Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
[platform/adaptation/renesas_rcar/renesas_kernel.git] / drivers / md / dm-thin.c
1 /*
2  * Copyright (C) 2011-2012 Red Hat UK.
3  *
4  * This file is released under the GPL.
5  */
6
7 #include "dm-thin-metadata.h"
8 #include "dm-bio-prison.h"
9 #include "dm.h"
10
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/list.h>
15 #include <linux/init.h>
16 #include <linux/module.h>
17 #include <linux/slab.h>
18
19 #define DM_MSG_PREFIX   "thin"
20
21 /*
22  * Tunable constants
23  */
24 #define ENDIO_HOOK_POOL_SIZE 1024
25 #define MAPPING_POOL_SIZE 1024
26 #define PRISON_CELLS 1024
27 #define COMMIT_PERIOD HZ
28
29 /*
30  * The block size of the device holding pool data must be
31  * between 64KB and 1GB.
32  */
33 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
34 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
35
36 /*
37  * Device id is restricted to 24 bits.
38  */
39 #define MAX_DEV_ID ((1 << 24) - 1)
40
41 /*
42  * How do we handle breaking sharing of data blocks?
43  * =================================================
44  *
45  * We use a standard copy-on-write btree to store the mappings for the
46  * devices (note I'm talking about copy-on-write of the metadata here, not
47  * the data).  When you take an internal snapshot you clone the root node
48  * of the origin btree.  After this there is no concept of an origin or a
49  * snapshot.  They are just two device trees that happen to point to the
50  * same data blocks.
51  *
52  * When we get a write in we decide if it's to a shared data block using
53  * some timestamp magic.  If it is, we have to break sharing.
54  *
55  * Let's say we write to a shared block in what was the origin.  The
56  * steps are:
57  *
58  * i) plug io further to this physical block. (see bio_prison code).
59  *
60  * ii) quiesce any read io to that shared data block.  Obviously
61  * including all devices that share this block.  (see dm_deferred_set code)
62  *
63  * iii) copy the data block to a newly allocate block.  This step can be
64  * missed out if the io covers the block. (schedule_copy).
65  *
66  * iv) insert the new mapping into the origin's btree
67  * (process_prepared_mapping).  This act of inserting breaks some
68  * sharing of btree nodes between the two devices.  Breaking sharing only
69  * effects the btree of that specific device.  Btrees for the other
70  * devices that share the block never change.  The btree for the origin
71  * device as it was after the last commit is untouched, ie. we're using
72  * persistent data structures in the functional programming sense.
73  *
74  * v) unplug io to this physical block, including the io that triggered
75  * the breaking of sharing.
76  *
77  * Steps (ii) and (iii) occur in parallel.
78  *
79  * The metadata _doesn't_ need to be committed before the io continues.  We
80  * get away with this because the io is always written to a _new_ block.
81  * If there's a crash, then:
82  *
83  * - The origin mapping will point to the old origin block (the shared
84  * one).  This will contain the data as it was before the io that triggered
85  * the breaking of sharing came in.
86  *
87  * - The snap mapping still points to the old block.  As it would after
88  * the commit.
89  *
90  * The downside of this scheme is the timestamp magic isn't perfect, and
91  * will continue to think that data block in the snapshot device is shared
92  * even after the write to the origin has broken sharing.  I suspect data
93  * blocks will typically be shared by many different devices, so we're
94  * breaking sharing n + 1 times, rather than n, where n is the number of
95  * devices that reference this data block.  At the moment I think the
96  * benefits far, far outweigh the disadvantages.
97  */
98
99 /*----------------------------------------------------------------*/
100
101 /*
102  * Key building.
103  */
104 static void build_data_key(struct dm_thin_device *td,
105                            dm_block_t b, struct dm_cell_key *key)
106 {
107         key->virtual = 0;
108         key->dev = dm_thin_dev_id(td);
109         key->block = b;
110 }
111
112 static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
113                               struct dm_cell_key *key)
114 {
115         key->virtual = 1;
116         key->dev = dm_thin_dev_id(td);
117         key->block = b;
118 }
119
120 /*----------------------------------------------------------------*/
121
122 /*
123  * A pool device ties together a metadata device and a data device.  It
124  * also provides the interface for creating and destroying internal
125  * devices.
126  */
127 struct dm_thin_new_mapping;
128
129 /*
130  * The pool runs in 3 modes.  Ordered in degraded order for comparisons.
131  */
132 enum pool_mode {
133         PM_WRITE,               /* metadata may be changed */
134         PM_READ_ONLY,           /* metadata may not be changed */
135         PM_FAIL,                /* all I/O fails */
136 };
137
138 struct pool_features {
139         enum pool_mode mode;
140
141         bool zero_new_blocks:1;
142         bool discard_enabled:1;
143         bool discard_passdown:1;
144 };
145
146 struct thin_c;
147 typedef void (*process_bio_fn)(struct thin_c *tc, struct bio *bio);
148 typedef void (*process_mapping_fn)(struct dm_thin_new_mapping *m);
149
150 struct pool {
151         struct list_head list;
152         struct dm_target *ti;   /* Only set if a pool target is bound */
153
154         struct mapped_device *pool_md;
155         struct block_device *md_dev;
156         struct dm_pool_metadata *pmd;
157
158         dm_block_t low_water_blocks;
159         uint32_t sectors_per_block;
160         int sectors_per_block_shift;
161
162         struct pool_features pf;
163         unsigned low_water_triggered:1; /* A dm event has been sent */
164         unsigned no_free_space:1;       /* A -ENOSPC warning has been issued */
165
166         struct dm_bio_prison *prison;
167         struct dm_kcopyd_client *copier;
168
169         struct workqueue_struct *wq;
170         struct work_struct worker;
171         struct delayed_work waker;
172
173         unsigned long last_commit_jiffies;
174         unsigned ref_count;
175
176         spinlock_t lock;
177         struct bio_list deferred_bios;
178         struct bio_list deferred_flush_bios;
179         struct list_head prepared_mappings;
180         struct list_head prepared_discards;
181
182         struct bio_list retry_on_resume_list;
183
184         struct dm_deferred_set *shared_read_ds;
185         struct dm_deferred_set *all_io_ds;
186
187         struct dm_thin_new_mapping *next_mapping;
188         mempool_t *mapping_pool;
189
190         process_bio_fn process_bio;
191         process_bio_fn process_discard;
192
193         process_mapping_fn process_prepared_mapping;
194         process_mapping_fn process_prepared_discard;
195 };
196
197 static enum pool_mode get_pool_mode(struct pool *pool);
198 static void set_pool_mode(struct pool *pool, enum pool_mode mode);
199
200 /*
201  * Target context for a pool.
202  */
203 struct pool_c {
204         struct dm_target *ti;
205         struct pool *pool;
206         struct dm_dev *data_dev;
207         struct dm_dev *metadata_dev;
208         struct dm_target_callbacks callbacks;
209
210         dm_block_t low_water_blocks;
211         struct pool_features requested_pf; /* Features requested during table load */
212         struct pool_features adjusted_pf;  /* Features used after adjusting for constituent devices */
213 };
214
215 /*
216  * Target context for a thin.
217  */
218 struct thin_c {
219         struct dm_dev *pool_dev;
220         struct dm_dev *origin_dev;
221         dm_thin_id dev_id;
222
223         struct pool *pool;
224         struct dm_thin_device *td;
225 };
226
227 /*----------------------------------------------------------------*/
228
229 /*
230  * A global list of pools that uses a struct mapped_device as a key.
231  */
232 static struct dm_thin_pool_table {
233         struct mutex mutex;
234         struct list_head pools;
235 } dm_thin_pool_table;
236
237 static void pool_table_init(void)
238 {
239         mutex_init(&dm_thin_pool_table.mutex);
240         INIT_LIST_HEAD(&dm_thin_pool_table.pools);
241 }
242
243 static void __pool_table_insert(struct pool *pool)
244 {
245         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
246         list_add(&pool->list, &dm_thin_pool_table.pools);
247 }
248
249 static void __pool_table_remove(struct pool *pool)
250 {
251         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
252         list_del(&pool->list);
253 }
254
255 static struct pool *__pool_table_lookup(struct mapped_device *md)
256 {
257         struct pool *pool = NULL, *tmp;
258
259         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
260
261         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
262                 if (tmp->pool_md == md) {
263                         pool = tmp;
264                         break;
265                 }
266         }
267
268         return pool;
269 }
270
271 static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
272 {
273         struct pool *pool = NULL, *tmp;
274
275         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
276
277         list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
278                 if (tmp->md_dev == md_dev) {
279                         pool = tmp;
280                         break;
281                 }
282         }
283
284         return pool;
285 }
286
287 /*----------------------------------------------------------------*/
288
289 struct dm_thin_endio_hook {
290         struct thin_c *tc;
291         struct dm_deferred_entry *shared_read_entry;
292         struct dm_deferred_entry *all_io_entry;
293         struct dm_thin_new_mapping *overwrite_mapping;
294 };
295
296 static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
297 {
298         struct bio *bio;
299         struct bio_list bios;
300
301         bio_list_init(&bios);
302         bio_list_merge(&bios, master);
303         bio_list_init(master);
304
305         while ((bio = bio_list_pop(&bios))) {
306                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
307
308                 if (h->tc == tc)
309                         bio_endio(bio, DM_ENDIO_REQUEUE);
310                 else
311                         bio_list_add(master, bio);
312         }
313 }
314
315 static void requeue_io(struct thin_c *tc)
316 {
317         struct pool *pool = tc->pool;
318         unsigned long flags;
319
320         spin_lock_irqsave(&pool->lock, flags);
321         __requeue_bio_list(tc, &pool->deferred_bios);
322         __requeue_bio_list(tc, &pool->retry_on_resume_list);
323         spin_unlock_irqrestore(&pool->lock, flags);
324 }
325
326 /*
327  * This section of code contains the logic for processing a thin device's IO.
328  * Much of the code depends on pool object resources (lists, workqueues, etc)
329  * but most is exclusively called from the thin target rather than the thin-pool
330  * target.
331  */
332
333 static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
334 {
335         sector_t block_nr = bio->bi_sector;
336
337         if (tc->pool->sectors_per_block_shift < 0)
338                 (void) sector_div(block_nr, tc->pool->sectors_per_block);
339         else
340                 block_nr >>= tc->pool->sectors_per_block_shift;
341
342         return block_nr;
343 }
344
345 static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
346 {
347         struct pool *pool = tc->pool;
348         sector_t bi_sector = bio->bi_sector;
349
350         bio->bi_bdev = tc->pool_dev->bdev;
351         if (tc->pool->sectors_per_block_shift < 0)
352                 bio->bi_sector = (block * pool->sectors_per_block) +
353                                  sector_div(bi_sector, pool->sectors_per_block);
354         else
355                 bio->bi_sector = (block << pool->sectors_per_block_shift) |
356                                 (bi_sector & (pool->sectors_per_block - 1));
357 }
358
359 static void remap_to_origin(struct thin_c *tc, struct bio *bio)
360 {
361         bio->bi_bdev = tc->origin_dev->bdev;
362 }
363
364 static int bio_triggers_commit(struct thin_c *tc, struct bio *bio)
365 {
366         return (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
367                 dm_thin_changed_this_transaction(tc->td);
368 }
369
370 static void inc_all_io_entry(struct pool *pool, struct bio *bio)
371 {
372         struct dm_thin_endio_hook *h;
373
374         if (bio->bi_rw & REQ_DISCARD)
375                 return;
376
377         h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
378         h->all_io_entry = dm_deferred_entry_inc(pool->all_io_ds);
379 }
380
381 static void issue(struct thin_c *tc, struct bio *bio)
382 {
383         struct pool *pool = tc->pool;
384         unsigned long flags;
385
386         if (!bio_triggers_commit(tc, bio)) {
387                 generic_make_request(bio);
388                 return;
389         }
390
391         /*
392          * Complete bio with an error if earlier I/O caused changes to
393          * the metadata that can't be committed e.g, due to I/O errors
394          * on the metadata device.
395          */
396         if (dm_thin_aborted_changes(tc->td)) {
397                 bio_io_error(bio);
398                 return;
399         }
400
401         /*
402          * Batch together any bios that trigger commits and then issue a
403          * single commit for them in process_deferred_bios().
404          */
405         spin_lock_irqsave(&pool->lock, flags);
406         bio_list_add(&pool->deferred_flush_bios, bio);
407         spin_unlock_irqrestore(&pool->lock, flags);
408 }
409
410 static void remap_to_origin_and_issue(struct thin_c *tc, struct bio *bio)
411 {
412         remap_to_origin(tc, bio);
413         issue(tc, bio);
414 }
415
416 static void remap_and_issue(struct thin_c *tc, struct bio *bio,
417                             dm_block_t block)
418 {
419         remap(tc, bio, block);
420         issue(tc, bio);
421 }
422
423 /*
424  * wake_worker() is used when new work is queued and when pool_resume is
425  * ready to continue deferred IO processing.
426  */
427 static void wake_worker(struct pool *pool)
428 {
429         queue_work(pool->wq, &pool->worker);
430 }
431
432 /*----------------------------------------------------------------*/
433
434 /*
435  * Bio endio functions.
436  */
437 struct dm_thin_new_mapping {
438         struct list_head list;
439
440         unsigned quiesced:1;
441         unsigned prepared:1;
442         unsigned pass_discard:1;
443
444         struct thin_c *tc;
445         dm_block_t virt_block;
446         dm_block_t data_block;
447         struct dm_bio_prison_cell *cell, *cell2;
448         int err;
449
450         /*
451          * If the bio covers the whole area of a block then we can avoid
452          * zeroing or copying.  Instead this bio is hooked.  The bio will
453          * still be in the cell, so care has to be taken to avoid issuing
454          * the bio twice.
455          */
456         struct bio *bio;
457         bio_end_io_t *saved_bi_end_io;
458 };
459
460 static void __maybe_add_mapping(struct dm_thin_new_mapping *m)
461 {
462         struct pool *pool = m->tc->pool;
463
464         if (m->quiesced && m->prepared) {
465                 list_add(&m->list, &pool->prepared_mappings);
466                 wake_worker(pool);
467         }
468 }
469
470 static void copy_complete(int read_err, unsigned long write_err, void *context)
471 {
472         unsigned long flags;
473         struct dm_thin_new_mapping *m = context;
474         struct pool *pool = m->tc->pool;
475
476         m->err = read_err || write_err ? -EIO : 0;
477
478         spin_lock_irqsave(&pool->lock, flags);
479         m->prepared = 1;
480         __maybe_add_mapping(m);
481         spin_unlock_irqrestore(&pool->lock, flags);
482 }
483
484 static void overwrite_endio(struct bio *bio, int err)
485 {
486         unsigned long flags;
487         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
488         struct dm_thin_new_mapping *m = h->overwrite_mapping;
489         struct pool *pool = m->tc->pool;
490
491         m->err = err;
492
493         spin_lock_irqsave(&pool->lock, flags);
494         m->prepared = 1;
495         __maybe_add_mapping(m);
496         spin_unlock_irqrestore(&pool->lock, flags);
497 }
498
499 /*----------------------------------------------------------------*/
500
501 /*
502  * Workqueue.
503  */
504
505 /*
506  * Prepared mapping jobs.
507  */
508
509 /*
510  * This sends the bios in the cell back to the deferred_bios list.
511  */
512 static void cell_defer(struct thin_c *tc, struct dm_bio_prison_cell *cell)
513 {
514         struct pool *pool = tc->pool;
515         unsigned long flags;
516
517         spin_lock_irqsave(&pool->lock, flags);
518         dm_cell_release(cell, &pool->deferred_bios);
519         spin_unlock_irqrestore(&tc->pool->lock, flags);
520
521         wake_worker(pool);
522 }
523
524 /*
525  * Same as cell_defer except it omits the original holder of the cell.
526  */
527 static void cell_defer_no_holder(struct thin_c *tc, struct dm_bio_prison_cell *cell)
528 {
529         struct pool *pool = tc->pool;
530         unsigned long flags;
531
532         spin_lock_irqsave(&pool->lock, flags);
533         dm_cell_release_no_holder(cell, &pool->deferred_bios);
534         spin_unlock_irqrestore(&pool->lock, flags);
535
536         wake_worker(pool);
537 }
538
539 static void process_prepared_mapping_fail(struct dm_thin_new_mapping *m)
540 {
541         if (m->bio)
542                 m->bio->bi_end_io = m->saved_bi_end_io;
543         dm_cell_error(m->cell);
544         list_del(&m->list);
545         mempool_free(m, m->tc->pool->mapping_pool);
546 }
547 static void process_prepared_mapping(struct dm_thin_new_mapping *m)
548 {
549         struct thin_c *tc = m->tc;
550         struct bio *bio;
551         int r;
552
553         bio = m->bio;
554         if (bio)
555                 bio->bi_end_io = m->saved_bi_end_io;
556
557         if (m->err) {
558                 dm_cell_error(m->cell);
559                 goto out;
560         }
561
562         /*
563          * Commit the prepared block into the mapping btree.
564          * Any I/O for this block arriving after this point will get
565          * remapped to it directly.
566          */
567         r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
568         if (r) {
569                 DMERR_LIMIT("dm_thin_insert_block() failed");
570                 dm_cell_error(m->cell);
571                 goto out;
572         }
573
574         /*
575          * Release any bios held while the block was being provisioned.
576          * If we are processing a write bio that completely covers the block,
577          * we already processed it so can ignore it now when processing
578          * the bios in the cell.
579          */
580         if (bio) {
581                 cell_defer_no_holder(tc, m->cell);
582                 bio_endio(bio, 0);
583         } else
584                 cell_defer(tc, m->cell);
585
586 out:
587         list_del(&m->list);
588         mempool_free(m, tc->pool->mapping_pool);
589 }
590
591 static void process_prepared_discard_fail(struct dm_thin_new_mapping *m)
592 {
593         struct thin_c *tc = m->tc;
594
595         bio_io_error(m->bio);
596         cell_defer_no_holder(tc, m->cell);
597         cell_defer_no_holder(tc, m->cell2);
598         mempool_free(m, tc->pool->mapping_pool);
599 }
600
601 static void process_prepared_discard_passdown(struct dm_thin_new_mapping *m)
602 {
603         struct thin_c *tc = m->tc;
604
605         inc_all_io_entry(tc->pool, m->bio);
606         cell_defer_no_holder(tc, m->cell);
607         cell_defer_no_holder(tc, m->cell2);
608
609         if (m->pass_discard)
610                 remap_and_issue(tc, m->bio, m->data_block);
611         else
612                 bio_endio(m->bio, 0);
613
614         mempool_free(m, tc->pool->mapping_pool);
615 }
616
617 static void process_prepared_discard(struct dm_thin_new_mapping *m)
618 {
619         int r;
620         struct thin_c *tc = m->tc;
621
622         r = dm_thin_remove_block(tc->td, m->virt_block);
623         if (r)
624                 DMERR_LIMIT("dm_thin_remove_block() failed");
625
626         process_prepared_discard_passdown(m);
627 }
628
629 static void process_prepared(struct pool *pool, struct list_head *head,
630                              process_mapping_fn *fn)
631 {
632         unsigned long flags;
633         struct list_head maps;
634         struct dm_thin_new_mapping *m, *tmp;
635
636         INIT_LIST_HEAD(&maps);
637         spin_lock_irqsave(&pool->lock, flags);
638         list_splice_init(head, &maps);
639         spin_unlock_irqrestore(&pool->lock, flags);
640
641         list_for_each_entry_safe(m, tmp, &maps, list)
642                 (*fn)(m);
643 }
644
645 /*
646  * Deferred bio jobs.
647  */
648 static int io_overlaps_block(struct pool *pool, struct bio *bio)
649 {
650         return bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT);
651 }
652
653 static int io_overwrites_block(struct pool *pool, struct bio *bio)
654 {
655         return (bio_data_dir(bio) == WRITE) &&
656                 io_overlaps_block(pool, bio);
657 }
658
659 static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
660                                bio_end_io_t *fn)
661 {
662         *save = bio->bi_end_io;
663         bio->bi_end_io = fn;
664 }
665
666 static int ensure_next_mapping(struct pool *pool)
667 {
668         if (pool->next_mapping)
669                 return 0;
670
671         pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
672
673         return pool->next_mapping ? 0 : -ENOMEM;
674 }
675
676 static struct dm_thin_new_mapping *get_next_mapping(struct pool *pool)
677 {
678         struct dm_thin_new_mapping *r = pool->next_mapping;
679
680         BUG_ON(!pool->next_mapping);
681
682         pool->next_mapping = NULL;
683
684         return r;
685 }
686
687 static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
688                           struct dm_dev *origin, dm_block_t data_origin,
689                           dm_block_t data_dest,
690                           struct dm_bio_prison_cell *cell, struct bio *bio)
691 {
692         int r;
693         struct pool *pool = tc->pool;
694         struct dm_thin_new_mapping *m = get_next_mapping(pool);
695
696         INIT_LIST_HEAD(&m->list);
697         m->quiesced = 0;
698         m->prepared = 0;
699         m->tc = tc;
700         m->virt_block = virt_block;
701         m->data_block = data_dest;
702         m->cell = cell;
703         m->err = 0;
704         m->bio = NULL;
705
706         if (!dm_deferred_set_add_work(pool->shared_read_ds, &m->list))
707                 m->quiesced = 1;
708
709         /*
710          * IO to pool_dev remaps to the pool target's data_dev.
711          *
712          * If the whole block of data is being overwritten, we can issue the
713          * bio immediately. Otherwise we use kcopyd to clone the data first.
714          */
715         if (io_overwrites_block(pool, bio)) {
716                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
717
718                 h->overwrite_mapping = m;
719                 m->bio = bio;
720                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
721                 inc_all_io_entry(pool, bio);
722                 remap_and_issue(tc, bio, data_dest);
723         } else {
724                 struct dm_io_region from, to;
725
726                 from.bdev = origin->bdev;
727                 from.sector = data_origin * pool->sectors_per_block;
728                 from.count = pool->sectors_per_block;
729
730                 to.bdev = tc->pool_dev->bdev;
731                 to.sector = data_dest * pool->sectors_per_block;
732                 to.count = pool->sectors_per_block;
733
734                 r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
735                                    0, copy_complete, m);
736                 if (r < 0) {
737                         mempool_free(m, pool->mapping_pool);
738                         DMERR_LIMIT("dm_kcopyd_copy() failed");
739                         dm_cell_error(cell);
740                 }
741         }
742 }
743
744 static void schedule_internal_copy(struct thin_c *tc, dm_block_t virt_block,
745                                    dm_block_t data_origin, dm_block_t data_dest,
746                                    struct dm_bio_prison_cell *cell, struct bio *bio)
747 {
748         schedule_copy(tc, virt_block, tc->pool_dev,
749                       data_origin, data_dest, cell, bio);
750 }
751
752 static void schedule_external_copy(struct thin_c *tc, dm_block_t virt_block,
753                                    dm_block_t data_dest,
754                                    struct dm_bio_prison_cell *cell, struct bio *bio)
755 {
756         schedule_copy(tc, virt_block, tc->origin_dev,
757                       virt_block, data_dest, cell, bio);
758 }
759
760 static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
761                           dm_block_t data_block, struct dm_bio_prison_cell *cell,
762                           struct bio *bio)
763 {
764         struct pool *pool = tc->pool;
765         struct dm_thin_new_mapping *m = get_next_mapping(pool);
766
767         INIT_LIST_HEAD(&m->list);
768         m->quiesced = 1;
769         m->prepared = 0;
770         m->tc = tc;
771         m->virt_block = virt_block;
772         m->data_block = data_block;
773         m->cell = cell;
774         m->err = 0;
775         m->bio = NULL;
776
777         /*
778          * If the whole block of data is being overwritten or we are not
779          * zeroing pre-existing data, we can issue the bio immediately.
780          * Otherwise we use kcopyd to zero the data first.
781          */
782         if (!pool->pf.zero_new_blocks)
783                 process_prepared_mapping(m);
784
785         else if (io_overwrites_block(pool, bio)) {
786                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
787
788                 h->overwrite_mapping = m;
789                 m->bio = bio;
790                 save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
791                 inc_all_io_entry(pool, bio);
792                 remap_and_issue(tc, bio, data_block);
793         } else {
794                 int r;
795                 struct dm_io_region to;
796
797                 to.bdev = tc->pool_dev->bdev;
798                 to.sector = data_block * pool->sectors_per_block;
799                 to.count = pool->sectors_per_block;
800
801                 r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
802                 if (r < 0) {
803                         mempool_free(m, pool->mapping_pool);
804                         DMERR_LIMIT("dm_kcopyd_zero() failed");
805                         dm_cell_error(cell);
806                 }
807         }
808 }
809
810 static int commit(struct pool *pool)
811 {
812         int r;
813
814         r = dm_pool_commit_metadata(pool->pmd);
815         if (r)
816                 DMERR_LIMIT("commit failed: error = %d", r);
817
818         return r;
819 }
820
821 /*
822  * A non-zero return indicates read_only or fail_io mode.
823  * Many callers don't care about the return value.
824  */
825 static int commit_or_fallback(struct pool *pool)
826 {
827         int r;
828
829         if (get_pool_mode(pool) != PM_WRITE)
830                 return -EINVAL;
831
832         r = commit(pool);
833         if (r)
834                 set_pool_mode(pool, PM_READ_ONLY);
835
836         return r;
837 }
838
839 static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
840 {
841         int r;
842         dm_block_t free_blocks;
843         unsigned long flags;
844         struct pool *pool = tc->pool;
845
846         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
847         if (r)
848                 return r;
849
850         if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
851                 DMWARN("%s: reached low water mark, sending event.",
852                        dm_device_name(pool->pool_md));
853                 spin_lock_irqsave(&pool->lock, flags);
854                 pool->low_water_triggered = 1;
855                 spin_unlock_irqrestore(&pool->lock, flags);
856                 dm_table_event(pool->ti->table);
857         }
858
859         if (!free_blocks) {
860                 if (pool->no_free_space)
861                         return -ENOSPC;
862                 else {
863                         /*
864                          * Try to commit to see if that will free up some
865                          * more space.
866                          */
867                         (void) commit_or_fallback(pool);
868
869                         r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
870                         if (r)
871                                 return r;
872
873                         /*
874                          * If we still have no space we set a flag to avoid
875                          * doing all this checking and return -ENOSPC.
876                          */
877                         if (!free_blocks) {
878                                 DMWARN("%s: no free space available.",
879                                        dm_device_name(pool->pool_md));
880                                 spin_lock_irqsave(&pool->lock, flags);
881                                 pool->no_free_space = 1;
882                                 spin_unlock_irqrestore(&pool->lock, flags);
883                                 return -ENOSPC;
884                         }
885                 }
886         }
887
888         r = dm_pool_alloc_data_block(pool->pmd, result);
889         if (r)
890                 return r;
891
892         return 0;
893 }
894
895 /*
896  * If we have run out of space, queue bios until the device is
897  * resumed, presumably after having been reloaded with more space.
898  */
899 static void retry_on_resume(struct bio *bio)
900 {
901         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
902         struct thin_c *tc = h->tc;
903         struct pool *pool = tc->pool;
904         unsigned long flags;
905
906         spin_lock_irqsave(&pool->lock, flags);
907         bio_list_add(&pool->retry_on_resume_list, bio);
908         spin_unlock_irqrestore(&pool->lock, flags);
909 }
910
911 static void no_space(struct dm_bio_prison_cell *cell)
912 {
913         struct bio *bio;
914         struct bio_list bios;
915
916         bio_list_init(&bios);
917         dm_cell_release(cell, &bios);
918
919         while ((bio = bio_list_pop(&bios)))
920                 retry_on_resume(bio);
921 }
922
923 static void process_discard(struct thin_c *tc, struct bio *bio)
924 {
925         int r;
926         unsigned long flags;
927         struct pool *pool = tc->pool;
928         struct dm_bio_prison_cell *cell, *cell2;
929         struct dm_cell_key key, key2;
930         dm_block_t block = get_bio_block(tc, bio);
931         struct dm_thin_lookup_result lookup_result;
932         struct dm_thin_new_mapping *m;
933
934         build_virtual_key(tc->td, block, &key);
935         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
936                 return;
937
938         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
939         switch (r) {
940         case 0:
941                 /*
942                  * Check nobody is fiddling with this pool block.  This can
943                  * happen if someone's in the process of breaking sharing
944                  * on this block.
945                  */
946                 build_data_key(tc->td, lookup_result.block, &key2);
947                 if (dm_bio_detain(tc->pool->prison, &key2, bio, &cell2)) {
948                         cell_defer_no_holder(tc, cell);
949                         break;
950                 }
951
952                 if (io_overlaps_block(pool, bio)) {
953                         /*
954                          * IO may still be going to the destination block.  We must
955                          * quiesce before we can do the removal.
956                          */
957                         m = get_next_mapping(pool);
958                         m->tc = tc;
959                         m->pass_discard = (!lookup_result.shared) && pool->pf.discard_passdown;
960                         m->virt_block = block;
961                         m->data_block = lookup_result.block;
962                         m->cell = cell;
963                         m->cell2 = cell2;
964                         m->err = 0;
965                         m->bio = bio;
966
967                         if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list)) {
968                                 spin_lock_irqsave(&pool->lock, flags);
969                                 list_add(&m->list, &pool->prepared_discards);
970                                 spin_unlock_irqrestore(&pool->lock, flags);
971                                 wake_worker(pool);
972                         }
973                 } else {
974                         inc_all_io_entry(pool, bio);
975                         cell_defer_no_holder(tc, cell);
976                         cell_defer_no_holder(tc, cell2);
977
978                         /*
979                          * The DM core makes sure that the discard doesn't span
980                          * a block boundary.  So we submit the discard of a
981                          * partial block appropriately.
982                          */
983                         if ((!lookup_result.shared) && pool->pf.discard_passdown)
984                                 remap_and_issue(tc, bio, lookup_result.block);
985                         else
986                                 bio_endio(bio, 0);
987                 }
988                 break;
989
990         case -ENODATA:
991                 /*
992                  * It isn't provisioned, just forget it.
993                  */
994                 cell_defer_no_holder(tc, cell);
995                 bio_endio(bio, 0);
996                 break;
997
998         default:
999                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1000                             __func__, r);
1001                 cell_defer_no_holder(tc, cell);
1002                 bio_io_error(bio);
1003                 break;
1004         }
1005 }
1006
1007 static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
1008                           struct dm_cell_key *key,
1009                           struct dm_thin_lookup_result *lookup_result,
1010                           struct dm_bio_prison_cell *cell)
1011 {
1012         int r;
1013         dm_block_t data_block;
1014
1015         r = alloc_data_block(tc, &data_block);
1016         switch (r) {
1017         case 0:
1018                 schedule_internal_copy(tc, block, lookup_result->block,
1019                                        data_block, cell, bio);
1020                 break;
1021
1022         case -ENOSPC:
1023                 no_space(cell);
1024                 break;
1025
1026         default:
1027                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1028                             __func__, r);
1029                 dm_cell_error(cell);
1030                 break;
1031         }
1032 }
1033
1034 static void process_shared_bio(struct thin_c *tc, struct bio *bio,
1035                                dm_block_t block,
1036                                struct dm_thin_lookup_result *lookup_result)
1037 {
1038         struct dm_bio_prison_cell *cell;
1039         struct pool *pool = tc->pool;
1040         struct dm_cell_key key;
1041
1042         /*
1043          * If cell is already occupied, then sharing is already in the process
1044          * of being broken so we have nothing further to do here.
1045          */
1046         build_data_key(tc->td, lookup_result->block, &key);
1047         if (dm_bio_detain(pool->prison, &key, bio, &cell))
1048                 return;
1049
1050         if (bio_data_dir(bio) == WRITE && bio->bi_size)
1051                 break_sharing(tc, bio, block, &key, lookup_result, cell);
1052         else {
1053                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1054
1055                 h->shared_read_entry = dm_deferred_entry_inc(pool->shared_read_ds);
1056                 inc_all_io_entry(pool, bio);
1057                 cell_defer_no_holder(tc, cell);
1058
1059                 remap_and_issue(tc, bio, lookup_result->block);
1060         }
1061 }
1062
1063 static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
1064                             struct dm_bio_prison_cell *cell)
1065 {
1066         int r;
1067         dm_block_t data_block;
1068
1069         /*
1070          * Remap empty bios (flushes) immediately, without provisioning.
1071          */
1072         if (!bio->bi_size) {
1073                 inc_all_io_entry(tc->pool, bio);
1074                 cell_defer_no_holder(tc, cell);
1075
1076                 remap_and_issue(tc, bio, 0);
1077                 return;
1078         }
1079
1080         /*
1081          * Fill read bios with zeroes and complete them immediately.
1082          */
1083         if (bio_data_dir(bio) == READ) {
1084                 zero_fill_bio(bio);
1085                 cell_defer_no_holder(tc, cell);
1086                 bio_endio(bio, 0);
1087                 return;
1088         }
1089
1090         r = alloc_data_block(tc, &data_block);
1091         switch (r) {
1092         case 0:
1093                 if (tc->origin_dev)
1094                         schedule_external_copy(tc, block, data_block, cell, bio);
1095                 else
1096                         schedule_zero(tc, block, data_block, cell, bio);
1097                 break;
1098
1099         case -ENOSPC:
1100                 no_space(cell);
1101                 break;
1102
1103         default:
1104                 DMERR_LIMIT("%s: alloc_data_block() failed: error = %d",
1105                             __func__, r);
1106                 set_pool_mode(tc->pool, PM_READ_ONLY);
1107                 dm_cell_error(cell);
1108                 break;
1109         }
1110 }
1111
1112 static void process_bio(struct thin_c *tc, struct bio *bio)
1113 {
1114         int r;
1115         dm_block_t block = get_bio_block(tc, bio);
1116         struct dm_bio_prison_cell *cell;
1117         struct dm_cell_key key;
1118         struct dm_thin_lookup_result lookup_result;
1119
1120         /*
1121          * If cell is already occupied, then the block is already
1122          * being provisioned so we have nothing further to do here.
1123          */
1124         build_virtual_key(tc->td, block, &key);
1125         if (dm_bio_detain(tc->pool->prison, &key, bio, &cell))
1126                 return;
1127
1128         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1129         switch (r) {
1130         case 0:
1131                 if (lookup_result.shared) {
1132                         process_shared_bio(tc, bio, block, &lookup_result);
1133                         cell_defer_no_holder(tc, cell);
1134                 } else {
1135                         inc_all_io_entry(tc->pool, bio);
1136                         cell_defer_no_holder(tc, cell);
1137
1138                         remap_and_issue(tc, bio, lookup_result.block);
1139                 }
1140                 break;
1141
1142         case -ENODATA:
1143                 if (bio_data_dir(bio) == READ && tc->origin_dev) {
1144                         inc_all_io_entry(tc->pool, bio);
1145                         cell_defer_no_holder(tc, cell);
1146
1147                         remap_to_origin_and_issue(tc, bio);
1148                 } else
1149                         provision_block(tc, bio, block, cell);
1150                 break;
1151
1152         default:
1153                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1154                             __func__, r);
1155                 cell_defer_no_holder(tc, cell);
1156                 bio_io_error(bio);
1157                 break;
1158         }
1159 }
1160
1161 static void process_bio_read_only(struct thin_c *tc, struct bio *bio)
1162 {
1163         int r;
1164         int rw = bio_data_dir(bio);
1165         dm_block_t block = get_bio_block(tc, bio);
1166         struct dm_thin_lookup_result lookup_result;
1167
1168         r = dm_thin_find_block(tc->td, block, 1, &lookup_result);
1169         switch (r) {
1170         case 0:
1171                 if (lookup_result.shared && (rw == WRITE) && bio->bi_size)
1172                         bio_io_error(bio);
1173                 else {
1174                         inc_all_io_entry(tc->pool, bio);
1175                         remap_and_issue(tc, bio, lookup_result.block);
1176                 }
1177                 break;
1178
1179         case -ENODATA:
1180                 if (rw != READ) {
1181                         bio_io_error(bio);
1182                         break;
1183                 }
1184
1185                 if (tc->origin_dev) {
1186                         inc_all_io_entry(tc->pool, bio);
1187                         remap_to_origin_and_issue(tc, bio);
1188                         break;
1189                 }
1190
1191                 zero_fill_bio(bio);
1192                 bio_endio(bio, 0);
1193                 break;
1194
1195         default:
1196                 DMERR_LIMIT("%s: dm_thin_find_block() failed: error = %d",
1197                             __func__, r);
1198                 bio_io_error(bio);
1199                 break;
1200         }
1201 }
1202
1203 static void process_bio_fail(struct thin_c *tc, struct bio *bio)
1204 {
1205         bio_io_error(bio);
1206 }
1207
1208 static int need_commit_due_to_time(struct pool *pool)
1209 {
1210         return jiffies < pool->last_commit_jiffies ||
1211                jiffies > pool->last_commit_jiffies + COMMIT_PERIOD;
1212 }
1213
1214 static void process_deferred_bios(struct pool *pool)
1215 {
1216         unsigned long flags;
1217         struct bio *bio;
1218         struct bio_list bios;
1219
1220         bio_list_init(&bios);
1221
1222         spin_lock_irqsave(&pool->lock, flags);
1223         bio_list_merge(&bios, &pool->deferred_bios);
1224         bio_list_init(&pool->deferred_bios);
1225         spin_unlock_irqrestore(&pool->lock, flags);
1226
1227         while ((bio = bio_list_pop(&bios))) {
1228                 struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1229                 struct thin_c *tc = h->tc;
1230
1231                 /*
1232                  * If we've got no free new_mapping structs, and processing
1233                  * this bio might require one, we pause until there are some
1234                  * prepared mappings to process.
1235                  */
1236                 if (ensure_next_mapping(pool)) {
1237                         spin_lock_irqsave(&pool->lock, flags);
1238                         bio_list_merge(&pool->deferred_bios, &bios);
1239                         spin_unlock_irqrestore(&pool->lock, flags);
1240
1241                         break;
1242                 }
1243
1244                 if (bio->bi_rw & REQ_DISCARD)
1245                         pool->process_discard(tc, bio);
1246                 else
1247                         pool->process_bio(tc, bio);
1248         }
1249
1250         /*
1251          * If there are any deferred flush bios, we must commit
1252          * the metadata before issuing them.
1253          */
1254         bio_list_init(&bios);
1255         spin_lock_irqsave(&pool->lock, flags);
1256         bio_list_merge(&bios, &pool->deferred_flush_bios);
1257         bio_list_init(&pool->deferred_flush_bios);
1258         spin_unlock_irqrestore(&pool->lock, flags);
1259
1260         if (bio_list_empty(&bios) && !need_commit_due_to_time(pool))
1261                 return;
1262
1263         if (commit_or_fallback(pool)) {
1264                 while ((bio = bio_list_pop(&bios)))
1265                         bio_io_error(bio);
1266                 return;
1267         }
1268         pool->last_commit_jiffies = jiffies;
1269
1270         while ((bio = bio_list_pop(&bios)))
1271                 generic_make_request(bio);
1272 }
1273
1274 static void do_worker(struct work_struct *ws)
1275 {
1276         struct pool *pool = container_of(ws, struct pool, worker);
1277
1278         process_prepared(pool, &pool->prepared_mappings, &pool->process_prepared_mapping);
1279         process_prepared(pool, &pool->prepared_discards, &pool->process_prepared_discard);
1280         process_deferred_bios(pool);
1281 }
1282
1283 /*
1284  * We want to commit periodically so that not too much
1285  * unwritten data builds up.
1286  */
1287 static void do_waker(struct work_struct *ws)
1288 {
1289         struct pool *pool = container_of(to_delayed_work(ws), struct pool, waker);
1290         wake_worker(pool);
1291         queue_delayed_work(pool->wq, &pool->waker, COMMIT_PERIOD);
1292 }
1293
1294 /*----------------------------------------------------------------*/
1295
1296 static enum pool_mode get_pool_mode(struct pool *pool)
1297 {
1298         return pool->pf.mode;
1299 }
1300
1301 static void set_pool_mode(struct pool *pool, enum pool_mode mode)
1302 {
1303         int r;
1304
1305         pool->pf.mode = mode;
1306
1307         switch (mode) {
1308         case PM_FAIL:
1309                 DMERR("switching pool to failure mode");
1310                 pool->process_bio = process_bio_fail;
1311                 pool->process_discard = process_bio_fail;
1312                 pool->process_prepared_mapping = process_prepared_mapping_fail;
1313                 pool->process_prepared_discard = process_prepared_discard_fail;
1314                 break;
1315
1316         case PM_READ_ONLY:
1317                 DMERR("switching pool to read-only mode");
1318                 r = dm_pool_abort_metadata(pool->pmd);
1319                 if (r) {
1320                         DMERR("aborting transaction failed");
1321                         set_pool_mode(pool, PM_FAIL);
1322                 } else {
1323                         dm_pool_metadata_read_only(pool->pmd);
1324                         pool->process_bio = process_bio_read_only;
1325                         pool->process_discard = process_discard;
1326                         pool->process_prepared_mapping = process_prepared_mapping_fail;
1327                         pool->process_prepared_discard = process_prepared_discard_passdown;
1328                 }
1329                 break;
1330
1331         case PM_WRITE:
1332                 pool->process_bio = process_bio;
1333                 pool->process_discard = process_discard;
1334                 pool->process_prepared_mapping = process_prepared_mapping;
1335                 pool->process_prepared_discard = process_prepared_discard;
1336                 break;
1337         }
1338 }
1339
1340 /*----------------------------------------------------------------*/
1341
1342 /*
1343  * Mapping functions.
1344  */
1345
1346 /*
1347  * Called only while mapping a thin bio to hand it over to the workqueue.
1348  */
1349 static void thin_defer_bio(struct thin_c *tc, struct bio *bio)
1350 {
1351         unsigned long flags;
1352         struct pool *pool = tc->pool;
1353
1354         spin_lock_irqsave(&pool->lock, flags);
1355         bio_list_add(&pool->deferred_bios, bio);
1356         spin_unlock_irqrestore(&pool->lock, flags);
1357
1358         wake_worker(pool);
1359 }
1360
1361 static void thin_hook_bio(struct thin_c *tc, struct bio *bio)
1362 {
1363         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
1364
1365         h->tc = tc;
1366         h->shared_read_entry = NULL;
1367         h->all_io_entry = NULL;
1368         h->overwrite_mapping = NULL;
1369 }
1370
1371 /*
1372  * Non-blocking function called from the thin target's map function.
1373  */
1374 static int thin_bio_map(struct dm_target *ti, struct bio *bio)
1375 {
1376         int r;
1377         struct thin_c *tc = ti->private;
1378         dm_block_t block = get_bio_block(tc, bio);
1379         struct dm_thin_device *td = tc->td;
1380         struct dm_thin_lookup_result result;
1381         struct dm_bio_prison_cell *cell1, *cell2;
1382         struct dm_cell_key key;
1383
1384         thin_hook_bio(tc, bio);
1385
1386         if (get_pool_mode(tc->pool) == PM_FAIL) {
1387                 bio_io_error(bio);
1388                 return DM_MAPIO_SUBMITTED;
1389         }
1390
1391         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1392                 thin_defer_bio(tc, bio);
1393                 return DM_MAPIO_SUBMITTED;
1394         }
1395
1396         r = dm_thin_find_block(td, block, 0, &result);
1397
1398         /*
1399          * Note that we defer readahead too.
1400          */
1401         switch (r) {
1402         case 0:
1403                 if (unlikely(result.shared)) {
1404                         /*
1405                          * We have a race condition here between the
1406                          * result.shared value returned by the lookup and
1407                          * snapshot creation, which may cause new
1408                          * sharing.
1409                          *
1410                          * To avoid this always quiesce the origin before
1411                          * taking the snap.  You want to do this anyway to
1412                          * ensure a consistent application view
1413                          * (i.e. lockfs).
1414                          *
1415                          * More distant ancestors are irrelevant. The
1416                          * shared flag will be set in their case.
1417                          */
1418                         thin_defer_bio(tc, bio);
1419                         return DM_MAPIO_SUBMITTED;
1420                 }
1421
1422                 build_virtual_key(tc->td, block, &key);
1423                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1))
1424                         return DM_MAPIO_SUBMITTED;
1425
1426                 build_data_key(tc->td, result.block, &key);
1427                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2)) {
1428                         cell_defer_no_holder(tc, cell1);
1429                         return DM_MAPIO_SUBMITTED;
1430                 }
1431
1432                 inc_all_io_entry(tc->pool, bio);
1433                 cell_defer_no_holder(tc, cell2);
1434                 cell_defer_no_holder(tc, cell1);
1435
1436                 remap(tc, bio, result.block);
1437                 return DM_MAPIO_REMAPPED;
1438
1439         case -ENODATA:
1440                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1441                         /*
1442                          * This block isn't provisioned, and we have no way
1443                          * of doing so.  Just error it.
1444                          */
1445                         bio_io_error(bio);
1446                         return DM_MAPIO_SUBMITTED;
1447                 }
1448                 /* fall through */
1449
1450         case -EWOULDBLOCK:
1451                 /*
1452                  * In future, the failed dm_thin_find_block above could
1453                  * provide the hint to load the metadata into cache.
1454                  */
1455                 thin_defer_bio(tc, bio);
1456                 return DM_MAPIO_SUBMITTED;
1457
1458         default:
1459                 /*
1460                  * Must always call bio_io_error on failure.
1461                  * dm_thin_find_block can fail with -EINVAL if the
1462                  * pool is switched to fail-io mode.
1463                  */
1464                 bio_io_error(bio);
1465                 return DM_MAPIO_SUBMITTED;
1466         }
1467 }
1468
1469 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1470 {
1471         int r;
1472         unsigned long flags;
1473         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1474
1475         spin_lock_irqsave(&pt->pool->lock, flags);
1476         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1477         spin_unlock_irqrestore(&pt->pool->lock, flags);
1478
1479         if (!r) {
1480                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1481                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1482         }
1483
1484         return r;
1485 }
1486
1487 static void __requeue_bios(struct pool *pool)
1488 {
1489         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1490         bio_list_init(&pool->retry_on_resume_list);
1491 }
1492
1493 /*----------------------------------------------------------------
1494  * Binding of control targets to a pool object
1495  *--------------------------------------------------------------*/
1496 static bool data_dev_supports_discard(struct pool_c *pt)
1497 {
1498         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1499
1500         return q && blk_queue_discard(q);
1501 }
1502
1503 /*
1504  * If discard_passdown was enabled verify that the data device
1505  * supports discards.  Disable discard_passdown if not.
1506  */
1507 static void disable_passdown_if_not_supported(struct pool_c *pt)
1508 {
1509         struct pool *pool = pt->pool;
1510         struct block_device *data_bdev = pt->data_dev->bdev;
1511         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1512         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1513         const char *reason = NULL;
1514         char buf[BDEVNAME_SIZE];
1515
1516         if (!pt->adjusted_pf.discard_passdown)
1517                 return;
1518
1519         if (!data_dev_supports_discard(pt))
1520                 reason = "discard unsupported";
1521
1522         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1523                 reason = "max discard sectors smaller than a block";
1524
1525         else if (data_limits->discard_granularity > block_size)
1526                 reason = "discard granularity larger than a block";
1527
1528         else if (block_size & (data_limits->discard_granularity - 1))
1529                 reason = "discard granularity not a factor of block size";
1530
1531         if (reason) {
1532                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1533                 pt->adjusted_pf.discard_passdown = false;
1534         }
1535 }
1536
1537 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1538 {
1539         struct pool_c *pt = ti->private;
1540
1541         /*
1542          * We want to make sure that degraded pools are never upgraded.
1543          */
1544         enum pool_mode old_mode = pool->pf.mode;
1545         enum pool_mode new_mode = pt->adjusted_pf.mode;
1546
1547         if (old_mode > new_mode)
1548                 new_mode = old_mode;
1549
1550         pool->ti = ti;
1551         pool->low_water_blocks = pt->low_water_blocks;
1552         pool->pf = pt->adjusted_pf;
1553
1554         set_pool_mode(pool, new_mode);
1555
1556         return 0;
1557 }
1558
1559 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1560 {
1561         if (pool->ti == ti)
1562                 pool->ti = NULL;
1563 }
1564
1565 /*----------------------------------------------------------------
1566  * Pool creation
1567  *--------------------------------------------------------------*/
1568 /* Initialize pool features. */
1569 static void pool_features_init(struct pool_features *pf)
1570 {
1571         pf->mode = PM_WRITE;
1572         pf->zero_new_blocks = true;
1573         pf->discard_enabled = true;
1574         pf->discard_passdown = true;
1575 }
1576
1577 static void __pool_destroy(struct pool *pool)
1578 {
1579         __pool_table_remove(pool);
1580
1581         if (dm_pool_metadata_close(pool->pmd) < 0)
1582                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1583
1584         dm_bio_prison_destroy(pool->prison);
1585         dm_kcopyd_client_destroy(pool->copier);
1586
1587         if (pool->wq)
1588                 destroy_workqueue(pool->wq);
1589
1590         if (pool->next_mapping)
1591                 mempool_free(pool->next_mapping, pool->mapping_pool);
1592         mempool_destroy(pool->mapping_pool);
1593         dm_deferred_set_destroy(pool->shared_read_ds);
1594         dm_deferred_set_destroy(pool->all_io_ds);
1595         kfree(pool);
1596 }
1597
1598 static struct kmem_cache *_new_mapping_cache;
1599
1600 static struct pool *pool_create(struct mapped_device *pool_md,
1601                                 struct block_device *metadata_dev,
1602                                 unsigned long block_size,
1603                                 int read_only, char **error)
1604 {
1605         int r;
1606         void *err_p;
1607         struct pool *pool;
1608         struct dm_pool_metadata *pmd;
1609         bool format_device = read_only ? false : true;
1610
1611         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1612         if (IS_ERR(pmd)) {
1613                 *error = "Error creating metadata object";
1614                 return (struct pool *)pmd;
1615         }
1616
1617         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1618         if (!pool) {
1619                 *error = "Error allocating memory for pool";
1620                 err_p = ERR_PTR(-ENOMEM);
1621                 goto bad_pool;
1622         }
1623
1624         pool->pmd = pmd;
1625         pool->sectors_per_block = block_size;
1626         if (block_size & (block_size - 1))
1627                 pool->sectors_per_block_shift = -1;
1628         else
1629                 pool->sectors_per_block_shift = __ffs(block_size);
1630         pool->low_water_blocks = 0;
1631         pool_features_init(&pool->pf);
1632         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1633         if (!pool->prison) {
1634                 *error = "Error creating pool's bio prison";
1635                 err_p = ERR_PTR(-ENOMEM);
1636                 goto bad_prison;
1637         }
1638
1639         pool->copier = dm_kcopyd_client_create();
1640         if (IS_ERR(pool->copier)) {
1641                 r = PTR_ERR(pool->copier);
1642                 *error = "Error creating pool's kcopyd client";
1643                 err_p = ERR_PTR(r);
1644                 goto bad_kcopyd_client;
1645         }
1646
1647         /*
1648          * Create singlethreaded workqueue that will service all devices
1649          * that use this metadata.
1650          */
1651         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1652         if (!pool->wq) {
1653                 *error = "Error creating pool's workqueue";
1654                 err_p = ERR_PTR(-ENOMEM);
1655                 goto bad_wq;
1656         }
1657
1658         INIT_WORK(&pool->worker, do_worker);
1659         INIT_DELAYED_WORK(&pool->waker, do_waker);
1660         spin_lock_init(&pool->lock);
1661         bio_list_init(&pool->deferred_bios);
1662         bio_list_init(&pool->deferred_flush_bios);
1663         INIT_LIST_HEAD(&pool->prepared_mappings);
1664         INIT_LIST_HEAD(&pool->prepared_discards);
1665         pool->low_water_triggered = 0;
1666         pool->no_free_space = 0;
1667         bio_list_init(&pool->retry_on_resume_list);
1668
1669         pool->shared_read_ds = dm_deferred_set_create();
1670         if (!pool->shared_read_ds) {
1671                 *error = "Error creating pool's shared read deferred set";
1672                 err_p = ERR_PTR(-ENOMEM);
1673                 goto bad_shared_read_ds;
1674         }
1675
1676         pool->all_io_ds = dm_deferred_set_create();
1677         if (!pool->all_io_ds) {
1678                 *error = "Error creating pool's all io deferred set";
1679                 err_p = ERR_PTR(-ENOMEM);
1680                 goto bad_all_io_ds;
1681         }
1682
1683         pool->next_mapping = NULL;
1684         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1685                                                       _new_mapping_cache);
1686         if (!pool->mapping_pool) {
1687                 *error = "Error creating pool's mapping mempool";
1688                 err_p = ERR_PTR(-ENOMEM);
1689                 goto bad_mapping_pool;
1690         }
1691
1692         pool->ref_count = 1;
1693         pool->last_commit_jiffies = jiffies;
1694         pool->pool_md = pool_md;
1695         pool->md_dev = metadata_dev;
1696         __pool_table_insert(pool);
1697
1698         return pool;
1699
1700 bad_mapping_pool:
1701         dm_deferred_set_destroy(pool->all_io_ds);
1702 bad_all_io_ds:
1703         dm_deferred_set_destroy(pool->shared_read_ds);
1704 bad_shared_read_ds:
1705         destroy_workqueue(pool->wq);
1706 bad_wq:
1707         dm_kcopyd_client_destroy(pool->copier);
1708 bad_kcopyd_client:
1709         dm_bio_prison_destroy(pool->prison);
1710 bad_prison:
1711         kfree(pool);
1712 bad_pool:
1713         if (dm_pool_metadata_close(pmd))
1714                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1715
1716         return err_p;
1717 }
1718
1719 static void __pool_inc(struct pool *pool)
1720 {
1721         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1722         pool->ref_count++;
1723 }
1724
1725 static void __pool_dec(struct pool *pool)
1726 {
1727         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1728         BUG_ON(!pool->ref_count);
1729         if (!--pool->ref_count)
1730                 __pool_destroy(pool);
1731 }
1732
1733 static struct pool *__pool_find(struct mapped_device *pool_md,
1734                                 struct block_device *metadata_dev,
1735                                 unsigned long block_size, int read_only,
1736                                 char **error, int *created)
1737 {
1738         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1739
1740         if (pool) {
1741                 if (pool->pool_md != pool_md) {
1742                         *error = "metadata device already in use by a pool";
1743                         return ERR_PTR(-EBUSY);
1744                 }
1745                 __pool_inc(pool);
1746
1747         } else {
1748                 pool = __pool_table_lookup(pool_md);
1749                 if (pool) {
1750                         if (pool->md_dev != metadata_dev) {
1751                                 *error = "different pool cannot replace a pool";
1752                                 return ERR_PTR(-EINVAL);
1753                         }
1754                         __pool_inc(pool);
1755
1756                 } else {
1757                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1758                         *created = 1;
1759                 }
1760         }
1761
1762         return pool;
1763 }
1764
1765 /*----------------------------------------------------------------
1766  * Pool target methods
1767  *--------------------------------------------------------------*/
1768 static void pool_dtr(struct dm_target *ti)
1769 {
1770         struct pool_c *pt = ti->private;
1771
1772         mutex_lock(&dm_thin_pool_table.mutex);
1773
1774         unbind_control_target(pt->pool, ti);
1775         __pool_dec(pt->pool);
1776         dm_put_device(ti, pt->metadata_dev);
1777         dm_put_device(ti, pt->data_dev);
1778         kfree(pt);
1779
1780         mutex_unlock(&dm_thin_pool_table.mutex);
1781 }
1782
1783 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1784                                struct dm_target *ti)
1785 {
1786         int r;
1787         unsigned argc;
1788         const char *arg_name;
1789
1790         static struct dm_arg _args[] = {
1791                 {0, 3, "Invalid number of pool feature arguments"},
1792         };
1793
1794         /*
1795          * No feature arguments supplied.
1796          */
1797         if (!as->argc)
1798                 return 0;
1799
1800         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1801         if (r)
1802                 return -EINVAL;
1803
1804         while (argc && !r) {
1805                 arg_name = dm_shift_arg(as);
1806                 argc--;
1807
1808                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1809                         pf->zero_new_blocks = false;
1810
1811                 else if (!strcasecmp(arg_name, "ignore_discard"))
1812                         pf->discard_enabled = false;
1813
1814                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1815                         pf->discard_passdown = false;
1816
1817                 else if (!strcasecmp(arg_name, "read_only"))
1818                         pf->mode = PM_READ_ONLY;
1819
1820                 else {
1821                         ti->error = "Unrecognised pool feature requested";
1822                         r = -EINVAL;
1823                         break;
1824                 }
1825         }
1826
1827         return r;
1828 }
1829
1830 /*
1831  * thin-pool <metadata dev> <data dev>
1832  *           <data block size (sectors)>
1833  *           <low water mark (blocks)>
1834  *           [<#feature args> [<arg>]*]
1835  *
1836  * Optional feature arguments are:
1837  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1838  *           ignore_discard: disable discard
1839  *           no_discard_passdown: don't pass discards down to the data device
1840  */
1841 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1842 {
1843         int r, pool_created = 0;
1844         struct pool_c *pt;
1845         struct pool *pool;
1846         struct pool_features pf;
1847         struct dm_arg_set as;
1848         struct dm_dev *data_dev;
1849         unsigned long block_size;
1850         dm_block_t low_water_blocks;
1851         struct dm_dev *metadata_dev;
1852         sector_t metadata_dev_size;
1853         char b[BDEVNAME_SIZE];
1854
1855         /*
1856          * FIXME Remove validation from scope of lock.
1857          */
1858         mutex_lock(&dm_thin_pool_table.mutex);
1859
1860         if (argc < 4) {
1861                 ti->error = "Invalid argument count";
1862                 r = -EINVAL;
1863                 goto out_unlock;
1864         }
1865         as.argc = argc;
1866         as.argv = argv;
1867
1868         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1869         if (r) {
1870                 ti->error = "Error opening metadata block device";
1871                 goto out_unlock;
1872         }
1873
1874         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1875         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1876                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1877                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1878
1879         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1880         if (r) {
1881                 ti->error = "Error getting data device";
1882                 goto out_metadata;
1883         }
1884
1885         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1886             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1887             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1888             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1889                 ti->error = "Invalid block size";
1890                 r = -EINVAL;
1891                 goto out;
1892         }
1893
1894         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1895                 ti->error = "Invalid low water mark";
1896                 r = -EINVAL;
1897                 goto out;
1898         }
1899
1900         /*
1901          * Set default pool features.
1902          */
1903         pool_features_init(&pf);
1904
1905         dm_consume_args(&as, 4);
1906         r = parse_pool_features(&as, &pf, ti);
1907         if (r)
1908                 goto out;
1909
1910         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1911         if (!pt) {
1912                 r = -ENOMEM;
1913                 goto out;
1914         }
1915
1916         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1917                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1918         if (IS_ERR(pool)) {
1919                 r = PTR_ERR(pool);
1920                 goto out_free_pt;
1921         }
1922
1923         /*
1924          * 'pool_created' reflects whether this is the first table load.
1925          * Top level discard support is not allowed to be changed after
1926          * initial load.  This would require a pool reload to trigger thin
1927          * device changes.
1928          */
1929         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1930                 ti->error = "Discard support cannot be disabled once enabled";
1931                 r = -EINVAL;
1932                 goto out_flags_changed;
1933         }
1934
1935         pt->pool = pool;
1936         pt->ti = ti;
1937         pt->metadata_dev = metadata_dev;
1938         pt->data_dev = data_dev;
1939         pt->low_water_blocks = low_water_blocks;
1940         pt->adjusted_pf = pt->requested_pf = pf;
1941         ti->num_flush_requests = 1;
1942
1943         /*
1944          * Only need to enable discards if the pool should pass
1945          * them down to the data device.  The thin device's discard
1946          * processing will cause mappings to be removed from the btree.
1947          */
1948         if (pf.discard_enabled && pf.discard_passdown) {
1949                 ti->num_discard_requests = 1;
1950
1951                 /*
1952                  * Setting 'discards_supported' circumvents the normal
1953                  * stacking of discard limits (this keeps the pool and
1954                  * thin devices' discard limits consistent).
1955                  */
1956                 ti->discards_supported = true;
1957                 ti->discard_zeroes_data_unsupported = true;
1958         }
1959         ti->private = pt;
1960
1961         pt->callbacks.congested_fn = pool_is_congested;
1962         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1963
1964         mutex_unlock(&dm_thin_pool_table.mutex);
1965
1966         return 0;
1967
1968 out_flags_changed:
1969         __pool_dec(pool);
1970 out_free_pt:
1971         kfree(pt);
1972 out:
1973         dm_put_device(ti, data_dev);
1974 out_metadata:
1975         dm_put_device(ti, metadata_dev);
1976 out_unlock:
1977         mutex_unlock(&dm_thin_pool_table.mutex);
1978
1979         return r;
1980 }
1981
1982 static int pool_map(struct dm_target *ti, struct bio *bio)
1983 {
1984         int r;
1985         struct pool_c *pt = ti->private;
1986         struct pool *pool = pt->pool;
1987         unsigned long flags;
1988
1989         /*
1990          * As this is a singleton target, ti->begin is always zero.
1991          */
1992         spin_lock_irqsave(&pool->lock, flags);
1993         bio->bi_bdev = pt->data_dev->bdev;
1994         r = DM_MAPIO_REMAPPED;
1995         spin_unlock_irqrestore(&pool->lock, flags);
1996
1997         return r;
1998 }
1999
2000 /*
2001  * Retrieves the number of blocks of the data device from
2002  * the superblock and compares it to the actual device size,
2003  * thus resizing the data device in case it has grown.
2004  *
2005  * This both copes with opening preallocated data devices in the ctr
2006  * being followed by a resume
2007  * -and-
2008  * calling the resume method individually after userspace has
2009  * grown the data device in reaction to a table event.
2010  */
2011 static int pool_preresume(struct dm_target *ti)
2012 {
2013         int r;
2014         struct pool_c *pt = ti->private;
2015         struct pool *pool = pt->pool;
2016         sector_t data_size = ti->len;
2017         dm_block_t sb_data_size;
2018
2019         /*
2020          * Take control of the pool object.
2021          */
2022         r = bind_control_target(pool, ti);
2023         if (r)
2024                 return r;
2025
2026         (void) sector_div(data_size, pool->sectors_per_block);
2027
2028         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2029         if (r) {
2030                 DMERR("failed to retrieve data device size");
2031                 return r;
2032         }
2033
2034         if (data_size < sb_data_size) {
2035                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2036                       (unsigned long long)data_size, sb_data_size);
2037                 return -EINVAL;
2038
2039         } else if (data_size > sb_data_size) {
2040                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2041                 if (r) {
2042                         DMERR("failed to resize data device");
2043                         /* FIXME Stricter than necessary: Rollback transaction instead here */
2044                         set_pool_mode(pool, PM_READ_ONLY);
2045                         return r;
2046                 }
2047
2048                 (void) commit_or_fallback(pool);
2049         }
2050
2051         return 0;
2052 }
2053
2054 static void pool_resume(struct dm_target *ti)
2055 {
2056         struct pool_c *pt = ti->private;
2057         struct pool *pool = pt->pool;
2058         unsigned long flags;
2059
2060         spin_lock_irqsave(&pool->lock, flags);
2061         pool->low_water_triggered = 0;
2062         pool->no_free_space = 0;
2063         __requeue_bios(pool);
2064         spin_unlock_irqrestore(&pool->lock, flags);
2065
2066         do_waker(&pool->waker.work);
2067 }
2068
2069 static void pool_postsuspend(struct dm_target *ti)
2070 {
2071         struct pool_c *pt = ti->private;
2072         struct pool *pool = pt->pool;
2073
2074         cancel_delayed_work(&pool->waker);
2075         flush_workqueue(pool->wq);
2076         (void) commit_or_fallback(pool);
2077 }
2078
2079 static int check_arg_count(unsigned argc, unsigned args_required)
2080 {
2081         if (argc != args_required) {
2082                 DMWARN("Message received with %u arguments instead of %u.",
2083                        argc, args_required);
2084                 return -EINVAL;
2085         }
2086
2087         return 0;
2088 }
2089
2090 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2091 {
2092         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2093             *dev_id <= MAX_DEV_ID)
2094                 return 0;
2095
2096         if (warning)
2097                 DMWARN("Message received with invalid device id: %s", arg);
2098
2099         return -EINVAL;
2100 }
2101
2102 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2103 {
2104         dm_thin_id dev_id;
2105         int r;
2106
2107         r = check_arg_count(argc, 2);
2108         if (r)
2109                 return r;
2110
2111         r = read_dev_id(argv[1], &dev_id, 1);
2112         if (r)
2113                 return r;
2114
2115         r = dm_pool_create_thin(pool->pmd, dev_id);
2116         if (r) {
2117                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2118                        argv[1]);
2119                 return r;
2120         }
2121
2122         return 0;
2123 }
2124
2125 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2126 {
2127         dm_thin_id dev_id;
2128         dm_thin_id origin_dev_id;
2129         int r;
2130
2131         r = check_arg_count(argc, 3);
2132         if (r)
2133                 return r;
2134
2135         r = read_dev_id(argv[1], &dev_id, 1);
2136         if (r)
2137                 return r;
2138
2139         r = read_dev_id(argv[2], &origin_dev_id, 1);
2140         if (r)
2141                 return r;
2142
2143         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2144         if (r) {
2145                 DMWARN("Creation of new snapshot %s of device %s failed.",
2146                        argv[1], argv[2]);
2147                 return r;
2148         }
2149
2150         return 0;
2151 }
2152
2153 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2154 {
2155         dm_thin_id dev_id;
2156         int r;
2157
2158         r = check_arg_count(argc, 2);
2159         if (r)
2160                 return r;
2161
2162         r = read_dev_id(argv[1], &dev_id, 1);
2163         if (r)
2164                 return r;
2165
2166         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2167         if (r)
2168                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2169
2170         return r;
2171 }
2172
2173 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2174 {
2175         dm_thin_id old_id, new_id;
2176         int r;
2177
2178         r = check_arg_count(argc, 3);
2179         if (r)
2180                 return r;
2181
2182         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2183                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2184                 return -EINVAL;
2185         }
2186
2187         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2188                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2189                 return -EINVAL;
2190         }
2191
2192         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2193         if (r) {
2194                 DMWARN("Failed to change transaction id from %s to %s.",
2195                        argv[1], argv[2]);
2196                 return r;
2197         }
2198
2199         return 0;
2200 }
2201
2202 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2203 {
2204         int r;
2205
2206         r = check_arg_count(argc, 1);
2207         if (r)
2208                 return r;
2209
2210         (void) commit_or_fallback(pool);
2211
2212         r = dm_pool_reserve_metadata_snap(pool->pmd);
2213         if (r)
2214                 DMWARN("reserve_metadata_snap message failed.");
2215
2216         return r;
2217 }
2218
2219 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2220 {
2221         int r;
2222
2223         r = check_arg_count(argc, 1);
2224         if (r)
2225                 return r;
2226
2227         r = dm_pool_release_metadata_snap(pool->pmd);
2228         if (r)
2229                 DMWARN("release_metadata_snap message failed.");
2230
2231         return r;
2232 }
2233
2234 /*
2235  * Messages supported:
2236  *   create_thin        <dev_id>
2237  *   create_snap        <dev_id> <origin_id>
2238  *   delete             <dev_id>
2239  *   trim               <dev_id> <new_size_in_sectors>
2240  *   set_transaction_id <current_trans_id> <new_trans_id>
2241  *   reserve_metadata_snap
2242  *   release_metadata_snap
2243  */
2244 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2245 {
2246         int r = -EINVAL;
2247         struct pool_c *pt = ti->private;
2248         struct pool *pool = pt->pool;
2249
2250         if (!strcasecmp(argv[0], "create_thin"))
2251                 r = process_create_thin_mesg(argc, argv, pool);
2252
2253         else if (!strcasecmp(argv[0], "create_snap"))
2254                 r = process_create_snap_mesg(argc, argv, pool);
2255
2256         else if (!strcasecmp(argv[0], "delete"))
2257                 r = process_delete_mesg(argc, argv, pool);
2258
2259         else if (!strcasecmp(argv[0], "set_transaction_id"))
2260                 r = process_set_transaction_id_mesg(argc, argv, pool);
2261
2262         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2263                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2264
2265         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2266                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2267
2268         else
2269                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2270
2271         if (!r)
2272                 (void) commit_or_fallback(pool);
2273
2274         return r;
2275 }
2276
2277 static void emit_flags(struct pool_features *pf, char *result,
2278                        unsigned sz, unsigned maxlen)
2279 {
2280         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2281                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2282         DMEMIT("%u ", count);
2283
2284         if (!pf->zero_new_blocks)
2285                 DMEMIT("skip_block_zeroing ");
2286
2287         if (!pf->discard_enabled)
2288                 DMEMIT("ignore_discard ");
2289
2290         if (!pf->discard_passdown)
2291                 DMEMIT("no_discard_passdown ");
2292
2293         if (pf->mode == PM_READ_ONLY)
2294                 DMEMIT("read_only ");
2295 }
2296
2297 /*
2298  * Status line is:
2299  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2300  *    <used data sectors>/<total data sectors> <held metadata root>
2301  */
2302 static int pool_status(struct dm_target *ti, status_type_t type,
2303                        unsigned status_flags, char *result, unsigned maxlen)
2304 {
2305         int r;
2306         unsigned sz = 0;
2307         uint64_t transaction_id;
2308         dm_block_t nr_free_blocks_data;
2309         dm_block_t nr_free_blocks_metadata;
2310         dm_block_t nr_blocks_data;
2311         dm_block_t nr_blocks_metadata;
2312         dm_block_t held_root;
2313         char buf[BDEVNAME_SIZE];
2314         char buf2[BDEVNAME_SIZE];
2315         struct pool_c *pt = ti->private;
2316         struct pool *pool = pt->pool;
2317
2318         switch (type) {
2319         case STATUSTYPE_INFO:
2320                 if (get_pool_mode(pool) == PM_FAIL) {
2321                         DMEMIT("Fail");
2322                         break;
2323                 }
2324
2325                 /* Commit to ensure statistics aren't out-of-date */
2326                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2327                         (void) commit_or_fallback(pool);
2328
2329                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2330                                                         &transaction_id);
2331                 if (r)
2332                         return r;
2333
2334                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2335                                                           &nr_free_blocks_metadata);
2336                 if (r)
2337                         return r;
2338
2339                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2340                 if (r)
2341                         return r;
2342
2343                 r = dm_pool_get_free_block_count(pool->pmd,
2344                                                  &nr_free_blocks_data);
2345                 if (r)
2346                         return r;
2347
2348                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2349                 if (r)
2350                         return r;
2351
2352                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2353                 if (r)
2354                         return r;
2355
2356                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2357                        (unsigned long long)transaction_id,
2358                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2359                        (unsigned long long)nr_blocks_metadata,
2360                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2361                        (unsigned long long)nr_blocks_data);
2362
2363                 if (held_root)
2364                         DMEMIT("%llu ", held_root);
2365                 else
2366                         DMEMIT("- ");
2367
2368                 if (pool->pf.mode == PM_READ_ONLY)
2369                         DMEMIT("ro ");
2370                 else
2371                         DMEMIT("rw ");
2372
2373                 if (!pool->pf.discard_enabled)
2374                         DMEMIT("ignore_discard");
2375                 else if (pool->pf.discard_passdown)
2376                         DMEMIT("discard_passdown");
2377                 else
2378                         DMEMIT("no_discard_passdown");
2379
2380                 break;
2381
2382         case STATUSTYPE_TABLE:
2383                 DMEMIT("%s %s %lu %llu ",
2384                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2385                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2386                        (unsigned long)pool->sectors_per_block,
2387                        (unsigned long long)pt->low_water_blocks);
2388                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2389                 break;
2390         }
2391
2392         return 0;
2393 }
2394
2395 static int pool_iterate_devices(struct dm_target *ti,
2396                                 iterate_devices_callout_fn fn, void *data)
2397 {
2398         struct pool_c *pt = ti->private;
2399
2400         return fn(ti, pt->data_dev, 0, ti->len, data);
2401 }
2402
2403 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2404                       struct bio_vec *biovec, int max_size)
2405 {
2406         struct pool_c *pt = ti->private;
2407         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2408
2409         if (!q->merge_bvec_fn)
2410                 return max_size;
2411
2412         bvm->bi_bdev = pt->data_dev->bdev;
2413
2414         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2415 }
2416
2417 static bool block_size_is_power_of_two(struct pool *pool)
2418 {
2419         return pool->sectors_per_block_shift >= 0;
2420 }
2421
2422 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2423 {
2424         struct pool *pool = pt->pool;
2425         struct queue_limits *data_limits;
2426
2427         limits->max_discard_sectors = pool->sectors_per_block;
2428
2429         /*
2430          * discard_granularity is just a hint, and not enforced.
2431          */
2432         if (pt->adjusted_pf.discard_passdown) {
2433                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2434                 limits->discard_granularity = data_limits->discard_granularity;
2435         } else if (block_size_is_power_of_two(pool))
2436                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2437         else
2438                 /*
2439                  * Use largest power of 2 that is a factor of sectors_per_block
2440                  * but at least DATA_DEV_BLOCK_SIZE_MIN_SECTORS.
2441                  */
2442                 limits->discard_granularity = max(1 << (ffs(pool->sectors_per_block) - 1),
2443                                                   DATA_DEV_BLOCK_SIZE_MIN_SECTORS) << SECTOR_SHIFT;
2444 }
2445
2446 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2447 {
2448         struct pool_c *pt = ti->private;
2449         struct pool *pool = pt->pool;
2450
2451         blk_limits_io_min(limits, 0);
2452         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2453
2454         /*
2455          * pt->adjusted_pf is a staging area for the actual features to use.
2456          * They get transferred to the live pool in bind_control_target()
2457          * called from pool_preresume().
2458          */
2459         if (!pt->adjusted_pf.discard_enabled)
2460                 return;
2461
2462         disable_passdown_if_not_supported(pt);
2463
2464         set_discard_limits(pt, limits);
2465 }
2466
2467 static struct target_type pool_target = {
2468         .name = "thin-pool",
2469         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2470                     DM_TARGET_IMMUTABLE,
2471         .version = {1, 6, 0},
2472         .module = THIS_MODULE,
2473         .ctr = pool_ctr,
2474         .dtr = pool_dtr,
2475         .map = pool_map,
2476         .postsuspend = pool_postsuspend,
2477         .preresume = pool_preresume,
2478         .resume = pool_resume,
2479         .message = pool_message,
2480         .status = pool_status,
2481         .merge = pool_merge,
2482         .iterate_devices = pool_iterate_devices,
2483         .io_hints = pool_io_hints,
2484 };
2485
2486 /*----------------------------------------------------------------
2487  * Thin target methods
2488  *--------------------------------------------------------------*/
2489 static void thin_dtr(struct dm_target *ti)
2490 {
2491         struct thin_c *tc = ti->private;
2492
2493         mutex_lock(&dm_thin_pool_table.mutex);
2494
2495         __pool_dec(tc->pool);
2496         dm_pool_close_thin_device(tc->td);
2497         dm_put_device(ti, tc->pool_dev);
2498         if (tc->origin_dev)
2499                 dm_put_device(ti, tc->origin_dev);
2500         kfree(tc);
2501
2502         mutex_unlock(&dm_thin_pool_table.mutex);
2503 }
2504
2505 /*
2506  * Thin target parameters:
2507  *
2508  * <pool_dev> <dev_id> [origin_dev]
2509  *
2510  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2511  * dev_id: the internal device identifier
2512  * origin_dev: a device external to the pool that should act as the origin
2513  *
2514  * If the pool device has discards disabled, they get disabled for the thin
2515  * device as well.
2516  */
2517 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2518 {
2519         int r;
2520         struct thin_c *tc;
2521         struct dm_dev *pool_dev, *origin_dev;
2522         struct mapped_device *pool_md;
2523
2524         mutex_lock(&dm_thin_pool_table.mutex);
2525
2526         if (argc != 2 && argc != 3) {
2527                 ti->error = "Invalid argument count";
2528                 r = -EINVAL;
2529                 goto out_unlock;
2530         }
2531
2532         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2533         if (!tc) {
2534                 ti->error = "Out of memory";
2535                 r = -ENOMEM;
2536                 goto out_unlock;
2537         }
2538
2539         if (argc == 3) {
2540                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2541                 if (r) {
2542                         ti->error = "Error opening origin device";
2543                         goto bad_origin_dev;
2544                 }
2545                 tc->origin_dev = origin_dev;
2546         }
2547
2548         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2549         if (r) {
2550                 ti->error = "Error opening pool device";
2551                 goto bad_pool_dev;
2552         }
2553         tc->pool_dev = pool_dev;
2554
2555         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2556                 ti->error = "Invalid device id";
2557                 r = -EINVAL;
2558                 goto bad_common;
2559         }
2560
2561         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2562         if (!pool_md) {
2563                 ti->error = "Couldn't get pool mapped device";
2564                 r = -EINVAL;
2565                 goto bad_common;
2566         }
2567
2568         tc->pool = __pool_table_lookup(pool_md);
2569         if (!tc->pool) {
2570                 ti->error = "Couldn't find pool object";
2571                 r = -EINVAL;
2572                 goto bad_pool_lookup;
2573         }
2574         __pool_inc(tc->pool);
2575
2576         if (get_pool_mode(tc->pool) == PM_FAIL) {
2577                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2578                 goto bad_thin_open;
2579         }
2580
2581         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2582         if (r) {
2583                 ti->error = "Couldn't open thin internal device";
2584                 goto bad_thin_open;
2585         }
2586
2587         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2588         if (r)
2589                 goto bad_thin_open;
2590
2591         ti->num_flush_requests = 1;
2592         ti->flush_supported = true;
2593         ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
2594
2595         /* In case the pool supports discards, pass them on. */
2596         if (tc->pool->pf.discard_enabled) {
2597                 ti->discards_supported = true;
2598                 ti->num_discard_requests = 1;
2599                 ti->discard_zeroes_data_unsupported = true;
2600                 /* Discard requests must be split on a block boundary */
2601                 ti->split_discard_requests = true;
2602         }
2603
2604         dm_put(pool_md);
2605
2606         mutex_unlock(&dm_thin_pool_table.mutex);
2607
2608         return 0;
2609
2610 bad_thin_open:
2611         __pool_dec(tc->pool);
2612 bad_pool_lookup:
2613         dm_put(pool_md);
2614 bad_common:
2615         dm_put_device(ti, tc->pool_dev);
2616 bad_pool_dev:
2617         if (tc->origin_dev)
2618                 dm_put_device(ti, tc->origin_dev);
2619 bad_origin_dev:
2620         kfree(tc);
2621 out_unlock:
2622         mutex_unlock(&dm_thin_pool_table.mutex);
2623
2624         return r;
2625 }
2626
2627 static int thin_map(struct dm_target *ti, struct bio *bio)
2628 {
2629         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2630
2631         return thin_bio_map(ti, bio);
2632 }
2633
2634 static int thin_endio(struct dm_target *ti, struct bio *bio, int err)
2635 {
2636         unsigned long flags;
2637         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2638         struct list_head work;
2639         struct dm_thin_new_mapping *m, *tmp;
2640         struct pool *pool = h->tc->pool;
2641
2642         if (h->shared_read_entry) {
2643                 INIT_LIST_HEAD(&work);
2644                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2645
2646                 spin_lock_irqsave(&pool->lock, flags);
2647                 list_for_each_entry_safe(m, tmp, &work, list) {
2648                         list_del(&m->list);
2649                         m->quiesced = 1;
2650                         __maybe_add_mapping(m);
2651                 }
2652                 spin_unlock_irqrestore(&pool->lock, flags);
2653         }
2654
2655         if (h->all_io_entry) {
2656                 INIT_LIST_HEAD(&work);
2657                 dm_deferred_entry_dec(h->all_io_entry, &work);
2658                 if (!list_empty(&work)) {
2659                         spin_lock_irqsave(&pool->lock, flags);
2660                         list_for_each_entry_safe(m, tmp, &work, list)
2661                                 list_add(&m->list, &pool->prepared_discards);
2662                         spin_unlock_irqrestore(&pool->lock, flags);
2663                         wake_worker(pool);
2664                 }
2665         }
2666
2667         return 0;
2668 }
2669
2670 static void thin_postsuspend(struct dm_target *ti)
2671 {
2672         if (dm_noflush_suspending(ti))
2673                 requeue_io((struct thin_c *)ti->private);
2674 }
2675
2676 /*
2677  * <nr mapped sectors> <highest mapped sector>
2678  */
2679 static int thin_status(struct dm_target *ti, status_type_t type,
2680                        unsigned status_flags, char *result, unsigned maxlen)
2681 {
2682         int r;
2683         ssize_t sz = 0;
2684         dm_block_t mapped, highest;
2685         char buf[BDEVNAME_SIZE];
2686         struct thin_c *tc = ti->private;
2687
2688         if (get_pool_mode(tc->pool) == PM_FAIL) {
2689                 DMEMIT("Fail");
2690                 return 0;
2691         }
2692
2693         if (!tc->td)
2694                 DMEMIT("-");
2695         else {
2696                 switch (type) {
2697                 case STATUSTYPE_INFO:
2698                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2699                         if (r)
2700                                 return r;
2701
2702                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2703                         if (r < 0)
2704                                 return r;
2705
2706                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2707                         if (r)
2708                                 DMEMIT("%llu", ((highest + 1) *
2709                                                 tc->pool->sectors_per_block) - 1);
2710                         else
2711                                 DMEMIT("-");
2712                         break;
2713
2714                 case STATUSTYPE_TABLE:
2715                         DMEMIT("%s %lu",
2716                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2717                                (unsigned long) tc->dev_id);
2718                         if (tc->origin_dev)
2719                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2720                         break;
2721                 }
2722         }
2723
2724         return 0;
2725 }
2726
2727 static int thin_iterate_devices(struct dm_target *ti,
2728                                 iterate_devices_callout_fn fn, void *data)
2729 {
2730         sector_t blocks;
2731         struct thin_c *tc = ti->private;
2732         struct pool *pool = tc->pool;
2733
2734         /*
2735          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2736          * we follow a more convoluted path through to the pool's target.
2737          */
2738         if (!pool->ti)
2739                 return 0;       /* nothing is bound */
2740
2741         blocks = pool->ti->len;
2742         (void) sector_div(blocks, pool->sectors_per_block);
2743         if (blocks)
2744                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2745
2746         return 0;
2747 }
2748
2749 static struct target_type thin_target = {
2750         .name = "thin",
2751         .version = {1, 7, 0},
2752         .module = THIS_MODULE,
2753         .ctr = thin_ctr,
2754         .dtr = thin_dtr,
2755         .map = thin_map,
2756         .end_io = thin_endio,
2757         .postsuspend = thin_postsuspend,
2758         .status = thin_status,
2759         .iterate_devices = thin_iterate_devices,
2760 };
2761
2762 /*----------------------------------------------------------------*/
2763
2764 static int __init dm_thin_init(void)
2765 {
2766         int r;
2767
2768         pool_table_init();
2769
2770         r = dm_register_target(&thin_target);
2771         if (r)
2772                 return r;
2773
2774         r = dm_register_target(&pool_target);
2775         if (r)
2776                 goto bad_pool_target;
2777
2778         r = -ENOMEM;
2779
2780         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2781         if (!_new_mapping_cache)
2782                 goto bad_new_mapping_cache;
2783
2784         return 0;
2785
2786 bad_new_mapping_cache:
2787         dm_unregister_target(&pool_target);
2788 bad_pool_target:
2789         dm_unregister_target(&thin_target);
2790
2791         return r;
2792 }
2793
2794 static void dm_thin_exit(void)
2795 {
2796         dm_unregister_target(&thin_target);
2797         dm_unregister_target(&pool_target);
2798
2799         kmem_cache_destroy(_new_mapping_cache);
2800 }
2801
2802 module_init(dm_thin_init);
2803 module_exit(dm_thin_exit);
2804
2805 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2806 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2807 MODULE_LICENSE("GPL");