]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/md/dm-cache-target.c
Merge remote-tracking branch 'usb-chipidea-next/ci-for-usb-next'
[karo-tx-linux.git] / drivers / md / dm-cache-target.c
1 /*
2  * Copyright (C) 2012 Red Hat. All rights reserved.
3  *
4  * This file is released under the GPL.
5  */
6
7 #include "dm.h"
8 #include "dm-bio-prison.h"
9 #include "dm-bio-record.h"
10 #include "dm-cache-metadata.h"
11
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/vmalloc.h>
20
21 #define DM_MSG_PREFIX "cache"
22
23 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
24         "A percentage of time allocated for copying to and/or from cache");
25
26 /*----------------------------------------------------------------*/
27
28 #define IOT_RESOLUTION 4
29
30 struct io_tracker {
31         spinlock_t lock;
32
33         /*
34          * Sectors of in-flight IO.
35          */
36         sector_t in_flight;
37
38         /*
39          * The time, in jiffies, when this device became idle (if it is
40          * indeed idle).
41          */
42         unsigned long idle_time;
43         unsigned long last_update_time;
44 };
45
46 static void iot_init(struct io_tracker *iot)
47 {
48         spin_lock_init(&iot->lock);
49         iot->in_flight = 0ul;
50         iot->idle_time = 0ul;
51         iot->last_update_time = jiffies;
52 }
53
54 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
55 {
56         if (iot->in_flight)
57                 return false;
58
59         return time_after(jiffies, iot->idle_time + jifs);
60 }
61
62 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
63 {
64         bool r;
65         unsigned long flags;
66
67         spin_lock_irqsave(&iot->lock, flags);
68         r = __iot_idle_for(iot, jifs);
69         spin_unlock_irqrestore(&iot->lock, flags);
70
71         return r;
72 }
73
74 static void iot_io_begin(struct io_tracker *iot, sector_t len)
75 {
76         unsigned long flags;
77
78         spin_lock_irqsave(&iot->lock, flags);
79         iot->in_flight += len;
80         spin_unlock_irqrestore(&iot->lock, flags);
81 }
82
83 static void __iot_io_end(struct io_tracker *iot, sector_t len)
84 {
85         iot->in_flight -= len;
86         if (!iot->in_flight)
87                 iot->idle_time = jiffies;
88 }
89
90 static void iot_io_end(struct io_tracker *iot, sector_t len)
91 {
92         unsigned long flags;
93
94         spin_lock_irqsave(&iot->lock, flags);
95         __iot_io_end(iot, len);
96         spin_unlock_irqrestore(&iot->lock, flags);
97 }
98
99 /*----------------------------------------------------------------*/
100
101 /*
102  * Glossary:
103  *
104  * oblock: index of an origin block
105  * cblock: index of a cache block
106  * promotion: movement of a block from origin to cache
107  * demotion: movement of a block from cache to origin
108  * migration: movement of a block between the origin and cache device,
109  *            either direction
110  */
111
112 /*----------------------------------------------------------------*/
113
114 /*
115  * There are a couple of places where we let a bio run, but want to do some
116  * work before calling its endio function.  We do this by temporarily
117  * changing the endio fn.
118  */
119 struct dm_hook_info {
120         bio_end_io_t *bi_end_io;
121 };
122
123 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
124                         bio_end_io_t *bi_end_io, void *bi_private)
125 {
126         h->bi_end_io = bio->bi_end_io;
127
128         bio->bi_end_io = bi_end_io;
129         bio->bi_private = bi_private;
130 }
131
132 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
133 {
134         bio->bi_end_io = h->bi_end_io;
135 }
136
137 /*----------------------------------------------------------------*/
138
139 #define MIGRATION_POOL_SIZE 128
140 #define COMMIT_PERIOD HZ
141 #define MIGRATION_COUNT_WINDOW 10
142
143 /*
144  * The block size of the device holding cache data must be
145  * between 32KB and 1GB.
146  */
147 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
148 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
149
150 enum cache_metadata_mode {
151         CM_WRITE,               /* metadata may be changed */
152         CM_READ_ONLY,           /* metadata may not be changed */
153         CM_FAIL
154 };
155
156 enum cache_io_mode {
157         /*
158          * Data is written to cached blocks only.  These blocks are marked
159          * dirty.  If you lose the cache device you will lose data.
160          * Potential performance increase for both reads and writes.
161          */
162         CM_IO_WRITEBACK,
163
164         /*
165          * Data is written to both cache and origin.  Blocks are never
166          * dirty.  Potential performance benfit for reads only.
167          */
168         CM_IO_WRITETHROUGH,
169
170         /*
171          * A degraded mode useful for various cache coherency situations
172          * (eg, rolling back snapshots).  Reads and writes always go to the
173          * origin.  If a write goes to a cached oblock, then the cache
174          * block is invalidated.
175          */
176         CM_IO_PASSTHROUGH
177 };
178
179 struct cache_features {
180         enum cache_metadata_mode mode;
181         enum cache_io_mode io_mode;
182 };
183
184 struct cache_stats {
185         atomic_t read_hit;
186         atomic_t read_miss;
187         atomic_t write_hit;
188         atomic_t write_miss;
189         atomic_t demotion;
190         atomic_t promotion;
191         atomic_t copies_avoided;
192         atomic_t cache_cell_clash;
193         atomic_t commit_count;
194         atomic_t discard_count;
195 };
196
197 /*
198  * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
199  * the one-past-the-end value.
200  */
201 struct cblock_range {
202         dm_cblock_t begin;
203         dm_cblock_t end;
204 };
205
206 struct invalidation_request {
207         struct list_head list;
208         struct cblock_range *cblocks;
209
210         atomic_t complete;
211         int err;
212
213         wait_queue_head_t result_wait;
214 };
215
216 struct cache {
217         struct dm_target *ti;
218         struct dm_target_callbacks callbacks;
219
220         struct dm_cache_metadata *cmd;
221
222         /*
223          * Metadata is written to this device.
224          */
225         struct dm_dev *metadata_dev;
226
227         /*
228          * The slower of the two data devices.  Typically a spindle.
229          */
230         struct dm_dev *origin_dev;
231
232         /*
233          * The faster of the two data devices.  Typically an SSD.
234          */
235         struct dm_dev *cache_dev;
236
237         /*
238          * Size of the origin device in _complete_ blocks and native sectors.
239          */
240         dm_oblock_t origin_blocks;
241         sector_t origin_sectors;
242
243         /*
244          * Size of the cache device in blocks.
245          */
246         dm_cblock_t cache_size;
247
248         /*
249          * Fields for converting from sectors to blocks.
250          */
251         uint32_t sectors_per_block;
252         int sectors_per_block_shift;
253
254         spinlock_t lock;
255         struct list_head deferred_cells;
256         struct bio_list deferred_bios;
257         struct bio_list deferred_flush_bios;
258         struct bio_list deferred_writethrough_bios;
259         struct list_head quiesced_migrations;
260         struct list_head completed_migrations;
261         struct list_head need_commit_migrations;
262         sector_t migration_threshold;
263         wait_queue_head_t migration_wait;
264         atomic_t nr_allocated_migrations;
265
266         /*
267          * The number of in flight migrations that are performing
268          * background io. eg, promotion, writeback.
269          */
270         atomic_t nr_io_migrations;
271
272         wait_queue_head_t quiescing_wait;
273         atomic_t quiescing;
274         atomic_t quiescing_ack;
275
276         /*
277          * cache_size entries, dirty if set
278          */
279         atomic_t nr_dirty;
280         unsigned long *dirty_bitset;
281
282         /*
283          * origin_blocks entries, discarded if set.
284          */
285         dm_dblock_t discard_nr_blocks;
286         unsigned long *discard_bitset;
287         uint32_t discard_block_size; /* a power of 2 times sectors per block */
288
289         /*
290          * Rather than reconstructing the table line for the status we just
291          * save it and regurgitate.
292          */
293         unsigned nr_ctr_args;
294         const char **ctr_args;
295
296         struct dm_kcopyd_client *copier;
297         struct workqueue_struct *wq;
298         struct work_struct worker;
299
300         struct delayed_work waker;
301         unsigned long last_commit_jiffies;
302
303         struct dm_bio_prison *prison;
304         struct dm_deferred_set *all_io_ds;
305
306         mempool_t *migration_pool;
307
308         struct dm_cache_policy *policy;
309         unsigned policy_nr_args;
310
311         bool need_tick_bio:1;
312         bool sized:1;
313         bool invalidate:1;
314         bool commit_requested:1;
315         bool loaded_mappings:1;
316         bool loaded_discards:1;
317
318         /*
319          * Cache features such as write-through.
320          */
321         struct cache_features features;
322
323         struct cache_stats stats;
324
325         /*
326          * Invalidation fields.
327          */
328         spinlock_t invalidation_lock;
329         struct list_head invalidation_requests;
330
331         struct io_tracker origin_tracker;
332 };
333
334 struct per_bio_data {
335         bool tick:1;
336         unsigned req_nr:2;
337         struct dm_deferred_entry *all_io_entry;
338         struct dm_hook_info hook_info;
339         sector_t len;
340
341         /*
342          * writethrough fields.  These MUST remain at the end of this
343          * structure and the 'cache' member must be the first as it
344          * is used to determine the offset of the writethrough fields.
345          */
346         struct cache *cache;
347         dm_cblock_t cblock;
348         struct dm_bio_details bio_details;
349 };
350
351 struct dm_cache_migration {
352         struct list_head list;
353         struct cache *cache;
354
355         unsigned long start_jiffies;
356         dm_oblock_t old_oblock;
357         dm_oblock_t new_oblock;
358         dm_cblock_t cblock;
359
360         bool err:1;
361         bool discard:1;
362         bool writeback:1;
363         bool demote:1;
364         bool promote:1;
365         bool requeue_holder:1;
366         bool invalidate:1;
367
368         struct dm_bio_prison_cell *old_ocell;
369         struct dm_bio_prison_cell *new_ocell;
370 };
371
372 /*
373  * Processing a bio in the worker thread may require these memory
374  * allocations.  We prealloc to avoid deadlocks (the same worker thread
375  * frees them back to the mempool).
376  */
377 struct prealloc {
378         struct dm_cache_migration *mg;
379         struct dm_bio_prison_cell *cell1;
380         struct dm_bio_prison_cell *cell2;
381 };
382
383 static enum cache_metadata_mode get_cache_mode(struct cache *cache);
384
385 static void wake_worker(struct cache *cache)
386 {
387         queue_work(cache->wq, &cache->worker);
388 }
389
390 /*----------------------------------------------------------------*/
391
392 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
393 {
394         /* FIXME: change to use a local slab. */
395         return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
396 }
397
398 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
399 {
400         dm_bio_prison_free_cell(cache->prison, cell);
401 }
402
403 static struct dm_cache_migration *alloc_migration(struct cache *cache)
404 {
405         struct dm_cache_migration *mg;
406
407         mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
408         if (mg) {
409                 mg->cache = cache;
410                 atomic_inc(&mg->cache->nr_allocated_migrations);
411         }
412
413         return mg;
414 }
415
416 static void free_migration(struct dm_cache_migration *mg)
417 {
418         struct cache *cache = mg->cache;
419
420         if (atomic_dec_and_test(&cache->nr_allocated_migrations))
421                 wake_up(&cache->migration_wait);
422
423         mempool_free(mg, cache->migration_pool);
424 }
425
426 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
427 {
428         if (!p->mg) {
429                 p->mg = alloc_migration(cache);
430                 if (!p->mg)
431                         return -ENOMEM;
432         }
433
434         if (!p->cell1) {
435                 p->cell1 = alloc_prison_cell(cache);
436                 if (!p->cell1)
437                         return -ENOMEM;
438         }
439
440         if (!p->cell2) {
441                 p->cell2 = alloc_prison_cell(cache);
442                 if (!p->cell2)
443                         return -ENOMEM;
444         }
445
446         return 0;
447 }
448
449 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
450 {
451         if (p->cell2)
452                 free_prison_cell(cache, p->cell2);
453
454         if (p->cell1)
455                 free_prison_cell(cache, p->cell1);
456
457         if (p->mg)
458                 free_migration(p->mg);
459 }
460
461 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
462 {
463         struct dm_cache_migration *mg = p->mg;
464
465         BUG_ON(!mg);
466         p->mg = NULL;
467
468         return mg;
469 }
470
471 /*
472  * You must have a cell within the prealloc struct to return.  If not this
473  * function will BUG() rather than returning NULL.
474  */
475 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
476 {
477         struct dm_bio_prison_cell *r = NULL;
478
479         if (p->cell1) {
480                 r = p->cell1;
481                 p->cell1 = NULL;
482
483         } else if (p->cell2) {
484                 r = p->cell2;
485                 p->cell2 = NULL;
486         } else
487                 BUG();
488
489         return r;
490 }
491
492 /*
493  * You can't have more than two cells in a prealloc struct.  BUG() will be
494  * called if you try and overfill.
495  */
496 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
497 {
498         if (!p->cell2)
499                 p->cell2 = cell;
500
501         else if (!p->cell1)
502                 p->cell1 = cell;
503
504         else
505                 BUG();
506 }
507
508 /*----------------------------------------------------------------*/
509
510 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key *key)
511 {
512         key->virtual = 0;
513         key->dev = 0;
514         key->block_begin = from_oblock(begin);
515         key->block_end = from_oblock(end);
516 }
517
518 /*
519  * The caller hands in a preallocated cell, and a free function for it.
520  * The cell will be freed if there's an error, or if it wasn't used because
521  * a cell with that key already exists.
522  */
523 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
524
525 static int bio_detain_range(struct cache *cache, dm_oblock_t oblock_begin, dm_oblock_t oblock_end,
526                             struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
527                             cell_free_fn free_fn, void *free_context,
528                             struct dm_bio_prison_cell **cell_result)
529 {
530         int r;
531         struct dm_cell_key key;
532
533         build_key(oblock_begin, oblock_end, &key);
534         r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
535         if (r)
536                 free_fn(free_context, cell_prealloc);
537
538         return r;
539 }
540
541 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
542                       struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
543                       cell_free_fn free_fn, void *free_context,
544                       struct dm_bio_prison_cell **cell_result)
545 {
546         dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
547         return bio_detain_range(cache, oblock, end, bio,
548                                 cell_prealloc, free_fn, free_context, cell_result);
549 }
550
551 static int get_cell(struct cache *cache,
552                     dm_oblock_t oblock,
553                     struct prealloc *structs,
554                     struct dm_bio_prison_cell **cell_result)
555 {
556         int r;
557         struct dm_cell_key key;
558         struct dm_bio_prison_cell *cell_prealloc;
559
560         cell_prealloc = prealloc_get_cell(structs);
561
562         build_key(oblock, to_oblock(from_oblock(oblock) + 1ULL), &key);
563         r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
564         if (r)
565                 prealloc_put_cell(structs, cell_prealloc);
566
567         return r;
568 }
569
570 /*----------------------------------------------------------------*/
571
572 static bool is_dirty(struct cache *cache, dm_cblock_t b)
573 {
574         return test_bit(from_cblock(b), cache->dirty_bitset);
575 }
576
577 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
578 {
579         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
580                 atomic_inc(&cache->nr_dirty);
581                 policy_set_dirty(cache->policy, oblock);
582         }
583 }
584
585 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
586 {
587         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
588                 policy_clear_dirty(cache->policy, oblock);
589                 if (atomic_dec_return(&cache->nr_dirty) == 0)
590                         dm_table_event(cache->ti->table);
591         }
592 }
593
594 /*----------------------------------------------------------------*/
595
596 static bool block_size_is_power_of_two(struct cache *cache)
597 {
598         return cache->sectors_per_block_shift >= 0;
599 }
600
601 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
602 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
603 __always_inline
604 #endif
605 static dm_block_t block_div(dm_block_t b, uint32_t n)
606 {
607         do_div(b, n);
608
609         return b;
610 }
611
612 static dm_block_t oblocks_per_dblock(struct cache *cache)
613 {
614         dm_block_t oblocks = cache->discard_block_size;
615
616         if (block_size_is_power_of_two(cache))
617                 oblocks >>= cache->sectors_per_block_shift;
618         else
619                 oblocks = block_div(oblocks, cache->sectors_per_block);
620
621         return oblocks;
622 }
623
624 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
625 {
626         return to_dblock(block_div(from_oblock(oblock),
627                                    oblocks_per_dblock(cache)));
628 }
629
630 static dm_oblock_t dblock_to_oblock(struct cache *cache, dm_dblock_t dblock)
631 {
632         return to_oblock(from_dblock(dblock) * oblocks_per_dblock(cache));
633 }
634
635 static void set_discard(struct cache *cache, dm_dblock_t b)
636 {
637         unsigned long flags;
638
639         BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
640         atomic_inc(&cache->stats.discard_count);
641
642         spin_lock_irqsave(&cache->lock, flags);
643         set_bit(from_dblock(b), cache->discard_bitset);
644         spin_unlock_irqrestore(&cache->lock, flags);
645 }
646
647 static void clear_discard(struct cache *cache, dm_dblock_t b)
648 {
649         unsigned long flags;
650
651         spin_lock_irqsave(&cache->lock, flags);
652         clear_bit(from_dblock(b), cache->discard_bitset);
653         spin_unlock_irqrestore(&cache->lock, flags);
654 }
655
656 static bool is_discarded(struct cache *cache, dm_dblock_t b)
657 {
658         int r;
659         unsigned long flags;
660
661         spin_lock_irqsave(&cache->lock, flags);
662         r = test_bit(from_dblock(b), cache->discard_bitset);
663         spin_unlock_irqrestore(&cache->lock, flags);
664
665         return r;
666 }
667
668 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
669 {
670         int r;
671         unsigned long flags;
672
673         spin_lock_irqsave(&cache->lock, flags);
674         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
675                      cache->discard_bitset);
676         spin_unlock_irqrestore(&cache->lock, flags);
677
678         return r;
679 }
680
681 /*----------------------------------------------------------------*/
682
683 static void load_stats(struct cache *cache)
684 {
685         struct dm_cache_statistics stats;
686
687         dm_cache_metadata_get_stats(cache->cmd, &stats);
688         atomic_set(&cache->stats.read_hit, stats.read_hits);
689         atomic_set(&cache->stats.read_miss, stats.read_misses);
690         atomic_set(&cache->stats.write_hit, stats.write_hits);
691         atomic_set(&cache->stats.write_miss, stats.write_misses);
692 }
693
694 static void save_stats(struct cache *cache)
695 {
696         struct dm_cache_statistics stats;
697
698         if (get_cache_mode(cache) >= CM_READ_ONLY)
699                 return;
700
701         stats.read_hits = atomic_read(&cache->stats.read_hit);
702         stats.read_misses = atomic_read(&cache->stats.read_miss);
703         stats.write_hits = atomic_read(&cache->stats.write_hit);
704         stats.write_misses = atomic_read(&cache->stats.write_miss);
705
706         dm_cache_metadata_set_stats(cache->cmd, &stats);
707 }
708
709 /*----------------------------------------------------------------
710  * Per bio data
711  *--------------------------------------------------------------*/
712
713 /*
714  * If using writeback, leave out struct per_bio_data's writethrough fields.
715  */
716 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
717 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
718
719 static bool writethrough_mode(struct cache_features *f)
720 {
721         return f->io_mode == CM_IO_WRITETHROUGH;
722 }
723
724 static bool writeback_mode(struct cache_features *f)
725 {
726         return f->io_mode == CM_IO_WRITEBACK;
727 }
728
729 static bool passthrough_mode(struct cache_features *f)
730 {
731         return f->io_mode == CM_IO_PASSTHROUGH;
732 }
733
734 static size_t get_per_bio_data_size(struct cache *cache)
735 {
736         return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
737 }
738
739 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
740 {
741         struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
742         BUG_ON(!pb);
743         return pb;
744 }
745
746 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
747 {
748         struct per_bio_data *pb = get_per_bio_data(bio, data_size);
749
750         pb->tick = false;
751         pb->req_nr = dm_bio_get_target_bio_nr(bio);
752         pb->all_io_entry = NULL;
753         pb->len = 0;
754
755         return pb;
756 }
757
758 /*----------------------------------------------------------------
759  * Remapping
760  *--------------------------------------------------------------*/
761 static void remap_to_origin(struct cache *cache, struct bio *bio)
762 {
763         bio->bi_bdev = cache->origin_dev->bdev;
764 }
765
766 static void remap_to_cache(struct cache *cache, struct bio *bio,
767                            dm_cblock_t cblock)
768 {
769         sector_t bi_sector = bio->bi_iter.bi_sector;
770         sector_t block = from_cblock(cblock);
771
772         bio->bi_bdev = cache->cache_dev->bdev;
773         if (!block_size_is_power_of_two(cache))
774                 bio->bi_iter.bi_sector =
775                         (block * cache->sectors_per_block) +
776                         sector_div(bi_sector, cache->sectors_per_block);
777         else
778                 bio->bi_iter.bi_sector =
779                         (block << cache->sectors_per_block_shift) |
780                         (bi_sector & (cache->sectors_per_block - 1));
781 }
782
783 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
784 {
785         unsigned long flags;
786         size_t pb_data_size = get_per_bio_data_size(cache);
787         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
788
789         spin_lock_irqsave(&cache->lock, flags);
790         if (cache->need_tick_bio &&
791             !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
792                 pb->tick = true;
793                 cache->need_tick_bio = false;
794         }
795         spin_unlock_irqrestore(&cache->lock, flags);
796 }
797
798 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
799                                   dm_oblock_t oblock)
800 {
801         check_if_tick_bio_needed(cache, bio);
802         remap_to_origin(cache, bio);
803         if (bio_data_dir(bio) == WRITE)
804                 clear_discard(cache, oblock_to_dblock(cache, oblock));
805 }
806
807 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
808                                  dm_oblock_t oblock, dm_cblock_t cblock)
809 {
810         check_if_tick_bio_needed(cache, bio);
811         remap_to_cache(cache, bio, cblock);
812         if (bio_data_dir(bio) == WRITE) {
813                 set_dirty(cache, oblock, cblock);
814                 clear_discard(cache, oblock_to_dblock(cache, oblock));
815         }
816 }
817
818 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
819 {
820         sector_t block_nr = bio->bi_iter.bi_sector;
821
822         if (!block_size_is_power_of_two(cache))
823                 (void) sector_div(block_nr, cache->sectors_per_block);
824         else
825                 block_nr >>= cache->sectors_per_block_shift;
826
827         return to_oblock(block_nr);
828 }
829
830 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
831 {
832         return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
833 }
834
835 /*
836  * You must increment the deferred set whilst the prison cell is held.  To
837  * encourage this, we ask for 'cell' to be passed in.
838  */
839 static void inc_ds(struct cache *cache, struct bio *bio,
840                    struct dm_bio_prison_cell *cell)
841 {
842         size_t pb_data_size = get_per_bio_data_size(cache);
843         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
844
845         BUG_ON(!cell);
846         BUG_ON(pb->all_io_entry);
847
848         pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
849 }
850
851 static bool accountable_bio(struct cache *cache, struct bio *bio)
852 {
853         return ((bio->bi_bdev == cache->origin_dev->bdev) &&
854                 !(bio->bi_rw & REQ_DISCARD));
855 }
856
857 static void accounted_begin(struct cache *cache, struct bio *bio)
858 {
859         size_t pb_data_size = get_per_bio_data_size(cache);
860         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
861
862         if (accountable_bio(cache, bio)) {
863                 pb->len = bio_sectors(bio);
864                 iot_io_begin(&cache->origin_tracker, pb->len);
865         }
866 }
867
868 static void accounted_complete(struct cache *cache, struct bio *bio)
869 {
870         size_t pb_data_size = get_per_bio_data_size(cache);
871         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
872
873         iot_io_end(&cache->origin_tracker, pb->len);
874 }
875
876 static void accounted_request(struct cache *cache, struct bio *bio)
877 {
878         accounted_begin(cache, bio);
879         generic_make_request(bio);
880 }
881
882 static void issue(struct cache *cache, struct bio *bio)
883 {
884         unsigned long flags;
885
886         if (!bio_triggers_commit(cache, bio)) {
887                 accounted_request(cache, bio);
888                 return;
889         }
890
891         /*
892          * Batch together any bios that trigger commits and then issue a
893          * single commit for them in do_worker().
894          */
895         spin_lock_irqsave(&cache->lock, flags);
896         cache->commit_requested = true;
897         bio_list_add(&cache->deferred_flush_bios, bio);
898         spin_unlock_irqrestore(&cache->lock, flags);
899 }
900
901 static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
902 {
903         inc_ds(cache, bio, cell);
904         issue(cache, bio);
905 }
906
907 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
908 {
909         unsigned long flags;
910
911         spin_lock_irqsave(&cache->lock, flags);
912         bio_list_add(&cache->deferred_writethrough_bios, bio);
913         spin_unlock_irqrestore(&cache->lock, flags);
914
915         wake_worker(cache);
916 }
917
918 static void writethrough_endio(struct bio *bio)
919 {
920         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
921
922         dm_unhook_bio(&pb->hook_info, bio);
923
924         if (bio->bi_error) {
925                 bio_endio(bio);
926                 return;
927         }
928
929         dm_bio_restore(&pb->bio_details, bio);
930         remap_to_cache(pb->cache, bio, pb->cblock);
931
932         /*
933          * We can't issue this bio directly, since we're in interrupt
934          * context.  So it gets put on a bio list for processing by the
935          * worker thread.
936          */
937         defer_writethrough_bio(pb->cache, bio);
938 }
939
940 /*
941  * When running in writethrough mode we need to send writes to clean blocks
942  * to both the cache and origin devices.  In future we'd like to clone the
943  * bio and send them in parallel, but for now we're doing them in
944  * series as this is easier.
945  */
946 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
947                                        dm_oblock_t oblock, dm_cblock_t cblock)
948 {
949         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
950
951         pb->cache = cache;
952         pb->cblock = cblock;
953         dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
954         dm_bio_record(&pb->bio_details, bio);
955
956         remap_to_origin_clear_discard(pb->cache, bio, oblock);
957 }
958
959 /*----------------------------------------------------------------
960  * Failure modes
961  *--------------------------------------------------------------*/
962 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
963 {
964         return cache->features.mode;
965 }
966
967 static const char *cache_device_name(struct cache *cache)
968 {
969         return dm_device_name(dm_table_get_md(cache->ti->table));
970 }
971
972 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
973 {
974         const char *descs[] = {
975                 "write",
976                 "read-only",
977                 "fail"
978         };
979
980         dm_table_event(cache->ti->table);
981         DMINFO("%s: switching cache to %s mode",
982                cache_device_name(cache), descs[(int)mode]);
983 }
984
985 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
986 {
987         bool needs_check = dm_cache_metadata_needs_check(cache->cmd);
988         enum cache_metadata_mode old_mode = get_cache_mode(cache);
989
990         if (new_mode == CM_WRITE && needs_check) {
991                 DMERR("%s: unable to switch cache to write mode until repaired.",
992                       cache_device_name(cache));
993                 if (old_mode != new_mode)
994                         new_mode = old_mode;
995                 else
996                         new_mode = CM_READ_ONLY;
997         }
998
999         /* Never move out of fail mode */
1000         if (old_mode == CM_FAIL)
1001                 new_mode = CM_FAIL;
1002
1003         switch (new_mode) {
1004         case CM_FAIL:
1005         case CM_READ_ONLY:
1006                 dm_cache_metadata_set_read_only(cache->cmd);
1007                 break;
1008
1009         case CM_WRITE:
1010                 dm_cache_metadata_set_read_write(cache->cmd);
1011                 break;
1012         }
1013
1014         cache->features.mode = new_mode;
1015
1016         if (new_mode != old_mode)
1017                 notify_mode_switch(cache, new_mode);
1018 }
1019
1020 static void abort_transaction(struct cache *cache)
1021 {
1022         const char *dev_name = cache_device_name(cache);
1023
1024         if (get_cache_mode(cache) >= CM_READ_ONLY)
1025                 return;
1026
1027         if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1028                 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1029                 set_cache_mode(cache, CM_FAIL);
1030         }
1031
1032         DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1033         if (dm_cache_metadata_abort(cache->cmd)) {
1034                 DMERR("%s: failed to abort metadata transaction", dev_name);
1035                 set_cache_mode(cache, CM_FAIL);
1036         }
1037 }
1038
1039 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1040 {
1041         DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1042                     cache_device_name(cache), op, r);
1043         abort_transaction(cache);
1044         set_cache_mode(cache, CM_READ_ONLY);
1045 }
1046
1047 /*----------------------------------------------------------------
1048  * Migration processing
1049  *
1050  * Migration covers moving data from the origin device to the cache, or
1051  * vice versa.
1052  *--------------------------------------------------------------*/
1053 static void inc_io_migrations(struct cache *cache)
1054 {
1055         atomic_inc(&cache->nr_io_migrations);
1056 }
1057
1058 static void dec_io_migrations(struct cache *cache)
1059 {
1060         atomic_dec(&cache->nr_io_migrations);
1061 }
1062
1063 static bool discard_or_flush(struct bio *bio)
1064 {
1065         return bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD);
1066 }
1067
1068 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1069 {
1070         if (discard_or_flush(cell->holder)) {
1071                 /*
1072                  * We have to handle these bios individually.
1073                  */
1074                 dm_cell_release(cache->prison, cell, &cache->deferred_bios);
1075                 free_prison_cell(cache, cell);
1076         } else
1077                 list_add_tail(&cell->user_list, &cache->deferred_cells);
1078 }
1079
1080 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1081 {
1082         unsigned long flags;
1083
1084         if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1085                 /*
1086                  * There was no prisoner to promote to holder, the
1087                  * cell has been released.
1088                  */
1089                 free_prison_cell(cache, cell);
1090                 return;
1091         }
1092
1093         spin_lock_irqsave(&cache->lock, flags);
1094         __cell_defer(cache, cell);
1095         spin_unlock_irqrestore(&cache->lock, flags);
1096
1097         wake_worker(cache);
1098 }
1099
1100 static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1101 {
1102         dm_cell_error(cache->prison, cell, err);
1103         free_prison_cell(cache, cell);
1104 }
1105
1106 static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1107 {
1108         cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1109 }
1110
1111 static void free_io_migration(struct dm_cache_migration *mg)
1112 {
1113         struct cache *cache = mg->cache;
1114
1115         dec_io_migrations(cache);
1116         free_migration(mg);
1117         wake_worker(cache);
1118 }
1119
1120 static void migration_failure(struct dm_cache_migration *mg)
1121 {
1122         struct cache *cache = mg->cache;
1123         const char *dev_name = cache_device_name(cache);
1124
1125         if (mg->writeback) {
1126                 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1127                 set_dirty(cache, mg->old_oblock, mg->cblock);
1128                 cell_defer(cache, mg->old_ocell, false);
1129
1130         } else if (mg->demote) {
1131                 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1132                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1133
1134                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1135                 if (mg->promote)
1136                         cell_defer(cache, mg->new_ocell, true);
1137         } else {
1138                 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1139                 policy_remove_mapping(cache->policy, mg->new_oblock);
1140                 cell_defer(cache, mg->new_ocell, true);
1141         }
1142
1143         free_io_migration(mg);
1144 }
1145
1146 static void migration_success_pre_commit(struct dm_cache_migration *mg)
1147 {
1148         int r;
1149         unsigned long flags;
1150         struct cache *cache = mg->cache;
1151
1152         if (mg->writeback) {
1153                 clear_dirty(cache, mg->old_oblock, mg->cblock);
1154                 cell_defer(cache, mg->old_ocell, false);
1155                 free_io_migration(mg);
1156                 return;
1157
1158         } else if (mg->demote) {
1159                 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1160                 if (r) {
1161                         DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1162                                     cache_device_name(cache));
1163                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1164                         policy_force_mapping(cache->policy, mg->new_oblock,
1165                                              mg->old_oblock);
1166                         if (mg->promote)
1167                                 cell_defer(cache, mg->new_ocell, true);
1168                         free_io_migration(mg);
1169                         return;
1170                 }
1171         } else {
1172                 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1173                 if (r) {
1174                         DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1175                                     cache_device_name(cache));
1176                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1177                         policy_remove_mapping(cache->policy, mg->new_oblock);
1178                         free_io_migration(mg);
1179                         return;
1180                 }
1181         }
1182
1183         spin_lock_irqsave(&cache->lock, flags);
1184         list_add_tail(&mg->list, &cache->need_commit_migrations);
1185         cache->commit_requested = true;
1186         spin_unlock_irqrestore(&cache->lock, flags);
1187 }
1188
1189 static void migration_success_post_commit(struct dm_cache_migration *mg)
1190 {
1191         unsigned long flags;
1192         struct cache *cache = mg->cache;
1193
1194         if (mg->writeback) {
1195                 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1196                              cache_device_name(cache));
1197                 return;
1198
1199         } else if (mg->demote) {
1200                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1201
1202                 if (mg->promote) {
1203                         mg->demote = false;
1204
1205                         spin_lock_irqsave(&cache->lock, flags);
1206                         list_add_tail(&mg->list, &cache->quiesced_migrations);
1207                         spin_unlock_irqrestore(&cache->lock, flags);
1208
1209                 } else {
1210                         if (mg->invalidate)
1211                                 policy_remove_mapping(cache->policy, mg->old_oblock);
1212                         free_io_migration(mg);
1213                 }
1214
1215         } else {
1216                 if (mg->requeue_holder) {
1217                         clear_dirty(cache, mg->new_oblock, mg->cblock);
1218                         cell_defer(cache, mg->new_ocell, true);
1219                 } else {
1220                         /*
1221                          * The block was promoted via an overwrite, so it's dirty.
1222                          */
1223                         set_dirty(cache, mg->new_oblock, mg->cblock);
1224                         bio_endio(mg->new_ocell->holder);
1225                         cell_defer(cache, mg->new_ocell, false);
1226                 }
1227                 free_io_migration(mg);
1228         }
1229 }
1230
1231 static void copy_complete(int read_err, unsigned long write_err, void *context)
1232 {
1233         unsigned long flags;
1234         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1235         struct cache *cache = mg->cache;
1236
1237         if (read_err || write_err)
1238                 mg->err = true;
1239
1240         spin_lock_irqsave(&cache->lock, flags);
1241         list_add_tail(&mg->list, &cache->completed_migrations);
1242         spin_unlock_irqrestore(&cache->lock, flags);
1243
1244         wake_worker(cache);
1245 }
1246
1247 static void issue_copy(struct dm_cache_migration *mg)
1248 {
1249         int r;
1250         struct dm_io_region o_region, c_region;
1251         struct cache *cache = mg->cache;
1252         sector_t cblock = from_cblock(mg->cblock);
1253
1254         o_region.bdev = cache->origin_dev->bdev;
1255         o_region.count = cache->sectors_per_block;
1256
1257         c_region.bdev = cache->cache_dev->bdev;
1258         c_region.sector = cblock * cache->sectors_per_block;
1259         c_region.count = cache->sectors_per_block;
1260
1261         if (mg->writeback || mg->demote) {
1262                 /* demote */
1263                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1264                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1265         } else {
1266                 /* promote */
1267                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1268                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1269         }
1270
1271         if (r < 0) {
1272                 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1273                 migration_failure(mg);
1274         }
1275 }
1276
1277 static void overwrite_endio(struct bio *bio)
1278 {
1279         struct dm_cache_migration *mg = bio->bi_private;
1280         struct cache *cache = mg->cache;
1281         size_t pb_data_size = get_per_bio_data_size(cache);
1282         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1283         unsigned long flags;
1284
1285         dm_unhook_bio(&pb->hook_info, bio);
1286
1287         if (bio->bi_error)
1288                 mg->err = true;
1289
1290         mg->requeue_holder = false;
1291
1292         spin_lock_irqsave(&cache->lock, flags);
1293         list_add_tail(&mg->list, &cache->completed_migrations);
1294         spin_unlock_irqrestore(&cache->lock, flags);
1295
1296         wake_worker(cache);
1297 }
1298
1299 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1300 {
1301         size_t pb_data_size = get_per_bio_data_size(mg->cache);
1302         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1303
1304         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1305         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1306
1307         /*
1308          * No need to inc_ds() here, since the cell will be held for the
1309          * duration of the io.
1310          */
1311         accounted_request(mg->cache, bio);
1312 }
1313
1314 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1315 {
1316         return (bio_data_dir(bio) == WRITE) &&
1317                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1318 }
1319
1320 static void avoid_copy(struct dm_cache_migration *mg)
1321 {
1322         atomic_inc(&mg->cache->stats.copies_avoided);
1323         migration_success_pre_commit(mg);
1324 }
1325
1326 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1327                                      dm_dblock_t *b, dm_dblock_t *e)
1328 {
1329         sector_t sb = bio->bi_iter.bi_sector;
1330         sector_t se = bio_end_sector(bio);
1331
1332         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1333
1334         if (se - sb < cache->discard_block_size)
1335                 *e = *b;
1336         else
1337                 *e = to_dblock(block_div(se, cache->discard_block_size));
1338 }
1339
1340 static void issue_discard(struct dm_cache_migration *mg)
1341 {
1342         dm_dblock_t b, e;
1343         struct bio *bio = mg->new_ocell->holder;
1344         struct cache *cache = mg->cache;
1345
1346         calc_discard_block_range(cache, bio, &b, &e);
1347         while (b != e) {
1348                 set_discard(cache, b);
1349                 b = to_dblock(from_dblock(b) + 1);
1350         }
1351
1352         bio_endio(bio);
1353         cell_defer(cache, mg->new_ocell, false);
1354         free_migration(mg);
1355         wake_worker(cache);
1356 }
1357
1358 static void issue_copy_or_discard(struct dm_cache_migration *mg)
1359 {
1360         bool avoid;
1361         struct cache *cache = mg->cache;
1362
1363         if (mg->discard) {
1364                 issue_discard(mg);
1365                 return;
1366         }
1367
1368         if (mg->writeback || mg->demote)
1369                 avoid = !is_dirty(cache, mg->cblock) ||
1370                         is_discarded_oblock(cache, mg->old_oblock);
1371         else {
1372                 struct bio *bio = mg->new_ocell->holder;
1373
1374                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1375
1376                 if (writeback_mode(&cache->features) &&
1377                     !avoid && bio_writes_complete_block(cache, bio)) {
1378                         issue_overwrite(mg, bio);
1379                         return;
1380                 }
1381         }
1382
1383         avoid ? avoid_copy(mg) : issue_copy(mg);
1384 }
1385
1386 static void complete_migration(struct dm_cache_migration *mg)
1387 {
1388         if (mg->err)
1389                 migration_failure(mg);
1390         else
1391                 migration_success_pre_commit(mg);
1392 }
1393
1394 static void process_migrations(struct cache *cache, struct list_head *head,
1395                                void (*fn)(struct dm_cache_migration *))
1396 {
1397         unsigned long flags;
1398         struct list_head list;
1399         struct dm_cache_migration *mg, *tmp;
1400
1401         INIT_LIST_HEAD(&list);
1402         spin_lock_irqsave(&cache->lock, flags);
1403         list_splice_init(head, &list);
1404         spin_unlock_irqrestore(&cache->lock, flags);
1405
1406         list_for_each_entry_safe(mg, tmp, &list, list)
1407                 fn(mg);
1408 }
1409
1410 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1411 {
1412         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1413 }
1414
1415 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1416 {
1417         unsigned long flags;
1418         struct cache *cache = mg->cache;
1419
1420         spin_lock_irqsave(&cache->lock, flags);
1421         __queue_quiesced_migration(mg);
1422         spin_unlock_irqrestore(&cache->lock, flags);
1423
1424         wake_worker(cache);
1425 }
1426
1427 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1428 {
1429         unsigned long flags;
1430         struct dm_cache_migration *mg, *tmp;
1431
1432         spin_lock_irqsave(&cache->lock, flags);
1433         list_for_each_entry_safe(mg, tmp, work, list)
1434                 __queue_quiesced_migration(mg);
1435         spin_unlock_irqrestore(&cache->lock, flags);
1436
1437         wake_worker(cache);
1438 }
1439
1440 static void check_for_quiesced_migrations(struct cache *cache,
1441                                           struct per_bio_data *pb)
1442 {
1443         struct list_head work;
1444
1445         if (!pb->all_io_entry)
1446                 return;
1447
1448         INIT_LIST_HEAD(&work);
1449         dm_deferred_entry_dec(pb->all_io_entry, &work);
1450
1451         if (!list_empty(&work))
1452                 queue_quiesced_migrations(cache, &work);
1453 }
1454
1455 static void quiesce_migration(struct dm_cache_migration *mg)
1456 {
1457         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1458                 queue_quiesced_migration(mg);
1459 }
1460
1461 static void promote(struct cache *cache, struct prealloc *structs,
1462                     dm_oblock_t oblock, dm_cblock_t cblock,
1463                     struct dm_bio_prison_cell *cell)
1464 {
1465         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1466
1467         mg->err = false;
1468         mg->discard = false;
1469         mg->writeback = false;
1470         mg->demote = false;
1471         mg->promote = true;
1472         mg->requeue_holder = true;
1473         mg->invalidate = false;
1474         mg->cache = cache;
1475         mg->new_oblock = oblock;
1476         mg->cblock = cblock;
1477         mg->old_ocell = NULL;
1478         mg->new_ocell = cell;
1479         mg->start_jiffies = jiffies;
1480
1481         inc_io_migrations(cache);
1482         quiesce_migration(mg);
1483 }
1484
1485 static void writeback(struct cache *cache, struct prealloc *structs,
1486                       dm_oblock_t oblock, dm_cblock_t cblock,
1487                       struct dm_bio_prison_cell *cell)
1488 {
1489         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1490
1491         mg->err = false;
1492         mg->discard = false;
1493         mg->writeback = true;
1494         mg->demote = false;
1495         mg->promote = false;
1496         mg->requeue_holder = true;
1497         mg->invalidate = false;
1498         mg->cache = cache;
1499         mg->old_oblock = oblock;
1500         mg->cblock = cblock;
1501         mg->old_ocell = cell;
1502         mg->new_ocell = NULL;
1503         mg->start_jiffies = jiffies;
1504
1505         inc_io_migrations(cache);
1506         quiesce_migration(mg);
1507 }
1508
1509 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1510                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1511                                 dm_cblock_t cblock,
1512                                 struct dm_bio_prison_cell *old_ocell,
1513                                 struct dm_bio_prison_cell *new_ocell)
1514 {
1515         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1516
1517         mg->err = false;
1518         mg->discard = false;
1519         mg->writeback = false;
1520         mg->demote = true;
1521         mg->promote = true;
1522         mg->requeue_holder = true;
1523         mg->invalidate = false;
1524         mg->cache = cache;
1525         mg->old_oblock = old_oblock;
1526         mg->new_oblock = new_oblock;
1527         mg->cblock = cblock;
1528         mg->old_ocell = old_ocell;
1529         mg->new_ocell = new_ocell;
1530         mg->start_jiffies = jiffies;
1531
1532         inc_io_migrations(cache);
1533         quiesce_migration(mg);
1534 }
1535
1536 /*
1537  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1538  * block are thrown away.
1539  */
1540 static void invalidate(struct cache *cache, struct prealloc *structs,
1541                        dm_oblock_t oblock, dm_cblock_t cblock,
1542                        struct dm_bio_prison_cell *cell)
1543 {
1544         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1545
1546         mg->err = false;
1547         mg->discard = false;
1548         mg->writeback = false;
1549         mg->demote = true;
1550         mg->promote = false;
1551         mg->requeue_holder = true;
1552         mg->invalidate = true;
1553         mg->cache = cache;
1554         mg->old_oblock = oblock;
1555         mg->cblock = cblock;
1556         mg->old_ocell = cell;
1557         mg->new_ocell = NULL;
1558         mg->start_jiffies = jiffies;
1559
1560         inc_io_migrations(cache);
1561         quiesce_migration(mg);
1562 }
1563
1564 static void discard(struct cache *cache, struct prealloc *structs,
1565                     struct dm_bio_prison_cell *cell)
1566 {
1567         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1568
1569         mg->err = false;
1570         mg->discard = true;
1571         mg->writeback = false;
1572         mg->demote = false;
1573         mg->promote = false;
1574         mg->requeue_holder = false;
1575         mg->invalidate = false;
1576         mg->cache = cache;
1577         mg->old_ocell = NULL;
1578         mg->new_ocell = cell;
1579         mg->start_jiffies = jiffies;
1580
1581         quiesce_migration(mg);
1582 }
1583
1584 /*----------------------------------------------------------------
1585  * bio processing
1586  *--------------------------------------------------------------*/
1587 static void defer_bio(struct cache *cache, struct bio *bio)
1588 {
1589         unsigned long flags;
1590
1591         spin_lock_irqsave(&cache->lock, flags);
1592         bio_list_add(&cache->deferred_bios, bio);
1593         spin_unlock_irqrestore(&cache->lock, flags);
1594
1595         wake_worker(cache);
1596 }
1597
1598 static void process_flush_bio(struct cache *cache, struct bio *bio)
1599 {
1600         size_t pb_data_size = get_per_bio_data_size(cache);
1601         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1602
1603         BUG_ON(bio->bi_iter.bi_size);
1604         if (!pb->req_nr)
1605                 remap_to_origin(cache, bio);
1606         else
1607                 remap_to_cache(cache, bio, 0);
1608
1609         /*
1610          * REQ_FLUSH is not directed at any particular block so we don't
1611          * need to inc_ds().  REQ_FUA's are split into a write + REQ_FLUSH
1612          * by dm-core.
1613          */
1614         issue(cache, bio);
1615 }
1616
1617 static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1618                                 struct bio *bio)
1619 {
1620         int r;
1621         dm_dblock_t b, e;
1622         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1623
1624         calc_discard_block_range(cache, bio, &b, &e);
1625         if (b == e) {
1626                 bio_endio(bio);
1627                 return;
1628         }
1629
1630         cell_prealloc = prealloc_get_cell(structs);
1631         r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1632                              (cell_free_fn) prealloc_put_cell,
1633                              structs, &new_ocell);
1634         if (r > 0)
1635                 return;
1636
1637         discard(cache, structs, new_ocell);
1638 }
1639
1640 static bool spare_migration_bandwidth(struct cache *cache)
1641 {
1642         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1643                 cache->sectors_per_block;
1644         return current_volume < cache->migration_threshold;
1645 }
1646
1647 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1648 {
1649         atomic_inc(bio_data_dir(bio) == READ ?
1650                    &cache->stats.read_hit : &cache->stats.write_hit);
1651 }
1652
1653 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1654 {
1655         atomic_inc(bio_data_dir(bio) == READ ?
1656                    &cache->stats.read_miss : &cache->stats.write_miss);
1657 }
1658
1659 /*----------------------------------------------------------------*/
1660
1661 struct inc_detail {
1662         struct cache *cache;
1663         struct bio_list bios_for_issue;
1664         struct bio_list unhandled_bios;
1665         bool any_writes;
1666 };
1667
1668 static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1669 {
1670         struct bio *bio;
1671         struct inc_detail *detail = context;
1672         struct cache *cache = detail->cache;
1673
1674         inc_ds(cache, cell->holder, cell);
1675         if (bio_data_dir(cell->holder) == WRITE)
1676                 detail->any_writes = true;
1677
1678         while ((bio = bio_list_pop(&cell->bios))) {
1679                 if (discard_or_flush(bio)) {
1680                         bio_list_add(&detail->unhandled_bios, bio);
1681                         continue;
1682                 }
1683
1684                 if (bio_data_dir(bio) == WRITE)
1685                         detail->any_writes = true;
1686
1687                 bio_list_add(&detail->bios_for_issue, bio);
1688                 inc_ds(cache, bio, cell);
1689         }
1690 }
1691
1692 // FIXME: refactor these two
1693 static void remap_cell_to_origin_clear_discard(struct cache *cache,
1694                                                struct dm_bio_prison_cell *cell,
1695                                                dm_oblock_t oblock, bool issue_holder)
1696 {
1697         struct bio *bio;
1698         unsigned long flags;
1699         struct inc_detail detail;
1700
1701         detail.cache = cache;
1702         bio_list_init(&detail.bios_for_issue);
1703         bio_list_init(&detail.unhandled_bios);
1704         detail.any_writes = false;
1705
1706         spin_lock_irqsave(&cache->lock, flags);
1707         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1708         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1709         spin_unlock_irqrestore(&cache->lock, flags);
1710
1711         remap_to_origin(cache, cell->holder);
1712         if (issue_holder)
1713                 issue(cache, cell->holder);
1714         else
1715                 accounted_begin(cache, cell->holder);
1716
1717         if (detail.any_writes)
1718                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1719
1720         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1721                 remap_to_origin(cache, bio);
1722                 issue(cache, bio);
1723         }
1724
1725         free_prison_cell(cache, cell);
1726 }
1727
1728 static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1729                                       dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1730 {
1731         struct bio *bio;
1732         unsigned long flags;
1733         struct inc_detail detail;
1734
1735         detail.cache = cache;
1736         bio_list_init(&detail.bios_for_issue);
1737         bio_list_init(&detail.unhandled_bios);
1738         detail.any_writes = false;
1739
1740         spin_lock_irqsave(&cache->lock, flags);
1741         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1742         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1743         spin_unlock_irqrestore(&cache->lock, flags);
1744
1745         remap_to_cache(cache, cell->holder, cblock);
1746         if (issue_holder)
1747                 issue(cache, cell->holder);
1748         else
1749                 accounted_begin(cache, cell->holder);
1750
1751         if (detail.any_writes) {
1752                 set_dirty(cache, oblock, cblock);
1753                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1754         }
1755
1756         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1757                 remap_to_cache(cache, bio, cblock);
1758                 issue(cache, bio);
1759         }
1760
1761         free_prison_cell(cache, cell);
1762 }
1763
1764 /*----------------------------------------------------------------*/
1765
1766 struct old_oblock_lock {
1767         struct policy_locker locker;
1768         struct cache *cache;
1769         struct prealloc *structs;
1770         struct dm_bio_prison_cell *cell;
1771 };
1772
1773 static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1774 {
1775         /* This should never be called */
1776         BUG();
1777         return 0;
1778 }
1779
1780 static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1781 {
1782         struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1783         struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1784
1785         return bio_detain(l->cache, b, NULL, cell_prealloc,
1786                           (cell_free_fn) prealloc_put_cell,
1787                           l->structs, &l->cell);
1788 }
1789
1790 static void process_cell(struct cache *cache, struct prealloc *structs,
1791                          struct dm_bio_prison_cell *new_ocell)
1792 {
1793         int r;
1794         bool release_cell = true;
1795         struct bio *bio = new_ocell->holder;
1796         dm_oblock_t block = get_bio_block(cache, bio);
1797         struct policy_result lookup_result;
1798         bool passthrough = passthrough_mode(&cache->features);
1799         bool fast_promotion, can_migrate;
1800         struct old_oblock_lock ool;
1801
1802         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1803         can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1804
1805         ool.locker.fn = cell_locker;
1806         ool.cache = cache;
1807         ool.structs = structs;
1808         ool.cell = NULL;
1809         r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1810                        bio, &ool.locker, &lookup_result);
1811
1812         if (r == -EWOULDBLOCK)
1813                 /* migration has been denied */
1814                 lookup_result.op = POLICY_MISS;
1815
1816         switch (lookup_result.op) {
1817         case POLICY_HIT:
1818                 if (passthrough) {
1819                         inc_miss_counter(cache, bio);
1820
1821                         /*
1822                          * Passthrough always maps to the origin,
1823                          * invalidating any cache blocks that are written
1824                          * to.
1825                          */
1826
1827                         if (bio_data_dir(bio) == WRITE) {
1828                                 atomic_inc(&cache->stats.demotion);
1829                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1830                                 release_cell = false;
1831
1832                         } else {
1833                                 /* FIXME: factor out issue_origin() */
1834                                 remap_to_origin_clear_discard(cache, bio, block);
1835                                 inc_and_issue(cache, bio, new_ocell);
1836                         }
1837                 } else {
1838                         inc_hit_counter(cache, bio);
1839
1840                         if (bio_data_dir(bio) == WRITE &&
1841                             writethrough_mode(&cache->features) &&
1842                             !is_dirty(cache, lookup_result.cblock)) {
1843                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1844                                 inc_and_issue(cache, bio, new_ocell);
1845
1846                         } else {
1847                                 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1848                                 release_cell = false;
1849                         }
1850                 }
1851
1852                 break;
1853
1854         case POLICY_MISS:
1855                 inc_miss_counter(cache, bio);
1856                 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1857                 release_cell = false;
1858                 break;
1859
1860         case POLICY_NEW:
1861                 atomic_inc(&cache->stats.promotion);
1862                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1863                 release_cell = false;
1864                 break;
1865
1866         case POLICY_REPLACE:
1867                 atomic_inc(&cache->stats.demotion);
1868                 atomic_inc(&cache->stats.promotion);
1869                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1870                                     block, lookup_result.cblock,
1871                                     ool.cell, new_ocell);
1872                 release_cell = false;
1873                 break;
1874
1875         default:
1876                 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1877                             cache_device_name(cache), __func__,
1878                             (unsigned) lookup_result.op);
1879                 bio_io_error(bio);
1880         }
1881
1882         if (release_cell)
1883                 cell_defer(cache, new_ocell, false);
1884 }
1885
1886 static void process_bio(struct cache *cache, struct prealloc *structs,
1887                         struct bio *bio)
1888 {
1889         int r;
1890         dm_oblock_t block = get_bio_block(cache, bio);
1891         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1892
1893         /*
1894          * Check to see if that block is currently migrating.
1895          */
1896         cell_prealloc = prealloc_get_cell(structs);
1897         r = bio_detain(cache, block, bio, cell_prealloc,
1898                        (cell_free_fn) prealloc_put_cell,
1899                        structs, &new_ocell);
1900         if (r > 0)
1901                 return;
1902
1903         process_cell(cache, structs, new_ocell);
1904 }
1905
1906 static int need_commit_due_to_time(struct cache *cache)
1907 {
1908         return jiffies < cache->last_commit_jiffies ||
1909                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1910 }
1911
1912 /*
1913  * A non-zero return indicates read_only or fail_io mode.
1914  */
1915 static int commit(struct cache *cache, bool clean_shutdown)
1916 {
1917         int r;
1918
1919         if (get_cache_mode(cache) >= CM_READ_ONLY)
1920                 return -EINVAL;
1921
1922         atomic_inc(&cache->stats.commit_count);
1923         r = dm_cache_commit(cache->cmd, clean_shutdown);
1924         if (r)
1925                 metadata_operation_failed(cache, "dm_cache_commit", r);
1926
1927         return r;
1928 }
1929
1930 static int commit_if_needed(struct cache *cache)
1931 {
1932         int r = 0;
1933
1934         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1935             dm_cache_changed_this_transaction(cache->cmd)) {
1936                 r = commit(cache, false);
1937                 cache->commit_requested = false;
1938                 cache->last_commit_jiffies = jiffies;
1939         }
1940
1941         return r;
1942 }
1943
1944 static void process_deferred_bios(struct cache *cache)
1945 {
1946         bool prealloc_used = false;
1947         unsigned long flags;
1948         struct bio_list bios;
1949         struct bio *bio;
1950         struct prealloc structs;
1951
1952         memset(&structs, 0, sizeof(structs));
1953         bio_list_init(&bios);
1954
1955         spin_lock_irqsave(&cache->lock, flags);
1956         bio_list_merge(&bios, &cache->deferred_bios);
1957         bio_list_init(&cache->deferred_bios);
1958         spin_unlock_irqrestore(&cache->lock, flags);
1959
1960         while (!bio_list_empty(&bios)) {
1961                 /*
1962                  * If we've got no free migration structs, and processing
1963                  * this bio might require one, we pause until there are some
1964                  * prepared mappings to process.
1965                  */
1966                 prealloc_used = true;
1967                 if (prealloc_data_structs(cache, &structs)) {
1968                         spin_lock_irqsave(&cache->lock, flags);
1969                         bio_list_merge(&cache->deferred_bios, &bios);
1970                         spin_unlock_irqrestore(&cache->lock, flags);
1971                         break;
1972                 }
1973
1974                 bio = bio_list_pop(&bios);
1975
1976                 if (bio->bi_rw & REQ_FLUSH)
1977                         process_flush_bio(cache, bio);
1978                 else if (bio->bi_rw & REQ_DISCARD)
1979                         process_discard_bio(cache, &structs, bio);
1980                 else
1981                         process_bio(cache, &structs, bio);
1982         }
1983
1984         if (prealloc_used)
1985                 prealloc_free_structs(cache, &structs);
1986 }
1987
1988 static void process_deferred_cells(struct cache *cache)
1989 {
1990         bool prealloc_used = false;
1991         unsigned long flags;
1992         struct dm_bio_prison_cell *cell, *tmp;
1993         struct list_head cells;
1994         struct prealloc structs;
1995
1996         memset(&structs, 0, sizeof(structs));
1997
1998         INIT_LIST_HEAD(&cells);
1999
2000         spin_lock_irqsave(&cache->lock, flags);
2001         list_splice_init(&cache->deferred_cells, &cells);
2002         spin_unlock_irqrestore(&cache->lock, flags);
2003
2004         list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2005                 /*
2006                  * If we've got no free migration structs, and processing
2007                  * this bio might require one, we pause until there are some
2008                  * prepared mappings to process.
2009                  */
2010                 prealloc_used = true;
2011                 if (prealloc_data_structs(cache, &structs)) {
2012                         spin_lock_irqsave(&cache->lock, flags);
2013                         list_splice(&cells, &cache->deferred_cells);
2014                         spin_unlock_irqrestore(&cache->lock, flags);
2015                         break;
2016                 }
2017
2018                 process_cell(cache, &structs, cell);
2019         }
2020
2021         if (prealloc_used)
2022                 prealloc_free_structs(cache, &structs);
2023 }
2024
2025 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2026 {
2027         unsigned long flags;
2028         struct bio_list bios;
2029         struct bio *bio;
2030
2031         bio_list_init(&bios);
2032
2033         spin_lock_irqsave(&cache->lock, flags);
2034         bio_list_merge(&bios, &cache->deferred_flush_bios);
2035         bio_list_init(&cache->deferred_flush_bios);
2036         spin_unlock_irqrestore(&cache->lock, flags);
2037
2038         /*
2039          * These bios have already been through inc_ds()
2040          */
2041         while ((bio = bio_list_pop(&bios)))
2042                 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2043 }
2044
2045 static void process_deferred_writethrough_bios(struct cache *cache)
2046 {
2047         unsigned long flags;
2048         struct bio_list bios;
2049         struct bio *bio;
2050
2051         bio_list_init(&bios);
2052
2053         spin_lock_irqsave(&cache->lock, flags);
2054         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2055         bio_list_init(&cache->deferred_writethrough_bios);
2056         spin_unlock_irqrestore(&cache->lock, flags);
2057
2058         /*
2059          * These bios have already been through inc_ds()
2060          */
2061         while ((bio = bio_list_pop(&bios)))
2062                 accounted_request(cache, bio);
2063 }
2064
2065 static void writeback_some_dirty_blocks(struct cache *cache)
2066 {
2067         bool prealloc_used = false;
2068         dm_oblock_t oblock;
2069         dm_cblock_t cblock;
2070         struct prealloc structs;
2071         struct dm_bio_prison_cell *old_ocell;
2072         bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2073
2074         memset(&structs, 0, sizeof(structs));
2075
2076         while (spare_migration_bandwidth(cache)) {
2077                 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2078                         break; /* no work to do */
2079
2080                 prealloc_used = true;
2081                 if (prealloc_data_structs(cache, &structs) ||
2082                     get_cell(cache, oblock, &structs, &old_ocell)) {
2083                         policy_set_dirty(cache->policy, oblock);
2084                         break;
2085                 }
2086
2087                 writeback(cache, &structs, oblock, cblock, old_ocell);
2088         }
2089
2090         if (prealloc_used)
2091                 prealloc_free_structs(cache, &structs);
2092 }
2093
2094 /*----------------------------------------------------------------
2095  * Invalidations.
2096  * Dropping something from the cache *without* writing back.
2097  *--------------------------------------------------------------*/
2098
2099 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2100 {
2101         int r = 0;
2102         uint64_t begin = from_cblock(req->cblocks->begin);
2103         uint64_t end = from_cblock(req->cblocks->end);
2104
2105         while (begin != end) {
2106                 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2107                 if (!r) {
2108                         r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2109                         if (r) {
2110                                 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2111                                 break;
2112                         }
2113
2114                 } else if (r == -ENODATA) {
2115                         /* harmless, already unmapped */
2116                         r = 0;
2117
2118                 } else {
2119                         DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2120                         break;
2121                 }
2122
2123                 begin++;
2124         }
2125
2126         cache->commit_requested = true;
2127
2128         req->err = r;
2129         atomic_set(&req->complete, 1);
2130
2131         wake_up(&req->result_wait);
2132 }
2133
2134 static void process_invalidation_requests(struct cache *cache)
2135 {
2136         struct list_head list;
2137         struct invalidation_request *req, *tmp;
2138
2139         INIT_LIST_HEAD(&list);
2140         spin_lock(&cache->invalidation_lock);
2141         list_splice_init(&cache->invalidation_requests, &list);
2142         spin_unlock(&cache->invalidation_lock);
2143
2144         list_for_each_entry_safe (req, tmp, &list, list)
2145                 process_invalidation_request(cache, req);
2146 }
2147
2148 /*----------------------------------------------------------------
2149  * Main worker loop
2150  *--------------------------------------------------------------*/
2151 static bool is_quiescing(struct cache *cache)
2152 {
2153         return atomic_read(&cache->quiescing);
2154 }
2155
2156 static void ack_quiescing(struct cache *cache)
2157 {
2158         if (is_quiescing(cache)) {
2159                 atomic_inc(&cache->quiescing_ack);
2160                 wake_up(&cache->quiescing_wait);
2161         }
2162 }
2163
2164 static void wait_for_quiescing_ack(struct cache *cache)
2165 {
2166         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2167 }
2168
2169 static void start_quiescing(struct cache *cache)
2170 {
2171         atomic_inc(&cache->quiescing);
2172         wait_for_quiescing_ack(cache);
2173 }
2174
2175 static void stop_quiescing(struct cache *cache)
2176 {
2177         atomic_set(&cache->quiescing, 0);
2178         atomic_set(&cache->quiescing_ack, 0);
2179 }
2180
2181 static void wait_for_migrations(struct cache *cache)
2182 {
2183         wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2184 }
2185
2186 static void stop_worker(struct cache *cache)
2187 {
2188         cancel_delayed_work(&cache->waker);
2189         flush_workqueue(cache->wq);
2190 }
2191
2192 static void requeue_deferred_cells(struct cache *cache)
2193 {
2194         unsigned long flags;
2195         struct list_head cells;
2196         struct dm_bio_prison_cell *cell, *tmp;
2197
2198         INIT_LIST_HEAD(&cells);
2199         spin_lock_irqsave(&cache->lock, flags);
2200         list_splice_init(&cache->deferred_cells, &cells);
2201         spin_unlock_irqrestore(&cache->lock, flags);
2202
2203         list_for_each_entry_safe(cell, tmp, &cells, user_list)
2204                 cell_requeue(cache, cell);
2205 }
2206
2207 static void requeue_deferred_bios(struct cache *cache)
2208 {
2209         struct bio *bio;
2210         struct bio_list bios;
2211
2212         bio_list_init(&bios);
2213         bio_list_merge(&bios, &cache->deferred_bios);
2214         bio_list_init(&cache->deferred_bios);
2215
2216         while ((bio = bio_list_pop(&bios))) {
2217                 bio->bi_error = DM_ENDIO_REQUEUE;
2218                 bio_endio(bio);
2219         }
2220 }
2221
2222 static int more_work(struct cache *cache)
2223 {
2224         if (is_quiescing(cache))
2225                 return !list_empty(&cache->quiesced_migrations) ||
2226                         !list_empty(&cache->completed_migrations) ||
2227                         !list_empty(&cache->need_commit_migrations);
2228         else
2229                 return !bio_list_empty(&cache->deferred_bios) ||
2230                         !list_empty(&cache->deferred_cells) ||
2231                         !bio_list_empty(&cache->deferred_flush_bios) ||
2232                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
2233                         !list_empty(&cache->quiesced_migrations) ||
2234                         !list_empty(&cache->completed_migrations) ||
2235                         !list_empty(&cache->need_commit_migrations) ||
2236                         cache->invalidate;
2237 }
2238
2239 static void do_worker(struct work_struct *ws)
2240 {
2241         struct cache *cache = container_of(ws, struct cache, worker);
2242
2243         do {
2244                 if (!is_quiescing(cache)) {
2245                         writeback_some_dirty_blocks(cache);
2246                         process_deferred_writethrough_bios(cache);
2247                         process_deferred_bios(cache);
2248                         process_deferred_cells(cache);
2249                         process_invalidation_requests(cache);
2250                 }
2251
2252                 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2253                 process_migrations(cache, &cache->completed_migrations, complete_migration);
2254
2255                 if (commit_if_needed(cache)) {
2256                         process_deferred_flush_bios(cache, false);
2257                         process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2258                 } else {
2259                         process_deferred_flush_bios(cache, true);
2260                         process_migrations(cache, &cache->need_commit_migrations,
2261                                            migration_success_post_commit);
2262                 }
2263
2264                 ack_quiescing(cache);
2265
2266         } while (more_work(cache));
2267 }
2268
2269 /*
2270  * We want to commit periodically so that not too much
2271  * unwritten metadata builds up.
2272  */
2273 static void do_waker(struct work_struct *ws)
2274 {
2275         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2276         policy_tick(cache->policy, true);
2277         wake_worker(cache);
2278         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2279 }
2280
2281 /*----------------------------------------------------------------*/
2282
2283 static int is_congested(struct dm_dev *dev, int bdi_bits)
2284 {
2285         struct request_queue *q = bdev_get_queue(dev->bdev);
2286         return bdi_congested(&q->backing_dev_info, bdi_bits);
2287 }
2288
2289 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2290 {
2291         struct cache *cache = container_of(cb, struct cache, callbacks);
2292
2293         return is_congested(cache->origin_dev, bdi_bits) ||
2294                 is_congested(cache->cache_dev, bdi_bits);
2295 }
2296
2297 /*----------------------------------------------------------------
2298  * Target methods
2299  *--------------------------------------------------------------*/
2300
2301 /*
2302  * This function gets called on the error paths of the constructor, so we
2303  * have to cope with a partially initialised struct.
2304  */
2305 static void destroy(struct cache *cache)
2306 {
2307         unsigned i;
2308
2309         mempool_destroy(cache->migration_pool);
2310
2311         if (cache->all_io_ds)
2312                 dm_deferred_set_destroy(cache->all_io_ds);
2313
2314         if (cache->prison)
2315                 dm_bio_prison_destroy(cache->prison);
2316
2317         if (cache->wq)
2318                 destroy_workqueue(cache->wq);
2319
2320         if (cache->dirty_bitset)
2321                 free_bitset(cache->dirty_bitset);
2322
2323         if (cache->discard_bitset)
2324                 free_bitset(cache->discard_bitset);
2325
2326         if (cache->copier)
2327                 dm_kcopyd_client_destroy(cache->copier);
2328
2329         if (cache->cmd)
2330                 dm_cache_metadata_close(cache->cmd);
2331
2332         if (cache->metadata_dev)
2333                 dm_put_device(cache->ti, cache->metadata_dev);
2334
2335         if (cache->origin_dev)
2336                 dm_put_device(cache->ti, cache->origin_dev);
2337
2338         if (cache->cache_dev)
2339                 dm_put_device(cache->ti, cache->cache_dev);
2340
2341         if (cache->policy)
2342                 dm_cache_policy_destroy(cache->policy);
2343
2344         for (i = 0; i < cache->nr_ctr_args ; i++)
2345                 kfree(cache->ctr_args[i]);
2346         kfree(cache->ctr_args);
2347
2348         kfree(cache);
2349 }
2350
2351 static void cache_dtr(struct dm_target *ti)
2352 {
2353         struct cache *cache = ti->private;
2354
2355         destroy(cache);
2356 }
2357
2358 static sector_t get_dev_size(struct dm_dev *dev)
2359 {
2360         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2361 }
2362
2363 /*----------------------------------------------------------------*/
2364
2365 /*
2366  * Construct a cache device mapping.
2367  *
2368  * cache <metadata dev> <cache dev> <origin dev> <block size>
2369  *       <#feature args> [<feature arg>]*
2370  *       <policy> <#policy args> [<policy arg>]*
2371  *
2372  * metadata dev    : fast device holding the persistent metadata
2373  * cache dev       : fast device holding cached data blocks
2374  * origin dev      : slow device holding original data blocks
2375  * block size      : cache unit size in sectors
2376  *
2377  * #feature args   : number of feature arguments passed
2378  * feature args    : writethrough.  (The default is writeback.)
2379  *
2380  * policy          : the replacement policy to use
2381  * #policy args    : an even number of policy arguments corresponding
2382  *                   to key/value pairs passed to the policy
2383  * policy args     : key/value pairs passed to the policy
2384  *                   E.g. 'sequential_threshold 1024'
2385  *                   See cache-policies.txt for details.
2386  *
2387  * Optional feature arguments are:
2388  *   writethrough  : write through caching that prohibits cache block
2389  *                   content from being different from origin block content.
2390  *                   Without this argument, the default behaviour is to write
2391  *                   back cache block contents later for performance reasons,
2392  *                   so they may differ from the corresponding origin blocks.
2393  */
2394 struct cache_args {
2395         struct dm_target *ti;
2396
2397         struct dm_dev *metadata_dev;
2398
2399         struct dm_dev *cache_dev;
2400         sector_t cache_sectors;
2401
2402         struct dm_dev *origin_dev;
2403         sector_t origin_sectors;
2404
2405         uint32_t block_size;
2406
2407         const char *policy_name;
2408         int policy_argc;
2409         const char **policy_argv;
2410
2411         struct cache_features features;
2412 };
2413
2414 static void destroy_cache_args(struct cache_args *ca)
2415 {
2416         if (ca->metadata_dev)
2417                 dm_put_device(ca->ti, ca->metadata_dev);
2418
2419         if (ca->cache_dev)
2420                 dm_put_device(ca->ti, ca->cache_dev);
2421
2422         if (ca->origin_dev)
2423                 dm_put_device(ca->ti, ca->origin_dev);
2424
2425         kfree(ca);
2426 }
2427
2428 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2429 {
2430         if (!as->argc) {
2431                 *error = "Insufficient args";
2432                 return false;
2433         }
2434
2435         return true;
2436 }
2437
2438 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2439                               char **error)
2440 {
2441         int r;
2442         sector_t metadata_dev_size;
2443         char b[BDEVNAME_SIZE];
2444
2445         if (!at_least_one_arg(as, error))
2446                 return -EINVAL;
2447
2448         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2449                           &ca->metadata_dev);
2450         if (r) {
2451                 *error = "Error opening metadata device";
2452                 return r;
2453         }
2454
2455         metadata_dev_size = get_dev_size(ca->metadata_dev);
2456         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2457                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2458                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2459
2460         return 0;
2461 }
2462
2463 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2464                            char **error)
2465 {
2466         int r;
2467
2468         if (!at_least_one_arg(as, error))
2469                 return -EINVAL;
2470
2471         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2472                           &ca->cache_dev);
2473         if (r) {
2474                 *error = "Error opening cache device";
2475                 return r;
2476         }
2477         ca->cache_sectors = get_dev_size(ca->cache_dev);
2478
2479         return 0;
2480 }
2481
2482 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2483                             char **error)
2484 {
2485         int r;
2486
2487         if (!at_least_one_arg(as, error))
2488                 return -EINVAL;
2489
2490         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2491                           &ca->origin_dev);
2492         if (r) {
2493                 *error = "Error opening origin device";
2494                 return r;
2495         }
2496
2497         ca->origin_sectors = get_dev_size(ca->origin_dev);
2498         if (ca->ti->len > ca->origin_sectors) {
2499                 *error = "Device size larger than cached device";
2500                 return -EINVAL;
2501         }
2502
2503         return 0;
2504 }
2505
2506 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2507                             char **error)
2508 {
2509         unsigned long block_size;
2510
2511         if (!at_least_one_arg(as, error))
2512                 return -EINVAL;
2513
2514         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2515             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2516             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2517             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2518                 *error = "Invalid data block size";
2519                 return -EINVAL;
2520         }
2521
2522         if (block_size > ca->cache_sectors) {
2523                 *error = "Data block size is larger than the cache device";
2524                 return -EINVAL;
2525         }
2526
2527         ca->block_size = block_size;
2528
2529         return 0;
2530 }
2531
2532 static void init_features(struct cache_features *cf)
2533 {
2534         cf->mode = CM_WRITE;
2535         cf->io_mode = CM_IO_WRITEBACK;
2536 }
2537
2538 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2539                           char **error)
2540 {
2541         static struct dm_arg _args[] = {
2542                 {0, 1, "Invalid number of cache feature arguments"},
2543         };
2544
2545         int r;
2546         unsigned argc;
2547         const char *arg;
2548         struct cache_features *cf = &ca->features;
2549
2550         init_features(cf);
2551
2552         r = dm_read_arg_group(_args, as, &argc, error);
2553         if (r)
2554                 return -EINVAL;
2555
2556         while (argc--) {
2557                 arg = dm_shift_arg(as);
2558
2559                 if (!strcasecmp(arg, "writeback"))
2560                         cf->io_mode = CM_IO_WRITEBACK;
2561
2562                 else if (!strcasecmp(arg, "writethrough"))
2563                         cf->io_mode = CM_IO_WRITETHROUGH;
2564
2565                 else if (!strcasecmp(arg, "passthrough"))
2566                         cf->io_mode = CM_IO_PASSTHROUGH;
2567
2568                 else {
2569                         *error = "Unrecognised cache feature requested";
2570                         return -EINVAL;
2571                 }
2572         }
2573
2574         return 0;
2575 }
2576
2577 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2578                         char **error)
2579 {
2580         static struct dm_arg _args[] = {
2581                 {0, 1024, "Invalid number of policy arguments"},
2582         };
2583
2584         int r;
2585
2586         if (!at_least_one_arg(as, error))
2587                 return -EINVAL;
2588
2589         ca->policy_name = dm_shift_arg(as);
2590
2591         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2592         if (r)
2593                 return -EINVAL;
2594
2595         ca->policy_argv = (const char **)as->argv;
2596         dm_consume_args(as, ca->policy_argc);
2597
2598         return 0;
2599 }
2600
2601 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2602                             char **error)
2603 {
2604         int r;
2605         struct dm_arg_set as;
2606
2607         as.argc = argc;
2608         as.argv = argv;
2609
2610         r = parse_metadata_dev(ca, &as, error);
2611         if (r)
2612                 return r;
2613
2614         r = parse_cache_dev(ca, &as, error);
2615         if (r)
2616                 return r;
2617
2618         r = parse_origin_dev(ca, &as, error);
2619         if (r)
2620                 return r;
2621
2622         r = parse_block_size(ca, &as, error);
2623         if (r)
2624                 return r;
2625
2626         r = parse_features(ca, &as, error);
2627         if (r)
2628                 return r;
2629
2630         r = parse_policy(ca, &as, error);
2631         if (r)
2632                 return r;
2633
2634         return 0;
2635 }
2636
2637 /*----------------------------------------------------------------*/
2638
2639 static struct kmem_cache *migration_cache;
2640
2641 #define NOT_CORE_OPTION 1
2642
2643 static int process_config_option(struct cache *cache, const char *key, const char *value)
2644 {
2645         unsigned long tmp;
2646
2647         if (!strcasecmp(key, "migration_threshold")) {
2648                 if (kstrtoul(value, 10, &tmp))
2649                         return -EINVAL;
2650
2651                 cache->migration_threshold = tmp;
2652                 return 0;
2653         }
2654
2655         return NOT_CORE_OPTION;
2656 }
2657
2658 static int set_config_value(struct cache *cache, const char *key, const char *value)
2659 {
2660         int r = process_config_option(cache, key, value);
2661
2662         if (r == NOT_CORE_OPTION)
2663                 r = policy_set_config_value(cache->policy, key, value);
2664
2665         if (r)
2666                 DMWARN("bad config value for %s: %s", key, value);
2667
2668         return r;
2669 }
2670
2671 static int set_config_values(struct cache *cache, int argc, const char **argv)
2672 {
2673         int r = 0;
2674
2675         if (argc & 1) {
2676                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2677                 return -EINVAL;
2678         }
2679
2680         while (argc) {
2681                 r = set_config_value(cache, argv[0], argv[1]);
2682                 if (r)
2683                         break;
2684
2685                 argc -= 2;
2686                 argv += 2;
2687         }
2688
2689         return r;
2690 }
2691
2692 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2693                                char **error)
2694 {
2695         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2696                                                            cache->cache_size,
2697                                                            cache->origin_sectors,
2698                                                            cache->sectors_per_block);
2699         if (IS_ERR(p)) {
2700                 *error = "Error creating cache's policy";
2701                 return PTR_ERR(p);
2702         }
2703         cache->policy = p;
2704
2705         return 0;
2706 }
2707
2708 /*
2709  * We want the discard block size to be at least the size of the cache
2710  * block size and have no more than 2^14 discard blocks across the origin.
2711  */
2712 #define MAX_DISCARD_BLOCKS (1 << 14)
2713
2714 static bool too_many_discard_blocks(sector_t discard_block_size,
2715                                     sector_t origin_size)
2716 {
2717         (void) sector_div(origin_size, discard_block_size);
2718
2719         return origin_size > MAX_DISCARD_BLOCKS;
2720 }
2721
2722 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2723                                              sector_t origin_size)
2724 {
2725         sector_t discard_block_size = cache_block_size;
2726
2727         if (origin_size)
2728                 while (too_many_discard_blocks(discard_block_size, origin_size))
2729                         discard_block_size *= 2;
2730
2731         return discard_block_size;
2732 }
2733
2734 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2735 {
2736         dm_block_t nr_blocks = from_cblock(size);
2737
2738         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2739                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2740                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2741                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2742                              (unsigned long long) nr_blocks);
2743
2744         cache->cache_size = size;
2745 }
2746
2747 #define DEFAULT_MIGRATION_THRESHOLD 2048
2748
2749 static int cache_create(struct cache_args *ca, struct cache **result)
2750 {
2751         int r = 0;
2752         char **error = &ca->ti->error;
2753         struct cache *cache;
2754         struct dm_target *ti = ca->ti;
2755         dm_block_t origin_blocks;
2756         struct dm_cache_metadata *cmd;
2757         bool may_format = ca->features.mode == CM_WRITE;
2758
2759         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2760         if (!cache)
2761                 return -ENOMEM;
2762
2763         cache->ti = ca->ti;
2764         ti->private = cache;
2765         ti->num_flush_bios = 2;
2766         ti->flush_supported = true;
2767
2768         ti->num_discard_bios = 1;
2769         ti->discards_supported = true;
2770         ti->discard_zeroes_data_unsupported = true;
2771         ti->split_discard_bios = false;
2772
2773         cache->features = ca->features;
2774         ti->per_io_data_size = get_per_bio_data_size(cache);
2775
2776         cache->callbacks.congested_fn = cache_is_congested;
2777         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2778
2779         cache->metadata_dev = ca->metadata_dev;
2780         cache->origin_dev = ca->origin_dev;
2781         cache->cache_dev = ca->cache_dev;
2782
2783         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2784
2785         /* FIXME: factor out this whole section */
2786         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2787         origin_blocks = block_div(origin_blocks, ca->block_size);
2788         cache->origin_blocks = to_oblock(origin_blocks);
2789
2790         cache->sectors_per_block = ca->block_size;
2791         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2792                 r = -EINVAL;
2793                 goto bad;
2794         }
2795
2796         if (ca->block_size & (ca->block_size - 1)) {
2797                 dm_block_t cache_size = ca->cache_sectors;
2798
2799                 cache->sectors_per_block_shift = -1;
2800                 cache_size = block_div(cache_size, ca->block_size);
2801                 set_cache_size(cache, to_cblock(cache_size));
2802         } else {
2803                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2804                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2805         }
2806
2807         r = create_cache_policy(cache, ca, error);
2808         if (r)
2809                 goto bad;
2810
2811         cache->policy_nr_args = ca->policy_argc;
2812         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2813
2814         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2815         if (r) {
2816                 *error = "Error setting cache policy's config values";
2817                 goto bad;
2818         }
2819
2820         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2821                                      ca->block_size, may_format,
2822                                      dm_cache_policy_get_hint_size(cache->policy));
2823         if (IS_ERR(cmd)) {
2824                 *error = "Error creating metadata object";
2825                 r = PTR_ERR(cmd);
2826                 goto bad;
2827         }
2828         cache->cmd = cmd;
2829         set_cache_mode(cache, CM_WRITE);
2830         if (get_cache_mode(cache) != CM_WRITE) {
2831                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2832                 r = -EINVAL;
2833                 goto bad;
2834         }
2835
2836         if (passthrough_mode(&cache->features)) {
2837                 bool all_clean;
2838
2839                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2840                 if (r) {
2841                         *error = "dm_cache_metadata_all_clean() failed";
2842                         goto bad;
2843                 }
2844
2845                 if (!all_clean) {
2846                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2847                         r = -EINVAL;
2848                         goto bad;
2849                 }
2850         }
2851
2852         spin_lock_init(&cache->lock);
2853         INIT_LIST_HEAD(&cache->deferred_cells);
2854         bio_list_init(&cache->deferred_bios);
2855         bio_list_init(&cache->deferred_flush_bios);
2856         bio_list_init(&cache->deferred_writethrough_bios);
2857         INIT_LIST_HEAD(&cache->quiesced_migrations);
2858         INIT_LIST_HEAD(&cache->completed_migrations);
2859         INIT_LIST_HEAD(&cache->need_commit_migrations);
2860         atomic_set(&cache->nr_allocated_migrations, 0);
2861         atomic_set(&cache->nr_io_migrations, 0);
2862         init_waitqueue_head(&cache->migration_wait);
2863
2864         init_waitqueue_head(&cache->quiescing_wait);
2865         atomic_set(&cache->quiescing, 0);
2866         atomic_set(&cache->quiescing_ack, 0);
2867
2868         r = -ENOMEM;
2869         atomic_set(&cache->nr_dirty, 0);
2870         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2871         if (!cache->dirty_bitset) {
2872                 *error = "could not allocate dirty bitset";
2873                 goto bad;
2874         }
2875         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2876
2877         cache->discard_block_size =
2878                 calculate_discard_block_size(cache->sectors_per_block,
2879                                              cache->origin_sectors);
2880         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2881                                                               cache->discard_block_size));
2882         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2883         if (!cache->discard_bitset) {
2884                 *error = "could not allocate discard bitset";
2885                 goto bad;
2886         }
2887         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2888
2889         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2890         if (IS_ERR(cache->copier)) {
2891                 *error = "could not create kcopyd client";
2892                 r = PTR_ERR(cache->copier);
2893                 goto bad;
2894         }
2895
2896         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2897         if (!cache->wq) {
2898                 *error = "could not create workqueue for metadata object";
2899                 goto bad;
2900         }
2901         INIT_WORK(&cache->worker, do_worker);
2902         INIT_DELAYED_WORK(&cache->waker, do_waker);
2903         cache->last_commit_jiffies = jiffies;
2904
2905         cache->prison = dm_bio_prison_create();
2906         if (!cache->prison) {
2907                 *error = "could not create bio prison";
2908                 goto bad;
2909         }
2910
2911         cache->all_io_ds = dm_deferred_set_create();
2912         if (!cache->all_io_ds) {
2913                 *error = "could not create all_io deferred set";
2914                 goto bad;
2915         }
2916
2917         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2918                                                          migration_cache);
2919         if (!cache->migration_pool) {
2920                 *error = "Error creating cache's migration mempool";
2921                 goto bad;
2922         }
2923
2924         cache->need_tick_bio = true;
2925         cache->sized = false;
2926         cache->invalidate = false;
2927         cache->commit_requested = false;
2928         cache->loaded_mappings = false;
2929         cache->loaded_discards = false;
2930
2931         load_stats(cache);
2932
2933         atomic_set(&cache->stats.demotion, 0);
2934         atomic_set(&cache->stats.promotion, 0);
2935         atomic_set(&cache->stats.copies_avoided, 0);
2936         atomic_set(&cache->stats.cache_cell_clash, 0);
2937         atomic_set(&cache->stats.commit_count, 0);
2938         atomic_set(&cache->stats.discard_count, 0);
2939
2940         spin_lock_init(&cache->invalidation_lock);
2941         INIT_LIST_HEAD(&cache->invalidation_requests);
2942
2943         iot_init(&cache->origin_tracker);
2944
2945         *result = cache;
2946         return 0;
2947
2948 bad:
2949         destroy(cache);
2950         return r;
2951 }
2952
2953 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2954 {
2955         unsigned i;
2956         const char **copy;
2957
2958         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2959         if (!copy)
2960                 return -ENOMEM;
2961         for (i = 0; i < argc; i++) {
2962                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2963                 if (!copy[i]) {
2964                         while (i--)
2965                                 kfree(copy[i]);
2966                         kfree(copy);
2967                         return -ENOMEM;
2968                 }
2969         }
2970
2971         cache->nr_ctr_args = argc;
2972         cache->ctr_args = copy;
2973
2974         return 0;
2975 }
2976
2977 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2978 {
2979         int r = -EINVAL;
2980         struct cache_args *ca;
2981         struct cache *cache = NULL;
2982
2983         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2984         if (!ca) {
2985                 ti->error = "Error allocating memory for cache";
2986                 return -ENOMEM;
2987         }
2988         ca->ti = ti;
2989
2990         r = parse_cache_args(ca, argc, argv, &ti->error);
2991         if (r)
2992                 goto out;
2993
2994         r = cache_create(ca, &cache);
2995         if (r)
2996                 goto out;
2997
2998         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2999         if (r) {
3000                 destroy(cache);
3001                 goto out;
3002         }
3003
3004         ti->private = cache;
3005
3006 out:
3007         destroy_cache_args(ca);
3008         return r;
3009 }
3010
3011 /*----------------------------------------------------------------*/
3012
3013 static int cache_map(struct dm_target *ti, struct bio *bio)
3014 {
3015         struct cache *cache = ti->private;
3016
3017         int r;
3018         struct dm_bio_prison_cell *cell = NULL;
3019         dm_oblock_t block = get_bio_block(cache, bio);
3020         size_t pb_data_size = get_per_bio_data_size(cache);
3021         bool can_migrate = false;
3022         bool fast_promotion;
3023         struct policy_result lookup_result;
3024         struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3025         struct old_oblock_lock ool;
3026
3027         ool.locker.fn = null_locker;
3028
3029         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3030                 /*
3031                  * This can only occur if the io goes to a partial block at
3032                  * the end of the origin device.  We don't cache these.
3033                  * Just remap to the origin and carry on.
3034                  */
3035                 remap_to_origin(cache, bio);
3036                 accounted_begin(cache, bio);
3037                 return DM_MAPIO_REMAPPED;
3038         }
3039
3040         if (discard_or_flush(bio)) {
3041                 defer_bio(cache, bio);
3042                 return DM_MAPIO_SUBMITTED;
3043         }
3044
3045         /*
3046          * Check to see if that block is currently migrating.
3047          */
3048         cell = alloc_prison_cell(cache);
3049         if (!cell) {
3050                 defer_bio(cache, bio);
3051                 return DM_MAPIO_SUBMITTED;
3052         }
3053
3054         r = bio_detain(cache, block, bio, cell,
3055                        (cell_free_fn) free_prison_cell,
3056                        cache, &cell);
3057         if (r) {
3058                 if (r < 0)
3059                         defer_bio(cache, bio);
3060
3061                 return DM_MAPIO_SUBMITTED;
3062         }
3063
3064         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3065
3066         r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3067                        bio, &ool.locker, &lookup_result);
3068         if (r == -EWOULDBLOCK) {
3069                 cell_defer(cache, cell, true);
3070                 return DM_MAPIO_SUBMITTED;
3071
3072         } else if (r) {
3073                 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3074                             cache_device_name(cache), r);
3075                 cell_defer(cache, cell, false);
3076                 bio_io_error(bio);
3077                 return DM_MAPIO_SUBMITTED;
3078         }
3079
3080         r = DM_MAPIO_REMAPPED;
3081         switch (lookup_result.op) {
3082         case POLICY_HIT:
3083                 if (passthrough_mode(&cache->features)) {
3084                         if (bio_data_dir(bio) == WRITE) {
3085                                 /*
3086                                  * We need to invalidate this block, so
3087                                  * defer for the worker thread.
3088                                  */
3089                                 cell_defer(cache, cell, true);
3090                                 r = DM_MAPIO_SUBMITTED;
3091
3092                         } else {
3093                                 inc_miss_counter(cache, bio);
3094                                 remap_to_origin_clear_discard(cache, bio, block);
3095                                 accounted_begin(cache, bio);
3096                                 inc_ds(cache, bio, cell);
3097                                 // FIXME: we want to remap hits or misses straight
3098                                 // away rather than passing over to the worker.
3099                                 cell_defer(cache, cell, false);
3100                         }
3101
3102                 } else {
3103                         inc_hit_counter(cache, bio);
3104                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3105                             !is_dirty(cache, lookup_result.cblock)) {
3106                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3107                                 accounted_begin(cache, bio);
3108                                 inc_ds(cache, bio, cell);
3109                                 cell_defer(cache, cell, false);
3110
3111                         } else
3112                                 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3113                 }
3114                 break;
3115
3116         case POLICY_MISS:
3117                 inc_miss_counter(cache, bio);
3118                 if (pb->req_nr != 0) {
3119                         /*
3120                          * This is a duplicate writethrough io that is no
3121                          * longer needed because the block has been demoted.
3122                          */
3123                         bio_endio(bio);
3124                         // FIXME: remap everything as a miss
3125                         cell_defer(cache, cell, false);
3126                         r = DM_MAPIO_SUBMITTED;
3127
3128                 } else
3129                         remap_cell_to_origin_clear_discard(cache, cell, block, false);
3130                 break;
3131
3132         default:
3133                 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3134                             cache_device_name(cache), __func__,
3135                             (unsigned) lookup_result.op);
3136                 cell_defer(cache, cell, false);
3137                 bio_io_error(bio);
3138                 r = DM_MAPIO_SUBMITTED;
3139         }
3140
3141         return r;
3142 }
3143
3144 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3145 {
3146         struct cache *cache = ti->private;
3147         unsigned long flags;
3148         size_t pb_data_size = get_per_bio_data_size(cache);
3149         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3150
3151         if (pb->tick) {
3152                 policy_tick(cache->policy, false);
3153
3154                 spin_lock_irqsave(&cache->lock, flags);
3155                 cache->need_tick_bio = true;
3156                 spin_unlock_irqrestore(&cache->lock, flags);
3157         }
3158
3159         check_for_quiesced_migrations(cache, pb);
3160         accounted_complete(cache, bio);
3161
3162         return 0;
3163 }
3164
3165 static int write_dirty_bitset(struct cache *cache)
3166 {
3167         unsigned i, r;
3168
3169         if (get_cache_mode(cache) >= CM_READ_ONLY)
3170                 return -EINVAL;
3171
3172         for (i = 0; i < from_cblock(cache->cache_size); i++) {
3173                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3174                                        is_dirty(cache, to_cblock(i)));
3175                 if (r) {
3176                         metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3177                         return r;
3178                 }
3179         }
3180
3181         return 0;
3182 }
3183
3184 static int write_discard_bitset(struct cache *cache)
3185 {
3186         unsigned i, r;
3187
3188         if (get_cache_mode(cache) >= CM_READ_ONLY)
3189                 return -EINVAL;
3190
3191         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3192                                            cache->discard_nr_blocks);
3193         if (r) {
3194                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3195                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3196                 return r;
3197         }
3198
3199         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3200                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3201                                          is_discarded(cache, to_dblock(i)));
3202                 if (r) {
3203                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
3204                         return r;
3205                 }
3206         }
3207
3208         return 0;
3209 }
3210
3211 static int write_hints(struct cache *cache)
3212 {
3213         int r;
3214
3215         if (get_cache_mode(cache) >= CM_READ_ONLY)
3216                 return -EINVAL;
3217
3218         r = dm_cache_write_hints(cache->cmd, cache->policy);
3219         if (r) {
3220                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3221                 return r;
3222         }
3223
3224         return 0;
3225 }
3226
3227 /*
3228  * returns true on success
3229  */
3230 static bool sync_metadata(struct cache *cache)
3231 {
3232         int r1, r2, r3, r4;
3233
3234         r1 = write_dirty_bitset(cache);
3235         if (r1)
3236                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3237
3238         r2 = write_discard_bitset(cache);
3239         if (r2)
3240                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3241
3242         save_stats(cache);
3243
3244         r3 = write_hints(cache);
3245         if (r3)
3246                 DMERR("%s: could not write hints", cache_device_name(cache));
3247
3248         /*
3249          * If writing the above metadata failed, we still commit, but don't
3250          * set the clean shutdown flag.  This will effectively force every
3251          * dirty bit to be set on reload.
3252          */
3253         r4 = commit(cache, !r1 && !r2 && !r3);
3254         if (r4)
3255                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3256
3257         return !r1 && !r2 && !r3 && !r4;
3258 }
3259
3260 static void cache_postsuspend(struct dm_target *ti)
3261 {
3262         struct cache *cache = ti->private;
3263
3264         start_quiescing(cache);
3265         wait_for_migrations(cache);
3266         stop_worker(cache);
3267         requeue_deferred_bios(cache);
3268         requeue_deferred_cells(cache);
3269         stop_quiescing(cache);
3270
3271         if (get_cache_mode(cache) == CM_WRITE)
3272                 (void) sync_metadata(cache);
3273 }
3274
3275 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3276                         bool dirty, uint32_t hint, bool hint_valid)
3277 {
3278         int r;
3279         struct cache *cache = context;
3280
3281         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3282         if (r)
3283                 return r;
3284
3285         if (dirty)
3286                 set_dirty(cache, oblock, cblock);
3287         else
3288                 clear_dirty(cache, oblock, cblock);
3289
3290         return 0;
3291 }
3292
3293 /*
3294  * The discard block size in the on disk metadata is not
3295  * neccessarily the same as we're currently using.  So we have to
3296  * be careful to only set the discarded attribute if we know it
3297  * covers a complete block of the new size.
3298  */
3299 struct discard_load_info {
3300         struct cache *cache;
3301
3302         /*
3303          * These blocks are sized using the on disk dblock size, rather
3304          * than the current one.
3305          */
3306         dm_block_t block_size;
3307         dm_block_t discard_begin, discard_end;
3308 };
3309
3310 static void discard_load_info_init(struct cache *cache,
3311                                    struct discard_load_info *li)
3312 {
3313         li->cache = cache;
3314         li->discard_begin = li->discard_end = 0;
3315 }
3316
3317 static void set_discard_range(struct discard_load_info *li)
3318 {
3319         sector_t b, e;
3320
3321         if (li->discard_begin == li->discard_end)
3322                 return;
3323
3324         /*
3325          * Convert to sectors.
3326          */
3327         b = li->discard_begin * li->block_size;
3328         e = li->discard_end * li->block_size;
3329
3330         /*
3331          * Then convert back to the current dblock size.
3332          */
3333         b = dm_sector_div_up(b, li->cache->discard_block_size);
3334         sector_div(e, li->cache->discard_block_size);
3335
3336         /*
3337          * The origin may have shrunk, so we need to check we're still in
3338          * bounds.
3339          */
3340         if (e > from_dblock(li->cache->discard_nr_blocks))
3341                 e = from_dblock(li->cache->discard_nr_blocks);
3342
3343         for (; b < e; b++)
3344                 set_discard(li->cache, to_dblock(b));
3345 }
3346
3347 static int load_discard(void *context, sector_t discard_block_size,
3348                         dm_dblock_t dblock, bool discard)
3349 {
3350         struct discard_load_info *li = context;
3351
3352         li->block_size = discard_block_size;
3353
3354         if (discard) {
3355                 if (from_dblock(dblock) == li->discard_end)
3356                         /*
3357                          * We're already in a discard range, just extend it.
3358                          */
3359                         li->discard_end = li->discard_end + 1ULL;
3360
3361                 else {
3362                         /*
3363                          * Emit the old range and start a new one.
3364                          */
3365                         set_discard_range(li);
3366                         li->discard_begin = from_dblock(dblock);
3367                         li->discard_end = li->discard_begin + 1ULL;
3368                 }
3369         } else {
3370                 set_discard_range(li);
3371                 li->discard_begin = li->discard_end = 0;
3372         }
3373
3374         return 0;
3375 }
3376
3377 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3378 {
3379         sector_t size = get_dev_size(cache->cache_dev);
3380         (void) sector_div(size, cache->sectors_per_block);
3381         return to_cblock(size);
3382 }
3383
3384 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3385 {
3386         if (from_cblock(new_size) > from_cblock(cache->cache_size))
3387                 return true;
3388
3389         /*
3390          * We can't drop a dirty block when shrinking the cache.
3391          */
3392         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3393                 new_size = to_cblock(from_cblock(new_size) + 1);
3394                 if (is_dirty(cache, new_size)) {
3395                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3396                               cache_device_name(cache),
3397                               (unsigned long long) from_cblock(new_size));
3398                         return false;
3399                 }
3400         }
3401
3402         return true;
3403 }
3404
3405 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3406 {
3407         int r;
3408
3409         r = dm_cache_resize(cache->cmd, new_size);
3410         if (r) {
3411                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3412                 metadata_operation_failed(cache, "dm_cache_resize", r);
3413                 return r;
3414         }
3415
3416         set_cache_size(cache, new_size);
3417
3418         return 0;
3419 }
3420
3421 static int cache_preresume(struct dm_target *ti)
3422 {
3423         int r = 0;
3424         struct cache *cache = ti->private;
3425         dm_cblock_t csize = get_cache_dev_size(cache);
3426
3427         /*
3428          * Check to see if the cache has resized.
3429          */
3430         if (!cache->sized) {
3431                 r = resize_cache_dev(cache, csize);
3432                 if (r)
3433                         return r;
3434
3435                 cache->sized = true;
3436
3437         } else if (csize != cache->cache_size) {
3438                 if (!can_resize(cache, csize))
3439                         return -EINVAL;
3440
3441                 r = resize_cache_dev(cache, csize);
3442                 if (r)
3443                         return r;
3444         }
3445
3446         if (!cache->loaded_mappings) {
3447                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3448                                            load_mapping, cache);
3449                 if (r) {
3450                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3451                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3452                         return r;
3453                 }
3454
3455                 cache->loaded_mappings = true;
3456         }
3457
3458         if (!cache->loaded_discards) {
3459                 struct discard_load_info li;
3460
3461                 /*
3462                  * The discard bitset could have been resized, or the
3463                  * discard block size changed.  To be safe we start by
3464                  * setting every dblock to not discarded.
3465                  */
3466                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3467
3468                 discard_load_info_init(cache, &li);
3469                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3470                 if (r) {
3471                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3472                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3473                         return r;
3474                 }
3475                 set_discard_range(&li);
3476
3477                 cache->loaded_discards = true;
3478         }
3479
3480         return r;
3481 }
3482
3483 static void cache_resume(struct dm_target *ti)
3484 {
3485         struct cache *cache = ti->private;
3486
3487         cache->need_tick_bio = true;
3488         do_waker(&cache->waker.work);
3489 }
3490
3491 /*
3492  * Status format:
3493  *
3494  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3495  * <cache block size> <#used cache blocks>/<#total cache blocks>
3496  * <#read hits> <#read misses> <#write hits> <#write misses>
3497  * <#demotions> <#promotions> <#dirty>
3498  * <#features> <features>*
3499  * <#core args> <core args>
3500  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3501  */
3502 static void cache_status(struct dm_target *ti, status_type_t type,
3503                          unsigned status_flags, char *result, unsigned maxlen)
3504 {
3505         int r = 0;
3506         unsigned i;
3507         ssize_t sz = 0;
3508         dm_block_t nr_free_blocks_metadata = 0;
3509         dm_block_t nr_blocks_metadata = 0;
3510         char buf[BDEVNAME_SIZE];
3511         struct cache *cache = ti->private;
3512         dm_cblock_t residency;
3513
3514         switch (type) {
3515         case STATUSTYPE_INFO:
3516                 if (get_cache_mode(cache) == CM_FAIL) {
3517                         DMEMIT("Fail");
3518                         break;
3519                 }
3520
3521                 /* Commit to ensure statistics aren't out-of-date */
3522                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3523                         (void) commit(cache, false);
3524
3525                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3526                 if (r) {
3527                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3528                               cache_device_name(cache), r);
3529                         goto err;
3530                 }
3531
3532                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3533                 if (r) {
3534                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3535                               cache_device_name(cache), r);
3536                         goto err;
3537                 }
3538
3539                 residency = policy_residency(cache->policy);
3540
3541                 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3542                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3543                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3544                        (unsigned long long)nr_blocks_metadata,
3545                        cache->sectors_per_block,
3546                        (unsigned long long) from_cblock(residency),
3547                        (unsigned long long) from_cblock(cache->cache_size),
3548                        (unsigned) atomic_read(&cache->stats.read_hit),
3549                        (unsigned) atomic_read(&cache->stats.read_miss),
3550                        (unsigned) atomic_read(&cache->stats.write_hit),
3551                        (unsigned) atomic_read(&cache->stats.write_miss),
3552                        (unsigned) atomic_read(&cache->stats.demotion),
3553                        (unsigned) atomic_read(&cache->stats.promotion),
3554                        (unsigned long) atomic_read(&cache->nr_dirty));
3555
3556                 if (writethrough_mode(&cache->features))
3557                         DMEMIT("1 writethrough ");
3558
3559                 else if (passthrough_mode(&cache->features))
3560                         DMEMIT("1 passthrough ");
3561
3562                 else if (writeback_mode(&cache->features))
3563                         DMEMIT("1 writeback ");
3564
3565                 else {
3566                         DMERR("%s: internal error: unknown io mode: %d",
3567                               cache_device_name(cache), (int) cache->features.io_mode);
3568                         goto err;
3569                 }
3570
3571                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3572
3573                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3574                 if (sz < maxlen) {
3575                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3576                         if (r)
3577                                 DMERR("%s: policy_emit_config_values returned %d",
3578                                       cache_device_name(cache), r);
3579                 }
3580
3581                 if (get_cache_mode(cache) == CM_READ_ONLY)
3582                         DMEMIT("ro ");
3583                 else
3584                         DMEMIT("rw ");
3585
3586                 if (dm_cache_metadata_needs_check(cache->cmd))
3587                         DMEMIT("needs_check ");
3588                 else
3589                         DMEMIT("- ");
3590
3591                 break;
3592
3593         case STATUSTYPE_TABLE:
3594                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3595                 DMEMIT("%s ", buf);
3596                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3597                 DMEMIT("%s ", buf);
3598                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3599                 DMEMIT("%s", buf);
3600
3601                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3602                         DMEMIT(" %s", cache->ctr_args[i]);
3603                 if (cache->nr_ctr_args)
3604                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3605         }
3606
3607         return;
3608
3609 err:
3610         DMEMIT("Error");
3611 }
3612
3613 /*
3614  * A cache block range can take two forms:
3615  *
3616  * i) A single cblock, eg. '3456'
3617  * ii) A begin and end cblock with dots between, eg. 123-234
3618  */
3619 static int parse_cblock_range(struct cache *cache, const char *str,
3620                               struct cblock_range *result)
3621 {
3622         char dummy;
3623         uint64_t b, e;
3624         int r;
3625
3626         /*
3627          * Try and parse form (ii) first.
3628          */
3629         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3630         if (r < 0)
3631                 return r;
3632
3633         if (r == 2) {
3634                 result->begin = to_cblock(b);
3635                 result->end = to_cblock(e);
3636                 return 0;
3637         }
3638
3639         /*
3640          * That didn't work, try form (i).
3641          */
3642         r = sscanf(str, "%llu%c", &b, &dummy);
3643         if (r < 0)
3644                 return r;
3645
3646         if (r == 1) {
3647                 result->begin = to_cblock(b);
3648                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3649                 return 0;
3650         }
3651
3652         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3653         return -EINVAL;
3654 }
3655
3656 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3657 {
3658         uint64_t b = from_cblock(range->begin);
3659         uint64_t e = from_cblock(range->end);
3660         uint64_t n = from_cblock(cache->cache_size);
3661
3662         if (b >= n) {
3663                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3664                       cache_device_name(cache), b, n);
3665                 return -EINVAL;
3666         }
3667
3668         if (e > n) {
3669                 DMERR("%s: end cblock out of range: %llu > %llu",
3670                       cache_device_name(cache), e, n);
3671                 return -EINVAL;
3672         }
3673
3674         if (b >= e) {
3675                 DMERR("%s: invalid cblock range: %llu >= %llu",
3676                       cache_device_name(cache), b, e);
3677                 return -EINVAL;
3678         }
3679
3680         return 0;
3681 }
3682
3683 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3684 {
3685         struct invalidation_request req;
3686
3687         INIT_LIST_HEAD(&req.list);
3688         req.cblocks = range;
3689         atomic_set(&req.complete, 0);
3690         req.err = 0;
3691         init_waitqueue_head(&req.result_wait);
3692
3693         spin_lock(&cache->invalidation_lock);
3694         list_add(&req.list, &cache->invalidation_requests);
3695         spin_unlock(&cache->invalidation_lock);
3696         wake_worker(cache);
3697
3698         wait_event(req.result_wait, atomic_read(&req.complete));
3699         return req.err;
3700 }
3701
3702 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3703                                               const char **cblock_ranges)
3704 {
3705         int r = 0;
3706         unsigned i;
3707         struct cblock_range range;
3708
3709         if (!passthrough_mode(&cache->features)) {
3710                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3711                       cache_device_name(cache));
3712                 return -EPERM;
3713         }
3714
3715         for (i = 0; i < count; i++) {
3716                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3717                 if (r)
3718                         break;
3719
3720                 r = validate_cblock_range(cache, &range);
3721                 if (r)
3722                         break;
3723
3724                 /*
3725                  * Pass begin and end origin blocks to the worker and wake it.
3726                  */
3727                 r = request_invalidation(cache, &range);
3728                 if (r)
3729                         break;
3730         }
3731
3732         return r;
3733 }
3734
3735 /*
3736  * Supports
3737  *      "<key> <value>"
3738  * and
3739  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3740  *
3741  * The key migration_threshold is supported by the cache target core.
3742  */
3743 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3744 {
3745         struct cache *cache = ti->private;
3746
3747         if (!argc)
3748                 return -EINVAL;
3749
3750         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3751                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3752                       cache_device_name(cache));
3753                 return -EOPNOTSUPP;
3754         }
3755
3756         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3757                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3758
3759         if (argc != 2)
3760                 return -EINVAL;
3761
3762         return set_config_value(cache, argv[0], argv[1]);
3763 }
3764
3765 static int cache_iterate_devices(struct dm_target *ti,
3766                                  iterate_devices_callout_fn fn, void *data)
3767 {
3768         int r = 0;
3769         struct cache *cache = ti->private;
3770
3771         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3772         if (!r)
3773                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3774
3775         return r;
3776 }
3777
3778 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3779 {
3780         /*
3781          * FIXME: these limits may be incompatible with the cache device
3782          */
3783         limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3784                                             cache->origin_sectors);
3785         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3786 }
3787
3788 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3789 {
3790         struct cache *cache = ti->private;
3791         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3792
3793         /*
3794          * If the system-determined stacked limits are compatible with the
3795          * cache's blocksize (io_opt is a factor) do not override them.
3796          */
3797         if (io_opt_sectors < cache->sectors_per_block ||
3798             do_div(io_opt_sectors, cache->sectors_per_block)) {
3799                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3800                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3801         }
3802         set_discard_limits(cache, limits);
3803 }
3804
3805 /*----------------------------------------------------------------*/
3806
3807 static struct target_type cache_target = {
3808         .name = "cache",
3809         .version = {1, 8, 0},
3810         .module = THIS_MODULE,
3811         .ctr = cache_ctr,
3812         .dtr = cache_dtr,
3813         .map = cache_map,
3814         .end_io = cache_end_io,
3815         .postsuspend = cache_postsuspend,
3816         .preresume = cache_preresume,
3817         .resume = cache_resume,
3818         .status = cache_status,
3819         .message = cache_message,
3820         .iterate_devices = cache_iterate_devices,
3821         .io_hints = cache_io_hints,
3822 };
3823
3824 static int __init dm_cache_init(void)
3825 {
3826         int r;
3827
3828         r = dm_register_target(&cache_target);
3829         if (r) {
3830                 DMERR("cache target registration failed: %d", r);
3831                 return r;
3832         }
3833
3834         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3835         if (!migration_cache) {
3836                 dm_unregister_target(&cache_target);
3837                 return -ENOMEM;
3838         }
3839
3840         return 0;
3841 }
3842
3843 static void __exit dm_cache_exit(void)
3844 {
3845         dm_unregister_target(&cache_target);
3846         kmem_cache_destroy(migration_cache);
3847 }
3848
3849 module_init(dm_cache_init);
3850 module_exit(dm_cache_exit);
3851
3852 MODULE_DESCRIPTION(DM_NAME " cache target");
3853 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3854 MODULE_LICENSE("GPL");