]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/md/dm-thin.c
Merge tag 'powerpc-4.13-2' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc...
[karo-tx-linux.git] / drivers / md / dm-thin.c
1 /*
2  * Copyright (C) 2011-2012 Red Hat UK.
3  *
4  * This file is released under the GPL.
5  */
6
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison-v1.h"
9 #include "dm.h"
10
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/log2.h>
16 #include <linux/list.h>
17 #include <linux/rculist.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
22 #include <linux/sort.h>
23 #include <linux/rbtree.h>
24
25 #define DM_MSG_PREFIX   "thin"
26
27 /*
28  * Tunable constants
29  */
30 #define ENDIO_HOOK_POOL_SIZE 1024
31 #define MAPPING_POOL_SIZE 1024
32 #define COMMIT_PERIOD HZ
33 #define NO_SPACE_TIMEOUT_SECS 60
34
35 static unsigned no_space_timeout_secs = NO_SPACE_TIMEOUT_SECS;
36
37 DECLARE_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  */
115 enum lock_space {
116         VIRTUAL,
117         PHYSICAL
118 };
119
120 static 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
129 static 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
135 static 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
145 struct throttle {
146         struct rw_semaphore lock;
147         unsigned long threshold;
148         bool throttle_applied;
149 };
150
151 static void throttle_init(struct throttle *t)
152 {
153         init_rwsem(&t->lock);
154         t->throttle_applied = false;
155 }
156
157 static void throttle_work_start(struct throttle *t)
158 {
159         t->threshold = jiffies + THROTTLE_THRESHOLD;
160 }
161
162 static 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
170 static 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
178 static void throttle_lock(struct throttle *t)
179 {
180         down_read(&t->lock);
181 }
182
183 static 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  */
195 struct dm_thin_new_mapping;
196
197 /*
198  * The pool runs in 4 modes.  Ordered in degraded order for comparisons.
199  */
200 enum 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
207 struct 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
216 struct thin_c;
217 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
218 typedef void (*process_cell_fn)(struct thin_c *tc, struct dm_bio_prison_cell *cell);
219 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
220
221 #define CELL_SORT_ARRAY_SIZE 8192
222
223 struct pool {
224         struct list_head list;
225         struct dm_target *ti;   /* Only set if a pool target is bound */
226
227         struct mapped_device *pool_md;
228         struct block_device *md_dev;
229         struct dm_pool_metadata *pmd;
230
231         dm_block_t low_water_blocks;
232         uint32_t sectors_per_block;
233         int sectors_per_block_shift;
234
235         struct pool_features pf;
236         bool low_water_triggered:1;     /* A dm event has been sent */
237         bool suspended:1;
238         bool out_of_data_space:1;
239
240         struct dm_bio_prison *prison;
241         struct dm_kcopyd_client *copier;
242
243         struct workqueue_struct *wq;
244         struct throttle throttle;
245         struct work_struct worker;
246         struct delayed_work waker;
247         struct delayed_work no_space_timeout;
248
249         unsigned long last_commit_jiffies;
250         unsigned ref_count;
251
252         spinlock_t lock;
253         struct bio_list deferred_flush_bios;
254         struct list_head prepared_mappings;
255         struct list_head prepared_discards;
256         struct list_head prepared_discards_pt2;
257         struct list_head active_thins;
258
259         struct dm_deferred_set *shared_read_ds;
260         struct dm_deferred_set *all_io_ds;
261
262         struct dm_thin_new_mapping *next_mapping;
263         mempool_t *mapping_pool;
264
265         process_bio_fn process_bio;
266         process_bio_fn process_discard;
267
268         process_cell_fn process_cell;
269         process_cell_fn process_discard_cell;
270
271         process_mapping_fn process_prepared_mapping;
272         process_mapping_fn process_prepared_discard;
273         process_mapping_fn process_prepared_discard_pt2;
274
275         struct dm_bio_prison_cell **cell_sort_array;
276 };
277
278 static enum pool_mode get_pool_mode(struct pool *pool);
279 static void metadata_operation_failed(struct pool *pool, const char *op, int r);
280
281 /*
282  * Target context for a pool.
283  */
284 struct pool_c {
285         struct dm_target *ti;
286         struct pool *pool;
287         struct dm_dev *data_dev;
288         struct dm_dev *metadata_dev;
289         struct dm_target_callbacks callbacks;
290
291         dm_block_t low_water_blocks;
292         struct pool_features requested_pf; /* Features requested during table load */
293         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
294 };
295
296 /*
297  * Target context for a thin.
298  */
299 struct thin_c {
300         struct list_head list;
301         struct dm_dev *pool_dev;
302         struct dm_dev *origin_dev;
303         sector_t origin_size;
304         dm_thin_id dev_id;
305
306         struct pool *pool;
307         struct dm_thin_device *td;
308         struct mapped_device *thin_md;
309
310         bool requeue_mode:1;
311         spinlock_t lock;
312         struct list_head deferred_cells;
313         struct bio_list deferred_bio_list;
314         struct bio_list retry_on_resume_list;
315         struct rb_root sort_bio_list; /* sorted list of deferred bios */
316
317         /*
318          * Ensures the thin is not destroyed until the worker has finished
319          * iterating the active_thins list.
320          */
321         atomic_t refcount;
322         struct completion can_destroy;
323 };
324
325 /*----------------------------------------------------------------*/
326
327 static bool block_size_is_power_of_two(struct pool *pool)
328 {
329         return pool->sectors_per_block_shift >= 0;
330 }
331
332 static sector_t block_to_sectors(struct pool *pool, dm_block_t b)
333 {
334         return block_size_is_power_of_two(pool) ?
335                 (b << pool->sectors_per_block_shift) :
336                 (b * pool->sectors_per_block);
337 }
338
339 /*----------------------------------------------------------------*/
340
341 struct discard_op {
342         struct thin_c *tc;
343         struct blk_plug plug;
344         struct bio *parent_bio;
345         struct bio *bio;
346 };
347
348 static void begin_discard(struct discard_op *op, struct thin_c *tc, struct bio *parent)
349 {
350         BUG_ON(!parent);
351
352         op->tc = tc;
353         blk_start_plug(&op->plug);
354         op->parent_bio = parent;
355         op->bio = NULL;
356 }
357
358 static int issue_discard(struct discard_op *op, dm_block_t data_b, dm_block_t data_e)
359 {
360         struct thin_c *tc = op->tc;
361         sector_t s = block_to_sectors(tc->pool, data_b);
362         sector_t len = block_to_sectors(tc->pool, data_e - data_b);
363
364         return __blkdev_issue_discard(tc->pool_dev->bdev, s, len,
365                                       GFP_NOWAIT, 0, &op->bio);
366 }
367
368 static void end_discard(struct discard_op *op, int r)
369 {
370         if (op->bio) {
371                 /*
372                  * Even if one of the calls to issue_discard failed, we
373                  * need to wait for the chain to complete.
374                  */
375                 bio_chain(op->bio, op->parent_bio);
376                 bio_set_op_attrs(op->bio, REQ_OP_DISCARD, 0);
377                 submit_bio(op->bio);
378         }
379
380         blk_finish_plug(&op->plug);
381
382         /*
383          * Even if r is set, there could be sub discards in flight that we
384          * need to wait for.
385          */
386         if (r && !op->parent_bio->bi_status)
387                 op->parent_bio->bi_status = errno_to_blk_status(r);
388         bio_endio(op->parent_bio);
389 }
390
391 /*----------------------------------------------------------------*/
392
393 /*
394  * wake_worker() is used when new work is queued and when pool_resume is
395  * ready to continue deferred IO processing.
396  */
397 static void wake_worker(struct pool *pool)
398 {
399         queue_work(pool->wq, &pool->worker);
400 }
401
402 /*----------------------------------------------------------------*/
403
404 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
405                       struct dm_bio_prison_cell **cell_result)
406 {
407         int r;
408         struct dm_bio_prison_cell *cell_prealloc;
409
410         /*
411          * Allocate a cell from the prison's mempool.
412          * This might block but it can't fail.
413          */
414         cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
415
416         r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
417         if (r)
418                 /*
419                  * We reused an old cell; we can get rid of
420                  * the new one.
421                  */
422                 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
423
424         return r;
425 }
426
427 static void cell_release(struct pool *pool,
428                          struct dm_bio_prison_cell *cell,
429                          struct bio_list *bios)
430 {
431         dm_cell_release(pool->prison, cell, bios);
432         dm_bio_prison_free_cell(pool->prison, cell);
433 }
434
435 static void cell_visit_release(struct pool *pool,
436                                void (*fn)(void *, struct dm_bio_prison_cell *),
437                                void *context,
438                                struct dm_bio_prison_cell *cell)
439 {
440         dm_cell_visit_release(pool->prison, fn, context, cell);
441         dm_bio_prison_free_cell(pool->prison, cell);
442 }
443
444 static void cell_release_no_holder(struct pool *pool,
445                                    struct dm_bio_prison_cell *cell,
446                                    struct bio_list *bios)
447 {
448         dm_cell_release_no_holder(pool->prison, cell, bios);
449         dm_bio_prison_free_cell(pool->prison, cell);
450 }
451
452 static void cell_error_with_code(struct pool *pool,
453                 struct dm_bio_prison_cell *cell, blk_status_t error_code)
454 {
455         dm_cell_error(pool->prison, cell, error_code);
456         dm_bio_prison_free_cell(pool->prison, cell);
457 }
458
459 static blk_status_t get_pool_io_error_code(struct pool *pool)
460 {
461         return pool->out_of_data_space ? BLK_STS_NOSPC : BLK_STS_IOERR;
462 }
463
464 static void cell_error(struct pool *pool, struct dm_bio_prison_cell *cell)
465 {
466         cell_error_with_code(pool, cell, get_pool_io_error_code(pool));
467 }
468
469 static void cell_success(struct pool *pool, struct dm_bio_prison_cell *cell)
470 {
471         cell_error_with_code(pool, cell, 0);
472 }
473
474 static void cell_requeue(struct pool *pool, struct dm_bio_prison_cell *cell)
475 {
476         cell_error_with_code(pool, cell, BLK_STS_DM_REQUEUE);
477 }
478
479 /*----------------------------------------------------------------*/
480
481 /*
482  * A global list of pools that uses a struct mapped_device as a key.
483  */
484 static struct dm_thin_pool_table {
485         struct mutex mutex;
486         struct list_head pools;
487 } dm_thin_pool_table;
488
489 static void pool_table_init(void)
490 {
491         mutex_init(&dm_thin_pool_table.mutex);
492         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
493 }
494
495 static void __pool_table_insert(struct pool *pool)
496 {
497         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
498         list_add(&pool->list, &dm_thin_pool_table.pools);
499 }
500
501 static void __pool_table_remove(struct pool *pool)
502 {
503         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
504         list_del(&pool->list);
505 }
506
507 static struct pool *__pool_table_lookup(struct mapped_device *md)
508 {
509         struct pool *pool = NULL, *tmp;
510
511         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
512
513         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
514                 if (tmp->pool_md == md) {
515                         pool = tmp;
516                         break;
517                 }
518         }
519
520         return pool;
521 }
522
523 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
524 {
525         struct pool *pool = NULL, *tmp;
526
527         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
528
529         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
530                 if (tmp->md_dev == md_dev) {
531                         pool = tmp;
532                         break;
533                 }
534         }
535
536         return pool;
537 }
538
539 /*----------------------------------------------------------------*/
540
541 struct dm_thin_endio_hook {
542         struct thin_c *tc;
543         struct dm_deferred_entry *shared_read_entry;
544         struct dm_deferred_entry *all_io_entry;
545         struct dm_thin_new_mapping *overwrite_mapping;
546         struct rb_node rb_node;
547         struct dm_bio_prison_cell *cell;
548 };
549
550 static void __merge_bio_list(struct bio_list *bios, struct bio_list *master)
551 {
552         bio_list_merge(bios, master);
553         bio_list_init(master);
554 }
555
556 static void error_bio_list(struct bio_list *bios, blk_status_t error)
557 {
558         struct bio *bio;
559
560         while ((bio = bio_list_pop(bios))) {
561                 bio->bi_status = error;
562                 bio_endio(bio);
563         }
564 }
565
566 static void error_thin_bio_list(struct thin_c *tc, struct bio_list *master,
567                 blk_status_t error)
568 {
569         struct bio_list bios;
570         unsigned long flags;
571
572         bio_list_init(&bios);
573
574         spin_lock_irqsave(&tc->lock, flags);
575         __merge_bio_list(&bios, master);
576         spin_unlock_irqrestore(&tc->lock, flags);
577
578         error_bio_list(&bios, error);
579 }
580
581 static void requeue_deferred_cells(struct thin_c *tc)
582 {
583         struct pool *pool = tc->pool;
584         unsigned long flags;
585         struct list_head cells;
586         struct dm_bio_prison_cell *cell, *tmp;
587
588         INIT_LIST_HEAD(&cells);
589
590         spin_lock_irqsave(&tc->lock, flags);
591         list_splice_init(&tc->deferred_cells, &cells);
592         spin_unlock_irqrestore(&tc->lock, flags);
593
594         list_for_each_entry_safe(cell, tmp, &cells, user_list)
595                 cell_requeue(pool, cell);
596 }
597
598 static void requeue_io(struct thin_c *tc)
599 {
600         struct bio_list bios;
601         unsigned long flags;
602
603         bio_list_init(&bios);
604
605         spin_lock_irqsave(&tc->lock, flags);
606         __merge_bio_list(&bios, &tc->deferred_bio_list);
607         __merge_bio_list(&bios, &tc->retry_on_resume_list);
608         spin_unlock_irqrestore(&tc->lock, flags);
609
610         error_bio_list(&bios, BLK_STS_DM_REQUEUE);
611         requeue_deferred_cells(tc);
612 }
613
614 static void error_retry_list_with_code(struct pool *pool, blk_status_t error)
615 {
616         struct thin_c *tc;
617
618         rcu_read_lock();
619         list_for_each_entry_rcu(tc, &pool->active_thins, list)
620                 error_thin_bio_list(tc, &tc->retry_on_resume_list, error);
621         rcu_read_unlock();
622 }
623
624 static void error_retry_list(struct pool *pool)
625 {
626         error_retry_list_with_code(pool, get_pool_io_error_code(pool));
627 }
628
629 /*
630  * This section of code contains the logic for processing a thin device's IO.
631  * Much of the code depends on pool object resources (lists, workqueues, etc)
632  * but most is exclusively called from the thin target rather than the thin-pool
633  * target.
634  */
635
636 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
637 {
638         struct pool *pool = tc->pool;
639         sector_t block_nr = bio->bi_iter.bi_sector;
640
641         if (block_size_is_power_of_two(pool))
642                 block_nr >>= pool->sectors_per_block_shift;
643         else
644                 (void) sector_div(block_nr, pool->sectors_per_block);
645
646         return block_nr;
647 }
648
649 /*
650  * Returns the _complete_ blocks that this bio covers.
651  */
652 static void get_bio_block_range(struct thin_c *tc, struct bio *bio,
653                                 dm_block_t *begin, dm_block_t *end)
654 {
655         struct pool *pool = tc->pool;
656         sector_t b = bio->bi_iter.bi_sector;
657         sector_t e = b + (bio->bi_iter.bi_size >> SECTOR_SHIFT);
658
659         b += pool->sectors_per_block - 1ull; /* so we round up */
660
661         if (block_size_is_power_of_two(pool)) {
662                 b >>= pool->sectors_per_block_shift;
663                 e >>= pool->sectors_per_block_shift;
664         } else {
665                 (void) sector_div(b, pool->sectors_per_block);
666                 (void) sector_div(e, pool->sectors_per_block);
667         }
668
669         if (e < b)
670                 /* Can happen if the bio is within a single block. */
671                 e = b;
672
673         *begin = b;
674         *end = e;
675 }
676
677 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
678 {
679         struct pool *pool = tc->pool;
680         sector_t bi_sector = bio->bi_iter.bi_sector;
681
682         bio->bi_bdev = tc->pool_dev->bdev;
683         if (block_size_is_power_of_two(pool))
684                 bio->bi_iter.bi_sector =
685                         (block << pool->sectors_per_block_shift) |
686                         (bi_sector & (pool->sectors_per_block - 1));
687         else
688                 bio->bi_iter.bi_sector = (block * pool->sectors_per_block) +
689                                  sector_div(bi_sector, pool->sectors_per_block);
690 }
691
692 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
693 {
694         bio->bi_bdev = tc->origin_dev->bdev;
695 }
696
697 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
698 {
699         return op_is_flush(bio->bi_opf) &&
700                 dm_thin_changed_this_transaction(tc->td);
701 }
702
703 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
704 {
705         struct dm_thin_endio_hook *h;
706
707         if (bio_op(bio) == REQ_OP_DISCARD)
708                 return;
709
710         h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
711         h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
712 }
713
714 static void issue(struct thin_c *tc, struct bio *bio)
715 {
716         struct pool *pool = tc->pool;
717         unsigned long flags;
718
719         if (!bio_triggers_commit(tc, bio)) {
720                 generic_make_request(bio);
721                 return;
722         }
723
724         /*
725          * Complete bio with an error if earlier I/O caused changes to
726          * the metadata that can't be committed e.g, due to I/O errors
727          * on the metadata device.
728          */
729         if (dm_thin_aborted_changes(tc->td)) {
730                 bio_io_error(bio);
731                 return;
732         }
733
734         /*
735          * Batch together any bios that trigger commits and then issue a
736          * single commit for them in process_deferred_bios().
737          */
738         spin_lock_irqsave(&pool->lock, flags);
739         bio_list_add(&pool->deferred_flush_bios, bio);
740         spin_unlock_irqrestore(&pool->lock, flags);
741 }
742
743 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
744 {
745         remap_to_origin(tc, bio);
746         issue(tc, bio);
747 }
748
749 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
750                             dm_block_t block)
751 {
752         remap(tc, bio, block);
753         issue(tc, bio);
754 }
755
756 /*----------------------------------------------------------------*/
757
758 /*
759  * Bio endio functions.
760  */
761 struct dm_thin_new_mapping {
762         struct list_head list;
763
764         bool pass_discard:1;
765         bool maybe_shared:1;
766
767         /*
768          * Track quiescing, copying and zeroing preparation actions.  When this
769          * counter hits zero the block is prepared and can be inserted into the
770          * btree.
771          */
772         atomic_t prepare_actions;
773
774         blk_status_t status;
775         struct thin_c *tc;
776         dm_block_t virt_begin, virt_end;
777         dm_block_t data_block;
778         struct dm_bio_prison_cell *cell;
779
780         /*
781          * If the bio covers the whole area of a block then we can avoid
782          * zeroing or copying.  Instead this bio is hooked.  The bio will
783          * still be in the cell, so care has to be taken to avoid issuing
784          * the bio twice.
785          */
786         struct bio *bio;
787         bio_end_io_t *saved_bi_end_io;
788 };
789
790 static void __complete_mapping_preparation(struct dm_thin_new_mapping *m)
791 {
792         struct pool *pool = m->tc->pool;
793
794         if (atomic_dec_and_test(&m->prepare_actions)) {
795                 list_add_tail(&m->list, &pool->prepared_mappings);
796                 wake_worker(pool);
797         }
798 }
799
800 static void complete_mapping_preparation(struct dm_thin_new_mapping *m)
801 {
802         unsigned long flags;
803         struct pool *pool = m->tc->pool;
804
805         spin_lock_irqsave(&pool->lock, flags);
806         __complete_mapping_preparation(m);
807         spin_unlock_irqrestore(&pool->lock, flags);
808 }
809
810 static void copy_complete(int read_err, unsigned long write_err, void *context)
811 {
812         struct dm_thin_new_mapping *m = context;
813
814         m->status = read_err || write_err ? BLK_STS_IOERR : 0;
815         complete_mapping_preparation(m);
816 }
817
818 static void overwrite_endio(struct bio *bio)
819 {
820         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
821         struct dm_thin_new_mapping *m = h->overwrite_mapping;
822
823         bio->bi_end_io = m->saved_bi_end_io;
824
825         m->status = bio->bi_status;
826         complete_mapping_preparation(m);
827 }
828
829 /*----------------------------------------------------------------*/
830
831 /*
832  * Workqueue.
833  */
834
835 /*
836  * Prepared mapping jobs.
837  */
838
839 /*
840  * This sends the bios in the cell, except the original holder, back
841  * to the deferred_bios list.
842  */
843 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
844 {
845         struct pool *pool = tc->pool;
846         unsigned long flags;
847
848         spin_lock_irqsave(&tc->lock, flags);
849         cell_release_no_holder(pool, cell, &tc->deferred_bio_list);
850         spin_unlock_irqrestore(&tc->lock, flags);
851
852         wake_worker(pool);
853 }
854
855 static void thin_defer_bio(struct thin_c *tc, struct bio *bio);
856
857 struct remap_info {
858         struct thin_c *tc;
859         struct bio_list defer_bios;
860         struct bio_list issue_bios;
861 };
862
863 static void __inc_remap_and_issue_cell(void *context,
864                                        struct dm_bio_prison_cell *cell)
865 {
866         struct remap_info *info = context;
867         struct bio *bio;
868
869         while ((bio = bio_list_pop(&cell->bios))) {
870                 if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD)
871                         bio_list_add(&info->defer_bios, bio);
872                 else {
873                         inc_all_io_entry(info->tc->pool, bio);
874
875                         /*
876                          * We can't issue the bios with the bio prison lock
877                          * held, so we add them to a list to issue on
878                          * return from this function.
879                          */
880                         bio_list_add(&info->issue_bios, bio);
881                 }
882         }
883 }
884
885 static void inc_remap_and_issue_cell(struct thin_c *tc,
886                                      struct dm_bio_prison_cell *cell,
887                                      dm_block_t block)
888 {
889         struct bio *bio;
890         struct remap_info info;
891
892         info.tc = tc;
893         bio_list_init(&info.defer_bios);
894         bio_list_init(&info.issue_bios);
895
896         /*
897          * We have to be careful to inc any bios we're about to issue
898          * before the cell is released, and avoid a race with new bios
899          * being added to the cell.
900          */
901         cell_visit_release(tc->pool, __inc_remap_and_issue_cell,
902                            &info, cell);
903
904         while ((bio = bio_list_pop(&info.defer_bios)))
905                 thin_defer_bio(tc, bio);
906
907         while ((bio = bio_list_pop(&info.issue_bios)))
908                 remap_and_issue(info.tc, bio, block);
909 }
910
911 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
912 {
913         cell_error(m->tc->pool, m->cell);
914         list_del(&m->list);
915         mempool_free(m, m->tc->pool->mapping_pool);
916 }
917
918 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
919 {
920         struct thin_c *tc = m->tc;
921         struct pool *pool = tc->pool;
922         struct bio *bio = m->bio;
923         int r;
924
925         if (m->status) {
926                 cell_error(pool, m->cell);
927                 goto out;
928         }
929
930         /*
931          * Commit the prepared block into the mapping btree.
932          * Any I/O for this block arriving after this point will get
933          * remapped to it directly.
934          */
935         r = dm_thin_insert_block(tc->td, m->virt_begin, m->data_block);
936         if (r) {
937                 metadata_operation_failed(pool, "dm_thin_insert_block", r);
938                 cell_error(pool, m->cell);
939                 goto out;
940         }
941
942         /*
943          * Release any bios held while the block was being provisioned.
944          * If we are processing a write bio that completely covers the block,
945          * we already processed it so can ignore it now when processing
946          * the bios in the cell.
947          */
948         if (bio) {
949                 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
950                 bio_endio(bio);
951         } else {
952                 inc_all_io_entry(tc->pool, m->cell->holder);
953                 remap_and_issue(tc, m->cell->holder, m->data_block);
954                 inc_remap_and_issue_cell(tc, m->cell, m->data_block);
955         }
956
957 out:
958         list_del(&m->list);
959         mempool_free(m, pool->mapping_pool);
960 }
961
962 /*----------------------------------------------------------------*/
963
964 static void free_discard_mapping(struct dm_thin_new_mapping *m)
965 {
966         struct thin_c *tc = m->tc;
967         if (m->cell)
968                 cell_defer_no_holder(tc, m->cell);
969         mempool_free(m, tc->pool->mapping_pool);
970 }
971
972 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
973 {
974         bio_io_error(m->bio);
975         free_discard_mapping(m);
976 }
977
978 static void process_prepared_discard_success(struct dm_thin_new_mapping *m)
979 {
980         bio_endio(m->bio);
981         free_discard_mapping(m);
982 }
983
984 static void process_prepared_discard_no_passdown(struct dm_thin_new_mapping *m)
985 {
986         int r;
987         struct thin_c *tc = m->tc;
988
989         r = dm_thin_remove_range(tc->td, m->cell->key.block_begin, m->cell->key.block_end);
990         if (r) {
991                 metadata_operation_failed(tc->pool, "dm_thin_remove_range", r);
992                 bio_io_error(m->bio);
993         } else
994                 bio_endio(m->bio);
995
996         cell_defer_no_holder(tc, m->cell);
997         mempool_free(m, tc->pool->mapping_pool);
998 }
999
1000 /*----------------------------------------------------------------*/
1001
1002 static void passdown_double_checking_shared_status(struct dm_thin_new_mapping *m,
1003                                                    struct bio *discard_parent)
1004 {
1005         /*
1006          * We've already unmapped this range of blocks, but before we
1007          * passdown we have to check that these blocks are now unused.
1008          */
1009         int r = 0;
1010         bool used = true;
1011         struct thin_c *tc = m->tc;
1012         struct pool *pool = tc->pool;
1013         dm_block_t b = m->data_block, e, end = m->data_block + m->virt_end - m->virt_begin;
1014         struct discard_op op;
1015
1016         begin_discard(&op, tc, discard_parent);
1017         while (b != end) {
1018                 /* find start of unmapped run */
1019                 for (; b < end; b++) {
1020                         r = dm_pool_block_is_used(pool->pmd, b, &used);
1021                         if (r)
1022                                 goto out;
1023
1024                         if (!used)
1025                                 break;
1026                 }
1027
1028                 if (b == end)
1029                         break;
1030
1031                 /* find end of run */
1032                 for (e = b + 1; e != end; e++) {
1033                         r = dm_pool_block_is_used(pool->pmd, e, &used);
1034                         if (r)
1035                                 goto out;
1036
1037                         if (used)
1038                                 break;
1039                 }
1040
1041                 r = issue_discard(&op, b, e);
1042                 if (r)
1043                         goto out;
1044
1045                 b = e;
1046         }
1047 out:
1048         end_discard(&op, r);
1049 }
1050
1051 static void queue_passdown_pt2(struct dm_thin_new_mapping *m)
1052 {
1053         unsigned long flags;
1054         struct pool *pool = m->tc->pool;
1055
1056         spin_lock_irqsave(&pool->lock, flags);
1057         list_add_tail(&m->list, &pool->prepared_discards_pt2);
1058         spin_unlock_irqrestore(&pool->lock, flags);
1059         wake_worker(pool);
1060 }
1061
1062 static void passdown_endio(struct bio *bio)
1063 {
1064         /*
1065          * It doesn't matter if the passdown discard failed, we still want
1066          * to unmap (we ignore err).
1067          */
1068         queue_passdown_pt2(bio->bi_private);
1069         bio_put(bio);
1070 }
1071
1072 static void process_prepared_discard_passdown_pt1(struct dm_thin_new_mapping *m)
1073 {
1074         int r;
1075         struct thin_c *tc = m->tc;
1076         struct pool *pool = tc->pool;
1077         struct bio *discard_parent;
1078         dm_block_t data_end = m->data_block + (m->virt_end - m->virt_begin);
1079
1080         /*
1081          * Only this thread allocates blocks, so we can be sure that the
1082          * newly unmapped blocks will not be allocated before the end of
1083          * the function.
1084          */
1085         r = dm_thin_remove_range(tc->td, m->virt_begin, m->virt_end);
1086         if (r) {
1087                 metadata_operation_failed(pool, "dm_thin_remove_range", r);
1088                 bio_io_error(m->bio);
1089                 cell_defer_no_holder(tc, m->cell);
1090                 mempool_free(m, pool->mapping_pool);
1091                 return;
1092         }
1093
1094         /*
1095          * Increment the unmapped blocks.  This prevents a race between the
1096          * passdown io and reallocation of freed blocks.
1097          */
1098         r = dm_pool_inc_data_range(pool->pmd, m->data_block, data_end);
1099         if (r) {
1100                 metadata_operation_failed(pool, "dm_pool_inc_data_range", r);
1101                 bio_io_error(m->bio);
1102                 cell_defer_no_holder(tc, m->cell);
1103                 mempool_free(m, pool->mapping_pool);
1104                 return;
1105         }
1106
1107         discard_parent = bio_alloc(GFP_NOIO, 1);
1108         if (!discard_parent) {
1109                 DMWARN("%s: unable to allocate top level discard bio for passdown. Skipping passdown.",
1110                        dm_device_name(tc->pool->pool_md));
1111                 queue_passdown_pt2(m);
1112
1113         } else {
1114                 discard_parent->bi_end_io = passdown_endio;
1115                 discard_parent->bi_private = m;
1116
1117                 if (m->maybe_shared)
1118                         passdown_double_checking_shared_status(m, discard_parent);
1119                 else {
1120                         struct discard_op op;
1121
1122                         begin_discard(&op, tc, discard_parent);
1123                         r = issue_discard(&op, m->data_block, data_end);
1124                         end_discard(&op, r);
1125                 }
1126         }
1127 }
1128
1129 static void process_prepared_discard_passdown_pt2(struct dm_thin_new_mapping *m)
1130 {
1131         int r;
1132         struct thin_c *tc = m->tc;
1133         struct pool *pool = tc->pool;
1134
1135         /*
1136          * The passdown has completed, so now we can decrement all those
1137          * unmapped blocks.
1138          */
1139         r = dm_pool_dec_data_range(pool->pmd, m->data_block,
1140                                    m->data_block + (m->virt_end - m->virt_begin));
1141         if (r) {
1142                 metadata_operation_failed(pool, "dm_pool_dec_data_range", r);
1143                 bio_io_error(m->bio);
1144         } else
1145                 bio_endio(m->bio);
1146
1147         cell_defer_no_holder(tc, m->cell);
1148         mempool_free(m, pool->mapping_pool);
1149 }
1150
1151 static void process_prepared(struct pool *pool, struct list_head *head,
1152                              process_mapping_fn *fn)
1153 {
1154         unsigned long flags;
1155         struct list_head maps;
1156         struct dm_thin_new_mapping *m, *tmp;
1157
1158         INIT_LIST_HEAD(&maps);
1159         spin_lock_irqsave(&pool->lock, flags);
1160         list_splice_init(head, &maps);
1161         spin_unlock_irqrestore(&pool->lock, flags);
1162
1163         list_for_each_entry_safe(m, tmp, &maps, list)
1164                 (*fn)(m);
1165 }
1166
1167 /*
1168  * Deferred bio jobs.
1169  */
1170 static int io_overlaps_block(struct pool *pool, struct bio *bio)
1171 {
1172         return bio->bi_iter.bi_size ==
1173                 (pool->sectors_per_block << SECTOR_SHIFT);
1174 }
1175
1176 static int io_overwrites_block(struct pool *pool, struct bio *bio)
1177 {
1178         return (bio_data_dir(bio) == WRITE) &&
1179                 io_overlaps_block(pool, bio);
1180 }
1181
1182 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
1183                                bio_end_io_t *fn)
1184 {
1185         *save = bio->bi_end_io;
1186         bio->bi_end_io = fn;
1187 }
1188
1189 static int ensure_next_mapping(struct pool *pool)
1190 {
1191         if (pool->next_mapping)
1192                 return 0;
1193
1194         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
1195
1196         return pool->next_mapping ? 0 : -ENOMEM;
1197 }
1198
1199 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
1200 {
1201         struct dm_thin_new_mapping *m = pool->next_mapping;
1202
1203         BUG_ON(!pool->next_mapping);
1204
1205         memset(m, 0, sizeof(struct dm_thin_new_mapping));
1206         INIT_LIST_HEAD(&m->list);
1207         m->bio = NULL;
1208
1209         pool->next_mapping = NULL;
1210
1211         return m;
1212 }
1213
1214 static void ll_zero(struct thin_c *tc, struct dm_thin_new_mapping *m,
1215                     sector_t begin, sector_t end)
1216 {
1217         int r;
1218         struct dm_io_region to;
1219
1220         to.bdev = tc->pool_dev->bdev;
1221         to.sector = begin;
1222         to.count = end - begin;
1223
1224         r = dm_kcopyd_zero(tc->pool->copier, 1, &to, 0, copy_complete, m);
1225         if (r < 0) {
1226                 DMERR_LIMIT("dm_kcopyd_zero() failed");
1227                 copy_complete(1, 1, m);
1228         }
1229 }
1230
1231 static void remap_and_issue_overwrite(struct thin_c *tc, struct bio *bio,
1232                                       dm_block_t data_begin,
1233                                       struct dm_thin_new_mapping *m)
1234 {
1235         struct pool *pool = tc->pool;
1236         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1237
1238         h->overwrite_mapping = m;
1239         m->bio = bio;
1240         save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1241         inc_all_io_entry(pool, bio);
1242         remap_and_issue(tc, bio, data_begin);
1243 }
1244
1245 /*
1246  * A partial copy also needs to zero the uncopied region.
1247  */
1248 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
1249                           struct dm_dev *origin, dm_block_t data_origin,
1250                           dm_block_t data_dest,
1251                           struct dm_bio_prison_cell *cell, struct bio *bio,
1252                           sector_t len)
1253 {
1254         int r;
1255         struct pool *pool = tc->pool;
1256         struct dm_thin_new_mapping *m = get_next_mapping(pool);
1257
1258         m->tc = tc;
1259         m->virt_begin = virt_block;
1260         m->virt_end = virt_block + 1u;
1261         m->data_block = data_dest;
1262         m->cell = cell;
1263
1264         /*
1265          * quiesce action + copy action + an extra reference held for the
1266          * duration of this function (we may need to inc later for a
1267          * partial zero).
1268          */
1269         atomic_set(&m->prepare_actions, 3);
1270
1271         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
1272                 complete_mapping_preparation(m); /* already quiesced */
1273
1274         /*
1275          * IO to pool_dev remaps to the pool target's data_dev.
1276          *
1277          * If the whole block of data is being overwritten, we can issue the
1278          * bio immediately. Otherwise we use kcopyd to clone the data first.
1279          */
1280         if (io_overwrites_block(pool, bio))
1281                 remap_and_issue_overwrite(tc, bio, data_dest, m);
1282         else {
1283                 struct dm_io_region from, to;
1284
1285                 from.bdev = origin->bdev;
1286                 from.sector = data_origin * pool->sectors_per_block;
1287                 from.count = len;
1288
1289                 to.bdev = tc->pool_dev->bdev;
1290                 to.sector = data_dest * pool->sectors_per_block;
1291                 to.count = len;
1292
1293                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1294                                    0, copy_complete, m);
1295                 if (r < 0) {
1296                         DMERR_LIMIT("dm_kcopyd_copy() failed");
1297                         copy_complete(1, 1, m);
1298
1299                         /*
1300                          * We allow the zero to be issued, to simplify the
1301                          * error path.  Otherwise we'd need to start
1302                          * worrying about decrementing the prepare_actions
1303                          * counter.
1304                          */
1305                 }
1306
1307                 /*
1308                  * Do we need to zero a tail region?
1309                  */
1310                 if (len < pool->sectors_per_block && pool->pf.zero_new_blocks) {
1311                         atomic_inc(&m->prepare_actions);
1312                         ll_zero(tc, m,
1313                                 data_dest * pool->sectors_per_block + len,
1314                                 (data_dest + 1) * pool->sectors_per_block);
1315                 }
1316         }
1317
1318         complete_mapping_preparation(m); /* drop our ref */
1319 }
1320
1321 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1322                                    dm_block_t data_origin, dm_block_t data_dest,
1323                                    struct dm_bio_prison_cell *cell, struct bio *bio)
1324 {
1325         schedule_copy(tc, virt_block, tc->pool_dev,
1326                       data_origin, data_dest, cell, bio,
1327                       tc->pool->sectors_per_block);
1328 }
1329
1330 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1331                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
1332                           struct bio *bio)
1333 {
1334         struct pool *pool = tc->pool;
1335         struct dm_thin_new_mapping *m = get_next_mapping(pool);
1336
1337         atomic_set(&m->prepare_actions, 1); /* no need to quiesce */
1338         m->tc = tc;
1339         m->virt_begin = virt_block;
1340         m->virt_end = virt_block + 1u;
1341         m->data_block = data_block;
1342         m->cell = cell;
1343
1344         /*
1345          * If the whole block of data is being overwritten or we are not
1346          * zeroing pre-existing data, we can issue the bio immediately.
1347          * Otherwise we use kcopyd to zero the data first.
1348          */
1349         if (pool->pf.zero_new_blocks) {
1350                 if (io_overwrites_block(pool, bio))
1351                         remap_and_issue_overwrite(tc, bio, data_block, m);
1352                 else
1353                         ll_zero(tc, m, data_block * pool->sectors_per_block,
1354                                 (data_block + 1) * pool->sectors_per_block);
1355         } else
1356                 process_prepared_mapping(m);
1357 }
1358
1359 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1360                                    dm_block_t data_dest,
1361                                    struct dm_bio_prison_cell *cell, struct bio *bio)
1362 {
1363         struct pool *pool = tc->pool;
1364         sector_t virt_block_begin = virt_block * pool->sectors_per_block;
1365         sector_t virt_block_end = (virt_block + 1) * pool->sectors_per_block;
1366
1367         if (virt_block_end <= tc->origin_size)
1368                 schedule_copy(tc, virt_block, tc->origin_dev,
1369                               virt_block, data_dest, cell, bio,
1370                               pool->sectors_per_block);
1371
1372         else if (virt_block_begin < tc->origin_size)
1373                 schedule_copy(tc, virt_block, tc->origin_dev,
1374                               virt_block, data_dest, cell, bio,
1375                               tc->origin_size - virt_block_begin);
1376
1377         else
1378                 schedule_zero(tc, virt_block, data_dest, cell, bio);
1379 }
1380
1381 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode);
1382
1383 static void check_for_space(struct pool *pool)
1384 {
1385         int r;
1386         dm_block_t nr_free;
1387
1388         if (get_pool_mode(pool) != PM_OUT_OF_DATA_SPACE)
1389                 return;
1390
1391         r = dm_pool_get_free_block_count(pool->pmd, &nr_free);
1392         if (r)
1393                 return;
1394
1395         if (nr_free)
1396                 set_pool_mode(pool, PM_WRITE);
1397 }
1398
1399 /*
1400  * A non-zero return indicates read_only or fail_io mode.
1401  * Many callers don't care about the return value.
1402  */
1403 static int commit(struct pool *pool)
1404 {
1405         int r;
1406
1407         if (get_pool_mode(pool) >= PM_READ_ONLY)
1408                 return -EINVAL;
1409
1410         r = dm_pool_commit_metadata(pool->pmd);
1411         if (r)
1412                 metadata_operation_failed(pool, "dm_pool_commit_metadata", r);
1413         else
1414                 check_for_space(pool);
1415
1416         return r;
1417 }
1418
1419 static void check_low_water_mark(struct pool *pool, dm_block_t free_blocks)
1420 {
1421         unsigned long flags;
1422
1423         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1424                 DMWARN("%s: reached low water mark for data device: sending event.",
1425                        dm_device_name(pool->pool_md));
1426                 spin_lock_irqsave(&pool->lock, flags);
1427                 pool->low_water_triggered = true;
1428                 spin_unlock_irqrestore(&pool->lock, flags);
1429                 dm_table_event(pool->ti->table);
1430         }
1431 }
1432
1433 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1434 {
1435         int r;
1436         dm_block_t free_blocks;
1437         struct pool *pool = tc->pool;
1438
1439         if (WARN_ON(get_pool_mode(pool) != PM_WRITE))
1440                 return -EINVAL;
1441
1442         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1443         if (r) {
1444                 metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1445                 return r;
1446         }
1447
1448         check_low_water_mark(pool, free_blocks);
1449
1450         if (!free_blocks) {
1451                 /*
1452                  * Try to commit to see if that will free up some
1453                  * more space.
1454                  */
1455                 r = commit(pool);
1456                 if (r)
1457                         return r;
1458
1459                 r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1460                 if (r) {
1461                         metadata_operation_failed(pool, "dm_pool_get_free_block_count", r);
1462                         return r;
1463                 }
1464
1465                 if (!free_blocks) {
1466                         set_pool_mode(pool, PM_OUT_OF_DATA_SPACE);
1467                         return -ENOSPC;
1468                 }
1469         }
1470
1471         r = dm_pool_alloc_data_block(pool->pmd, result);
1472         if (r) {
1473                 metadata_operation_failed(pool, "dm_pool_alloc_data_block", r);
1474                 return r;
1475         }
1476
1477         return 0;
1478 }
1479
1480 /*
1481  * If we have run out of space, queue bios until the device is
1482  * resumed, presumably after having been reloaded with more space.
1483  */
1484 static void retry_on_resume(struct bio *bio)
1485 {
1486         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1487         struct thin_c *tc = h->tc;
1488         unsigned long flags;
1489
1490         spin_lock_irqsave(&tc->lock, flags);
1491         bio_list_add(&tc->retry_on_resume_list, bio);
1492         spin_unlock_irqrestore(&tc->lock, flags);
1493 }
1494
1495 static blk_status_t should_error_unserviceable_bio(struct pool *pool)
1496 {
1497         enum pool_mode m = get_pool_mode(pool);
1498
1499         switch (m) {
1500         case PM_WRITE:
1501                 /* Shouldn't get here */
1502                 DMERR_LIMIT("bio unserviceable, yet pool is in PM_WRITE mode");
1503                 return BLK_STS_IOERR;
1504
1505         case PM_OUT_OF_DATA_SPACE:
1506                 return pool->pf.error_if_no_space ? BLK_STS_NOSPC : 0;
1507
1508         case PM_READ_ONLY:
1509         case PM_FAIL:
1510                 return BLK_STS_IOERR;
1511         default:
1512                 /* Shouldn't get here */
1513                 DMERR_LIMIT("bio unserviceable, yet pool has an unknown mode");
1514                 return BLK_STS_IOERR;
1515         }
1516 }
1517
1518 static void handle_unserviceable_bio(struct pool *pool, struct bio *bio)
1519 {
1520         blk_status_t error = should_error_unserviceable_bio(pool);
1521
1522         if (error) {
1523                 bio->bi_status = error;
1524                 bio_endio(bio);
1525         } else
1526                 retry_on_resume(bio);
1527 }
1528
1529 static void retry_bios_on_resume(struct pool *pool, struct dm_bio_prison_cell *cell)
1530 {
1531         struct bio *bio;
1532         struct bio_list bios;
1533         blk_status_t error;
1534
1535         error = should_error_unserviceable_bio(pool);
1536         if (error) {
1537                 cell_error_with_code(pool, cell, error);
1538                 return;
1539         }
1540
1541         bio_list_init(&bios);
1542         cell_release(pool, cell, &bios);
1543
1544         while ((bio = bio_list_pop(&bios)))
1545                 retry_on_resume(bio);
1546 }
1547
1548 static void process_discard_cell_no_passdown(struct thin_c *tc,
1549                                              struct dm_bio_prison_cell *virt_cell)
1550 {
1551         struct pool *pool = tc->pool;
1552         struct dm_thin_new_mapping *m = get_next_mapping(pool);
1553
1554         /*
1555          * We don't need to lock the data blocks, since there's no
1556          * passdown.  We only lock data blocks for allocation and breaking sharing.
1557          */
1558         m->tc = tc;
1559         m->virt_begin = virt_cell->key.block_begin;
1560         m->virt_end = virt_cell->key.block_end;
1561         m->cell = virt_cell;
1562         m->bio = virt_cell->holder;
1563
1564         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1565                 pool->process_prepared_discard(m);
1566 }
1567
1568 static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t end,
1569                                  struct bio *bio)
1570 {
1571         struct pool *pool = tc->pool;
1572
1573         int r;
1574         bool maybe_shared;
1575         struct dm_cell_key data_key;
1576         struct dm_bio_prison_cell *data_cell;
1577         struct dm_thin_new_mapping *m;
1578         dm_block_t virt_begin, virt_end, data_begin;
1579
1580         while (begin != end) {
1581                 r = ensure_next_mapping(pool);
1582                 if (r)
1583                         /* we did our best */
1584                         return;
1585
1586                 r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
1587                                               &data_begin, &maybe_shared);
1588                 if (r)
1589                         /*
1590                          * Silently fail, letting any mappings we've
1591                          * created complete.
1592                          */
1593                         break;
1594
1595                 build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
1596                 if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
1597                         /* contention, we'll give up with this range */
1598                         begin = virt_end;
1599                         continue;
1600                 }
1601
1602                 /*
1603                  * IO may still be going to the destination block.  We must
1604                  * quiesce before we can do the removal.
1605                  */
1606                 m = get_next_mapping(pool);
1607                 m->tc = tc;
1608                 m->maybe_shared = maybe_shared;
1609                 m->virt_begin = virt_begin;
1610                 m->virt_end = virt_end;
1611                 m->data_block = data_begin;
1612                 m->cell = data_cell;
1613                 m->bio = bio;
1614
1615                 /*
1616                  * The parent bio must not complete before sub discard bios are
1617                  * chained to it (see end_discard's bio_chain)!
1618                  *
1619                  * This per-mapping bi_remaining increment is paired with
1620                  * the implicit decrement that occurs via bio_endio() in
1621                  * end_discard().
1622                  */
1623                 bio_inc_remaining(bio);
1624                 if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
1625                         pool->process_prepared_discard(m);
1626
1627                 begin = virt_end;
1628         }
1629 }
1630
1631 static void process_discard_cell_passdown(struct thin_c *tc, struct dm_bio_prison_cell *virt_cell)
1632 {
1633         struct bio *bio = virt_cell->holder;
1634         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1635
1636         /*
1637          * The virt_cell will only get freed once the origin bio completes.
1638          * This means it will remain locked while all the individual
1639          * passdown bios are in flight.
1640          */
1641         h->cell = virt_cell;
1642         break_up_discard_bio(tc, virt_cell->key.block_begin, virt_cell->key.block_end, bio);
1643
1644         /*
1645          * We complete the bio now, knowing that the bi_remaining field
1646          * will prevent completion until the sub range discards have
1647          * completed.
1648          */
1649         bio_endio(bio);
1650 }
1651
1652 static void process_discard_bio(struct thin_c *tc, struct bio *bio)
1653 {
1654         dm_block_t begin, end;
1655         struct dm_cell_key virt_key;
1656         struct dm_bio_prison_cell *virt_cell;
1657
1658         get_bio_block_range(tc, bio, &begin, &end);
1659         if (begin == end) {
1660                 /*
1661                  * The discard covers less than a block.
1662                  */
1663                 bio_endio(bio);
1664                 return;
1665         }
1666
1667         build_key(tc->td, VIRTUAL, begin, end, &virt_key);
1668         if (bio_detain(tc->pool, &virt_key, bio, &virt_cell))
1669                 /*
1670                  * Potential starvation issue: We're relying on the
1671                  * fs/application being well behaved, and not trying to
1672                  * send IO to a region at the same time as discarding it.
1673                  * If they do this persistently then it's possible this
1674                  * cell will never be granted.
1675                  */
1676                 return;
1677
1678         tc->pool->process_discard_cell(tc, virt_cell);
1679 }
1680
1681 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1682                           struct dm_cell_key *key,
1683                           struct dm_thin_lookup_result *lookup_result,
1684                           struct dm_bio_prison_cell *cell)
1685 {
1686         int r;
1687         dm_block_t data_block;
1688         struct pool *pool = tc->pool;
1689
1690         r = alloc_data_block(tc, &data_block);
1691         switch (r) {
1692         case 0:
1693                 schedule_internal_copy(tc, block, lookup_result->block,
1694                                        data_block, cell, bio);
1695                 break;
1696
1697         case -ENOSPC:
1698                 retry_bios_on_resume(pool, cell);
1699                 break;
1700
1701         default:
1702                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1703                             __func__, r);
1704                 cell_error(pool, cell);
1705                 break;
1706         }
1707 }
1708
1709 static void __remap_and_issue_shared_cell(void *context,
1710                                           struct dm_bio_prison_cell *cell)
1711 {
1712         struct remap_info *info = context;
1713         struct bio *bio;
1714
1715         while ((bio = bio_list_pop(&cell->bios))) {
1716                 if (bio_data_dir(bio) == WRITE || op_is_flush(bio->bi_opf) ||
1717                     bio_op(bio) == REQ_OP_DISCARD)
1718                         bio_list_add(&info->defer_bios, bio);
1719                 else {
1720                         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));;
1721
1722                         h->shared_read_entry = dm_deferred_entry_inc(info->tc->pool->shared_read_ds);
1723                         inc_all_io_entry(info->tc->pool, bio);
1724                         bio_list_add(&info->issue_bios, bio);
1725                 }
1726         }
1727 }
1728
1729 static void remap_and_issue_shared_cell(struct thin_c *tc,
1730                                         struct dm_bio_prison_cell *cell,
1731                                         dm_block_t block)
1732 {
1733         struct bio *bio;
1734         struct remap_info info;
1735
1736         info.tc = tc;
1737         bio_list_init(&info.defer_bios);
1738         bio_list_init(&info.issue_bios);
1739
1740         cell_visit_release(tc->pool, __remap_and_issue_shared_cell,
1741                            &info, cell);
1742
1743         while ((bio = bio_list_pop(&info.defer_bios)))
1744                 thin_defer_bio(tc, bio);
1745
1746         while ((bio = bio_list_pop(&info.issue_bios)))
1747                 remap_and_issue(tc, bio, block);
1748 }
1749
1750 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1751                                dm_block_t block,
1752                                struct dm_thin_lookup_result *lookup_result,
1753                                struct dm_bio_prison_cell *virt_cell)
1754 {
1755         struct dm_bio_prison_cell *data_cell;
1756         struct pool *pool = tc->pool;
1757         struct dm_cell_key key;
1758
1759         /*
1760          * If cell is already occupied, then sharing is already in the process
1761          * of being broken so we have nothing further to do here.
1762          */
1763         build_data_key(tc->td, lookup_result->block, &key);
1764         if (bio_detain(pool, &key, bio, &data_cell)) {
1765                 cell_defer_no_holder(tc, virt_cell);
1766                 return;
1767         }
1768
1769         if (bio_data_dir(bio) == WRITE && bio->bi_iter.bi_size) {
1770                 break_sharing(tc, bio, block, &key, lookup_result, data_cell);
1771                 cell_defer_no_holder(tc, virt_cell);
1772         } else {
1773                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1774
1775                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1776                 inc_all_io_entry(pool, bio);
1777                 remap_and_issue(tc, bio, lookup_result->block);
1778
1779                 remap_and_issue_shared_cell(tc, data_cell, lookup_result->block);
1780                 remap_and_issue_shared_cell(tc, virt_cell, lookup_result->block);
1781         }
1782 }
1783
1784 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1785                             struct dm_bio_prison_cell *cell)
1786 {
1787         int r;
1788         dm_block_t data_block;
1789         struct pool *pool = tc->pool;
1790
1791         /*
1792          * Remap empty bios (flushes) immediately, without provisioning.
1793          */
1794         if (!bio->bi_iter.bi_size) {
1795                 inc_all_io_entry(pool, bio);
1796                 cell_defer_no_holder(tc, cell);
1797
1798                 remap_and_issue(tc, bio, 0);
1799                 return;
1800         }
1801
1802         /*
1803          * Fill read bios with zeroes and complete them immediately.
1804          */
1805         if (bio_data_dir(bio) == READ) {
1806                 zero_fill_bio(bio);
1807                 cell_defer_no_holder(tc, cell);
1808                 bio_endio(bio);
1809                 return;
1810         }
1811
1812         r = alloc_data_block(tc, &data_block);
1813         switch (r) {
1814         case 0:
1815                 if (tc->origin_dev)
1816                         schedule_external_copy(tc, block, data_block, cell, bio);
1817                 else
1818                         schedule_zero(tc, block, data_block, cell, bio);
1819                 break;
1820
1821         case -ENOSPC:
1822                 retry_bios_on_resume(pool, cell);
1823                 break;
1824
1825         default:
1826                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1827                             __func__, r);
1828                 cell_error(pool, cell);
1829                 break;
1830         }
1831 }
1832
1833 static void process_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1834 {
1835         int r;
1836         struct pool *pool = tc->pool;
1837         struct bio *bio = cell->holder;
1838         dm_block_t block = get_bio_block(tc, bio);
1839         struct dm_thin_lookup_result lookup_result;
1840
1841         if (tc->requeue_mode) {
1842                 cell_requeue(pool, cell);
1843                 return;
1844         }
1845
1846         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1847         switch (r) {
1848         case 0:
1849                 if (lookup_result.shared)
1850                         process_shared_bio(tc, bio, block, &lookup_result, cell);
1851                 else {
1852                         inc_all_io_entry(pool, bio);
1853                         remap_and_issue(tc, bio, lookup_result.block);
1854                         inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1855                 }
1856                 break;
1857
1858         case -ENODATA:
1859                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1860                         inc_all_io_entry(pool, bio);
1861                         cell_defer_no_holder(tc, cell);
1862
1863                         if (bio_end_sector(bio) <= tc->origin_size)
1864                                 remap_to_origin_and_issue(tc, bio);
1865
1866                         else if (bio->bi_iter.bi_sector < tc->origin_size) {
1867                                 zero_fill_bio(bio);
1868                                 bio->bi_iter.bi_size = (tc->origin_size - bio->bi_iter.bi_sector) << SECTOR_SHIFT;
1869                                 remap_to_origin_and_issue(tc, bio);
1870
1871                         } else {
1872                                 zero_fill_bio(bio);
1873                                 bio_endio(bio);
1874                         }
1875                 } else
1876                         provision_block(tc, bio, block, cell);
1877                 break;
1878
1879         default:
1880                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1881                             __func__, r);
1882                 cell_defer_no_holder(tc, cell);
1883                 bio_io_error(bio);
1884                 break;
1885         }
1886 }
1887
1888 static void process_bio(struct thin_c *tc, struct bio *bio)
1889 {
1890         struct pool *pool = tc->pool;
1891         dm_block_t block = get_bio_block(tc, bio);
1892         struct dm_bio_prison_cell *cell;
1893         struct dm_cell_key key;
1894
1895         /*
1896          * If cell is already occupied, then the block is already
1897          * being provisioned so we have nothing further to do here.
1898          */
1899         build_virtual_key(tc->td, block, &key);
1900         if (bio_detain(pool, &key, bio, &cell))
1901                 return;
1902
1903         process_cell(tc, cell);
1904 }
1905
1906 static void __process_bio_read_only(struct thin_c *tc, struct bio *bio,
1907                                     struct dm_bio_prison_cell *cell)
1908 {
1909         int r;
1910         int rw = bio_data_dir(bio);
1911         dm_block_t block = get_bio_block(tc, bio);
1912         struct dm_thin_lookup_result lookup_result;
1913
1914         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1915         switch (r) {
1916         case 0:
1917                 if (lookup_result.shared && (rw == WRITE) && bio->bi_iter.bi_size) {
1918                         handle_unserviceable_bio(tc->pool, bio);
1919                         if (cell)
1920                                 cell_defer_no_holder(tc, cell);
1921                 } else {
1922                         inc_all_io_entry(tc->pool, bio);
1923                         remap_and_issue(tc, bio, lookup_result.block);
1924                         if (cell)
1925                                 inc_remap_and_issue_cell(tc, cell, lookup_result.block);
1926                 }
1927                 break;
1928
1929         case -ENODATA:
1930                 if (cell)
1931                         cell_defer_no_holder(tc, cell);
1932                 if (rw != READ) {
1933                         handle_unserviceable_bio(tc->pool, bio);
1934                         break;
1935                 }
1936
1937                 if (tc->origin_dev) {
1938                         inc_all_io_entry(tc->pool, bio);
1939                         remap_to_origin_and_issue(tc, bio);
1940                         break;
1941                 }
1942
1943                 zero_fill_bio(bio);
1944                 bio_endio(bio);
1945                 break;
1946
1947         default:
1948                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1949                             __func__, r);
1950                 if (cell)
1951                         cell_defer_no_holder(tc, cell);
1952                 bio_io_error(bio);
1953                 break;
1954         }
1955 }
1956
1957 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1958 {
1959         __process_bio_read_only(tc, bio, NULL);
1960 }
1961
1962 static void process_cell_read_only(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1963 {
1964         __process_bio_read_only(tc, cell->holder, cell);
1965 }
1966
1967 static void process_bio_success(struct thin_c *tc, struct bio *bio)
1968 {
1969         bio_endio(bio);
1970 }
1971
1972 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1973 {
1974         bio_io_error(bio);
1975 }
1976
1977 static void process_cell_success(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1978 {
1979         cell_success(tc->pool, cell);
1980 }
1981
1982 static void process_cell_fail(struct thin_c *tc, struct dm_bio_prison_cell *cell)
1983 {
1984         cell_error(tc->pool, cell);
1985 }
1986
1987 /*
1988  * FIXME: should we also commit due to size of transaction, measured in
1989  * metadata blocks?
1990  */
1991 static int need_commit_due_to_time(struct pool *pool)
1992 {
1993         return !time_in_range(jiffies, pool->last_commit_jiffies,
1994                               pool->last_commit_jiffies + COMMIT_PERIOD);
1995 }
1996
1997 #define thin_pbd(node) rb_entry((node), struct dm_thin_endio_hook, rb_node)
1998 #define thin_bio(pbd) dm_bio_from_per_bio_data((pbd), sizeof(struct dm_thin_endio_hook))
1999
2000 static void __thin_bio_rb_add(struct thin_c *tc, struct bio *bio)
2001 {
2002         struct rb_node **rbp, *parent;
2003         struct dm_thin_endio_hook *pbd;
2004         sector_t bi_sector = bio->bi_iter.bi_sector;
2005
2006         rbp = &tc->sort_bio_list.rb_node;
2007         parent = NULL;
2008         while (*rbp) {
2009                 parent = *rbp;
2010                 pbd = thin_pbd(parent);
2011
2012                 if (bi_sector < thin_bio(pbd)->bi_iter.bi_sector)
2013                         rbp = &(*rbp)->rb_left;
2014                 else
2015                         rbp = &(*rbp)->rb_right;
2016         }
2017
2018         pbd = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2019         rb_link_node(&pbd->rb_node, parent, rbp);
2020         rb_insert_color(&pbd->rb_node, &tc->sort_bio_list);
2021 }
2022
2023 static void __extract_sorted_bios(struct thin_c *tc)
2024 {
2025         struct rb_node *node;
2026         struct dm_thin_endio_hook *pbd;
2027         struct bio *bio;
2028
2029         for (node = rb_first(&tc->sort_bio_list); node; node = rb_next(node)) {
2030                 pbd = thin_pbd(node);
2031                 bio = thin_bio(pbd);
2032
2033                 bio_list_add(&tc->deferred_bio_list, bio);
2034                 rb_erase(&pbd->rb_node, &tc->sort_bio_list);
2035         }
2036
2037         WARN_ON(!RB_EMPTY_ROOT(&tc->sort_bio_list));
2038 }
2039
2040 static void __sort_thin_deferred_bios(struct thin_c *tc)
2041 {
2042         struct bio *bio;
2043         struct bio_list bios;
2044
2045         bio_list_init(&bios);
2046         bio_list_merge(&bios, &tc->deferred_bio_list);
2047         bio_list_init(&tc->deferred_bio_list);
2048
2049         /* Sort deferred_bio_list using rb-tree */
2050         while ((bio = bio_list_pop(&bios)))
2051                 __thin_bio_rb_add(tc, bio);
2052
2053         /*
2054          * Transfer the sorted bios in sort_bio_list back to
2055          * deferred_bio_list to allow lockless submission of
2056          * all bios.
2057          */
2058         __extract_sorted_bios(tc);
2059 }
2060
2061 static void process_thin_deferred_bios(struct thin_c *tc)
2062 {
2063         struct pool *pool = tc->pool;
2064         unsigned long flags;
2065         struct bio *bio;
2066         struct bio_list bios;
2067         struct blk_plug plug;
2068         unsigned count = 0;
2069
2070         if (tc->requeue_mode) {
2071                 error_thin_bio_list(tc, &tc->deferred_bio_list,
2072                                 BLK_STS_DM_REQUEUE);
2073                 return;
2074         }
2075
2076         bio_list_init(&bios);
2077
2078         spin_lock_irqsave(&tc->lock, flags);
2079
2080         if (bio_list_empty(&tc->deferred_bio_list)) {
2081                 spin_unlock_irqrestore(&tc->lock, flags);
2082                 return;
2083         }
2084
2085         __sort_thin_deferred_bios(tc);
2086
2087         bio_list_merge(&bios, &tc->deferred_bio_list);
2088         bio_list_init(&tc->deferred_bio_list);
2089
2090         spin_unlock_irqrestore(&tc->lock, flags);
2091
2092         blk_start_plug(&plug);
2093         while ((bio = bio_list_pop(&bios))) {
2094                 /*
2095                  * If we've got no free new_mapping structs, and processing
2096                  * this bio might require one, we pause until there are some
2097                  * prepared mappings to process.
2098                  */
2099                 if (ensure_next_mapping(pool)) {
2100                         spin_lock_irqsave(&tc->lock, flags);
2101                         bio_list_add(&tc->deferred_bio_list, bio);
2102                         bio_list_merge(&tc->deferred_bio_list, &bios);
2103                         spin_unlock_irqrestore(&tc->lock, flags);
2104                         break;
2105                 }
2106
2107                 if (bio_op(bio) == REQ_OP_DISCARD)
2108                         pool->process_discard(tc, bio);
2109                 else
2110                         pool->process_bio(tc, bio);
2111
2112                 if ((count++ & 127) == 0) {
2113                         throttle_work_update(&pool->throttle);
2114                         dm_pool_issue_prefetches(pool->pmd);
2115                 }
2116         }
2117         blk_finish_plug(&plug);
2118 }
2119
2120 static int cmp_cells(const void *lhs, const void *rhs)
2121 {
2122         struct dm_bio_prison_cell *lhs_cell = *((struct dm_bio_prison_cell **) lhs);
2123         struct dm_bio_prison_cell *rhs_cell = *((struct dm_bio_prison_cell **) rhs);
2124
2125         BUG_ON(!lhs_cell->holder);
2126         BUG_ON(!rhs_cell->holder);
2127
2128         if (lhs_cell->holder->bi_iter.bi_sector < rhs_cell->holder->bi_iter.bi_sector)
2129                 return -1;
2130
2131         if (lhs_cell->holder->bi_iter.bi_sector > rhs_cell->holder->bi_iter.bi_sector)
2132                 return 1;
2133
2134         return 0;
2135 }
2136
2137 static unsigned sort_cells(struct pool *pool, struct list_head *cells)
2138 {
2139         unsigned count = 0;
2140         struct dm_bio_prison_cell *cell, *tmp;
2141
2142         list_for_each_entry_safe(cell, tmp, cells, user_list) {
2143                 if (count >= CELL_SORT_ARRAY_SIZE)
2144                         break;
2145
2146                 pool->cell_sort_array[count++] = cell;
2147                 list_del(&cell->user_list);
2148         }
2149
2150         sort(pool->cell_sort_array, count, sizeof(cell), cmp_cells, NULL);
2151
2152         return count;
2153 }
2154
2155 static void process_thin_deferred_cells(struct thin_c *tc)
2156 {
2157         struct pool *pool = tc->pool;
2158         unsigned long flags;
2159         struct list_head cells;
2160         struct dm_bio_prison_cell *cell;
2161         unsigned i, j, count;
2162
2163         INIT_LIST_HEAD(&cells);
2164
2165         spin_lock_irqsave(&tc->lock, flags);
2166         list_splice_init(&tc->deferred_cells, &cells);
2167         spin_unlock_irqrestore(&tc->lock, flags);
2168
2169         if (list_empty(&cells))
2170                 return;
2171
2172         do {
2173                 count = sort_cells(tc->pool, &cells);
2174
2175                 for (i = 0; i < count; i++) {
2176                         cell = pool->cell_sort_array[i];
2177                         BUG_ON(!cell->holder);
2178
2179                         /*
2180                          * If we've got no free new_mapping structs, and processing
2181                          * this bio might require one, we pause until there are some
2182                          * prepared mappings to process.
2183                          */
2184                         if (ensure_next_mapping(pool)) {
2185                                 for (j = i; j < count; j++)
2186                                         list_add(&pool->cell_sort_array[j]->user_list, &cells);
2187
2188                                 spin_lock_irqsave(&tc->lock, flags);
2189                                 list_splice(&cells, &tc->deferred_cells);
2190                                 spin_unlock_irqrestore(&tc->lock, flags);
2191                                 return;
2192                         }
2193
2194                         if (bio_op(cell->holder) == REQ_OP_DISCARD)
2195                                 pool->process_discard_cell(tc, cell);
2196                         else
2197                                 pool->process_cell(tc, cell);
2198                 }
2199         } while (!list_empty(&cells));
2200 }
2201
2202 static void thin_get(struct thin_c *tc);
2203 static void thin_put(struct thin_c *tc);
2204
2205 /*
2206  * We can't hold rcu_read_lock() around code that can block.  So we
2207  * find a thin with the rcu lock held; bump a refcount; then drop
2208  * the lock.
2209  */
2210 static struct thin_c *get_first_thin(struct pool *pool)
2211 {
2212         struct thin_c *tc = NULL;
2213
2214         rcu_read_lock();
2215         if (!list_empty(&pool->active_thins)) {
2216                 tc = list_entry_rcu(pool->active_thins.next, struct thin_c, list);
2217                 thin_get(tc);
2218         }
2219         rcu_read_unlock();
2220
2221         return tc;
2222 }
2223
2224 static struct thin_c *get_next_thin(struct pool *pool, struct thin_c *tc)
2225 {
2226         struct thin_c *old_tc = tc;
2227
2228         rcu_read_lock();
2229         list_for_each_entry_continue_rcu(tc, &pool->active_thins, list) {
2230                 thin_get(tc);
2231                 thin_put(old_tc);
2232                 rcu_read_unlock();
2233                 return tc;
2234         }
2235         thin_put(old_tc);
2236         rcu_read_unlock();
2237
2238         return NULL;
2239 }
2240
2241 static void process_deferred_bios(struct pool *pool)
2242 {
2243         unsigned long flags;
2244         struct bio *bio;
2245         struct bio_list bios;
2246         struct thin_c *tc;
2247
2248         tc = get_first_thin(pool);
2249         while (tc) {
2250                 process_thin_deferred_cells(tc);
2251                 process_thin_deferred_bios(tc);
2252                 tc = get_next_thin(pool, tc);
2253         }
2254
2255         /*
2256          * If there are any deferred flush bios, we must commit
2257          * the metadata before issuing them.
2258          */
2259         bio_list_init(&bios);
2260         spin_lock_irqsave(&pool->lock, flags);
2261         bio_list_merge(&bios, &pool->deferred_flush_bios);
2262         bio_list_init(&pool->deferred_flush_bios);
2263         spin_unlock_irqrestore(&pool->lock, flags);
2264
2265         if (bio_list_empty(&bios) &&
2266             !(dm_pool_changed_this_transaction(pool->pmd) && need_commit_due_to_time(pool)))
2267                 return;
2268
2269         if (commit(pool)) {
2270                 while ((bio = bio_list_pop(&bios)))
2271                         bio_io_error(bio);
2272                 return;
2273         }
2274         pool->last_commit_jiffies = jiffies;
2275
2276         while ((bio = bio_list_pop(&bios)))
2277                 generic_make_request(bio);
2278 }
2279
2280 static void do_worker(struct work_struct *ws)
2281 {
2282         struct pool *pool = container_of(ws, struct pool, worker);
2283
2284         throttle_work_start(&pool->throttle);
2285         dm_pool_issue_prefetches(pool->pmd);
2286         throttle_work_update(&pool->throttle);
2287         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
2288         throttle_work_update(&pool->throttle);
2289         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
2290         throttle_work_update(&pool->throttle);
2291         process_prepared(pool, &pool->prepared_discards_pt2, &pool->process_prepared_discard_pt2);
2292         throttle_work_update(&pool->throttle);
2293         process_deferred_bios(pool);
2294         throttle_work_complete(&pool->throttle);
2295 }
2296
2297 /*
2298  * We want to commit periodically so that not too much
2299  * unwritten data builds up.
2300  */
2301 static void do_waker(struct work_struct *ws)
2302 {
2303         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
2304         wake_worker(pool);
2305         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
2306 }
2307
2308 static void notify_of_pool_mode_change_to_oods(struct pool *pool);
2309
2310 /*
2311  * We're holding onto IO to allow userland time to react.  After the
2312  * timeout either the pool will have been resized (and thus back in
2313  * PM_WRITE mode), or we degrade to PM_OUT_OF_DATA_SPACE w/ error_if_no_space.
2314  */
2315 static void do_no_space_timeout(struct work_struct *ws)
2316 {
2317         struct pool *pool = container_of(to_delayed_work(ws), struct pool,
2318                                          no_space_timeout);
2319
2320         if (get_pool_mode(pool) == PM_OUT_OF_DATA_SPACE && !pool->pf.error_if_no_space) {
2321                 pool->pf.error_if_no_space = true;
2322                 notify_of_pool_mode_change_to_oods(pool);
2323                 error_retry_list_with_code(pool, BLK_STS_NOSPC);
2324         }
2325 }
2326
2327 /*----------------------------------------------------------------*/
2328
2329 struct pool_work {
2330         struct work_struct worker;
2331         struct completion complete;
2332 };
2333
2334 static struct pool_work *to_pool_work(struct work_struct *ws)
2335 {
2336         return container_of(ws, struct pool_work, worker);
2337 }
2338
2339 static void pool_work_complete(struct pool_work *pw)
2340 {
2341         complete(&pw->complete);
2342 }
2343
2344 static void pool_work_wait(struct pool_work *pw, struct pool *pool,
2345                            void (*fn)(struct work_struct *))
2346 {
2347         INIT_WORK_ONSTACK(&pw->worker, fn);
2348         init_completion(&pw->complete);
2349         queue_work(pool->wq, &pw->worker);
2350         wait_for_completion(&pw->complete);
2351 }
2352
2353 /*----------------------------------------------------------------*/
2354
2355 struct noflush_work {
2356         struct pool_work pw;
2357         struct thin_c *tc;
2358 };
2359
2360 static struct noflush_work *to_noflush(struct work_struct *ws)
2361 {
2362         return container_of(to_pool_work(ws), struct noflush_work, pw);
2363 }
2364
2365 static void do_noflush_start(struct work_struct *ws)
2366 {
2367         struct noflush_work *w = to_noflush(ws);
2368         w->tc->requeue_mode = true;
2369         requeue_io(w->tc);
2370         pool_work_complete(&w->pw);
2371 }
2372
2373 static void do_noflush_stop(struct work_struct *ws)
2374 {
2375         struct noflush_work *w = to_noflush(ws);
2376         w->tc->requeue_mode = false;
2377         pool_work_complete(&w->pw);
2378 }
2379
2380 static void noflush_work(struct thin_c *tc, void (*fn)(struct work_struct *))
2381 {
2382         struct noflush_work w;
2383
2384         w.tc = tc;
2385         pool_work_wait(&w.pw, tc->pool, fn);
2386 }
2387
2388 /*----------------------------------------------------------------*/
2389
2390 static enum pool_mode get_pool_mode(struct pool *pool)
2391 {
2392         return pool->pf.mode;
2393 }
2394
2395 static void notify_of_pool_mode_change(struct pool *pool, const char *new_mode)
2396 {
2397         dm_table_event(pool->ti->table);
2398         DMINFO("%s: switching pool to %s mode",
2399                dm_device_name(pool->pool_md), new_mode);
2400 }
2401
2402 static void notify_of_pool_mode_change_to_oods(struct pool *pool)
2403 {
2404         if (!pool->pf.error_if_no_space)
2405                 notify_of_pool_mode_change(pool, "out-of-data-space (queue IO)");
2406         else
2407                 notify_of_pool_mode_change(pool, "out-of-data-space (error IO)");
2408 }
2409
2410 static bool passdown_enabled(struct pool_c *pt)
2411 {
2412         return pt->adjusted_pf.discard_passdown;
2413 }
2414
2415 static void set_discard_callbacks(struct pool *pool)
2416 {
2417         struct pool_c *pt = pool->ti->private;
2418
2419         if (passdown_enabled(pt)) {
2420                 pool->process_discard_cell = process_discard_cell_passdown;
2421                 pool->process_prepared_discard = process_prepared_discard_passdown_pt1;
2422                 pool->process_prepared_discard_pt2 = process_prepared_discard_passdown_pt2;
2423         } else {
2424                 pool->process_discard_cell = process_discard_cell_no_passdown;
2425                 pool->process_prepared_discard = process_prepared_discard_no_passdown;
2426         }
2427 }
2428
2429 static void set_pool_mode(struct pool *pool, enum pool_mode new_mode)
2430 {
2431         struct pool_c *pt = pool->ti->private;
2432         bool needs_check = dm_pool_metadata_needs_check(pool->pmd);
2433         enum pool_mode old_mode = get_pool_mode(pool);
2434         unsigned long no_space_timeout = ACCESS_ONCE(no_space_timeout_secs) * HZ;
2435
2436         /*
2437          * Never allow the pool to transition to PM_WRITE mode if user
2438          * intervention is required to verify metadata and data consistency.
2439          */
2440         if (new_mode == PM_WRITE && needs_check) {
2441                 DMERR("%s: unable to switch pool to write mode until repaired.",
2442                       dm_device_name(pool->pool_md));
2443                 if (old_mode != new_mode)
2444                         new_mode = old_mode;
2445                 else
2446                         new_mode = PM_READ_ONLY;
2447         }
2448         /*
2449          * If we were in PM_FAIL mode, rollback of metadata failed.  We're
2450          * not going to recover without a thin_repair.  So we never let the
2451          * pool move out of the old mode.
2452          */
2453         if (old_mode == PM_FAIL)
2454                 new_mode = old_mode;
2455
2456         switch (new_mode) {
2457         case PM_FAIL:
2458                 if (old_mode != new_mode)
2459                         notify_of_pool_mode_change(pool, "failure");
2460                 dm_pool_metadata_read_only(pool->pmd);
2461                 pool->process_bio = process_bio_fail;
2462                 pool->process_discard = process_bio_fail;
2463                 pool->process_cell = process_cell_fail;
2464                 pool->process_discard_cell = process_cell_fail;
2465                 pool->process_prepared_mapping = process_prepared_mapping_fail;
2466                 pool->process_prepared_discard = process_prepared_discard_fail;
2467
2468                 error_retry_list(pool);
2469                 break;
2470
2471         case PM_READ_ONLY:
2472                 if (old_mode != new_mode)
2473                         notify_of_pool_mode_change(pool, "read-only");
2474                 dm_pool_metadata_read_only(pool->pmd);
2475                 pool->process_bio = process_bio_read_only;
2476                 pool->process_discard = process_bio_success;
2477                 pool->process_cell = process_cell_read_only;
2478                 pool->process_discard_cell = process_cell_success;
2479                 pool->process_prepared_mapping = process_prepared_mapping_fail;
2480                 pool->process_prepared_discard = process_prepared_discard_success;
2481
2482                 error_retry_list(pool);
2483                 break;
2484
2485         case PM_OUT_OF_DATA_SPACE:
2486                 /*
2487                  * Ideally we'd never hit this state; the low water mark
2488                  * would trigger userland to extend the pool before we
2489                  * completely run out of data space.  However, many small
2490                  * IOs to unprovisioned space can consume data space at an
2491                  * alarming rate.  Adjust your low water mark if you're
2492                  * frequently seeing this mode.
2493                  */
2494                 if (old_mode != new_mode)
2495                         notify_of_pool_mode_change_to_oods(pool);
2496                 pool->out_of_data_space = true;
2497                 pool->process_bio = process_bio_read_only;
2498                 pool->process_discard = process_discard_bio;
2499                 pool->process_cell = process_cell_read_only;
2500                 pool->process_prepared_mapping = process_prepared_mapping;
2501                 set_discard_callbacks(pool);
2502
2503                 if (!pool->pf.error_if_no_space && no_space_timeout)
2504                         queue_delayed_work(pool->wq, &pool->no_space_timeout, no_space_timeout);
2505                 break;
2506
2507         case PM_WRITE:
2508                 if (old_mode != new_mode)
2509                         notify_of_pool_mode_change(pool, "write");
2510                 pool->out_of_data_space = false;
2511                 pool->pf.error_if_no_space = pt->requested_pf.error_if_no_space;
2512                 dm_pool_metadata_read_write(pool->pmd);
2513                 pool->process_bio = process_bio;
2514                 pool->process_discard = process_discard_bio;
2515                 pool->process_cell = process_cell;
2516                 pool->process_prepared_mapping = process_prepared_mapping;
2517                 set_discard_callbacks(pool);
2518                 break;
2519         }
2520
2521         pool->pf.mode = new_mode;
2522         /*
2523          * The pool mode may have changed, sync it so bind_control_target()
2524          * doesn't cause an unexpected mode transition on resume.
2525          */
2526         pt->adjusted_pf.mode = new_mode;
2527 }
2528
2529 static void abort_transaction(struct pool *pool)
2530 {
2531         const char *dev_name = dm_device_name(pool->pool_md);
2532
2533         DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
2534         if (dm_pool_abort_metadata(pool->pmd)) {
2535                 DMERR("%s: failed to abort metadata transaction", dev_name);
2536                 set_pool_mode(pool, PM_FAIL);
2537         }
2538
2539         if (dm_pool_metadata_set_needs_check(pool->pmd)) {
2540                 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
2541                 set_pool_mode(pool, PM_FAIL);
2542         }
2543 }
2544
2545 static void metadata_operation_failed(struct pool *pool, const char *op, int r)
2546 {
2547         DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
2548                     dm_device_name(pool->pool_md), op, r);
2549
2550         abort_transaction(pool);
2551         set_pool_mode(pool, PM_READ_ONLY);
2552 }
2553
2554 /*----------------------------------------------------------------*/
2555
2556 /*
2557  * Mapping functions.
2558  */
2559
2560 /*
2561  * Called only while mapping a thin bio to hand it over to the workqueue.
2562  */
2563 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
2564 {
2565         unsigned long flags;
2566         struct pool *pool = tc->pool;
2567
2568         spin_lock_irqsave(&tc->lock, flags);
2569         bio_list_add(&tc->deferred_bio_list, bio);
2570         spin_unlock_irqrestore(&tc->lock, flags);
2571
2572         wake_worker(pool);
2573 }
2574
2575 static void thin_defer_bio_with_throttle(struct thin_c *tc, struct bio *bio)
2576 {
2577         struct pool *pool = tc->pool;
2578
2579         throttle_lock(&pool->throttle);
2580         thin_defer_bio(tc, bio);
2581         throttle_unlock(&pool->throttle);
2582 }
2583
2584 static void thin_defer_cell(struct thin_c *tc, struct dm_bio_prison_cell *cell)
2585 {
2586         unsigned long flags;
2587         struct pool *pool = tc->pool;
2588
2589         throttle_lock(&pool->throttle);
2590         spin_lock_irqsave(&tc->lock, flags);
2591         list_add_tail(&cell->user_list, &tc->deferred_cells);
2592         spin_unlock_irqrestore(&tc->lock, flags);
2593         throttle_unlock(&pool->throttle);
2594
2595         wake_worker(pool);
2596 }
2597
2598 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
2599 {
2600         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2601
2602         h->tc = tc;
2603         h->shared_read_entry = NULL;
2604         h->all_io_entry = NULL;
2605         h->overwrite_mapping = NULL;
2606         h->cell = NULL;
2607 }
2608
2609 /*
2610  * Non-blocking function called from the thin target's map function.
2611  */
2612 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
2613 {
2614         int r;
2615         struct thin_c *tc = ti->private;
2616         dm_block_t block = get_bio_block(tc, bio);
2617         struct dm_thin_device *td = tc->td;
2618         struct dm_thin_lookup_result result;
2619         struct dm_bio_prison_cell *virt_cell, *data_cell;
2620         struct dm_cell_key key;
2621
2622         thin_hook_bio(tc, bio);
2623
2624         if (tc->requeue_mode) {
2625                 bio->bi_status = BLK_STS_DM_REQUEUE;
2626                 bio_endio(bio);
2627                 return DM_MAPIO_SUBMITTED;
2628         }
2629
2630         if (get_pool_mode(tc->pool) == PM_FAIL) {
2631                 bio_io_error(bio);
2632                 return DM_MAPIO_SUBMITTED;
2633         }
2634
2635         if (op_is_flush(bio->bi_opf) || bio_op(bio) == REQ_OP_DISCARD) {
2636                 thin_defer_bio_with_throttle(tc, bio);
2637                 return DM_MAPIO_SUBMITTED;
2638         }
2639
2640         /*
2641          * We must hold the virtual cell before doing the lookup, otherwise
2642          * there's a race with discard.
2643          */
2644         build_virtual_key(tc->td, block, &key);
2645         if (bio_detain(tc->pool, &key, bio, &virt_cell))
2646                 return DM_MAPIO_SUBMITTED;
2647
2648         r = dm_thin_find_block(td, block, 0, &result);
2649
2650         /*
2651          * Note that we defer readahead too.
2652          */
2653         switch (r) {
2654         case 0:
2655                 if (unlikely(result.shared)) {
2656                         /*
2657                          * We have a race condition here between the
2658                          * result.shared value returned by the lookup and
2659                          * snapshot creation, which may cause new
2660                          * sharing.
2661                          *
2662                          * To avoid this always quiesce the origin before
2663                          * taking the snap.  You want to do this anyway to
2664                          * ensure a consistent application view
2665                          * (i.e. lockfs).
2666                          *
2667                          * More distant ancestors are irrelevant. The
2668                          * shared flag will be set in their case.
2669                          */
2670                         thin_defer_cell(tc, virt_cell);
2671                         return DM_MAPIO_SUBMITTED;
2672                 }
2673
2674                 build_data_key(tc->td, result.block, &key);
2675                 if (bio_detain(tc->pool, &key, bio, &data_cell)) {
2676                         cell_defer_no_holder(tc, virt_cell);
2677                         return DM_MAPIO_SUBMITTED;
2678                 }
2679
2680                 inc_all_io_entry(tc->pool, bio);
2681                 cell_defer_no_holder(tc, data_cell);
2682                 cell_defer_no_holder(tc, virt_cell);
2683
2684                 remap(tc, bio, result.block);
2685                 return DM_MAPIO_REMAPPED;
2686
2687         case -ENODATA:
2688         case -EWOULDBLOCK:
2689                 thin_defer_cell(tc, virt_cell);
2690                 return DM_MAPIO_SUBMITTED;
2691
2692         default:
2693                 /*
2694                  * Must always call bio_io_error on failure.
2695                  * dm_thin_find_block can fail with -EINVAL if the
2696                  * pool is switched to fail-io mode.
2697                  */
2698                 bio_io_error(bio);
2699                 cell_defer_no_holder(tc, virt_cell);
2700                 return DM_MAPIO_SUBMITTED;
2701         }
2702 }
2703
2704 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2705 {
2706         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
2707         struct request_queue *q;
2708
2709         if (get_pool_mode(pt->pool) == PM_OUT_OF_DATA_SPACE)
2710                 return 1;
2711
2712         q = bdev_get_queue(pt->data_dev->bdev);
2713         return bdi_congested(q->backing_dev_info, bdi_bits);
2714 }
2715
2716 static void requeue_bios(struct pool *pool)
2717 {
2718         unsigned long flags;
2719         struct thin_c *tc;
2720
2721         rcu_read_lock();
2722         list_for_each_entry_rcu(tc, &pool->active_thins, list) {
2723                 spin_lock_irqsave(&tc->lock, flags);
2724                 bio_list_merge(&tc->deferred_bio_list, &tc->retry_on_resume_list);
2725                 bio_list_init(&tc->retry_on_resume_list);
2726                 spin_unlock_irqrestore(&tc->lock, flags);
2727         }
2728         rcu_read_unlock();
2729 }
2730
2731 /*----------------------------------------------------------------
2732  * Binding of control targets to a pool object
2733  *--------------------------------------------------------------*/
2734 static bool data_dev_supports_discard(struct pool_c *pt)
2735 {
2736         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2737
2738         return q && blk_queue_discard(q);
2739 }
2740
2741 static bool is_factor(sector_t block_size, uint32_t n)
2742 {
2743         return !sector_div(block_size, n);
2744 }
2745
2746 /*
2747  * If discard_passdown was enabled verify that the data device
2748  * supports discards.  Disable discard_passdown if not.
2749  */
2750 static void disable_passdown_if_not_supported(struct pool_c *pt)
2751 {
2752         struct pool *pool = pt->pool;
2753         struct block_device *data_bdev = pt->data_dev->bdev;
2754         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
2755         const char *reason = NULL;
2756         char buf[BDEVNAME_SIZE];
2757
2758         if (!pt->adjusted_pf.discard_passdown)
2759                 return;
2760
2761         if (!data_dev_supports_discard(pt))
2762                 reason = "discard unsupported";
2763
2764         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
2765                 reason = "max discard sectors smaller than a block";
2766
2767         if (reason) {
2768                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
2769                 pt->adjusted_pf.discard_passdown = false;
2770         }
2771 }
2772
2773 static int bind_control_target(struct pool *pool, struct dm_target *ti)
2774 {
2775         struct pool_c *pt = ti->private;
2776
2777         /*
2778          * We want to make sure that a pool in PM_FAIL mode is never upgraded.
2779          */
2780         enum pool_mode old_mode = get_pool_mode(pool);
2781         enum pool_mode new_mode = pt->adjusted_pf.mode;
2782
2783         /*
2784          * Don't change the pool's mode until set_pool_mode() below.
2785          * Otherwise the pool's process_* function pointers may
2786          * not match the desired pool mode.
2787          */
2788         pt->adjusted_pf.mode = old_mode;
2789
2790         pool->ti = ti;
2791         pool->pf = pt->adjusted_pf;
2792         pool->low_water_blocks = pt->low_water_blocks;
2793
2794         set_pool_mode(pool, new_mode);
2795
2796         return 0;
2797 }
2798
2799 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
2800 {
2801         if (pool->ti == ti)
2802                 pool->ti = NULL;
2803 }
2804
2805 /*----------------------------------------------------------------
2806  * Pool creation
2807  *--------------------------------------------------------------*/
2808 /* Initialize pool features. */
2809 static void pool_features_init(struct pool_features *pf)
2810 {
2811         pf->mode = PM_WRITE;
2812         pf->zero_new_blocks = true;
2813         pf->discard_enabled = true;
2814         pf->discard_passdown = true;
2815         pf->error_if_no_space = false;
2816 }
2817
2818 static void __pool_destroy(struct pool *pool)
2819 {
2820         __pool_table_remove(pool);
2821
2822         vfree(pool->cell_sort_array);
2823         if (dm_pool_metadata_close(pool->pmd) < 0)
2824                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2825
2826         dm_bio_prison_destroy(pool->prison);
2827         dm_kcopyd_client_destroy(pool->copier);
2828
2829         if (pool->wq)
2830                 destroy_workqueue(pool->wq);
2831
2832         if (pool->next_mapping)
2833                 mempool_free(pool->next_mapping, pool->mapping_pool);
2834         mempool_destroy(pool->mapping_pool);
2835         dm_deferred_set_destroy(pool->shared_read_ds);
2836         dm_deferred_set_destroy(pool->all_io_ds);
2837         kfree(pool);
2838 }
2839
2840 static struct kmem_cache *_new_mapping_cache;
2841
2842 static struct pool *pool_create(struct mapped_device *pool_md,
2843                                 struct block_device *metadata_dev,
2844                                 unsigned long block_size,
2845                                 int read_only, char **error)
2846 {
2847         int r;
2848         void *err_p;
2849         struct pool *pool;
2850         struct dm_pool_metadata *pmd;
2851         bool format_device = read_only ? false : true;
2852
2853         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
2854         if (IS_ERR(pmd)) {
2855                 *error = "Error creating metadata object";
2856                 return (struct pool *)pmd;
2857         }
2858
2859         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
2860         if (!pool) {
2861                 *error = "Error allocating memory for pool";
2862                 err_p = ERR_PTR(-ENOMEM);
2863                 goto bad_pool;
2864         }
2865
2866         pool->pmd = pmd;
2867         pool->sectors_per_block = block_size;
2868         if (block_size & (block_size - 1))
2869                 pool->sectors_per_block_shift = -1;
2870         else
2871                 pool->sectors_per_block_shift = __ffs(block_size);
2872         pool->low_water_blocks = 0;
2873         pool_features_init(&pool->pf);
2874         pool->prison = dm_bio_prison_create();
2875         if (!pool->prison) {
2876                 *error = "Error creating pool's bio prison";
2877                 err_p = ERR_PTR(-ENOMEM);
2878                 goto bad_prison;
2879         }
2880
2881         pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2882         if (IS_ERR(pool->copier)) {
2883                 r = PTR_ERR(pool->copier);
2884                 *error = "Error creating pool's kcopyd client";
2885                 err_p = ERR_PTR(r);
2886                 goto bad_kcopyd_client;
2887         }
2888
2889         /*
2890          * Create singlethreaded workqueue that will service all devices
2891          * that use this metadata.
2892          */
2893         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2894         if (!pool->wq) {
2895                 *error = "Error creating pool's workqueue";
2896                 err_p = ERR_PTR(-ENOMEM);
2897                 goto bad_wq;
2898         }
2899
2900         throttle_init(&pool->throttle);
2901         INIT_WORK(&pool->worker, do_worker);
2902         INIT_DELAYED_WORK(&pool->waker, do_waker);
2903         INIT_DELAYED_WORK(&pool->no_space_timeout, do_no_space_timeout);
2904         spin_lock_init(&pool->lock);
2905         bio_list_init(&pool->deferred_flush_bios);
2906         INIT_LIST_HEAD(&pool->prepared_mappings);
2907         INIT_LIST_HEAD(&pool->prepared_discards);
2908         INIT_LIST_HEAD(&pool->prepared_discards_pt2);
2909         INIT_LIST_HEAD(&pool->active_thins);
2910         pool->low_water_triggered = false;
2911         pool->suspended = true;
2912         pool->out_of_data_space = false;
2913
2914         pool->shared_read_ds = dm_deferred_set_create();
2915         if (!pool->shared_read_ds) {
2916                 *error = "Error creating pool's shared read deferred set";
2917                 err_p = ERR_PTR(-ENOMEM);
2918                 goto bad_shared_read_ds;
2919         }
2920
2921         pool->all_io_ds = dm_deferred_set_create();
2922         if (!pool->all_io_ds) {
2923                 *error = "Error creating pool's all io deferred set";
2924                 err_p = ERR_PTR(-ENOMEM);
2925                 goto bad_all_io_ds;
2926         }
2927
2928         pool->next_mapping = NULL;
2929         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
2930                                                       _new_mapping_cache);
2931         if (!pool->mapping_pool) {
2932                 *error = "Error creating pool's mapping mempool";
2933                 err_p = ERR_PTR(-ENOMEM);
2934                 goto bad_mapping_pool;
2935         }
2936
2937         pool->cell_sort_array = vmalloc(sizeof(*pool->cell_sort_array) * CELL_SORT_ARRAY_SIZE);
2938         if (!pool->cell_sort_array) {
2939                 *error = "Error allocating cell sort array";
2940                 err_p = ERR_PTR(-ENOMEM);
2941                 goto bad_sort_array;
2942         }
2943
2944         pool->ref_count = 1;
2945         pool->last_commit_jiffies = jiffies;
2946         pool->pool_md = pool_md;
2947         pool->md_dev = metadata_dev;
2948         __pool_table_insert(pool);
2949
2950         return pool;
2951
2952 bad_sort_array:
2953         mempool_destroy(pool->mapping_pool);
2954 bad_mapping_pool:
2955         dm_deferred_set_destroy(pool->all_io_ds);
2956 bad_all_io_ds:
2957         dm_deferred_set_destroy(pool->shared_read_ds);
2958 bad_shared_read_ds:
2959         destroy_workqueue(pool->wq);
2960 bad_wq:
2961         dm_kcopyd_client_destroy(pool->copier);
2962 bad_kcopyd_client:
2963         dm_bio_prison_destroy(pool->prison);
2964 bad_prison:
2965         kfree(pool);
2966 bad_pool:
2967         if (dm_pool_metadata_close(pmd))
2968                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
2969
2970         return err_p;
2971 }
2972
2973 static void __pool_inc(struct pool *pool)
2974 {
2975         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2976         pool->ref_count++;
2977 }
2978
2979 static void __pool_dec(struct pool *pool)
2980 {
2981         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
2982         BUG_ON(!pool->ref_count);
2983         if (!--pool->ref_count)
2984                 __pool_destroy(pool);
2985 }
2986
2987 static struct pool *__pool_find(struct mapped_device *pool_md,
2988                                 struct block_device *metadata_dev,
2989                                 unsigned long block_size, int read_only,
2990                                 char **error, int *created)
2991 {
2992         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
2993
2994         if (pool) {
2995                 if (pool->pool_md != pool_md) {
2996                         *error = "metadata device already in use by a pool";
2997                         return ERR_PTR(-EBUSY);
2998                 }
2999                 __pool_inc(pool);
3000
3001         } else {
3002                 pool = __pool_table_lookup(pool_md);
3003                 if (pool) {
3004                         if (pool->md_dev != metadata_dev) {
3005                                 *error = "different pool cannot replace a pool";
3006                                 return ERR_PTR(-EINVAL);
3007                         }
3008                         __pool_inc(pool);
3009
3010                 } else {
3011                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
3012                         *created = 1;
3013                 }
3014         }
3015
3016         return pool;
3017 }
3018
3019 /*----------------------------------------------------------------
3020  * Pool target methods
3021  *--------------------------------------------------------------*/
3022 static void pool_dtr(struct dm_target *ti)
3023 {
3024         struct pool_c *pt = ti->private;
3025
3026         mutex_lock(&dm_thin_pool_table.mutex);
3027
3028         unbind_control_target(pt->pool, ti);
3029         __pool_dec(pt->pool);
3030         dm_put_device(ti, pt->metadata_dev);
3031         dm_put_device(ti, pt->data_dev);
3032         kfree(pt);
3033
3034         mutex_unlock(&dm_thin_pool_table.mutex);
3035 }
3036
3037 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
3038                                struct dm_target *ti)
3039 {
3040         int r;
3041         unsigned argc;
3042         const char *arg_name;
3043
3044         static struct dm_arg _args[] = {
3045                 {0, 4, "Invalid number of pool feature arguments"},
3046         };
3047
3048         /*
3049          * No feature arguments supplied.
3050          */
3051         if (!as->argc)
3052                 return 0;
3053
3054         r = dm_read_arg_group(_args, as, &argc, &ti->error);
3055         if (r)
3056                 return -EINVAL;
3057
3058         while (argc && !r) {
3059                 arg_name = dm_shift_arg(as);
3060                 argc--;
3061
3062                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
3063                         pf->zero_new_blocks = false;
3064
3065                 else if (!strcasecmp(arg_name, "ignore_discard"))
3066                         pf->discard_enabled = false;
3067
3068                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
3069                         pf->discard_passdown = false;
3070
3071                 else if (!strcasecmp(arg_name, "read_only"))
3072                         pf->mode = PM_READ_ONLY;
3073
3074                 else if (!strcasecmp(arg_name, "error_if_no_space"))
3075                         pf->error_if_no_space = true;
3076
3077                 else {
3078                         ti->error = "Unrecognised pool feature requested";
3079                         r = -EINVAL;
3080                         break;
3081                 }
3082         }
3083
3084         return r;
3085 }
3086
3087 static void metadata_low_callback(void *context)
3088 {
3089         struct pool *pool = context;
3090
3091         DMWARN("%s: reached low water mark for metadata device: sending event.",
3092                dm_device_name(pool->pool_md));
3093
3094         dm_table_event(pool->ti->table);
3095 }
3096
3097 static sector_t get_dev_size(struct block_device *bdev)
3098 {
3099         return i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
3100 }
3101
3102 static void warn_if_metadata_device_too_big(struct block_device *bdev)
3103 {
3104         sector_t metadata_dev_size = get_dev_size(bdev);
3105         char buffer[BDEVNAME_SIZE];
3106
3107         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
3108                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
3109                        bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
3110 }
3111
3112 static sector_t get_metadata_dev_size(struct block_device *bdev)
3113 {
3114         sector_t metadata_dev_size = get_dev_size(bdev);
3115
3116         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS)
3117                 metadata_dev_size = THIN_METADATA_MAX_SECTORS;
3118
3119         return metadata_dev_size;
3120 }
3121
3122 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
3123 {
3124         sector_t metadata_dev_size = get_metadata_dev_size(bdev);
3125
3126         sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE);
3127
3128         return metadata_dev_size;
3129 }
3130
3131 /*
3132  * When a metadata threshold is crossed a dm event is triggered, and
3133  * userland should respond by growing the metadata device.  We could let
3134  * userland set the threshold, like we do with the data threshold, but I'm
3135  * not sure they know enough to do this well.
3136  */
3137 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
3138 {
3139         /*
3140          * 4M is ample for all ops with the possible exception of thin
3141          * device deletion which is harmless if it fails (just retry the
3142          * delete after you've grown the device).
3143          */
3144         dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
3145         return min((dm_block_t)1024ULL /* 4M */, quarter);
3146 }
3147
3148 /*
3149  * thin-pool <metadata dev> <data dev>
3150  *           <data block size (sectors)>
3151  *           <low water mark (blocks)>
3152  *           [<#feature args> [<arg>]*]
3153  *
3154  * Optional feature arguments are:
3155  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
3156  *           ignore_discard: disable discard
3157  *           no_discard_passdown: don't pass discards down to the data device
3158  *           read_only: Don't allow any changes to be made to the pool metadata.
3159  *           error_if_no_space: error IOs, instead of queueing, if no space.
3160  */
3161 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
3162 {
3163         int r, pool_created = 0;
3164         struct pool_c *pt;
3165         struct pool *pool;
3166         struct pool_features pf;
3167         struct dm_arg_set as;
3168         struct dm_dev *data_dev;
3169         unsigned long block_size;
3170         dm_block_t low_water_blocks;
3171         struct dm_dev *metadata_dev;
3172         fmode_t metadata_mode;
3173
3174         /*
3175          * FIXME Remove validation from scope of lock.
3176          */
3177         mutex_lock(&dm_thin_pool_table.mutex);
3178
3179         if (argc < 4) {
3180                 ti->error = "Invalid argument count";
3181                 r = -EINVAL;
3182                 goto out_unlock;
3183         }
3184
3185         as.argc = argc;
3186         as.argv = argv;
3187
3188         /*
3189          * Set default pool features.
3190          */
3191         pool_features_init(&pf);
3192
3193         dm_consume_args(&as, 4);
3194         r = parse_pool_features(&as, &pf, ti);
3195         if (r)
3196                 goto out_unlock;
3197
3198         metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
3199         r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
3200         if (r) {
3201                 ti->error = "Error opening metadata block device";
3202                 goto out_unlock;
3203         }
3204         warn_if_metadata_device_too_big(metadata_dev->bdev);
3205
3206         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
3207         if (r) {
3208                 ti->error = "Error getting data device";
3209                 goto out_metadata;
3210         }
3211
3212         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
3213             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
3214             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
3215             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
3216                 ti->error = "Invalid block size";
3217                 r = -EINVAL;
3218                 goto out;
3219         }
3220
3221         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
3222                 ti->error = "Invalid low water mark";
3223                 r = -EINVAL;
3224                 goto out;
3225         }
3226
3227         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
3228         if (!pt) {
3229                 r = -ENOMEM;
3230                 goto out;
3231         }
3232
3233         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
3234                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
3235         if (IS_ERR(pool)) {
3236                 r = PTR_ERR(pool);
3237                 goto out_free_pt;
3238         }
3239
3240         /*
3241          * 'pool_created' reflects whether this is the first table load.
3242          * Top level discard support is not allowed to be changed after
3243          * initial load.  This would require a pool reload to trigger thin
3244          * device changes.
3245          */
3246         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
3247                 ti->error = "Discard support cannot be disabled once enabled";
3248                 r = -EINVAL;
3249                 goto out_flags_changed;
3250         }
3251
3252         pt->pool = pool;
3253         pt->ti = ti;
3254         pt->metadata_dev = metadata_dev;
3255         pt->data_dev = data_dev;
3256         pt->low_water_blocks = low_water_blocks;
3257         pt->adjusted_pf = pt->requested_pf = pf;
3258         ti->num_flush_bios = 1;
3259
3260         /*
3261          * Only need to enable discards if the pool should pass
3262          * them down to the data device.  The thin device's discard
3263          * processing will cause mappings to be removed from the btree.
3264          */
3265         if (pf.discard_enabled && pf.discard_passdown) {
3266                 ti->num_discard_bios = 1;
3267
3268                 /*
3269                  * Setting 'discards_supported' circumvents the normal
3270                  * stacking of discard limits (this keeps the pool and
3271                  * thin devices' discard limits consistent).
3272                  */
3273                 ti->discards_supported = true;
3274         }
3275         ti->private = pt;
3276
3277         r = dm_pool_register_metadata_threshold(pt->pool->pmd,
3278                                                 calc_metadata_threshold(pt),
3279                                                 metadata_low_callback,
3280                                                 pool);
3281         if (r)
3282                 goto out_flags_changed;
3283
3284         pt->callbacks.congested_fn = pool_is_congested;
3285         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
3286
3287         mutex_unlock(&dm_thin_pool_table.mutex);
3288
3289         return 0;
3290
3291 out_flags_changed:
3292         __pool_dec(pool);
3293 out_free_pt:
3294         kfree(pt);
3295 out:
3296         dm_put_device(ti, data_dev);
3297 out_metadata:
3298         dm_put_device(ti, metadata_dev);
3299 out_unlock:
3300         mutex_unlock(&dm_thin_pool_table.mutex);
3301
3302         return r;
3303 }
3304
3305 static int pool_map(struct dm_target *ti, struct bio *bio)
3306 {
3307         int r;
3308         struct pool_c *pt = ti->private;
3309         struct pool *pool = pt->pool;
3310         unsigned long flags;
3311
3312         /*
3313          * As this is a singleton target, ti->begin is always zero.
3314          */
3315         spin_lock_irqsave(&pool->lock, flags);
3316         bio->bi_bdev = pt->data_dev->bdev;
3317         r = DM_MAPIO_REMAPPED;
3318         spin_unlock_irqrestore(&pool->lock, flags);
3319
3320         return r;
3321 }
3322
3323 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
3324 {
3325         int r;
3326         struct pool_c *pt = ti->private;
3327         struct pool *pool = pt->pool;
3328         sector_t data_size = ti->len;
3329         dm_block_t sb_data_size;
3330
3331         *need_commit = false;
3332
3333         (void) sector_div(data_size, pool->sectors_per_block);
3334
3335         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
3336         if (r) {
3337                 DMERR("%s: failed to retrieve data device size",
3338                       dm_device_name(pool->pool_md));
3339                 return r;
3340         }
3341
3342         if (data_size < sb_data_size) {
3343                 DMERR("%s: pool target (%llu blocks) too small: expected %llu",
3344                       dm_device_name(pool->pool_md),
3345                       (unsigned long long)data_size, sb_data_size);
3346                 return -EINVAL;
3347
3348         } else if (data_size > sb_data_size) {
3349                 if (dm_pool_metadata_needs_check(pool->pmd)) {
3350                         DMERR("%s: unable to grow the data device until repaired.",
3351                               dm_device_name(pool->pool_md));
3352                         return 0;
3353                 }
3354
3355                 if (sb_data_size)
3356                         DMINFO("%s: growing the data device from %llu to %llu blocks",
3357                                dm_device_name(pool->pool_md),
3358                                sb_data_size, (unsigned long long)data_size);
3359                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
3360                 if (r) {
3361                         metadata_operation_failed(pool, "dm_pool_resize_data_dev", r);
3362                         return r;
3363                 }
3364
3365                 *need_commit = true;
3366         }
3367
3368         return 0;
3369 }
3370
3371 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
3372 {
3373         int r;
3374         struct pool_c *pt = ti->private;
3375         struct pool *pool = pt->pool;
3376         dm_block_t metadata_dev_size, sb_metadata_dev_size;
3377
3378         *need_commit = false;
3379
3380         metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
3381
3382         r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
3383         if (r) {
3384                 DMERR("%s: failed to retrieve metadata device size",
3385                       dm_device_name(pool->pool_md));
3386                 return r;
3387         }
3388
3389         if (metadata_dev_size < sb_metadata_dev_size) {
3390                 DMERR("%s: metadata device (%llu blocks) too small: expected %llu",
3391                       dm_device_name(pool->pool_md),
3392                       metadata_dev_size, sb_metadata_dev_size);
3393                 return -EINVAL;
3394
3395         } else if (metadata_dev_size > sb_metadata_dev_size) {
3396                 if (dm_pool_metadata_needs_check(pool->pmd)) {
3397                         DMERR("%s: unable to grow the metadata device until repaired.",
3398                               dm_device_name(pool->pool_md));
3399                         return 0;
3400                 }
3401
3402                 warn_if_metadata_device_too_big(pool->md_dev);
3403                 DMINFO("%s: growing the metadata device from %llu to %llu blocks",
3404                        dm_device_name(pool->pool_md),
3405                        sb_metadata_dev_size, metadata_dev_size);
3406                 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
3407                 if (r) {
3408                         metadata_operation_failed(pool, "dm_pool_resize_metadata_dev", r);
3409                         return r;
3410                 }
3411
3412                 *need_commit = true;
3413         }
3414
3415         return 0;
3416 }
3417
3418 /*
3419  * Retrieves the number of blocks of the data device from
3420  * the superblock and compares it to the actual device size,
3421  * thus resizing the data device in case it has grown.
3422  *
3423  * This both copes with opening preallocated data devices in the ctr
3424  * being followed by a resume
3425  * -and-
3426  * calling the resume method individually after userspace has
3427  * grown the data device in reaction to a table event.
3428  */
3429 static int pool_preresume(struct dm_target *ti)
3430 {
3431         int r;
3432         bool need_commit1, need_commit2;
3433         struct pool_c *pt = ti->private;
3434         struct pool *pool = pt->pool;
3435
3436         /*
3437          * Take control of the pool object.
3438          */
3439         r = bind_control_target(pool, ti);
3440         if (r)
3441                 return r;
3442
3443         r = maybe_resize_data_dev(ti, &need_commit1);
3444         if (r)
3445                 return r;
3446
3447         r = maybe_resize_metadata_dev(ti, &need_commit2);
3448         if (r)
3449                 return r;
3450
3451         if (need_commit1 || need_commit2)
3452                 (void) commit(pool);
3453
3454         return 0;
3455 }
3456
3457 static void pool_suspend_active_thins(struct pool *pool)
3458 {
3459         struct thin_c *tc;
3460
3461         /* Suspend all active thin devices */
3462         tc = get_first_thin(pool);
3463         while (tc) {
3464                 dm_internal_suspend_noflush(tc->thin_md);
3465                 tc = get_next_thin(pool, tc);
3466         }
3467 }
3468
3469 static void pool_resume_active_thins(struct pool *pool)
3470 {
3471         struct thin_c *tc;
3472
3473         /* Resume all active thin devices */
3474         tc = get_first_thin(pool);
3475         while (tc) {
3476                 dm_internal_resume(tc->thin_md);
3477                 tc = get_next_thin(pool, tc);
3478         }
3479 }
3480
3481 static void pool_resume(struct dm_target *ti)
3482 {
3483         struct pool_c *pt = ti->private;
3484         struct pool *pool = pt->pool;
3485         unsigned long flags;
3486
3487         /*
3488          * Must requeue active_thins' bios and then resume
3489          * active_thins _before_ clearing 'suspend' flag.
3490          */
3491         requeue_bios(pool);
3492         pool_resume_active_thins(pool);
3493
3494         spin_lock_irqsave(&pool->lock, flags);
3495         pool->low_water_triggered = false;
3496         pool->suspended = false;
3497         spin_unlock_irqrestore(&pool->lock, flags);
3498
3499         do_waker(&pool->waker.work);
3500 }
3501
3502 static void pool_presuspend(struct dm_target *ti)
3503 {
3504         struct pool_c *pt = ti->private;
3505         struct pool *pool = pt->pool;
3506         unsigned long flags;
3507
3508         spin_lock_irqsave(&pool->lock, flags);
3509         pool->suspended = true;
3510         spin_unlock_irqrestore(&pool->lock, flags);
3511
3512         pool_suspend_active_thins(pool);
3513 }
3514
3515 static void pool_presuspend_undo(struct dm_target *ti)
3516 {
3517         struct pool_c *pt = ti->private;
3518         struct pool *pool = pt->pool;
3519         unsigned long flags;
3520
3521         pool_resume_active_thins(pool);
3522
3523         spin_lock_irqsave(&pool->lock, flags);
3524         pool->suspended = false;
3525         spin_unlock_irqrestore(&pool->lock, flags);
3526 }
3527
3528 static void pool_postsuspend(struct dm_target *ti)
3529 {
3530         struct pool_c *pt = ti->private;
3531         struct pool *pool = pt->pool;
3532
3533         cancel_delayed_work_sync(&pool->waker);
3534         cancel_delayed_work_sync(&pool->no_space_timeout);
3535         flush_workqueue(pool->wq);
3536         (void) commit(pool);
3537 }
3538
3539 static int check_arg_count(unsigned argc, unsigned args_required)
3540 {
3541         if (argc != args_required) {
3542                 DMWARN("Message received with %u arguments instead of %u.",
3543                        argc, args_required);
3544                 return -EINVAL;
3545         }
3546
3547         return 0;
3548 }
3549
3550 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
3551 {
3552         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
3553             *dev_id <= MAX_DEV_ID)
3554                 return 0;
3555
3556         if (warning)
3557                 DMWARN("Message received with invalid device id: %s", arg);
3558
3559         return -EINVAL;
3560 }
3561
3562 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
3563 {
3564         dm_thin_id dev_id;
3565         int r;
3566
3567         r = check_arg_count(argc, 2);
3568         if (r)
3569                 return r;
3570
3571         r = read_dev_id(argv[1], &dev_id, 1);
3572         if (r)
3573                 return r;
3574
3575         r = dm_pool_create_thin(pool->pmd, dev_id);
3576         if (r) {
3577                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
3578                        argv[1]);
3579                 return r;
3580         }
3581
3582         return 0;
3583 }
3584
3585 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3586 {
3587         dm_thin_id dev_id;
3588         dm_thin_id origin_dev_id;
3589         int r;
3590
3591         r = check_arg_count(argc, 3);
3592         if (r)
3593                 return r;
3594
3595         r = read_dev_id(argv[1], &dev_id, 1);
3596         if (r)
3597                 return r;
3598
3599         r = read_dev_id(argv[2], &origin_dev_id, 1);
3600         if (r)
3601                 return r;
3602
3603         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
3604         if (r) {
3605                 DMWARN("Creation of new snapshot %s of device %s failed.",
3606                        argv[1], argv[2]);
3607                 return r;
3608         }
3609
3610         return 0;
3611 }
3612
3613 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
3614 {
3615         dm_thin_id dev_id;
3616         int r;
3617
3618         r = check_arg_count(argc, 2);
3619         if (r)
3620                 return r;
3621
3622         r = read_dev_id(argv[1], &dev_id, 1);
3623         if (r)
3624                 return r;
3625
3626         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
3627         if (r)
3628                 DMWARN("Deletion of thin device %s failed.", argv[1]);
3629
3630         return r;
3631 }
3632
3633 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
3634 {
3635         dm_thin_id old_id, new_id;
3636         int r;
3637
3638         r = check_arg_count(argc, 3);
3639         if (r)
3640                 return r;
3641
3642         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
3643                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
3644                 return -EINVAL;
3645         }
3646
3647         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
3648                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
3649                 return -EINVAL;
3650         }
3651
3652         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
3653         if (r) {
3654                 DMWARN("Failed to change transaction id from %s to %s.",
3655                        argv[1], argv[2]);
3656                 return r;
3657         }
3658
3659         return 0;
3660 }
3661
3662 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3663 {
3664         int r;
3665
3666         r = check_arg_count(argc, 1);
3667         if (r)
3668                 return r;
3669
3670         (void) commit(pool);
3671
3672         r = dm_pool_reserve_metadata_snap(pool->pmd);
3673         if (r)
3674                 DMWARN("reserve_metadata_snap message failed.");
3675
3676         return r;
3677 }
3678
3679 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
3680 {
3681         int r;
3682
3683         r = check_arg_count(argc, 1);
3684         if (r)
3685                 return r;
3686
3687         r = dm_pool_release_metadata_snap(pool->pmd);
3688         if (r)
3689                 DMWARN("release_metadata_snap message failed.");
3690
3691         return r;
3692 }
3693
3694 /*
3695  * Messages supported:
3696  *   create_thin        <dev_id>
3697  *   create_snap        <dev_id> <origin_id>
3698  *   delete             <dev_id>
3699  *   set_transaction_id <current_trans_id> <new_trans_id>
3700  *   reserve_metadata_snap
3701  *   release_metadata_snap
3702  */
3703 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
3704 {
3705         int r = -EINVAL;
3706         struct pool_c *pt = ti->private;
3707         struct pool *pool = pt->pool;
3708
3709         if (get_pool_mode(pool) >= PM_READ_ONLY) {
3710                 DMERR("%s: unable to service pool target messages in READ_ONLY or FAIL mode",
3711                       dm_device_name(pool->pool_md));
3712                 return -EOPNOTSUPP;
3713         }
3714
3715         if (!strcasecmp(argv[0], "create_thin"))
3716                 r = process_create_thin_mesg(argc, argv, pool);
3717
3718         else if (!strcasecmp(argv[0], "create_snap"))
3719                 r = process_create_snap_mesg(argc, argv, pool);
3720
3721         else if (!strcasecmp(argv[0], "delete"))
3722                 r = process_delete_mesg(argc, argv, pool);
3723
3724         else if (!strcasecmp(argv[0], "set_transaction_id"))
3725                 r = process_set_transaction_id_mesg(argc, argv, pool);
3726
3727         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
3728                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
3729
3730         else if (!strcasecmp(argv[0], "release_metadata_snap"))
3731                 r = process_release_metadata_snap_mesg(argc, argv, pool);
3732
3733         else
3734                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
3735
3736         if (!r)
3737                 (void) commit(pool);
3738
3739         return r;
3740 }
3741
3742 static void emit_flags(struct pool_features *pf, char *result,
3743                        unsigned sz, unsigned maxlen)
3744 {
3745         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
3746                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY) +
3747                 pf->error_if_no_space;
3748         DMEMIT("%u ", count);
3749
3750         if (!pf->zero_new_blocks)
3751                 DMEMIT("skip_block_zeroing ");
3752
3753         if (!pf->discard_enabled)
3754                 DMEMIT("ignore_discard ");
3755
3756         if (!pf->discard_passdown)
3757                 DMEMIT("no_discard_passdown ");
3758
3759         if (pf->mode == PM_READ_ONLY)
3760                 DMEMIT("read_only ");
3761
3762         if (pf->error_if_no_space)
3763                 DMEMIT("error_if_no_space ");
3764 }
3765
3766 /*
3767  * Status line is:
3768  *    <transaction id> <used metadata sectors>/<total metadata sectors>
3769  *    <used data sectors>/<total data sectors> <held metadata root>
3770  *    <pool mode> <discard config> <no space config> <needs_check>
3771  */
3772 static void pool_status(struct dm_target *ti, status_type_t type,
3773                         unsigned status_flags, char *result, unsigned maxlen)
3774 {
3775         int r;
3776         unsigned sz = 0;
3777         uint64_t transaction_id;
3778         dm_block_t nr_free_blocks_data;
3779         dm_block_t nr_free_blocks_metadata;
3780         dm_block_t nr_blocks_data;
3781         dm_block_t nr_blocks_metadata;
3782         dm_block_t held_root;
3783         char buf[BDEVNAME_SIZE];
3784         char buf2[BDEVNAME_SIZE];
3785         struct pool_c *pt = ti->private;
3786         struct pool *pool = pt->pool;
3787
3788         switch (type) {
3789         case STATUSTYPE_INFO:
3790                 if (get_pool_mode(pool) == PM_FAIL) {
3791                         DMEMIT("Fail");
3792                         break;
3793                 }
3794
3795                 /* Commit to ensure statistics aren't out-of-date */
3796                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3797                         (void) commit(pool);
3798
3799                 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
3800                 if (r) {
3801                         DMERR("%s: dm_pool_get_metadata_transaction_id returned %d",
3802                               dm_device_name(pool->pool_md), r);
3803                         goto err;
3804                 }
3805
3806                 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
3807                 if (r) {
3808                         DMERR("%s: dm_pool_get_free_metadata_block_count returned %d",
3809                               dm_device_name(pool->pool_md), r);
3810                         goto err;
3811                 }
3812
3813                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
3814                 if (r) {
3815                         DMERR("%s: dm_pool_get_metadata_dev_size returned %d",
3816                               dm_device_name(pool->pool_md), r);
3817                         goto err;
3818                 }
3819
3820                 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
3821                 if (r) {
3822                         DMERR("%s: dm_pool_get_free_block_count returned %d",
3823                               dm_device_name(pool->pool_md), r);
3824                         goto err;
3825                 }
3826
3827                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
3828                 if (r) {
3829                         DMERR("%s: dm_pool_get_data_dev_size returned %d",
3830                               dm_device_name(pool->pool_md), r);
3831                         goto err;
3832                 }
3833
3834                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
3835                 if (r) {
3836                         DMERR("%s: dm_pool_get_metadata_snap returned %d",
3837                               dm_device_name(pool->pool_md), r);
3838                         goto err;
3839                 }
3840
3841                 DMEMIT("%llu %llu/%llu %llu/%llu ",
3842                        (unsigned long long)transaction_id,
3843                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3844                        (unsigned long long)nr_blocks_metadata,
3845                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
3846                        (unsigned long long)nr_blocks_data);
3847
3848                 if (held_root)
3849                         DMEMIT("%llu ", held_root);
3850                 else
3851                         DMEMIT("- ");
3852
3853                 if (pool->pf.mode == PM_OUT_OF_DATA_SPACE)
3854                         DMEMIT("out_of_data_space ");
3855                 else if (pool->pf.mode == PM_READ_ONLY)
3856                         DMEMIT("ro ");
3857                 else
3858                         DMEMIT("rw ");
3859
3860                 if (!pool->pf.discard_enabled)
3861                         DMEMIT("ignore_discard ");
3862                 else if (pool->pf.discard_passdown)
3863                         DMEMIT("discard_passdown ");
3864                 else
3865                         DMEMIT("no_discard_passdown ");
3866
3867                 if (pool->pf.error_if_no_space)
3868                         DMEMIT("error_if_no_space ");
3869                 else
3870                         DMEMIT("queue_if_no_space ");
3871
3872                 if (dm_pool_metadata_needs_check(pool->pmd))
3873                         DMEMIT("needs_check ");
3874                 else
3875                         DMEMIT("- ");
3876
3877                 break;
3878
3879         case STATUSTYPE_TABLE:
3880                 DMEMIT("%s %s %lu %llu ",
3881                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
3882                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
3883                        (unsigned long)pool->sectors_per_block,
3884                        (unsigned long long)pt->low_water_blocks);
3885                 emit_flags(&pt->requested_pf, result, sz, maxlen);
3886                 break;
3887         }
3888         return;
3889
3890 err:
3891         DMEMIT("Error");
3892 }
3893
3894 static int pool_iterate_devices(struct dm_target *ti,
3895                                 iterate_devices_callout_fn fn, void *data)
3896 {
3897         struct pool_c *pt = ti->private;
3898
3899         return fn(ti, pt->data_dev, 0, ti->len, data);
3900 }
3901
3902 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
3903 {
3904         struct pool_c *pt = ti->private;
3905         struct pool *pool = pt->pool;
3906         sector_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3907
3908         /*
3909          * If max_sectors is smaller than pool->sectors_per_block adjust it
3910          * to the highest possible power-of-2 factor of pool->sectors_per_block.
3911          * This is especially beneficial when the pool's data device is a RAID
3912          * device that has a full stripe width that matches pool->sectors_per_block
3913          * -- because even though partial RAID stripe-sized IOs will be issued to a
3914          *    single RAID stripe; when aggregated they will end on a full RAID stripe
3915          *    boundary.. which avoids additional partial RAID stripe writes cascading
3916          */
3917         if (limits->max_sectors < pool->sectors_per_block) {
3918                 while (!is_factor(pool->sectors_per_block, limits->max_sectors)) {
3919                         if ((limits->max_sectors & (limits->max_sectors - 1)) == 0)
3920                                 limits->max_sectors--;
3921                         limits->max_sectors = rounddown_pow_of_two(limits->max_sectors);
3922                 }
3923         }
3924
3925         /*
3926          * If the system-determined stacked limits are compatible with the
3927          * pool's blocksize (io_opt is a factor) do not override them.
3928          */
3929         if (io_opt_sectors < pool->sectors_per_block ||
3930             !is_factor(io_opt_sectors, pool->sectors_per_block)) {
3931                 if (is_factor(pool->sectors_per_block, limits->max_sectors))
3932                         blk_limits_io_min(limits, limits->max_sectors << SECTOR_SHIFT);
3933                 else
3934                         blk_limits_io_min(limits, pool->sectors_per_block << SECTOR_SHIFT);
3935                 blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
3936         }
3937
3938         /*
3939          * pt->adjusted_pf is a staging area for the actual features to use.
3940          * They get transferred to the live pool in bind_control_target()
3941          * called from pool_preresume().
3942          */
3943         if (!pt->adjusted_pf.discard_enabled) {
3944                 /*
3945                  * Must explicitly disallow stacking discard limits otherwise the
3946                  * block layer will stack them if pool's data device has support.
3947                  * QUEUE_FLAG_DISCARD wouldn't be set but there is no way for the
3948                  * user to see that, so make sure to set all discard limits to 0.
3949                  */
3950                 limits->discard_granularity = 0;
3951                 return;
3952         }
3953
3954         disable_passdown_if_not_supported(pt);
3955
3956         /*
3957          * The pool uses the same discard limits as the underlying data
3958          * device.  DM core has already set this up.
3959          */
3960 }
3961
3962 static struct target_type pool_target = {
3963         .name = "thin-pool",
3964         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
3965                     DM_TARGET_IMMUTABLE,
3966         .version = {1, 19, 0},
3967         .module = THIS_MODULE,
3968         .ctr = pool_ctr,
3969         .dtr = pool_dtr,
3970         .map = pool_map,
3971         .presuspend = pool_presuspend,
3972         .presuspend_undo = pool_presuspend_undo,
3973         .postsuspend = pool_postsuspend,
3974         .preresume = pool_preresume,
3975         .resume = pool_resume,
3976         .message = pool_message,
3977         .status = pool_status,
3978         .iterate_devices = pool_iterate_devices,
3979         .io_hints = pool_io_hints,
3980 };
3981
3982 /*----------------------------------------------------------------
3983  * Thin target methods
3984  *--------------------------------------------------------------*/
3985 static void thin_get(struct thin_c *tc)
3986 {
3987         atomic_inc(&tc->refcount);
3988 }
3989
3990 static void thin_put(struct thin_c *tc)
3991 {
3992         if (atomic_dec_and_test(&tc->refcount))
3993                 complete(&tc->can_destroy);
3994 }
3995
3996 static void thin_dtr(struct dm_target *ti)
3997 {
3998         struct thin_c *tc = ti->private;
3999         unsigned long flags;
4000
4001         spin_lock_irqsave(&tc->pool->lock, flags);
4002         list_del_rcu(&tc->list);
4003         spin_unlock_irqrestore(&tc->pool->lock, flags);
4004         synchronize_rcu();
4005
4006         thin_put(tc);
4007         wait_for_completion(&tc->can_destroy);
4008
4009         mutex_lock(&dm_thin_pool_table.mutex);
4010
4011         __pool_dec(tc->pool);
4012         dm_pool_close_thin_device(tc->td);
4013         dm_put_device(ti, tc->pool_dev);
4014         if (tc->origin_dev)
4015                 dm_put_device(ti, tc->origin_dev);
4016         kfree(tc);
4017
4018         mutex_unlock(&dm_thin_pool_table.mutex);
4019 }
4020
4021 /*
4022  * Thin target parameters:
4023  *
4024  * <pool_dev> <dev_id> [origin_dev]
4025  *
4026  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
4027  * dev_id: the internal device identifier
4028  * origin_dev: a device external to the pool that should act as the origin
4029  *
4030  * If the pool device has discards disabled, they get disabled for the thin
4031  * device as well.
4032  */
4033 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
4034 {
4035         int r;
4036         struct thin_c *tc;
4037         struct dm_dev *pool_dev, *origin_dev;
4038         struct mapped_device *pool_md;
4039         unsigned long flags;
4040
4041         mutex_lock(&dm_thin_pool_table.mutex);
4042
4043         if (argc != 2 && argc != 3) {
4044                 ti->error = "Invalid argument count";
4045                 r = -EINVAL;
4046                 goto out_unlock;
4047         }
4048
4049         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
4050         if (!tc) {
4051                 ti->error = "Out of memory";
4052                 r = -ENOMEM;
4053                 goto out_unlock;
4054         }
4055         tc->thin_md = dm_table_get_md(ti->table);
4056         spin_lock_init(&tc->lock);
4057         INIT_LIST_HEAD(&tc->deferred_cells);
4058         bio_list_init(&tc->deferred_bio_list);
4059         bio_list_init(&tc->retry_on_resume_list);
4060         tc->sort_bio_list = RB_ROOT;
4061
4062         if (argc == 3) {
4063                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
4064                 if (r) {
4065                         ti->error = "Error opening origin device";
4066                         goto bad_origin_dev;
4067                 }
4068                 tc->origin_dev = origin_dev;
4069         }
4070
4071         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
4072         if (r) {
4073                 ti->error = "Error opening pool device";
4074                 goto bad_pool_dev;
4075         }
4076         tc->pool_dev = pool_dev;
4077
4078         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
4079                 ti->error = "Invalid device id";
4080                 r = -EINVAL;
4081                 goto bad_common;
4082         }
4083
4084         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
4085         if (!pool_md) {
4086                 ti->error = "Couldn't get pool mapped device";
4087                 r = -EINVAL;
4088                 goto bad_common;
4089         }
4090
4091         tc->pool = __pool_table_lookup(pool_md);
4092         if (!tc->pool) {
4093                 ti->error = "Couldn't find pool object";
4094                 r = -EINVAL;
4095                 goto bad_pool_lookup;
4096         }
4097         __pool_inc(tc->pool);
4098
4099         if (get_pool_mode(tc->pool) == PM_FAIL) {
4100                 ti->error = "Couldn't open thin device, Pool is in fail mode";
4101                 r = -EINVAL;
4102                 goto bad_pool;
4103         }
4104
4105         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
4106         if (r) {
4107                 ti->error = "Couldn't open thin internal device";
4108                 goto bad_pool;
4109         }
4110
4111         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
4112         if (r)
4113                 goto bad;
4114
4115         ti->num_flush_bios = 1;
4116         ti->flush_supported = true;
4117         ti->per_io_data_size = sizeof(struct dm_thin_endio_hook);
4118
4119         /* In case the pool supports discards, pass them on. */
4120         if (tc->pool->pf.discard_enabled) {
4121                 ti->discards_supported = true;
4122                 ti->num_discard_bios = 1;
4123                 ti->split_discard_bios = false;
4124         }
4125
4126         mutex_unlock(&dm_thin_pool_table.mutex);
4127
4128         spin_lock_irqsave(&tc->pool->lock, flags);
4129         if (tc->pool->suspended) {
4130                 spin_unlock_irqrestore(&tc->pool->lock, flags);
4131                 mutex_lock(&dm_thin_pool_table.mutex); /* reacquire for __pool_dec */
4132                 ti->error = "Unable to activate thin device while pool is suspended";
4133                 r = -EINVAL;
4134                 goto bad;
4135         }
4136         atomic_set(&tc->refcount, 1);
4137         init_completion(&tc->can_destroy);
4138         list_add_tail_rcu(&tc->list, &tc->pool->active_thins);
4139         spin_unlock_irqrestore(&tc->pool->lock, flags);
4140         /*
4141          * This synchronize_rcu() call is needed here otherwise we risk a
4142          * wake_worker() call finding no bios to process (because the newly
4143          * added tc isn't yet visible).  So this reduces latency since we
4144          * aren't then dependent on the periodic commit to wake_worker().
4145          */
4146         synchronize_rcu();
4147
4148         dm_put(pool_md);
4149
4150         return 0;
4151
4152 bad:
4153         dm_pool_close_thin_device(tc->td);
4154 bad_pool:
4155         __pool_dec(tc->pool);
4156 bad_pool_lookup:
4157         dm_put(pool_md);
4158 bad_common:
4159         dm_put_device(ti, tc->pool_dev);
4160 bad_pool_dev:
4161         if (tc->origin_dev)
4162                 dm_put_device(ti, tc->origin_dev);
4163 bad_origin_dev:
4164         kfree(tc);
4165 out_unlock:
4166         mutex_unlock(&dm_thin_pool_table.mutex);
4167
4168         return r;
4169 }
4170
4171 static int thin_map(struct dm_target *ti, struct bio *bio)
4172 {
4173         bio->bi_iter.bi_sector = dm_target_offset(ti, bio->bi_iter.bi_sector);
4174
4175         return thin_bio_map(ti, bio);
4176 }
4177
4178 static int thin_endio(struct dm_target *ti, struct bio *bio,
4179                 blk_status_t *err)
4180 {
4181         unsigned long flags;
4182         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
4183         struct list_head work;
4184         struct dm_thin_new_mapping *m, *tmp;
4185         struct pool *pool = h->tc->pool;
4186
4187         if (h->shared_read_entry) {
4188                 INIT_LIST_HEAD(&work);
4189                 dm_deferred_entry_dec(h->shared_read_entry, &work);
4190
4191                 spin_lock_irqsave(&pool->lock, flags);
4192                 list_for_each_entry_safe(m, tmp, &work, list) {
4193                         list_del(&m->list);
4194                         __complete_mapping_preparation(m);
4195                 }
4196                 spin_unlock_irqrestore(&pool->lock, flags);
4197         }
4198
4199         if (h->all_io_entry) {
4200                 INIT_LIST_HEAD(&work);
4201                 dm_deferred_entry_dec(h->all_io_entry, &work);
4202                 if (!list_empty(&work)) {
4203                         spin_lock_irqsave(&pool->lock, flags);
4204                         list_for_each_entry_safe(m, tmp, &work, list)
4205                                 list_add_tail(&m->list, &pool->prepared_discards);
4206                         spin_unlock_irqrestore(&pool->lock, flags);
4207                         wake_worker(pool);
4208                 }
4209         }
4210
4211         if (h->cell)
4212                 cell_defer_no_holder(h->tc, h->cell);
4213
4214         return DM_ENDIO_DONE;
4215 }
4216
4217 static void thin_presuspend(struct dm_target *ti)
4218 {
4219         struct thin_c *tc = ti->private;
4220
4221         if (dm_noflush_suspending(ti))
4222                 noflush_work(tc, do_noflush_start);
4223 }
4224
4225 static void thin_postsuspend(struct dm_target *ti)
4226 {
4227         struct thin_c *tc = ti->private;
4228
4229         /*
4230          * The dm_noflush_suspending flag has been cleared by now, so
4231          * unfortunately we must always run this.
4232          */
4233         noflush_work(tc, do_noflush_stop);
4234 }
4235
4236 static int thin_preresume(struct dm_target *ti)
4237 {
4238         struct thin_c *tc = ti->private;
4239
4240         if (tc->origin_dev)
4241                 tc->origin_size = get_dev_size(tc->origin_dev->bdev);
4242
4243         return 0;
4244 }
4245
4246 /*
4247  * <nr mapped sectors> <highest mapped sector>
4248  */
4249 static void thin_status(struct dm_target *ti, status_type_t type,
4250                         unsigned status_flags, char *result, unsigned maxlen)
4251 {
4252         int r;
4253         ssize_t sz = 0;
4254         dm_block_t mapped, highest;
4255         char buf[BDEVNAME_SIZE];
4256         struct thin_c *tc = ti->private;
4257
4258         if (get_pool_mode(tc->pool) == PM_FAIL) {
4259                 DMEMIT("Fail");
4260                 return;
4261         }
4262
4263         if (!tc->td)
4264                 DMEMIT("-");
4265         else {
4266                 switch (type) {
4267                 case STATUSTYPE_INFO:
4268                         r = dm_thin_get_mapped_count(tc->td, &mapped);
4269                         if (r) {
4270                                 DMERR("dm_thin_get_mapped_count returned %d", r);
4271                                 goto err;
4272                         }
4273
4274                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
4275                         if (r < 0) {
4276                                 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
4277                                 goto err;
4278                         }
4279
4280                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
4281                         if (r)
4282                                 DMEMIT("%llu", ((highest + 1) *
4283                                                 tc->pool->sectors_per_block) - 1);
4284                         else
4285                                 DMEMIT("-");
4286                         break;
4287
4288                 case STATUSTYPE_TABLE:
4289                         DMEMIT("%s %lu",
4290                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
4291                                (unsigned long) tc->dev_id);
4292                         if (tc->origin_dev)
4293                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
4294                         break;
4295                 }
4296         }
4297
4298         return;
4299
4300 err:
4301         DMEMIT("Error");
4302 }
4303
4304 static int thin_iterate_devices(struct dm_target *ti,
4305                                 iterate_devices_callout_fn fn, void *data)
4306 {
4307         sector_t blocks;
4308         struct thin_c *tc = ti->private;
4309         struct pool *pool = tc->pool;
4310
4311         /*
4312          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
4313          * we follow a more convoluted path through to the pool's target.
4314          */
4315         if (!pool->ti)
4316                 return 0;       /* nothing is bound */
4317
4318         blocks = pool->ti->len;
4319         (void) sector_div(blocks, pool->sectors_per_block);
4320         if (blocks)
4321                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
4322
4323         return 0;
4324 }
4325
4326 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
4327 {
4328         struct thin_c *tc = ti->private;
4329         struct pool *pool = tc->pool;
4330
4331         if (!pool->pf.discard_enabled)
4332                 return;
4333
4334         limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
4335         limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
4336 }
4337
4338 static struct target_type thin_target = {
4339         .name = "thin",
4340         .version = {1, 19, 0},
4341         .module = THIS_MODULE,
4342         .ctr = thin_ctr,
4343         .dtr = thin_dtr,
4344         .map = thin_map,
4345         .end_io = thin_endio,
4346         .preresume = thin_preresume,
4347         .presuspend = thin_presuspend,
4348         .postsuspend = thin_postsuspend,
4349         .status = thin_status,
4350         .iterate_devices = thin_iterate_devices,
4351         .io_hints = thin_io_hints,
4352 };
4353
4354 /*----------------------------------------------------------------*/
4355
4356 static int __init dm_thin_init(void)
4357 {
4358         int r;
4359
4360         pool_table_init();
4361
4362         r = dm_register_target(&thin_target);
4363         if (r)
4364                 return r;
4365
4366         r = dm_register_target(&pool_target);
4367         if (r)
4368                 goto bad_pool_target;
4369
4370         r = -ENOMEM;
4371
4372         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
4373         if (!_new_mapping_cache)
4374                 goto bad_new_mapping_cache;
4375
4376         return 0;
4377
4378 bad_new_mapping_cache:
4379         dm_unregister_target(&pool_target);
4380 bad_pool_target:
4381         dm_unregister_target(&thin_target);
4382
4383         return r;
4384 }
4385
4386 static void dm_thin_exit(void)
4387 {
4388         dm_unregister_target(&thin_target);
4389         dm_unregister_target(&pool_target);
4390
4391         kmem_cache_destroy(_new_mapping_cache);
4392 }
4393
4394 module_init(dm_thin_init);
4395 module_exit(dm_thin_exit);
4396
4397 module_param_named(no_space_timeout, no_space_timeout_secs, uint, S_IRUGO | S_IWUSR);
4398 MODULE_PARM_DESC(no_space_timeout, "Out of data space queue IO timeout in seconds");
4399
4400 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
4401 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
4402 MODULE_LICENSE("GPL");