m->list);
924 mempool_free(m, m->tc->pool->mapping_pool);
925}
926
927static void process_prepared_mapping(struct dm_thin_new_mapping *m)
928{
929 struct thin_c *tc = m->tc;
930 struct pool *pool = tc->pool;
931 struct bio *bio = m->bio;
932 int r;
933
934 if (m->err) {
935 cell_error(pool, m->cell);
936 goto out;
937 }
938
939 /*
940 * Commit the prepared block into the mapping btree.
941 * Any I/O for this block arriving after this point will get
942 * remapped to it directly.
943 */
944 r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
945 if (r) {
946 metadata_operation_failed(pool, "dm_thin_insert_block", r);
947 cell_error(pool, m->cell);
948 goto out;
949 }
950
951 /*
952 * Release any bios held while the block was being provisioned.
953 * If we are processing a write bio that completely covers the block,
954 * we already processed it so can ignore it now when processing
955 * the bios in the cell.
956 */
957 if (bio) {
958 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
959 bio_endio(bio);
960 } else {
961 inc_all_io_entry(tc->pool, m->cell->holder);
962 remap_and_issue(tc, m->cell->holder, m->data_block);
963 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
964 }
965
966out:
967 list_del(&m->list);
968 mempool_free(m, pool->mapping_pool);
969}
970
971/*----------------------------------------------------------------*/
972
973static void free_discard_mapping(struct dm_thin_new_mapping *m)
974{
975 struct thin_c *tc = m->tc;
976 if (m->cell)
977 cell_defer_no_holder(tc, m->cell);
978 mempool_free(m, tc->pool->mapping_pool);
979}
980
981static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
982{
983 bio_io_error(m->bio);
984 free_discard_mapping(m);
985}
986
987static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
988{
989 bio_endio(m->bio);
990 free_discard_mapping(m);
991}
992
993static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
994{
995 int r;
996 struct thin_c *tc = m->tc;
997
998 r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
999 if (r) {
1000 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
1001 bio_io_error(m->bio);
1002 } else
1003 bio_endio(m->bio);
1004
1005 cell_defer_no_holder(tc, m->cell);
1006 mempool_free(m, tc->pool->mapping_pool);
1007}
1008
1009static int passdown_double_checking_shared_status(struct dm_thin_new_mapping *m)
1010{
1011 /*
1012 * We've already unmapped this range of blocks, but before we
1013 * passdown we have to check that these blocks are now unused.
1014 */
1015 int r;
1016 bool used = true;
1017 struct thin_c *tc = m->tc;
1018 struct pool *pool = tc->pool;
1019 dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1020
1021 while (b != end) {
1022 /* find start of unmapped run */
1023 for (; b < end; b++) {
1024 r = dm_pool_block_is_used(pool->pmd, b, &used);
1025 if (r)
1026 return r;
1027
1028 if (!used)
1029 break;
1030 }
1031
1032 if (b == end)
1033 break;
1034
1035 /* find end of run */
1036 for (e = b + 1; e != end; e++) {
1037 r = dm_pool_block_is_used(pool->pmd, e, &used);
1038 if (r)
1039 return r;
1040
1041 if (used)
1042 break;
1043 }
1044
1045 r = issue_discard(tc, b, e, m->bio);
1046 if (r)
1047 return r;
1048
1049 b = e;
1050 }
1051
1052 return 0;
1053}
1054
1055static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
1056{
1057 int r;
1058 struct thin_c *tc = m->tc;
1059 struct pool *pool = tc->pool;
1060
1061 r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1062 if (r)
1063 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1064
1065 else if (m->maybe_shared)
1066 r = passdown_double_checking_shared_status(m);
1067 else
1068 r = issue_discard(tc, m->data_block, m->data_block + (m->virt_end - m->virt_begin), m->bio);
1069
1070 /*
1071 * Even if r is set, there could be sub discards in flight that we
1072 * need to wait for.
1073 */
1074 m->bio->bi_error = r;
1075 bio_endio(m->bio);
1076 cell_defer_no_holder(tc, m->cell);
1077 mempool_free(m, pool->mapping_pool);
1078}
1079
1080static void process_prepared(struct pool *pool, struct list_head *head,
1081 process_mapping_fn *fn)
1082{
1083 unsigned long flags;
1084 struct list_head maps;
1085 struct dm_thin_new_mapping *m, *tmp;
1086
1087 INIT_LIST_HEAD(&maps);
1088 spin_lock_irqsave(&pool->lock, flags);
1089 list_splice_init(head, &maps);
1090 spin_unlock_irqrestore(&pool->lock, flags);
1091
1092 list_for_each_entry_safe(m, tmp, &maps, list)
1093 (*fn)(m);
1094}
1095
1096/*
1097 * Deferred bio jobs.
1098 */
1099static int io_overlaps_block(struct pool *pool, struct bio *bio)
1100{
1101 return bio->bi_iter.bi_size ==
1102 (pool->sectors_per_block << SECTOR_SHIFT);
1103}
1104
1105static int io_overwrites_block(struct pool *pool, struct bio *bio)
1106{
1107 return (bio_data_dir(bio) == WRITE) &&
1108 io_overlaps_block(pool, bio);
1109}
1110
1111static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1112 bio_end_io_t *fn)
1113{
1114 *save = bio->bi_end_io;
1115 bio->bi_end_io = fn;
1116}
1117
1118static int ensure_next_mapping(struct pool *pool)
1119{
1120 if (pool->next_mapping)
1121 return 0;
1122
1123 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
1124
1125 return pool->next_mapping ? 0 : -ENOMEM;
1126}
1127
1128static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1129{
1130 struct dm_thin_new_mapping *m = pool->next_mapping;
1131
1132 BUG_ON(!pool->next_mapping);
1133
1134 memset(m, 0, sizeof(struct dm_thin_new_mapping));
1135 INIT_LIST_HEAD(&m->list);
1136 m->bio = NULL;
1137
1138 pool->next_mapping = NULL;
1139
1140 return m;
1141}
1142
1143static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1144 sector_t begin, sector_t end)
1145{
1146 int r;
1147 struct dm_io_region to;
1148
1149 to.bdev = tc->pool_dev->bdev;
1150 to.sector = begin;
1151 to.count = end - begin;
1152
1153 r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1154 if (r < 0) {
1155 DMERR_LIMIT("dm_kcopyd_zero() failed");
1156 copy_complete(1, 1, m);
1157 }
1158}
1159
1160static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1161 dm_block_t data_begin,
1162 struct dm_thin_new_mapping *m)
1163{
1164 struct pool *pool = tc->pool;
1165 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1166
1167 h->overwrite_mapping = m;
1168 m->bio = bio;
1169 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1170 inc_all_io_entry(pool, bio);
1171 remap_and_issue(tc, bio, data_begin);
1172}
1173
1174/*
1175 * A partial copy also needs to zero the uncopied region.
1176 */
1177static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1178 struct dm_dev *origin, dm_block_t data_origin,
1179 dm_block_t data_dest,
1180 struct dm_bio_prison_cell *cell, struct bio *bio,
1181 sector_t len)
1182{
1183 int r;
1184 struct pool *pool = tc->pool;
1185 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1186
1187 m->tc = tc;
1188 m->virt_begin = virt_block;
1189 m->virt_end = virt_block + 1u;
1190 m->data_block = data_dest;
1191 m->cell = cell;
1192
1193 /*
1194 * quiesce action + copy action + an extra reference held for the
1195 * duration of this function (we may need to inc later for a
1196 * partial zero).
1197 */
1198 atomic_set(&m->prepare_actions, 3);
1199
1200 if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1201 complete_mapping_preparation(m); /* already quiesced */
1202
1203 /*
1204 * IO to pool_dev remaps to the pool target's data_dev.
1205 *
1206 * If the whole block of data is being overwritten, we can issue the
1207 * bio immediately. Otherwise we use kcopyd to clone the data first.
1208 */
1209 if (io_overwrites_block(pool, bio))
1210 remap_and_issue_overwrite(tc, bio, data_dest, m);
1211 else {
1212 struct dm_io_region from, to;
1213
1214 from.bdev = origin->bdev;
1215 from.sector = data_origin * pool->sectors_per_block;
1216 from.count = len;
1217
1218 to.bdev = tc->pool_dev->bdev;
1219 to.sector = data_dest * pool->sectors_per_block;
1220 to.count = len;
1221
1222 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1223 0, copy_complete, m);
1224 if (r < 0) {
1225 DMERR_LIMIT("dm_kcopyd_copy() failed");
1226 copy_complete(1, 1, m);
1227
1228 /*
1229 * We allow the zero to be issued, to simplify the
1230 * error path. Otherwise we'd need to start
1231 * worrying about decrementing the prepare_actions
1232 * counter.
1233 */
1234 }
1235
1236 /*
1237 * Do we need to zero a tail region?
1238 */
1239 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1240 atomic_inc(&m->prepare_actions);
1241 ll_zero(tc, m,
1242 data_dest * pool->sectors_per_block + len,
1243 (data_dest + 1) * pool->sectors_per_block);
1244 }
1245 }
1246
1247 complete_mapping_preparation(m); /* drop our ref */
1248}
1249
1250static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1251 dm_block_t data_origin, dm_block_t data_dest,
1252 struct dm_bio_prison_cell *cell, struct bio *bio)
1253{
1254 schedule_copy(tc, virt_block, tc->pool_dev,
1255 data_origin, data_dest, cell, bio,
1256 tc->pool->sectors_per_block);
1257}
1258
1259static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1260 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1261 struct bio *bio)
1262{
1263 struct pool *pool = tc->pool;
1264 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1265
1266 atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1267 m->tc = tc;
1268 m->virt_begin = virt_block;
1269 m->virt_end = virt_block + 1u;
1270 m->data_block = data_block;
1271 m->cell = cell;
1272
1273 /*
1274 * If the whole block of data is being overwritten or we are not
1275 * zeroing pre-existing data, we can issue the bio immediately.
1276 * Otherwise we use kcopyd to zero the data first.
1277 */
1278 if (pool->pf.zero_new_blocks) {
1279 if (io_overwrites_block(pool, bio))
1280 remap_and_issue_overwrite(tc, bio, data_block, m);
1281 else
1282 ll_zero(tc, m, data_block * pool->sectors_per_block,
1283 (data_block + 1) * pool->sectors_per_block);
1284 } else
1285 process_prepared_mapping(m);
1286}
1287
1288static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1289 dm_block_t data_dest,
1290 struct dm_bio_prison_cell *cell, struct bio *bio)
1291{
1292 struct pool *pool = tc->pool;
1293 sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1294 sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1295
1296 if (virt_block_end <= tc->origin_size)
1297 schedule_copy(tc, virt_block, tc->origin_dev,
1298 virt_block, data_dest, cell, bio,
1299 pool->sectors_per_block);
1300
1301 else if (virt_block_begin < tc->origin_size)
1302 schedule_copy(tc, virt_block, tc->origin_dev,
1303 virt_block, data_dest, cell, bio,
1304 tc->origin_size - virt_block_begin);
1305
1306 else
1307 schedule_zero(tc, virt_block, data_dest, cell, bio);
1308}
1309
1310static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1311
1312static void check_for_space(struct pool *pool)
1313{
1314 int r;
1315 dm_block_t nr_free;
1316
1317 if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1318 return;
1319
1320 r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1321 if (r)
1322 return;
1323
1324 if (nr_free)
1325 set_pool_mode(pool, PM_WRITE);
1326}
1327
1328/*
1329 * A non-zero return indicates read_only or fail_io mode.
1330 * Many callers don't care about the return value.
1331 */
1332static int commit(struct pool *pool)
1333{
1334 int r;
1335
1336 if (get_pool_mode(pool) >= PM_READ_ONLY)
1337 return -EINVAL;
1338
1339 r = dm_pool_commit_metadata(pool->pmd);
1340 if (r)
1341 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1342 else
1343 check_for_space(pool);
1344
1345 return r;
1346}
1347
1348static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1349{
1350 unsigned long flags;
1351
1352 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1353 DMWARN("%s: reached low water mark for data device: sending event.",
1354 dm_device_name(pool->pool_md));
1355 spin_lock_irqsave(&pool->lock, flags);
1356 pool->low_water_triggered = true;
1357 spin_unlock_irqrestore(&pool->lock, flags);
1358 dm_table_event(pool->ti->table);
1359 }
1360}
1361
1362static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1363{
1364 int r;
1365 dm_block_t free_blocks;
1366 struct pool *pool = tc->pool;
1367
1368 if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1369 return -EINVAL;
1370
1371 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1372 if (r) {
1373 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1374 return r;
1375 }
1376
1377 check_low_water_mark(pool, free_blocks);
1378
1379 if (!free_blocks) {
1380 /*
1381 * Try to commit to see if that will free up some
1382 * more space.
1383 */
1384 r = commit(pool);
1385 if (r)
1386 return r;
1387
1388 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1389 if (r) {
1390 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1391 return r;
1392 }
1393
1394 if (!free_blocks) {
1395 set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1396 return -ENOSPC;
1397 }
1398 }
1399
1400 r = dm_pool_alloc_data_block(pool->pmd, result);
1401 if (r) {
1402 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1403 return r;
1404 }
1405
1406 return 0;
1407}
1408
1409/*
1410 * If we have run out of space, queue bios until the device is
1411 * resumed, presumably after having been reloaded with more space.
1412 */
1413static void retry_on_resume(struct bio *bio)
1414{
1415 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1416 struct thin_c *tc = h->tc;
1417 unsigned long flags;
1418
1419 spin_lock_irqsave(&tc->lock, flags);
1420 bio_list_add(&tc->retry_on_resume_list, bio);
1421 spin_unlock_irqrestore(&tc->lock, flags);
1422}
1423
1424static int should_error_unserviceable_bio(struct pool *pool)
1425{
1426 enum pool_mode m = get_pool_mode(pool);
1427
1428 switch (m) {
1429 case PM_WRITE:
1430 /* Shouldn't get here */
1431 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1432 return -EIO;
1433
1434 case PM_OUT_OF_DATA_SPACE:
1435 return pool->pf.error_if_no_space ? -ENOSPC : 0;
1436
1437 case PM_READ_ONLY:
1438 case PM_FAIL:
1439 return -EIO;
1440 default:
1441 /* Shouldn't get here */
1442 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1443 return -EIO;
1444 }
1445}
1446
1447static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1448{
1449 int error = should_error_unserviceable_bio(pool);
1450
1451 if (error) {
1452 bio->bi_error = error;
1453 bio_endio(bio);
1454 } else
1455 retry_on_resume(bio);
1456}
1457
1458static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1459{
1460 struct bio *bio;
1461 struct bio_list bios;
1462 int error;
1463
1464 error = should_error_unserviceable_bio(pool);
1465 if (error) {
1466 cell_error_with_code(pool, cell, error);
1467 return;
1468 }
1469
1470 bio_list_init(&bios);
1471 cell_release(pool, cell, &bios);
1472
1473 while ((bio = bio_list_pop(&bios)))
1474 retry_on_resume(bio);
1475}
1476
1477static void process_discard_cell_no_passdown(struct thin_c *tc,
1478 struct dm_bio_prison_cell *virt_cell)
1479{
1480 struct pool *pool = tc->pool;
1481 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1482
1483 /*
1484 * We don't need to lock the data blocks, since there's no
1485 * passdown. We only lock data blocks for allocation and breaking sharing.
1486 */
1487 m->tc = tc;
1488 m->virt_begin = virt_cell->key.block_begin;
1489 m->virt_end = virt_cell->key.block_end;
1490 m->cell = virt_cell;
1491 m->bio = virt_cell->holder;
1492
1493 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1494 pool->process_prepared_discard(m);
1495}
1496
1497/*
1498 * __bio_inc_remaining() is used to defer parent bios's end_io until
1499 * we _know_ all chained sub range discard bios have completed.
1500 */
1501static inline void __bio_inc_remaining(struct bio *bio)
1502{
1503 bio->bi_flags |= (1 << BIO_CHAIN);
1504 smp_mb__before_atomic();
1505 atomic_inc(&bio->__bi_remaining);
1506}
1507
1508static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1509 struct bio *bio)
1510{
1511 struct pool *pool = tc->pool;
1512
1513 int r;
1514 bool maybe_shared;
1515 struct dm_cell_key data_key;
1516 struct dm_bio_prison_cell *data_cell;
1517 struct dm_thin_new_mapping *m;
1518 dm_block_t virt_begin, virt_end, data_begin;
1519
1520 while (begin != end) {
1521 r = ensure_next_mapping(pool);
1522 if (r)
1523 /* we did our best */
1524 return;
1525
1526 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1527 &data_begin, &maybe_shared);
1528 if (r)
1529 /*
1530 * Silently fail, letting any mappings we've
1531 * created complete.
1532 */
1533 break;
1534
1535 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1536 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1537 /* contention, we'll give up with this range */
1538 begin = virt_end;
1539 continue;
1540 }
1541
1542 /*
1543 * IO may still be going to the destination block. We must
1544 * quiesce before we can do the removal.
1545 */
1546 m = get_next_mapping(pool);
1547 m->tc = tc;
1548 m->maybe_shared = maybe_shared;
1549 m->virt_begin = virt_begin;
1550 m->virt_end = virt_end;
1551 m->data_block = data_begin;
1552 m->cell = data_cell;
1553 m->bio = bio;
1554
1555 /*
1556 * The parent bio must not complete before sub discard bios are
1557 * chained to it (see __blkdev_issue_discard_async's bio_chain)!
1558 *
1559 * This per-mapping bi_remaining increment is paired with
1560 * the implicit decrement that occurs via bio_endio() in
1561 * process_prepared_discard_{passdown,no_passdown}.
1562 */
1563 __bio_inc_remaining(bio);
1564 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1565 pool->process_prepared_discard(m);
1566
1567 begin = virt_end;
1568 }
1569}
1570
1571static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1572{
1573 struct bio *bio = virt_cell->holder;
1574 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1575
1576 /*
1577 * The virt_cell will only get freed once the origin bio completes.
1578 * This means it will remain locked while all the individual
1579 * passdown bios are in flight.
1580 */
1581 h->cell = virt_cell;
1582 break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1583
1584 /*
1585 * We complete the bio now, knowing that the bi_remaining field
1586 * will prevent completion until the sub range discards have
1587 * completed.
1588 */
1589 bio_endio(bio);
1590}
1591
1592static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1593{
1594 dm_block_t begin, end;
1595 struct dm_cell_key virt_key;
1596 struct dm_bio_prison_cell *virt_cell;
1597
1598 get_bio_block_range(tc, bio, &begin, &end);
1599 if (begin == end) {
1600 /*
1601 * The discard covers less than a block.
1602 */
1603 bio_endio(bio);
1604 return;
1605 }
1606
1607 build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1608 if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1609 /*
1610 * Potential starvation issue: We're relying on the
1611 * fs/application being well behaved, and not trying to
1612 * send IO to a region at the same time as discarding it.
1613 * If they do this persistently then it's possible this
1614 * cell will never be granted.
1615 */
1616 return;
1617
1618 tc->pool->process_discard_cell(tc, virt_cell);
1619}
1620
1621static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1622 struct dm_cell_key *key,
1623 struct dm_thin_lookup_result *lookup_result,
1624 struct dm_bio_prison_cell *cell)
1625{
1626 int r;
1627 dm_block_t data_block;
1628 struct pool *pool = tc->pool;
1629
1630 r = alloc_data_block(tc, &data_block);
1631 switch (r) {
1632 case 0:
1633 schedule_internal_copy(tc, block, lookup_result->block,
1634 data_block, cell, bio);
1635 break;
1636
1637 case -ENOSPC:
1638 retry_bios_on_resume(pool, cell);
1639 break;
1640
1641 default:
1642 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1643 __func__, r);
1644 cell_error(pool, cell);
1645 break;
1646 }
1647}
1648
1649static void __remap_and_issue_shared_cell(void *context,
1650 struct dm_bio_prison_cell *cell)
1651{
1652 struct remap_info *info = context;
1653 struct bio *bio;
1654
1655 while ((bio = bio_list_pop(&cell->bios))) {
1656 if ((bio_data_dir(bio) == WRITE) ||
1657 (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)))
1658 bio_list_add(&info->defer_bios, bio);
1659 else {
1660 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1661
1662 h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1663 inc_all_io_entry(info->tc->pool, bio);
1664 bio_list_add(&info->issue_bios, bio);
1665 }
1666 }
1667}
1668
1669static void remap_and_issue_shared_cell(struct thin_c *tc,
1670 struct dm_bio_prison_cell *cell,
1671 dm_block_t block)
1672{
1673 struct bio *bio;
1674 struct remap_info info;
1675
1676 info.tc = tc;
1677 bio_list_init(&info.defer_bios);
1678 bio_list_init(&info.issue_bios);
1679
1680 cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1681 &info, cell);
1682
1683 while ((bio = bio_list_pop(&info.defer_bios)))
1684 thin_defer_bio(tc, bio);
1685
1686 while ((bio = bio_list_pop(&info.issue_bios)))
1687 remap_and_issue(tc, bio, block);
1688}
1689
1690static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1691 dm_block_t block,
1692 struct dm_thin_lookup_result *lookup_result,
1693 struct dm_bio_prison_cell *virt_cell)
1694{
1695 struct dm_bio_prison_cell *data_cell;
1696 struct pool *pool = tc->pool;
1697 struct dm_cell_key key;
1698
1699 /*
1700 * If cell is already occupied, then sharing is already in the process
1701 * of being broken so we have nothing further to do here.
1702 */
1703 build_data_key(tc->td, lookup_result->block, &key);
1704 if (bio_detain(pool, &key, bio, &data_cell)) {
1705 cell_defer_no_holder(tc, virt_cell);
1706 return;
1707 }
1708
1709 if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1710 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1711 cell_defer_no_holder(tc, virt_cell);
1712 } else {
1713 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1714
1715 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1716 inc_all_io_entry(pool, bio);
1717 remap_and_issue(tc, bio, lookup_result->block);
1718
1719 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1720 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1721 }
1722}
1723
1724static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1725 struct dm_bio_prison_cell *cell)
1726{
1727 int r;
1728 dm_block_t data_block;
1729 struct pool *pool = tc->pool;
1730
1731 /*
1732 * Remap empty bios (flushes) immediately, without provisioning.
1733 */
1734 if (!bio->bi_iter.bi_size) {
1735 inc_all_io_entry(pool, bio);
1736 cell_defer_no_holder(tc, cell);
1737
1738 remap_and_issue(tc, bio, 0);
1739 return;
1740 }
1741
1742 /*
1743 * Fill read bios with zeroes and complete them immediately.
1744 */
1745 if (bio_data_dir(bio) == READ) {
1746 zero_fill_bio(bio);
1747 cell_defer_no_holder(tc, cell);
1748 bio_endio(bio);
1749 return;
1750 }
1751
1752 r = alloc_data_block(tc, &data_block);
1753 switch (r) {
1754 case 0:
1755 if (tc->origin_dev)
1756 schedule_external_copy(tc, block, data_block, cell, bio);
1757 else
1758 schedule_zero(tc, block, data_block, cell, bio);
1759 break;
1760
1761 case -ENOSPC:
1762 retry_bios_on_resume(pool, cell);
1763 break;
1764
1765 default:
1766 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1767 __func__, r);
1768 cell_error(pool, cell);
1769 break;
1770 }
1771}
1772
1773static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1774{
1775 int r;
1776 struct pool *pool = tc->pool;
1777 struct bio *bio = cell->holder;
1778 dm_block_t block = get_bio_block(tc, bio);
1779 struct dm_thin_lookup_result lookup_result;
1780
1781 if (tc->requeue_mode) {
1782 cell_requeue(pool, cell);
1783 return;
1784 }
1785
1786 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1787 switch (r) {
1788 case 0:
1789 if (lookup_result.shared)
1790 process_shared_bio(tc, bio, block, &lookup_result, cell);
1791 else {
1792 inc_all_io_entry(pool, bio);
1793 remap_and_issue(tc, bio, lookup_result.block);
1794 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1795 }
1796 break;
1797
1798 case -ENODATA:
1799 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1800 inc_all_io_entry(pool, bio);
1801 cell_defer_no_holder(tc, cell);
1802
1803 if (bio_end_sector(bio) <= tc->origin_size)
1804 remap_to_origin_and_issue(tc, bio);
1805
1806 else if (bio->bi_iter.bi_sector < tc->origin_size) {
1807 zero_fill_bio(bio);
1808 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1809 remap_to_origin_and_issue(tc, bio);
1810
1811 } else {
1812 zero_fill_bio(bio);
1813 bio_endio(bio);
1814 }
1815 } else
1816 provision_block(tc, bio, block, cell);
1817 break;
1818
1819 default:
1820 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1821 __func__, r);
1822 cell_defer_no_holder(tc, cell);
1823 bio_io_error(bio);
1824 break;
1825 }
1826}
1827
1828static void process_bio(struct thin_c *tc, struct bio *bio)
1829{
1830 struct pool *pool = tc->pool;
1831 dm_block_t block = get_bio_block(tc, bio);
1832 struct dm_bio_prison_cell *cell;
1833 struct dm_cell_key key;
1834
1835 /*
1836 * If cell is already occupied, then the block is already
1837 * being provisioned so we have nothing further to do here.
1838 */
1839 build_virtual_key(tc->td, block, &key);
1840 if (bio_detain(pool, &key, bio, &cell))
1841 return;
1842
1843 process_cell(tc, cell);
1844}
1845
1846static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1847 struct dm_bio_prison_cell *cell)
1848{
1849 int r;
1850 int rw = bio_data_dir(bio);
1851 dm_block_t block = get_bio_block(tc, bio);
1852 struct dm_thin_lookup_result lookup_result;
1853
1854 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1855 switch (r) {
1856 case 0:
1857 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1858 handle_unserviceable_bio(tc->pool, bio);
1859 if (cell)
1860 cell_defer_no_holder(tc, cell);
1861 } else {
1862 inc_all_io_entry(tc->pool, bio);
1863 remap_and_issue(tc, bio, lookup_result.block);
1864 if (cell)
1865 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1866 }
1867 break;
1868
1869 case -ENODATA:
1870 if (cell)
1871 cell_defer_no_holder(tc, cell);
1872 if (rw != READ) {
1873 handle_unserviceable_bio(tc->pool, bio);
1874 break;
1875 }
1876
1877 if (tc->origin_dev) {
1878 inc_all_io_entry(tc->pool, bio);
1879 remap_to_origin_and_issue(tc, bio);
1880 break;
1881 }
1882
1883 zero_fill_bio(bio);
1884 bio_endio(bio);
1885 break;
1886
1887 default:
1888 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1889 __func__, r);
1890 if (cell)
1891 cell_defer_no_holder(tc, cell);
1892 bio_io_error(bio);
1893 break;
1894 }
1895}
1896
1897static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1898{
1899 __process_bio_read_only(tc, bio, NULL);
1900}
1901
1902static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1903{
1904 __process_bio_read_only(tc, cell->holder, cell);
1905}
1906
1907static void process_bio_success(struct thin_c *tc, struct bio *bio)
1908{
1909 bio_endio(bio);
1910}
1911
1912static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1913{
1914 bio_io_error(bio);
1915}
1916
1917static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1918{
1919 cell_success(tc->pool, cell);
1920}
1921
1922static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1923{
1924 cell_error(tc->pool, cell);
1925}
1926
1927/*
1928 * FIXME: should we also commit due to size of transaction, measured in
1929 * metadata blocks?
1930 */
1931static int need_commit_due_to_time(struct pool *pool)
1932{
1933 return !time_in_range(jiffies, pool->last_commit_jiffies,
1934 pool->last_commit_jiffies + COMMIT_PERIOD);
1935}
1936
1937#define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1938#define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1939
1940static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
1941{
1942 struct rb_node **rbp, *parent;
1943 struct dm_thin_endio_hook *pbd;
1944 sector_t bi_sector = bio->bi_iter.bi_sector;
1945
1946 rbp = &tc->sort_bio_list.rb_node;
1947 parent = NULL;
1948 while (*rbp) {
1949 parent = *rbp;
1950 pbd = thin_pbd(parent);
1951
1952 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
1953 rbp = &(*rbp)->rb_left;
1954 else
1955 rbp = &(*rbp)->rb_right;
1956 }
1957
1958 pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1959 rb_link_node(&pbd->rb_node, parent, rbp);
1960 rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
1961}
1962
1963static void __extract_sorted_bios(struct thin_c *tc)
1964{
1965 struct rb_node *node;
1966 struct dm_thin_endio_hook *pbd;
1967 struct bio *bio;
1968
1969 for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
1970 pbd = thin_pbd(node);
1971 bio = thin_bio(pbd);
1972
1973 bio_list_add(&tc->deferred_bio_list, bio);
1974 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
1975 }
1976
1977 WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
1978}
1979
1980static void __sort_thin_deferred_bios(struct thin_c *tc)
1981{
1982 struct bio *bio;
1983 struct bio_list bios;
1984
1985 bio_list_init(&bios);
1986 bio_list_merge(&bios, &tc->deferred_bio_list);
1987 bio_list_init(&tc->deferred_bio_list);
1988
1989 /* Sort deferred_bio_list using rb-tree */
1990 while ((bio = bio_list_pop(&bios)))
1991 __thin_bio_rb_add(tc, bio);
1992
1993 /*
1994 * Transfer the sorted bios in sort_bio_list back to
1995 * deferred_bio_list to allow lockless submission of
1996 * all bios.
1997 */
1998 __extract_sorted_bios(tc);
1999}
2000
2001static void process_thin_deferred_bios(struct thin_c *tc)
2002{
2003 struct pool *pool = tc->pool;
2004 unsigned long flags;
2005 struct bio *bio;
2006 struct bio_list bios;
2007 struct blk_plug plug;
2008 unsigned count = 0;
2009
2010 if (tc->requeue_mode) {
2011 error_thin_bio_list(tc, &tc->deferred_bio_list, DM_ENDIO_REQUEUE);
2012 return;
2013 }
2014
2015 bio_list_init(&bios);
2016
2017 spin_lock_irqsave(&tc->lock, flags);
2018
2019 if (bio_list_empty(&tc->deferred_bio_list)) {
2020 spin_unlock_irqrestore(&tc->lock, flags);
2021 return;
2022 }
2023
2024 __sort_thin_deferred_bios(tc);
2025
2026 bio_list_merge(&bios, &tc->deferred_bio_list);
2027 bio_list_init(&tc->deferred_bio_list);
2028
2029 spin_unlock_irqrestore(&tc->lock, flags);
2030
2031 blk_start_plug(&plug);
2032 while ((bio = bio_list_pop(&bios))) {
2033 /*
2034 * If we've got no free new_mapping structs, and processing
2035 * this bio might require one, we pause until there are some
2036 * prepared mappings to process.
2037 */
2038 if (ensure_next_mapping(pool)) {
2039 spin_lock_irqsave(&tc->lock, flags);
2040 bio_list_add(&tc->deferred_bio_list, bio);
2041 bio_list_merge(&tc->deferred_bio_list, &bios);
2042 spin_unlock_irqrestore(&tc->lock, flags);
2043 break;
2044 }
2045
2046 if (bio->bi_rw & REQ_DISCARD)
2047 pool->process_discard(tc, bio);
2048 else
2049 pool->process_bio(tc, bio);
2050
2051 if ((count++ & 127) == 0) {
2052 throttle_work_update(&pool->throttle);
2053 dm_pool_issue_prefetches(pool->pmd);
2054 }
2055 }
2056 blk_finish_plug(&plug);
2057}
2058
2059static int cmp_cells(const void *lhs, const void *rhs)
2060{
2061 struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2062 struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2063
2064 BUG_ON(!lhs_cell->holder);
2065 BUG_ON(!rhs_cell->holder);
2066
2067 if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2068 return -1;
2069
2070 if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2071 return 1;
2072
2073 return 0;
2074}
2075
2076static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2077{
2078 unsigned count = 0;
2079 struct dm_bio_prison_cell *cell, *tmp;
2080
2081 list_for_each_entry_safe(cell, tmp, cells, user_list) {
2082 if (count >= CELL_SORT_ARRAY_SIZE)
2083 break;
2084
2085 pool->cell_sort_array[count++] = cell;
2086 list_del(&cell->user_list);
2087 }
2088
2089 sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2090
2091 return count;
2092}
2093
2094static void process_thin_deferred_cells(struct thin_c *tc)
2095{
2096 struct pool *pool = tc->pool;
2097 unsigned long flags;
2098 struct list_head cells;
2099 struct dm_bio_prison_cell *cell;
2100 unsigned i, j, count;
2101
2102 INIT_LIST_HEAD(&cells);
2103
2104 spin_lock_irqsave(&tc->lock, flags);
2105 list_splice_init(&tc->deferred_cells, &cells);
2106 spin_unlock_irqrestore(&tc->lock, flags);
2107
2108 if (list_empty(&cells))
2109 return;
2110
2111 do {
2112 count = sort_cells(tc->pool, &cells);
2113
2114 for (i = 0; i < count; i++) {
2115 cell = pool->cell_sort_array[i];
2116 BUG_ON(!cell->holder);
2117
2118 /*
2119 * If we've got no free new_mapping structs, and processing
2120 * this bio might require one, we pause until there are some
2121 * prepared mappings to process.
2122 */
2123 if (ensure_next_mapping(pool)) {
2124 for (j = i; j < count; j++)
2125 list_add(&pool->cell_sort_array[j]->user_list, &cells);
2126
2127 spin_lock_irqsave(&tc->lock, flags);
2128 list_splice(&cells, &tc->deferred_cells);
2129 spin_unlock_irqrestore(&tc->lock, flags);
2130 return;
2131 }
2132
2133 if (cell->holder->bi_rw & REQ_DISCARD)
2134 pool->process_discard_cell(tc, cell);
2135 else
2136 pool->process_cell(tc, cell);
2137 }
2138 } while (!list_empty(&cells));
2139}
2140
2141static void thin_get(struct thin_c *tc);
2142static void thin_put(struct thin_c *tc);
2143
2144/*
2145 * We can't hold rcu_read_lock() around code that can block. So we
2146 * find a thin with the rcu lock held; bump a refcount; then drop
2147 * the lock.
2148 */
2149static struct thin_c *get_first_thin(struct pool *pool)
2150{
2151 struct thin_c *tc = NULL;
2152
2153 rcu_read_lock();
2154 if (!list_empty(&pool->active_thins)) {
2155 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2156 thin_get(tc);
2157 }
2158 rcu_read_unlock();
2159
2160 return tc;
2161}
2162
2163static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2164{
2165 struct thin_c *old_tc = tc;
2166
2167 rcu_read_lock();
2168 list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2169 thin_get(tc);
2170 thin_put(old_tc);
2171 rcu_read_unlock();
2172 return tc;
2173 }
2174 thin_put(old_tc);
2175 rcu_read_unlock();
2176
2177 return NULL;
2178}
2179
2180static void process_deferred_bios(struct pool *pool)
2181{
2182 unsigned long flags;
2183 struct bio *bio;
2184 struct bio_list bios;
2185 struct thin_c *tc;
2186
2187 tc = get_first_thin(pool);
2188 while (tc) {
2189 process_thin_deferred_cells(tc);
2190 process_thin_deferred_bios(tc);
2191 tc = get_next_thin(pool, tc);
2192 }
2193
2194 /*
2195 * If there are any deferred flush bios, we must commit
2196 * the metadata before issuing them.
2197 */
2198 bio_list_init(&bios);
2199 spin_lock_irqsave(&pool->lock, flags);
2200 bio_list_merge(&bios, &pool->deferred_flush_bios);
2201 bio_list_init(&pool->deferred_flush_bios);
2202 spin_unlock_irqrestore(&pool->lock, flags);
2203
2204 if (bio_list_empty(&bios) &&
2205 !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2206 return;
2207
2208 if (commit(pool)) {
2209 while ((bio = bio_list_pop(&bios)))
2210 bio_io_error(bio);
2211 return;
2212 }
2213 pool->last_commit_jiffies = jiffies;
2214
2215 while ((bio = bio_list_pop(&bios)))
2216 generic_make_request(bio);
2217}
2218
2219static void do_worker(struct work_struct *ws)
2220{
2221 struct pool *pool = container_of(ws, struct pool, worker);
2222
2223 throttle_work_start(&pool->throttle);
2224 dm_pool_issue_prefetches(pool->pmd);
2225 throttle_work_update(&pool->throttle);
2226 process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2227 throttle_work_update(&pool->throttle);
2228 process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2229 throttle_work_update(&pool->throttle);
2230 process_deferred_bios(pool);
2231 throttle_work_complete(&pool->throttle);
2232}
2233
2234/*
2235 * We want to commit periodically so that not too much
2236 * unwritten data builds up.
2237 */
2238static void do_waker(struct work_struct *ws)
2239{
2240 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2241 wake_worker(pool);
2242 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2243}
2244
2245static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2246
2247/*
2248 * We're holding onto IO to allow userland time to react. After the
2249 * timeout either the pool will have been resized (and thus back in
2250 * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2251 */
2252static void do_no_space_timeout(struct work_struct *ws)
2253{
2254 struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2255 no_space_timeout);
2256
2257 if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2258 pool->pf.error_if_no_space = true;
2259 notify_of_pool_mode_change_to_oods(pool);
2260 error_retry_list_with_code(pool, -ENOSPC);
2261 }
2262}
2263
2264/*----------------------------------------------------------------*/
2265
2266struct pool_work {
2267 struct work_struct worker;
2268 struct completion complete;
2269};
2270
2271static struct pool_work *to_pool_work(struct work_struct *ws)
2272{
2273 return container_of(ws, struct pool_work, worker);
2274}
2275
2276static void pool_work_complete(struct pool_work *pw)
2277{
2278 complete(&pw->complete);
2279}
2280
2281static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2282 void (*fn)(struct work_struct *))
2283{
2284 INIT_WORK_ONSTACK(&pw->worker, fn);
2285 init_completion(&pw->complete);
2286 queue_work(pool->wq, &pw->worker);
2287 wait_for_completion(&pw->complete);
2288}
2289
2290/*----------------------------------------------------------------*/
2291
2292struct noflush_work {
2293 struct pool_work pw;
2294 struct thin_c *tc;
2295};
2296
2297static struct noflush_work *to_noflush(struct work_struct *ws)
2298{
2299 return container_of(to_pool_work(ws), struct noflush_work, pw);
2300}
2301
2302static void do_noflush_start(struct work_struct *ws)
2303{
2304 struct noflush_work *w = to_noflush(ws);
2305 w->tc->requeue_mode = true;
2306 requeue_io(w->tc);
2307 pool_work_complete(&w->pw);
2308}
2309
2310static void do_noflush_stop(struct work_struct *ws)
2311{
2312 struct noflush_work *w = to_noflush(ws);
2313 w->tc->requeue_mode = false;
2314 pool_work_complete(&w->pw);
2315}
2316
2317static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2318{
2319 struct noflush_work w;
2320
2321 w.tc = tc;
2322 pool_work_wait(&w.pw, tc->pool, fn);
2323}
2324
2325/*----------------------------------------------------------------*/
2326
2327static enum pool_mode get_pool_mode(struct pool *pool)
2328{
2329 return pool->pf.mode;
2330}
2331
2332static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2333{
2334 dm_table_event(pool->ti->table);
2335 DMINFO("%s: switching pool to %s mode",
2336 dm_device_name(pool->pool_md), new_mode);
2337}
2338
2339static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2340{
2341 if (!pool->pf.error_if_no_space)
2342 notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2343 else
2344 notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2345}
2346
2347static bool passdown_enabled(struct pool_c *pt)
2348{
2349 return pt->adjusted_pf.discard_passdown;
2350}
2351
2352static void set_discard_callbacks(struct pool *pool)
2353{
2354 struct pool_c *pt = pool->ti->private;
2355
2356 if (passdown_enabled(pt)) {
2357 pool->process_discard_cell = process_discard_cell_passdown;
2358 pool->process_prepared_discard = process_prepared_discard_passdown;
2359 } else {
2360 pool->process_discard_cell = process_discard_cell_no_passdown;
2361 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2362 }
2363}
2364
2365static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2366{
2367 struct pool_c *pt = pool->ti->private;
2368 bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2369 enum pool_mode old_mode = get_pool_mode(pool);
2370 unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2371
2372 /*
2373 * Never allow the pool to transition to PM_WRITE mode if user
2374 * intervention is required to verify metadata and data consistency.
2375 */
2376 if (new_mode == PM_WRITE && needs_check) {
2377 DMERR("%s: unable to switch pool to write mode until repaired.",
2378 dm_device_name(pool->pool_md));
2379 if (old_mode != new_mode)
2380 new_mode = old_mode;
2381 else
2382 new_mode = PM_READ_ONLY;
2383 }
2384 /*
2385 * If we were in PM_FAIL mode, rollback of metadata failed. We're
2386 * not going to recover without a thin_repair. So we never let the
2387 * pool move out of the old mode.
2388 */
2389 if (old_mode == PM_FAIL)
2390 new_mode = old_mode;
2391
2392 switch (new_mode) {
2393 case PM_FAIL:
2394 if (old_mode != new_mode)
2395 notify_of_pool_mode_change(pool, "failure");
2396 dm_pool_metadata_read_only(pool->pmd);
2397 pool->process_bio = process_bio_fail;
2398 pool->process_discard = process_bio_fail;
2399 pool->process_cell = process_cell_fail;
2400 pool->process_discard_cell = process_cell_fail;
2401 pool->process_prepared_mapping = process_prepared_mapping_fail;
2402 pool->process_prepared_discard = process_prepared_discard_fail;
2403
2404 error_retry_list(pool);
2405 break;
2406
2407 case PM_READ_ONLY:
2408 if (old_mode != new_mode)
2409 notify_of_pool_mode_change(pool, "read-only");
2410 dm_pool_metadata_read_only(pool->pmd);
2411 pool->process_bio = process_bio_read_only;
2412 pool->process_discard = process_bio_success;
2413 pool->process_cell = process_cell_read_only;
2414 pool->process_discard_cell = process_cell_success;
2415 pool->process_prepared_mapping = process_prepared_mapping_fail;
2416 pool->process_prepared_discard = process_prepared_discard_success;
2417
2418 error_retry_list(pool);
2419 break;
2420
2421 case PM_OUT_OF_DATA_SPACE:
2422 /*
2423 * Ideally we'd never hit this state; the low water mark
2424 * would trigger userland to extend the pool before we
2425 * completely run out of data space. However, many small
2426 * IOs to unprovisioned space can consume data space at an
2427 * alarming rate. Adjust your low water mark if you're
2428 * frequently seeing this mode.
2429 */
2430 if (old_mode != new_mode)
2431 notify_of_pool_mode_change_to_oods(pool);
2432 pool->out_of_data_space = true;
2433 pool->process_bio = process_bio_read_only;
2434 pool->process_discard = process_discard_bio;
2435 pool->process_cell = process_cell_read_only;
2436 pool->process_prepared_mapping = process_prepared_mapping;
2437 set_discard_callbacks(pool);
2438
2439 if (!pool->pf.error_if_no_space && no_space_timeout)
2440 queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2441 break;
2442
2443 case PM_WRITE:
2444 if (old_mode != new_mode)
2445 notify_of_pool_mode_change(pool, "write");
2446 pool->out_of_data_space = false;
2447 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2448 dm_pool_metadata_read_write(pool->pmd);
2449 pool->process_bio = process_bio;
2450 pool->process_discard = process_discard_bio;
2451 pool->process_cell = process_cell;
2452 pool->process_prepared_mapping = process_prepared_mapping;
2453 set_discard_callbacks(pool);
2454 break;
2455 }
2456
2457 pool->pf.mode = new_mode;
2458 /*
2459 * The pool mode may have changed, sync it so bind_control_target()
2460 * doesn't cause an unexpected mode transition on resume.
2461 */
2462 pt->adjusted_pf.mode = new_mode;
2463}
2464
2465static void abort_transaction(struct pool *pool)
2466{
2467 const char *dev_name = dm_device_name(pool->pool_md);
2468
2469 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2470 if (dm_pool_abort_metadata(pool->pmd)) {
2471 DMERR("%s: failed to abort metadata transaction", dev_name);
2472 set_pool_mode(pool, PM_FAIL);
2473 }
2474
2475 if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2476 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2477 set_pool_mode(pool, PM_FAIL);
2478 }
2479}
2480
2481static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2482{
2483 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2484 dm_device_name(pool->pool_md), op, r);
2485
2486 abort_transaction(pool);
2487 set_pool_mode(pool, PM_READ_ONLY);
2488}
2489
2490/*----------------------------------------------------------------*/
2491
2492/*
2493 * Mapping functions.
2494 */
2495
2496/*
2497 * Called only while mapping a thin bio to hand it over to the workqueue.
2498 */
2499static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2500{
2501 unsigned long flags;
2502 struct pool *pool = tc->pool;
2503
2504 spin_lock_irqsave(&tc->lock, flags);
2505 bio_list_add(&tc->deferred_bio_list, bio);
2506 spin_unlock_irqrestore(&tc->lock, flags);
2507
2508 wake_worker(pool);
2509}
2510
2511static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2512{
2513 struct pool *pool = tc->pool;
2514
2515 throttle_lock(&pool->throttle);
2516 thin_defer_bio(tc, bio);
2517 throttle_unlock(&pool->throttle);
2518}
2519
2520static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2521{
2522 unsigned long flags;
2523 struct pool *pool = tc->pool;
2524
2525 throttle_lock(&pool->throttle);
2526 spin_lock_irqsave(&tc->lock, flags);
2527 list_add_tail(&cell->user_list, &tc->deferred_cells);
2528 spin_unlock_irqrestore(&tc->lock, flags);
2529 throttle_unlock(&pool->throttle);
2530
2531 wake_worker(pool);
2532}
2533
2534static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2535{
2536 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2537
2538 h->tc = tc;
2539 h->shared_read_entry = NULL;
2540 h->all_io_entry = NULL;
2541 h->overwrite_mapping = NULL;
2542 h->cell = NULL;
2543}
2544
2545/*
2546 * Non-blocking function called from the thin target's map function.
2547 */
2548static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2549{
2550 int r;
2551 struct thin_c *tc = ti->private;
2552 dm_block_t block = get_bio_block(tc, bio);
2553 struct dm_thin_device *td = tc->td;
2554 struct dm_thin_lookup_result result;
2555 struct dm_bio_prison_cell *virt_cell, *data_cell;
2556 struct dm_cell_key key;
2557
2558 thin_hook_bio(tc, bio);
2559
2560 if (tc->requeue_mode) {
2561 bio->bi_error = DM_ENDIO_REQUEUE;
2562 bio_endio(bio);
2563 return DM_MAPIO_SUBMITTED;
2564 }
2565
2566 if (get_pool_mode(tc->pool) == PM_FAIL) {
2567 bio_io_error(bio);
2568 return DM_MAPIO_SUBMITTED;
2569 }
2570
2571 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
2572 thin_defer_bio_with_throttle(tc, bio);
2573 return DM_MAPIO_SUBMITTED;
2574 }
2575
2576 /*
2577 * We must hold the virtual cell before doing the lookup, otherwise
2578 * there's a race with discard.
2579 */
2580 build_virtual_key(tc->td, block, &key);
2581 if (bio_detain(tc->pool, &key, bio, &virt_cell))
2582 return DM_MAPIO_SUBMITTED;
2583
2584 r = dm_thin_find_block(td, block, 0, &result);
2585
2586 /*
2587 * Note that we defer readahead too.
2588 */
2589 switch (r) {
2590 case 0:
2591 if (unlikely(result.shared)) {
2592 /*
2593 * We have a race condition here between the
2594 * result.shared value returned by the lookup and
2595 * snapshot creation, which may cause new
2596 * sharing.
2597 *
2598 * To avoid this always quiesce the origin before
2599 * taking the snap. You want to do this anyway to
2600 * ensure a consistent application view
2601 * (i.e. lockfs).
2602 *
2603 * More distant ancestors are irrelevant. The
2604 * shared flag will be set in their case.
2605 */
2606 thin_defer_cell(tc, virt_cell);
2607 return DM_MAPIO_SUBMITTED;
2608 }
2609
2610 build_data_key(tc->td, result.block, &key);
2611 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2612 cell_defer_no_holder(tc, virt_cell);
2613 return DM_MAPIO_SUBMITTED;
2614 }
2615
2616 inc_all_io_entry(tc->pool, bio);
2617 cell_defer_no_holder(tc, data_cell);
2618 cell_defer_no_holder(tc, virt_cell);
2619
2620 remap(tc, bio, result.block);
2621 return DM_MAPIO_REMAPPED;
2622
2623 case -ENODATA:
2624 case -EWOULDBLOCK:
2625 thin_defer_cell(tc, virt_cell);
2626 return DM_MAPIO_SUBMITTED;
2627
2628 default:
2629 /*
2630 * Must always call bio_io_error on failure.
2631 * dm_thin_find_block can fail with -EINVAL if the
2632 * pool is switched to fail-io mode.
2633 */
2634 bio_io_error(bio);
2635 cell_defer_no_holder(tc, virt_cell);
2636 return DM_MAPIO_SUBMITTED;
2637 }
2638}
2639
2640static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2641{
2642 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2643 struct request_queue *q;
2644
2645 if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2646 return 1;
2647
2648 q = bdev_get_queue(pt->data_dev->bdev);
2649 return bdi_congested(&q->backing_dev_info, bdi_bits);
2650}
2651
2652static void requeue_bios(struct pool *pool)
2653{
2654 unsigned long flags;
2655 struct thin_c *tc;
2656
2657 rcu_read_lock();
2658 list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2659 spin_lock_irqsave(&tc->lock, flags);
2660 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2661 bio_list_init(&tc->retry_on_resume_list);
2662 spin_unlock_irqrestore(&tc->lock, flags);
2663 }
2664 rcu_read_unlock();
2665}
2666
2667/*----------------------------------------------------------------
2668 * Binding of control targets to a pool object
2669 *--------------------------------------------------------------*/
2670static bool data_dev_supports_discard(struct pool_c *pt)
2671{
2672 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2673
2674 return q && blk_queue_discard(q);
2675}
2676
2677static bool is_factor(sector_t block_size, uint32_t n)
2678{
2679 return !sector_div(block_size, n);
2680}
2681
2682/*
2683 * If discard_passdown was enabled verify that the data device
2684 * supports discards. Disable discard_passdown if not.
2685 */
2686static void disable_passdown_if_not_supported(struct pool_c *pt)
2687{
2688 struct pool *pool = pt->pool;
2689 struct block_device *data_bdev = pt->data_dev->bdev;
2690 struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2691 const char *reason = NULL;
2692 char buf[BDEVNAME_SIZE];
2693
2694 if (!pt->adjusted_pf.discard_passdown)
2695 return;
2696
2697 if (!data_dev_supports_discard(pt))
2698 reason = "discard unsupported";
2699
2700 else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2701 reason = "max discard sectors smaller than a block";
2702
2703 if (reason) {
2704 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2705 pt->adjusted_pf.discard_passdown = false;
2706 }
2707}
2708
2709static int bind_control_target(struct pool *pool, struct dm_target *ti)
2710{
2711 struct pool_c *pt = ti->private;
2712
2713 /*
2714 * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2715 */
2716 enum pool_mode old_mode = get_pool_mode(pool);
2717 enum pool_mode new_mode = pt->adjusted_pf.mode;
2718
2719 /*
2720 * Don't change the pool's mode until set_pool_mode() below.
2721 * Otherwise the pool's process_* function pointers may
2722 * not match the desired pool mode.
2723 */
2724 pt->adjusted_pf.mode = old_mode;
2725
2726 pool->ti = ti;
2727 pool->pf = pt->adjusted_pf;
2728 pool->low_water_blocks = pt->low_water_blocks;
2729
2730 set_pool_mode(pool, new_mode);
2731
2732 return 0;
2733}
2734
2735static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2736{
2737 if (pool->ti == ti)
2738 pool->ti = NULL;
2739}
2740
2741/*----------------------------------------------------------------
2742 * Pool creation
2743 *--------------------------------------------------------------*/
2744/* Initialize pool features. */
2745static void pool_features_init(struct pool_features *pf)
2746{
2747 pf->mode = PM_WRITE;
2748 pf->zero_new_blocks = true;
2749 pf->discard_enabled = true;
2750 pf->discard_passdown = true;
2751 pf->error_if_no_space = false;
2752}
2753
2754static void __pool_destroy(struct pool *pool)
2755{
2756 __pool_table_remove(pool);
2757
2758 vfree(pool->cell_sort_array);
2759 if (dm_pool_metadata_close(pool->pmd) < 0)
2760 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2761
2762 dm_bio_prison_destroy(pool->prison);
2763 dm_kcopyd_client_destroy(pool->copier);
2764
2765 if (pool->wq)
2766 destroy_workqueue(pool->wq);
2767
2768 if (pool->next_mapping)
2769 mempool_free(pool->next_mapping, pool->mapping_pool);
2770 mempool_destroy(pool->mapping_pool);
2771 dm_deferred_set_destroy(pool->shared_read_ds);
2772 dm_deferred_set_destroy(pool->all_io_ds);
2773 kfree(pool);
2774}
2775
2776static struct kmem_cache *_new_mapping_cache;
2777
2778static struct pool *pool_create(struct mapped_device *pool_md,
2779 struct block_device *metadata_dev,
2780 unsigned long block_size,
2781 int read_only, char **error)
2782{
2783 int r;
2784 void *err_p;
2785 struct pool *pool;
2786 struct dm_pool_metadata *pmd;
2787 bool format_device = read_only ? false : true;
2788
2789 pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2790 if (IS_ERR(pmd)) {
2791 *error = "Error creating metadata object";
2792 return (struct pool *)pmd;
2793 }
2794
2795 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2796 if (!pool) {
2797 *error = "Error allocating memory for pool";
2798 err_p = ERR_PTR(-ENOMEM);
2799 goto bad_pool;
2800 }
2801
2802 pool->pmd = pmd;
2803 pool->sectors_per_block = block_size;
2804 if (block_size & (block_size - 1))
2805 pool->sectors_per_block_shift = -1;
2806 else
2807 pool->sectors_per_block_shift = __ffs(block_size);
2808 pool->low_water_blocks = 0;
2809 pool_features_init(&pool->pf);
2810 pool->prison = dm_bio_prison_create();
2811 if (!pool->prison) {
2812 *error = "Error creating pool's bio prison";
2813 err_p = ERR_PTR(-ENOMEM);
2814 goto bad_prison;
2815 }
2816
2817 pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2818 if (IS_ERR(pool->copier)) {
2819 r = PTR_ERR(pool->copier);
2820 *error = "Error creating pool's kcopyd client";
2821 err_p = ERR_PTR(r);
2822 goto bad_kcopyd_client;
2823 }
2824
2825 /*
2826 * Create singlethreaded workqueue that will service all devices
2827 * that use this metadata.
2828 */
2829 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2830 if (!pool->wq) {
2831 *error = "Error creating pool's workqueue";
2832 err_p = ERR_PTR(-ENOMEM);
2833 goto bad_wq;
2834 }
2835
2836 throttle_init(&pool->throttle);
2837 INIT_WORK(&pool->worker, do_worker);
2838 INIT_DELAYED_WORK(&pool->waker, do_waker);
2839 INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2840 spin_lock_init(&pool->lock);
2841 bio_list_init(&pool->deferred_flush_bios);
2842 INIT_LIST_HEAD(&pool->prepared_mappings);
2843 INIT_LIST_HEAD(&pool->prepared_discards);
2844 INIT_LIST_HEAD(&pool->active_thins);
2845 pool->low_water_triggered = false;
2846 pool->suspended = true;
2847 pool->out_of_data_space = false;
2848
2849 pool->shared_read_ds = dm_deferred_set_create();
2850 if (!pool->shared_read_ds) {
2851 *error = "Error creating pool's shared read deferred set";
2852 err_p = ERR_PTR(-ENOMEM);
2853 goto bad_shared_read_ds;
2854 }
2855
2856 pool->all_io_ds = dm_deferred_set_create();
2857 if (!pool->all_io_ds) {
2858 *error = "Error creating pool's all io deferred set";
2859 err_p = ERR_PTR(-ENOMEM);
2860 goto bad_all_io_ds;
2861 }
2862
2863 pool->next_mapping = NULL;
2864 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2865 _new_mapping_cache);
2866 if (!pool->mapping_pool) {
2867 *error = "Error creating pool's mapping mempool";
2868 err_p = ERR_PTR(-ENOMEM);
2869 goto bad_mapping_pool;
2870 }
2871
2872 pool->cell_sort_array = vmalloc(sizeof(*pool->cell_sort_array) * CELL_SORT_ARRAY_SIZE);
2873 if (!pool->cell_sort_array) {
2874 *error = "Error allocating cell sort array";
2875 err_p = ERR_PTR(-ENOMEM);
2876 goto bad_sort_array;
2877 }
2878
2879 pool->ref_count = 1;
2880 pool->last_commit_jiffies = jiffies;
2881 pool->pool_md = pool_md;
2882 pool->md_dev = metadata_dev;
2883 __pool_table_insert(pool);
2884
2885 return pool;
2886
2887bad_sort_array:
2888 mempool_destroy(pool->mapping_pool);
2889bad_mapping_pool:
2890 dm_deferred_set_destroy(pool->all_io_ds);
2891bad_all_io_ds:
2892 dm_deferred_set_destroy(pool->shared_read_ds);
2893bad_shared_read_ds:
2894 destroy_workqueue(pool->wq);
2895bad_wq:
2896 dm_kcopyd_client_destroy(pool->copier);
2897bad_kcopyd_client:
2898 dm_bio_prison_destroy(pool->prison);
2899bad_prison:
2900 kfree(pool);
2901bad_pool:
2902 if (dm_pool_metadata_close(pmd))
2903 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2904
2905 return err_p;
2906}
2907
2908static void __pool_inc(struct pool *pool)
2909{
2910 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2911 pool->ref_count++;
2912}
2913
2914static void __pool_dec(struct pool *pool)
2915{
2916 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2917 BUG_ON(!pool->ref_count);
2918 if (!--pool->ref_count)
2919 __pool_destroy(pool);
2920}
2921
2922static struct pool *__pool_find(struct mapped_device *pool_md,
2923 struct block_device *metadata_dev,
2924 unsigned long block_size, int read_only,
2925 char **error, int *created)
2926{
2927 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2928
2929 if (pool) {
2930 if (pool->pool_md != pool_md) {
2931 *error = "metadata device already in use by a pool";
2932 return ERR_PTR(-EBUSY);
2933 }
2934 __pool_inc(pool);
2935
2936 } else {
2937 pool = __pool_table_lookup(pool_md);
2938 if (pool) {
2939 if (pool->md_dev != metadata_dev) {
2940 *error = "different pool cannot replace a pool";
2941 return ERR_PTR(-EINVAL);
2942 }
2943 __pool_inc(pool);
2944
2945 } else {
2946 pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
2947 *created = 1;
2948 }
2949 }
2950
2951 return pool;
2952}
2953
2954/*----------------------------------------------------------------
2955 * Pool target methods
2956 *--------------------------------------------------------------*/
2957static void pool_dtr(struct dm_target *ti)
2958{
2959 struct pool_c *pt = ti->private;
2960
2961 mutex_lock(&dm_thin_pool_table.mutex);
2962
2963 unbind_control_target(pt->pool, ti);
2964 __pool_dec(pt->pool);
2965 dm_put_device(ti, pt->metadata_dev);
2966 dm_put_device(ti, pt->data_dev);
2967 kfree(pt);
2968
2969 mutex_unlock(&dm_thin_pool_table.mutex);
2970}
2971
2972static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
2973 struct dm_target *ti)
2974{
2975 int r;
2976 unsigned argc;
2977 const char *arg_name;
2978
2979 static struct dm_arg _args[] = {
2980 {0, 4, "Invalid number of pool feature arguments"},
2981 };
2982
2983 /*
2984 * No feature arguments supplied.
2985 */
2986 if (!as->argc)
2987 return 0;
2988
2989 r = dm_read_arg_group(_args, as, &argc, &ti->error);
2990 if (r)
2991 return -EINVAL;
2992
2993 while (argc && !r) {
2994 arg_name = dm_shift_arg(as);
2995 argc--;
2996
2997 if (!strcasecmp(arg_name, "skip_block_zeroing"))
2998 pf->zero_new_blocks = false;
2999
3000 else if (!strcasecmp(arg_name, "ignore_discard"))
3001 pf->discard_enabled = false;
3002
3003 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3004 pf->discard_passdown = false;
3005
3006 else if (!strcasecmp(arg_name, "read_only"))
3007 pf->mode = PM_READ_ONLY;
3008
3009 else if (!strcasecmp(arg_name, "error_if_no_space"))
3010 pf->error_if_no_space = true;
3011
3012 else {
3013 ti->error = "Unrecognised pool feature requested";
3014 r = -EINVAL;
3015 break;
3016 }
3017 }
3018
3019 return r;
3020}
3021
3022static void metadata_low_callback(void *context)
3023{
3024 struct pool *pool = context;
3025
3026 DMWARN("%s: reached low water mark for metadata device: sending event.",
3027 dm_device_name(pool->pool_md));
3028
3029 dm_table_event(pool->ti->table);
3030}
3031
3032static sector_t get_dev_size(struct block_device *bdev)
3033{
3034 return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3035}
3036
3037static void warn_if_metadata_device_too_big(struct block_device *bdev)
3038{
3039 sector_t metadata_dev_size = get_dev_size(bdev);
3040 char buffer[BDEVNAME_SIZE];
3041
3042 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3043 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3044 bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3045}
3046
3047static sector_t get_metadata_dev_size(struct block_device *bdev)
3048{
3049 sector_t metadata_dev_size = get_dev_size(bdev);
3050
3051 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3052 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3053
3054 return metadata_dev_size;
3055}
3056
3057static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3058{
3059 sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3060
3061 sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3062
3063 return metadata_dev_size;
3064}
3065
3066/*
3067 * When a metadata threshold is crossed a dm event is triggered, and
3068 * userland should respond by growing the metadata device. We could let
3069 * userland set the threshold, like we do with the data threshold, but I'm
3070 * not sure they know enough to do this well.
3071 */
3072static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3073{
3074 /*
3075 * 4M is ample for all ops with the possible exception of thin
3076 * device deletion which is harmless if it fails (just retry the
3077 * delete after you've grown the device).
3078 */
3079 dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3080 return min((dm_block_t)1024ULL /* 4M */, quarter);
3081}
3082
3083/*
3084 * thin-pool <metadata dev> <data dev>
3085 * <data block size (sectors)>
3086 * <low water mark (blocks)>
3087 * [<#feature args> [<arg>]*]
3088 *
3089 * Optional feature arguments are:
3090 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3091 * ignore_discard: disable discard
3092 * no_discard_passdown: don't pass discards down to the data device
3093 * read_only: Don't allow any changes to be made to the pool metadata.
3094 * error_if_no_space: error IOs, instead of queueing, if no space.
3095 */
3096static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3097{
3098 int r, pool_created = 0;
3099 struct pool_c *pt;
3100 struct pool *pool;
3101 struct pool_features pf;
3102 struct dm_arg_set as;
3103 struct dm_dev *data_dev;
3104 unsigned long block_size;
3105 dm_block_t low_water_blocks;
3106 struct dm_dev *metadata_dev;
3107 fmode_t metadata_mode;
3108
3109 /*
3110 * FIXME Remove validation from scope of lock.
3111 */
3112 mutex_lock(&dm_thin_pool_table.mutex);
3113
3114 if (argc < 4) {
3115 ti->error = "Invalid argument count";
3116 r = -EINVAL;
3117 goto out_unlock;
3118 }
3119
3120 as.argc = argc;
3121 as.argv = argv;
3122
3123 /*
3124 * Set default pool features.
3125 */
3126 pool_features_init(&pf);
3127
3128 dm_consume_args(&as, 4);
3129 r = parse_pool_features(&as, &pf, ti);
3130 if (r)
3131 goto out_unlock;
3132
3133 metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3134 r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3135 if (r) {
3136 ti->error = "Error opening metadata block device";
3137 goto out_unlock;
3138 }
3139 warn_if_metadata_device_too_big(metadata_dev->bdev);
3140
3141 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3142 if (r) {
3143 ti->error = "Error getting data device";
3144 goto out_metadata;
3145 }
3146
3147 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3148 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3149 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3150 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3151 ti->error = "Invalid block size";
3152 r = -EINVAL;
3153 goto out;
3154 }
3155
3156 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3157 ti->error = "Invalid low water mark";
3158 r = -EINVAL;
3159 goto out;
3160 }
3161
3162 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3163 if (!pt) {
3164 r = -ENOMEM;
3165 goto out;
3166 }
3167
3168 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3169 block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3170 if (IS_ERR(pool)) {
3171 r = PTR_ERR(pool);
3172 goto out_free_pt;
3173 }
3174
3175 /*
3176 * 'pool_created' reflects whether this is the first table load.
3177 * Top level discard support is not allowed to be changed after
3178 * initial load. This would require a pool reload to trigger thin
3179 * device changes.
3180 */
3181 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3182 ti->error = "Discard support cannot be disabled once enabled";
3183 r = -EINVAL;
3184 goto out_flags_changed;
3185 }
3186
3187 pt->pool = pool;
3188 pt->ti = ti;
3189 pt->metadata_dev = metadata_dev;
3190 pt->data_dev = data_dev;
3191 pt->low_water_blocks = low_water_blocks;
3192 pt->adjusted_pf = pt->requested_pf = pf;
3193 ti->num_flush_bios = 1;
3194
3195 /*
3196 * Only need to enable discards if the pool should pass
3197 * them down to the data device. The thin device's discard
3198 * processing will cause mappings to be removed from the btree.
3199 */
3200 ti->discard_zeroes_data_unsupported = true;
3201 if (pf.discard_enabled && pf.discard_passdown) {
3202 ti->num_discard_bios = 1;
3203
3204 /*
3205 * Setting 'discards_supported' circumvents the normal
3206 * stacking of discard limits (this keeps the pool and
3207 * thin devices' discard limits consistent).
3208 */
3209 ti->discards_supported = true;
3210 }
3211 ti->private = pt;
3212
3213 r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3214 calc_metadata_threshold(pt),
3215 metadata_low_callback,
3216 pool);
3217 if (r)
3218 goto out_flags_changed;
3219
3220 pt->callbacks.congested_fn = pool_is_congested;
3221 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3222
3223 mutex_unlock(&dm_thin_pool_table.mutex);
3224
3225 return 0;
3226
3227out_flags_changed:
3228 __pool_dec(pool);
3229out_free_pt:
3230 kfree(pt);
3231out:
3232 dm_put_device(ti, data_dev);
3233out_metadata:
3234 dm_put_device(ti, metadata_dev);
3235out_unlock:
3236 mutex_unlock(&dm_thin_pool_table.mutex);
3237
3238 return r;
3239}
3240
3241static int pool_map(struct dm_target *ti, struct bio *bio)
3242{
3243 int r;
3244 struct pool_c *pt = ti->private;
3245 struct pool *pool = pt->pool;
3246 unsigned long flags;
3247
3248 /*
3249 * As this is a singleton target, ti->begin is always zero.
3250 */
3251 spin_lock_irqsave(&pool->lock, flags);
3252 bio->bi_bdev = pt->data_dev->bdev;
3253 r = DM_MAPIO_REMAPPED;
3254 spin_unlock_irqrestore(&pool->lock, flags);
3255
3256 return r;
3257}
3258
3259static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3260{
3261 int r;
3262 struct pool_c *pt = ti->private;
3263 struct pool *pool = pt->pool;
3264 sector_t data_size = ti->len;
3265 dm_block_t sb_data_size;
3266
3267 *need_commit = false;
3268
3269 (void) sector_div(data_size, pool->sectors_per_block);
3270
3271 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3272 if (r) {
3273 DMERR("%s: failed to retrieve data device size",
3274 dm_device_name(pool->pool_md));
3275 return r;
3276 }
3277
3278 if (data_size < sb_data_size) {
3279 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3280 dm_device_name(pool->pool_md),
3281 (unsigned long long)data_size, sb_data_size);
3282 return -EINVAL;
3283
3284 } else if (data_size > sb_data_size) {
3285 if (dm_pool_metadata_needs_check(pool->pmd)) {
3286 DMERR("%s: unable to grow the data device until repaired.",
3287 dm_device_name(pool->pool_md));
3288 return 0;
3289 }
3290
3291 if (sb_data_size)
3292 DMINFO("%s: growing the data device from %llu to %llu blocks",
3293 dm_device_name(pool->pool_md),
3294 sb_data_size, (unsigned long long)data_size);
3295 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3296 if (r) {
3297 metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3298 return r;
3299 }
3300
3301 *need_commit = true;
3302 }
3303
3304 return 0;
3305}
3306
3307static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3308{
3309 int r;
3310 struct pool_c *pt = ti->private;
3311 struct pool *pool = pt->pool;
3312 dm_block_t metadata_dev_size, sb_metadata_dev_size;
3313
3314 *need_commit = false;
3315
3316 metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3317
3318 r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3319 if (r) {
3320 DMERR("%s: failed to retrieve metadata device size",
3321 dm_device_name(pool->pool_md));
3322 return r;
3323 }
3324
3325 if (metadata_dev_size < sb_metadata_dev_size) {
3326 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3327 dm_device_name(pool->pool_md),
3328 metadata_dev_size, sb_metadata_dev_size);
3329 return -EINVAL;
3330
3331 } else if (metadata_dev_size > sb_metadata_dev_size) {
3332 if (dm_pool_metadata_needs_check(pool->pmd)) {
3333 DMERR("%s: unable to grow the metadata device until repaired.",
3334 dm_device_name(pool->pool_md));
3335 return 0;
3336 }
3337
3338 warn_if_metadata_device_too_big(pool->md_dev);
3339 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3340 dm_device_name(pool->pool_md),
3341 sb_metadata_dev_size, metadata_dev_size);
3342 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3343 if (r) {
3344 metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3345 return r;
3346 }
3347
3348 *need_commit = true;
3349 }
3350
3351 return 0;
3352}
3353
3354/*
3355 * Retrieves the number of blocks of the data device from
3356 * the superblock and compares it to the actual device size,
3357 * thus resizing the data device in case it has grown.
3358 *
3359 * This both copes with opening preallocated data devices in the ctr
3360 * being followed by a resume
3361 * -and-
3362 * calling the resume method individually after userspace has
3363 * grown the data device in reaction to a table event.
3364 */
3365static int pool_preresume(struct dm_target *ti)
3366{
3367 int r;
3368 bool need_commit1, need_commit2;
3369 struct pool_c *pt = ti->private;
3370 struct pool *pool = pt->pool;
3371
3372 /*
3373 * Take control of the pool object.
3374 */
3375 r = bind_control_target(pool, ti);
3376 if (r)
3377 return r;
3378
3379 r = maybe_resize_data_dev(ti, &need_commit1);
3380 if (r)
3381 return r;
3382
3383 r = maybe_resize_metadata_dev(ti, &need_commit2);
3384 if (r)
3385 return r;
3386
3387 if (need_commit1 || need_commit2)
3388 (void) commit(pool);
3389
3390 return 0;
3391}
3392
3393static void pool_suspend_active_thins(struct pool *pool)
3394{
3395 struct thin_c *tc;
3396
3397 /* Suspend all active thin devices */
3398 tc = get_first_thin(pool);
3399 while (tc) {
3400 dm_internal_suspend_noflush(tc->thin_md);
3401 tc = get_next_thin(pool, tc);
3402 }
3403}
3404
3405static void pool_resume_active_thins(struct pool *pool)
3406{
3407 struct thin_c *tc;
3408
3409 /* Resume all active thin devices */
3410 tc = get_first_thin(pool);
3411 while (tc) {
3412 dm_internal_resume(tc->thin_md);
3413 tc = get_next_thin(pool, tc);
3414 }
3415}
3416
3417static void pool_resume(struct dm_target *ti)
3418{
3419 struct pool_c *pt = ti->private;
3420 struct pool *pool = pt->pool;
3421 unsigned long flags;
3422
3423 /*
3424 * Must requeue active_thins' bios and then resume
3425 * active_thins _before_ clearing 'suspend' flag.
3426 */
3427 requeue_bios(pool);
3428 pool_resume_active_thins(pool);
3429
3430 spin_lock_irqsave(&pool->lock, flags);
3431 pool->low_water_triggered = false;
3432 pool->suspended = false;
3433 spin_unlock_irqrestore(&pool->lock, flags);
3434
3435 do_waker(&pool->waker.work);
3436}
3437
3438static void pool_presuspend(struct dm_target *ti)
3439{
3440 struct pool_c *pt = ti->private;
3441 struct pool *pool = pt->pool;
3442 unsigned long flags;
3443
3444 spin_lock_irqsave(&pool->lock, flags);
3445 pool->suspended = true;
3446 spin_unlock_irqrestore(&pool->lock, flags);
3447
3448 pool_suspend_active_thins(pool);
3449}
3450
3451static void pool_presuspend_undo(struct dm_target *ti)
3452{
3453 struct pool_c *pt = ti->private;
3454 struct pool *pool = pt->pool;
3455 unsigned long flags;
3456
3457 pool_resume_active_thins(pool);
3458
3459 spin_lock_irqsave(&pool->lock, flags);
3460 pool->suspended = false;
3461 spin_unlock_irqrestore(&pool->lock, flags);
3462}
3463
3464static void pool_postsuspend(struct dm_target *ti)
3465{
3466 struct pool_c *pt = ti->private;
3467 struct pool *pool = pt->pool;
3468
3469 cancel_delayed_work_sync(&pool->waker);
3470 cancel_delayed_work_sync(&pool->no_space_timeout);
3471 flush_workqueue(pool->wq);
3472 (void) commit(pool);
3473}
3474
3475static int check_arg_count(unsigned argc, unsigned args_required)
3476{
3477 if (argc != args_required) {
3478 DMWARN("Message received with %u arguments instead of %u.",
3479 argc, args_required);
3480 return -EINVAL;
3481 }
3482
3483 return 0;
3484}
3485
3486static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3487{
3488 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3489 *dev_id <= MAX_DEV_ID)
3490 return 0;
3491
3492 if (warning)
3493 DMWARN("Message received with invalid device id: %s", arg);
3494
3495 return -EINVAL;
3496}
3497
3498static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3499{
3500 dm_thin_id dev_id;
3501 int r;
3502
3503 r = check_arg_count(argc, 2);
3504 if (r)
3505 return r;
3506
3507 r = read_dev_id(argv[1], &dev_id, 1);
3508 if (r)
3509 return r;
3510
3511 r = dm_pool_create_thin(pool->pmd, dev_id);
3512 if (r) {
3513 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3514 argv[1]);
3515 return r;
3516 }
3517
3518 return 0;
3519}
3520
3521static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3522{
3523 dm_thin_id dev_id;
3524 dm_thin_id origin_dev_id;
3525 int r;
3526
3527 r = check_arg_count(argc, 3);
3528 if (r)
3529 return r;
3530
3531 r = read_dev_id(argv[1], &dev_id, 1);
3532 if (r)
3533 return r;
3534
3535 r = read_dev_id(argv[2], &origin_dev_id, 1);
3536 if (r)
3537 return r;
3538
3539 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3540 if (r) {
3541 DMWARN("Creation of new snapshot %s of device %s failed.",
3542 argv[1], argv[2]);
3543 return r;
3544 }
3545
3546 return 0;
3547}
3548
3549static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3550{
3551 dm_thin_id dev_id;
3552 int r;
3553
3554 r = check_arg_count(argc, 2);
3555 if (r)
3556 return r;
3557
3558 r = read_dev_id(argv[1], &dev_id, 1);
3559 if (r)
3560 return r;
3561
3562 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3563 if (r)
3564 DMWARN("Deletion of thin device %s failed.", argv[1]);
3565
3566 return r;
3567}
3568
3569static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3570{
3571 dm_thin_id old_id, new_id;
3572 int r;
3573
3574 r = check_arg_count(argc, 3);
3575 if (r)
3576 return r;
3577
3578 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3579 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3580 return -EINVAL;
3581 }
3582
3583 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3584 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3585 return -EINVAL;
3586 }
3587
3588 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3589 if (r) {
3590 DMWARN("Failed to change transaction id from %s to %s.",
3591 argv[1], argv[2]);
3592 return r;
3593 }
3594
3595 return 0;
3596}
3597
3598static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3599{
3600 int r;
3601
3602 r = check_arg_count(argc, 1);
3603 if (r)
3604 return r;
3605
3606 (void) commit(pool);
3607
3608 r = dm_pool_reserve_metadata_snap(pool->pmd);
3609 if (r)
3610 DMWARN("reserve_metadata_snap message failed.");
3611
3612 return r;
3613}
3614
3615static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3616{
3617 int r;
3618
3619 r = check_arg_count(argc, 1);
3620 if (r)
3621 return r;
3622
3623 r = dm_pool_release_metadata_snap(pool->pmd);
3624 if (r)
3625 DMWARN("release_metadata_snap message failed.");
3626
3627 return r;
3628}
3629
3630/*
3631 * Messages supported:
3632 * create_thin <dev_id>
3633 * create_snap <dev_id> <origin_id>
3634 * delete <dev_id>
3635 * set_transaction_id <current_trans_id> <new_trans_id>
3636 * reserve_metadata_snap
3637 * release_metadata_snap
3638 */
3639static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3640{
3641 int r = -EINVAL;
3642 struct pool_c *pt = ti->private;
3643 struct pool *pool = pt->pool;
3644
3645 if (get_pool_mode(pool) >= PM_READ_ONLY) {
3646 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3647 dm_device_name(pool->pool_md));
3648 return -EOPNOTSUPP;
3649 }
3650
3651 if (!strcasecmp(argv[0], "create_thin"))
3652 r = process_create_thin_mesg(argc, argv, pool);
3653
3654 else if (!strcasecmp(argv[0], "create_snap"))
3655 r = process_create_snap_mesg(argc, argv, pool);
3656
3657 else if (!strcasecmp(argv[0], "delete"))
3658 r = process_delete_mesg(argc, argv, pool);
3659
3660 else if (!strcasecmp(argv[0], "set_transaction_id"))
3661 r = process_set_transaction_id_mesg(argc, argv, pool);
3662
3663 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3664 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3665
3666 else if (!strcasecmp(argv[0], "release_metadata_snap"))
3667 r = process_release_metadata_snap_mesg(argc, argv, pool);
3668
3669 else
3670 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3671
3672 if (!r)
3673 (void) commit(pool);
3674
3675 return r;
3676}
3677
3678static void emit_flags(struct pool_features *pf, char *result,
3679 unsigned sz, unsigned maxlen)
3680{
3681 unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3682 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3683 pf->error_if_no_space;
3684 DMEMIT("%u ", count);
3685
3686 if (!pf->zero_new_blocks)
3687 DMEMIT("skip_block_zeroing ");
3688
3689 if (!pf->discard_enabled)
3690 DMEMIT("ignore_discard ");
3691
3692 if (!pf->discard_passdown)
3693 DMEMIT("no_discard_passdown ");
3694
3695 if (pf->mode == PM_READ_ONLY)
3696 DMEMIT("read_only ");
3697
3698 if (pf->error_if_no_space)
3699 DMEMIT("error_if_no_space ");
3700}
3701
3702/*
3703 * Status line is:
3704 * <transaction id> <used metadata sectors>/<total metadata sectors>
3705 * <used data sectors>/<total data sectors> <held metadata root>
3706 * <pool mode> <discard config> <no space config> <needs_check>
3707 */
3708static void pool_status(struct dm_target *ti, status_type_t type,
3709 unsigned status_flags, char *result, unsigned maxlen)
3710{
3711 int r;
3712 unsigned sz = 0;
3713 uint64_t transaction_id;
3714 dm_block_t nr_free_blocks_data;
3715 dm_block_t nr_free_blocks_metadata;
3716 dm_block_t nr_blocks_data;
3717 dm_block_t nr_blocks_metadata;
3718 dm_block_t held_root;
3719 char buf[BDEVNAME_SIZE];
3720 char buf2[BDEVNAME_SIZE];
3721 struct pool_c *pt = ti->private;
3722 struct pool *pool = pt->pool;
3723
3724 switch (type) {
3725 case STATUSTYPE_INFO:
3726 if (get_pool_mode(pool) == PM_FAIL) {
3727 DMEMIT("Fail");
3728 break;
3729 }
3730
3731 /* Commit to ensure statistics aren't out-of-date */
3732 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3733 (void) commit(pool);
3734
3735 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3736 if (r) {
3737 DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3738 dm_device_name(pool->pool_md), r);
3739 goto err;
3740 }
3741
3742 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3743 if (r) {
3744 DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3745 dm_device_name(pool->pool_md), r);
3746 goto err;
3747 }
3748
3749 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3750 if (r) {
3751 DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3752 dm_device_name(pool->pool_md), r);
3753 goto err;
3754 }
3755
3756 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3757 if (r) {
3758 DMERR("%s: dm_pool_get_free_block_count returned %d",
3759 dm_device_name(pool->pool_md), r);
3760 goto err;
3761 }
3762
3763 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3764 if (r) {
3765 DMERR("%s: dm_pool_get_data_dev_size returned %d",
3766 dm_device_name(pool->pool_md), r);
3767 goto err;
3768 }
3769
3770 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3771 if (r) {
3772 DMERR("%s: dm_pool_get_metadata_snap returned %d",
3773 dm_device_name(pool->pool_md), r);
3774 goto err;
3775 }
3776
3777 DMEMIT("%llu %llu/%llu %llu/%llu ",
3778 (unsigned long long)transaction_id,
3779 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3780 (unsigned long long)nr_blocks_metadata,
3781 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3782 (unsigned long long)nr_blocks_data);
3783
3784 if (held_root)
3785 DMEMIT("%llu ", held_root);
3786 else
3787 DMEMIT("- ");
3788
3789 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3790 DMEMIT("out_of_data_space ");
3791 else if (pool->pf.mode == PM_READ_ONLY)
3792 DMEMIT("ro ");
3793 else
3794 DMEMIT("rw ");
3795
3796 if (!pool->pf.discard_enabled)
3797 DMEMIT("ignore_discard ");
3798 else if (pool->pf.discard_passdown)
3799 DMEMIT("discard_passdown ");
3800 else
3801 DMEMIT("no_discard_passdown ");
3802
3803 if (pool->pf.error_if_no_space)
3804 DMEMIT("error_if_no_space ");
3805 else
3806 DMEMIT("queue_if_no_space ");
3807
3808 if (dm_pool_metadata_needs_check(pool->pmd))
3809 DMEMIT("needs_check ");
3810 else
3811 DMEMIT("- ");
3812
3813 break;
3814
3815 case STATUSTYPE_TABLE:
3816 DMEMIT("%s %s %lu %llu ",
3817 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3818 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3819 (unsigned long)pool->sectors_per_block,
3820 (unsigned long long)pt->low_water_blocks);
3821 emit_flags(&pt->requested_pf, result, sz, maxlen);
3822 break;
3823 }
3824 return;
3825
3826err:
3827 DMEMIT("Error");
3828}
3829
3830static int pool_iterate_devices(struct dm_target *ti,
3831 iterate_devices_callout_fn fn, void *data)
3832{
3833 struct pool_c *pt = ti->private;
3834
3835 return fn(ti, pt->data_dev, 0, ti->len, data);
3836}
3837
3838static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3839{
3840 struct pool_c *pt = ti->private;
3841 struct pool *pool = pt->pool;
3842 sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3843
3844 /*
3845 * If max_sectors is smaller than pool->sectors_per_block adjust it
3846 * to the highest possible power-of-2 factor of pool->sectors_per_block.
3847 * This is especially beneficial when the pool's data device is a RAID
3848 * device that has a full stripe width that matches pool->sectors_per_block
3849 * -- because even though partial RAID stripe-sized IOs will be issued to a
3850 * single RAID stripe; when aggregated they will end on a full RAID stripe
3851 * boundary.. which avoids additional partial RAID stripe writes cascading
3852 */
3853 if (limits->max_sectors < pool->sectors_per_block) {
3854 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3855 if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3856 limits->max_sectors--;
3857 limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3858 }
3859 }
3860
3861 /*
3862 * If the system-determined stacked limits are compatible with the
3863 * pool's blocksize (io_opt is a factor) do not override them.
3864 */
3865 if (io_opt_sectors < pool->sectors_per_block ||
3866 !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3867 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3868 blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3869 else
3870 blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3871 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3872 }
3873
3874 /*
3875 * pt->adjusted_pf is a staging area for the actual features to use.
3876 * They get transferred to the live pool in bind_control_target()
3877 * called from pool_preresume().
3878 */
3879 if (!pt->adjusted_pf.discard_enabled) {
3880 /*
3881 * Must explicitly disallow stacking discard limits otherwise the
3882 * block layer will stack them if pool's data device has support.
3883 * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3884 * user to see that, so make sure to set all discard limits to 0.
3885 */
3886 limits->discard_granularity = 0;
3887 return;
3888 }
3889
3890 disable_passdown_if_not_supported(pt);
3891
3892 /*
3893 * The pool uses the same discard limits as the underlying data
3894 * device. DM core has already set this up.
3895 */
3896}
3897
3898static struct target_type pool_target = {
3899 .name = "thin-pool",
3900 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3901 DM_TARGET_IMMUTABLE,
3902 .version = {1, 18, 0},
3903 .module = THIS_MODULE,
3904 .ctr = pool_ctr,
3905 .dtr = pool_dtr,
3906 .map = pool_map,
3907 .presuspend = pool_presuspend,
3908 .presuspend_undo = pool_presuspend_undo,
3909 .postsuspend = pool_postsuspend,
3910 .preresume = pool_preresume,
3911 .resume = pool_resume,
3912 .message = pool_message,
3913 .status = pool_status,
3914 .iterate_devices = pool_iterate_devices,
3915 .io_hints = pool_io_hints,
3916};
3917
3918/*----------------------------------------------------------------
3919 * Thin target methods
3920 *--------------------------------------------------------------*/
3921static void thin_get(struct thin_c *tc)
3922{
3923 atomic_inc(&tc->refcount);
3924}
3925
3926static void thin_put(struct thin_c *tc)
3927{
3928 if (atomic_dec_and_test(&tc->refcount))
3929 complete(&tc->can_destroy);
3930}
3931
3932static void thin_dtr(struct dm_target *ti)
3933{
3934 struct thin_c *tc = ti->private;
3935 unsigned long flags;
3936
3937 spin_lock_irqsave(&tc->pool->lock, flags);
3938 list_del_rcu(&tc->list);
3939 spin_unlock_irqrestore(&tc->pool->lock, flags);
3940 synchronize_rcu();
3941
3942 thin_put(tc);
3943 wait_for_completion(&tc->can_destroy);
3944
3945 mutex_lock(&dm_thin_pool_table.mutex);
3946
3947 __pool_dec(tc->pool);
3948 dm_pool_close_thin_device(tc->td);
3949 dm_put_device(ti, tc->pool_dev);
3950 if (tc->origin_dev)
3951 dm_put_device(ti, tc->origin_dev);
3952 kfree(tc);
3953
3954 mutex_unlock(&dm_thin_pool_table.mutex);
3955}
3956
3957/*
3958 * Thin target parameters:
3959 *
3960 * <pool_dev> <dev_id> [origin_dev]
3961 *
3962 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
3963 * dev_id: the internal device identifier
3964 * origin_dev: a device external to the pool that should act as the origin
3965 *
3966 * If the pool device has discards disabled, they get disabled for the thin
3967 * device as well.
3968 */
3969static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
3970{
3971 int r;
3972 struct thin_c *tc;
3973 struct dm_dev *pool_dev, *origin_dev;
3974 struct mapped_device *pool_md;
3975 unsigned long flags;
3976
3977 mutex_lock(&dm_thin_pool_table.mutex);
3978
3979 if (argc != 2 && argc != 3) {
3980 ti->error = "Invalid argument count";
3981 r = -EINVAL;
3982 goto out_unlock;
3983 }
3984
3985 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
3986 if (!tc) {
3987 ti->error = "Out of memory";
3988 r = -ENOMEM;
3989 goto out_unlock;
3990 }
3991 tc->thin_md = dm_table_get_md(ti->table);
3992 spin_lock_init(&tc->lock);
3993 INIT_LIST_HEAD(&tc->deferred_cells);
3994 bio_list_init(&tc->deferred_bio_list);
3995 bio_list_init(&tc->retry_on_resume_list);
3996 tc->sort_bio_list = RB_ROOT;
3997
3998 if (argc == 3) {
3999 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4000 if (r) {
4001 ti->error = "Error opening origin device";
4002 goto bad_origin_dev;
4003 }
4004 tc->origin_dev = origin_dev;
4005 }
4006
4007 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4008 if (r) {
4009 ti->error = "Error opening pool device";
4010 goto bad_pool_dev;
4011 }
4012 tc->pool_dev = pool_dev;
4013
4014 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4015 ti->error = "Invalid device id";
4016 r = -EINVAL;
4017 goto bad_common;
4018 }
4019
4020 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4021 if (!pool_md) {
4022 ti->error = "Couldn't get pool mapped device";
4023 r = -EINVAL;
4024 goto bad_common;
4025 }
4026
4027 tc->pool = __pool_table_lookup(pool_md);
4028 if (!tc->pool) {
4029 ti->error = "Couldn't find pool object";
4030 r = -EINVAL;
4031 goto bad_pool_lookup;
4032 }
4033 __pool_inc(tc->pool);
4034
4035 if (get_pool_mode(tc->pool) == PM_FAIL) {
4036 ti->error = "Couldn't open thin device, Pool is in fail mode";
4037 r = -EINVAL;
4038 goto bad_pool;
4039 }
4040
4041 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4042 if (r) {
4043 ti->error = "Couldn't open thin internal device";
4044 goto bad_pool;
4045 }
4046
4047 r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4048 if (r)
4049 goto bad;
4050
4051 ti->num_flush_bios = 1;
4052 ti->flush_supported = true;
4053 ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4054
4055 /* In case the pool supports discards, pass them on. */
4056 ti->discard_zeroes_data_unsupported = true;
4057 if (tc->pool->pf.discard_enabled) {
4058 ti->discards_supported = true;
4059 ti->num_discard_bios = 1;
4060 ti->split_discard_bios = false;
4061 }
4062
4063 mutex_unlock(&dm_thin_pool_table.mutex);
4064
4065 spin_lock_irqsave(&tc->pool->lock, flags);
4066 if (tc->pool->suspended) {
4067 spin_unlock_irqrestore(&tc->pool->lock, flags);
4068 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4069 ti->error = "Unable to activate thin device while pool is suspended";
4070 r = -EINVAL;
4071 goto bad;
4072 }
4073 atomic_set(&tc->refcount, 1);
4074 init_completion(&tc->can_destroy);
4075 list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4076 spin_unlock_irqrestore(&tc->pool->lock, flags);
4077 /*
4078 * This synchronize_rcu() call is needed here otherwise we risk a
4079 * wake_worker() call finding no bios to process (because the newly
4080 * added tc isn't yet visible). So this reduces latency since we
4081 * aren't then dependent on the periodic commit to wake_worker().
4082 */
4083 synchronize_rcu();
4084
4085 dm_put(pool_md);
4086
4087 return 0;
4088
4089bad:
4090 dm_pool_close_thin_device(tc->td);
4091bad_pool:
4092 __pool_dec(tc->pool);
4093bad_pool_lookup:
4094 dm_put(pool_md);
4095bad_common:
4096 dm_put_device(ti, tc->pool_dev);
4097bad_pool_dev:
4098 if (tc->origin_dev)
4099 dm_put_device(ti, tc->origin_dev);
4100bad_origin_dev:
4101 kfree(tc);
4102out_unlock:
4103 mutex_unlock(&dm_thin_pool_table.mutex);
4104
4105 return r;
4106}
4107
4108static int thin_map(struct dm_target *ti, struct bio *bio)
4109{
4110 bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4111
4112 return thin_bio_map(ti, bio);
4113}
4114
4115static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
4116{
4117 unsigned long flags;
4118 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4119 struct list_head work;
4120 struct dm_thin_new_mapping *m, *tmp;
4121 struct pool *pool = h->tc->pool;
4122
4123 if (h->shared_read_entry) {
4124 INIT_LIST_HEAD(&work);
4125 dm_deferred_entry_dec(h->shared_read_entry, &work);
4126
4127 spin_lock_irqsave(&pool->lock, flags);
4128 list_for_each_entry_safe(m, tmp, &work, list) {
4129 list_del(&m->list);
4130 __complete_mapping_preparation(m);
4131 }
4132 spin_unlock_irqrestore(&pool->lock, flags);
4133 }
4134
4135 if (h->all_io_entry) {
4136 INIT_LIST_HEAD(&work);
4137 dm_deferred_entry_dec(h->all_io_entry, &work);
4138 if (!list_empty(&work)) {
4139 spin_lock_irqsave(&pool->lock, flags);
4140 list_for_each_entry_safe(m, tmp, &work, list)
4141 list_add_tail(&m->list, &pool->prepared_discards);
4142 spin_unlock_irqrestore(&pool->lock, flags);
4143 wake_worker(pool);
4144 }
4145 }
4146
4147 if (h->cell)
4148 cell_defer_no_holder(h->tc, h->cell);
4149
4150 return 0;
4151}
4152
4153static void thin_presuspend(struct dm_target *ti)
4154{
4155 struct thin_c *tc = ti->private;
4156
4157 if (dm_noflush_suspending(ti))
4158 noflush_work(tc, do_noflush_start);
4159}
4160
4161static void thin_postsuspend(struct dm_target *ti)
4162{
4163 struct thin_c *tc = ti->private;
4164
4165 /*
4166 * The dm_noflush_suspending flag has been cleared by now, so
4167 * unfortunately we must always run this.
4168 */
4169 noflush_work(tc, do_noflush_stop);
4170}
4171
4172static int thin_preresume(struct dm_target *ti)
4173{
4174 struct thin_c *tc = ti->private;
4175
4176 if (tc->origin_dev)
4177 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4178
4179 return 0;
4180}
4181
4182/*
4183 * <nr mapped sectors> <highest mapped sector>
4184 */
4185static void thin_status(struct dm_target *ti, status_type_t type,
4186 unsigned status_flags, char *result, unsigned maxlen)
4187{
4188 int r;
4189 ssize_t sz = 0;
4190 dm_block_t mapped, highest;
4191 char buf[BDEVNAME_SIZE];
4192 struct thin_c *tc = ti->private;
4193
4194 if (get_pool_mode(tc->pool) == PM_FAIL) {
4195 DMEMIT("Fail");
4196 return;
4197 }
4198
4199 if (!tc->td)
4200 DMEMIT("-");
4201 else {
4202 switch (type) {
4203 case STATUSTYPE_INFO:
4204 r = dm_thin_get_mapped_count(tc->td, &mapped);
4205 if (r) {
4206 DMERR("dm_thin_get_mapped_count returned %d", r);
4207 goto err;
4208 }
4209
4210 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4211 if (r < 0) {
4212 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4213 goto err;
4214 }
4215
4216 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4217 if (r)
4218 DMEMIT("%llu", ((highest + 1) *
4219 tc->pool->sectors_per_block) - 1);
4220 else
4221 DMEMIT("-");
4222 break;
4223
4224 case STATUSTYPE_TABLE:
4225 DMEMIT("%s %lu",
4226 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4227 (unsigned long) tc->dev_id);
4228 if (tc->origin_dev)
4229 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4230 break;
4231 }
4232 }
4233
4234 return;
4235
4236err:
4237 DMEMIT("Error");
4238}
4239
4240static int thin_iterate_devices(struct dm_target *ti,
4241 iterate_devices_callout_fn fn, void *data)
4242{
4243 sector_t blocks;
4244 struct thin_c *tc = ti->private;
4245 struct pool *pool = tc->pool;
4246
4247 /*
4248 * We can't call dm_pool_get_data_dev_size() since that blocks. So
4249 * we follow a more convoluted path through to the pool's target.
4250 */
4251 if (!pool->ti)
4252 return 0; /* nothing is bound */
4253
4254 blocks = pool->ti->len;
4255 (void) sector_div(blocks, pool->sectors_per_block);
4256 if (blocks)
4257 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4258
4259 return 0;
4260}
4261
4262static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4263{
4264 struct thin_c *tc = ti->private;
4265 struct pool *pool = tc->pool;
4266
4267 if (!pool->pf.discard_enabled)
4268 return;
4269
4270 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4271 limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4272}
4273
4274static struct target_type thin_target = {
4275 .name = "thin",
4276 .version = {1, 18, 0},
4277 .module = THIS_MODULE,
4278 .ctr = thin_ctr,
4279 .dtr = thin_dtr,
4280 .map = thin_map,
4281 .end_io = thin_endio,
4282 .preresume = thin_preresume,
4283 .presuspend = thin_presuspend,
4284 .postsuspend = thin_postsuspend,
4285 .status = thin_status,
4286 .iterate_devices = thin_iterate_devices,
4287 .io_hints = thin_io_hints,
4288};
4289
4290/*----------------------------------------------------------------*/
4291
4292static int __init dm_thin_init(void)
4293{
4294 int r;
4295
4296 pool_table_init();
4297
4298 r = dm_register_target(&thin_target);
4299 if (r)
4300 return r;
4301
4302 r = dm_register_target(&pool_target);
4303 if (r)
4304 goto bad_pool_target;
4305
4306 r = -ENOMEM;
4307
4308 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4309 if (!_new_mapping_cache)
4310 goto bad_new_mapping_cache;
4311
4312 return 0;
4313
4314bad_new_mapping_cache:
4315 dm_unregister_target(&pool_target);
4316bad_pool_target:
4317 dm_unregister_target(&thin_target);
4318
4319 return r;
4320}
4321
4322static void dm_thin_exit(void)
4323{
4324 dm_unregister_target(&thin_target);
4325 dm_unregister_target(&pool_target);
4326
4327 kmem_cache_destroy(_new_mapping_cache);
4328}
4329
4330module_init(dm_thin_init);
4331module_exit(dm_thin_exit);
4332
4333module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4334MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4335
4336MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4337MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4338MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2011 Red Hat UK.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm-thin-metadata.h"
8
9#include <linux/device-mapper.h>
10#include <linux/dm-io.h>
11#include <linux/dm-kcopyd.h>
12#include <linux/list.h>
13#include <linux/init.h>
14#include <linux/module.h>
15#include <linux/slab.h>
16
17#define DM_MSG_PREFIX "thin"
18
19/*
20 * Tunable constants
21 */
22#define ENDIO_HOOK_POOL_SIZE 1024
23#define DEFERRED_SET_SIZE 64
24#define MAPPING_POOL_SIZE 1024
25#define PRISON_CELLS 1024
26#define COMMIT_PERIOD HZ
27
28/*
29 * The block size of the device holding pool data must be
30 * between 64KB and 1GB.
31 */
32#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
33#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
34
35/*
36 * Device id is restricted to 24 bits.
37 */
38#define MAX_DEV_ID ((1 << 24) - 1)
39
40/*
41 * How do we handle breaking sharing of data blocks?
42 * =================================================
43 *
44 * We use a standard copy-on-write btree to store the mappings for the
45 * devices (note I'm talking about copy-on-write of the metadata here, not
46 * the data). When you take an internal snapshot you clone the root node
47 * of the origin btree. After this there is no concept of an origin or a
48 * snapshot. They are just two device trees that happen to point to the
49 * same data blocks.
50 *
51 * When we get a write in we decide if it's to a shared data block using
52 * some timestamp magic. If it is, we have to break sharing.
53 *
54 * Let's say we write to a shared block in what was the origin. The
55 * steps are:
56 *
57 * i) plug io further to this physical block. (see bio_prison code).
58 *
59 * ii) quiesce any read io to that shared data block. Obviously
60 * including all devices that share this block. (see deferred_set code)
61 *
62 * iii) copy the data block to a newly allocate block. This step can be
63 * missed out if the io covers the block. (schedule_copy).
64 *
65 * iv) insert the new mapping into the origin's btree
66 * (process_prepared_mapping). This act of inserting breaks some
67 * sharing of btree nodes between the two devices. Breaking sharing only
68 * effects the btree of that specific device. Btrees for the other
69 * devices that share the block never change. The btree for the origin
70 * device as it was after the last commit is untouched, ie. we're using
71 * persistent data structures in the functional programming sense.
72 *
73 * v) unplug io to this physical block, including the io that triggered
74 * the breaking of sharing.
75 *
76 * Steps (ii) and (iii) occur in parallel.
77 *
78 * The metadata _doesn't_ need to be committed before the io continues. We
79 * get away with this because the io is always written to a _new_ block.
80 * If there's a crash, then:
81 *
82 * - The origin mapping will point to the old origin block (the shared
83 * one). This will contain the data as it was before the io that triggered
84 * the breaking of sharing came in.
85 *
86 * - The snap mapping still points to the old block. As it would after
87 * the commit.
88 *
89 * The downside of this scheme is the timestamp magic isn't perfect, and
90 * will continue to think that data block in the snapshot device is shared
91 * even after the write to the origin has broken sharing. I suspect data
92 * blocks will typically be shared by many different devices, so we're
93 * breaking sharing n + 1 times, rather than n, where n is the number of
94 * devices that reference this data block. At the moment I think the
95 * benefits far, far outweigh the disadvantages.
96 */
97
98/*----------------------------------------------------------------*/
99
100/*
101 * Sometimes we can't deal with a bio straight away. We put them in prison
102 * where they can't cause any mischief. Bios are put in a cell identified
103 * by a key, multiple bios can be in the same cell. When the cell is
104 * subsequently unlocked the bios become available.
105 */
106struct bio_prison;
107
108struct cell_key {
109 int virtual;
110 dm_thin_id dev;
111 dm_block_t block;
112};
113
114struct dm_bio_prison_cell {
115 struct hlist_node list;
116 struct bio_prison *prison;
117 struct cell_key key;
118 struct bio *holder;
119 struct bio_list bios;
120};
121
122struct bio_prison {
123 spinlock_t lock;
124 mempool_t *cell_pool;
125
126 unsigned nr_buckets;
127 unsigned hash_mask;
128 struct hlist_head *cells;
129};
130
131static uint32_t calc_nr_buckets(unsigned nr_cells)
132{
133 uint32_t n = 128;
134
135 nr_cells /= 4;
136 nr_cells = min(nr_cells, 8192u);
137
138 while (n < nr_cells)
139 n <<= 1;
140
141 return n;
142}
143
144static struct kmem_cache *_cell_cache;
145
146/*
147 * @nr_cells should be the number of cells you want in use _concurrently_.
148 * Don't confuse it with the number of distinct keys.
149 */
150static struct bio_prison *prison_create(unsigned nr_cells)
151{
152 unsigned i;
153 uint32_t nr_buckets = calc_nr_buckets(nr_cells);
154 size_t len = sizeof(struct bio_prison) +
155 (sizeof(struct hlist_head) * nr_buckets);
156 struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
157
158 if (!prison)
159 return NULL;
160
161 spin_lock_init(&prison->lock);
162 prison->cell_pool = mempool_create_slab_pool(nr_cells, _cell_cache);
163 if (!prison->cell_pool) {
164 kfree(prison);
165 return NULL;
166 }
167
168 prison->nr_buckets = nr_buckets;
169 prison->hash_mask = nr_buckets - 1;
170 prison->cells = (struct hlist_head *) (prison + 1);
171 for (i = 0; i < nr_buckets; i++)
172 INIT_HLIST_HEAD(prison->cells + i);
173
174 return prison;
175}
176
177static void prison_destroy(struct bio_prison *prison)
178{
179 mempool_destroy(prison->cell_pool);
180 kfree(prison);
181}
182
183static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
184{
185 const unsigned long BIG_PRIME = 4294967291UL;
186 uint64_t hash = key->block * BIG_PRIME;
187
188 return (uint32_t) (hash & prison->hash_mask);
189}
190
191static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
192{
193 return (lhs->virtual == rhs->virtual) &&
194 (lhs->dev == rhs->dev) &&
195 (lhs->block == rhs->block);
196}
197
198static struct dm_bio_prison_cell *__search_bucket(struct hlist_head *bucket,
199 struct cell_key *key)
200{
201 struct dm_bio_prison_cell *cell;
202 struct hlist_node *tmp;
203
204 hlist_for_each_entry(cell, tmp, bucket, list)
205 if (keys_equal(&cell->key, key))
206 return cell;
207
208 return NULL;
209}
210
211/*
212 * This may block if a new cell needs allocating. You must ensure that
213 * cells will be unlocked even if the calling thread is blocked.
214 *
215 * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
216 */
217static int bio_detain(struct bio_prison *prison, struct cell_key *key,
218 struct bio *inmate, struct dm_bio_prison_cell **ref)
219{
220 int r = 1;
221 unsigned long flags;
222 uint32_t hash = hash_key(prison, key);
223 struct dm_bio_prison_cell *cell, *cell2;
224
225 BUG_ON(hash > prison->nr_buckets);
226
227 spin_lock_irqsave(&prison->lock, flags);
228
229 cell = __search_bucket(prison->cells + hash, key);
230 if (cell) {
231 bio_list_add(&cell->bios, inmate);
232 goto out;
233 }
234
235 /*
236 * Allocate a new cell
237 */
238 spin_unlock_irqrestore(&prison->lock, flags);
239 cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
240 spin_lock_irqsave(&prison->lock, flags);
241
242 /*
243 * We've been unlocked, so we have to double check that
244 * nobody else has inserted this cell in the meantime.
245 */
246 cell = __search_bucket(prison->cells + hash, key);
247 if (cell) {
248 mempool_free(cell2, prison->cell_pool);
249 bio_list_add(&cell->bios, inmate);
250 goto out;
251 }
252
253 /*
254 * Use new cell.
255 */
256 cell = cell2;
257
258 cell->prison = prison;
259 memcpy(&cell->key, key, sizeof(cell->key));
260 cell->holder = inmate;
261 bio_list_init(&cell->bios);
262 hlist_add_head(&cell->list, prison->cells + hash);
263
264 r = 0;
265
266out:
267 spin_unlock_irqrestore(&prison->lock, flags);
268
269 *ref = cell;
270
271 return r;
272}
273
274/*
275 * @inmates must have been initialised prior to this call
276 */
277static void __cell_release(struct dm_bio_prison_cell *cell, struct bio_list *inmates)
278{
279 struct bio_prison *prison = cell->prison;
280
281 hlist_del(&cell->list);
282
283 if (inmates) {
284 bio_list_add(inmates, cell->holder);
285 bio_list_merge(inmates, &cell->bios);
286 }
287
288 mempool_free(cell, prison->cell_pool);
289}
290
291static void cell_release(struct dm_bio_prison_cell *cell, struct bio_list *bios)
292{
293 unsigned long flags;
294 struct bio_prison *prison = cell->prison;
295
296 spin_lock_irqsave(&prison->lock, flags);
297 __cell_release(cell, bios);
298 spin_unlock_irqrestore(&prison->lock, flags);
299}
300
301/*
302 * There are a couple of places where we put a bio into a cell briefly
303 * before taking it out again. In these situations we know that no other
304 * bio may be in the cell. This function releases the cell, and also does
305 * a sanity check.
306 */
307static void __cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
308{
309 BUG_ON(cell->holder != bio);
310 BUG_ON(!bio_list_empty(&cell->bios));
311
312 __cell_release(cell, NULL);
313}
314
315static void cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
316{
317 unsigned long flags;
318 struct bio_prison *prison = cell->prison;
319
320 spin_lock_irqsave(&prison->lock, flags);
321 __cell_release_singleton(cell, bio);
322 spin_unlock_irqrestore(&prison->lock, flags);
323}
324
325/*
326 * Sometimes we don't want the holder, just the additional bios.
327 */
328static void __cell_release_no_holder(struct dm_bio_prison_cell *cell,
329 struct bio_list *inmates)
330{
331 struct bio_prison *prison = cell->prison;
332
333 hlist_del(&cell->list);
334 bio_list_merge(inmates, &cell->bios);
335
336 mempool_free(cell, prison->cell_pool);
337}
338
339static void cell_release_no_holder(struct dm_bio_prison_cell *cell,
340 struct bio_list *inmates)
341{
342 unsigned long flags;
343 struct bio_prison *prison = cell->prison;
344
345 spin_lock_irqsave(&prison->lock, flags);
346 __cell_release_no_holder(cell, inmates);
347 spin_unlock_irqrestore(&prison->lock, flags);
348}
349
350static void cell_error(struct dm_bio_prison_cell *cell)
351{
352 struct bio_prison *prison = cell->prison;
353 struct bio_list bios;
354 struct bio *bio;
355 unsigned long flags;
356
357 bio_list_init(&bios);
358
359 spin_lock_irqsave(&prison->lock, flags);
360 __cell_release(cell, &bios);
361 spin_unlock_irqrestore(&prison->lock, flags);
362
363 while ((bio = bio_list_pop(&bios)))
364 bio_io_error(bio);
365}
366
367/*----------------------------------------------------------------*/
368
369/*
370 * We use the deferred set to keep track of pending reads to shared blocks.
371 * We do this to ensure the new mapping caused by a write isn't performed
372 * until these prior reads have completed. Otherwise the insertion of the
373 * new mapping could free the old block that the read bios are mapped to.
374 */
375
376struct deferred_set;
377struct deferred_entry {
378 struct deferred_set *ds;
379 unsigned count;
380 struct list_head work_items;
381};
382
383struct deferred_set {
384 spinlock_t lock;
385 unsigned current_entry;
386 unsigned sweeper;
387 struct deferred_entry entries[DEFERRED_SET_SIZE];
388};
389
390static void ds_init(struct deferred_set *ds)
391{
392 int i;
393
394 spin_lock_init(&ds->lock);
395 ds->current_entry = 0;
396 ds->sweeper = 0;
397 for (i = 0; i < DEFERRED_SET_SIZE; i++) {
398 ds->entries[i].ds = ds;
399 ds->entries[i].count = 0;
400 INIT_LIST_HEAD(&ds->entries[i].work_items);
401 }
402}
403
404static struct deferred_entry *ds_inc(struct deferred_set *ds)
405{
406 unsigned long flags;
407 struct deferred_entry *entry;
408
409 spin_lock_irqsave(&ds->lock, flags);
410 entry = ds->entries + ds->current_entry;
411 entry->count++;
412 spin_unlock_irqrestore(&ds->lock, flags);
413
414 return entry;
415}
416
417static unsigned ds_next(unsigned index)
418{
419 return (index + 1) % DEFERRED_SET_SIZE;
420}
421
422static void __sweep(struct deferred_set *ds, struct list_head *head)
423{
424 while ((ds->sweeper != ds->current_entry) &&
425 !ds->entries[ds->sweeper].count) {
426 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
427 ds->sweeper = ds_next(ds->sweeper);
428 }
429
430 if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
431 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
432}
433
434static void ds_dec(struct deferred_entry *entry, struct list_head *head)
435{
436 unsigned long flags;
437
438 spin_lock_irqsave(&entry->ds->lock, flags);
439 BUG_ON(!entry->count);
440 --entry->count;
441 __sweep(entry->ds, head);
442 spin_unlock_irqrestore(&entry->ds->lock, flags);
443}
444
445/*
446 * Returns 1 if deferred or 0 if no pending items to delay job.
447 */
448static int ds_add_work(struct deferred_set *ds, struct list_head *work)
449{
450 int r = 1;
451 unsigned long flags;
452 unsigned next_entry;
453
454 spin_lock_irqsave(&ds->lock, flags);
455 if ((ds->sweeper == ds->current_entry) &&
456 !ds->entries[ds->current_entry].count)
457 r = 0;
458 else {
459 list_add(work, &ds->entries[ds->current_entry].work_items);
460 next_entry = ds_next(ds->current_entry);
461 if (!ds->entries[next_entry].count)
462 ds->current_entry = next_entry;
463 }
464 spin_unlock_irqrestore(&ds->lock, flags);
465
466 return r;
467}
468
469/*----------------------------------------------------------------*/
470
471/*
472 * Key building.
473 */
474static void build_data_key(struct dm_thin_device *td,
475 dm_block_t b, struct cell_key *key)
476{
477 key->virtual = 0;
478 key->dev = dm_thin_dev_id(td);
479 key->block = b;
480}
481
482static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
483 struct cell_key *key)
484{
485 key->virtual = 1;
486 key->dev = dm_thin_dev_id(td);
487 key->block = b;
488}
489
490/*----------------------------------------------------------------*/
491
492/*
493 * A pool device ties together a metadata device and a data device. It
494 * also provides the interface for creating and destroying internal
495 * devices.
496 */
497struct dm_thin_new_mapping;
498
499struct pool_features {
500 unsigned zero_new_blocks:1;
501 unsigned discard_enabled:1;
502 unsigned discard_passdown:1;
503};
504
505struct pool {
506 struct list_head list;
507 struct dm_target *ti; /* Only set if a pool target is bound */
508
509 struct mapped_device *pool_md;
510 struct block_device *md_dev;
511 struct dm_pool_metadata *pmd;
512
513 uint32_t sectors_per_block;
514 unsigned block_shift;
515 dm_block_t offset_mask;
516 dm_block_t low_water_blocks;
517
518 struct pool_features pf;
519 unsigned low_water_triggered:1; /* A dm event has been sent */
520 unsigned no_free_space:1; /* A -ENOSPC warning has been issued */
521
522 struct bio_prison *prison;
523 struct dm_kcopyd_client *copier;
524
525 struct workqueue_struct *wq;
526 struct work_struct worker;
527 struct delayed_work waker;
528
529 unsigned ref_count;
530 unsigned long last_commit_jiffies;
531
532 spinlock_t lock;
533 struct bio_list deferred_bios;
534 struct bio_list deferred_flush_bios;
535 struct list_head prepared_mappings;
536 struct list_head prepared_discards;
537
538 struct bio_list retry_on_resume_list;
539
540 struct deferred_set shared_read_ds;
541 struct deferred_set all_io_ds;
542
543 struct dm_thin_new_mapping *next_mapping;
544 mempool_t *mapping_pool;
545 mempool_t *endio_hook_pool;
546};
547
548/*
549 * Target context for a pool.
550 */
551struct pool_c {
552 struct dm_target *ti;
553 struct pool *pool;
554 struct dm_dev *data_dev;
555 struct dm_dev *metadata_dev;
556 struct dm_target_callbacks callbacks;
557
558 dm_block_t low_water_blocks;
559 struct pool_features pf;
560};
561
562/*
563 * Target context for a thin.
564 */
565struct thin_c {
566 struct dm_dev *pool_dev;
567 struct dm_dev *origin_dev;
568 dm_thin_id dev_id;
569
570 struct pool *pool;
571 struct dm_thin_device *td;
572};
573
574/*----------------------------------------------------------------*/
575
576/*
577 * A global list of pools that uses a struct mapped_device as a key.
578 */
579static struct dm_thin_pool_table {
580 struct mutex mutex;
581 struct list_head pools;
582} dm_thin_pool_table;
583
584static void pool_table_init(void)
585{
586 mutex_init(&dm_thin_pool_table.mutex);
587 INIT_LIST_HEAD(&dm_thin_pool_table.pools);
588}
589
590static void __pool_table_insert(struct pool *pool)
591{
592 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
593 list_add(&pool->list, &dm_thin_pool_table.pools);
594}
595
596static void __pool_table_remove(struct pool *pool)
597{
598 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
599 list_del(&pool->list);
600}
601
602static struct pool *__pool_table_lookup(struct mapped_device *md)
603{
604 struct pool *pool = NULL, *tmp;
605
606 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
607
608 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
609 if (tmp->pool_md == md) {
610 pool = tmp;
611 break;
612 }
613 }
614
615 return pool;
616}
617
618static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
619{
620 struct pool *pool = NULL, *tmp;
621
622 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
623
624 list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
625 if (tmp->md_dev == md_dev) {
626 pool = tmp;
627 break;
628 }
629 }
630
631 return pool;
632}
633
634/*----------------------------------------------------------------*/
635
636struct dm_thin_endio_hook {
637 struct thin_c *tc;
638 struct deferred_entry *shared_read_entry;
639 struct deferred_entry *all_io_entry;
640 struct dm_thin_new_mapping *overwrite_mapping;
641};
642
643static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
644{
645 struct bio *bio;
646 struct bio_list bios;
647
648 bio_list_init(&bios);
649 bio_list_merge(&bios, master);
650 bio_list_init(master);
651
652 while ((bio = bio_list_pop(&bios))) {
653 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
654
655 if (h->tc == tc)
656 bio_endio(bio, DM_ENDIO_REQUEUE);
657 else
658 bio_list_add(master, bio);
659 }
660}
661
662static void requeue_io(struct thin_c *tc)
663{
664 struct pool *pool = tc->pool;
665 unsigned long flags;
666
667 spin_lock_irqsave(&pool->lock, flags);
668 __requeue_bio_list(tc, &pool->deferred_bios);
669 __requeue_bio_list(tc, &pool->retry_on_resume_list);
670 spin_unlock_irqrestore(&pool->lock, flags);
671}
672
673/*
674 * This section of code contains the logic for processing a thin device's IO.
675 * Much of the code depends on pool object resources (lists, workqueues, etc)
676 * but most is exclusively called from the thin target rather than the thin-pool
677 * target.
678 */
679
680static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
681{
682 return bio->bi_sector >> tc->pool->block_shift;
683}
684
685static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
686{
687 struct pool *pool = tc->pool;
688
689 bio->bi_bdev = tc->pool_dev->bdev;
690 bio->bi_sector = (block << pool->block_shift) +
691 (bio->bi_sector & pool->offset_mask);
692}
693
694static void remap_to_origin(struct thin_c *tc, struct bio *bio)
695{
696 bio->bi_bdev = tc->origin_dev->bdev;
697}
698
699static void issue(struct thin_c *tc, struct bio *bio)
700{
701 struct pool *pool = tc->pool;
702 unsigned long flags;
703
704 /*
705 * Batch together any FUA/FLUSH bios we find and then issue
706 * a single commit for them in process_deferred_bios().
707 */
708 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
709 spin_lock_irqsave(&pool->lock, flags);
710 bio_list_add(&pool->deferred_flush_bios, bio);
711 spin_unlock_irqrestore(&pool->lock, flags);
712 } else
713 generic_make_request(bio);
714}
715
716static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
717{
718 remap_to_origin(tc, bio);
719 issue(tc, bio);
720}
721
722static void remap_and_issue(struct thin_c *tc, struct bio *bio,
723 dm_block_t block)
724{
725 remap(tc, bio, block);
726 issue(tc, bio);
727}
728
729/*
730 * wake_worker() is used when new work is queued and when pool_resume is
731 * ready to continue deferred IO processing.
732 */
733static void wake_worker(struct pool *pool)
734{
735 queue_work(pool->wq, &pool->worker);
736}
737
738/*----------------------------------------------------------------*/
739
740/*
741 * Bio endio functions.
742 */
743struct dm_thin_new_mapping {
744 struct list_head list;
745
746 unsigned quiesced:1;
747 unsigned prepared:1;
748 unsigned pass_discard:1;
749
750 struct thin_c *tc;
751 dm_block_t virt_block;
752 dm_block_t data_block;
753 struct dm_bio_prison_cell *cell, *cell2;
754 int err;
755
756 /*
757 * If the bio covers the whole area of a block then we can avoid
758 * zeroing or copying. Instead this bio is hooked. The bio will
759 * still be in the cell, so care has to be taken to avoid issuing
760 * the bio twice.
761 */
762 struct bio *bio;
763 bio_end_io_t *saved_bi_end_io;
764};
765
766static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
767{
768 struct pool *pool = m->tc->pool;
769
770 if (m->quiesced && m->prepared) {
771 list_add(&m->list, &pool->prepared_mappings);
772 wake_worker(pool);
773 }
774}
775
776static void copy_complete(int read_err, unsigned long write_err, void *context)
777{
778 unsigned long flags;
779 struct dm_thin_new_mapping *m = context;
780 struct pool *pool = m->tc->pool;
781
782 m->err = read_err || write_err ? -EIO : 0;
783
784 spin_lock_irqsave(&pool->lock, flags);
785 m->prepared = 1;
786 __maybe_add_mapping(m);
787 spin_unlock_irqrestore(&pool->lock, flags);
788}
789
790static void overwrite_endio(struct bio *bio, int err)
791{
792 unsigned long flags;
793 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
794 struct dm_thin_new_mapping *m = h->overwrite_mapping;
795 struct pool *pool = m->tc->pool;
796
797 m->err = err;
798
799 spin_lock_irqsave(&pool->lock, flags);
800 m->prepared = 1;
801 __maybe_add_mapping(m);
802 spin_unlock_irqrestore(&pool->lock, flags);
803}
804
805/*----------------------------------------------------------------*/
806
807/*
808 * Workqueue.
809 */
810
811/*
812 * Prepared mapping jobs.
813 */
814
815/*
816 * This sends the bios in the cell back to the deferred_bios list.
817 */
818static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell,
819 dm_block_t data_block)
820{
821 struct pool *pool = tc->pool;
822 unsigned long flags;
823
824 spin_lock_irqsave(&pool->lock, flags);
825 cell_release(cell, &pool->deferred_bios);
826 spin_unlock_irqrestore(&tc->pool->lock, flags);
827
828 wake_worker(pool);
829}
830
831/*
832 * Same as cell_defer above, except it omits one particular detainee,
833 * a write bio that covers the block and has already been processed.
834 */
835static void cell_defer_except(struct thin_c *tc, struct dm_bio_prison_cell *cell)
836{
837 struct bio_list bios;
838 struct pool *pool = tc->pool;
839 unsigned long flags;
840
841 bio_list_init(&bios);
842
843 spin_lock_irqsave(&pool->lock, flags);
844 cell_release_no_holder(cell, &pool->deferred_bios);
845 spin_unlock_irqrestore(&pool->lock, flags);
846
847 wake_worker(pool);
848}
849
850static void process_prepared_mapping(struct dm_thin_new_mapping *m)
851{
852 struct thin_c *tc = m->tc;
853 struct bio *bio;
854 int r;
855
856 bio = m->bio;
857 if (bio)
858 bio->bi_end_io = m->saved_bi_end_io;
859
860 if (m->err) {
861 cell_error(m->cell);
862 goto out;
863 }
864
865 /*
866 * Commit the prepared block into the mapping btree.
867 * Any I/O for this block arriving after this point will get
868 * remapped to it directly.
869 */
870 r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
871 if (r) {
872 DMERR("dm_thin_insert_block() failed");
873 cell_error(m->cell);
874 goto out;
875 }
876
877 /*
878 * Release any bios held while the block was being provisioned.
879 * If we are processing a write bio that completely covers the block,
880 * we already processed it so can ignore it now when processing
881 * the bios in the cell.
882 */
883 if (bio) {
884 cell_defer_except(tc, m->cell);
885 bio_endio(bio, 0);
886 } else
887 cell_defer(tc, m->cell, m->data_block);
888
889out:
890 list_del(&m->list);
891 mempool_free(m, tc->pool->mapping_pool);
892}
893
894static void process_prepared_discard(struct dm_thin_new_mapping *m)
895{
896 int r;
897 struct thin_c *tc = m->tc;
898
899 r = dm_thin_remove_block(tc->td, m->virt_block);
900 if (r)
901 DMERR("dm_thin_remove_block() failed");
902
903 /*
904 * Pass the discard down to the underlying device?
905 */
906 if (m->pass_discard)
907 remap_and_issue(tc, m->bio, m->data_block);
908 else
909 bio_endio(m->bio, 0);
910
911 cell_defer_except(tc, m->cell);
912 cell_defer_except(tc, m->cell2);
913 mempool_free(m, tc->pool->mapping_pool);
914}
915
916static void process_prepared(struct pool *pool, struct list_head *head,
917 void (*fn)(struct dm_thin_new_mapping *))
918{
919 unsigned long flags;
920 struct list_head maps;
921 struct dm_thin_new_mapping *m, *tmp;
922
923 INIT_LIST_HEAD(&maps);
924 spin_lock_irqsave(&pool->lock, flags);
925 list_splice_init(head, &maps);
926 spin_unlock_irqrestore(&pool->lock, flags);
927
928 list_for_each_entry_safe(m, tmp, &maps, list)
929 fn(m);
930}
931
932/*
933 * Deferred bio jobs.
934 */
935static int io_overlaps_block(struct pool *pool, struct bio *bio)
936{
937 return !(bio->bi_sector & pool->offset_mask) &&
938 (bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
939
940}
941
942static int io_overwrites_block(struct pool *pool, struct bio *bio)
943{
944 return (bio_data_dir(bio) == WRITE) &&
945 io_overlaps_block(pool, bio);
946}
947
948static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
949 bio_end_io_t *fn)
950{
951 *save = bio->bi_end_io;
952 bio->bi_end_io = fn;
953}
954
955static int ensure_next_mapping(struct pool *pool)
956{
957 if (pool->next_mapping)
958 return 0;
959
960 pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
961
962 return pool->next_mapping ? 0 : -ENOMEM;
963}
964
965static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
966{
967 struct dm_thin_new_mapping *r = pool->next_mapping;
968
969 BUG_ON(!pool->next_mapping);
970
971 pool->next_mapping = NULL;
972
973 return r;
974}
975
976static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
977 struct dm_dev *origin, dm_block_t data_origin,
978 dm_block_t data_dest,
979 struct dm_bio_prison_cell *cell, struct bio *bio)
980{
981 int r;
982 struct pool *pool = tc->pool;
983 struct dm_thin_new_mapping *m = get_next_mapping(pool);
984
985 INIT_LIST_HEAD(&m->list);
986 m->quiesced = 0;
987 m->prepared = 0;
988 m->tc = tc;
989 m->virt_block = virt_block;
990 m->data_block = data_dest;
991 m->cell = cell;
992 m->err = 0;
993 m->bio = NULL;
994
995 if (!ds_add_work(&pool->shared_read_ds, &m->list))
996 m->quiesced = 1;
997
998 /*
999 * IO to pool_dev remaps to the pool target's data_dev.
1000 *
1001 * If the whole block of data is being overwritten, we can issue the
1002 * bio immediately. Otherwise we use kcopyd to clone the data first.
1003 */
1004 if (io_overwrites_block(pool, bio)) {
1005 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1006
1007 h->overwrite_mapping = m;
1008 m->bio = bio;
1009 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1010 remap_and_issue(tc, bio, data_dest);
1011 } else {
1012 struct dm_io_region from, to;
1013
1014 from.bdev = origin->bdev;
1015 from.sector = data_origin * pool->sectors_per_block;
1016 from.count = pool->sectors_per_block;
1017
1018 to.bdev = tc->pool_dev->bdev;
1019 to.sector = data_dest * pool->sectors_per_block;
1020 to.count = pool->sectors_per_block;
1021
1022 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1023 0, copy_complete, m);
1024 if (r < 0) {
1025 mempool_free(m, pool->mapping_pool);
1026 DMERR("dm_kcopyd_copy() failed");
1027 cell_error(cell);
1028 }
1029 }
1030}
1031
1032static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1033 dm_block_t data_origin, dm_block_t data_dest,
1034 struct dm_bio_prison_cell *cell, struct bio *bio)
1035{
1036 schedule_copy(tc, virt_block, tc->pool_dev,
1037 data_origin, data_dest, cell, bio);
1038}
1039
1040static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1041 dm_block_t data_dest,
1042 struct dm_bio_prison_cell *cell, struct bio *bio)
1043{
1044 schedule_copy(tc, virt_block, tc->origin_dev,
1045 virt_block, data_dest, cell, bio);
1046}
1047
1048static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1049 dm_block_t data_block, struct dm_bio_prison_cell *cell,
1050 struct bio *bio)
1051{
1052 struct pool *pool = tc->pool;
1053 struct dm_thin_new_mapping *m = get_next_mapping(pool);
1054
1055 INIT_LIST_HEAD(&m->list);
1056 m->quiesced = 1;
1057 m->prepared = 0;
1058 m->tc = tc;
1059 m->virt_block = virt_block;
1060 m->data_block = data_block;
1061 m->cell = cell;
1062 m->err = 0;
1063 m->bio = NULL;
1064
1065 /*
1066 * If the whole block of data is being overwritten or we are not
1067 * zeroing pre-existing data, we can issue the bio immediately.
1068 * Otherwise we use kcopyd to zero the data first.
1069 */
1070 if (!pool->pf.zero_new_blocks)
1071 process_prepared_mapping(m);
1072
1073 else if (io_overwrites_block(pool, bio)) {
1074 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1075
1076 h->overwrite_mapping = m;
1077 m->bio = bio;
1078 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1079 remap_and_issue(tc, bio, data_block);
1080 } else {
1081 int r;
1082 struct dm_io_region to;
1083
1084 to.bdev = tc->pool_dev->bdev;
1085 to.sector = data_block * pool->sectors_per_block;
1086 to.count = pool->sectors_per_block;
1087
1088 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
1089 if (r < 0) {
1090 mempool_free(m, pool->mapping_pool);
1091 DMERR("dm_kcopyd_zero() failed");
1092 cell_error(cell);
1093 }
1094 }
1095}
1096
1097static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1098{
1099 int r;
1100 dm_block_t free_blocks;
1101 unsigned long flags;
1102 struct pool *pool = tc->pool;
1103
1104 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1105 if (r)
1106 return r;
1107
1108 if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1109 DMWARN("%s: reached low water mark, sending event.",
1110 dm_device_name(pool->pool_md));
1111 spin_lock_irqsave(&pool->lock, flags);
1112 pool->low_water_triggered = 1;
1113 spin_unlock_irqrestore(&pool->lock, flags);
1114 dm_table_event(pool->ti->table);
1115 }
1116
1117 if (!free_blocks) {
1118 if (pool->no_free_space)
1119 return -ENOSPC;
1120 else {
1121 /*
1122 * Try to commit to see if that will free up some
1123 * more space.
1124 */
1125 r = dm_pool_commit_metadata(pool->pmd);
1126 if (r) {
1127 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1128 __func__, r);
1129 return r;
1130 }
1131
1132 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1133 if (r)
1134 return r;
1135
1136 /*
1137 * If we still have no space we set a flag to avoid
1138 * doing all this checking and return -ENOSPC.
1139 */
1140 if (!free_blocks) {
1141 DMWARN("%s: no free space available.",
1142 dm_device_name(pool->pool_md));
1143 spin_lock_irqsave(&pool->lock, flags);
1144 pool->no_free_space = 1;
1145 spin_unlock_irqrestore(&pool->lock, flags);
1146 return -ENOSPC;
1147 }
1148 }
1149 }
1150
1151 r = dm_pool_alloc_data_block(pool->pmd, result);
1152 if (r)
1153 return r;
1154
1155 return 0;
1156}
1157
1158/*
1159 * If we have run out of space, queue bios until the device is
1160 * resumed, presumably after having been reloaded with more space.
1161 */
1162static void retry_on_resume(struct bio *bio)
1163{
1164 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1165 struct thin_c *tc = h->tc;
1166 struct pool *pool = tc->pool;
1167 unsigned long flags;
1168
1169 spin_lock_irqsave(&pool->lock, flags);
1170 bio_list_add(&pool->retry_on_resume_list, bio);
1171 spin_unlock_irqrestore(&pool->lock, flags);
1172}
1173
1174static void no_space(struct dm_bio_prison_cell *cell)
1175{
1176 struct bio *bio;
1177 struct bio_list bios;
1178
1179 bio_list_init(&bios);
1180 cell_release(cell, &bios);
1181
1182 while ((bio = bio_list_pop(&bios)))
1183 retry_on_resume(bio);
1184}
1185
1186static void process_discard(struct thin_c *tc, struct bio *bio)
1187{
1188 int r;
1189 unsigned long flags;
1190 struct pool *pool = tc->pool;
1191 struct dm_bio_prison_cell *cell, *cell2;
1192 struct cell_key key, key2;
1193 dm_block_t block = get_bio_block(tc, bio);
1194 struct dm_thin_lookup_result lookup_result;
1195 struct dm_thin_new_mapping *m;
1196
1197 build_virtual_key(tc->td, block, &key);
1198 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1199 return;
1200
1201 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1202 switch (r) {
1203 case 0:
1204 /*
1205 * Check nobody is fiddling with this pool block. This can
1206 * happen if someone's in the process of breaking sharing
1207 * on this block.
1208 */
1209 build_data_key(tc->td, lookup_result.block, &key2);
1210 if (bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
1211 cell_release_singleton(cell, bio);
1212 break;
1213 }
1214
1215 if (io_overlaps_block(pool, bio)) {
1216 /*
1217 * IO may still be going to the destination block. We must
1218 * quiesce before we can do the removal.
1219 */
1220 m = get_next_mapping(pool);
1221 m->tc = tc;
1222 m->pass_discard = (!lookup_result.shared) & pool->pf.discard_passdown;
1223 m->virt_block = block;
1224 m->data_block = lookup_result.block;
1225 m->cell = cell;
1226 m->cell2 = cell2;
1227 m->err = 0;
1228 m->bio = bio;
1229
1230 if (!ds_add_work(&pool->all_io_ds, &m->list)) {
1231 spin_lock_irqsave(&pool->lock, flags);
1232 list_add(&m->list, &pool->prepared_discards);
1233 spin_unlock_irqrestore(&pool->lock, flags);
1234 wake_worker(pool);
1235 }
1236 } else {
1237 /*
1238 * This path is hit if people are ignoring
1239 * limits->discard_granularity. It ignores any
1240 * part of the discard that is in a subsequent
1241 * block.
1242 */
1243 sector_t offset = bio->bi_sector - (block << pool->block_shift);
1244 unsigned remaining = (pool->sectors_per_block - offset) << 9;
1245 bio->bi_size = min(bio->bi_size, remaining);
1246
1247 cell_release_singleton(cell, bio);
1248 cell_release_singleton(cell2, bio);
1249 if ((!lookup_result.shared) && pool->pf.discard_passdown)
1250 remap_and_issue(tc, bio, lookup_result.block);
1251 else
1252 bio_endio(bio, 0);
1253 }
1254 break;
1255
1256 case -ENODATA:
1257 /*
1258 * It isn't provisioned, just forget it.
1259 */
1260 cell_release_singleton(cell, bio);
1261 bio_endio(bio, 0);
1262 break;
1263
1264 default:
1265 DMERR("discard: find block unexpectedly returned %d", r);
1266 cell_release_singleton(cell, bio);
1267 bio_io_error(bio);
1268 break;
1269 }
1270}
1271
1272static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1273 struct cell_key *key,
1274 struct dm_thin_lookup_result *lookup_result,
1275 struct dm_bio_prison_cell *cell)
1276{
1277 int r;
1278 dm_block_t data_block;
1279
1280 r = alloc_data_block(tc, &data_block);
1281 switch (r) {
1282 case 0:
1283 schedule_internal_copy(tc, block, lookup_result->block,
1284 data_block, cell, bio);
1285 break;
1286
1287 case -ENOSPC:
1288 no_space(cell);
1289 break;
1290
1291 default:
1292 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1293 cell_error(cell);
1294 break;
1295 }
1296}
1297
1298static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1299 dm_block_t block,
1300 struct dm_thin_lookup_result *lookup_result)
1301{
1302 struct dm_bio_prison_cell *cell;
1303 struct pool *pool = tc->pool;
1304 struct cell_key key;
1305
1306 /*
1307 * If cell is already occupied, then sharing is already in the process
1308 * of being broken so we have nothing further to do here.
1309 */
1310 build_data_key(tc->td, lookup_result->block, &key);
1311 if (bio_detain(pool->prison, &key, bio, &cell))
1312 return;
1313
1314 if (bio_data_dir(bio) == WRITE)
1315 break_sharing(tc, bio, block, &key, lookup_result, cell);
1316 else {
1317 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1318
1319 h->shared_read_entry = ds_inc(&pool->shared_read_ds);
1320
1321 cell_release_singleton(cell, bio);
1322 remap_and_issue(tc, bio, lookup_result->block);
1323 }
1324}
1325
1326static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1327 struct dm_bio_prison_cell *cell)
1328{
1329 int r;
1330 dm_block_t data_block;
1331
1332 /*
1333 * Remap empty bios (flushes) immediately, without provisioning.
1334 */
1335 if (!bio->bi_size) {
1336 cell_release_singleton(cell, bio);
1337 remap_and_issue(tc, bio, 0);
1338 return;
1339 }
1340
1341 /*
1342 * Fill read bios with zeroes and complete them immediately.
1343 */
1344 if (bio_data_dir(bio) == READ) {
1345 zero_fill_bio(bio);
1346 cell_release_singleton(cell, bio);
1347 bio_endio(bio, 0);
1348 return;
1349 }
1350
1351 r = alloc_data_block(tc, &data_block);
1352 switch (r) {
1353 case 0:
1354 if (tc->origin_dev)
1355 schedule_external_copy(tc, block, data_block, cell, bio);
1356 else
1357 schedule_zero(tc, block, data_block, cell, bio);
1358 break;
1359
1360 case -ENOSPC:
1361 no_space(cell);
1362 break;
1363
1364 default:
1365 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1366 cell_error(cell);
1367 break;
1368 }
1369}
1370
1371static void process_bio(struct thin_c *tc, struct bio *bio)
1372{
1373 int r;
1374 dm_block_t block = get_bio_block(tc, bio);
1375 struct dm_bio_prison_cell *cell;
1376 struct cell_key key;
1377 struct dm_thin_lookup_result lookup_result;
1378
1379 /*
1380 * If cell is already occupied, then the block is already
1381 * being provisioned so we have nothing further to do here.
1382 */
1383 build_virtual_key(tc->td, block, &key);
1384 if (bio_detain(tc->pool->prison, &key, bio, &cell))
1385 return;
1386
1387 r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1388 switch (r) {
1389 case 0:
1390 /*
1391 * We can release this cell now. This thread is the only
1392 * one that puts bios into a cell, and we know there were
1393 * no preceding bios.
1394 */
1395 /*
1396 * TODO: this will probably have to change when discard goes
1397 * back in.
1398 */
1399 cell_release_singleton(cell, bio);
1400
1401 if (lookup_result.shared)
1402 process_shared_bio(tc, bio, block, &lookup_result);
1403 else
1404 remap_and_issue(tc, bio, lookup_result.block);
1405 break;
1406
1407 case -ENODATA:
1408 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1409 cell_release_singleton(cell, bio);
1410 remap_to_origin_and_issue(tc, bio);
1411 } else
1412 provision_block(tc, bio, block, cell);
1413 break;
1414
1415 default:
1416 DMERR("dm_thin_find_block() failed, error = %d", r);
1417 cell_release_singleton(cell, bio);
1418 bio_io_error(bio);
1419 break;
1420 }
1421}
1422
1423static int need_commit_due_to_time(struct pool *pool)
1424{
1425 return jiffies < pool->last_commit_jiffies ||
1426 jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1427}
1428
1429static void process_deferred_bios(struct pool *pool)
1430{
1431 unsigned long flags;
1432 struct bio *bio;
1433 struct bio_list bios;
1434 int r;
1435
1436 bio_list_init(&bios);
1437
1438 spin_lock_irqsave(&pool->lock, flags);
1439 bio_list_merge(&bios, &pool->deferred_bios);
1440 bio_list_init(&pool->deferred_bios);
1441 spin_unlock_irqrestore(&pool->lock, flags);
1442
1443 while ((bio = bio_list_pop(&bios))) {
1444 struct dm_thin_endio_hook *h = dm_get_mapinfo(bio)->ptr;
1445 struct thin_c *tc = h->tc;
1446
1447 /*
1448 * If we've got no free new_mapping structs, and processing
1449 * this bio might require one, we pause until there are some
1450 * prepared mappings to process.
1451 */
1452 if (ensure_next_mapping(pool)) {
1453 spin_lock_irqsave(&pool->lock, flags);
1454 bio_list_merge(&pool->deferred_bios, &bios);
1455 spin_unlock_irqrestore(&pool->lock, flags);
1456
1457 break;
1458 }
1459
1460 if (bio->bi_rw & REQ_DISCARD)
1461 process_discard(tc, bio);
1462 else
1463 process_bio(tc, bio);
1464 }
1465
1466 /*
1467 * If there are any deferred flush bios, we must commit
1468 * the metadata before issuing them.
1469 */
1470 bio_list_init(&bios);
1471 spin_lock_irqsave(&pool->lock, flags);
1472 bio_list_merge(&bios, &pool->deferred_flush_bios);
1473 bio_list_init(&pool->deferred_flush_bios);
1474 spin_unlock_irqrestore(&pool->lock, flags);
1475
1476 if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1477 return;
1478
1479 r = dm_pool_commit_metadata(pool->pmd);
1480 if (r) {
1481 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1482 __func__, r);
1483 while ((bio = bio_list_pop(&bios)))
1484 bio_io_error(bio);
1485 return;
1486 }
1487 pool->last_commit_jiffies = jiffies;
1488
1489 while ((bio = bio_list_pop(&bios)))
1490 generic_make_request(bio);
1491}
1492
1493static void do_worker(struct work_struct *ws)
1494{
1495 struct pool *pool = container_of(ws, struct pool, worker);
1496
1497 process_prepared(pool, &pool->prepared_mappings, process_prepared_mapping);
1498 process_prepared(pool, &pool->prepared_discards, process_prepared_discard);
1499 process_deferred_bios(pool);
1500}
1501
1502/*
1503 * We want to commit periodically so that not too much
1504 * unwritten data builds up.
1505 */
1506static void do_waker(struct work_struct *ws)
1507{
1508 struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1509 wake_worker(pool);
1510 queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1511}
1512
1513/*----------------------------------------------------------------*/
1514
1515/*
1516 * Mapping functions.
1517 */
1518
1519/*
1520 * Called only while mapping a thin bio to hand it over to the workqueue.
1521 */
1522static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1523{
1524 unsigned long flags;
1525 struct pool *pool = tc->pool;
1526
1527 spin_lock_irqsave(&pool->lock, flags);
1528 bio_list_add(&pool->deferred_bios, bio);
1529 spin_unlock_irqrestore(&pool->lock, flags);
1530
1531 wake_worker(pool);
1532}
1533
1534static struct dm_thin_endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1535{
1536 struct pool *pool = tc->pool;
1537 struct dm_thin_endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1538
1539 h->tc = tc;
1540 h->shared_read_entry = NULL;
1541 h->all_io_entry = bio->bi_rw & REQ_DISCARD ? NULL : ds_inc(&pool->all_io_ds);
1542 h->overwrite_mapping = NULL;
1543
1544 return h;
1545}
1546
1547/*
1548 * Non-blocking function called from the thin target's map function.
1549 */
1550static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1551 union map_info *map_context)
1552{
1553 int r;
1554 struct thin_c *tc = ti->private;
1555 dm_block_t block = get_bio_block(tc, bio);
1556 struct dm_thin_device *td = tc->td;
1557 struct dm_thin_lookup_result result;
1558
1559 map_context->ptr = thin_hook_bio(tc, bio);
1560 if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1561 thin_defer_bio(tc, bio);
1562 return DM_MAPIO_SUBMITTED;
1563 }
1564
1565 r = dm_thin_find_block(td, block, 0, &result);
1566
1567 /*
1568 * Note that we defer readahead too.
1569 */
1570 switch (r) {
1571 case 0:
1572 if (unlikely(result.shared)) {
1573 /*
1574 * We have a race condition here between the
1575 * result.shared value returned by the lookup and
1576 * snapshot creation, which may cause new
1577 * sharing.
1578 *
1579 * To avoid this always quiesce the origin before
1580 * taking the snap. You want to do this anyway to
1581 * ensure a consistent application view
1582 * (i.e. lockfs).
1583 *
1584 * More distant ancestors are irrelevant. The
1585 * shared flag will be set in their case.
1586 */
1587 thin_defer_bio(tc, bio);
1588 r = DM_MAPIO_SUBMITTED;
1589 } else {
1590 remap(tc, bio, result.block);
1591 r = DM_MAPIO_REMAPPED;
1592 }
1593 break;
1594
1595 case -ENODATA:
1596 /*
1597 * In future, the failed dm_thin_find_block above could
1598 * provide the hint to load the metadata into cache.
1599 */
1600 case -EWOULDBLOCK:
1601 thin_defer_bio(tc, bio);
1602 r = DM_MAPIO_SUBMITTED;
1603 break;
1604 }
1605
1606 return r;
1607}
1608
1609static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1610{
1611 int r;
1612 unsigned long flags;
1613 struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1614
1615 spin_lock_irqsave(&pt->pool->lock, flags);
1616 r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1617 spin_unlock_irqrestore(&pt->pool->lock, flags);
1618
1619 if (!r) {
1620 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1621 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1622 }
1623
1624 return r;
1625}
1626
1627static void __requeue_bios(struct pool *pool)
1628{
1629 bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1630 bio_list_init(&pool->retry_on_resume_list);
1631}
1632
1633/*----------------------------------------------------------------
1634 * Binding of control targets to a pool object
1635 *--------------------------------------------------------------*/
1636static int bind_control_target(struct pool *pool, struct dm_target *ti)
1637{
1638 struct pool_c *pt = ti->private;
1639
1640 pool->ti = ti;
1641 pool->low_water_blocks = pt->low_water_blocks;
1642 pool->pf = pt->pf;
1643
1644 /*
1645 * If discard_passdown was enabled verify that the data device
1646 * supports discards. Disable discard_passdown if not; otherwise
1647 * -EOPNOTSUPP will be returned.
1648 */
1649 if (pt->pf.discard_passdown) {
1650 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1651 if (!q || !blk_queue_discard(q)) {
1652 char buf[BDEVNAME_SIZE];
1653 DMWARN("Discard unsupported by data device (%s): Disabling discard passdown.",
1654 bdevname(pt->data_dev->bdev, buf));
1655 pool->pf.discard_passdown = 0;
1656 }
1657 }
1658
1659 return 0;
1660}
1661
1662static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1663{
1664 if (pool->ti == ti)
1665 pool->ti = NULL;
1666}
1667
1668/*----------------------------------------------------------------
1669 * Pool creation
1670 *--------------------------------------------------------------*/
1671/* Initialize pool features. */
1672static void pool_features_init(struct pool_features *pf)
1673{
1674 pf->zero_new_blocks = 1;
1675 pf->discard_enabled = 1;
1676 pf->discard_passdown = 1;
1677}
1678
1679static void __pool_destroy(struct pool *pool)
1680{
1681 __pool_table_remove(pool);
1682
1683 if (dm_pool_metadata_close(pool->pmd) < 0)
1684 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1685
1686 prison_destroy(pool->prison);
1687 dm_kcopyd_client_destroy(pool->copier);
1688
1689 if (pool->wq)
1690 destroy_workqueue(pool->wq);
1691
1692 if (pool->next_mapping)
1693 mempool_free(pool->next_mapping, pool->mapping_pool);
1694 mempool_destroy(pool->mapping_pool);
1695 mempool_destroy(pool->endio_hook_pool);
1696 kfree(pool);
1697}
1698
1699static struct kmem_cache *_new_mapping_cache;
1700static struct kmem_cache *_endio_hook_cache;
1701
1702static struct pool *pool_create(struct mapped_device *pool_md,
1703 struct block_device *metadata_dev,
1704 unsigned long block_size, char **error)
1705{
1706 int r;
1707 void *err_p;
1708 struct pool *pool;
1709 struct dm_pool_metadata *pmd;
1710
1711 pmd = dm_pool_metadata_open(metadata_dev, block_size);
1712 if (IS_ERR(pmd)) {
1713 *error = "Error creating metadata object";
1714 return (struct pool *)pmd;
1715 }
1716
1717 pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1718 if (!pool) {
1719 *error = "Error allocating memory for pool";
1720 err_p = ERR_PTR(-ENOMEM);
1721 goto bad_pool;
1722 }
1723
1724 pool->pmd = pmd;
1725 pool->sectors_per_block = block_size;
1726 pool->block_shift = ffs(block_size) - 1;
1727 pool->offset_mask = block_size - 1;
1728 pool->low_water_blocks = 0;
1729 pool_features_init(&pool->pf);
1730 pool->prison = prison_create(PRISON_CELLS);
1731 if (!pool->prison) {
1732 *error = "Error creating pool's bio prison";
1733 err_p = ERR_PTR(-ENOMEM);
1734 goto bad_prison;
1735 }
1736
1737 pool->copier = dm_kcopyd_client_create();
1738 if (IS_ERR(pool->copier)) {
1739 r = PTR_ERR(pool->copier);
1740 *error = "Error creating pool's kcopyd client";
1741 err_p = ERR_PTR(r);
1742 goto bad_kcopyd_client;
1743 }
1744
1745 /*
1746 * Create singlethreaded workqueue that will service all devices
1747 * that use this metadata.
1748 */
1749 pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1750 if (!pool->wq) {
1751 *error = "Error creating pool's workqueue";
1752 err_p = ERR_PTR(-ENOMEM);
1753 goto bad_wq;
1754 }
1755
1756 INIT_WORK(&pool->worker, do_worker);
1757 INIT_DELAYED_WORK(&pool->waker, do_waker);
1758 spin_lock_init(&pool->lock);
1759 bio_list_init(&pool->deferred_bios);
1760 bio_list_init(&pool->deferred_flush_bios);
1761 INIT_LIST_HEAD(&pool->prepared_mappings);
1762 INIT_LIST_HEAD(&pool->prepared_discards);
1763 pool->low_water_triggered = 0;
1764 pool->no_free_space = 0;
1765 bio_list_init(&pool->retry_on_resume_list);
1766 ds_init(&pool->shared_read_ds);
1767 ds_init(&pool->all_io_ds);
1768
1769 pool->next_mapping = NULL;
1770 pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1771 _new_mapping_cache);
1772 if (!pool->mapping_pool) {
1773 *error = "Error creating pool's mapping mempool";
1774 err_p = ERR_PTR(-ENOMEM);
1775 goto bad_mapping_pool;
1776 }
1777
1778 pool->endio_hook_pool = mempool_create_slab_pool(ENDIO_HOOK_POOL_SIZE,
1779 _endio_hook_cache);
1780 if (!pool->endio_hook_pool) {
1781 *error = "Error creating pool's endio_hook mempool";
1782 err_p = ERR_PTR(-ENOMEM);
1783 goto bad_endio_hook_pool;
1784 }
1785 pool->ref_count = 1;
1786 pool->last_commit_jiffies = jiffies;
1787 pool->pool_md = pool_md;
1788 pool->md_dev = metadata_dev;
1789 __pool_table_insert(pool);
1790
1791 return pool;
1792
1793bad_endio_hook_pool:
1794 mempool_destroy(pool->mapping_pool);
1795bad_mapping_pool:
1796 destroy_workqueue(pool->wq);
1797bad_wq:
1798 dm_kcopyd_client_destroy(pool->copier);
1799bad_kcopyd_client:
1800 prison_destroy(pool->prison);
1801bad_prison:
1802 kfree(pool);
1803bad_pool:
1804 if (dm_pool_metadata_close(pmd))
1805 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1806
1807 return err_p;
1808}
1809
1810static void __pool_inc(struct pool *pool)
1811{
1812 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1813 pool->ref_count++;
1814}
1815
1816static void __pool_dec(struct pool *pool)
1817{
1818 BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1819 BUG_ON(!pool->ref_count);
1820 if (!--pool->ref_count)
1821 __pool_destroy(pool);
1822}
1823
1824static struct pool *__pool_find(struct mapped_device *pool_md,
1825 struct block_device *metadata_dev,
1826 unsigned long block_size, char **error,
1827 int *created)
1828{
1829 struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1830
1831 if (pool) {
1832 if (pool->pool_md != pool_md)
1833 return ERR_PTR(-EBUSY);
1834 __pool_inc(pool);
1835
1836 } else {
1837 pool = __pool_table_lookup(pool_md);
1838 if (pool) {
1839 if (pool->md_dev != metadata_dev)
1840 return ERR_PTR(-EINVAL);
1841 __pool_inc(pool);
1842
1843 } else {
1844 pool = pool_create(pool_md, metadata_dev, block_size, error);
1845 *created = 1;
1846 }
1847 }
1848
1849 return pool;
1850}
1851
1852/*----------------------------------------------------------------
1853 * Pool target methods
1854 *--------------------------------------------------------------*/
1855static void pool_dtr(struct dm_target *ti)
1856{
1857 struct pool_c *pt = ti->private;
1858
1859 mutex_lock(&dm_thin_pool_table.mutex);
1860
1861 unbind_control_target(pt->pool, ti);
1862 __pool_dec(pt->pool);
1863 dm_put_device(ti, pt->metadata_dev);
1864 dm_put_device(ti, pt->data_dev);
1865 kfree(pt);
1866
1867 mutex_unlock(&dm_thin_pool_table.mutex);
1868}
1869
1870static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1871 struct dm_target *ti)
1872{
1873 int r;
1874 unsigned argc;
1875 const char *arg_name;
1876
1877 static struct dm_arg _args[] = {
1878 {0, 3, "Invalid number of pool feature arguments"},
1879 };
1880
1881 /*
1882 * No feature arguments supplied.
1883 */
1884 if (!as->argc)
1885 return 0;
1886
1887 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1888 if (r)
1889 return -EINVAL;
1890
1891 while (argc && !r) {
1892 arg_name = dm_shift_arg(as);
1893 argc--;
1894
1895 if (!strcasecmp(arg_name, "skip_block_zeroing")) {
1896 pf->zero_new_blocks = 0;
1897 continue;
1898 } else if (!strcasecmp(arg_name, "ignore_discard")) {
1899 pf->discard_enabled = 0;
1900 continue;
1901 } else if (!strcasecmp(arg_name, "no_discard_passdown")) {
1902 pf->discard_passdown = 0;
1903 continue;
1904 }
1905
1906 ti->error = "Unrecognised pool feature requested";
1907 r = -EINVAL;
1908 }
1909
1910 return r;
1911}
1912
1913/*
1914 * thin-pool <metadata dev> <data dev>
1915 * <data block size (sectors)>
1916 * <low water mark (blocks)>
1917 * [<#feature args> [<arg>]*]
1918 *
1919 * Optional feature arguments are:
1920 * skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1921 * ignore_discard: disable discard
1922 * no_discard_passdown: don't pass discards down to the data device
1923 */
1924static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1925{
1926 int r, pool_created = 0;
1927 struct pool_c *pt;
1928 struct pool *pool;
1929 struct pool_features pf;
1930 struct dm_arg_set as;
1931 struct dm_dev *data_dev;
1932 unsigned long block_size;
1933 dm_block_t low_water_blocks;
1934 struct dm_dev *metadata_dev;
1935 sector_t metadata_dev_size;
1936 char b[BDEVNAME_SIZE];
1937
1938 /*
1939 * FIXME Remove validation from scope of lock.
1940 */
1941 mutex_lock(&dm_thin_pool_table.mutex);
1942
1943 if (argc < 4) {
1944 ti->error = "Invalid argument count";
1945 r = -EINVAL;
1946 goto out_unlock;
1947 }
1948 as.argc = argc;
1949 as.argv = argv;
1950
1951 r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1952 if (r) {
1953 ti->error = "Error opening metadata block device";
1954 goto out_unlock;
1955 }
1956
1957 metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1958 if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1959 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1960 bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1961
1962 r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1963 if (r) {
1964 ti->error = "Error getting data device";
1965 goto out_metadata;
1966 }
1967
1968 if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1969 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1970 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1971 !is_power_of_2(block_size)) {
1972 ti->error = "Invalid block size";
1973 r = -EINVAL;
1974 goto out;
1975 }
1976
1977 if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1978 ti->error = "Invalid low water mark";
1979 r = -EINVAL;
1980 goto out;
1981 }
1982
1983 /*
1984 * Set default pool features.
1985 */
1986 pool_features_init(&pf);
1987
1988 dm_consume_args(&as, 4);
1989 r = parse_pool_features(&as, &pf, ti);
1990 if (r)
1991 goto out;
1992
1993 pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1994 if (!pt) {
1995 r = -ENOMEM;
1996 goto out;
1997 }
1998
1999 pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2000 block_size, &ti->error, &pool_created);
2001 if (IS_ERR(pool)) {
2002 r = PTR_ERR(pool);
2003 goto out_free_pt;
2004 }
2005
2006 /*
2007 * 'pool_created' reflects whether this is the first table load.
2008 * Top level discard support is not allowed to be changed after
2009 * initial load. This would require a pool reload to trigger thin
2010 * device changes.
2011 */
2012 if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2013 ti->error = "Discard support cannot be disabled once enabled";
2014 r = -EINVAL;
2015 goto out_flags_changed;
2016 }
2017
2018 pt->pool = pool;
2019 pt->ti = ti;
2020 pt->metadata_dev = metadata_dev;
2021 pt->data_dev = data_dev;
2022 pt->low_water_blocks = low_water_blocks;
2023 pt->pf = pf;
2024 ti->num_flush_requests = 1;
2025 /*
2026 * Only need to enable discards if the pool should pass
2027 * them down to the data device. The thin device's discard
2028 * processing will cause mappings to be removed from the btree.
2029 */
2030 if (pf.discard_enabled && pf.discard_passdown) {
2031 ti->num_discard_requests = 1;
2032 /*
2033 * Setting 'discards_supported' circumvents the normal
2034 * stacking of discard limits (this keeps the pool and
2035 * thin devices' discard limits consistent).
2036 */
2037 ti->discards_supported = 1;
2038 }
2039 ti->private = pt;
2040
2041 pt->callbacks.congested_fn = pool_is_congested;
2042 dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2043
2044 mutex_unlock(&dm_thin_pool_table.mutex);
2045
2046 return 0;
2047
2048out_flags_changed:
2049 __pool_dec(pool);
2050out_free_pt:
2051 kfree(pt);
2052out:
2053 dm_put_device(ti, data_dev);
2054out_metadata:
2055 dm_put_device(ti, metadata_dev);
2056out_unlock:
2057 mutex_unlock(&dm_thin_pool_table.mutex);
2058
2059 return r;
2060}
2061
2062static int pool_map(struct dm_target *ti, struct bio *bio,
2063 union map_info *map_context)
2064{
2065 int r;
2066 struct pool_c *pt = ti->private;
2067 struct pool *pool = pt->pool;
2068 unsigned long flags;
2069
2070 /*
2071 * As this is a singleton target, ti->begin is always zero.
2072 */
2073 spin_lock_irqsave(&pool->lock, flags);
2074 bio->bi_bdev = pt->data_dev->bdev;
2075 r = DM_MAPIO_REMAPPED;
2076 spin_unlock_irqrestore(&pool->lock, flags);
2077
2078 return r;
2079}
2080
2081/*
2082 * Retrieves the number of blocks of the data device from
2083 * the superblock and compares it to the actual device size,
2084 * thus resizing the data device in case it has grown.
2085 *
2086 * This both copes with opening preallocated data devices in the ctr
2087 * being followed by a resume
2088 * -and-
2089 * calling the resume method individually after userspace has
2090 * grown the data device in reaction to a table event.
2091 */
2092static int pool_preresume(struct dm_target *ti)
2093{
2094 int r;
2095 struct pool_c *pt = ti->private;
2096 struct pool *pool = pt->pool;
2097 dm_block_t data_size, sb_data_size;
2098
2099 /*
2100 * Take control of the pool object.
2101 */
2102 r = bind_control_target(pool, ti);
2103 if (r)
2104 return r;
2105
2106 data_size = ti->len >> pool->block_shift;
2107 r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2108 if (r) {
2109 DMERR("failed to retrieve data device size");
2110 return r;
2111 }
2112
2113 if (data_size < sb_data_size) {
2114 DMERR("pool target too small, is %llu blocks (expected %llu)",
2115 data_size, sb_data_size);
2116 return -EINVAL;
2117
2118 } else if (data_size > sb_data_size) {
2119 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2120 if (r) {
2121 DMERR("failed to resize data device");
2122 return r;
2123 }
2124
2125 r = dm_pool_commit_metadata(pool->pmd);
2126 if (r) {
2127 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2128 __func__, r);
2129 return r;
2130 }
2131 }
2132
2133 return 0;
2134}
2135
2136static void pool_resume(struct dm_target *ti)
2137{
2138 struct pool_c *pt = ti->private;
2139 struct pool *pool = pt->pool;
2140 unsigned long flags;
2141
2142 spin_lock_irqsave(&pool->lock, flags);
2143 pool->low_water_triggered = 0;
2144 pool->no_free_space = 0;
2145 __requeue_bios(pool);
2146 spin_unlock_irqrestore(&pool->lock, flags);
2147
2148 do_waker(&pool->waker.work);
2149}
2150
2151static void pool_postsuspend(struct dm_target *ti)
2152{
2153 int r;
2154 struct pool_c *pt = ti->private;
2155 struct pool *pool = pt->pool;
2156
2157 cancel_delayed_work(&pool->waker);
2158 flush_workqueue(pool->wq);
2159
2160 r = dm_pool_commit_metadata(pool->pmd);
2161 if (r < 0) {
2162 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2163 __func__, r);
2164 /* FIXME: invalidate device? error the next FUA or FLUSH bio ?*/
2165 }
2166}
2167
2168static int check_arg_count(unsigned argc, unsigned args_required)
2169{
2170 if (argc != args_required) {
2171 DMWARN("Message received with %u arguments instead of %u.",
2172 argc, args_required);
2173 return -EINVAL;
2174 }
2175
2176 return 0;
2177}
2178
2179static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2180{
2181 if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2182 *dev_id <= MAX_DEV_ID)
2183 return 0;
2184
2185 if (warning)
2186 DMWARN("Message received with invalid device id: %s", arg);
2187
2188 return -EINVAL;
2189}
2190
2191static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2192{
2193 dm_thin_id dev_id;
2194 int r;
2195
2196 r = check_arg_count(argc, 2);
2197 if (r)
2198 return r;
2199
2200 r = read_dev_id(argv[1], &dev_id, 1);
2201 if (r)
2202 return r;
2203
2204 r = dm_pool_create_thin(pool->pmd, dev_id);
2205 if (r) {
2206 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2207 argv[1]);
2208 return r;
2209 }
2210
2211 return 0;
2212}
2213
2214static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2215{
2216 dm_thin_id dev_id;
2217 dm_thin_id origin_dev_id;
2218 int r;
2219
2220 r = check_arg_count(argc, 3);
2221 if (r)
2222 return r;
2223
2224 r = read_dev_id(argv[1], &dev_id, 1);
2225 if (r)
2226 return r;
2227
2228 r = read_dev_id(argv[2], &origin_dev_id, 1);
2229 if (r)
2230 return r;
2231
2232 r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2233 if (r) {
2234 DMWARN("Creation of new snapshot %s of device %s failed.",
2235 argv[1], argv[2]);
2236 return r;
2237 }
2238
2239 return 0;
2240}
2241
2242static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2243{
2244 dm_thin_id dev_id;
2245 int r;
2246
2247 r = check_arg_count(argc, 2);
2248 if (r)
2249 return r;
2250
2251 r = read_dev_id(argv[1], &dev_id, 1);
2252 if (r)
2253 return r;
2254
2255 r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2256 if (r)
2257 DMWARN("Deletion of thin device %s failed.", argv[1]);
2258
2259 return r;
2260}
2261
2262static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2263{
2264 dm_thin_id old_id, new_id;
2265 int r;
2266
2267 r = check_arg_count(argc, 3);
2268 if (r)
2269 return r;
2270
2271 if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2272 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2273 return -EINVAL;
2274 }
2275
2276 if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2277 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2278 return -EINVAL;
2279 }
2280
2281 r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2282 if (r) {
2283 DMWARN("Failed to change transaction id from %s to %s.",
2284 argv[1], argv[2]);
2285 return r;
2286 }
2287
2288 return 0;
2289}
2290
2291static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2292{
2293 int r;
2294
2295 r = check_arg_count(argc, 1);
2296 if (r)
2297 return r;
2298
2299 r = dm_pool_commit_metadata(pool->pmd);
2300 if (r) {
2301 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2302 __func__, r);
2303 return r;
2304 }
2305
2306 r = dm_pool_reserve_metadata_snap(pool->pmd);
2307 if (r)
2308 DMWARN("reserve_metadata_snap message failed.");
2309
2310 return r;
2311}
2312
2313static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2314{
2315 int r;
2316
2317 r = check_arg_count(argc, 1);
2318 if (r)
2319 return r;
2320
2321 r = dm_pool_release_metadata_snap(pool->pmd);
2322 if (r)
2323 DMWARN("release_metadata_snap message failed.");
2324
2325 return r;
2326}
2327
2328/*
2329 * Messages supported:
2330 * create_thin <dev_id>
2331 * create_snap <dev_id> <origin_id>
2332 * delete <dev_id>
2333 * trim <dev_id> <new_size_in_sectors>
2334 * set_transaction_id <current_trans_id> <new_trans_id>
2335 * reserve_metadata_snap
2336 * release_metadata_snap
2337 */
2338static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2339{
2340 int r = -EINVAL;
2341 struct pool_c *pt = ti->private;
2342 struct pool *pool = pt->pool;
2343
2344 if (!strcasecmp(argv[0], "create_thin"))
2345 r = process_create_thin_mesg(argc, argv, pool);
2346
2347 else if (!strcasecmp(argv[0], "create_snap"))
2348 r = process_create_snap_mesg(argc, argv, pool);
2349
2350 else if (!strcasecmp(argv[0], "delete"))
2351 r = process_delete_mesg(argc, argv, pool);
2352
2353 else if (!strcasecmp(argv[0], "set_transaction_id"))
2354 r = process_set_transaction_id_mesg(argc, argv, pool);
2355
2356 else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2357 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2358
2359 else if (!strcasecmp(argv[0], "release_metadata_snap"))
2360 r = process_release_metadata_snap_mesg(argc, argv, pool);
2361
2362 else
2363 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2364
2365 if (!r) {
2366 r = dm_pool_commit_metadata(pool->pmd);
2367 if (r)
2368 DMERR("%s message: dm_pool_commit_metadata() failed, error = %d",
2369 argv[0], r);
2370 }
2371
2372 return r;
2373}
2374
2375/*
2376 * Status line is:
2377 * <transaction id> <used metadata sectors>/<total metadata sectors>
2378 * <used data sectors>/<total data sectors> <held metadata root>
2379 */
2380static int pool_status(struct dm_target *ti, status_type_t type,
2381 char *result, unsigned maxlen)
2382{
2383 int r, count;
2384 unsigned sz = 0;
2385 uint64_t transaction_id;
2386 dm_block_t nr_free_blocks_data;
2387 dm_block_t nr_free_blocks_metadata;
2388 dm_block_t nr_blocks_data;
2389 dm_block_t nr_blocks_metadata;
2390 dm_block_t held_root;
2391 char buf[BDEVNAME_SIZE];
2392 char buf2[BDEVNAME_SIZE];
2393 struct pool_c *pt = ti->private;
2394 struct pool *pool = pt->pool;
2395
2396 switch (type) {
2397 case STATUSTYPE_INFO:
2398 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2399 &transaction_id);
2400 if (r)
2401 return r;
2402
2403 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2404 &nr_free_blocks_metadata);
2405 if (r)
2406 return r;
2407
2408 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2409 if (r)
2410 return r;
2411
2412 r = dm_pool_get_free_block_count(pool->pmd,
2413 &nr_free_blocks_data);
2414 if (r)
2415 return r;
2416
2417 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2418 if (r)
2419 return r;
2420
2421 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2422 if (r)
2423 return r;
2424
2425 DMEMIT("%llu %llu/%llu %llu/%llu ",
2426 (unsigned long long)transaction_id,
2427 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2428 (unsigned long long)nr_blocks_metadata,
2429 (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2430 (unsigned long long)nr_blocks_data);
2431
2432 if (held_root)
2433 DMEMIT("%llu", held_root);
2434 else
2435 DMEMIT("-");
2436
2437 break;
2438
2439 case STATUSTYPE_TABLE:
2440 DMEMIT("%s %s %lu %llu ",
2441 format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2442 format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2443 (unsigned long)pool->sectors_per_block,
2444 (unsigned long long)pt->low_water_blocks);
2445
2446 count = !pool->pf.zero_new_blocks + !pool->pf.discard_enabled +
2447 !pt->pf.discard_passdown;
2448 DMEMIT("%u ", count);
2449
2450 if (!pool->pf.zero_new_blocks)
2451 DMEMIT("skip_block_zeroing ");
2452
2453 if (!pool->pf.discard_enabled)
2454 DMEMIT("ignore_discard ");
2455
2456 if (!pt->pf.discard_passdown)
2457 DMEMIT("no_discard_passdown ");
2458
2459 break;
2460 }
2461
2462 return 0;
2463}
2464
2465static int pool_iterate_devices(struct dm_target *ti,
2466 iterate_devices_callout_fn fn, void *data)
2467{
2468 struct pool_c *pt = ti->private;
2469
2470 return fn(ti, pt->data_dev, 0, ti->len, data);
2471}
2472
2473static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2474 struct bio_vec *biovec, int max_size)
2475{
2476 struct pool_c *pt = ti->private;
2477 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2478
2479 if (!q->merge_bvec_fn)
2480 return max_size;
2481
2482 bvm->bi_bdev = pt->data_dev->bdev;
2483
2484 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2485}
2486
2487static void set_discard_limits(struct pool *pool, struct queue_limits *limits)
2488{
2489 /*
2490 * FIXME: these limits may be incompatible with the pool's data device
2491 */
2492 limits->max_discard_sectors = pool->sectors_per_block;
2493
2494 /*
2495 * This is just a hint, and not enforced. We have to cope with
2496 * bios that overlap 2 blocks.
2497 */
2498 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2499 limits->discard_zeroes_data = pool->pf.zero_new_blocks;
2500}
2501
2502static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2503{
2504 struct pool_c *pt = ti->private;
2505 struct pool *pool = pt->pool;
2506
2507 blk_limits_io_min(limits, 0);
2508 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2509 if (pool->pf.discard_enabled)
2510 set_discard_limits(pool, limits);
2511}
2512
2513static struct target_type pool_target = {
2514 .name = "thin-pool",
2515 .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2516 DM_TARGET_IMMUTABLE,
2517 .version = {1, 2, 0},
2518 .module = THIS_MODULE,
2519 .ctr = pool_ctr,
2520 .dtr = pool_dtr,
2521 .map = pool_map,
2522 .postsuspend = pool_postsuspend,
2523 .preresume = pool_preresume,
2524 .resume = pool_resume,
2525 .message = pool_message,
2526 .status = pool_status,
2527 .merge = pool_merge,
2528 .iterate_devices = pool_iterate_devices,
2529 .io_hints = pool_io_hints,
2530};
2531
2532/*----------------------------------------------------------------
2533 * Thin target methods
2534 *--------------------------------------------------------------*/
2535static void thin_dtr(struct dm_target *ti)
2536{
2537 struct thin_c *tc = ti->private;
2538
2539 mutex_lock(&dm_thin_pool_table.mutex);
2540
2541 __pool_dec(tc->pool);
2542 dm_pool_close_thin_device(tc->td);
2543 dm_put_device(ti, tc->pool_dev);
2544 if (tc->origin_dev)
2545 dm_put_device(ti, tc->origin_dev);
2546 kfree(tc);
2547
2548 mutex_unlock(&dm_thin_pool_table.mutex);
2549}
2550
2551/*
2552 * Thin target parameters:
2553 *
2554 * <pool_dev> <dev_id> [origin_dev]
2555 *
2556 * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2557 * dev_id: the internal device identifier
2558 * origin_dev: a device external to the pool that should act as the origin
2559 *
2560 * If the pool device has discards disabled, they get disabled for the thin
2561 * device as well.
2562 */
2563static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2564{
2565 int r;
2566 struct thin_c *tc;
2567 struct dm_dev *pool_dev, *origin_dev;
2568 struct mapped_device *pool_md;
2569
2570 mutex_lock(&dm_thin_pool_table.mutex);
2571
2572 if (argc != 2 && argc != 3) {
2573 ti->error = "Invalid argument count";
2574 r = -EINVAL;
2575 goto out_unlock;
2576 }
2577
2578 tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2579 if (!tc) {
2580 ti->error = "Out of memory";
2581 r = -ENOMEM;
2582 goto out_unlock;
2583 }
2584
2585 if (argc == 3) {
2586 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2587 if (r) {
2588 ti->error = "Error opening origin device";
2589 goto bad_origin_dev;
2590 }
2591 tc->origin_dev = origin_dev;
2592 }
2593
2594 r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2595 if (r) {
2596 ti->error = "Error opening pool device";
2597 goto bad_pool_dev;
2598 }
2599 tc->pool_dev = pool_dev;
2600
2601 if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2602 ti->error = "Invalid device id";
2603 r = -EINVAL;
2604 goto bad_common;
2605 }
2606
2607 pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2608 if (!pool_md) {
2609 ti->error = "Couldn't get pool mapped device";
2610 r = -EINVAL;
2611 goto bad_common;
2612 }
2613
2614 tc->pool = __pool_table_lookup(pool_md);
2615 if (!tc->pool) {
2616 ti->error = "Couldn't find pool object";
2617 r = -EINVAL;
2618 goto bad_pool_lookup;
2619 }
2620 __pool_inc(tc->pool);
2621
2622 r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2623 if (r) {
2624 ti->error = "Couldn't open thin internal device";
2625 goto bad_thin_open;
2626 }
2627
2628 ti->split_io = tc->pool->sectors_per_block;
2629 ti->num_flush_requests = 1;
2630
2631 /* In case the pool supports discards, pass them on. */
2632 if (tc->pool->pf.discard_enabled) {
2633 ti->discards_supported = 1;
2634 ti->num_discard_requests = 1;
2635 ti->discard_zeroes_data_unsupported = 1;
2636 }
2637
2638 dm_put(pool_md);
2639
2640 mutex_unlock(&dm_thin_pool_table.mutex);
2641
2642 return 0;
2643
2644bad_thin_open:
2645 __pool_dec(tc->pool);
2646bad_pool_lookup:
2647 dm_put(pool_md);
2648bad_common:
2649 dm_put_device(ti, tc->pool_dev);
2650bad_pool_dev:
2651 if (tc->origin_dev)
2652 dm_put_device(ti, tc->origin_dev);
2653bad_origin_dev:
2654 kfree(tc);
2655out_unlock:
2656 mutex_unlock(&dm_thin_pool_table.mutex);
2657
2658 return r;
2659}
2660
2661static int thin_map(struct dm_target *ti, struct bio *bio,
2662 union map_info *map_context)
2663{
2664 bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2665
2666 return thin_bio_map(ti, bio, map_context);
2667}
2668
2669static int thin_endio(struct dm_target *ti,
2670 struct bio *bio, int err,
2671 union map_info *map_context)
2672{
2673 unsigned long flags;
2674 struct dm_thin_endio_hook *h = map_context->ptr;
2675 struct list_head work;
2676 struct dm_thin_new_mapping *m, *tmp;
2677 struct pool *pool = h->tc->pool;
2678
2679 if (h->shared_read_entry) {
2680 INIT_LIST_HEAD(&work);
2681 ds_dec(h->shared_read_entry, &work);
2682
2683 spin_lock_irqsave(&pool->lock, flags);
2684 list_for_each_entry_safe(m, tmp, &work, list) {
2685 list_del(&m->list);
2686 m->quiesced = 1;
2687 __maybe_add_mapping(m);
2688 }
2689 spin_unlock_irqrestore(&pool->lock, flags);
2690 }
2691
2692 if (h->all_io_entry) {
2693 INIT_LIST_HEAD(&work);
2694 ds_dec(h->all_io_entry, &work);
2695 spin_lock_irqsave(&pool->lock, flags);
2696 list_for_each_entry_safe(m, tmp, &work, list)
2697 list_add(&m->list, &pool->prepared_discards);
2698 spin_unlock_irqrestore(&pool->lock, flags);
2699 }
2700
2701 mempool_free(h, pool->endio_hook_pool);
2702
2703 return 0;
2704}
2705
2706static void thin_postsuspend(struct dm_target *ti)
2707{
2708 if (dm_noflush_suspending(ti))
2709 requeue_io((struct thin_c *)ti->private);
2710}
2711
2712/*
2713 * <nr mapped sectors> <highest mapped sector>
2714 */
2715static int thin_status(struct dm_target *ti, status_type_t type,
2716 char *result, unsigned maxlen)
2717{
2718 int r;
2719 ssize_t sz = 0;
2720 dm_block_t mapped, highest;
2721 char buf[BDEVNAME_SIZE];
2722 struct thin_c *tc = ti->private;
2723
2724 if (!tc->td)
2725 DMEMIT("-");
2726 else {
2727 switch (type) {
2728 case STATUSTYPE_INFO:
2729 r = dm_thin_get_mapped_count(tc->td, &mapped);
2730 if (r)
2731 return r;
2732
2733 r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2734 if (r < 0)
2735 return r;
2736
2737 DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2738 if (r)
2739 DMEMIT("%llu", ((highest + 1) *
2740 tc->pool->sectors_per_block) - 1);
2741 else
2742 DMEMIT("-");
2743 break;
2744
2745 case STATUSTYPE_TABLE:
2746 DMEMIT("%s %lu",
2747 format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2748 (unsigned long) tc->dev_id);
2749 if (tc->origin_dev)
2750 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2751 break;
2752 }
2753 }
2754
2755 return 0;
2756}
2757
2758static int thin_iterate_devices(struct dm_target *ti,
2759 iterate_devices_callout_fn fn, void *data)
2760{
2761 dm_block_t blocks;
2762 struct thin_c *tc = ti->private;
2763
2764 /*
2765 * We can't call dm_pool_get_data_dev_size() since that blocks. So
2766 * we follow a more convoluted path through to the pool's target.
2767 */
2768 if (!tc->pool->ti)
2769 return 0; /* nothing is bound */
2770
2771 blocks = tc->pool->ti->len >> tc->pool->block_shift;
2772 if (blocks)
2773 return fn(ti, tc->pool_dev, 0, tc->pool->sectors_per_block * blocks, data);
2774
2775 return 0;
2776}
2777
2778static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2779{
2780 struct thin_c *tc = ti->private;
2781 struct pool *pool = tc->pool;
2782
2783 blk_limits_io_min(limits, 0);
2784 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2785 set_discard_limits(pool, limits);
2786}
2787
2788static struct target_type thin_target = {
2789 .name = "thin",
2790 .version = {1, 1, 0},
2791 .module = THIS_MODULE,
2792 .ctr = thin_ctr,
2793 .dtr = thin_dtr,
2794 .map = thin_map,
2795 .end_io = thin_endio,
2796 .postsuspend = thin_postsuspend,
2797 .status = thin_status,
2798 .iterate_devices = thin_iterate_devices,
2799 .io_hints = thin_io_hints,
2800};
2801
2802/*----------------------------------------------------------------*/
2803
2804static int __init dm_thin_init(void)
2805{
2806 int r;
2807
2808 pool_table_init();
2809
2810 r = dm_register_target(&thin_target);
2811 if (r)
2812 return r;
2813
2814 r = dm_register_target(&pool_target);
2815 if (r)
2816 goto bad_pool_target;
2817
2818 r = -ENOMEM;
2819
2820 _cell_cache = KMEM_CACHE(dm_bio_prison_cell, 0);
2821 if (!_cell_cache)
2822 goto bad_cell_cache;
2823
2824 _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2825 if (!_new_mapping_cache)
2826 goto bad_new_mapping_cache;
2827
2828 _endio_hook_cache = KMEM_CACHE(dm_thin_endio_hook, 0);
2829 if (!_endio_hook_cache)
2830 goto bad_endio_hook_cache;
2831
2832 return 0;
2833
2834bad_endio_hook_cache:
2835 kmem_cache_destroy(_new_mapping_cache);
2836bad_new_mapping_cache:
2837 kmem_cache_destroy(_cell_cache);
2838bad_cell_cache:
2839 dm_unregister_target(&pool_target);
2840bad_pool_target:
2841 dm_unregister_target(&thin_target);
2842
2843 return r;
2844}
2845
2846static void dm_thin_exit(void)
2847{
2848 dm_unregister_target(&thin_target);
2849 dm_unregister_target(&pool_target);
2850
2851 kmem_cache_destroy(_cell_cache);
2852 kmem_cache_destroy(_new_mapping_cache);
2853 kmem_cache_destroy(_endio_hook_cache);
2854}
2855
2856module_init(dm_thin_init);
2857module_exit(dm_thin_exit);
2858
2859MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2860MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2861MODULE_LICENSE("GPL");