Merge branch 'for-3.11' of git://linux-nfs.org/~bfields/linux
[platform/adaptation/renesas_rcar/renesas_kernel.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.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/list.h>
15 #include <linux/init.h>
16 #include <linux/module.h>
17 #include <linux/slab.h>
18
19 #define DM_MSG_PREFIX   "thin"
20
21 /*
22  * Tunable constants
23  */
24 #define ENDIO_HOOK_POOL_SIZE 1024
25 #define MAPPING_POOL_SIZE 1024
26 #define PRISON_CELLS 1024
27 #define COMMIT_PERIOD HZ
28
29 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(snapshot_copy_throttle,
30                 "A percentage of time allocated for copy on write");
31
32 /*
33  * The block size of the device holding pool data must be
34  * between 64KB and 1GB.
35  */
36 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
37 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
38
39 /*
40  * Device id is restricted to 24 bits.
41  */
42 #define MAX_DEV_ID ((1 << 24) - 1)
43
44 /*
45  * How do we handle breaking sharing of data blocks?
46  * =================================================
47  *
48  * We use a standard copy-on-write btree to store the mappings for the
49  * devices (note I'm talking about copy-on-write of the metadata here, not
50  * the data).  When you take an internal snapshot you clone the root node
51  * of the origin btree.  After this there is no concept of an origin or a
52  * snapshot.  They are just two device trees that happen to point to the
53  * same data blocks.
54  *
55  * When we get a write in we decide if it's to a shared data block using
56  * some timestamp magic.  If it is, we have to break sharing.
57  *
58  * Let's say we write to a shared block in what was the origin.  The
59  * steps are:
60  *
61  * i) plug io further to this physical block. (see bio_prison code).
62  *
63  * ii) quiesce any read io to that shared data block.  Obviously
64  * including all devices that share this block.  (see dm_deferred_set code)
65  *
66  * iii) copy the data block to a newly allocate block.  This step can be
67  * missed out if the io covers the block. (schedule_copy).
68  *
69  * iv) insert the new mapping into the origin's btree
70  * (process_prepared_mapping).  This act of inserting breaks some
71  * sharing of btree nodes between the two devices.  Breaking sharing only
72  * effects the btree of that specific device.  Btrees for the other
73  * devices that share the block never change.  The btree for the origin
74  * device as it was after the last commit is untouched, ie. we're using
75  * persistent data structures in the functional programming sense.
76  *
77  * v) unplug io to this physical block, including the io that triggered
78  * the breaking of sharing.
79  *
80  * Steps (ii) and (iii) occur in parallel.
81  *
82  * The metadata _doesn't_ need to be committed before the io continues.  We
83  * get away with this because the io is always written to a _new_ block.
84  * If there's a crash, then:
85  *
86  * - The origin mapping will point to the old origin block (the shared
87  * one).  This will contain the data as it was before the io that triggered
88  * the breaking of sharing came in.
89  *
90  * - The snap mapping still points to the old block.  As it would after
91  * the commit.
92  *
93  * The downside of this scheme is the timestamp magic isn't perfect, and
94  * will continue to think that data block in the snapshot device is shared
95  * even after the write to the origin has broken sharing.  I suspect data
96  * blocks will typically be shared by many different devices, so we're
97  * breaking sharing n + 1 times, rather than n, where n is the number of
98  * devices that reference this data block.  At the moment I think the
99  * benefits far, far outweigh the disadvantages.
100  */
101
102 /*----------------------------------------------------------------*/
103
104 /*
105  * Key building.
106  */
107 static void build_data_key(struct dm_thin_device *td,
108                            dm_block_t b, struct dm_cell_key *key)
109 {
110         key->virtual = 0;
111         key->dev = dm_thin_dev_id(td);
112         key->block = b;
113 }
114
115 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
116                               struct dm_cell_key *key)
117 {
118         key->virtual = 1;
119         key->dev = dm_thin_dev_id(td);
120         key->block = b;
121 }
122
123 /*----------------------------------------------------------------*/
124
125 /*
126  * A pool device ties together a metadata device and a data device.  It
127  * also provides the interface for creating and destroying internal
128  * devices.
129  */
130 struct dm_thin_new_mapping;
131
132 /*
133  * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
134  */
135 enum pool_mode {
136         PM_WRITE,               /* metadata may be changed */
137         PM_READ_ONLY,           /* metadata may not be changed */
138         PM_FAIL,                /* all I/O fails */
139 };
140
141 struct pool_features {
142         enum pool_mode mode;
143
144         bool zero_new_blocks:1;
145         bool discard_enabled:1;
146         bool discard_passdown:1;
147 };
148
149 struct thin_c;
150 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
151 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
152
153 struct pool {
154         struct list_head list;
155         struct dm_target *ti;   /* Only set if a pool target is bound */
156
157         struct mapped_device *pool_md;
158         struct block_device *md_dev;
159         struct dm_pool_metadata *pmd;
160
161         dm_block_t low_water_blocks;
162         uint32_t sectors_per_block;
163         int sectors_per_block_shift;
164
165         struct pool_features pf;
166         unsigned low_water_triggered:1; /* A dm event has been sent */
167         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
168
169         struct dm_bio_prison *prison;
170         struct dm_kcopyd_client *copier;
171
172         struct workqueue_struct *wq;
173         struct work_struct worker;
174         struct delayed_work waker;
175
176         unsigned long last_commit_jiffies;
177         unsigned ref_count;
178
179         spinlock_t lock;
180         struct bio_list deferred_bios;
181         struct bio_list deferred_flush_bios;
182         struct list_head prepared_mappings;
183         struct list_head prepared_discards;
184
185         struct bio_list retry_on_resume_list;
186
187         struct dm_deferred_set *shared_read_ds;
188         struct dm_deferred_set *all_io_ds;
189
190         struct dm_thin_new_mapping *next_mapping;
191         mempool_t *mapping_pool;
192
193         process_bio_fn process_bio;
194         process_bio_fn process_discard;
195
196         process_mapping_fn process_prepared_mapping;
197         process_mapping_fn process_prepared_discard;
198 };
199
200 static enum pool_mode get_pool_mode(struct pool *pool);
201 static void set_pool_mode(struct pool *pool, enum pool_mode mode);
202
203 /*
204  * Target context for a pool.
205  */
206 struct pool_c {
207         struct dm_target *ti;
208         struct pool *pool;
209         struct dm_dev *data_dev;
210         struct dm_dev *metadata_dev;
211         struct dm_target_callbacks callbacks;
212
213         dm_block_t low_water_blocks;
214         struct pool_features requested_pf; /* Features requested during table load */
215         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
216 };
217
218 /*
219  * Target context for a thin.
220  */
221 struct thin_c {
222         struct dm_dev *pool_dev;
223         struct dm_dev *origin_dev;
224         dm_thin_id dev_id;
225
226         struct pool *pool;
227         struct dm_thin_device *td;
228 };
229
230 /*----------------------------------------------------------------*/
231
232 /*
233  * wake_worker() is used when new work is queued and when pool_resume is
234  * ready to continue deferred IO processing.
235  */
236 static void wake_worker(struct pool *pool)
237 {
238         queue_work(pool->wq, &pool->worker);
239 }
240
241 /*----------------------------------------------------------------*/
242
243 static int bio_detain(struct pool *pool, struct dm_cell_key *key, struct bio *bio,
244                       struct dm_bio_prison_cell **cell_result)
245 {
246         int r;
247         struct dm_bio_prison_cell *cell_prealloc;
248
249         /*
250          * Allocate a cell from the prison's mempool.
251          * This might block but it can't fail.
252          */
253         cell_prealloc = dm_bio_prison_alloc_cell(pool->prison, GFP_NOIO);
254
255         r = dm_bio_detain(pool->prison, key, bio, cell_prealloc, cell_result);
256         if (r)
257                 /*
258                  * We reused an old cell; we can get rid of
259                  * the new one.
260                  */
261                 dm_bio_prison_free_cell(pool->prison, cell_prealloc);
262
263         return r;
264 }
265
266 static void cell_release(struct pool *pool,
267                          struct dm_bio_prison_cell *cell,
268                          struct bio_list *bios)
269 {
270         dm_cell_release(pool->prison, cell, bios);
271         dm_bio_prison_free_cell(pool->prison, cell);
272 }
273
274 static void cell_release_no_holder(struct pool *pool,
275                                    struct dm_bio_prison_cell *cell,
276                                    struct bio_list *bios)
277 {
278         dm_cell_release_no_holder(pool->prison, cell, bios);
279         dm_bio_prison_free_cell(pool->prison, cell);
280 }
281
282 static void cell_defer_no_holder_no_free(struct thin_c *tc,
283                                          struct dm_bio_prison_cell *cell)
284 {
285         struct pool *pool = tc->pool;
286         unsigned long flags;
287
288         spin_lock_irqsave(&pool->lock, flags);
289         dm_cell_release_no_holder(pool->prison, cell, &pool->deferred_bios);
290         spin_unlock_irqrestore(&pool->lock, flags);
291
292         wake_worker(pool);
293 }
294
295 static void cell_error(struct pool *pool,
296                        struct dm_bio_prison_cell *cell)
297 {
298         dm_cell_error(pool->prison, cell);
299         dm_bio_prison_free_cell(pool->prison, cell);
300 }
301
302 /*----------------------------------------------------------------*/
303
304 /*
305  * A global list of pools that uses a struct mapped_device as a key.
306  */
307 static struct dm_thin_pool_table {
308         struct mutex mutex;
309         struct list_head pools;
310 } dm_thin_pool_table;
311
312 static void pool_table_init(void)
313 {
314         mutex_init(&dm_thin_pool_table.mutex);
315         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
316 }
317
318 static void __pool_table_insert(struct pool *pool)
319 {
320         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
321         list_add(&pool->list, &dm_thin_pool_table.pools);
322 }
323
324 static void __pool_table_remove(struct pool *pool)
325 {
326         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
327         list_del(&pool->list);
328 }
329
330 static struct pool *__pool_table_lookup(struct mapped_device *md)
331 {
332         struct pool *pool = NULL, *tmp;
333
334         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
335
336         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
337                 if (tmp->pool_md == md) {
338                         pool = tmp;
339                         break;
340                 }
341         }
342
343         return pool;
344 }
345
346 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
347 {
348         struct pool *pool = NULL, *tmp;
349
350         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
351
352         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
353                 if (tmp->md_dev == md_dev) {
354                         pool = tmp;
355                         break;
356                 }
357         }
358
359         return pool;
360 }
361
362 /*----------------------------------------------------------------*/
363
364 struct dm_thin_endio_hook {
365         struct thin_c *tc;
366         struct dm_deferred_entry *shared_read_entry;
367         struct dm_deferred_entry *all_io_entry;
368         struct dm_thin_new_mapping *overwrite_mapping;
369 };
370
371 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
372 {
373         struct bio *bio;
374         struct bio_list bios;
375
376         bio_list_init(&bios);
377         bio_list_merge(&bios, master);
378         bio_list_init(master);
379
380         while ((bio = bio_list_pop(&bios))) {
381                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
382
383                 if (h->tc == tc)
384                         bio_endio(bio, DM_ENDIO_REQUEUE);
385                 else
386                         bio_list_add(master, bio);
387         }
388 }
389
390 static void requeue_io(struct thin_c *tc)
391 {
392         struct pool *pool = tc->pool;
393         unsigned long flags;
394
395         spin_lock_irqsave(&pool->lock, flags);
396         __requeue_bio_list(tc, &pool->deferred_bios);
397         __requeue_bio_list(tc, &pool->retry_on_resume_list);
398         spin_unlock_irqrestore(&pool->lock, flags);
399 }
400
401 /*
402  * This section of code contains the logic for processing a thin device's IO.
403  * Much of the code depends on pool object resources (lists, workqueues, etc)
404  * but most is exclusively called from the thin target rather than the thin-pool
405  * target.
406  */
407
408 static bool block_size_is_power_of_two(struct pool *pool)
409 {
410         return pool->sectors_per_block_shift >= 0;
411 }
412
413 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
414 {
415         struct pool *pool = tc->pool;
416         sector_t block_nr = bio->bi_sector;
417
418         if (block_size_is_power_of_two(pool))
419                 block_nr >>= pool->sectors_per_block_shift;
420         else
421                 (void) sector_div(block_nr, pool->sectors_per_block);
422
423         return block_nr;
424 }
425
426 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
427 {
428         struct pool *pool = tc->pool;
429         sector_t bi_sector = bio->bi_sector;
430
431         bio->bi_bdev = tc->pool_dev->bdev;
432         if (block_size_is_power_of_two(pool))
433                 bio->bi_sector = (block << pool->sectors_per_block_shift) |
434                                 (bi_sector & (pool->sectors_per_block - 1));
435         else
436                 bio->bi_sector = (block * pool->sectors_per_block) +
437                                  sector_div(bi_sector, pool->sectors_per_block);
438 }
439
440 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
441 {
442         bio->bi_bdev = tc->origin_dev->bdev;
443 }
444
445 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
446 {
447         return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
448                 dm_thin_changed_this_transaction(tc->td);
449 }
450
451 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
452 {
453         struct dm_thin_endio_hook *h;
454
455         if (bio->bi_rw & REQ_DISCARD)
456                 return;
457
458         h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
459         h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
460 }
461
462 static void issue(struct thin_c *tc, struct bio *bio)
463 {
464         struct pool *pool = tc->pool;
465         unsigned long flags;
466
467         if (!bio_triggers_commit(tc, bio)) {
468                 generic_make_request(bio);
469                 return;
470         }
471
472         /*
473          * Complete bio with an error if earlier I/O caused changes to
474          * the metadata that can't be committed e.g, due to I/O errors
475          * on the metadata device.
476          */
477         if (dm_thin_aborted_changes(tc->td)) {
478                 bio_io_error(bio);
479                 return;
480         }
481
482         /*
483          * Batch together any bios that trigger commits and then issue a
484          * single commit for them in process_deferred_bios().
485          */
486         spin_lock_irqsave(&pool->lock, flags);
487         bio_list_add(&pool->deferred_flush_bios, bio);
488         spin_unlock_irqrestore(&pool->lock, flags);
489 }
490
491 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
492 {
493         remap_to_origin(tc, bio);
494         issue(tc, bio);
495 }
496
497 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
498                             dm_block_t block)
499 {
500         remap(tc, bio, block);
501         issue(tc, bio);
502 }
503
504 /*----------------------------------------------------------------*/
505
506 /*
507  * Bio endio functions.
508  */
509 struct dm_thin_new_mapping {
510         struct list_head list;
511
512         unsigned quiesced:1;
513         unsigned prepared:1;
514         unsigned pass_discard:1;
515
516         struct thin_c *tc;
517         dm_block_t virt_block;
518         dm_block_t data_block;
519         struct dm_bio_prison_cell *cell, *cell2;
520         int err;
521
522         /*
523          * If the bio covers the whole area of a block then we can avoid
524          * zeroing or copying.  Instead this bio is hooked.  The bio will
525          * still be in the cell, so care has to be taken to avoid issuing
526          * the bio twice.
527          */
528         struct bio *bio;
529         bio_end_io_t *saved_bi_end_io;
530 };
531
532 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
533 {
534         struct pool *pool = m->tc->pool;
535
536         if (m->quiesced && m->prepared) {
537                 list_add(&m->list, &pool->prepared_mappings);
538                 wake_worker(pool);
539         }
540 }
541
542 static void copy_complete(int read_err, unsigned long write_err, void *context)
543 {
544         unsigned long flags;
545         struct dm_thin_new_mapping *m = context;
546         struct pool *pool = m->tc->pool;
547
548         m->err = read_err || write_err ? -EIO : 0;
549
550         spin_lock_irqsave(&pool->lock, flags);
551         m->prepared = 1;
552         __maybe_add_mapping(m);
553         spin_unlock_irqrestore(&pool->lock, flags);
554 }
555
556 static void overwrite_endio(struct bio *bio, int err)
557 {
558         unsigned long flags;
559         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
560         struct dm_thin_new_mapping *m = h->overwrite_mapping;
561         struct pool *pool = m->tc->pool;
562
563         m->err = err;
564
565         spin_lock_irqsave(&pool->lock, flags);
566         m->prepared = 1;
567         __maybe_add_mapping(m);
568         spin_unlock_irqrestore(&pool->lock, flags);
569 }
570
571 /*----------------------------------------------------------------*/
572
573 /*
574  * Workqueue.
575  */
576
577 /*
578  * Prepared mapping jobs.
579  */
580
581 /*
582  * This sends the bios in the cell back to the deferred_bios list.
583  */
584 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
585 {
586         struct pool *pool = tc->pool;
587         unsigned long flags;
588
589         spin_lock_irqsave(&pool->lock, flags);
590         cell_release(pool, cell, &pool->deferred_bios);
591         spin_unlock_irqrestore(&tc->pool->lock, flags);
592
593         wake_worker(pool);
594 }
595
596 /*
597  * Same as cell_defer above, except it omits the original holder of the cell.
598  */
599 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
600 {
601         struct pool *pool = tc->pool;
602         unsigned long flags;
603
604         spin_lock_irqsave(&pool->lock, flags);
605         cell_release_no_holder(pool, cell, &pool->deferred_bios);
606         spin_unlock_irqrestore(&pool->lock, flags);
607
608         wake_worker(pool);
609 }
610
611 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
612 {
613         if (m->bio)
614                 m->bio->bi_end_io = m->saved_bi_end_io;
615         cell_error(m->tc->pool, m->cell);
616         list_del(&m->list);
617         mempool_free(m, m->tc->pool->mapping_pool);
618 }
619
620 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
621 {
622         struct thin_c *tc = m->tc;
623         struct pool *pool = tc->pool;
624         struct bio *bio;
625         int r;
626
627         bio = m->bio;
628         if (bio)
629                 bio->bi_end_io = m->saved_bi_end_io;
630
631         if (m->err) {
632                 cell_error(pool, m->cell);
633                 goto out;
634         }
635
636         /*
637          * Commit the prepared block into the mapping btree.
638          * Any I/O for this block arriving after this point will get
639          * remapped to it directly.
640          */
641         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
642         if (r) {
643                 DMERR_LIMIT("dm_thin_insert_block() failed");
644                 cell_error(pool, m->cell);
645                 goto out;
646         }
647
648         /*
649          * Release any bios held while the block was being provisioned.
650          * If we are processing a write bio that completely covers the block,
651          * we already processed it so can ignore it now when processing
652          * the bios in the cell.
653          */
654         if (bio) {
655                 cell_defer_no_holder(tc, m->cell);
656                 bio_endio(bio, 0);
657         } else
658                 cell_defer(tc, m->cell);
659
660 out:
661         list_del(&m->list);
662         mempool_free(m, pool->mapping_pool);
663 }
664
665 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
666 {
667         struct thin_c *tc = m->tc;
668
669         bio_io_error(m->bio);
670         cell_defer_no_holder(tc, m->cell);
671         cell_defer_no_holder(tc, m->cell2);
672         mempool_free(m, tc->pool->mapping_pool);
673 }
674
675 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
676 {
677         struct thin_c *tc = m->tc;
678
679         inc_all_io_entry(tc->pool, m->bio);
680         cell_defer_no_holder(tc, m->cell);
681         cell_defer_no_holder(tc, m->cell2);
682
683         if (m->pass_discard)
684                 remap_and_issue(tc, m->bio, m->data_block);
685         else
686                 bio_endio(m->bio, 0);
687
688         mempool_free(m, tc->pool->mapping_pool);
689 }
690
691 static void process_prepared_discard(struct dm_thin_new_mapping *m)
692 {
693         int r;
694         struct thin_c *tc = m->tc;
695
696         r = dm_thin_remove_block(tc->td, m->virt_block);
697         if (r)
698                 DMERR_LIMIT("dm_thin_remove_block() failed");
699
700         process_prepared_discard_passdown(m);
701 }
702
703 static void process_prepared(struct pool *pool, struct list_head *head,
704                              process_mapping_fn *fn)
705 {
706         unsigned long flags;
707         struct list_head maps;
708         struct dm_thin_new_mapping *m, *tmp;
709
710         INIT_LIST_HEAD(&maps);
711         spin_lock_irqsave(&pool->lock, flags);
712         list_splice_init(head, &maps);
713         spin_unlock_irqrestore(&pool->lock, flags);
714
715         list_for_each_entry_safe(m, tmp, &maps, list)
716                 (*fn)(m);
717 }
718
719 /*
720  * Deferred bio jobs.
721  */
722 static int io_overlaps_block(struct pool *pool, struct bio *bio)
723 {
724         return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
725 }
726
727 static int io_overwrites_block(struct pool *pool, struct bio *bio)
728 {
729         return (bio_data_dir(bio) == WRITE) &&
730                 io_overlaps_block(pool, bio);
731 }
732
733 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
734                                bio_end_io_t *fn)
735 {
736         *save = bio->bi_end_io;
737         bio->bi_end_io = fn;
738 }
739
740 static int ensure_next_mapping(struct pool *pool)
741 {
742         if (pool->next_mapping)
743                 return 0;
744
745         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
746
747         return pool->next_mapping ? 0 : -ENOMEM;
748 }
749
750 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
751 {
752         struct dm_thin_new_mapping *r = pool->next_mapping;
753
754         BUG_ON(!pool->next_mapping);
755
756         pool->next_mapping = NULL;
757
758         return r;
759 }
760
761 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
762                           struct dm_dev *origin, dm_block_t data_origin,
763                           dm_block_t data_dest,
764                           struct dm_bio_prison_cell *cell, struct bio *bio)
765 {
766         int r;
767         struct pool *pool = tc->pool;
768         struct dm_thin_new_mapping *m = get_next_mapping(pool);
769
770         INIT_LIST_HEAD(&m->list);
771         m->quiesced = 0;
772         m->prepared = 0;
773         m->tc = tc;
774         m->virt_block = virt_block;
775         m->data_block = data_dest;
776         m->cell = cell;
777         m->err = 0;
778         m->bio = NULL;
779
780         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
781                 m->quiesced = 1;
782
783         /*
784          * IO to pool_dev remaps to the pool target's data_dev.
785          *
786          * If the whole block of data is being overwritten, we can issue the
787          * bio immediately. Otherwise we use kcopyd to clone the data first.
788          */
789         if (io_overwrites_block(pool, bio)) {
790                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
791
792                 h->overwrite_mapping = m;
793                 m->bio = bio;
794                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
795                 inc_all_io_entry(pool, bio);
796                 remap_and_issue(tc, bio, data_dest);
797         } else {
798                 struct dm_io_region from, to;
799
800                 from.bdev = origin->bdev;
801                 from.sector = data_origin * pool->sectors_per_block;
802                 from.count = pool->sectors_per_block;
803
804                 to.bdev = tc->pool_dev->bdev;
805                 to.sector = data_dest * pool->sectors_per_block;
806                 to.count = pool->sectors_per_block;
807
808                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
809                                    0, copy_complete, m);
810                 if (r < 0) {
811                         mempool_free(m, pool->mapping_pool);
812                         DMERR_LIMIT("dm_kcopyd_copy() failed");
813                         cell_error(pool, cell);
814                 }
815         }
816 }
817
818 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
819                                    dm_block_t data_origin, dm_block_t data_dest,
820                                    struct dm_bio_prison_cell *cell, struct bio *bio)
821 {
822         schedule_copy(tc, virt_block, tc->pool_dev,
823                       data_origin, data_dest, cell, bio);
824 }
825
826 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
827                                    dm_block_t data_dest,
828                                    struct dm_bio_prison_cell *cell, struct bio *bio)
829 {
830         schedule_copy(tc, virt_block, tc->origin_dev,
831                       virt_block, data_dest, cell, bio);
832 }
833
834 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
835                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
836                           struct bio *bio)
837 {
838         struct pool *pool = tc->pool;
839         struct dm_thin_new_mapping *m = get_next_mapping(pool);
840
841         INIT_LIST_HEAD(&m->list);
842         m->quiesced = 1;
843         m->prepared = 0;
844         m->tc = tc;
845         m->virt_block = virt_block;
846         m->data_block = data_block;
847         m->cell = cell;
848         m->err = 0;
849         m->bio = NULL;
850
851         /*
852          * If the whole block of data is being overwritten or we are not
853          * zeroing pre-existing data, we can issue the bio immediately.
854          * Otherwise we use kcopyd to zero the data first.
855          */
856         if (!pool->pf.zero_new_blocks)
857                 process_prepared_mapping(m);
858
859         else if (io_overwrites_block(pool, bio)) {
860                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
861
862                 h->overwrite_mapping = m;
863                 m->bio = bio;
864                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
865                 inc_all_io_entry(pool, bio);
866                 remap_and_issue(tc, bio, data_block);
867         } else {
868                 int r;
869                 struct dm_io_region to;
870
871                 to.bdev = tc->pool_dev->bdev;
872                 to.sector = data_block * pool->sectors_per_block;
873                 to.count = pool->sectors_per_block;
874
875                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
876                 if (r < 0) {
877                         mempool_free(m, pool->mapping_pool);
878                         DMERR_LIMIT("dm_kcopyd_zero() failed");
879                         cell_error(pool, cell);
880                 }
881         }
882 }
883
884 static int commit(struct pool *pool)
885 {
886         int r;
887
888         r = dm_pool_commit_metadata(pool->pmd);
889         if (r)
890                 DMERR_LIMIT("commit failed: error = %d", r);
891
892         return r;
893 }
894
895 /*
896  * A non-zero return indicates read_only or fail_io mode.
897  * Many callers don't care about the return value.
898  */
899 static int commit_or_fallback(struct pool *pool)
900 {
901         int r;
902
903         if (get_pool_mode(pool) != PM_WRITE)
904                 return -EINVAL;
905
906         r = commit(pool);
907         if (r)
908                 set_pool_mode(pool, PM_READ_ONLY);
909
910         return r;
911 }
912
913 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
914 {
915         int r;
916         dm_block_t free_blocks;
917         unsigned long flags;
918         struct pool *pool = tc->pool;
919
920         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
921         if (r)
922                 return r;
923
924         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
925                 DMWARN("%s: reached low water mark for data device: sending event.",
926                        dm_device_name(pool->pool_md));
927                 spin_lock_irqsave(&pool->lock, flags);
928                 pool->low_water_triggered = 1;
929                 spin_unlock_irqrestore(&pool->lock, flags);
930                 dm_table_event(pool->ti->table);
931         }
932
933         if (!free_blocks) {
934                 if (pool->no_free_space)
935                         return -ENOSPC;
936                 else {
937                         /*
938                          * Try to commit to see if that will free up some
939                          * more space.
940                          */
941                         (void) commit_or_fallback(pool);
942
943                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
944                         if (r)
945                                 return r;
946
947                         /*
948                          * If we still have no space we set a flag to avoid
949                          * doing all this checking and return -ENOSPC.
950                          */
951                         if (!free_blocks) {
952                                 DMWARN("%s: no free space available.",
953                                        dm_device_name(pool->pool_md));
954                                 spin_lock_irqsave(&pool->lock, flags);
955                                 pool->no_free_space = 1;
956                                 spin_unlock_irqrestore(&pool->lock, flags);
957                                 return -ENOSPC;
958                         }
959                 }
960         }
961
962         r = dm_pool_alloc_data_block(pool->pmd, result);
963         if (r)
964                 return r;
965
966         return 0;
967 }
968
969 /*
970  * If we have run out of space, queue bios until the device is
971  * resumed, presumably after having been reloaded with more space.
972  */
973 static void retry_on_resume(struct bio *bio)
974 {
975         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
976         struct thin_c *tc = h->tc;
977         struct pool *pool = tc->pool;
978         unsigned long flags;
979
980         spin_lock_irqsave(&pool->lock, flags);
981         bio_list_add(&pool->retry_on_resume_list, bio);
982         spin_unlock_irqrestore(&pool->lock, flags);
983 }
984
985 static void no_space(struct pool *pool, struct dm_bio_prison_cell *cell)
986 {
987         struct bio *bio;
988         struct bio_list bios;
989
990         bio_list_init(&bios);
991         cell_release(pool, cell, &bios);
992
993         while ((bio = bio_list_pop(&bios)))
994                 retry_on_resume(bio);
995 }
996
997 static void process_discard(struct thin_c *tc, struct bio *bio)
998 {
999         int r;
1000         unsigned long flags;
1001         struct pool *pool = tc->pool;
1002         struct dm_bio_prison_cell *cell, *cell2;
1003         struct dm_cell_key key, key2;
1004         dm_block_t block = get_bio_block(tc, bio);
1005         struct dm_thin_lookup_result lookup_result;
1006         struct dm_thin_new_mapping *m;
1007
1008         build_virtual_key(tc->td, block, &key);
1009         if (bio_detain(tc->pool, &key, bio, &cell))
1010                 return;
1011
1012         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1013         switch (r) {
1014         case 0:
1015                 /*
1016                  * Check nobody is fiddling with this pool block.  This can
1017                  * happen if someone's in the process of breaking sharing
1018                  * on this block.
1019                  */
1020                 build_data_key(tc->td, lookup_result.block, &key2);
1021                 if (bio_detain(tc->pool, &key2, bio, &cell2)) {
1022                         cell_defer_no_holder(tc, cell);
1023                         break;
1024                 }
1025
1026                 if (io_overlaps_block(pool, bio)) {
1027                         /*
1028                          * IO may still be going to the destination block.  We must
1029                          * quiesce before we can do the removal.
1030                          */
1031                         m = get_next_mapping(pool);
1032                         m->tc = tc;
1033                         m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
1034                         m->virt_block = block;
1035                         m->data_block = lookup_result.block;
1036                         m->cell = cell;
1037                         m->cell2 = cell2;
1038                         m->err = 0;
1039                         m->bio = bio;
1040
1041                         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
1042                                 spin_lock_irqsave(&pool->lock, flags);
1043                                 list_add(&m->list, &pool->prepared_discards);
1044                                 spin_unlock_irqrestore(&pool->lock, flags);
1045                                 wake_worker(pool);
1046                         }
1047                 } else {
1048                         inc_all_io_entry(pool, bio);
1049                         cell_defer_no_holder(tc, cell);
1050                         cell_defer_no_holder(tc, cell2);
1051
1052                         /*
1053                          * The DM core makes sure that the discard doesn't span
1054                          * a block boundary.  So we submit the discard of a
1055                          * partial block appropriately.
1056                          */
1057                         if ((!lookup_result.shared) && pool->pf.discard_passdown)
1058                                 remap_and_issue(tc, bio, lookup_result.block);
1059                         else
1060                                 bio_endio(bio, 0);
1061                 }
1062                 break;
1063
1064         case -ENODATA:
1065                 /*
1066                  * It isn't provisioned, just forget it.
1067                  */
1068                 cell_defer_no_holder(tc, cell);
1069                 bio_endio(bio, 0);
1070                 break;
1071
1072         default:
1073                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1074                             __func__, r);
1075                 cell_defer_no_holder(tc, cell);
1076                 bio_io_error(bio);
1077                 break;
1078         }
1079 }
1080
1081 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1082                           struct dm_cell_key *key,
1083                           struct dm_thin_lookup_result *lookup_result,
1084                           struct dm_bio_prison_cell *cell)
1085 {
1086         int r;
1087         dm_block_t data_block;
1088
1089         r = alloc_data_block(tc, &data_block);
1090         switch (r) {
1091         case 0:
1092                 schedule_internal_copy(tc, block, lookup_result->block,
1093                                        data_block, cell, bio);
1094                 break;
1095
1096         case -ENOSPC:
1097                 no_space(tc->pool, cell);
1098                 break;
1099
1100         default:
1101                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1102                             __func__, r);
1103                 cell_error(tc->pool, cell);
1104                 break;
1105         }
1106 }
1107
1108 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1109                                dm_block_t block,
1110                                struct dm_thin_lookup_result *lookup_result)
1111 {
1112         struct dm_bio_prison_cell *cell;
1113         struct pool *pool = tc->pool;
1114         struct dm_cell_key key;
1115
1116         /*
1117          * If cell is already occupied, then sharing is already in the process
1118          * of being broken so we have nothing further to do here.
1119          */
1120         build_data_key(tc->td, lookup_result->block, &key);
1121         if (bio_detain(pool, &key, bio, &cell))
1122                 return;
1123
1124         if (bio_data_dir(bio) == WRITE && bio->bi_size)
1125                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1126         else {
1127                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1128
1129                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1130                 inc_all_io_entry(pool, bio);
1131                 cell_defer_no_holder(tc, cell);
1132
1133                 remap_and_issue(tc, bio, lookup_result->block);
1134         }
1135 }
1136
1137 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1138                             struct dm_bio_prison_cell *cell)
1139 {
1140         int r;
1141         dm_block_t data_block;
1142         struct pool *pool = tc->pool;
1143
1144         /*
1145          * Remap empty bios (flushes) immediately, without provisioning.
1146          */
1147         if (!bio->bi_size) {
1148                 inc_all_io_entry(pool, bio);
1149                 cell_defer_no_holder(tc, cell);
1150
1151                 remap_and_issue(tc, bio, 0);
1152                 return;
1153         }
1154
1155         /*
1156          * Fill read bios with zeroes and complete them immediately.
1157          */
1158         if (bio_data_dir(bio) == READ) {
1159                 zero_fill_bio(bio);
1160                 cell_defer_no_holder(tc, cell);
1161                 bio_endio(bio, 0);
1162                 return;
1163         }
1164
1165         r = alloc_data_block(tc, &data_block);
1166         switch (r) {
1167         case 0:
1168                 if (tc->origin_dev)
1169                         schedule_external_copy(tc, block, data_block, cell, bio);
1170                 else
1171                         schedule_zero(tc, block, data_block, cell, bio);
1172                 break;
1173
1174         case -ENOSPC:
1175                 no_space(pool, cell);
1176                 break;
1177
1178         default:
1179                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1180                             __func__, r);
1181                 set_pool_mode(pool, PM_READ_ONLY);
1182                 cell_error(pool, cell);
1183                 break;
1184         }
1185 }
1186
1187 static void process_bio(struct thin_c *tc, struct bio *bio)
1188 {
1189         int r;
1190         struct pool *pool = tc->pool;
1191         dm_block_t block = get_bio_block(tc, bio);
1192         struct dm_bio_prison_cell *cell;
1193         struct dm_cell_key key;
1194         struct dm_thin_lookup_result lookup_result;
1195
1196         /*
1197          * If cell is already occupied, then the block is already
1198          * being provisioned so we have nothing further to do here.
1199          */
1200         build_virtual_key(tc->td, block, &key);
1201         if (bio_detain(pool, &key, bio, &cell))
1202                 return;
1203
1204         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1205         switch (r) {
1206         case 0:
1207                 if (lookup_result.shared) {
1208                         process_shared_bio(tc, bio, block, &lookup_result);
1209                         cell_defer_no_holder(tc, cell); /* FIXME: pass this cell into process_shared? */
1210                 } else {
1211                         inc_all_io_entry(pool, bio);
1212                         cell_defer_no_holder(tc, cell);
1213
1214                         remap_and_issue(tc, bio, lookup_result.block);
1215                 }
1216                 break;
1217
1218         case -ENODATA:
1219                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1220                         inc_all_io_entry(pool, bio);
1221                         cell_defer_no_holder(tc, cell);
1222
1223                         remap_to_origin_and_issue(tc, bio);
1224                 } else
1225                         provision_block(tc, bio, block, cell);
1226                 break;
1227
1228         default:
1229                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1230                             __func__, r);
1231                 cell_defer_no_holder(tc, cell);
1232                 bio_io_error(bio);
1233                 break;
1234         }
1235 }
1236
1237 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1238 {
1239         int r;
1240         int rw = bio_data_dir(bio);
1241         dm_block_t block = get_bio_block(tc, bio);
1242         struct dm_thin_lookup_result lookup_result;
1243
1244         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1245         switch (r) {
1246         case 0:
1247                 if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1248                         bio_io_error(bio);
1249                 else {
1250                         inc_all_io_entry(tc->pool, bio);
1251                         remap_and_issue(tc, bio, lookup_result.block);
1252                 }
1253                 break;
1254
1255         case -ENODATA:
1256                 if (rw != READ) {
1257                         bio_io_error(bio);
1258                         break;
1259                 }
1260
1261                 if (tc->origin_dev) {
1262                         inc_all_io_entry(tc->pool, bio);
1263                         remap_to_origin_and_issue(tc, bio);
1264                         break;
1265                 }
1266
1267                 zero_fill_bio(bio);
1268                 bio_endio(bio, 0);
1269                 break;
1270
1271         default:
1272                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1273                             __func__, r);
1274                 bio_io_error(bio);
1275                 break;
1276         }
1277 }
1278
1279 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1280 {
1281         bio_io_error(bio);
1282 }
1283
1284 /*
1285  * FIXME: should we also commit due to size of transaction, measured in
1286  * metadata blocks?
1287  */
1288 static int need_commit_due_to_time(struct pool *pool)
1289 {
1290         return jiffies < pool->last_commit_jiffies ||
1291                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1292 }
1293
1294 static void process_deferred_bios(struct pool *pool)
1295 {
1296         unsigned long flags;
1297         struct bio *bio;
1298         struct bio_list bios;
1299
1300         bio_list_init(&bios);
1301
1302         spin_lock_irqsave(&pool->lock, flags);
1303         bio_list_merge(&bios, &pool->deferred_bios);
1304         bio_list_init(&pool->deferred_bios);
1305         spin_unlock_irqrestore(&pool->lock, flags);
1306
1307         while ((bio = bio_list_pop(&bios))) {
1308                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1309                 struct thin_c *tc = h->tc;
1310
1311                 /*
1312                  * If we've got no free new_mapping structs, and processing
1313                  * this bio might require one, we pause until there are some
1314                  * prepared mappings to process.
1315                  */
1316                 if (ensure_next_mapping(pool)) {
1317                         spin_lock_irqsave(&pool->lock, flags);
1318                         bio_list_merge(&pool->deferred_bios, &bios);
1319                         spin_unlock_irqrestore(&pool->lock, flags);
1320
1321                         break;
1322                 }
1323
1324                 if (bio->bi_rw & REQ_DISCARD)
1325                         pool->process_discard(tc, bio);
1326                 else
1327                         pool->process_bio(tc, bio);
1328         }
1329
1330         /*
1331          * If there are any deferred flush bios, we must commit
1332          * the metadata before issuing them.
1333          */
1334         bio_list_init(&bios);
1335         spin_lock_irqsave(&pool->lock, flags);
1336         bio_list_merge(&bios, &pool->deferred_flush_bios);
1337         bio_list_init(&pool->deferred_flush_bios);
1338         spin_unlock_irqrestore(&pool->lock, flags);
1339
1340         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1341                 return;
1342
1343         if (commit_or_fallback(pool)) {
1344                 while ((bio = bio_list_pop(&bios)))
1345                         bio_io_error(bio);
1346                 return;
1347         }
1348         pool->last_commit_jiffies = jiffies;
1349
1350         while ((bio = bio_list_pop(&bios)))
1351                 generic_make_request(bio);
1352 }
1353
1354 static void do_worker(struct work_struct *ws)
1355 {
1356         struct pool *pool = container_of(ws, struct pool, worker);
1357
1358         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1359         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1360         process_deferred_bios(pool);
1361 }
1362
1363 /*
1364  * We want to commit periodically so that not too much
1365  * unwritten data builds up.
1366  */
1367 static void do_waker(struct work_struct *ws)
1368 {
1369         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1370         wake_worker(pool);
1371         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1372 }
1373
1374 /*----------------------------------------------------------------*/
1375
1376 static enum pool_mode get_pool_mode(struct pool *pool)
1377 {
1378         return pool->pf.mode;
1379 }
1380
1381 static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1382 {
1383         int r;
1384
1385         pool->pf.mode = mode;
1386
1387         switch (mode) {
1388         case PM_FAIL:
1389                 DMERR("switching pool to failure mode");
1390                 pool->process_bio = process_bio_fail;
1391                 pool->process_discard = process_bio_fail;
1392                 pool->process_prepared_mapping = process_prepared_mapping_fail;
1393                 pool->process_prepared_discard = process_prepared_discard_fail;
1394                 break;
1395
1396         case PM_READ_ONLY:
1397                 DMERR("switching pool to read-only mode");
1398                 r = dm_pool_abort_metadata(pool->pmd);
1399                 if (r) {
1400                         DMERR("aborting transaction failed");
1401                         set_pool_mode(pool, PM_FAIL);
1402                 } else {
1403                         dm_pool_metadata_read_only(pool->pmd);
1404                         pool->process_bio = process_bio_read_only;
1405                         pool->process_discard = process_discard;
1406                         pool->process_prepared_mapping = process_prepared_mapping_fail;
1407                         pool->process_prepared_discard = process_prepared_discard_passdown;
1408                 }
1409                 break;
1410
1411         case PM_WRITE:
1412                 pool->process_bio = process_bio;
1413                 pool->process_discard = process_discard;
1414                 pool->process_prepared_mapping = process_prepared_mapping;
1415                 pool->process_prepared_discard = process_prepared_discard;
1416                 break;
1417         }
1418 }
1419
1420 /*----------------------------------------------------------------*/
1421
1422 /*
1423  * Mapping functions.
1424  */
1425
1426 /*
1427  * Called only while mapping a thin bio to hand it over to the workqueue.
1428  */
1429 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1430 {
1431         unsigned long flags;
1432         struct pool *pool = tc->pool;
1433
1434         spin_lock_irqsave(&pool->lock, flags);
1435         bio_list_add(&pool->deferred_bios, bio);
1436         spin_unlock_irqrestore(&pool->lock, flags);
1437
1438         wake_worker(pool);
1439 }
1440
1441 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1442 {
1443         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1444
1445         h->tc = tc;
1446         h->shared_read_entry = NULL;
1447         h->all_io_entry = NULL;
1448         h->overwrite_mapping = NULL;
1449 }
1450
1451 /*
1452  * Non-blocking function called from the thin target's map function.
1453  */
1454 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1455 {
1456         int r;
1457         struct thin_c *tc = ti->private;
1458         dm_block_t block = get_bio_block(tc, bio);
1459         struct dm_thin_device *td = tc->td;
1460         struct dm_thin_lookup_result result;
1461         struct dm_bio_prison_cell cell1, cell2;
1462         struct dm_bio_prison_cell *cell_result;
1463         struct dm_cell_key key;
1464
1465         thin_hook_bio(tc, bio);
1466
1467         if (get_pool_mode(tc->pool) == PM_FAIL) {
1468                 bio_io_error(bio);
1469                 return DM_MAPIO_SUBMITTED;
1470         }
1471
1472         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1473                 thin_defer_bio(tc, bio);
1474                 return DM_MAPIO_SUBMITTED;
1475         }
1476
1477         r = dm_thin_find_block(td, block, 0, &result);
1478
1479         /*
1480          * Note that we defer readahead too.
1481          */
1482         switch (r) {
1483         case 0:
1484                 if (unlikely(result.shared)) {
1485                         /*
1486                          * We have a race condition here between the
1487                          * result.shared value returned by the lookup and
1488                          * snapshot creation, which may cause new
1489                          * sharing.
1490                          *
1491                          * To avoid this always quiesce the origin before
1492                          * taking the snap.  You want to do this anyway to
1493                          * ensure a consistent application view
1494                          * (i.e. lockfs).
1495                          *
1496                          * More distant ancestors are irrelevant. The
1497                          * shared flag will be set in their case.
1498                          */
1499                         thin_defer_bio(tc, bio);
1500                         return DM_MAPIO_SUBMITTED;
1501                 }
1502
1503                 build_virtual_key(tc->td, block, &key);
1504                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1, &cell_result))
1505                         return DM_MAPIO_SUBMITTED;
1506
1507                 build_data_key(tc->td, result.block, &key);
1508                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2, &cell_result)) {
1509                         cell_defer_no_holder_no_free(tc, &cell1);
1510                         return DM_MAPIO_SUBMITTED;
1511                 }
1512
1513                 inc_all_io_entry(tc->pool, bio);
1514                 cell_defer_no_holder_no_free(tc, &cell2);
1515                 cell_defer_no_holder_no_free(tc, &cell1);
1516
1517                 remap(tc, bio, result.block);
1518                 return DM_MAPIO_REMAPPED;
1519
1520         case -ENODATA:
1521                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1522                         /*
1523                          * This block isn't provisioned, and we have no way
1524                          * of doing so.  Just error it.
1525                          */
1526                         bio_io_error(bio);
1527                         return DM_MAPIO_SUBMITTED;
1528                 }
1529                 /* fall through */
1530
1531         case -EWOULDBLOCK:
1532                 /*
1533                  * In future, the failed dm_thin_find_block above could
1534                  * provide the hint to load the metadata into cache.
1535                  */
1536                 thin_defer_bio(tc, bio);
1537                 return DM_MAPIO_SUBMITTED;
1538
1539         default:
1540                 /*
1541                  * Must always call bio_io_error on failure.
1542                  * dm_thin_find_block can fail with -EINVAL if the
1543                  * pool is switched to fail-io mode.
1544                  */
1545                 bio_io_error(bio);
1546                 return DM_MAPIO_SUBMITTED;
1547         }
1548 }
1549
1550 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1551 {
1552         int r;
1553         unsigned long flags;
1554         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1555
1556         spin_lock_irqsave(&pt->pool->lock, flags);
1557         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1558         spin_unlock_irqrestore(&pt->pool->lock, flags);
1559
1560         if (!r) {
1561                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1562                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1563         }
1564
1565         return r;
1566 }
1567
1568 static void __requeue_bios(struct pool *pool)
1569 {
1570         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1571         bio_list_init(&pool->retry_on_resume_list);
1572 }
1573
1574 /*----------------------------------------------------------------
1575  * Binding of control targets to a pool object
1576  *--------------------------------------------------------------*/
1577 static bool data_dev_supports_discard(struct pool_c *pt)
1578 {
1579         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1580
1581         return q && blk_queue_discard(q);
1582 }
1583
1584 static bool is_factor(sector_t block_size, uint32_t n)
1585 {
1586         return !sector_div(block_size, n);
1587 }
1588
1589 /*
1590  * If discard_passdown was enabled verify that the data device
1591  * supports discards.  Disable discard_passdown if not.
1592  */
1593 static void disable_passdown_if_not_supported(struct pool_c *pt)
1594 {
1595         struct pool *pool = pt->pool;
1596         struct block_device *data_bdev = pt->data_dev->bdev;
1597         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1598         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1599         const char *reason = NULL;
1600         char buf[BDEVNAME_SIZE];
1601
1602         if (!pt->adjusted_pf.discard_passdown)
1603                 return;
1604
1605         if (!data_dev_supports_discard(pt))
1606                 reason = "discard unsupported";
1607
1608         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1609                 reason = "max discard sectors smaller than a block";
1610
1611         else if (data_limits->discard_granularity > block_size)
1612                 reason = "discard granularity larger than a block";
1613
1614         else if (!is_factor(block_size, data_limits->discard_granularity))
1615                 reason = "discard granularity not a factor of block size";
1616
1617         if (reason) {
1618                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1619                 pt->adjusted_pf.discard_passdown = false;
1620         }
1621 }
1622
1623 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1624 {
1625         struct pool_c *pt = ti->private;
1626
1627         /*
1628          * We want to make sure that degraded pools are never upgraded.
1629          */
1630         enum pool_mode old_mode = pool->pf.mode;
1631         enum pool_mode new_mode = pt->adjusted_pf.mode;
1632
1633         if (old_mode > new_mode)
1634                 new_mode = old_mode;
1635
1636         pool->ti = ti;
1637         pool->low_water_blocks = pt->low_water_blocks;
1638         pool->pf = pt->adjusted_pf;
1639
1640         set_pool_mode(pool, new_mode);
1641
1642         return 0;
1643 }
1644
1645 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1646 {
1647         if (pool->ti == ti)
1648                 pool->ti = NULL;
1649 }
1650
1651 /*----------------------------------------------------------------
1652  * Pool creation
1653  *--------------------------------------------------------------*/
1654 /* Initialize pool features. */
1655 static void pool_features_init(struct pool_features *pf)
1656 {
1657         pf->mode = PM_WRITE;
1658         pf->zero_new_blocks = true;
1659         pf->discard_enabled = true;
1660         pf->discard_passdown = true;
1661 }
1662
1663 static void __pool_destroy(struct pool *pool)
1664 {
1665         __pool_table_remove(pool);
1666
1667         if (dm_pool_metadata_close(pool->pmd) < 0)
1668                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1669
1670         dm_bio_prison_destroy(pool->prison);
1671         dm_kcopyd_client_destroy(pool->copier);
1672
1673         if (pool->wq)
1674                 destroy_workqueue(pool->wq);
1675
1676         if (pool->next_mapping)
1677                 mempool_free(pool->next_mapping, pool->mapping_pool);
1678         mempool_destroy(pool->mapping_pool);
1679         dm_deferred_set_destroy(pool->shared_read_ds);
1680         dm_deferred_set_destroy(pool->all_io_ds);
1681         kfree(pool);
1682 }
1683
1684 static struct kmem_cache *_new_mapping_cache;
1685
1686 static struct pool *pool_create(struct mapped_device *pool_md,
1687                                 struct block_device *metadata_dev,
1688                                 unsigned long block_size,
1689                                 int read_only, char **error)
1690 {
1691         int r;
1692         void *err_p;
1693         struct pool *pool;
1694         struct dm_pool_metadata *pmd;
1695         bool format_device = read_only ? false : true;
1696
1697         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1698         if (IS_ERR(pmd)) {
1699                 *error = "Error creating metadata object";
1700                 return (struct pool *)pmd;
1701         }
1702
1703         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1704         if (!pool) {
1705                 *error = "Error allocating memory for pool";
1706                 err_p = ERR_PTR(-ENOMEM);
1707                 goto bad_pool;
1708         }
1709
1710         pool->pmd = pmd;
1711         pool->sectors_per_block = block_size;
1712         if (block_size & (block_size - 1))
1713                 pool->sectors_per_block_shift = -1;
1714         else
1715                 pool->sectors_per_block_shift = __ffs(block_size);
1716         pool->low_water_blocks = 0;
1717         pool_features_init(&pool->pf);
1718         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1719         if (!pool->prison) {
1720                 *error = "Error creating pool's bio prison";
1721                 err_p = ERR_PTR(-ENOMEM);
1722                 goto bad_prison;
1723         }
1724
1725         pool->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1726         if (IS_ERR(pool->copier)) {
1727                 r = PTR_ERR(pool->copier);
1728                 *error = "Error creating pool's kcopyd client";
1729                 err_p = ERR_PTR(r);
1730                 goto bad_kcopyd_client;
1731         }
1732
1733         /*
1734          * Create singlethreaded workqueue that will service all devices
1735          * that use this metadata.
1736          */
1737         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1738         if (!pool->wq) {
1739                 *error = "Error creating pool's workqueue";
1740                 err_p = ERR_PTR(-ENOMEM);
1741                 goto bad_wq;
1742         }
1743
1744         INIT_WORK(&pool->worker, do_worker);
1745         INIT_DELAYED_WORK(&pool->waker, do_waker);
1746         spin_lock_init(&pool->lock);
1747         bio_list_init(&pool->deferred_bios);
1748         bio_list_init(&pool->deferred_flush_bios);
1749         INIT_LIST_HEAD(&pool->prepared_mappings);
1750         INIT_LIST_HEAD(&pool->prepared_discards);
1751         pool->low_water_triggered = 0;
1752         pool->no_free_space = 0;
1753         bio_list_init(&pool->retry_on_resume_list);
1754
1755         pool->shared_read_ds = dm_deferred_set_create();
1756         if (!pool->shared_read_ds) {
1757                 *error = "Error creating pool's shared read deferred set";
1758                 err_p = ERR_PTR(-ENOMEM);
1759                 goto bad_shared_read_ds;
1760         }
1761
1762         pool->all_io_ds = dm_deferred_set_create();
1763         if (!pool->all_io_ds) {
1764                 *error = "Error creating pool's all io deferred set";
1765                 err_p = ERR_PTR(-ENOMEM);
1766                 goto bad_all_io_ds;
1767         }
1768
1769         pool->next_mapping = NULL;
1770         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1771                                                       _new_mapping_cache);
1772         if (!pool->mapping_pool) {
1773                 *error = "Error creating pool's mapping mempool";
1774                 err_p = ERR_PTR(-ENOMEM);
1775                 goto bad_mapping_pool;
1776         }
1777
1778         pool->ref_count = 1;
1779         pool->last_commit_jiffies = jiffies;
1780         pool->pool_md = pool_md;
1781         pool->md_dev = metadata_dev;
1782         __pool_table_insert(pool);
1783
1784         return pool;
1785
1786 bad_mapping_pool:
1787         dm_deferred_set_destroy(pool->all_io_ds);
1788 bad_all_io_ds:
1789         dm_deferred_set_destroy(pool->shared_read_ds);
1790 bad_shared_read_ds:
1791         destroy_workqueue(pool->wq);
1792 bad_wq:
1793         dm_kcopyd_client_destroy(pool->copier);
1794 bad_kcopyd_client:
1795         dm_bio_prison_destroy(pool->prison);
1796 bad_prison:
1797         kfree(pool);
1798 bad_pool:
1799         if (dm_pool_metadata_close(pmd))
1800                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1801
1802         return err_p;
1803 }
1804
1805 static void __pool_inc(struct pool *pool)
1806 {
1807         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1808         pool->ref_count++;
1809 }
1810
1811 static void __pool_dec(struct pool *pool)
1812 {
1813         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1814         BUG_ON(!pool->ref_count);
1815         if (!--pool->ref_count)
1816                 __pool_destroy(pool);
1817 }
1818
1819 static struct pool *__pool_find(struct mapped_device *pool_md,
1820                                 struct block_device *metadata_dev,
1821                                 unsigned long block_size, int read_only,
1822                                 char **error, int *created)
1823 {
1824         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1825
1826         if (pool) {
1827                 if (pool->pool_md != pool_md) {
1828                         *error = "metadata device already in use by a pool";
1829                         return ERR_PTR(-EBUSY);
1830                 }
1831                 __pool_inc(pool);
1832
1833         } else {
1834                 pool = __pool_table_lookup(pool_md);
1835                 if (pool) {
1836                         if (pool->md_dev != metadata_dev) {
1837                                 *error = "different pool cannot replace a pool";
1838                                 return ERR_PTR(-EINVAL);
1839                         }
1840                         __pool_inc(pool);
1841
1842                 } else {
1843                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1844                         *created = 1;
1845                 }
1846         }
1847
1848         return pool;
1849 }
1850
1851 /*----------------------------------------------------------------
1852  * Pool target methods
1853  *--------------------------------------------------------------*/
1854 static void pool_dtr(struct dm_target *ti)
1855 {
1856         struct pool_c *pt = ti->private;
1857
1858         mutex_lock(&dm_thin_pool_table.mutex);
1859
1860         unbind_control_target(pt->pool, ti);
1861         __pool_dec(pt->pool);
1862         dm_put_device(ti, pt->metadata_dev);
1863         dm_put_device(ti, pt->data_dev);
1864         kfree(pt);
1865
1866         mutex_unlock(&dm_thin_pool_table.mutex);
1867 }
1868
1869 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1870                                struct dm_target *ti)
1871 {
1872         int r;
1873         unsigned argc;
1874         const char *arg_name;
1875
1876         static struct dm_arg _args[] = {
1877                 {0, 3, "Invalid number of pool feature arguments"},
1878         };
1879
1880         /*
1881          * No feature arguments supplied.
1882          */
1883         if (!as->argc)
1884                 return 0;
1885
1886         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1887         if (r)
1888                 return -EINVAL;
1889
1890         while (argc && !r) {
1891                 arg_name = dm_shift_arg(as);
1892                 argc--;
1893
1894                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1895                         pf->zero_new_blocks = false;
1896
1897                 else if (!strcasecmp(arg_name, "ignore_discard"))
1898                         pf->discard_enabled = false;
1899
1900                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1901                         pf->discard_passdown = false;
1902
1903                 else if (!strcasecmp(arg_name, "read_only"))
1904                         pf->mode = PM_READ_ONLY;
1905
1906                 else {
1907                         ti->error = "Unrecognised pool feature requested";
1908                         r = -EINVAL;
1909                         break;
1910                 }
1911         }
1912
1913         return r;
1914 }
1915
1916 static void metadata_low_callback(void *context)
1917 {
1918         struct pool *pool = context;
1919
1920         DMWARN("%s: reached low water mark for metadata device: sending event.",
1921                dm_device_name(pool->pool_md));
1922
1923         dm_table_event(pool->ti->table);
1924 }
1925
1926 static sector_t get_metadata_dev_size(struct block_device *bdev)
1927 {
1928         sector_t metadata_dev_size = i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
1929         char buffer[BDEVNAME_SIZE];
1930
1931         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING) {
1932                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1933                        bdevname(bdev, buffer), THIN_METADATA_MAX_SECTORS);
1934                 metadata_dev_size = THIN_METADATA_MAX_SECTORS_WARNING;
1935         }
1936
1937         return metadata_dev_size;
1938 }
1939
1940 static dm_block_t get_metadata_dev_size_in_blocks(struct block_device *bdev)
1941 {
1942         sector_t metadata_dev_size = get_metadata_dev_size(bdev);
1943
1944         sector_div(metadata_dev_size, THIN_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
1945
1946         return metadata_dev_size;
1947 }
1948
1949 /*
1950  * When a metadata threshold is crossed a dm event is triggered, and
1951  * userland should respond by growing the metadata device.  We could let
1952  * userland set the threshold, like we do with the data threshold, but I'm
1953  * not sure they know enough to do this well.
1954  */
1955 static dm_block_t calc_metadata_threshold(struct pool_c *pt)
1956 {
1957         /*
1958          * 4M is ample for all ops with the possible exception of thin
1959          * device deletion which is harmless if it fails (just retry the
1960          * delete after you've grown the device).
1961          */
1962         dm_block_t quarter = get_metadata_dev_size_in_blocks(pt->metadata_dev->bdev) / 4;
1963         return min((dm_block_t)1024ULL /* 4M */, quarter);
1964 }
1965
1966 /*
1967  * thin-pool <metadata dev> <data dev>
1968  *           <data block size (sectors)>
1969  *           <low water mark (blocks)>
1970  *           [<#feature args> [<arg>]*]
1971  *
1972  * Optional feature arguments are:
1973  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1974  *           ignore_discard: disable discard
1975  *           no_discard_passdown: don't pass discards down to the data device
1976  */
1977 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1978 {
1979         int r, pool_created = 0;
1980         struct pool_c *pt;
1981         struct pool *pool;
1982         struct pool_features pf;
1983         struct dm_arg_set as;
1984         struct dm_dev *data_dev;
1985         unsigned long block_size;
1986         dm_block_t low_water_blocks;
1987         struct dm_dev *metadata_dev;
1988         fmode_t metadata_mode;
1989
1990         /*
1991          * FIXME Remove validation from scope of lock.
1992          */
1993         mutex_lock(&dm_thin_pool_table.mutex);
1994
1995         if (argc < 4) {
1996                 ti->error = "Invalid argument count";
1997                 r = -EINVAL;
1998                 goto out_unlock;
1999         }
2000
2001         as.argc = argc;
2002         as.argv = argv;
2003
2004         /*
2005          * Set default pool features.
2006          */
2007         pool_features_init(&pf);
2008
2009         dm_consume_args(&as, 4);
2010         r = parse_pool_features(&as, &pf, ti);
2011         if (r)
2012                 goto out_unlock;
2013
2014         metadata_mode = FMODE_READ | ((pf.mode == PM_READ_ONLY) ? 0 : FMODE_WRITE);
2015         r = dm_get_device(ti, argv[0], metadata_mode, &metadata_dev);
2016         if (r) {
2017                 ti->error = "Error opening metadata block device";
2018                 goto out_unlock;
2019         }
2020
2021         /*
2022          * Run for the side-effect of possibly issuing a warning if the
2023          * device is too big.
2024          */
2025         (void) get_metadata_dev_size(metadata_dev->bdev);
2026
2027         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
2028         if (r) {
2029                 ti->error = "Error getting data device";
2030                 goto out_metadata;
2031         }
2032
2033         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
2034             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2035             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2036             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2037                 ti->error = "Invalid block size";
2038                 r = -EINVAL;
2039                 goto out;
2040         }
2041
2042         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
2043                 ti->error = "Invalid low water mark";
2044                 r = -EINVAL;
2045                 goto out;
2046         }
2047
2048         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
2049         if (!pt) {
2050                 r = -ENOMEM;
2051                 goto out;
2052         }
2053
2054         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
2055                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
2056         if (IS_ERR(pool)) {
2057                 r = PTR_ERR(pool);
2058                 goto out_free_pt;
2059         }
2060
2061         /*
2062          * 'pool_created' reflects whether this is the first table load.
2063          * Top level discard support is not allowed to be changed after
2064          * initial load.  This would require a pool reload to trigger thin
2065          * device changes.
2066          */
2067         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
2068                 ti->error = "Discard support cannot be disabled once enabled";
2069                 r = -EINVAL;
2070                 goto out_flags_changed;
2071         }
2072
2073         pt->pool = pool;
2074         pt->ti = ti;
2075         pt->metadata_dev = metadata_dev;
2076         pt->data_dev = data_dev;
2077         pt->low_water_blocks = low_water_blocks;
2078         pt->adjusted_pf = pt->requested_pf = pf;
2079         ti->num_flush_bios = 1;
2080
2081         /*
2082          * Only need to enable discards if the pool should pass
2083          * them down to the data device.  The thin device's discard
2084          * processing will cause mappings to be removed from the btree.
2085          */
2086         if (pf.discard_enabled && pf.discard_passdown) {
2087                 ti->num_discard_bios = 1;
2088
2089                 /*
2090                  * Setting 'discards_supported' circumvents the normal
2091                  * stacking of discard limits (this keeps the pool and
2092                  * thin devices' discard limits consistent).
2093                  */
2094                 ti->discards_supported = true;
2095                 ti->discard_zeroes_data_unsupported = true;
2096         }
2097         ti->private = pt;
2098
2099         r = dm_pool_register_metadata_threshold(pt->pool->pmd,
2100                                                 calc_metadata_threshold(pt),
2101                                                 metadata_low_callback,
2102                                                 pool);
2103         if (r)
2104                 goto out_free_pt;
2105
2106         pt->callbacks.congested_fn = pool_is_congested;
2107         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2108
2109         mutex_unlock(&dm_thin_pool_table.mutex);
2110
2111         return 0;
2112
2113 out_flags_changed:
2114         __pool_dec(pool);
2115 out_free_pt:
2116         kfree(pt);
2117 out:
2118         dm_put_device(ti, data_dev);
2119 out_metadata:
2120         dm_put_device(ti, metadata_dev);
2121 out_unlock:
2122         mutex_unlock(&dm_thin_pool_table.mutex);
2123
2124         return r;
2125 }
2126
2127 static int pool_map(struct dm_target *ti, struct bio *bio)
2128 {
2129         int r;
2130         struct pool_c *pt = ti->private;
2131         struct pool *pool = pt->pool;
2132         unsigned long flags;
2133
2134         /*
2135          * As this is a singleton target, ti->begin is always zero.
2136          */
2137         spin_lock_irqsave(&pool->lock, flags);
2138         bio->bi_bdev = pt->data_dev->bdev;
2139         r = DM_MAPIO_REMAPPED;
2140         spin_unlock_irqrestore(&pool->lock, flags);
2141
2142         return r;
2143 }
2144
2145 static int maybe_resize_data_dev(struct dm_target *ti, bool *need_commit)
2146 {
2147         int r;
2148         struct pool_c *pt = ti->private;
2149         struct pool *pool = pt->pool;
2150         sector_t data_size = ti->len;
2151         dm_block_t sb_data_size;
2152
2153         *need_commit = false;
2154
2155         (void) sector_div(data_size, pool->sectors_per_block);
2156
2157         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2158         if (r) {
2159                 DMERR("failed to retrieve data device size");
2160                 return r;
2161         }
2162
2163         if (data_size < sb_data_size) {
2164                 DMERR("pool target (%llu blocks) too small: expected %llu",
2165                       (unsigned long long)data_size, sb_data_size);
2166                 return -EINVAL;
2167
2168         } else if (data_size > sb_data_size) {
2169                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2170                 if (r) {
2171                         DMERR("failed to resize data device");
2172                         set_pool_mode(pool, PM_READ_ONLY);
2173                         return r;
2174                 }
2175
2176                 *need_commit = true;
2177         }
2178
2179         return 0;
2180 }
2181
2182 static int maybe_resize_metadata_dev(struct dm_target *ti, bool *need_commit)
2183 {
2184         int r;
2185         struct pool_c *pt = ti->private;
2186         struct pool *pool = pt->pool;
2187         dm_block_t metadata_dev_size, sb_metadata_dev_size;
2188
2189         *need_commit = false;
2190
2191         metadata_dev_size = get_metadata_dev_size_in_blocks(pool->md_dev);
2192
2193         r = dm_pool_get_metadata_dev_size(pool->pmd, &sb_metadata_dev_size);
2194         if (r) {
2195                 DMERR("failed to retrieve data device size");
2196                 return r;
2197         }
2198
2199         if (metadata_dev_size < sb_metadata_dev_size) {
2200                 DMERR("metadata device (%llu blocks) too small: expected %llu",
2201                       metadata_dev_size, sb_metadata_dev_size);
2202                 return -EINVAL;
2203
2204         } else if (metadata_dev_size > sb_metadata_dev_size) {
2205                 r = dm_pool_resize_metadata_dev(pool->pmd, metadata_dev_size);
2206                 if (r) {
2207                         DMERR("failed to resize metadata device");
2208                         return r;
2209                 }
2210
2211                 *need_commit = true;
2212         }
2213
2214         return 0;
2215 }
2216
2217 /*
2218  * Retrieves the number of blocks of the data device from
2219  * the superblock and compares it to the actual device size,
2220  * thus resizing the data device in case it has grown.
2221  *
2222  * This both copes with opening preallocated data devices in the ctr
2223  * being followed by a resume
2224  * -and-
2225  * calling the resume method individually after userspace has
2226  * grown the data device in reaction to a table event.
2227  */
2228 static int pool_preresume(struct dm_target *ti)
2229 {
2230         int r;
2231         bool need_commit1, need_commit2;
2232         struct pool_c *pt = ti->private;
2233         struct pool *pool = pt->pool;
2234
2235         /*
2236          * Take control of the pool object.
2237          */
2238         r = bind_control_target(pool, ti);
2239         if (r)
2240                 return r;
2241
2242         r = maybe_resize_data_dev(ti, &need_commit1);
2243         if (r)
2244                 return r;
2245
2246         r = maybe_resize_metadata_dev(ti, &need_commit2);
2247         if (r)
2248                 return r;
2249
2250         if (need_commit1 || need_commit2)
2251                 (void) commit_or_fallback(pool);
2252
2253         return 0;
2254 }
2255
2256 static void pool_resume(struct dm_target *ti)
2257 {
2258         struct pool_c *pt = ti->private;
2259         struct pool *pool = pt->pool;
2260         unsigned long flags;
2261
2262         spin_lock_irqsave(&pool->lock, flags);
2263         pool->low_water_triggered = 0;
2264         pool->no_free_space = 0;
2265         __requeue_bios(pool);
2266         spin_unlock_irqrestore(&pool->lock, flags);
2267
2268         do_waker(&pool->waker.work);
2269 }
2270
2271 static void pool_postsuspend(struct dm_target *ti)
2272 {
2273         struct pool_c *pt = ti->private;
2274         struct pool *pool = pt->pool;
2275
2276         cancel_delayed_work(&pool->waker);
2277         flush_workqueue(pool->wq);
2278         (void) commit_or_fallback(pool);
2279 }
2280
2281 static int check_arg_count(unsigned argc, unsigned args_required)
2282 {
2283         if (argc != args_required) {
2284                 DMWARN("Message received with %u arguments instead of %u.",
2285                        argc, args_required);
2286                 return -EINVAL;
2287         }
2288
2289         return 0;
2290 }
2291
2292 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2293 {
2294         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2295             *dev_id <= MAX_DEV_ID)
2296                 return 0;
2297
2298         if (warning)
2299                 DMWARN("Message received with invalid device id: %s", arg);
2300
2301         return -EINVAL;
2302 }
2303
2304 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2305 {
2306         dm_thin_id dev_id;
2307         int r;
2308
2309         r = check_arg_count(argc, 2);
2310         if (r)
2311                 return r;
2312
2313         r = read_dev_id(argv[1], &dev_id, 1);
2314         if (r)
2315                 return r;
2316
2317         r = dm_pool_create_thin(pool->pmd, dev_id);
2318         if (r) {
2319                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2320                        argv[1]);
2321                 return r;
2322         }
2323
2324         return 0;
2325 }
2326
2327 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2328 {
2329         dm_thin_id dev_id;
2330         dm_thin_id origin_dev_id;
2331         int r;
2332
2333         r = check_arg_count(argc, 3);
2334         if (r)
2335                 return r;
2336
2337         r = read_dev_id(argv[1], &dev_id, 1);
2338         if (r)
2339                 return r;
2340
2341         r = read_dev_id(argv[2], &origin_dev_id, 1);
2342         if (r)
2343                 return r;
2344
2345         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2346         if (r) {
2347                 DMWARN("Creation of new snapshot %s of device %s failed.",
2348                        argv[1], argv[2]);
2349                 return r;
2350         }
2351
2352         return 0;
2353 }
2354
2355 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2356 {
2357         dm_thin_id dev_id;
2358         int r;
2359
2360         r = check_arg_count(argc, 2);
2361         if (r)
2362                 return r;
2363
2364         r = read_dev_id(argv[1], &dev_id, 1);
2365         if (r)
2366                 return r;
2367
2368         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2369         if (r)
2370                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2371
2372         return r;
2373 }
2374
2375 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2376 {
2377         dm_thin_id old_id, new_id;
2378         int r;
2379
2380         r = check_arg_count(argc, 3);
2381         if (r)
2382                 return r;
2383
2384         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2385                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2386                 return -EINVAL;
2387         }
2388
2389         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2390                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2391                 return -EINVAL;
2392         }
2393
2394         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2395         if (r) {
2396                 DMWARN("Failed to change transaction id from %s to %s.",
2397                        argv[1], argv[2]);
2398                 return r;
2399         }
2400
2401         return 0;
2402 }
2403
2404 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2405 {
2406         int r;
2407
2408         r = check_arg_count(argc, 1);
2409         if (r)
2410                 return r;
2411
2412         (void) commit_or_fallback(pool);
2413
2414         r = dm_pool_reserve_metadata_snap(pool->pmd);
2415         if (r)
2416                 DMWARN("reserve_metadata_snap message failed.");
2417
2418         return r;
2419 }
2420
2421 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2422 {
2423         int r;
2424
2425         r = check_arg_count(argc, 1);
2426         if (r)
2427                 return r;
2428
2429         r = dm_pool_release_metadata_snap(pool->pmd);
2430         if (r)
2431                 DMWARN("release_metadata_snap message failed.");
2432
2433         return r;
2434 }
2435
2436 /*
2437  * Messages supported:
2438  *   create_thin        <dev_id>
2439  *   create_snap        <dev_id> <origin_id>
2440  *   delete             <dev_id>
2441  *   trim               <dev_id> <new_size_in_sectors>
2442  *   set_transaction_id <current_trans_id> <new_trans_id>
2443  *   reserve_metadata_snap
2444  *   release_metadata_snap
2445  */
2446 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2447 {
2448         int r = -EINVAL;
2449         struct pool_c *pt = ti->private;
2450         struct pool *pool = pt->pool;
2451
2452         if (!strcasecmp(argv[0], "create_thin"))
2453                 r = process_create_thin_mesg(argc, argv, pool);
2454
2455         else if (!strcasecmp(argv[0], "create_snap"))
2456                 r = process_create_snap_mesg(argc, argv, pool);
2457
2458         else if (!strcasecmp(argv[0], "delete"))
2459                 r = process_delete_mesg(argc, argv, pool);
2460
2461         else if (!strcasecmp(argv[0], "set_transaction_id"))
2462                 r = process_set_transaction_id_mesg(argc, argv, pool);
2463
2464         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2465                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2466
2467         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2468                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2469
2470         else
2471                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2472
2473         if (!r)
2474                 (void) commit_or_fallback(pool);
2475
2476         return r;
2477 }
2478
2479 static void emit_flags(struct pool_features *pf, char *result,
2480                        unsigned sz, unsigned maxlen)
2481 {
2482         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2483                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2484         DMEMIT("%u ", count);
2485
2486         if (!pf->zero_new_blocks)
2487                 DMEMIT("skip_block_zeroing ");
2488
2489         if (!pf->discard_enabled)
2490                 DMEMIT("ignore_discard ");
2491
2492         if (!pf->discard_passdown)
2493                 DMEMIT("no_discard_passdown ");
2494
2495         if (pf->mode == PM_READ_ONLY)
2496                 DMEMIT("read_only ");
2497 }
2498
2499 /*
2500  * Status line is:
2501  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2502  *    <used data sectors>/<total data sectors> <held metadata root>
2503  */
2504 static void pool_status(struct dm_target *ti, status_type_t type,
2505                         unsigned status_flags, char *result, unsigned maxlen)
2506 {
2507         int r;
2508         unsigned sz = 0;
2509         uint64_t transaction_id;
2510         dm_block_t nr_free_blocks_data;
2511         dm_block_t nr_free_blocks_metadata;
2512         dm_block_t nr_blocks_data;
2513         dm_block_t nr_blocks_metadata;
2514         dm_block_t held_root;
2515         char buf[BDEVNAME_SIZE];
2516         char buf2[BDEVNAME_SIZE];
2517         struct pool_c *pt = ti->private;
2518         struct pool *pool = pt->pool;
2519
2520         switch (type) {
2521         case STATUSTYPE_INFO:
2522                 if (get_pool_mode(pool) == PM_FAIL) {
2523                         DMEMIT("Fail");
2524                         break;
2525                 }
2526
2527                 /* Commit to ensure statistics aren't out-of-date */
2528                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2529                         (void) commit_or_fallback(pool);
2530
2531                 r = dm_pool_get_metadata_transaction_id(pool->pmd, &transaction_id);
2532                 if (r) {
2533                         DMERR("dm_pool_get_metadata_transaction_id returned %d", r);
2534                         goto err;
2535                 }
2536
2537                 r = dm_pool_get_free_metadata_block_count(pool->pmd, &nr_free_blocks_metadata);
2538                 if (r) {
2539                         DMERR("dm_pool_get_free_metadata_block_count returned %d", r);
2540                         goto err;
2541                 }
2542
2543                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2544                 if (r) {
2545                         DMERR("dm_pool_get_metadata_dev_size returned %d", r);
2546                         goto err;
2547                 }
2548
2549                 r = dm_pool_get_free_block_count(pool->pmd, &nr_free_blocks_data);
2550                 if (r) {
2551                         DMERR("dm_pool_get_free_block_count returned %d", r);
2552                         goto err;
2553                 }
2554
2555                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2556                 if (r) {
2557                         DMERR("dm_pool_get_data_dev_size returned %d", r);
2558                         goto err;
2559                 }
2560
2561                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2562                 if (r) {
2563                         DMERR("dm_pool_get_metadata_snap returned %d", r);
2564                         goto err;
2565                 }
2566
2567                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2568                        (unsigned long long)transaction_id,
2569                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2570                        (unsigned long long)nr_blocks_metadata,
2571                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2572                        (unsigned long long)nr_blocks_data);
2573
2574                 if (held_root)
2575                         DMEMIT("%llu ", held_root);
2576                 else
2577                         DMEMIT("- ");
2578
2579                 if (pool->pf.mode == PM_READ_ONLY)
2580                         DMEMIT("ro ");
2581                 else
2582                         DMEMIT("rw ");
2583
2584                 if (!pool->pf.discard_enabled)
2585                         DMEMIT("ignore_discard");
2586                 else if (pool->pf.discard_passdown)
2587                         DMEMIT("discard_passdown");
2588                 else
2589                         DMEMIT("no_discard_passdown");
2590
2591                 break;
2592
2593         case STATUSTYPE_TABLE:
2594                 DMEMIT("%s %s %lu %llu ",
2595                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2596                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2597                        (unsigned long)pool->sectors_per_block,
2598                        (unsigned long long)pt->low_water_blocks);
2599                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2600                 break;
2601         }
2602         return;
2603
2604 err:
2605         DMEMIT("Error");
2606 }
2607
2608 static int pool_iterate_devices(struct dm_target *ti,
2609                                 iterate_devices_callout_fn fn, void *data)
2610 {
2611         struct pool_c *pt = ti->private;
2612
2613         return fn(ti, pt->data_dev, 0, ti->len, data);
2614 }
2615
2616 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2617                       struct bio_vec *biovec, int max_size)
2618 {
2619         struct pool_c *pt = ti->private;
2620         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2621
2622         if (!q->merge_bvec_fn)
2623                 return max_size;
2624
2625         bvm->bi_bdev = pt->data_dev->bdev;
2626
2627         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2628 }
2629
2630 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2631 {
2632         struct pool *pool = pt->pool;
2633         struct queue_limits *data_limits;
2634
2635         limits->max_discard_sectors = pool->sectors_per_block;
2636
2637         /*
2638          * discard_granularity is just a hint, and not enforced.
2639          */
2640         if (pt->adjusted_pf.discard_passdown) {
2641                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2642                 limits->discard_granularity = data_limits->discard_granularity;
2643         } else
2644                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2645 }
2646
2647 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2648 {
2649         struct pool_c *pt = ti->private;
2650         struct pool *pool = pt->pool;
2651
2652         blk_limits_io_min(limits, 0);
2653         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2654
2655         /*
2656          * pt->adjusted_pf is a staging area for the actual features to use.
2657          * They get transferred to the live pool in bind_control_target()
2658          * called from pool_preresume().
2659          */
2660         if (!pt->adjusted_pf.discard_enabled)
2661                 return;
2662
2663         disable_passdown_if_not_supported(pt);
2664
2665         set_discard_limits(pt, limits);
2666 }
2667
2668 static struct target_type pool_target = {
2669         .name = "thin-pool",
2670         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2671                     DM_TARGET_IMMUTABLE,
2672         .version = {1, 8, 0},
2673         .module = THIS_MODULE,
2674         .ctr = pool_ctr,
2675         .dtr = pool_dtr,
2676         .map = pool_map,
2677         .postsuspend = pool_postsuspend,
2678         .preresume = pool_preresume,
2679         .resume = pool_resume,
2680         .message = pool_message,
2681         .status = pool_status,
2682         .merge = pool_merge,
2683         .iterate_devices = pool_iterate_devices,
2684         .io_hints = pool_io_hints,
2685 };
2686
2687 /*----------------------------------------------------------------
2688  * Thin target methods
2689  *--------------------------------------------------------------*/
2690 static void thin_dtr(struct dm_target *ti)
2691 {
2692         struct thin_c *tc = ti->private;
2693
2694         mutex_lock(&dm_thin_pool_table.mutex);
2695
2696         __pool_dec(tc->pool);
2697         dm_pool_close_thin_device(tc->td);
2698         dm_put_device(ti, tc->pool_dev);
2699         if (tc->origin_dev)
2700                 dm_put_device(ti, tc->origin_dev);
2701         kfree(tc);
2702
2703         mutex_unlock(&dm_thin_pool_table.mutex);
2704 }
2705
2706 /*
2707  * Thin target parameters:
2708  *
2709  * <pool_dev> <dev_id> [origin_dev]
2710  *
2711  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2712  * dev_id: the internal device identifier
2713  * origin_dev: a device external to the pool that should act as the origin
2714  *
2715  * If the pool device has discards disabled, they get disabled for the thin
2716  * device as well.
2717  */
2718 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2719 {
2720         int r;
2721         struct thin_c *tc;
2722         struct dm_dev *pool_dev, *origin_dev;
2723         struct mapped_device *pool_md;
2724
2725         mutex_lock(&dm_thin_pool_table.mutex);
2726
2727         if (argc != 2 && argc != 3) {
2728                 ti->error = "Invalid argument count";
2729                 r = -EINVAL;
2730                 goto out_unlock;
2731         }
2732
2733         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2734         if (!tc) {
2735                 ti->error = "Out of memory";
2736                 r = -ENOMEM;
2737                 goto out_unlock;
2738         }
2739
2740         if (argc == 3) {
2741                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2742                 if (r) {
2743                         ti->error = "Error opening origin device";
2744                         goto bad_origin_dev;
2745                 }
2746                 tc->origin_dev = origin_dev;
2747         }
2748
2749         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2750         if (r) {
2751                 ti->error = "Error opening pool device";
2752                 goto bad_pool_dev;
2753         }
2754         tc->pool_dev = pool_dev;
2755
2756         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2757                 ti->error = "Invalid device id";
2758                 r = -EINVAL;
2759                 goto bad_common;
2760         }
2761
2762         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2763         if (!pool_md) {
2764                 ti->error = "Couldn't get pool mapped device";
2765                 r = -EINVAL;
2766                 goto bad_common;
2767         }
2768
2769         tc->pool = __pool_table_lookup(pool_md);
2770         if (!tc->pool) {
2771                 ti->error = "Couldn't find pool object";
2772                 r = -EINVAL;
2773                 goto bad_pool_lookup;
2774         }
2775         __pool_inc(tc->pool);
2776
2777         if (get_pool_mode(tc->pool) == PM_FAIL) {
2778                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2779                 goto bad_thin_open;
2780         }
2781
2782         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2783         if (r) {
2784                 ti->error = "Couldn't open thin internal device";
2785                 goto bad_thin_open;
2786         }
2787
2788         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2789         if (r)
2790                 goto bad_thin_open;
2791
2792         ti->num_flush_bios = 1;
2793         ti->flush_supported = true;
2794         ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
2795
2796         /* In case the pool supports discards, pass them on. */
2797         if (tc->pool->pf.discard_enabled) {
2798                 ti->discards_supported = true;
2799                 ti->num_discard_bios = 1;
2800                 ti->discard_zeroes_data_unsupported = true;
2801                 /* Discard bios must be split on a block boundary */
2802                 ti->split_discard_bios = true;
2803         }
2804
2805         dm_put(pool_md);
2806
2807         mutex_unlock(&dm_thin_pool_table.mutex);
2808
2809         return 0;
2810
2811 bad_thin_open:
2812         __pool_dec(tc->pool);
2813 bad_pool_lookup:
2814         dm_put(pool_md);
2815 bad_common:
2816         dm_put_device(ti, tc->pool_dev);
2817 bad_pool_dev:
2818         if (tc->origin_dev)
2819                 dm_put_device(ti, tc->origin_dev);
2820 bad_origin_dev:
2821         kfree(tc);
2822 out_unlock:
2823         mutex_unlock(&dm_thin_pool_table.mutex);
2824
2825         return r;
2826 }
2827
2828 static int thin_map(struct dm_target *ti, struct bio *bio)
2829 {
2830         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2831
2832         return thin_bio_map(ti, bio);
2833 }
2834
2835 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
2836 {
2837         unsigned long flags;
2838         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2839         struct list_head work;
2840         struct dm_thin_new_mapping *m, *tmp;
2841         struct pool *pool = h->tc->pool;
2842
2843         if (h->shared_read_entry) {
2844                 INIT_LIST_HEAD(&work);
2845                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2846
2847                 spin_lock_irqsave(&pool->lock, flags);
2848                 list_for_each_entry_safe(m, tmp, &work, list) {
2849                         list_del(&m->list);
2850                         m->quiesced = 1;
2851                         __maybe_add_mapping(m);
2852                 }
2853                 spin_unlock_irqrestore(&pool->lock, flags);
2854         }
2855
2856         if (h->all_io_entry) {
2857                 INIT_LIST_HEAD(&work);
2858                 dm_deferred_entry_dec(h->all_io_entry, &work);
2859                 if (!list_empty(&work)) {
2860                         spin_lock_irqsave(&pool->lock, flags);
2861                         list_for_each_entry_safe(m, tmp, &work, list)
2862                                 list_add(&m->list, &pool->prepared_discards);
2863                         spin_unlock_irqrestore(&pool->lock, flags);
2864                         wake_worker(pool);
2865                 }
2866         }
2867
2868         return 0;
2869 }
2870
2871 static void thin_postsuspend(struct dm_target *ti)
2872 {
2873         if (dm_noflush_suspending(ti))
2874                 requeue_io((struct thin_c *)ti->private);
2875 }
2876
2877 /*
2878  * <nr mapped sectors> <highest mapped sector>
2879  */
2880 static void thin_status(struct dm_target *ti, status_type_t type,
2881                         unsigned status_flags, char *result, unsigned maxlen)
2882 {
2883         int r;
2884         ssize_t sz = 0;
2885         dm_block_t mapped, highest;
2886         char buf[BDEVNAME_SIZE];
2887         struct thin_c *tc = ti->private;
2888
2889         if (get_pool_mode(tc->pool) == PM_FAIL) {
2890                 DMEMIT("Fail");
2891                 return;
2892         }
2893
2894         if (!tc->td)
2895                 DMEMIT("-");
2896         else {
2897                 switch (type) {
2898                 case STATUSTYPE_INFO:
2899                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2900                         if (r) {
2901                                 DMERR("dm_thin_get_mapped_count returned %d", r);
2902                                 goto err;
2903                         }
2904
2905                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2906                         if (r < 0) {
2907                                 DMERR("dm_thin_get_highest_mapped_block returned %d", r);
2908                                 goto err;
2909                         }
2910
2911                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2912                         if (r)
2913                                 DMEMIT("%llu", ((highest + 1) *
2914                                                 tc->pool->sectors_per_block) - 1);
2915                         else
2916                                 DMEMIT("-");
2917                         break;
2918
2919                 case STATUSTYPE_TABLE:
2920                         DMEMIT("%s %lu",
2921                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2922                                (unsigned long) tc->dev_id);
2923                         if (tc->origin_dev)
2924                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2925                         break;
2926                 }
2927         }
2928
2929         return;
2930
2931 err:
2932         DMEMIT("Error");
2933 }
2934
2935 static int thin_iterate_devices(struct dm_target *ti,
2936                                 iterate_devices_callout_fn fn, void *data)
2937 {
2938         sector_t blocks;
2939         struct thin_c *tc = ti->private;
2940         struct pool *pool = tc->pool;
2941
2942         /*
2943          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2944          * we follow a more convoluted path through to the pool's target.
2945          */
2946         if (!pool->ti)
2947                 return 0;       /* nothing is bound */
2948
2949         blocks = pool->ti->len;
2950         (void) sector_div(blocks, pool->sectors_per_block);
2951         if (blocks)
2952                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2953
2954         return 0;
2955 }
2956
2957 static struct target_type thin_target = {
2958         .name = "thin",
2959         .version = {1, 8, 0},
2960         .module = THIS_MODULE,
2961         .ctr = thin_ctr,
2962         .dtr = thin_dtr,
2963         .map = thin_map,
2964         .end_io = thin_endio,
2965         .postsuspend = thin_postsuspend,
2966         .status = thin_status,
2967         .iterate_devices = thin_iterate_devices,
2968 };
2969
2970 /*----------------------------------------------------------------*/
2971
2972 static int __init dm_thin_init(void)
2973 {
2974         int r;
2975
2976         pool_table_init();
2977
2978         r = dm_register_target(&thin_target);
2979         if (r)
2980                 return r;
2981
2982         r = dm_register_target(&pool_target);
2983         if (r)
2984                 goto bad_pool_target;
2985
2986         r = -ENOMEM;
2987
2988         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2989         if (!_new_mapping_cache)
2990                 goto bad_new_mapping_cache;
2991
2992         return 0;
2993
2994 bad_new_mapping_cache:
2995         dm_unregister_target(&pool_target);
2996 bad_pool_target:
2997         dm_unregister_target(&thin_target);
2998
2999         return r;
3000 }
3001
3002 static void dm_thin_exit(void)
3003 {
3004         dm_unregister_target(&thin_target);
3005         dm_unregister_target(&pool_target);
3006
3007         kmem_cache_destroy(_new_mapping_cache);
3008 }
3009
3010 module_init(dm_thin_init);
3011 module_exit(dm_thin_exit);
3012
3013 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
3014 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
3015 MODULE_LICENSE("GPL");