]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - drivers/md/dm-thin.c
dm thin: dont use map_context
[~shefty/rdma-dev.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                         union map_info *map_context)
1376 {
1377         int r;
1378         struct thin_c *tc = ti->private;
1379         dm_block_t block = get_bio_block(tc, bio);
1380         struct dm_thin_device *td = tc->td;
1381         struct dm_thin_lookup_result result;
1382         struct dm_bio_prison_cell *cell1, *cell2;
1383         struct dm_cell_key key;
1384
1385         thin_hook_bio(tc, bio);
1386
1387         if (get_pool_mode(tc->pool) == PM_FAIL) {
1388                 bio_io_error(bio);
1389                 return DM_MAPIO_SUBMITTED;
1390         }
1391
1392         if (bio->bi_rw & (REQ_DISCARD | REQ_FLUSH | REQ_FUA)) {
1393                 thin_defer_bio(tc, bio);
1394                 return DM_MAPIO_SUBMITTED;
1395         }
1396
1397         r = dm_thin_find_block(td, block, 0, &result);
1398
1399         /*
1400          * Note that we defer readahead too.
1401          */
1402         switch (r) {
1403         case 0:
1404                 if (unlikely(result.shared)) {
1405                         /*
1406                          * We have a race condition here between the
1407                          * result.shared value returned by the lookup and
1408                          * snapshot creation, which may cause new
1409                          * sharing.
1410                          *
1411                          * To avoid this always quiesce the origin before
1412                          * taking the snap.  You want to do this anyway to
1413                          * ensure a consistent application view
1414                          * (i.e. lockfs).
1415                          *
1416                          * More distant ancestors are irrelevant. The
1417                          * shared flag will be set in their case.
1418                          */
1419                         thin_defer_bio(tc, bio);
1420                         return DM_MAPIO_SUBMITTED;
1421                 }
1422
1423                 build_virtual_key(tc->td, block, &key);
1424                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell1))
1425                         return DM_MAPIO_SUBMITTED;
1426
1427                 build_data_key(tc->td, result.block, &key);
1428                 if (dm_bio_detain(tc->pool->prison, &key, bio, &cell2)) {
1429                         cell_defer_no_holder(tc, cell1);
1430                         return DM_MAPIO_SUBMITTED;
1431                 }
1432
1433                 inc_all_io_entry(tc->pool, bio);
1434                 cell_defer_no_holder(tc, cell2);
1435                 cell_defer_no_holder(tc, cell1);
1436
1437                 remap(tc, bio, result.block);
1438                 return DM_MAPIO_REMAPPED;
1439
1440         case -ENODATA:
1441                 if (get_pool_mode(tc->pool) == PM_READ_ONLY) {
1442                         /*
1443                          * This block isn't provisioned, and we have no way
1444                          * of doing so.  Just error it.
1445                          */
1446                         bio_io_error(bio);
1447                         return DM_MAPIO_SUBMITTED;
1448                 }
1449                 /* fall through */
1450
1451         case -EWOULDBLOCK:
1452                 /*
1453                  * In future, the failed dm_thin_find_block above could
1454                  * provide the hint to load the metadata into cache.
1455                  */
1456                 thin_defer_bio(tc, bio);
1457                 return DM_MAPIO_SUBMITTED;
1458
1459         default:
1460                 /*
1461                  * Must always call bio_io_error on failure.
1462                  * dm_thin_find_block can fail with -EINVAL if the
1463                  * pool is switched to fail-io mode.
1464                  */
1465                 bio_io_error(bio);
1466                 return DM_MAPIO_SUBMITTED;
1467         }
1468 }
1469
1470 static int pool_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1471 {
1472         int r;
1473         unsigned long flags;
1474         struct pool_c *pt = container_of(cb, struct pool_c, callbacks);
1475
1476         spin_lock_irqsave(&pt->pool->lock, flags);
1477         r = !bio_list_empty(&pt->pool->retry_on_resume_list);
1478         spin_unlock_irqrestore(&pt->pool->lock, flags);
1479
1480         if (!r) {
1481                 struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1482                 r = bdi_congested(&q->backing_dev_info, bdi_bits);
1483         }
1484
1485         return r;
1486 }
1487
1488 static void __requeue_bios(struct pool *pool)
1489 {
1490         bio_list_merge(&pool->deferred_bios, &pool->retry_on_resume_list);
1491         bio_list_init(&pool->retry_on_resume_list);
1492 }
1493
1494 /*----------------------------------------------------------------
1495  * Binding of control targets to a pool object
1496  *--------------------------------------------------------------*/
1497 static bool data_dev_supports_discard(struct pool_c *pt)
1498 {
1499         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
1500
1501         return q && blk_queue_discard(q);
1502 }
1503
1504 /*
1505  * If discard_passdown was enabled verify that the data device
1506  * supports discards.  Disable discard_passdown if not.
1507  */
1508 static void disable_passdown_if_not_supported(struct pool_c *pt)
1509 {
1510         struct pool *pool = pt->pool;
1511         struct block_device *data_bdev = pt->data_dev->bdev;
1512         struct queue_limits *data_limits = &bdev_get_queue(data_bdev)->limits;
1513         sector_t block_size = pool->sectors_per_block << SECTOR_SHIFT;
1514         const char *reason = NULL;
1515         char buf[BDEVNAME_SIZE];
1516
1517         if (!pt->adjusted_pf.discard_passdown)
1518                 return;
1519
1520         if (!data_dev_supports_discard(pt))
1521                 reason = "discard unsupported";
1522
1523         else if (data_limits->max_discard_sectors < pool->sectors_per_block)
1524                 reason = "max discard sectors smaller than a block";
1525
1526         else if (data_limits->discard_granularity > block_size)
1527                 reason = "discard granularity larger than a block";
1528
1529         else if (block_size & (data_limits->discard_granularity - 1))
1530                 reason = "discard granularity not a factor of block size";
1531
1532         if (reason) {
1533                 DMWARN("Data device (%s) %s: Disabling discard passdown.", bdevname(data_bdev, buf), reason);
1534                 pt->adjusted_pf.discard_passdown = false;
1535         }
1536 }
1537
1538 static int bind_control_target(struct pool *pool, struct dm_target *ti)
1539 {
1540         struct pool_c *pt = ti->private;
1541
1542         /*
1543          * We want to make sure that degraded pools are never upgraded.
1544          */
1545         enum pool_mode old_mode = pool->pf.mode;
1546         enum pool_mode new_mode = pt->adjusted_pf.mode;
1547
1548         if (old_mode > new_mode)
1549                 new_mode = old_mode;
1550
1551         pool->ti = ti;
1552         pool->low_water_blocks = pt->low_water_blocks;
1553         pool->pf = pt->adjusted_pf;
1554
1555         set_pool_mode(pool, new_mode);
1556
1557         return 0;
1558 }
1559
1560 static void unbind_control_target(struct pool *pool, struct dm_target *ti)
1561 {
1562         if (pool->ti == ti)
1563                 pool->ti = NULL;
1564 }
1565
1566 /*----------------------------------------------------------------
1567  * Pool creation
1568  *--------------------------------------------------------------*/
1569 /* Initialize pool features. */
1570 static void pool_features_init(struct pool_features *pf)
1571 {
1572         pf->mode = PM_WRITE;
1573         pf->zero_new_blocks = true;
1574         pf->discard_enabled = true;
1575         pf->discard_passdown = true;
1576 }
1577
1578 static void __pool_destroy(struct pool *pool)
1579 {
1580         __pool_table_remove(pool);
1581
1582         if (dm_pool_metadata_close(pool->pmd) < 0)
1583                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1584
1585         dm_bio_prison_destroy(pool->prison);
1586         dm_kcopyd_client_destroy(pool->copier);
1587
1588         if (pool->wq)
1589                 destroy_workqueue(pool->wq);
1590
1591         if (pool->next_mapping)
1592                 mempool_free(pool->next_mapping, pool->mapping_pool);
1593         mempool_destroy(pool->mapping_pool);
1594         dm_deferred_set_destroy(pool->shared_read_ds);
1595         dm_deferred_set_destroy(pool->all_io_ds);
1596         kfree(pool);
1597 }
1598
1599 static struct kmem_cache *_new_mapping_cache;
1600
1601 static struct pool *pool_create(struct mapped_device *pool_md,
1602                                 struct block_device *metadata_dev,
1603                                 unsigned long block_size,
1604                                 int read_only, char **error)
1605 {
1606         int r;
1607         void *err_p;
1608         struct pool *pool;
1609         struct dm_pool_metadata *pmd;
1610         bool format_device = read_only ? false : true;
1611
1612         pmd = dm_pool_metadata_open(metadata_dev, block_size, format_device);
1613         if (IS_ERR(pmd)) {
1614                 *error = "Error creating metadata object";
1615                 return (struct pool *)pmd;
1616         }
1617
1618         pool = kmalloc(sizeof(*pool), GFP_KERNEL);
1619         if (!pool) {
1620                 *error = "Error allocating memory for pool";
1621                 err_p = ERR_PTR(-ENOMEM);
1622                 goto bad_pool;
1623         }
1624
1625         pool->pmd = pmd;
1626         pool->sectors_per_block = block_size;
1627         if (block_size & (block_size - 1))
1628                 pool->sectors_per_block_shift = -1;
1629         else
1630                 pool->sectors_per_block_shift = __ffs(block_size);
1631         pool->low_water_blocks = 0;
1632         pool_features_init(&pool->pf);
1633         pool->prison = dm_bio_prison_create(PRISON_CELLS);
1634         if (!pool->prison) {
1635                 *error = "Error creating pool's bio prison";
1636                 err_p = ERR_PTR(-ENOMEM);
1637                 goto bad_prison;
1638         }
1639
1640         pool->copier = dm_kcopyd_client_create();
1641         if (IS_ERR(pool->copier)) {
1642                 r = PTR_ERR(pool->copier);
1643                 *error = "Error creating pool's kcopyd client";
1644                 err_p = ERR_PTR(r);
1645                 goto bad_kcopyd_client;
1646         }
1647
1648         /*
1649          * Create singlethreaded workqueue that will service all devices
1650          * that use this metadata.
1651          */
1652         pool->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1653         if (!pool->wq) {
1654                 *error = "Error creating pool's workqueue";
1655                 err_p = ERR_PTR(-ENOMEM);
1656                 goto bad_wq;
1657         }
1658
1659         INIT_WORK(&pool->worker, do_worker);
1660         INIT_DELAYED_WORK(&pool->waker, do_waker);
1661         spin_lock_init(&pool->lock);
1662         bio_list_init(&pool->deferred_bios);
1663         bio_list_init(&pool->deferred_flush_bios);
1664         INIT_LIST_HEAD(&pool->prepared_mappings);
1665         INIT_LIST_HEAD(&pool->prepared_discards);
1666         pool->low_water_triggered = 0;
1667         pool->no_free_space = 0;
1668         bio_list_init(&pool->retry_on_resume_list);
1669
1670         pool->shared_read_ds = dm_deferred_set_create();
1671         if (!pool->shared_read_ds) {
1672                 *error = "Error creating pool's shared read deferred set";
1673                 err_p = ERR_PTR(-ENOMEM);
1674                 goto bad_shared_read_ds;
1675         }
1676
1677         pool->all_io_ds = dm_deferred_set_create();
1678         if (!pool->all_io_ds) {
1679                 *error = "Error creating pool's all io deferred set";
1680                 err_p = ERR_PTR(-ENOMEM);
1681                 goto bad_all_io_ds;
1682         }
1683
1684         pool->next_mapping = NULL;
1685         pool->mapping_pool = mempool_create_slab_pool(MAPPING_POOL_SIZE,
1686                                                       _new_mapping_cache);
1687         if (!pool->mapping_pool) {
1688                 *error = "Error creating pool's mapping mempool";
1689                 err_p = ERR_PTR(-ENOMEM);
1690                 goto bad_mapping_pool;
1691         }
1692
1693         pool->ref_count = 1;
1694         pool->last_commit_jiffies = jiffies;
1695         pool->pool_md = pool_md;
1696         pool->md_dev = metadata_dev;
1697         __pool_table_insert(pool);
1698
1699         return pool;
1700
1701 bad_mapping_pool:
1702         dm_deferred_set_destroy(pool->all_io_ds);
1703 bad_all_io_ds:
1704         dm_deferred_set_destroy(pool->shared_read_ds);
1705 bad_shared_read_ds:
1706         destroy_workqueue(pool->wq);
1707 bad_wq:
1708         dm_kcopyd_client_destroy(pool->copier);
1709 bad_kcopyd_client:
1710         dm_bio_prison_destroy(pool->prison);
1711 bad_prison:
1712         kfree(pool);
1713 bad_pool:
1714         if (dm_pool_metadata_close(pmd))
1715                 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
1716
1717         return err_p;
1718 }
1719
1720 static void __pool_inc(struct pool *pool)
1721 {
1722         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1723         pool->ref_count++;
1724 }
1725
1726 static void __pool_dec(struct pool *pool)
1727 {
1728         BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
1729         BUG_ON(!pool->ref_count);
1730         if (!--pool->ref_count)
1731                 __pool_destroy(pool);
1732 }
1733
1734 static struct pool *__pool_find(struct mapped_device *pool_md,
1735                                 struct block_device *metadata_dev,
1736                                 unsigned long block_size, int read_only,
1737                                 char **error, int *created)
1738 {
1739         struct pool *pool = __pool_table_lookup_metadata_dev(metadata_dev);
1740
1741         if (pool) {
1742                 if (pool->pool_md != pool_md) {
1743                         *error = "metadata device already in use by a pool";
1744                         return ERR_PTR(-EBUSY);
1745                 }
1746                 __pool_inc(pool);
1747
1748         } else {
1749                 pool = __pool_table_lookup(pool_md);
1750                 if (pool) {
1751                         if (pool->md_dev != metadata_dev) {
1752                                 *error = "different pool cannot replace a pool";
1753                                 return ERR_PTR(-EINVAL);
1754                         }
1755                         __pool_inc(pool);
1756
1757                 } else {
1758                         pool = pool_create(pool_md, metadata_dev, block_size, read_only, error);
1759                         *created = 1;
1760                 }
1761         }
1762
1763         return pool;
1764 }
1765
1766 /*----------------------------------------------------------------
1767  * Pool target methods
1768  *--------------------------------------------------------------*/
1769 static void pool_dtr(struct dm_target *ti)
1770 {
1771         struct pool_c *pt = ti->private;
1772
1773         mutex_lock(&dm_thin_pool_table.mutex);
1774
1775         unbind_control_target(pt->pool, ti);
1776         __pool_dec(pt->pool);
1777         dm_put_device(ti, pt->metadata_dev);
1778         dm_put_device(ti, pt->data_dev);
1779         kfree(pt);
1780
1781         mutex_unlock(&dm_thin_pool_table.mutex);
1782 }
1783
1784 static int parse_pool_features(struct dm_arg_set *as, struct pool_features *pf,
1785                                struct dm_target *ti)
1786 {
1787         int r;
1788         unsigned argc;
1789         const char *arg_name;
1790
1791         static struct dm_arg _args[] = {
1792                 {0, 3, "Invalid number of pool feature arguments"},
1793         };
1794
1795         /*
1796          * No feature arguments supplied.
1797          */
1798         if (!as->argc)
1799                 return 0;
1800
1801         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1802         if (r)
1803                 return -EINVAL;
1804
1805         while (argc && !r) {
1806                 arg_name = dm_shift_arg(as);
1807                 argc--;
1808
1809                 if (!strcasecmp(arg_name, "skip_block_zeroing"))
1810                         pf->zero_new_blocks = false;
1811
1812                 else if (!strcasecmp(arg_name, "ignore_discard"))
1813                         pf->discard_enabled = false;
1814
1815                 else if (!strcasecmp(arg_name, "no_discard_passdown"))
1816                         pf->discard_passdown = false;
1817
1818                 else if (!strcasecmp(arg_name, "read_only"))
1819                         pf->mode = PM_READ_ONLY;
1820
1821                 else {
1822                         ti->error = "Unrecognised pool feature requested";
1823                         r = -EINVAL;
1824                         break;
1825                 }
1826         }
1827
1828         return r;
1829 }
1830
1831 /*
1832  * thin-pool <metadata dev> <data dev>
1833  *           <data block size (sectors)>
1834  *           <low water mark (blocks)>
1835  *           [<#feature args> [<arg>]*]
1836  *
1837  * Optional feature arguments are:
1838  *           skip_block_zeroing: skips the zeroing of newly-provisioned blocks.
1839  *           ignore_discard: disable discard
1840  *           no_discard_passdown: don't pass discards down to the data device
1841  */
1842 static int pool_ctr(struct dm_target *ti, unsigned argc, char **argv)
1843 {
1844         int r, pool_created = 0;
1845         struct pool_c *pt;
1846         struct pool *pool;
1847         struct pool_features pf;
1848         struct dm_arg_set as;
1849         struct dm_dev *data_dev;
1850         unsigned long block_size;
1851         dm_block_t low_water_blocks;
1852         struct dm_dev *metadata_dev;
1853         sector_t metadata_dev_size;
1854         char b[BDEVNAME_SIZE];
1855
1856         /*
1857          * FIXME Remove validation from scope of lock.
1858          */
1859         mutex_lock(&dm_thin_pool_table.mutex);
1860
1861         if (argc < 4) {
1862                 ti->error = "Invalid argument count";
1863                 r = -EINVAL;
1864                 goto out_unlock;
1865         }
1866         as.argc = argc;
1867         as.argv = argv;
1868
1869         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &metadata_dev);
1870         if (r) {
1871                 ti->error = "Error opening metadata block device";
1872                 goto out_unlock;
1873         }
1874
1875         metadata_dev_size = i_size_read(metadata_dev->bdev->bd_inode) >> SECTOR_SHIFT;
1876         if (metadata_dev_size > THIN_METADATA_MAX_SECTORS_WARNING)
1877                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1878                        bdevname(metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1879
1880         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &data_dev);
1881         if (r) {
1882                 ti->error = "Error getting data device";
1883                 goto out_metadata;
1884         }
1885
1886         if (kstrtoul(argv[2], 10, &block_size) || !block_size ||
1887             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1888             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1889             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1890                 ti->error = "Invalid block size";
1891                 r = -EINVAL;
1892                 goto out;
1893         }
1894
1895         if (kstrtoull(argv[3], 10, (unsigned long long *)&low_water_blocks)) {
1896                 ti->error = "Invalid low water mark";
1897                 r = -EINVAL;
1898                 goto out;
1899         }
1900
1901         /*
1902          * Set default pool features.
1903          */
1904         pool_features_init(&pf);
1905
1906         dm_consume_args(&as, 4);
1907         r = parse_pool_features(&as, &pf, ti);
1908         if (r)
1909                 goto out;
1910
1911         pt = kzalloc(sizeof(*pt), GFP_KERNEL);
1912         if (!pt) {
1913                 r = -ENOMEM;
1914                 goto out;
1915         }
1916
1917         pool = __pool_find(dm_table_get_md(ti->table), metadata_dev->bdev,
1918                            block_size, pf.mode == PM_READ_ONLY, &ti->error, &pool_created);
1919         if (IS_ERR(pool)) {
1920                 r = PTR_ERR(pool);
1921                 goto out_free_pt;
1922         }
1923
1924         /*
1925          * 'pool_created' reflects whether this is the first table load.
1926          * Top level discard support is not allowed to be changed after
1927          * initial load.  This would require a pool reload to trigger thin
1928          * device changes.
1929          */
1930         if (!pool_created && pf.discard_enabled != pool->pf.discard_enabled) {
1931                 ti->error = "Discard support cannot be disabled once enabled";
1932                 r = -EINVAL;
1933                 goto out_flags_changed;
1934         }
1935
1936         pt->pool = pool;
1937         pt->ti = ti;
1938         pt->metadata_dev = metadata_dev;
1939         pt->data_dev = data_dev;
1940         pt->low_water_blocks = low_water_blocks;
1941         pt->adjusted_pf = pt->requested_pf = pf;
1942         ti->num_flush_requests = 1;
1943
1944         /*
1945          * Only need to enable discards if the pool should pass
1946          * them down to the data device.  The thin device's discard
1947          * processing will cause mappings to be removed from the btree.
1948          */
1949         if (pf.discard_enabled && pf.discard_passdown) {
1950                 ti->num_discard_requests = 1;
1951
1952                 /*
1953                  * Setting 'discards_supported' circumvents the normal
1954                  * stacking of discard limits (this keeps the pool and
1955                  * thin devices' discard limits consistent).
1956                  */
1957                 ti->discards_supported = true;
1958                 ti->discard_zeroes_data_unsupported = true;
1959         }
1960         ti->private = pt;
1961
1962         pt->callbacks.congested_fn = pool_is_congested;
1963         dm_table_add_target_callbacks(ti->table, &pt->callbacks);
1964
1965         mutex_unlock(&dm_thin_pool_table.mutex);
1966
1967         return 0;
1968
1969 out_flags_changed:
1970         __pool_dec(pool);
1971 out_free_pt:
1972         kfree(pt);
1973 out:
1974         dm_put_device(ti, data_dev);
1975 out_metadata:
1976         dm_put_device(ti, metadata_dev);
1977 out_unlock:
1978         mutex_unlock(&dm_thin_pool_table.mutex);
1979
1980         return r;
1981 }
1982
1983 static int pool_map(struct dm_target *ti, struct bio *bio,
1984                     union map_info *map_context)
1985 {
1986         int r;
1987         struct pool_c *pt = ti->private;
1988         struct pool *pool = pt->pool;
1989         unsigned long flags;
1990
1991         /*
1992          * As this is a singleton target, ti->begin is always zero.
1993          */
1994         spin_lock_irqsave(&pool->lock, flags);
1995         bio->bi_bdev = pt->data_dev->bdev;
1996         r = DM_MAPIO_REMAPPED;
1997         spin_unlock_irqrestore(&pool->lock, flags);
1998
1999         return r;
2000 }
2001
2002 /*
2003  * Retrieves the number of blocks of the data device from
2004  * the superblock and compares it to the actual device size,
2005  * thus resizing the data device in case it has grown.
2006  *
2007  * This both copes with opening preallocated data devices in the ctr
2008  * being followed by a resume
2009  * -and-
2010  * calling the resume method individually after userspace has
2011  * grown the data device in reaction to a table event.
2012  */
2013 static int pool_preresume(struct dm_target *ti)
2014 {
2015         int r;
2016         struct pool_c *pt = ti->private;
2017         struct pool *pool = pt->pool;
2018         sector_t data_size = ti->len;
2019         dm_block_t sb_data_size;
2020
2021         /*
2022          * Take control of the pool object.
2023          */
2024         r = bind_control_target(pool, ti);
2025         if (r)
2026                 return r;
2027
2028         (void) sector_div(data_size, pool->sectors_per_block);
2029
2030         r = dm_pool_get_data_dev_size(pool->pmd, &sb_data_size);
2031         if (r) {
2032                 DMERR("failed to retrieve data device size");
2033                 return r;
2034         }
2035
2036         if (data_size < sb_data_size) {
2037                 DMERR("pool target too small, is %llu blocks (expected %llu)",
2038                       (unsigned long long)data_size, sb_data_size);
2039                 return -EINVAL;
2040
2041         } else if (data_size > sb_data_size) {
2042                 r = dm_pool_resize_data_dev(pool->pmd, data_size);
2043                 if (r) {
2044                         DMERR("failed to resize data device");
2045                         /* FIXME Stricter than necessary: Rollback transaction instead here */
2046                         set_pool_mode(pool, PM_READ_ONLY);
2047                         return r;
2048                 }
2049
2050                 (void) commit_or_fallback(pool);
2051         }
2052
2053         return 0;
2054 }
2055
2056 static void pool_resume(struct dm_target *ti)
2057 {
2058         struct pool_c *pt = ti->private;
2059         struct pool *pool = pt->pool;
2060         unsigned long flags;
2061
2062         spin_lock_irqsave(&pool->lock, flags);
2063         pool->low_water_triggered = 0;
2064         pool->no_free_space = 0;
2065         __requeue_bios(pool);
2066         spin_unlock_irqrestore(&pool->lock, flags);
2067
2068         do_waker(&pool->waker.work);
2069 }
2070
2071 static void pool_postsuspend(struct dm_target *ti)
2072 {
2073         struct pool_c *pt = ti->private;
2074         struct pool *pool = pt->pool;
2075
2076         cancel_delayed_work(&pool->waker);
2077         flush_workqueue(pool->wq);
2078         (void) commit_or_fallback(pool);
2079 }
2080
2081 static int check_arg_count(unsigned argc, unsigned args_required)
2082 {
2083         if (argc != args_required) {
2084                 DMWARN("Message received with %u arguments instead of %u.",
2085                        argc, args_required);
2086                 return -EINVAL;
2087         }
2088
2089         return 0;
2090 }
2091
2092 static int read_dev_id(char *arg, dm_thin_id *dev_id, int warning)
2093 {
2094         if (!kstrtoull(arg, 10, (unsigned long long *)dev_id) &&
2095             *dev_id <= MAX_DEV_ID)
2096                 return 0;
2097
2098         if (warning)
2099                 DMWARN("Message received with invalid device id: %s", arg);
2100
2101         return -EINVAL;
2102 }
2103
2104 static int process_create_thin_mesg(unsigned argc, char **argv, struct pool *pool)
2105 {
2106         dm_thin_id dev_id;
2107         int r;
2108
2109         r = check_arg_count(argc, 2);
2110         if (r)
2111                 return r;
2112
2113         r = read_dev_id(argv[1], &dev_id, 1);
2114         if (r)
2115                 return r;
2116
2117         r = dm_pool_create_thin(pool->pmd, dev_id);
2118         if (r) {
2119                 DMWARN("Creation of new thinly-provisioned device with id %s failed.",
2120                        argv[1]);
2121                 return r;
2122         }
2123
2124         return 0;
2125 }
2126
2127 static int process_create_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2128 {
2129         dm_thin_id dev_id;
2130         dm_thin_id origin_dev_id;
2131         int r;
2132
2133         r = check_arg_count(argc, 3);
2134         if (r)
2135                 return r;
2136
2137         r = read_dev_id(argv[1], &dev_id, 1);
2138         if (r)
2139                 return r;
2140
2141         r = read_dev_id(argv[2], &origin_dev_id, 1);
2142         if (r)
2143                 return r;
2144
2145         r = dm_pool_create_snap(pool->pmd, dev_id, origin_dev_id);
2146         if (r) {
2147                 DMWARN("Creation of new snapshot %s of device %s failed.",
2148                        argv[1], argv[2]);
2149                 return r;
2150         }
2151
2152         return 0;
2153 }
2154
2155 static int process_delete_mesg(unsigned argc, char **argv, struct pool *pool)
2156 {
2157         dm_thin_id dev_id;
2158         int r;
2159
2160         r = check_arg_count(argc, 2);
2161         if (r)
2162                 return r;
2163
2164         r = read_dev_id(argv[1], &dev_id, 1);
2165         if (r)
2166                 return r;
2167
2168         r = dm_pool_delete_thin_device(pool->pmd, dev_id);
2169         if (r)
2170                 DMWARN("Deletion of thin device %s failed.", argv[1]);
2171
2172         return r;
2173 }
2174
2175 static int process_set_transaction_id_mesg(unsigned argc, char **argv, struct pool *pool)
2176 {
2177         dm_thin_id old_id, new_id;
2178         int r;
2179
2180         r = check_arg_count(argc, 3);
2181         if (r)
2182                 return r;
2183
2184         if (kstrtoull(argv[1], 10, (unsigned long long *)&old_id)) {
2185                 DMWARN("set_transaction_id message: Unrecognised id %s.", argv[1]);
2186                 return -EINVAL;
2187         }
2188
2189         if (kstrtoull(argv[2], 10, (unsigned long long *)&new_id)) {
2190                 DMWARN("set_transaction_id message: Unrecognised new id %s.", argv[2]);
2191                 return -EINVAL;
2192         }
2193
2194         r = dm_pool_set_metadata_transaction_id(pool->pmd, old_id, new_id);
2195         if (r) {
2196                 DMWARN("Failed to change transaction id from %s to %s.",
2197                        argv[1], argv[2]);
2198                 return r;
2199         }
2200
2201         return 0;
2202 }
2203
2204 static int process_reserve_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2205 {
2206         int r;
2207
2208         r = check_arg_count(argc, 1);
2209         if (r)
2210                 return r;
2211
2212         (void) commit_or_fallback(pool);
2213
2214         r = dm_pool_reserve_metadata_snap(pool->pmd);
2215         if (r)
2216                 DMWARN("reserve_metadata_snap message failed.");
2217
2218         return r;
2219 }
2220
2221 static int process_release_metadata_snap_mesg(unsigned argc, char **argv, struct pool *pool)
2222 {
2223         int r;
2224
2225         r = check_arg_count(argc, 1);
2226         if (r)
2227                 return r;
2228
2229         r = dm_pool_release_metadata_snap(pool->pmd);
2230         if (r)
2231                 DMWARN("release_metadata_snap message failed.");
2232
2233         return r;
2234 }
2235
2236 /*
2237  * Messages supported:
2238  *   create_thin        <dev_id>
2239  *   create_snap        <dev_id> <origin_id>
2240  *   delete             <dev_id>
2241  *   trim               <dev_id> <new_size_in_sectors>
2242  *   set_transaction_id <current_trans_id> <new_trans_id>
2243  *   reserve_metadata_snap
2244  *   release_metadata_snap
2245  */
2246 static int pool_message(struct dm_target *ti, unsigned argc, char **argv)
2247 {
2248         int r = -EINVAL;
2249         struct pool_c *pt = ti->private;
2250         struct pool *pool = pt->pool;
2251
2252         if (!strcasecmp(argv[0], "create_thin"))
2253                 r = process_create_thin_mesg(argc, argv, pool);
2254
2255         else if (!strcasecmp(argv[0], "create_snap"))
2256                 r = process_create_snap_mesg(argc, argv, pool);
2257
2258         else if (!strcasecmp(argv[0], "delete"))
2259                 r = process_delete_mesg(argc, argv, pool);
2260
2261         else if (!strcasecmp(argv[0], "set_transaction_id"))
2262                 r = process_set_transaction_id_mesg(argc, argv, pool);
2263
2264         else if (!strcasecmp(argv[0], "reserve_metadata_snap"))
2265                 r = process_reserve_metadata_snap_mesg(argc, argv, pool);
2266
2267         else if (!strcasecmp(argv[0], "release_metadata_snap"))
2268                 r = process_release_metadata_snap_mesg(argc, argv, pool);
2269
2270         else
2271                 DMWARN("Unrecognised thin pool target message received: %s", argv[0]);
2272
2273         if (!r)
2274                 (void) commit_or_fallback(pool);
2275
2276         return r;
2277 }
2278
2279 static void emit_flags(struct pool_features *pf, char *result,
2280                        unsigned sz, unsigned maxlen)
2281 {
2282         unsigned count = !pf->zero_new_blocks + !pf->discard_enabled +
2283                 !pf->discard_passdown + (pf->mode == PM_READ_ONLY);
2284         DMEMIT("%u ", count);
2285
2286         if (!pf->zero_new_blocks)
2287                 DMEMIT("skip_block_zeroing ");
2288
2289         if (!pf->discard_enabled)
2290                 DMEMIT("ignore_discard ");
2291
2292         if (!pf->discard_passdown)
2293                 DMEMIT("no_discard_passdown ");
2294
2295         if (pf->mode == PM_READ_ONLY)
2296                 DMEMIT("read_only ");
2297 }
2298
2299 /*
2300  * Status line is:
2301  *    <transaction id> <used metadata sectors>/<total metadata sectors>
2302  *    <used data sectors>/<total data sectors> <held metadata root>
2303  */
2304 static int pool_status(struct dm_target *ti, status_type_t type,
2305                        unsigned status_flags, char *result, unsigned maxlen)
2306 {
2307         int r;
2308         unsigned sz = 0;
2309         uint64_t transaction_id;
2310         dm_block_t nr_free_blocks_data;
2311         dm_block_t nr_free_blocks_metadata;
2312         dm_block_t nr_blocks_data;
2313         dm_block_t nr_blocks_metadata;
2314         dm_block_t held_root;
2315         char buf[BDEVNAME_SIZE];
2316         char buf2[BDEVNAME_SIZE];
2317         struct pool_c *pt = ti->private;
2318         struct pool *pool = pt->pool;
2319
2320         switch (type) {
2321         case STATUSTYPE_INFO:
2322                 if (get_pool_mode(pool) == PM_FAIL) {
2323                         DMEMIT("Fail");
2324                         break;
2325                 }
2326
2327                 /* Commit to ensure statistics aren't out-of-date */
2328                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
2329                         (void) commit_or_fallback(pool);
2330
2331                 r = dm_pool_get_metadata_transaction_id(pool->pmd,
2332                                                         &transaction_id);
2333                 if (r)
2334                         return r;
2335
2336                 r = dm_pool_get_free_metadata_block_count(pool->pmd,
2337                                                           &nr_free_blocks_metadata);
2338                 if (r)
2339                         return r;
2340
2341                 r = dm_pool_get_metadata_dev_size(pool->pmd, &nr_blocks_metadata);
2342                 if (r)
2343                         return r;
2344
2345                 r = dm_pool_get_free_block_count(pool->pmd,
2346                                                  &nr_free_blocks_data);
2347                 if (r)
2348                         return r;
2349
2350                 r = dm_pool_get_data_dev_size(pool->pmd, &nr_blocks_data);
2351                 if (r)
2352                         return r;
2353
2354                 r = dm_pool_get_metadata_snap(pool->pmd, &held_root);
2355                 if (r)
2356                         return r;
2357
2358                 DMEMIT("%llu %llu/%llu %llu/%llu ",
2359                        (unsigned long long)transaction_id,
2360                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2361                        (unsigned long long)nr_blocks_metadata,
2362                        (unsigned long long)(nr_blocks_data - nr_free_blocks_data),
2363                        (unsigned long long)nr_blocks_data);
2364
2365                 if (held_root)
2366                         DMEMIT("%llu ", held_root);
2367                 else
2368                         DMEMIT("- ");
2369
2370                 if (pool->pf.mode == PM_READ_ONLY)
2371                         DMEMIT("ro ");
2372                 else
2373                         DMEMIT("rw ");
2374
2375                 if (!pool->pf.discard_enabled)
2376                         DMEMIT("ignore_discard");
2377                 else if (pool->pf.discard_passdown)
2378                         DMEMIT("discard_passdown");
2379                 else
2380                         DMEMIT("no_discard_passdown");
2381
2382                 break;
2383
2384         case STATUSTYPE_TABLE:
2385                 DMEMIT("%s %s %lu %llu ",
2386                        format_dev_t(buf, pt->metadata_dev->bdev->bd_dev),
2387                        format_dev_t(buf2, pt->data_dev->bdev->bd_dev),
2388                        (unsigned long)pool->sectors_per_block,
2389                        (unsigned long long)pt->low_water_blocks);
2390                 emit_flags(&pt->requested_pf, result, sz, maxlen);
2391                 break;
2392         }
2393
2394         return 0;
2395 }
2396
2397 static int pool_iterate_devices(struct dm_target *ti,
2398                                 iterate_devices_callout_fn fn, void *data)
2399 {
2400         struct pool_c *pt = ti->private;
2401
2402         return fn(ti, pt->data_dev, 0, ti->len, data);
2403 }
2404
2405 static int pool_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
2406                       struct bio_vec *biovec, int max_size)
2407 {
2408         struct pool_c *pt = ti->private;
2409         struct request_queue *q = bdev_get_queue(pt->data_dev->bdev);
2410
2411         if (!q->merge_bvec_fn)
2412                 return max_size;
2413
2414         bvm->bi_bdev = pt->data_dev->bdev;
2415
2416         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2417 }
2418
2419 static bool block_size_is_power_of_two(struct pool *pool)
2420 {
2421         return pool->sectors_per_block_shift >= 0;
2422 }
2423
2424 static void set_discard_limits(struct pool_c *pt, struct queue_limits *limits)
2425 {
2426         struct pool *pool = pt->pool;
2427         struct queue_limits *data_limits;
2428
2429         limits->max_discard_sectors = pool->sectors_per_block;
2430
2431         /*
2432          * discard_granularity is just a hint, and not enforced.
2433          */
2434         if (pt->adjusted_pf.discard_passdown) {
2435                 data_limits = &bdev_get_queue(pt->data_dev->bdev)->limits;
2436                 limits->discard_granularity = data_limits->discard_granularity;
2437         } else if (block_size_is_power_of_two(pool))
2438                 limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
2439         else
2440                 /*
2441                  * Use largest power of 2 that is a factor of sectors_per_block
2442                  * but at least DATA_DEV_BLOCK_SIZE_MIN_SECTORS.
2443                  */
2444                 limits->discard_granularity = max(1 << (ffs(pool->sectors_per_block) - 1),
2445                                                   DATA_DEV_BLOCK_SIZE_MIN_SECTORS) << SECTOR_SHIFT;
2446 }
2447
2448 static void pool_io_hints(struct dm_target *ti, struct queue_limits *limits)
2449 {
2450         struct pool_c *pt = ti->private;
2451         struct pool *pool = pt->pool;
2452
2453         blk_limits_io_min(limits, 0);
2454         blk_limits_io_opt(limits, pool->sectors_per_block << SECTOR_SHIFT);
2455
2456         /*
2457          * pt->adjusted_pf is a staging area for the actual features to use.
2458          * They get transferred to the live pool in bind_control_target()
2459          * called from pool_preresume().
2460          */
2461         if (!pt->adjusted_pf.discard_enabled)
2462                 return;
2463
2464         disable_passdown_if_not_supported(pt);
2465
2466         set_discard_limits(pt, limits);
2467 }
2468
2469 static struct target_type pool_target = {
2470         .name = "thin-pool",
2471         .features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
2472                     DM_TARGET_IMMUTABLE,
2473         .version = {1, 6, 0},
2474         .module = THIS_MODULE,
2475         .ctr = pool_ctr,
2476         .dtr = pool_dtr,
2477         .map = pool_map,
2478         .postsuspend = pool_postsuspend,
2479         .preresume = pool_preresume,
2480         .resume = pool_resume,
2481         .message = pool_message,
2482         .status = pool_status,
2483         .merge = pool_merge,
2484         .iterate_devices = pool_iterate_devices,
2485         .io_hints = pool_io_hints,
2486 };
2487
2488 /*----------------------------------------------------------------
2489  * Thin target methods
2490  *--------------------------------------------------------------*/
2491 static void thin_dtr(struct dm_target *ti)
2492 {
2493         struct thin_c *tc = ti->private;
2494
2495         mutex_lock(&dm_thin_pool_table.mutex);
2496
2497         __pool_dec(tc->pool);
2498         dm_pool_close_thin_device(tc->td);
2499         dm_put_device(ti, tc->pool_dev);
2500         if (tc->origin_dev)
2501                 dm_put_device(ti, tc->origin_dev);
2502         kfree(tc);
2503
2504         mutex_unlock(&dm_thin_pool_table.mutex);
2505 }
2506
2507 /*
2508  * Thin target parameters:
2509  *
2510  * <pool_dev> <dev_id> [origin_dev]
2511  *
2512  * pool_dev: the path to the pool (eg, /dev/mapper/my_pool)
2513  * dev_id: the internal device identifier
2514  * origin_dev: a device external to the pool that should act as the origin
2515  *
2516  * If the pool device has discards disabled, they get disabled for the thin
2517  * device as well.
2518  */
2519 static int thin_ctr(struct dm_target *ti, unsigned argc, char **argv)
2520 {
2521         int r;
2522         struct thin_c *tc;
2523         struct dm_dev *pool_dev, *origin_dev;
2524         struct mapped_device *pool_md;
2525
2526         mutex_lock(&dm_thin_pool_table.mutex);
2527
2528         if (argc != 2 && argc != 3) {
2529                 ti->error = "Invalid argument count";
2530                 r = -EINVAL;
2531                 goto out_unlock;
2532         }
2533
2534         tc = ti->private = kzalloc(sizeof(*tc), GFP_KERNEL);
2535         if (!tc) {
2536                 ti->error = "Out of memory";
2537                 r = -ENOMEM;
2538                 goto out_unlock;
2539         }
2540
2541         if (argc == 3) {
2542                 r = dm_get_device(ti, argv[2], FMODE_READ, &origin_dev);
2543                 if (r) {
2544                         ti->error = "Error opening origin device";
2545                         goto bad_origin_dev;
2546                 }
2547                 tc->origin_dev = origin_dev;
2548         }
2549
2550         r = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &pool_dev);
2551         if (r) {
2552                 ti->error = "Error opening pool device";
2553                 goto bad_pool_dev;
2554         }
2555         tc->pool_dev = pool_dev;
2556
2557         if (read_dev_id(argv[1], (unsigned long long *)&tc->dev_id, 0)) {
2558                 ti->error = "Invalid device id";
2559                 r = -EINVAL;
2560                 goto bad_common;
2561         }
2562
2563         pool_md = dm_get_md(tc->pool_dev->bdev->bd_dev);
2564         if (!pool_md) {
2565                 ti->error = "Couldn't get pool mapped device";
2566                 r = -EINVAL;
2567                 goto bad_common;
2568         }
2569
2570         tc->pool = __pool_table_lookup(pool_md);
2571         if (!tc->pool) {
2572                 ti->error = "Couldn't find pool object";
2573                 r = -EINVAL;
2574                 goto bad_pool_lookup;
2575         }
2576         __pool_inc(tc->pool);
2577
2578         if (get_pool_mode(tc->pool) == PM_FAIL) {
2579                 ti->error = "Couldn't open thin device, Pool is in fail mode";
2580                 goto bad_thin_open;
2581         }
2582
2583         r = dm_pool_open_thin_device(tc->pool->pmd, tc->dev_id, &tc->td);
2584         if (r) {
2585                 ti->error = "Couldn't open thin internal device";
2586                 goto bad_thin_open;
2587         }
2588
2589         r = dm_set_target_max_io_len(ti, tc->pool->sectors_per_block);
2590         if (r)
2591                 goto bad_thin_open;
2592
2593         ti->num_flush_requests = 1;
2594         ti->flush_supported = true;
2595         ti->per_bio_data_size = sizeof(struct dm_thin_endio_hook);
2596
2597         /* In case the pool supports discards, pass them on. */
2598         if (tc->pool->pf.discard_enabled) {
2599                 ti->discards_supported = true;
2600                 ti->num_discard_requests = 1;
2601                 ti->discard_zeroes_data_unsupported = true;
2602                 /* Discard requests must be split on a block boundary */
2603                 ti->split_discard_requests = true;
2604         }
2605
2606         dm_put(pool_md);
2607
2608         mutex_unlock(&dm_thin_pool_table.mutex);
2609
2610         return 0;
2611
2612 bad_thin_open:
2613         __pool_dec(tc->pool);
2614 bad_pool_lookup:
2615         dm_put(pool_md);
2616 bad_common:
2617         dm_put_device(ti, tc->pool_dev);
2618 bad_pool_dev:
2619         if (tc->origin_dev)
2620                 dm_put_device(ti, tc->origin_dev);
2621 bad_origin_dev:
2622         kfree(tc);
2623 out_unlock:
2624         mutex_unlock(&dm_thin_pool_table.mutex);
2625
2626         return r;
2627 }
2628
2629 static int thin_map(struct dm_target *ti, struct bio *bio,
2630                     union map_info *map_context)
2631 {
2632         bio->bi_sector = dm_target_offset(ti, bio->bi_sector);
2633
2634         return thin_bio_map(ti, bio, map_context);
2635 }
2636
2637 static int thin_endio(struct dm_target *ti,
2638                       struct bio *bio, int err,
2639                       union map_info *map_context)
2640 {
2641         unsigned long flags;
2642         struct dm_thin_endio_hook *h = dm_per_bio_data(bio, sizeof(struct dm_thin_endio_hook));
2643         struct list_head work;
2644         struct dm_thin_new_mapping *m, *tmp;
2645         struct pool *pool = h->tc->pool;
2646
2647         if (h->shared_read_entry) {
2648                 INIT_LIST_HEAD(&work);
2649                 dm_deferred_entry_dec(h->shared_read_entry, &work);
2650
2651                 spin_lock_irqsave(&pool->lock, flags);
2652                 list_for_each_entry_safe(m, tmp, &work, list) {
2653                         list_del(&m->list);
2654                         m->quiesced = 1;
2655                         __maybe_add_mapping(m);
2656                 }
2657                 spin_unlock_irqrestore(&pool->lock, flags);
2658         }
2659
2660         if (h->all_io_entry) {
2661                 INIT_LIST_HEAD(&work);
2662                 dm_deferred_entry_dec(h->all_io_entry, &work);
2663                 if (!list_empty(&work)) {
2664                         spin_lock_irqsave(&pool->lock, flags);
2665                         list_for_each_entry_safe(m, tmp, &work, list)
2666                                 list_add(&m->list, &pool->prepared_discards);
2667                         spin_unlock_irqrestore(&pool->lock, flags);
2668                         wake_worker(pool);
2669                 }
2670         }
2671
2672         return 0;
2673 }
2674
2675 static void thin_postsuspend(struct dm_target *ti)
2676 {
2677         if (dm_noflush_suspending(ti))
2678                 requeue_io((struct thin_c *)ti->private);
2679 }
2680
2681 /*
2682  * <nr mapped sectors> <highest mapped sector>
2683  */
2684 static int thin_status(struct dm_target *ti, status_type_t type,
2685                        unsigned status_flags, char *result, unsigned maxlen)
2686 {
2687         int r;
2688         ssize_t sz = 0;
2689         dm_block_t mapped, highest;
2690         char buf[BDEVNAME_SIZE];
2691         struct thin_c *tc = ti->private;
2692
2693         if (get_pool_mode(tc->pool) == PM_FAIL) {
2694                 DMEMIT("Fail");
2695                 return 0;
2696         }
2697
2698         if (!tc->td)
2699                 DMEMIT("-");
2700         else {
2701                 switch (type) {
2702                 case STATUSTYPE_INFO:
2703                         r = dm_thin_get_mapped_count(tc->td, &mapped);
2704                         if (r)
2705                                 return r;
2706
2707                         r = dm_thin_get_highest_mapped_block(tc->td, &highest);
2708                         if (r < 0)
2709                                 return r;
2710
2711                         DMEMIT("%llu ", mapped * tc->pool->sectors_per_block);
2712                         if (r)
2713                                 DMEMIT("%llu", ((highest + 1) *
2714                                                 tc->pool->sectors_per_block) - 1);
2715                         else
2716                                 DMEMIT("-");
2717                         break;
2718
2719                 case STATUSTYPE_TABLE:
2720                         DMEMIT("%s %lu",
2721                                format_dev_t(buf, tc->pool_dev->bdev->bd_dev),
2722                                (unsigned long) tc->dev_id);
2723                         if (tc->origin_dev)
2724                                 DMEMIT(" %s", format_dev_t(buf, tc->origin_dev->bdev->bd_dev));
2725                         break;
2726                 }
2727         }
2728
2729         return 0;
2730 }
2731
2732 static int thin_iterate_devices(struct dm_target *ti,
2733                                 iterate_devices_callout_fn fn, void *data)
2734 {
2735         sector_t blocks;
2736         struct thin_c *tc = ti->private;
2737         struct pool *pool = tc->pool;
2738
2739         /*
2740          * We can't call dm_pool_get_data_dev_size() since that blocks.  So
2741          * we follow a more convoluted path through to the pool's target.
2742          */
2743         if (!pool->ti)
2744                 return 0;       /* nothing is bound */
2745
2746         blocks = pool->ti->len;
2747         (void) sector_div(blocks, pool->sectors_per_block);
2748         if (blocks)
2749                 return fn(ti, tc->pool_dev, 0, pool->sectors_per_block * blocks, data);
2750
2751         return 0;
2752 }
2753
2754 /*
2755  * A thin device always inherits its queue limits from its pool.
2756  */
2757 static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
2758 {
2759         struct thin_c *tc = ti->private;
2760
2761         *limits = bdev_get_queue(tc->pool_dev->bdev)->limits;
2762 }
2763
2764 static struct target_type thin_target = {
2765         .name = "thin",
2766         .version = {1, 6, 0},
2767         .module = THIS_MODULE,
2768         .ctr = thin_ctr,
2769         .dtr = thin_dtr,
2770         .map = thin_map,
2771         .end_io = thin_endio,
2772         .postsuspend = thin_postsuspend,
2773         .status = thin_status,
2774         .iterate_devices = thin_iterate_devices,
2775         .io_hints = thin_io_hints,
2776 };
2777
2778 /*----------------------------------------------------------------*/
2779
2780 static int __init dm_thin_init(void)
2781 {
2782         int r;
2783
2784         pool_table_init();
2785
2786         r = dm_register_target(&thin_target);
2787         if (r)
2788                 return r;
2789
2790         r = dm_register_target(&pool_target);
2791         if (r)
2792                 goto bad_pool_target;
2793
2794         r = -ENOMEM;
2795
2796         _new_mapping_cache = KMEM_CACHE(dm_thin_new_mapping, 0);
2797         if (!_new_mapping_cache)
2798                 goto bad_new_mapping_cache;
2799
2800         return 0;
2801
2802 bad_new_mapping_cache:
2803         dm_unregister_target(&pool_target);
2804 bad_pool_target:
2805         dm_unregister_target(&thin_target);
2806
2807         return r;
2808 }
2809
2810 static void dm_thin_exit(void)
2811 {
2812         dm_unregister_target(&thin_target);
2813         dm_unregister_target(&pool_target);
2814
2815         kmem_cache_destroy(_new_mapping_cache);
2816 }
2817
2818 module_init(dm_thin_init);
2819 module_exit(dm_thin_exit);
2820
2821 MODULE_DESCRIPTION(DM_NAME " thin provisioning target");
2822 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
2823 MODULE_LICENSE("GPL");