dm thin: add pool target flags to control discard
[linux-2.6.git] / drivers / md / dm-thin.c
1 /*
2  * Copyright (C) 2011 Red Hat UK.
3  *
4  * This file is released under the GPL.
5  */
6
7 #include "dm-thin-metadata.h"
8
9 #include <linux/device-mapper.h>
10 #include <linux/dm-io.h>
11 #include <linux/dm-kcopyd.h>
12 #include <linux/list.h>
13 #include <linux/init.h>
14 #include <linux/module.h>
15 #include <linux/slab.h>
16
17 #define DM_MSG_PREFIX   "thin"
18
19 /*
20  * Tunable constants
21  */
22 #define ENDIO_HOOK_POOL_SIZE 10240
23 #define DEFERRED_SET_SIZE 64
24 #define MAPPING_POOL_SIZE 1024
25 #define PRISON_CELLS 1024
26 #define COMMIT_PERIOD HZ
27
28 /*
29  * The block size of the device holding pool data must be
30  * between 64KB and 1GB.
31  */
32 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
33 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
34
35 /*
36  * Device id is restricted to 24 bits.
37  */
38 #define MAX_DEV_ID ((1 << 24) - 1)
39
40 /*
41  * How do we handle breaking sharing of data blocks?
42  * =================================================
43  *
44  * We use a standard copy-on-write btree to store the mappings for the
45  * devices (note I'm talking about copy-on-write of the metadata here, not
46  * the data).  When you take an internal snapshot you clone the root node
47  * of the origin btree.  After this there is no concept of an origin or a
48  * snapshot.  They are just two device trees that happen to point to the
49  * same data blocks.
50  *
51  * When we get a write in we decide if it's to a shared data block using
52  * some timestamp magic.  If it is, we have to break sharing.
53  *
54  * Let's say we write to a shared block in what was the origin.  The
55  * steps are:
56  *
57  * i) plug io further to this physical block. (see bio_prison code).
58  *
59  * ii) quiesce any read io to that shared data block.  Obviously
60  * including all devices that share this block.  (see deferred_set code)
61  *
62  * iii) copy the data block to a newly allocate block.  This step can be
63  * missed out if the io covers the block. (schedule_copy).
64  *
65  * iv) insert the new mapping into the origin's btree
66  * (process_prepared_mapping).  This act of inserting breaks some
67  * sharing of btree nodes between the two devices.  Breaking sharing only
68  * effects the btree of that specific device.  Btrees for the other
69  * devices that share the block never change.  The btree for the origin
70  * device as it was after the last commit is untouched, ie. we're using
71  * persistent data structures in the functional programming sense.
72  *
73  * v) unplug io to this physical block, including the io that triggered
74  * the breaking of sharing.
75  *
76  * Steps (ii) and (iii) occur in parallel.
77  *
78  * The metadata _doesn't_ need to be committed before the io continues.  We
79  * get away with this because the io is always written to a _new_ block.
80  * If there's a crash, then:
81  *
82  * - The origin mapping will point to the old origin block (the shared
83  * one).  This will contain the data as it was before the io that triggered
84  * the breaking of sharing came in.
85  *
86  * - The snap mapping still points to the old block.  As it would after
87  * the commit.
88  *
89  * The downside of this scheme is the timestamp magic isn't perfect, and
90  * will continue to think that data block in the snapshot device is shared
91  * even after the write to the origin has broken sharing.  I suspect data
92  * blocks will typically be shared by many different devices, so we're
93  * breaking sharing n + 1 times, rather than n, where n is the number of
94  * devices that reference this data block.  At the moment I think the
95  * benefits far, far outweigh the disadvantages.
96  */
97
98 /*----------------------------------------------------------------*/
99
100 /*
101  * Sometimes we can't deal with a bio straight away.  We put them in prison
102  * where they can't cause any mischief.  Bios are put in a cell identified
103  * by a key, multiple bios can be in the same cell.  When the cell is
104  * subsequently unlocked the bios become available.
105  */
106 struct bio_prison;
107
108 struct cell_key {
109         int virtual;
110         dm_thin_id dev;
111         dm_block_t block;
112 };
113
114 struct cell {
115         struct hlist_node list;
116         struct bio_prison *prison;
117         struct cell_key key;
118         struct bio *holder;
119         struct bio_list bios;
120 };
121
122 struct bio_prison {
123         spinlock_t lock;
124         mempool_t *cell_pool;
125
126         unsigned nr_buckets;
127         unsigned hash_mask;
128         struct hlist_head *cells;
129 };
130
131 static uint32_t calc_nr_buckets(unsigned nr_cells)
132 {
133         uint32_t n = 128;
134
135         nr_cells /= 4;
136         nr_cells = min(nr_cells, 8192u);
137
138         while (n < nr_cells)
139                 n <<= 1;
140
141         return n;
142 }
143
144 /*
145  * @nr_cells should be the number of cells you want in use _concurrently_.
146  * Don't confuse it with the number of distinct keys.
147  */
148 static struct bio_prison *prison_create(unsigned nr_cells)
149 {
150         unsigned i;
151         uint32_t nr_buckets = calc_nr_buckets(nr_cells);
152         size_t len = sizeof(struct bio_prison) +
153                 (sizeof(struct hlist_head) * nr_buckets);
154         struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
155
156         if (!prison)
157                 return NULL;
158
159         spin_lock_init(&prison->lock);
160         prison->cell_pool = mempool_create_kmalloc_pool(nr_cells,
161                                                         sizeof(struct cell));
162         if (!prison->cell_pool) {
163                 kfree(prison);
164                 return NULL;
165         }
166
167         prison->nr_buckets = nr_buckets;
168         prison->hash_mask = nr_buckets - 1;
169         prison->cells = (struct hlist_head *) (prison + 1);
170         for (i = 0; i < nr_buckets; i++)
171                 INIT_HLIST_HEAD(prison->cells + i);
172
173         return prison;
174 }
175
176 static void prison_destroy(struct bio_prison *prison)
177 {
178         mempool_destroy(prison->cell_pool);
179         kfree(prison);
180 }
181
182 static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
183 {
184         const unsigned long BIG_PRIME = 4294967291UL;
185         uint64_t hash = key->block * BIG_PRIME;
186
187         return (uint32_t) (hash & prison->hash_mask);
188 }
189
190 static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
191 {
192                return (lhs->virtual == rhs->virtual) &&
193                        (lhs->dev == rhs->dev) &&
194                        (lhs->block == rhs->block);
195 }
196
197 static struct cell *__search_bucket(struct hlist_head *bucket,
198                                     struct cell_key *key)
199 {
200         struct cell *cell;
201         struct hlist_node *tmp;
202
203         hlist_for_each_entry(cell, tmp, bucket, list)
204                 if (keys_equal(&cell->key, key))
205                         return cell;
206
207         return NULL;
208 }
209
210 /*
211  * This may block if a new cell needs allocating.  You must ensure that
212  * cells will be unlocked even if the calling thread is blocked.
213  *
214  * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
215  */
216 static int bio_detain(struct bio_prison *prison, struct cell_key *key,
217                       struct bio *inmate, struct cell **ref)
218 {
219         int r = 1;
220         unsigned long flags;
221         uint32_t hash = hash_key(prison, key);
222         struct cell *cell, *cell2;
223
224         BUG_ON(hash > prison->nr_buckets);
225
226         spin_lock_irqsave(&prison->lock, flags);
227
228         cell = __search_bucket(prison->cells + hash, key);
229         if (cell) {
230                 bio_list_add(&cell->bios, inmate);
231                 goto out;
232         }
233
234         /*
235          * Allocate a new cell
236          */
237         spin_unlock_irqrestore(&prison->lock, flags);
238         cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
239         spin_lock_irqsave(&prison->lock, flags);
240
241         /*
242          * We've been unlocked, so we have to double check that
243          * nobody else has inserted this cell in the meantime.
244          */
245         cell = __search_bucket(prison->cells + hash, key);
246         if (cell) {
247                 mempool_free(cell2, prison->cell_pool);
248                 bio_list_add(&cell->bios, inmate);
249                 goto out;
250         }
251
252         /*
253          * Use new cell.
254          */
255         cell = cell2;
256
257         cell->prison = prison;
258         memcpy(&cell->key, key, sizeof(cell->key));
259         cell->holder = inmate;
260         bio_list_init(&cell->bios);
261         hlist_add_head(&cell->list, prison->cells + hash);
262
263         r = 0;
264
265 out:
266         spin_unlock_irqrestore(&prison->lock, flags);
267
268         *ref = cell;
269
270         return r;
271 }
272
273 /*
274  * @inmates must have been initialised prior to this call
275  */
276 static void __cell_release(struct cell *cell, struct bio_list *inmates)
277 {
278         struct bio_prison *prison = cell->prison;
279
280         hlist_del(&cell->list);
281
282         bio_list_add(inmates, cell->holder);
283         bio_list_merge(inmates, &cell->bios);
284
285         mempool_free(cell, prison->cell_pool);
286 }
287
288 static void cell_release(struct cell *cell, struct bio_list *bios)
289 {
290         unsigned long flags;
291         struct bio_prison *prison = cell->prison;
292
293         spin_lock_irqsave(&prison->lock, flags);
294         __cell_release(cell, bios);
295         spin_unlock_irqrestore(&prison->lock, flags);
296 }
297
298 /*
299  * There are a couple of places where we put a bio into a cell briefly
300  * before taking it out again.  In these situations we know that no other
301  * bio may be in the cell.  This function releases the cell, and also does
302  * a sanity check.
303  */
304 static void __cell_release_singleton(struct cell *cell, struct bio *bio)
305 {
306         hlist_del(&cell->list);
307         BUG_ON(cell->holder != bio);
308         BUG_ON(!bio_list_empty(&cell->bios));
309 }
310
311 static void cell_release_singleton(struct cell *cell, struct bio *bio)
312 {
313         unsigned long flags;
314         struct bio_prison *prison = cell->prison;
315
316         spin_lock_irqsave(&prison->lock, flags);
317         __cell_release_singleton(cell, bio);
318         spin_unlock_irqrestore(&prison->lock, flags);
319 }
320
321 /*
322  * Sometimes we don't want the holder, just the additional bios.
323  */
324 static void __cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
325 {
326         struct bio_prison *prison = cell->prison;
327
328         hlist_del(&cell->list);
329         bio_list_merge(inmates, &cell->bios);
330
331         mempool_free(cell, prison->cell_pool);
332 }
333
334 static void cell_release_no_holder(struct cell *cell, struct bio_list *inmates)
335 {
336         unsigned long flags;
337         struct bio_prison *prison = cell->prison;
338
339         spin_lock_irqsave(&prison->lock, flags);
340         __cell_release_no_holder(cell, inmates);
341         spin_unlock_irqrestore(&prison->lock, flags);
342 }
343
344 static void cell_error(struct cell *cell)
345 {
346         struct bio_prison *prison = cell->prison;
347         struct bio_list bios;
348         struct bio *bio;
349         unsigned long flags;
350
351         bio_list_init(&bios);
352
353         spin_lock_irqsave(&prison->lock, flags);
354         __cell_release(cell, &bios);
355         spin_unlock_irqrestore(&prison->lock, flags);
356
357         while ((bio = bio_list_pop(&bios)))
358                 bio_io_error(bio);
359 }
360
361 /*----------------------------------------------------------------*/
362
363 /*
364  * We use the deferred set to keep track of pending reads to shared blocks.
365  * We do this to ensure the new mapping caused by a write isn't performed
366  * until these prior reads have completed.  Otherwise the insertion of the
367  * new mapping could free the old block that the read bios are mapped to.
368  */
369
370 struct deferred_set;
371 struct deferred_entry {
372         struct deferred_set *ds;
373         unsigned count;
374         struct list_head work_items;
375 };
376
377 struct deferred_set {
378         spinlock_t lock;
379         unsigned current_entry;
380         unsigned sweeper;
381         struct deferred_entry entries[DEFERRED_SET_SIZE];
382 };
383
384 static void ds_init(struct deferred_set *ds)
385 {
386         int i;
387
388         spin_lock_init(&ds->lock);
389         ds->current_entry = 0;
390         ds->sweeper = 0;
391         for (i = 0; i < DEFERRED_SET_SIZE; i++) {
392                 ds->entries[i].ds = ds;
393                 ds->entries[i].count = 0;
394                 INIT_LIST_HEAD(&ds->entries[i].work_items);
395         }
396 }
397
398 static struct deferred_entry *ds_inc(struct deferred_set *ds)
399 {
400         unsigned long flags;
401         struct deferred_entry *entry;
402
403         spin_lock_irqsave(&ds->lock, flags);
404         entry = ds->entries + ds->current_entry;
405         entry->count++;
406         spin_unlock_irqrestore(&ds->lock, flags);
407
408         return entry;
409 }
410
411 static unsigned ds_next(unsigned index)
412 {
413         return (index + 1) % DEFERRED_SET_SIZE;
414 }
415
416 static void __sweep(struct deferred_set *ds, struct list_head *head)
417 {
418         while ((ds->sweeper != ds->current_entry) &&
419                !ds->entries[ds->sweeper].count) {
420                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
421                 ds->sweeper = ds_next(ds->sweeper);
422         }
423
424         if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
425                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
426 }
427
428 static void ds_dec(struct deferred_entry *entry, struct list_head *head)
429 {
430         unsigned long flags;
431
432         spin_lock_irqsave(&entry->ds->lock, flags);
433         BUG_ON(!entry->count);
434         --entry->count;
435         __sweep(entry->ds, head);
436         spin_unlock_irqrestore(&entry->ds->lock, flags);
437 }
438
439 /*
440  * Returns 1 if deferred or 0 if no pending items to delay job.
441  */
442 static int ds_add_work(struct deferred_set *ds, struct list_head *work)
443 {
444         int r = 1;
445         unsigned long flags;
446         unsigned next_entry;
447
448         spin_lock_irqsave(&ds->lock, flags);
449         if ((ds->sweeper == ds->current_entry) &&
450             !ds->entries[ds->current_entry].count)
451                 r = 0;
452         else {
453                 list_add(work, &ds->entries[ds->current_entry].work_items);
454                 next_entry = ds_next(ds->current_entry);
455                 if (!ds->entries[next_entry].count)
456                         ds->current_entry = next_entry;
457         }
458         spin_unlock_irqrestore(&ds->lock, flags);
459
460         return r;
461 }
462
463 /*----------------------------------------------------------------*/
464
465 /*
466  * Key building.
467  */
468 static void build_data_key(struct dm_thin_device *td,
469                            dm_block_t b, struct cell_key *key)
470 {
471         key->virtual = 0;
472         key->dev = dm_thin_dev_id(td);
473         key->block = b;
474 }
475
476 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
477                               struct cell_key *key)
478 {
479         key->virtual = 1;
480         key->dev = dm_thin_dev_id(td);
481         key->block = b;
482 }
483
484 /*----------------------------------------------------------------*/
485
486 /*
487  * A pool device ties together a metadata device and a data device.  It
488  * also provides the interface for creating and destroying internal
489  * devices.
490  */
491 struct new_mapping;
492
493 struct pool_features {
494         unsigned zero_new_blocks:1;
495         unsigned discard_enabled:1;
496         unsigned discard_passdown:1;
497 };
498
499 struct pool {
500         struct list_head list;
501         struct dm_target *ti;   /* Only set if a pool target is bound */
502
503         struct mapped_device *pool_md;
504         struct block_device *md_dev;
505         struct dm_pool_metadata *pmd;
506
507         uint32_t sectors_per_block;
508         unsigned block_shift;
509         dm_block_t offset_mask;
510         dm_block_t low_water_blocks;
511
512         struct pool_features pf;
513         unsigned low_water_triggered:1; /* A dm event has been sent */
514         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
515
516         struct bio_prison *prison;
517         struct dm_kcopyd_client *copier;
518
519         struct workqueue_struct *wq;
520         struct work_struct worker;
521         struct delayed_work waker;
522
523         unsigned ref_count;
524         unsigned long last_commit_jiffies;
525
526         spinlock_t lock;
527         struct bio_list deferred_bios;
528         struct bio_list deferred_flush_bios;
529         struct list_head prepared_mappings;
530         struct list_head prepared_discards;
531
532         struct bio_list retry_on_resume_list;
533
534         struct deferred_set shared_read_ds;
535         struct deferred_set all_io_ds;
536
537         struct new_mapping *next_mapping;
538         mempool_t *mapping_pool;
539         mempool_t *endio_hook_pool;
540 };
541
542 /*
543  * Target context for a pool.
544  */
545 struct pool_c {
546         struct dm_target *ti;
547         struct pool *pool;
548         struct dm_dev *data_dev;
549         struct dm_dev *metadata_dev;
550         struct dm_target_callbacks callbacks;
551
552         dm_block_t low_water_blocks;
553         struct pool_features pf;
554 };
555
556 /*
557  * Target context for a thin.
558  */
559 struct thin_c {
560         struct dm_dev *pool_dev;
561         struct dm_dev *origin_dev;
562         dm_thin_id dev_id;
563
564         struct pool *pool;
565         struct dm_thin_device *td;
566 };
567
568 /*----------------------------------------------------------------*/
569
570 /*
571  * A global list of pools that uses a struct mapped_device as a key.
572  */
573 static struct dm_thin_pool_table {
574         struct mutex mutex;
575         struct list_head pools;
576 } dm_thin_pool_table;
577
578 static void pool_table_init(void)
579 {
580         mutex_init(&dm_thin_pool_table.mutex);
581         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
582 }
583
584 static void __pool_table_insert(struct pool *pool)
585 {
586         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
587         list_add(&pool->list, &dm_thin_pool_table.pools);
588 }
589
590 static void __pool_table_remove(struct pool *pool)
591 {
592         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
593         list_del(&pool->list);
594 }
595
596 static struct pool *__pool_table_lookup(struct mapped_device *md)
597 {
598         struct pool *pool = NULL, *tmp;
599
600         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
601
602         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
603                 if (tmp->pool_md == md) {
604                         pool = tmp;
605                         break;
606                 }
607         }
608
609         return pool;
610 }
611
612 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
613 {
614         struct pool *pool = NULL, *tmp;
615
616         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
617
618         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
619                 if (tmp->md_dev == md_dev) {
620                         pool = tmp;
621                         break;
622                 }
623         }
624
625         return pool;
626 }
627
628 /*----------------------------------------------------------------*/
629
630 struct endio_hook {
631         struct thin_c *tc;
632         struct deferred_entry *shared_read_entry;
633         struct deferred_entry *all_io_entry;
634         struct new_mapping *overwrite_mapping;
635 };
636
637 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
638 {
639         struct bio *bio;
640         struct bio_list bios;
641
642         bio_list_init(&bios);
643         bio_list_merge(&bios, master);
644         bio_list_init(master);
645
646         while ((bio = bio_list_pop(&bios))) {
647                 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
648                 if (h->tc == tc)
649                         bio_endio(bio, DM_ENDIO_REQUEUE);
650                 else
651                         bio_list_add(master, bio);
652         }
653 }
654
655 static void requeue_io(struct thin_c *tc)
656 {
657         struct pool *pool = tc->pool;
658         unsigned long flags;
659
660         spin_lock_irqsave(&pool->lock, flags);
661         __requeue_bio_list(tc, &pool->deferred_bios);
662         __requeue_bio_list(tc, &pool->retry_on_resume_list);
663         spin_unlock_irqrestore(&pool->lock, flags);
664 }
665
666 /*
667  * This section of code contains the logic for processing a thin device's IO.
668  * Much of the code depends on pool object resources (lists, workqueues, etc)
669  * but most is exclusively called from the thin target rather than the thin-pool
670  * target.
671  */
672
673 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
674 {
675         return bio->bi_sector >> tc->pool->block_shift;
676 }
677
678 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
679 {
680         struct pool *pool = tc->pool;
681
682         bio->bi_bdev = tc->pool_dev->bdev;
683         bio->bi_sector = (block << pool->block_shift) +
684                 (bio->bi_sector & pool->offset_mask);
685 }
686
687 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
688 {
689         bio->bi_bdev = tc->origin_dev->bdev;
690 }
691
692 static void issue(struct thin_c *tc, struct bio *bio)
693 {
694         struct pool *pool = tc->pool;
695         unsigned long flags;
696
697         /*
698          * Batch together any FUA/FLUSH bios we find and then issue
699          * a single commit for them in process_deferred_bios().
700          */
701         if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
702                 spin_lock_irqsave(&pool->lock, flags);
703                 bio_list_add(&pool->deferred_flush_bios, bio);
704                 spin_unlock_irqrestore(&pool->lock, flags);
705         } else
706                 generic_make_request(bio);
707 }
708
709 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
710 {
711         remap_to_origin(tc, bio);
712         issue(tc, bio);
713 }
714
715 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
716                             dm_block_t block)
717 {
718         remap(tc, bio, block);
719         issue(tc, bio);
720 }
721
722 /*
723  * wake_worker() is used when new work is queued and when pool_resume is
724  * ready to continue deferred IO processing.
725  */
726 static void wake_worker(struct pool *pool)
727 {
728         queue_work(pool->wq, &pool->worker);
729 }
730
731 /*----------------------------------------------------------------*/
732
733 /*
734  * Bio endio functions.
735  */
736 struct new_mapping {
737         struct list_head list;
738
739         unsigned quiesced:1;
740         unsigned prepared:1;
741         unsigned pass_discard:1;
742
743         struct thin_c *tc;
744         dm_block_t virt_block;
745         dm_block_t data_block;
746         struct cell *cell, *cell2;
747         int err;
748
749         /*
750          * If the bio covers the whole area of a block then we can avoid
751          * zeroing or copying.  Instead this bio is hooked.  The bio will
752          * still be in the cell, so care has to be taken to avoid issuing
753          * the bio twice.
754          */
755         struct bio *bio;
756         bio_end_io_t *saved_bi_end_io;
757 };
758
759 static void __maybe_add_mapping(struct new_mapping *m)
760 {
761         struct pool *pool = m->tc->pool;
762
763         if (m->quiesced && m->prepared) {
764                 list_add(&m->list, &pool->prepared_mappings);
765                 wake_worker(pool);
766         }
767 }
768
769 static void copy_complete(int read_err, unsigned long write_err, void *context)
770 {
771         unsigned long flags;
772         struct new_mapping *m = context;
773         struct pool *pool = m->tc->pool;
774
775         m->err = read_err || write_err ? -EIO : 0;
776
777         spin_lock_irqsave(&pool->lock, flags);
778         m->prepared = 1;
779         __maybe_add_mapping(m);
780         spin_unlock_irqrestore(&pool->lock, flags);
781 }
782
783 static void overwrite_endio(struct bio *bio, int err)
784 {
785         unsigned long flags;
786         struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
787         struct new_mapping *m = h->overwrite_mapping;
788         struct pool *pool = m->tc->pool;
789
790         m->err = err;
791
792         spin_lock_irqsave(&pool->lock, flags);
793         m->prepared = 1;
794         __maybe_add_mapping(m);
795         spin_unlock_irqrestore(&pool->lock, flags);
796 }
797
798 /*----------------------------------------------------------------*/
799
800 /*
801  * Workqueue.
802  */
803
804 /*
805  * Prepared mapping jobs.
806  */
807
808 /*
809  * This sends the bios in the cell back to the deferred_bios list.
810  */
811 static void cell_defer(struct thin_c *tc, struct cell *cell,
812                        dm_block_t data_block)
813 {
814         struct pool *pool = tc->pool;
815         unsigned long flags;
816
817         spin_lock_irqsave(&pool->lock, flags);
818         cell_release(cell, &pool->deferred_bios);
819         spin_unlock_irqrestore(&tc->pool->lock, flags);
820
821         wake_worker(pool);
822 }
823
824 /*
825  * Same as cell_defer above, except it omits one particular detainee,
826  * a write bio that covers the block and has already been processed.
827  */
828 static void cell_defer_except(struct thin_c *tc, struct cell *cell)
829 {
830         struct bio_list bios;
831         struct pool *pool = tc->pool;
832         unsigned long flags;
833
834         bio_list_init(&bios);
835
836         spin_lock_irqsave(&pool->lock, flags);
837         cell_release_no_holder(cell, &pool->deferred_bios);
838         spin_unlock_irqrestore(&pool->lock, flags);
839
840         wake_worker(pool);
841 }
842
843 static void process_prepared_mapping(struct new_mapping *m)
844 {
845         struct thin_c *tc = m->tc;
846         struct bio *bio;
847         int r;
848
849         bio = m->bio;
850         if (bio)
851                 bio->bi_end_io = m->saved_bi_end_io;
852
853         if (m->err) {
854                 cell_error(m->cell);
855                 return;
856         }
857
858         /*
859          * Commit the prepared block into the mapping btree.
860          * Any I/O for this block arriving after this point will get
861          * remapped to it directly.
862          */
863         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
864         if (r) {
865                 DMERR("dm_thin_insert_block() failed");
866                 cell_error(m->cell);
867                 return;
868         }
869
870         /*
871          * Release any bios held while the block was being provisioned.
872          * If we are processing a write bio that completely covers the block,
873          * we already processed it so can ignore it now when processing
874          * the bios in the cell.
875          */
876         if (bio) {
877                 cell_defer_except(tc, m->cell);
878                 bio_endio(bio, 0);
879         } else
880                 cell_defer(tc, m->cell, m->data_block);
881
882         list_del(&m->list);
883         mempool_free(m, tc->pool->mapping_pool);
884 }
885
886 static void process_prepared_discard(struct new_mapping *m)
887 {
888         int r;
889         struct thin_c *tc = m->tc;
890
891         r = dm_thin_remove_block(tc->td, m->virt_block);
892         if (r)
893                 DMERR("dm_thin_remove_block() failed");
894
895         /*
896          * Pass the discard down to the underlying device?
897          */
898         if (m->pass_discard)
899                 remap_and_issue(tc, m->bio, m->data_block);
900         else
901                 bio_endio(m->bio, 0);
902
903         cell_defer_except(tc, m->cell);
904         cell_defer_except(tc, m->cell2);
905         mempool_free(m, tc->pool->mapping_pool);
906 }
907
908 static void process_prepared(struct pool *pool, struct list_head *head,
909                              void (*fn)(struct new_mapping *))
910 {
911         unsigned long flags;
912         struct list_head maps;
913         struct new_mapping *m, *tmp;
914
915         INIT_LIST_HEAD(&maps);
916         spin_lock_irqsave(&pool->lock, flags);
917         list_splice_init(head, &maps);
918         spin_unlock_irqrestore(&pool->lock, flags);
919
920         list_for_each_entry_safe(m, tmp, &maps, list)
921                 fn(m);
922 }
923
924 /*
925  * Deferred bio jobs.
926  */
927 static int io_overlaps_block(struct pool *pool, struct bio *bio)
928 {
929         return !(bio->bi_sector & pool->offset_mask) &&
930                 (bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
931
932 }
933
934 static int io_overwrites_block(struct pool *pool, struct bio *bio)
935 {
936         return (bio_data_dir(bio) == WRITE) &&
937                 io_overlaps_block(pool, bio);
938 }
939
940 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
941                                bio_end_io_t *fn)
942 {
943         *save = bio->bi_end_io;
944         bio->bi_end_io = fn;
945 }
946
947 static int ensure_next_mapping(struct pool *pool)
948 {
949         if (pool->next_mapping)
950                 return 0;
951
952         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
953
954         return pool->next_mapping ? 0 : -ENOMEM;
955 }
956
957 static struct new_mapping *get_next_mapping(struct pool *pool)
958 {
959         struct new_mapping *r = pool->next_mapping;
960
961         BUG_ON(!pool->next_mapping);
962
963         pool->next_mapping = NULL;
964
965         return r;
966 }
967
968 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
969                           struct dm_dev *origin, dm_block_t data_origin,
970                           dm_block_t data_dest,
971                           struct cell *cell, struct bio *bio)
972 {
973         int r;
974         struct pool *pool = tc->pool;
975         struct new_mapping *m = get_next_mapping(pool);
976
977         INIT_LIST_HEAD(&m->list);
978         m->quiesced = 0;
979         m->prepared = 0;
980         m->tc = tc;
981         m->virt_block = virt_block;
982         m->data_block = data_dest;
983         m->cell = cell;
984         m->err = 0;
985         m->bio = NULL;
986
987         if (!ds_add_work(&pool->shared_read_ds, &m->list))
988                 m->quiesced = 1;
989
990         /*
991          * IO to pool_dev remaps to the pool target's data_dev.
992          *
993          * If the whole block of data is being overwritten, we can issue the
994          * bio immediately. Otherwise we use kcopyd to clone the data first.
995          */
996         if (io_overwrites_block(pool, bio)) {
997                 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
998                 h->overwrite_mapping = m;
999                 m->bio = bio;
1000                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1001                 remap_and_issue(tc, bio, data_dest);
1002         } else {
1003                 struct dm_io_region from, to;
1004
1005                 from.bdev = origin->bdev;
1006                 from.sector = data_origin * pool->sectors_per_block;
1007                 from.count = pool->sectors_per_block;
1008
1009                 to.bdev = tc->pool_dev->bdev;
1010                 to.sector = data_dest * pool->sectors_per_block;
1011                 to.count = pool->sectors_per_block;
1012
1013                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
1014                                    0, copy_complete, m);
1015                 if (r < 0) {
1016                         mempool_free(m, pool->mapping_pool);
1017                         DMERR("dm_kcopyd_copy() failed");
1018                         cell_error(cell);
1019                 }
1020         }
1021 }
1022
1023 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
1024                                    dm_block_t data_origin, dm_block_t data_dest,
1025                                    struct cell *cell, struct bio *bio)
1026 {
1027         schedule_copy(tc, virt_block, tc->pool_dev,
1028                       data_origin, data_dest, cell, bio);
1029 }
1030
1031 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
1032                                    dm_block_t data_dest,
1033                                    struct cell *cell, struct bio *bio)
1034 {
1035         schedule_copy(tc, virt_block, tc->origin_dev,
1036                       virt_block, data_dest, cell, bio);
1037 }
1038
1039 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
1040                           dm_block_t data_block, struct cell *cell,
1041                           struct bio *bio)
1042 {
1043         struct pool *pool = tc->pool;
1044         struct new_mapping *m = get_next_mapping(pool);
1045
1046         INIT_LIST_HEAD(&m->list);
1047         m->quiesced = 1;
1048         m->prepared = 0;
1049         m->tc = tc;
1050         m->virt_block = virt_block;
1051         m->data_block = data_block;
1052         m->cell = cell;
1053         m->err = 0;
1054         m->bio = NULL;
1055
1056         /*
1057          * If the whole block of data is being overwritten or we are not
1058          * zeroing pre-existing data, we can issue the bio immediately.
1059          * Otherwise we use kcopyd to zero the data first.
1060          */
1061         if (!pool->pf.zero_new_blocks)
1062                 process_prepared_mapping(m);
1063
1064         else if (io_overwrites_block(pool, bio)) {
1065                 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1066                 h->overwrite_mapping = m;
1067                 m->bio = bio;
1068                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
1069                 remap_and_issue(tc, bio, data_block);
1070
1071         } else {
1072                 int r;
1073                 struct dm_io_region to;
1074
1075                 to.bdev = tc->pool_dev->bdev;
1076                 to.sector = data_block * pool->sectors_per_block;
1077                 to.count = pool->sectors_per_block;
1078
1079                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
1080                 if (r < 0) {
1081                         mempool_free(m, pool->mapping_pool);
1082                         DMERR("dm_kcopyd_zero() failed");
1083                         cell_error(cell);
1084                 }
1085         }
1086 }
1087
1088 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
1089 {
1090         int r;
1091         dm_block_t free_blocks;
1092         unsigned long flags;
1093         struct pool *pool = tc->pool;
1094
1095         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1096         if (r)
1097                 return r;
1098
1099         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
1100                 DMWARN("%s: reached low water mark, sending event.",
1101                        dm_device_name(pool->pool_md));
1102                 spin_lock_irqsave(&pool->lock, flags);
1103                 pool->low_water_triggered = 1;
1104                 spin_unlock_irqrestore(&pool->lock, flags);
1105                 dm_table_event(pool->ti->table);
1106         }
1107
1108         if (!free_blocks) {
1109                 if (pool->no_free_space)
1110                         return -ENOSPC;
1111                 else {
1112                         /*
1113                          * Try to commit to see if that will free up some
1114                          * more space.
1115                          */
1116                         r = dm_pool_commit_metadata(pool->pmd);
1117                         if (r) {
1118                                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1119                                       __func__, r);
1120                                 return r;
1121                         }
1122
1123                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
1124                         if (r)
1125                                 return r;
1126
1127                         /*
1128                          * If we still have no space we set a flag to avoid
1129                          * doing all this checking and return -ENOSPC.
1130                          */
1131                         if (!free_blocks) {
1132                                 DMWARN("%s: no free space available.",
1133                                        dm_device_name(pool->pool_md));
1134                                 spin_lock_irqsave(&pool->lock, flags);
1135                                 pool->no_free_space = 1;
1136                                 spin_unlock_irqrestore(&pool->lock, flags);
1137                                 return -ENOSPC;
1138                         }
1139                 }
1140         }
1141
1142         r = dm_pool_alloc_data_block(pool->pmd, result);
1143         if (r)
1144                 return r;
1145
1146         return 0;
1147 }
1148
1149 /*
1150  * If we have run out of space, queue bios until the device is
1151  * resumed, presumably after having been reloaded with more space.
1152  */
1153 static void retry_on_resume(struct bio *bio)
1154 {
1155         struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1156         struct thin_c *tc = h->tc;
1157         struct pool *pool = tc->pool;
1158         unsigned long flags;
1159
1160         spin_lock_irqsave(&pool->lock, flags);
1161         bio_list_add(&pool->retry_on_resume_list, bio);
1162         spin_unlock_irqrestore(&pool->lock, flags);
1163 }
1164
1165 static void no_space(struct cell *cell)
1166 {
1167         struct bio *bio;
1168         struct bio_list bios;
1169
1170         bio_list_init(&bios);
1171         cell_release(cell, &bios);
1172
1173         while ((bio = bio_list_pop(&bios)))
1174                 retry_on_resume(bio);
1175 }
1176
1177 static void process_discard(struct thin_c *tc, struct bio *bio)
1178 {
1179         int r;
1180         struct pool *pool = tc->pool;
1181         struct cell *cell, *cell2;
1182         struct cell_key key, key2;
1183         dm_block_t block = get_bio_block(tc, bio);
1184         struct dm_thin_lookup_result lookup_result;
1185         struct new_mapping *m;
1186
1187         build_virtual_key(tc->td, block, &key);
1188         if (bio_detain(tc->pool->prison, &key, bio, &cell))
1189                 return;
1190
1191         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1192         switch (r) {
1193         case 0:
1194                 /*
1195                  * Check nobody is fiddling with this pool block.  This can
1196                  * happen if someone's in the process of breaking sharing
1197                  * on this block.
1198                  */
1199                 build_data_key(tc->td, lookup_result.block, &key2);
1200                 if (bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
1201                         cell_release_singleton(cell, bio);
1202                         break;
1203                 }
1204
1205                 if (io_overlaps_block(pool, bio)) {
1206                         /*
1207                          * IO may still be going to the destination block.  We must
1208                          * quiesce before we can do the removal.
1209                          */
1210                         m = get_next_mapping(pool);
1211                         m->tc = tc;
1212                         m->pass_discard = (!lookup_result.shared) & pool->pf.discard_passdown;
1213                         m->virt_block = block;
1214                         m->data_block = lookup_result.block;
1215                         m->cell = cell;
1216                         m->cell2 = cell2;
1217                         m->err = 0;
1218                         m->bio = bio;
1219
1220                         if (!ds_add_work(&pool->all_io_ds, &m->list)) {
1221                                 list_add(&m->list, &pool->prepared_discards);
1222                                 wake_worker(pool);
1223                         }
1224                 } else {
1225                         /*
1226                          * This path is hit if people are ignoring
1227                          * limits->discard_granularity.  It ignores any
1228                          * part of the discard that is in a subsequent
1229                          * block.
1230                          */
1231                         sector_t offset = bio->bi_sector - (block << pool->block_shift);
1232                         unsigned remaining = (pool->sectors_per_block - offset) << 9;
1233                         bio->bi_size = min(bio->bi_size, remaining);
1234
1235                         cell_release_singleton(cell, bio);
1236                         cell_release_singleton(cell2, bio);
1237                         remap_and_issue(tc, bio, lookup_result.block);
1238                 }
1239                 break;
1240
1241         case -ENODATA:
1242                 /*
1243                  * It isn't provisioned, just forget it.
1244                  */
1245                 cell_release_singleton(cell, bio);
1246                 bio_endio(bio, 0);
1247                 break;
1248
1249         default:
1250                 DMERR("discard: find block unexpectedly returned %d", r);
1251                 cell_release_singleton(cell, bio);
1252                 bio_io_error(bio);
1253                 break;
1254         }
1255 }
1256
1257 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1258                           struct cell_key *key,
1259                           struct dm_thin_lookup_result *lookup_result,
1260                           struct cell *cell)
1261 {
1262         int r;
1263         dm_block_t data_block;
1264
1265         r = alloc_data_block(tc, &data_block);
1266         switch (r) {
1267         case 0:
1268                 schedule_internal_copy(tc, block, lookup_result->block,
1269                                        data_block, cell, bio);
1270                 break;
1271
1272         case -ENOSPC:
1273                 no_space(cell);
1274                 break;
1275
1276         default:
1277                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1278                 cell_error(cell);
1279                 break;
1280         }
1281 }
1282
1283 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1284                                dm_block_t block,
1285                                struct dm_thin_lookup_result *lookup_result)
1286 {
1287         struct cell *cell;
1288         struct pool *pool = tc->pool;
1289         struct cell_key key;
1290
1291         /*
1292          * If cell is already occupied, then sharing is already in the process
1293          * of being broken so we have nothing further to do here.
1294          */
1295         build_data_key(tc->td, lookup_result->block, &key);
1296         if (bio_detain(pool->prison, &key, bio, &cell))
1297                 return;
1298
1299         if (bio_data_dir(bio) == WRITE)
1300                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1301         else {
1302                 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1303
1304                 h->shared_read_entry = ds_inc(&pool->shared_read_ds);
1305
1306                 cell_release_singleton(cell, bio);
1307                 remap_and_issue(tc, bio, lookup_result->block);
1308         }
1309 }
1310
1311 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1312                             struct cell *cell)
1313 {
1314         int r;
1315         dm_block_t data_block;
1316
1317         /*
1318          * Remap empty bios (flushes) immediately, without provisioning.
1319          */
1320         if (!bio->bi_size) {
1321                 cell_release_singleton(cell, bio);
1322                 remap_and_issue(tc, bio, 0);
1323                 return;
1324         }
1325
1326         /*
1327          * Fill read bios with zeroes and complete them immediately.
1328          */
1329         if (bio_data_dir(bio) == READ) {
1330                 zero_fill_bio(bio);
1331                 cell_release_singleton(cell, bio);
1332                 bio_endio(bio, 0);
1333                 return;
1334         }
1335
1336         r = alloc_data_block(tc, &data_block);
1337         switch (r) {
1338         case 0:
1339                 if (tc->origin_dev)
1340                         schedule_external_copy(tc, block, data_block, cell, bio);
1341                 else
1342                         schedule_zero(tc, block, data_block, cell, bio);
1343                 break;
1344
1345         case -ENOSPC:
1346                 no_space(cell);
1347                 break;
1348
1349         default:
1350                 DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
1351                 cell_error(cell);
1352                 break;
1353         }
1354 }
1355
1356 static void process_bio(struct thin_c *tc, struct bio *bio)
1357 {
1358         int r;
1359         dm_block_t block = get_bio_block(tc, bio);
1360         struct cell *cell;
1361         struct cell_key key;
1362         struct dm_thin_lookup_result lookup_result;
1363
1364         /*
1365          * If cell is already occupied, then the block is already
1366          * being provisioned so we have nothing further to do here.
1367          */
1368         build_virtual_key(tc->td, block, &key);
1369         if (bio_detain(tc->pool->prison, &key, bio, &cell))
1370                 return;
1371
1372         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1373         switch (r) {
1374         case 0:
1375                 /*
1376                  * We can release this cell now.  This thread is the only
1377                  * one that puts bios into a cell, and we know there were
1378                  * no preceding bios.
1379                  */
1380                 /*
1381                  * TODO: this will probably have to change when discard goes
1382                  * back in.
1383                  */
1384                 cell_release_singleton(cell, bio);
1385
1386                 if (lookup_result.shared)
1387                         process_shared_bio(tc, bio, block, &lookup_result);
1388                 else
1389                         remap_and_issue(tc, bio, lookup_result.block);
1390                 break;
1391
1392         case -ENODATA:
1393                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1394                         cell_release_singleton(cell, bio);
1395                         remap_to_origin_and_issue(tc, bio);
1396                 } else
1397                         provision_block(tc, bio, block, cell);
1398                 break;
1399
1400         default:
1401                 DMERR("dm_thin_find_block() failed, error = %d", r);
1402                 cell_release_singleton(cell, bio);
1403                 bio_io_error(bio);
1404                 break;
1405         }
1406 }
1407
1408 static int need_commit_due_to_time(struct pool *pool)
1409 {
1410         return jiffies < pool->last_commit_jiffies ||
1411                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1412 }
1413
1414 static void process_deferred_bios(struct pool *pool)
1415 {
1416         unsigned long flags;
1417         struct bio *bio;
1418         struct bio_list bios;
1419         int r;
1420
1421         bio_list_init(&bios);
1422
1423         spin_lock_irqsave(&pool->lock, flags);
1424         bio_list_merge(&bios, &pool->deferred_bios);
1425         bio_list_init(&pool->deferred_bios);
1426         spin_unlock_irqrestore(&pool->lock, flags);
1427
1428         while ((bio = bio_list_pop(&bios))) {
1429                 struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
1430                 struct thin_c *tc = h->tc;
1431
1432                 /*
1433                  * If we've got no free new_mapping structs, and processing
1434                  * this bio might require one, we pause until there are some
1435                  * prepared mappings to process.
1436                  */
1437                 if (ensure_next_mapping(pool)) {
1438                         spin_lock_irqsave(&pool->lock, flags);
1439                         bio_list_merge(&pool->deferred_bios, &bios);
1440                         spin_unlock_irqrestore(&pool->lock, flags);
1441
1442                         break;
1443                 }
1444
1445                 if (bio->bi_rw & REQ_DISCARD)
1446                         process_discard(tc, bio);
1447                 else
1448                         process_bio(tc, bio);
1449         }
1450
1451         /*
1452          * If there are any deferred flush bios, we must commit
1453          * the metadata before issuing them.
1454          */
1455         bio_list_init(&bios);
1456         spin_lock_irqsave(&pool->lock, flags);
1457         bio_list_merge(&bios, &pool->deferred_flush_bios);
1458         bio_list_init(&pool->deferred_flush_bios);
1459         spin_unlock_irqrestore(&pool->lock, flags);
1460
1461         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1462                 return;
1463
1464         r = dm_pool_commit_metadata(pool->pmd);
1465         if (r) {
1466                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
1467                       __func__, r);
1468                 while ((bio = bio_list_pop(&bios)))
1469                         bio_io_error(bio);
1470                 return;
1471         }
1472         pool->last_commit_jiffies = jiffies;
1473
1474         while ((bio = bio_list_pop(&bios)))
1475                 generic_make_request(bio);
1476 }
1477
1478 static void do_worker(struct work_struct *ws)
1479 {
1480         struct pool *pool = container_of(ws, struct pool, worker);
1481
1482         process_prepared(pool, &pool->prepared_mappings, process_prepared_mapping);
1483         process_prepared(pool, &pool->prepared_discards, process_prepared_discard);
1484         process_deferred_bios(pool);
1485 }
1486
1487 /*
1488  * We want to commit periodically so that not too much
1489  * unwritten data builds up.
1490  */
1491 static void do_waker(struct work_struct *ws)
1492 {
1493         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1494         wake_worker(pool);
1495         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1496 }
1497
1498 /*----------------------------------------------------------------*/
1499
1500 /*
1501  * Mapping functions.
1502  */
1503
1504 /*
1505  * Called only while mapping a thin bio to hand it over to the workqueue.
1506  */
1507 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1508 {
1509         unsigned long flags;
1510         struct pool *pool = tc->pool;
1511
1512         spin_lock_irqsave(&pool->lock, flags);
1513         bio_list_add(&pool->deferred_bios, bio);
1514         spin_unlock_irqrestore(&pool->lock, flags);
1515
1516         wake_worker(pool);
1517 }
1518
1519 static struct endio_hook *thin_hook_bio(struct thin_c *tc, struct bio *bio)
1520 {
1521         struct pool *pool = tc->pool;
1522         struct endio_hook *h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
1523
1524         h->tc = tc;
1525         h->shared_read_entry = NULL;
1526         h->all_io_entry = bio->bi_rw & REQ_DISCARD ? NULL : ds_inc(&pool->all_io_ds);
1527         h->overwrite_mapping = NULL;
1528
1529         return h;
1530 }
1531
1532 /*
1533  * Non-blocking function called from the thin target's map function.
1534  */
1535 static int thin_bio_map(struct dm_target *ti, struct bio *bio,
1536                         union map_info *map_context)
1537 {
1538         int r;
1539         struct thin_c *tc = ti->private;
1540         dm_block_t block = get_bio_block(tc, bio);
1541         struct dm_thin_device *td = tc->td;
1542         struct dm_thin_lookup_result result;
1543
1544         map_context->ptr = thin_hook_bio(tc, bio);
1545         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1546                 thin_defer_bio(tc, bio);
1547                 return DM_MAPIO_SUBMITTED;
1548         }
1549
1550         r = dm_thin_find_block(td, block, 0, &result);
1551
1552         /*
1553          * Note that we defer readahead too.
1554          */
1555         switch (r) {
1556         case 0:
1557                 if (unlikely(result.shared)) {
1558                         /*
1559                          * We have a race condition here between the
1560                          * result.shared value returned by the lookup and
1561                          * snapshot creation, which may cause new
1562                          * sharing.
1563                          *
1564                          * To avoid this always quiesce the origin before
1565                          * taking the snap.  You want to do this anyway to
1566                          * ensure a consistent application view
1567                          * (i.e. lockfs).
1568                          *
1569                          * More distant ancestors are irrelevant. The
1570                          * shared flag will be set in their case.
1571                          */
1572                         thin_defer_bio(tc, bio);
1573                         r = DM_MAPIO_SUBMITTED;
1574                 } else {
1575                         remap(tc, bio, result.block);
1576                         r = DM_MAPIO_REMAPPED;
1577                 }
1578                 break;
1579
1580         case -ENODATA:
1581                 /*
1582                  * In future, the failed dm_thin_find_block above could
1583                  * provide the hint to load the metadata into cache.
1584                  */
1585         case -EWOULDBLOCK:
1586                 thin_defer_bio(tc, bio);
1587                 r = DM_MAPIO_SUBMITTED;
1588                 break;
1589         }
1590
1591         return r;
1592 }
1593
1594 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1595 {
1596         int r;
1597         unsigned long flags;
1598         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1599
1600         spin_lock_irqsave(&pt->pool->lock, flags);
1601         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1602         spin_unlock_irqrestore(&pt->pool->lock, flags);
1603
1604         if (!r) {
1605                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1606                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1607         }
1608
1609         return r;
1610 }
1611
1612 static void __requeue_bios(struct pool *pool)
1613 {
1614         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1615         bio_list_init(&pool->retry_on_resume_list);
1616 }
1617
1618 /*----------------------------------------------------------------
1619  * Binding of control targets to a pool object
1620  *--------------------------------------------------------------*/
1621 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1622 {
1623         struct pool_c *pt = ti->private;
1624
1625         pool->ti = ti;
1626         pool->low_water_blocks = pt->low_water_blocks;
1627         pool->pf = pt->pf;
1628
1629         return 0;
1630 }
1631
1632 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1633 {
1634         if (pool->ti == ti)
1635                 pool->ti = NULL;
1636 }
1637
1638 /*----------------------------------------------------------------
1639  * Pool creation
1640  *--------------------------------------------------------------*/
1641 /* Initialize pool features. */
1642 static void pool_features_init(struct pool_features *pf)
1643 {
1644         pf->zero_new_blocks = 1;
1645         pf->discard_enabled = 1;
1646         pf->discard_passdown = 1;
1647 }
1648
1649 static void __pool_destroy(struct pool *pool)
1650 {
1651         __pool_table_remove(pool);
1652
1653         if (dm_pool_metadata_close(pool->pmd) < 0)
1654                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1655
1656         prison_destroy(pool->prison);
1657         dm_kcopyd_client_destroy(pool->copier);
1658
1659         if (pool->wq)
1660                 destroy_workqueue(pool->wq);
1661
1662         if (pool->next_mapping)
1663                 mempool_free(pool->next_mapping, pool->mapping_pool);
1664         mempool_destroy(pool->mapping_pool);
1665         mempool_destroy(pool->endio_hook_pool);
1666         kfree(pool);
1667 }
1668
1669 static struct pool *pool_create(struct mapped_device *pool_md,
1670                                 struct block_device *metadata_dev,
1671                                 unsigned long block_size, char **error)
1672 {
1673         int r;
1674         void *err_p;
1675         struct pool *pool;
1676         struct dm_pool_metadata *pmd;
1677
1678         pmd = dm_pool_metadata_open(metadata_dev, block_size);
1679         if (IS_ERR(pmd)) {
1680                 *error = "Error creating metadata object";
1681                 return (struct pool *)pmd;
1682         }
1683
1684         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1685         if (!pool) {
1686                 *error = "Error allocating memory for pool";
1687                 err_p = ERR_PTR(-ENOMEM);
1688                 goto bad_pool;
1689         }
1690
1691         pool->pmd = pmd;
1692         pool->sectors_per_block = block_size;
1693         pool->block_shift = ffs(block_size) - 1;
1694         pool->offset_mask = block_size - 1;
1695         pool->low_water_blocks = 0;
1696         pool_features_init(&pool->pf);
1697         pool->prison = prison_create(PRISON_CELLS);
1698         if (!pool->prison) {
1699                 *error = "Error creating pool's bio prison";
1700                 err_p = ERR_PTR(-ENOMEM);
1701                 goto bad_prison;
1702         }
1703
1704         pool->copier = dm_kcopyd_client_create();
1705         if (IS_ERR(pool->copier)) {
1706                 r = PTR_ERR(pool->copier);
1707                 *error = "Error creating pool's kcopyd client";
1708                 err_p = ERR_PTR(r);
1709                 goto bad_kcopyd_client;
1710         }
1711
1712         /*
1713          * Create singlethreaded workqueue that will service all devices
1714          * that use this metadata.
1715          */
1716         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1717         if (!pool->wq) {
1718                 *error = "Error creating pool's workqueue";
1719                 err_p = ERR_PTR(-ENOMEM);
1720                 goto bad_wq;
1721         }
1722
1723         INIT_WORK(&pool->worker, do_worker);
1724         INIT_DELAYED_WORK(&pool->waker, do_waker);
1725         spin_lock_init(&pool->lock);
1726         bio_list_init(&pool->deferred_bios);
1727         bio_list_init(&pool->deferred_flush_bios);
1728         INIT_LIST_HEAD(&pool->prepared_mappings);
1729         INIT_LIST_HEAD(&pool->prepared_discards);
1730         pool->low_water_triggered = 0;
1731         pool->no_free_space = 0;
1732         bio_list_init(&pool->retry_on_resume_list);
1733         ds_init(&pool->shared_read_ds);
1734         ds_init(&pool->all_io_ds);
1735
1736         pool->next_mapping = NULL;
1737         pool->mapping_pool =
1738                 mempool_create_kmalloc_pool(MAPPING_POOL_SIZE, sizeof(struct new_mapping));
1739         if (!pool->mapping_pool) {
1740                 *error = "Error creating pool's mapping mempool";
1741                 err_p = ERR_PTR(-ENOMEM);
1742                 goto bad_mapping_pool;
1743         }
1744
1745         pool->endio_hook_pool =
1746                 mempool_create_kmalloc_pool(ENDIO_HOOK_POOL_SIZE, sizeof(struct endio_hook));
1747         if (!pool->endio_hook_pool) {
1748                 *error = "Error creating pool's endio_hook mempool";
1749                 err_p = ERR_PTR(-ENOMEM);
1750                 goto bad_endio_hook_pool;
1751         }
1752         pool->ref_count = 1;
1753         pool->last_commit_jiffies = jiffies;
1754         pool->pool_md = pool_md;
1755         pool->md_dev = metadata_dev;
1756         __pool_table_insert(pool);
1757
1758         return pool;
1759
1760 bad_endio_hook_pool:
1761         mempool_destroy(pool->mapping_pool);
1762 bad_mapping_pool:
1763         destroy_workqueue(pool->wq);
1764 bad_wq:
1765         dm_kcopyd_client_destroy(pool->copier);
1766 bad_kcopyd_client:
1767         prison_destroy(pool->prison);
1768 bad_prison:
1769         kfree(pool);
1770 bad_pool:
1771         if (dm_pool_metadata_close(pmd))
1772                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1773
1774         return err_p;
1775 }
1776
1777 static void __pool_inc(struct pool *pool)
1778 {
1779         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1780         pool->ref_count++;
1781 }
1782
1783 static void __pool_dec(struct pool *pool)
1784 {
1785         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1786         BUG_ON(!pool->ref_count);
1787         if (!--pool->ref_count)
1788                 __pool_destroy(pool);
1789 }
1790
1791 static struct pool *__pool_find(struct mapped_device *pool_md,
1792                                 struct block_device *metadata_dev,
1793                                 unsigned long block_size, char **error,
1794                                 int *created)
1795 {
1796         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1797
1798         if (pool) {
1799                 if (pool->pool_md != pool_md)
1800                         return ERR_PTR(-EBUSY);
1801                 __pool_inc(pool);
1802
1803         } else {
1804                 pool = __pool_table_lookup(pool_md);
1805                 if (pool) {
1806                         if (pool->md_dev != metadata_dev)
1807                                 return ERR_PTR(-EINVAL);
1808                         __pool_inc(pool);
1809
1810                 } else {
1811                         pool = pool_create(pool_md, metadata_dev, block_size, error);
1812                         *created = 1;
1813                 }
1814         }
1815
1816         return pool;
1817 }
1818
1819 /*----------------------------------------------------------------
1820  * Pool target methods
1821  *--------------------------------------------------------------*/
1822 static void pool_dtr(struct dm_target *ti)
1823 {
1824         struct pool_c *pt = ti->private;
1825
1826         mutex_lock(&dm_thin_pool_table.mutex);
1827
1828         unbind_control_target(pt->pool, ti);
1829         __pool_dec(pt->pool);
1830         dm_put_device(ti, pt->metadata_dev);
1831         dm_put_device(ti, pt->data_dev);
1832         kfree(pt);
1833
1834         mutex_unlock(&dm_thin_pool_table.mutex);
1835 }
1836
1837 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1838                                struct dm_target *ti)
1839 {
1840         int r;
1841         unsigned argc;
1842         const char *arg_name;
1843
1844         static struct dm_arg _args[] = {
1845                 {0, 3, "Invalid number of pool feature arguments"},
1846         };
1847
1848         /*
1849          * No feature arguments supplied.
1850          */
1851         if (!as->argc)
1852                 return 0;
1853
1854         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1855         if (r)
1856                 return -EINVAL;
1857
1858         while (argc && !r) {
1859                 arg_name = dm_shift_arg(as);
1860                 argc--;
1861
1862                 if (!strcasecmp(arg_name, "skip_block_zeroing")) {
1863                         pf->zero_new_blocks = 0;
1864                         continue;
1865                 } else if (!strcasecmp(arg_name, "ignore_discard")) {
1866                         pf->discard_enabled = 0;
1867                         continue;
1868                 } else if (!strcasecmp(arg_name, "no_discard_passdown")) {
1869                         pf->discard_passdown = 0;
1870                         continue;
1871                 }
1872
1873                 ti->error = "Unrecognised pool feature requested";
1874                 r = -EINVAL;
1875         }
1876
1877         return r;
1878 }
1879
1880 /*
1881  * thin-pool <metadata dev> <data dev>
1882  *           <data block size (sectors)>
1883  *           <low water mark (blocks)>
1884  *           [<#feature args> [<arg>]*]
1885  *
1886  * Optional feature arguments are:
1887  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1888  *           ignore_discard: disable discard
1889  *           no_discard_passdown: don't pass discards down to the data device
1890  */
1891 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1892 {
1893         int r, pool_created = 0;
1894         struct pool_c *pt;
1895         struct pool *pool;
1896         struct pool_features pf;
1897         struct dm_arg_set as;
1898         struct dm_dev *data_dev;
1899         unsigned long block_size;
1900         dm_block_t low_water_blocks;
1901         struct dm_dev *metadata_dev;
1902         sector_t metadata_dev_size;
1903         char b[BDEVNAME_SIZE];
1904
1905         /*
1906          * FIXME Remove validation from scope of lock.
1907          */
1908         mutex_lock(&dm_thin_pool_table.mutex);
1909
1910         if (argc < 4) {
1911                 ti->error = "Invalid argument count";
1912                 r = -EINVAL;
1913                 goto out_unlock;
1914         }
1915         as.argc = argc;
1916         as.argv = argv;
1917
1918         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1919         if (r) {
1920                 ti->error = "Error opening metadata block device";
1921                 goto out_unlock;
1922         }
1923
1924         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1925         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1926                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1927                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1928
1929         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1930         if (r) {
1931                 ti->error = "Error getting data device";
1932                 goto out_metadata;
1933         }
1934
1935         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1936             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1937             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1938             !is_power_of_2(block_size)) {
1939                 ti->error = "Invalid block size";
1940                 r = -EINVAL;
1941                 goto out;
1942         }
1943
1944         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1945                 ti->error = "Invalid low water mark";
1946                 r = -EINVAL;
1947                 goto out;
1948         }
1949
1950         /*
1951          * Set default pool features.
1952          */
1953         pool_features_init(&pf);
1954
1955         dm_consume_args(&as, 4);
1956         r = parse_pool_features(&as, &pf, ti);
1957         if (r)
1958                 goto out;
1959
1960         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1961         if (!pt) {
1962                 r = -ENOMEM;
1963                 goto out;
1964         }
1965
1966         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1967                            block_size, &ti->error, &pool_created);
1968         if (IS_ERR(pool)) {
1969                 r = PTR_ERR(pool);
1970                 goto out_free_pt;
1971         }
1972
1973         /*
1974          * 'pool_created' reflects whether this is the first table load.
1975          * Top level discard support is not allowed to be changed after
1976          * initial load.  This would require a pool reload to trigger thin
1977          * device changes.
1978          */
1979         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1980                 ti->error = "Discard support cannot be disabled once enabled";
1981                 r = -EINVAL;
1982                 goto out_flags_changed;
1983         }
1984
1985         /*
1986          * If discard_passdown was enabled verify that the data device
1987          * supports discards.  Disable discard_passdown if not; otherwise
1988          * -EOPNOTSUPP will be returned.
1989          */
1990         if (pf.discard_passdown) {
1991                 struct request_queue *q = bdev_get_queue(data_dev->bdev);
1992                 if (!q || !blk_queue_discard(q)) {
1993                         DMWARN("Discard unsupported by data device: Disabling discard passdown.");
1994                         pf.discard_passdown = 0;
1995                 }
1996         }
1997
1998         pt->pool = pool;
1999         pt->ti = ti;
2000         pt->metadata_dev = metadata_dev;
2001         pt->data_dev = data_dev;
2002         pt->low_water_blocks = low_water_blocks;
2003         pt->pf = pf;
2004         ti->num_flush_requests = 1;
2005         /*
2006          * Only need to enable discards if the pool should pass
2007          * them down to the data device.  The thin device's discard
2008          * processing will cause mappings to be removed from the btree.
2009          */
2010         if (pf.discard_enabled && pf.discard_passdown) {
2011                 ti->num_discard_requests = 1;
2012                 /*
2013                  * Setting 'discards_supported' circumvents the normal
2014                  * stacking of discard limits (this keeps the pool and
2015                  * thin devices' discard limits consistent).
2016                  */
2017                 ti->discards_supported = 1;
2018         }
2019         ti->private = pt;
2020
2021         pt->callbacks.congested_fn = pool_is_congested;
2022         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
2023
2024         mutex_unlock(&dm_thin_pool_table.mutex);
2025
2026         return 0;
2027
2028 out_flags_changed:
2029         __pool_dec(pool);
2030 out_free_pt:
2031         kfree(pt);
2032 out:
2033         dm_put_device(ti, data_dev);
2034 out_metadata:
2035         dm_put_device(ti, metadata_dev);
2036 out_unlock:
2037         mutex_unlock(&dm_thin_pool_table.mutex);
2038
2039         return r;
2040 }
2041
2042 static int pool_map(struct dm_target *ti, struct bio *bio,
2043                     union map_info *map_context)
2044 {
2045         int r;
2046         struct pool_c *pt = ti->private;
2047         struct pool *pool = pt->pool;
2048         unsigned long flags;
2049
2050         /*
2051          * As this is a singleton target, ti->begin is always zero.
2052          */
2053         spin_lock_irqsave(&pool->lock, flags);
2054         bio->bi_bdev = pt->data_dev->bdev;
2055         r = DM_MAPIO_REMAPPED;
2056         spin_unlock_irqrestore(&pool->lock, flags);
2057
2058         return r;
2059 }
2060
2061 /*
2062  * Retrieves the number of blocks of the data device from
2063  * the superblock and compares it to the actual device size,
2064  * thus resizing the data device in case it has grown.
2065  *
2066  * This both copes with opening preallocated data devices in the ctr
2067  * being followed by a resume
2068  * -and-
2069  * calling the resume method individually after userspace has
2070  * grown the data device in reaction to a table event.
2071  */
2072 static int pool_preresume(struct dm_target *ti)
2073 {
2074         int r;
2075         struct pool_c *pt = ti->private;
2076         struct pool *pool = pt->pool;
2077         dm_block_t data_size, sb_data_size;
2078
2079         /*
2080          * Take control of the pool object.
2081          */
2082         r = bind_control_target(pool, ti);
2083         if (r)
2084                 return r;
2085
2086         data_size = ti->len >> pool->block_shift;
2087         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2088         if (r) {
2089                 DMERR("failed to retrieve data device size");
2090                 return r;
2091         }
2092
2093         if (data_size < sb_data_size) {
2094                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2095                       data_size, sb_data_size);
2096                 return -EINVAL;
2097
2098         } else if (data_size > sb_data_size) {
2099                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2100                 if (r) {
2101                         DMERR("failed to resize data device");
2102                         return r;
2103                 }
2104
2105                 r = dm_pool_commit_metadata(pool->pmd);
2106                 if (r) {
2107                         DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2108                               __func__, r);
2109                         return r;
2110                 }
2111         }
2112
2113         return 0;
2114 }
2115
2116 static void pool_resume(struct dm_target *ti)
2117 {
2118         struct pool_c *pt = ti->private;
2119         struct pool *pool = pt->pool;
2120         unsigned long flags;
2121
2122         spin_lock_irqsave(&pool->lock, flags);
2123         pool->low_water_triggered = 0;
2124         pool->no_free_space = 0;
2125         __requeue_bios(pool);
2126         spin_unlock_irqrestore(&pool->lock, flags);
2127
2128         do_waker(&pool->waker.work);
2129 }
2130
2131 static void pool_postsuspend(struct dm_target *ti)
2132 {
2133         int r;
2134         struct pool_c *pt = ti->private;
2135         struct pool *pool = pt->pool;
2136
2137         cancel_delayed_work(&pool->waker);
2138         flush_workqueue(pool->wq);
2139
2140         r = dm_pool_commit_metadata(pool->pmd);
2141         if (r < 0) {
2142                 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
2143                       __func__, r);
2144                 /* FIXME: invalidate device? error the next FUA or FLUSH bio ?*/
2145         }
2146 }
2147
2148 static int check_arg_count(unsigned argc, unsigned args_required)
2149 {
2150         if (argc != args_required) {
2151                 DMWARN("Message received with %u arguments instead of %u.",
2152                        argc, args_required);
2153                 return -EINVAL;
2154         }
2155
2156         return 0;
2157 }
2158
2159 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2160 {
2161         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2162             *dev_id <= MAX_DEV_ID)
2163                 return 0;
2164
2165         if (warning)
2166                 DMWARN("Message received with invalid device id: %s", arg);
2167
2168         return -EINVAL;
2169 }
2170
2171 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2172 {
2173         dm_thin_id dev_id;
2174         int r;
2175
2176         r = check_arg_count(argc, 2);
2177         if (r)
2178                 return r;
2179
2180         r = read_dev_id(argv[1], &dev_id, 1);
2181         if (r)
2182                 return r;
2183
2184         r = dm_pool_create_thin(pool->pmd, dev_id);
2185         if (r) {
2186                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2187                        argv[1]);
2188                 return r;
2189         }
2190
2191         return 0;
2192 }
2193
2194 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2195 {
2196         dm_thin_id dev_id;
2197         dm_thin_id origin_dev_id;
2198         int r;
2199
2200         r = check_arg_count(argc, 3);
2201         if (r)
2202                 return r;
2203
2204         r = read_dev_id(argv[1], &dev_id, 1);
2205         if (r)
2206                 return r;
2207
2208         r = read_dev_id(argv[2], &origin_dev_id, 1);
2209         if (r)
2210                 return r;
2211
2212         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2213         if (r) {
2214                 DMWARN("Creation of new snapshot %s of device %s failed.",
2215                        argv[1], argv[2]);
2216                 return r;
2217         }
2218
2219         return 0;
2220 }
2221
2222 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2223 {
2224         dm_thin_id dev_id;
2225         int r;
2226
2227         r = check_arg_count(argc, 2);
2228         if (r)
2229                 return r;
2230
2231         r = read_dev_id(argv[1], &dev_id, 1);
2232         if (r)
2233                 return r;
2234
2235         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2236         if (r)
2237                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2238
2239         return r;
2240 }
2241
2242 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2243 {
2244         dm_thin_id old_id, new_id;
2245         int r;
2246
2247         r = check_arg_count(argc, 3);
2248         if (r)
2249                 return r;
2250
2251         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2252                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2253                 return -EINVAL;
2254         }
2255
2256         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2257                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2258                 return -EINVAL;
2259         }
2260
2261         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2262         if (r) {
2263                 DMWARN("Failed to change transaction id from %s to %s.",
2264                        argv[1], argv[2]);
2265                 return r;
2266         }
2267
2268         return 0;
2269 }
2270
2271 /*
2272  * Messages supported:
2273  *   create_thin        <dev_id>
2274  *   create_snap        <dev_id> <origin_id>
2275  *   delete             <dev_id>
2276  *   trim               <dev_id> <new_size_in_sectors>
2277  *   set_transaction_id <current_trans_id> <new_trans_id>
2278  */
2279 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2280 {
2281         int r = -EINVAL;
2282         struct pool_c *pt = ti->private;
2283         struct pool *pool = pt->pool;
2284
2285         if (!strcasecmp(argv[0], "create_thin"))
2286                 r = process_create_thin_mesg(argc, argv, pool);
2287
2288         else if (!strcasecmp(argv[0], "create_snap"))
2289                 r = process_create_snap_mesg(argc, argv, pool);
2290
2291         else if (!strcasecmp(argv[0], "delete"))
2292                 r = process_delete_mesg(argc, argv, pool);
2293
2294         else if (!strcasecmp(argv[0], "set_transaction_id"))
2295                 r = process_set_transaction_id_mesg(argc, argv, pool);
2296
2297         else
2298                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2299
2300         if (!r) {
2301                 r = dm_pool_commit_metadata(pool->pmd);
2302                 if (r)
2303                         DMERR("%s message: dm_pool_commit_metadata() failed, error = %d",
2304                               argv[0], r);
2305         }
2306
2307         return r;
2308 }
2309
2310 /*
2311  * Status line is:
2312  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2313  *    <used data sectors>/<total data sectors> <held metadata root>
2314  */
2315 static int pool_status(struct dm_target *ti, status_type_t type,
2316                        char *result, unsigned maxlen)
2317 {
2318         int r, count;
2319         unsigned sz = 0;
2320         uint64_t transaction_id;
2321         dm_block_t nr_free_blocks_data;
2322         dm_block_t nr_free_blocks_metadata;
2323         dm_block_t nr_blocks_data;
2324         dm_block_t nr_blocks_metadata;
2325         dm_block_t held_root;
2326         char buf[BDEVNAME_SIZE];
2327         char buf2[BDEVNAME_SIZE];
2328         struct pool_c *pt = ti->private;
2329         struct pool *pool = pt->pool;
2330
2331         switch (type) {
2332         case STATUSTYPE_INFO:
2333                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2334                                                         &transaction_id);
2335                 if (r)
2336                         return r;
2337
2338                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2339                                                           &nr_free_blocks_metadata);
2340                 if (r)
2341                         return r;
2342
2343                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2344                 if (r)
2345                         return r;
2346
2347                 r = dm_pool_get_free_block_count(pool->pmd,
2348                                                  &nr_free_blocks_data);
2349                 if (r)
2350                         return r;
2351
2352                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2353                 if (r)
2354                         return r;
2355
2356                 r = dm_pool_get_held_metadata_root(pool->pmd, &held_root);
2357                 if (r)
2358                         return r;
2359
2360                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2361                        (unsigned long long)transaction_id,
2362                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2363                        (unsigned long long)nr_blocks_metadata,
2364                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2365                        (unsigned long long)nr_blocks_data);
2366
2367                 if (held_root)
2368                         DMEMIT("%llu", held_root);
2369                 else
2370                         DMEMIT("-");
2371
2372                 break;
2373
2374         case STATUSTYPE_TABLE:
2375                 DMEMIT("%s %s %lu %llu ",
2376                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2377                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2378                        (unsigned long)pool->sectors_per_block,
2379                        (unsigned long long)pt->low_water_blocks);
2380
2381                 count = !pool->pf.zero_new_blocks + !pool->pf.discard_enabled +
2382                         !pool->pf.discard_passdown;
2383                 DMEMIT("%u ", count);
2384
2385                 if (!pool->pf.zero_new_blocks)
2386                         DMEMIT("skip_block_zeroing ");
2387
2388                 if (!pool->pf.discard_enabled)
2389                         DMEMIT("ignore_discard ");
2390
2391                 if (!pool->pf.discard_passdown)
2392                         DMEMIT("no_discard_passdown ");
2393
2394                 break;
2395         }
2396
2397         return 0;
2398 }
2399
2400 static int pool_iterate_devices(struct dm_target *ti,
2401                                 iterate_devices_callout_fn fn, void *data)
2402 {
2403         struct pool_c *pt = ti->private;
2404
2405         return fn(ti, pt->data_dev, 0, ti->len, data);
2406 }
2407
2408 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2409                       struct bio_vec *biovec, int max_size)
2410 {
2411         struct pool_c *pt = ti->private;
2412         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2413
2414         if (!q->merge_bvec_fn)
2415                 return max_size;
2416
2417         bvm->bi_bdev = pt->data_dev->bdev;
2418
2419         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2420 }
2421
2422 static void set_discard_limits(struct pool *pool, struct queue_limits *limits)
2423 {
2424         /*
2425          * FIXME: these limits may be incompatible with the pool's data device
2426          */
2427         limits->max_discard_sectors = pool->sectors_per_block;
2428
2429         /*
2430          * This is just a hint, and not enforced.  We have to cope with
2431          * bios that overlap 2 blocks.
2432          */
2433         limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2434         limits->discard_zeroes_data = pool->pf.zero_new_blocks;
2435 }
2436
2437 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2438 {
2439         struct pool_c *pt = ti->private;
2440         struct pool *pool = pt->pool;
2441
2442         blk_limits_io_min(limits, 0);
2443         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2444         if (pool->pf.discard_enabled)
2445                 set_discard_limits(pool, limits);
2446 }
2447
2448 static struct target_type pool_target = {
2449         .name = "thin-pool",
2450         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2451                     DM_TARGET_IMMUTABLE,
2452         .version = {1, 1, 0},
2453         .module = THIS_MODULE,
2454         .ctr = pool_ctr,
2455         .dtr = pool_dtr,
2456         .map = pool_map,
2457         .postsuspend = pool_postsuspend,
2458         .preresume = pool_preresume,
2459         .resume = pool_resume,
2460         .message = pool_message,
2461         .status = pool_status,
2462         .merge = pool_merge,
2463         .iterate_devices = pool_iterate_devices,
2464         .io_hints = pool_io_hints,
2465 };
2466
2467 /*----------------------------------------------------------------
2468  * Thin target methods
2469  *--------------------------------------------------------------*/
2470 static void thin_dtr(struct dm_target *ti)
2471 {
2472         struct thin_c *tc = ti->private;
2473
2474         mutex_lock(&dm_thin_pool_table.mutex);
2475
2476         __pool_dec(tc->pool);
2477         dm_pool_close_thin_device(tc->td);
2478         dm_put_device(ti, tc->pool_dev);
2479         if (tc->origin_dev)
2480                 dm_put_device(ti, tc->origin_dev);
2481         kfree(tc);
2482
2483         mutex_unlock(&dm_thin_pool_table.mutex);
2484 }
2485
2486 /*
2487  * Thin target parameters:
2488  *
2489  * <pool_dev> <dev_id> [origin_dev]
2490  *
2491  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2492  * dev_id: the internal device identifier
2493  * origin_dev: a device external to the pool that should act as the origin
2494  *
2495  * If the pool device has discards disabled, they get disabled for the thin
2496  * device as well.
2497  */
2498 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2499 {
2500         int r;
2501         struct thin_c *tc;
2502         struct dm_dev *pool_dev, *origin_dev;
2503         struct mapped_device *pool_md;
2504
2505         mutex_lock(&dm_thin_pool_table.mutex);
2506
2507         if (argc != 2 && argc != 3) {
2508                 ti->error = "Invalid argument count";
2509                 r = -EINVAL;
2510                 goto out_unlock;
2511         }
2512
2513         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2514         if (!tc) {
2515                 ti->error = "Out of memory";
2516                 r = -ENOMEM;
2517                 goto out_unlock;
2518         }
2519
2520         if (argc == 3) {
2521                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2522                 if (r) {
2523                         ti->error = "Error opening origin device";
2524                         goto bad_origin_dev;
2525                 }
2526                 tc->origin_dev = origin_dev;
2527         }
2528
2529         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2530         if (r) {
2531                 ti->error = "Error opening pool device";
2532                 goto bad_pool_dev;
2533         }
2534         tc->pool_dev = pool_dev;
2535
2536         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2537                 ti->error = "Invalid device id";
2538                 r = -EINVAL;
2539                 goto bad_common;
2540         }
2541
2542         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2543         if (!pool_md) {
2544                 ti->error = "Couldn't get pool mapped device";
2545                 r = -EINVAL;
2546                 goto bad_common;
2547         }
2548
2549         tc->pool = __pool_table_lookup(pool_md);
2550         if (!tc->pool) {
2551                 ti->error = "Couldn't find pool object";
2552                 r = -EINVAL;
2553                 goto bad_pool_lookup;
2554         }
2555         __pool_inc(tc->pool);
2556
2557         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2558         if (r) {
2559                 ti->error = "Couldn't open thin internal device";
2560                 goto bad_thin_open;
2561         }
2562
2563         ti->split_io = tc->pool->sectors_per_block;
2564         ti->num_flush_requests = 1;
2565
2566         /* In case the pool supports discards, pass them on. */
2567         if (tc->pool->pf.discard_enabled) {
2568                 ti->discards_supported = 1;
2569                 ti->num_discard_requests = 1;
2570         }
2571
2572         dm_put(pool_md);
2573
2574         mutex_unlock(&dm_thin_pool_table.mutex);
2575
2576         return 0;
2577
2578 bad_thin_open:
2579         __pool_dec(tc->pool);
2580 bad_pool_lookup:
2581         dm_put(pool_md);
2582 bad_common:
2583         dm_put_device(ti, tc->pool_dev);
2584 bad_pool_dev:
2585         if (tc->origin_dev)
2586                 dm_put_device(ti, tc->origin_dev);
2587 bad_origin_dev:
2588         kfree(tc);
2589 out_unlock:
2590         mutex_unlock(&dm_thin_pool_table.mutex);
2591
2592         return r;
2593 }
2594
2595 static int thin_map(struct dm_target *ti, struct bio *bio,
2596                     union map_info *map_context)
2597 {
2598         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2599
2600         return thin_bio_map(ti, bio, map_context);
2601 }
2602
2603 static int thin_endio(struct dm_target *ti,
2604                       struct bio *bio, int err,
2605                       union map_info *map_context)
2606 {
2607         unsigned long flags;
2608         struct endio_hook *h = map_context->ptr;
2609         struct list_head work;
2610         struct new_mapping *m, *tmp;
2611         struct pool *pool = h->tc->pool;
2612
2613         if (h->shared_read_entry) {
2614                 INIT_LIST_HEAD(&work);
2615                 ds_dec(h->shared_read_entry, &work);
2616
2617                 spin_lock_irqsave(&pool->lock, flags);
2618                 list_for_each_entry_safe(m, tmp, &work, list) {
2619                         list_del(&m->list);
2620                         m->quiesced = 1;
2621                         __maybe_add_mapping(m);
2622                 }
2623                 spin_unlock_irqrestore(&pool->lock, flags);
2624         }
2625
2626         if (h->all_io_entry) {
2627                 INIT_LIST_HEAD(&work);
2628                 ds_dec(h->all_io_entry, &work);
2629                 list_for_each_entry_safe(m, tmp, &work, list)
2630                         list_add(&m->list, &pool->prepared_discards);
2631         }
2632
2633         mempool_free(h, pool->endio_hook_pool);
2634
2635         return 0;
2636 }
2637
2638 static void thin_postsuspend(struct dm_target *ti)
2639 {
2640         if (dm_noflush_suspending(ti))
2641                 requeue_io((struct thin_c *)ti->private);
2642 }
2643
2644 /*
2645  * <nr mapped sectors> <highest mapped sector>
2646  */
2647 static int thin_status(struct dm_target *ti, status_type_t type,
2648                        char *result, unsigned maxlen)
2649 {
2650         int r;
2651         ssize_t sz = 0;
2652         dm_block_t mapped, highest;
2653         char buf[BDEVNAME_SIZE];
2654         struct thin_c *tc = ti->private;
2655
2656         if (!tc->td)
2657                 DMEMIT("-");
2658         else {
2659                 switch (type) {
2660                 case STATUSTYPE_INFO:
2661                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2662                         if (r)
2663                                 return r;
2664
2665                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2666                         if (r < 0)
2667                                 return r;
2668
2669                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2670                         if (r)
2671                                 DMEMIT("%llu", ((highest + 1) *
2672                                                 tc->pool->sectors_per_block) - 1);
2673                         else
2674                                 DMEMIT("-");
2675                         break;
2676
2677                 case STATUSTYPE_TABLE:
2678                         DMEMIT("%s %lu",
2679                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2680                                (unsigned long) tc->dev_id);
2681                         if (tc->origin_dev)
2682                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2683                         break;
2684                 }
2685         }
2686
2687         return 0;
2688 }
2689
2690 static int thin_iterate_devices(struct dm_target *ti,
2691                                 iterate_devices_callout_fn fn, void *data)
2692 {
2693         dm_block_t blocks;
2694         struct thin_c *tc = ti->private;
2695
2696         /*
2697          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2698          * we follow a more convoluted path through to the pool's target.
2699          */
2700         if (!tc->pool->ti)
2701                 return 0;       /* nothing is bound */
2702
2703         blocks = tc->pool->ti->len >> tc->pool->block_shift;
2704         if (blocks)
2705                 return fn(ti, tc->pool_dev, 0, tc->pool->sectors_per_block * blocks, data);
2706
2707         return 0;
2708 }
2709
2710 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2711 {
2712         struct thin_c *tc = ti->private;
2713         struct pool *pool = tc->pool;
2714
2715         blk_limits_io_min(limits, 0);
2716         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2717         set_discard_limits(pool, limits);
2718 }
2719
2720 static struct target_type thin_target = {
2721         .name = "thin",
2722         .version = {1, 1, 0},
2723         .module = THIS_MODULE,
2724         .ctr = thin_ctr,
2725         .dtr = thin_dtr,
2726         .map = thin_map,
2727         .end_io = thin_endio,
2728         .postsuspend = thin_postsuspend,
2729         .status = thin_status,
2730         .iterate_devices = thin_iterate_devices,
2731         .io_hints = thin_io_hints,
2732 };
2733
2734 /*----------------------------------------------------------------*/
2735
2736 static int __init dm_thin_init(void)
2737 {
2738         int r;
2739
2740         pool_table_init();
2741
2742         r = dm_register_target(&thin_target);
2743         if (r)
2744                 return r;
2745
2746         r = dm_register_target(&pool_target);
2747         if (r)
2748                 dm_unregister_target(&thin_target);
2749
2750         return r;
2751 }
2752
2753 static void dm_thin_exit(void)
2754 {
2755         dm_unregister_target(&thin_target);
2756         dm_unregister_target(&pool_target);
2757 }
2758
2759 module_init(dm_thin_init);
2760 module_exit(dm_thin_exit);
2761
2762 MODULE_DESCRIPTION(DM_NAME "device-mapper thin provisioning target");
2763 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2764 MODULE_LICENSE("GPL");