Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2012 Red Hat. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.h"
9#include "dm-bio-prison-v2.h"
10#include "dm-bio-record.h"
11#include "dm-cache-metadata.h"
12#include "dm-io-tracker.h"
13
14#include <linux/dm-io.h>
15#include <linux/dm-kcopyd.h>
16#include <linux/jiffies.h>
17#include <linux/init.h>
18#include <linux/mempool.h>
19#include <linux/module.h>
20#include <linux/rwsem.h>
21#include <linux/slab.h>
22#include <linux/vmalloc.h>
23
24#define DM_MSG_PREFIX "cache"
25
26DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
27 "A percentage of time allocated for copying to and/or from cache");
28
29/*----------------------------------------------------------------*/
30
31/*
32 * Glossary:
33 *
34 * oblock: index of an origin block
35 * cblock: index of a cache block
36 * promotion: movement of a block from origin to cache
37 * demotion: movement of a block from cache to origin
38 * migration: movement of a block between the origin and cache device,
39 * either direction
40 */
41
42/*----------------------------------------------------------------*/
43
44/*
45 * Represents a chunk of future work. 'input' allows continuations to pass
46 * values between themselves, typically error values.
47 */
48struct continuation {
49 struct work_struct ws;
50 blk_status_t input;
51};
52
53static inline void init_continuation(struct continuation *k,
54 void (*fn)(struct work_struct *))
55{
56 INIT_WORK(&k->ws, fn);
57 k->input = 0;
58}
59
60static inline void queue_continuation(struct workqueue_struct *wq,
61 struct continuation *k)
62{
63 queue_work(wq, &k->ws);
64}
65
66/*----------------------------------------------------------------*/
67
68/*
69 * The batcher collects together pieces of work that need a particular
70 * operation to occur before they can proceed (typically a commit).
71 */
72struct batcher {
73 /*
74 * The operation that everyone is waiting for.
75 */
76 blk_status_t (*commit_op)(void *context);
77 void *commit_context;
78
79 /*
80 * This is how bios should be issued once the commit op is complete
81 * (accounted_request).
82 */
83 void (*issue_op)(struct bio *bio, void *context);
84 void *issue_context;
85
86 /*
87 * Queued work gets put on here after commit.
88 */
89 struct workqueue_struct *wq;
90
91 spinlock_t lock;
92 struct list_head work_items;
93 struct bio_list bios;
94 struct work_struct commit_work;
95
96 bool commit_scheduled;
97};
98
99static void __commit(struct work_struct *_ws)
100{
101 struct batcher *b = container_of(_ws, struct batcher, commit_work);
102 blk_status_t r;
103 struct list_head work_items;
104 struct work_struct *ws, *tmp;
105 struct continuation *k;
106 struct bio *bio;
107 struct bio_list bios;
108
109 INIT_LIST_HEAD(&work_items);
110 bio_list_init(&bios);
111
112 /*
113 * We have to grab these before the commit_op to avoid a race
114 * condition.
115 */
116 spin_lock_irq(&b->lock);
117 list_splice_init(&b->work_items, &work_items);
118 bio_list_merge(&bios, &b->bios);
119 bio_list_init(&b->bios);
120 b->commit_scheduled = false;
121 spin_unlock_irq(&b->lock);
122
123 r = b->commit_op(b->commit_context);
124
125 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
126 k = container_of(ws, struct continuation, ws);
127 k->input = r;
128 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
129 queue_work(b->wq, ws);
130 }
131
132 while ((bio = bio_list_pop(&bios))) {
133 if (r) {
134 bio->bi_status = r;
135 bio_endio(bio);
136 } else
137 b->issue_op(bio, b->issue_context);
138 }
139}
140
141static void batcher_init(struct batcher *b,
142 blk_status_t (*commit_op)(void *),
143 void *commit_context,
144 void (*issue_op)(struct bio *bio, void *),
145 void *issue_context,
146 struct workqueue_struct *wq)
147{
148 b->commit_op = commit_op;
149 b->commit_context = commit_context;
150 b->issue_op = issue_op;
151 b->issue_context = issue_context;
152 b->wq = wq;
153
154 spin_lock_init(&b->lock);
155 INIT_LIST_HEAD(&b->work_items);
156 bio_list_init(&b->bios);
157 INIT_WORK(&b->commit_work, __commit);
158 b->commit_scheduled = false;
159}
160
161static void async_commit(struct batcher *b)
162{
163 queue_work(b->wq, &b->commit_work);
164}
165
166static void continue_after_commit(struct batcher *b, struct continuation *k)
167{
168 bool commit_scheduled;
169
170 spin_lock_irq(&b->lock);
171 commit_scheduled = b->commit_scheduled;
172 list_add_tail(&k->ws.entry, &b->work_items);
173 spin_unlock_irq(&b->lock);
174
175 if (commit_scheduled)
176 async_commit(b);
177}
178
179/*
180 * Bios are errored if commit failed.
181 */
182static void issue_after_commit(struct batcher *b, struct bio *bio)
183{
184 bool commit_scheduled;
185
186 spin_lock_irq(&b->lock);
187 commit_scheduled = b->commit_scheduled;
188 bio_list_add(&b->bios, bio);
189 spin_unlock_irq(&b->lock);
190
191 if (commit_scheduled)
192 async_commit(b);
193}
194
195/*
196 * Call this if some urgent work is waiting for the commit to complete.
197 */
198static void schedule_commit(struct batcher *b)
199{
200 bool immediate;
201
202 spin_lock_irq(&b->lock);
203 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
204 b->commit_scheduled = true;
205 spin_unlock_irq(&b->lock);
206
207 if (immediate)
208 async_commit(b);
209}
210
211/*
212 * There are a couple of places where we let a bio run, but want to do some
213 * work before calling its endio function. We do this by temporarily
214 * changing the endio fn.
215 */
216struct dm_hook_info {
217 bio_end_io_t *bi_end_io;
218};
219
220static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
221 bio_end_io_t *bi_end_io, void *bi_private)
222{
223 h->bi_end_io = bio->bi_end_io;
224
225 bio->bi_end_io = bi_end_io;
226 bio->bi_private = bi_private;
227}
228
229static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
230{
231 bio->bi_end_io = h->bi_end_io;
232}
233
234/*----------------------------------------------------------------*/
235
236#define MIGRATION_POOL_SIZE 128
237#define COMMIT_PERIOD HZ
238#define MIGRATION_COUNT_WINDOW 10
239
240/*
241 * The block size of the device holding cache data must be
242 * between 32KB and 1GB.
243 */
244#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
245#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
246
247enum cache_metadata_mode {
248 CM_WRITE, /* metadata may be changed */
249 CM_READ_ONLY, /* metadata may not be changed */
250 CM_FAIL
251};
252
253enum cache_io_mode {
254 /*
255 * Data is written to cached blocks only. These blocks are marked
256 * dirty. If you lose the cache device you will lose data.
257 * Potential performance increase for both reads and writes.
258 */
259 CM_IO_WRITEBACK,
260
261 /*
262 * Data is written to both cache and origin. Blocks are never
263 * dirty. Potential performance benfit for reads only.
264 */
265 CM_IO_WRITETHROUGH,
266
267 /*
268 * A degraded mode useful for various cache coherency situations
269 * (eg, rolling back snapshots). Reads and writes always go to the
270 * origin. If a write goes to a cached oblock, then the cache
271 * block is invalidated.
272 */
273 CM_IO_PASSTHROUGH
274};
275
276struct cache_features {
277 enum cache_metadata_mode mode;
278 enum cache_io_mode io_mode;
279 unsigned int metadata_version;
280 bool discard_passdown:1;
281};
282
283struct cache_stats {
284 atomic_t read_hit;
285 atomic_t read_miss;
286 atomic_t write_hit;
287 atomic_t write_miss;
288 atomic_t demotion;
289 atomic_t promotion;
290 atomic_t writeback;
291 atomic_t copies_avoided;
292 atomic_t cache_cell_clash;
293 atomic_t commit_count;
294 atomic_t discard_count;
295};
296
297struct cache {
298 struct dm_target *ti;
299 spinlock_t lock;
300
301 /*
302 * Fields for converting from sectors to blocks.
303 */
304 int sectors_per_block_shift;
305 sector_t sectors_per_block;
306
307 struct dm_cache_metadata *cmd;
308
309 /*
310 * Metadata is written to this device.
311 */
312 struct dm_dev *metadata_dev;
313
314 /*
315 * The slower of the two data devices. Typically a spindle.
316 */
317 struct dm_dev *origin_dev;
318
319 /*
320 * The faster of the two data devices. Typically an SSD.
321 */
322 struct dm_dev *cache_dev;
323
324 /*
325 * Size of the origin device in _complete_ blocks and native sectors.
326 */
327 dm_oblock_t origin_blocks;
328 sector_t origin_sectors;
329
330 /*
331 * Size of the cache device in blocks.
332 */
333 dm_cblock_t cache_size;
334
335 /*
336 * Invalidation fields.
337 */
338 spinlock_t invalidation_lock;
339 struct list_head invalidation_requests;
340
341 sector_t migration_threshold;
342 wait_queue_head_t migration_wait;
343 atomic_t nr_allocated_migrations;
344
345 /*
346 * The number of in flight migrations that are performing
347 * background io. eg, promotion, writeback.
348 */
349 atomic_t nr_io_migrations;
350
351 struct bio_list deferred_bios;
352
353 struct rw_semaphore quiesce_lock;
354
355 /*
356 * origin_blocks entries, discarded if set.
357 */
358 dm_dblock_t discard_nr_blocks;
359 unsigned long *discard_bitset;
360 uint32_t discard_block_size; /* a power of 2 times sectors per block */
361
362 /*
363 * Rather than reconstructing the table line for the status we just
364 * save it and regurgitate.
365 */
366 unsigned int nr_ctr_args;
367 const char **ctr_args;
368
369 struct dm_kcopyd_client *copier;
370 struct work_struct deferred_bio_worker;
371 struct work_struct migration_worker;
372 struct workqueue_struct *wq;
373 struct delayed_work waker;
374 struct dm_bio_prison_v2 *prison;
375
376 /*
377 * cache_size entries, dirty if set
378 */
379 unsigned long *dirty_bitset;
380 atomic_t nr_dirty;
381
382 unsigned int policy_nr_args;
383 struct dm_cache_policy *policy;
384
385 /*
386 * Cache features such as write-through.
387 */
388 struct cache_features features;
389
390 struct cache_stats stats;
391
392 bool need_tick_bio:1;
393 bool sized:1;
394 bool invalidate:1;
395 bool commit_requested:1;
396 bool loaded_mappings:1;
397 bool loaded_discards:1;
398
399 struct rw_semaphore background_work_lock;
400
401 struct batcher committer;
402 struct work_struct commit_ws;
403
404 struct dm_io_tracker tracker;
405
406 mempool_t migration_pool;
407
408 struct bio_set bs;
409};
410
411struct per_bio_data {
412 bool tick:1;
413 unsigned int req_nr:2;
414 struct dm_bio_prison_cell_v2 *cell;
415 struct dm_hook_info hook_info;
416 sector_t len;
417};
418
419struct dm_cache_migration {
420 struct continuation k;
421 struct cache *cache;
422
423 struct policy_work *op;
424 struct bio *overwrite_bio;
425 struct dm_bio_prison_cell_v2 *cell;
426
427 dm_cblock_t invalidate_cblock;
428 dm_oblock_t invalidate_oblock;
429};
430
431/*----------------------------------------------------------------*/
432
433static bool writethrough_mode(struct cache *cache)
434{
435 return cache->features.io_mode == CM_IO_WRITETHROUGH;
436}
437
438static bool writeback_mode(struct cache *cache)
439{
440 return cache->features.io_mode == CM_IO_WRITEBACK;
441}
442
443static inline bool passthrough_mode(struct cache *cache)
444{
445 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
446}
447
448/*----------------------------------------------------------------*/
449
450static void wake_deferred_bio_worker(struct cache *cache)
451{
452 queue_work(cache->wq, &cache->deferred_bio_worker);
453}
454
455static void wake_migration_worker(struct cache *cache)
456{
457 if (passthrough_mode(cache))
458 return;
459
460 queue_work(cache->wq, &cache->migration_worker);
461}
462
463/*----------------------------------------------------------------*/
464
465static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
466{
467 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
468}
469
470static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
471{
472 dm_bio_prison_free_cell_v2(cache->prison, cell);
473}
474
475static struct dm_cache_migration *alloc_migration(struct cache *cache)
476{
477 struct dm_cache_migration *mg;
478
479 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
480
481 memset(mg, 0, sizeof(*mg));
482
483 mg->cache = cache;
484 atomic_inc(&cache->nr_allocated_migrations);
485
486 return mg;
487}
488
489static void free_migration(struct dm_cache_migration *mg)
490{
491 struct cache *cache = mg->cache;
492
493 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
494 wake_up(&cache->migration_wait);
495
496 mempool_free(mg, &cache->migration_pool);
497}
498
499/*----------------------------------------------------------------*/
500
501static inline dm_oblock_t oblock_succ(dm_oblock_t b)
502{
503 return to_oblock(from_oblock(b) + 1ull);
504}
505
506static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
507{
508 key->virtual = 0;
509 key->dev = 0;
510 key->block_begin = from_oblock(begin);
511 key->block_end = from_oblock(end);
512}
513
514/*
515 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
516 * level 1 which prevents *both* READs and WRITEs.
517 */
518#define WRITE_LOCK_LEVEL 0
519#define READ_WRITE_LOCK_LEVEL 1
520
521static unsigned int lock_level(struct bio *bio)
522{
523 return bio_data_dir(bio) == WRITE ?
524 WRITE_LOCK_LEVEL :
525 READ_WRITE_LOCK_LEVEL;
526}
527
528/*
529 *--------------------------------------------------------------
530 * Per bio data
531 *--------------------------------------------------------------
532 */
533
534static struct per_bio_data *get_per_bio_data(struct bio *bio)
535{
536 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
537
538 BUG_ON(!pb);
539 return pb;
540}
541
542static struct per_bio_data *init_per_bio_data(struct bio *bio)
543{
544 struct per_bio_data *pb = get_per_bio_data(bio);
545
546 pb->tick = false;
547 pb->req_nr = dm_bio_get_target_bio_nr(bio);
548 pb->cell = NULL;
549 pb->len = 0;
550
551 return pb;
552}
553
554/*----------------------------------------------------------------*/
555
556static void defer_bio(struct cache *cache, struct bio *bio)
557{
558 spin_lock_irq(&cache->lock);
559 bio_list_add(&cache->deferred_bios, bio);
560 spin_unlock_irq(&cache->lock);
561
562 wake_deferred_bio_worker(cache);
563}
564
565static void defer_bios(struct cache *cache, struct bio_list *bios)
566{
567 spin_lock_irq(&cache->lock);
568 bio_list_merge(&cache->deferred_bios, bios);
569 bio_list_init(bios);
570 spin_unlock_irq(&cache->lock);
571
572 wake_deferred_bio_worker(cache);
573}
574
575/*----------------------------------------------------------------*/
576
577static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
578{
579 bool r;
580 struct per_bio_data *pb;
581 struct dm_cell_key_v2 key;
582 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
583 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
584
585 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
586
587 build_key(oblock, end, &key);
588 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
589 if (!r) {
590 /*
591 * Failed to get the lock.
592 */
593 free_prison_cell(cache, cell_prealloc);
594 return r;
595 }
596
597 if (cell != cell_prealloc)
598 free_prison_cell(cache, cell_prealloc);
599
600 pb = get_per_bio_data(bio);
601 pb->cell = cell;
602
603 return r;
604}
605
606/*----------------------------------------------------------------*/
607
608static bool is_dirty(struct cache *cache, dm_cblock_t b)
609{
610 return test_bit(from_cblock(b), cache->dirty_bitset);
611}
612
613static void set_dirty(struct cache *cache, dm_cblock_t cblock)
614{
615 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
616 atomic_inc(&cache->nr_dirty);
617 policy_set_dirty(cache->policy, cblock);
618 }
619}
620
621/*
622 * These two are called when setting after migrations to force the policy
623 * and dirty bitset to be in sync.
624 */
625static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
626{
627 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
628 atomic_inc(&cache->nr_dirty);
629 policy_set_dirty(cache->policy, cblock);
630}
631
632static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
633{
634 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
635 if (atomic_dec_return(&cache->nr_dirty) == 0)
636 dm_table_event(cache->ti->table);
637 }
638
639 policy_clear_dirty(cache->policy, cblock);
640}
641
642/*----------------------------------------------------------------*/
643
644static bool block_size_is_power_of_two(struct cache *cache)
645{
646 return cache->sectors_per_block_shift >= 0;
647}
648
649static dm_block_t block_div(dm_block_t b, uint32_t n)
650{
651 do_div(b, n);
652
653 return b;
654}
655
656static dm_block_t oblocks_per_dblock(struct cache *cache)
657{
658 dm_block_t oblocks = cache->discard_block_size;
659
660 if (block_size_is_power_of_two(cache))
661 oblocks >>= cache->sectors_per_block_shift;
662 else
663 oblocks = block_div(oblocks, cache->sectors_per_block);
664
665 return oblocks;
666}
667
668static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
669{
670 return to_dblock(block_div(from_oblock(oblock),
671 oblocks_per_dblock(cache)));
672}
673
674static void set_discard(struct cache *cache, dm_dblock_t b)
675{
676 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
677 atomic_inc(&cache->stats.discard_count);
678
679 spin_lock_irq(&cache->lock);
680 set_bit(from_dblock(b), cache->discard_bitset);
681 spin_unlock_irq(&cache->lock);
682}
683
684static void clear_discard(struct cache *cache, dm_dblock_t b)
685{
686 spin_lock_irq(&cache->lock);
687 clear_bit(from_dblock(b), cache->discard_bitset);
688 spin_unlock_irq(&cache->lock);
689}
690
691static bool is_discarded(struct cache *cache, dm_dblock_t b)
692{
693 int r;
694
695 spin_lock_irq(&cache->lock);
696 r = test_bit(from_dblock(b), cache->discard_bitset);
697 spin_unlock_irq(&cache->lock);
698
699 return r;
700}
701
702static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
703{
704 int r;
705
706 spin_lock_irq(&cache->lock);
707 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
708 cache->discard_bitset);
709 spin_unlock_irq(&cache->lock);
710
711 return r;
712}
713
714/*
715 * -------------------------------------------------------------
716 * Remapping
717 *--------------------------------------------------------------
718 */
719static void remap_to_origin(struct cache *cache, struct bio *bio)
720{
721 bio_set_dev(bio, cache->origin_dev->bdev);
722}
723
724static void remap_to_cache(struct cache *cache, struct bio *bio,
725 dm_cblock_t cblock)
726{
727 sector_t bi_sector = bio->bi_iter.bi_sector;
728 sector_t block = from_cblock(cblock);
729
730 bio_set_dev(bio, cache->cache_dev->bdev);
731 if (!block_size_is_power_of_two(cache))
732 bio->bi_iter.bi_sector =
733 (block * cache->sectors_per_block) +
734 sector_div(bi_sector, cache->sectors_per_block);
735 else
736 bio->bi_iter.bi_sector =
737 (block << cache->sectors_per_block_shift) |
738 (bi_sector & (cache->sectors_per_block - 1));
739}
740
741static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
742{
743 struct per_bio_data *pb;
744
745 spin_lock_irq(&cache->lock);
746 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
747 bio_op(bio) != REQ_OP_DISCARD) {
748 pb = get_per_bio_data(bio);
749 pb->tick = true;
750 cache->need_tick_bio = false;
751 }
752 spin_unlock_irq(&cache->lock);
753}
754
755static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
756 dm_oblock_t oblock)
757{
758 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
759 check_if_tick_bio_needed(cache, bio);
760 remap_to_origin(cache, bio);
761 if (bio_data_dir(bio) == WRITE)
762 clear_discard(cache, oblock_to_dblock(cache, oblock));
763}
764
765static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
766 dm_oblock_t oblock, dm_cblock_t cblock)
767{
768 check_if_tick_bio_needed(cache, bio);
769 remap_to_cache(cache, bio, cblock);
770 if (bio_data_dir(bio) == WRITE) {
771 set_dirty(cache, cblock);
772 clear_discard(cache, oblock_to_dblock(cache, oblock));
773 }
774}
775
776static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
777{
778 sector_t block_nr = bio->bi_iter.bi_sector;
779
780 if (!block_size_is_power_of_two(cache))
781 (void) sector_div(block_nr, cache->sectors_per_block);
782 else
783 block_nr >>= cache->sectors_per_block_shift;
784
785 return to_oblock(block_nr);
786}
787
788static bool accountable_bio(struct cache *cache, struct bio *bio)
789{
790 return bio_op(bio) != REQ_OP_DISCARD;
791}
792
793static void accounted_begin(struct cache *cache, struct bio *bio)
794{
795 struct per_bio_data *pb;
796
797 if (accountable_bio(cache, bio)) {
798 pb = get_per_bio_data(bio);
799 pb->len = bio_sectors(bio);
800 dm_iot_io_begin(&cache->tracker, pb->len);
801 }
802}
803
804static void accounted_complete(struct cache *cache, struct bio *bio)
805{
806 struct per_bio_data *pb = get_per_bio_data(bio);
807
808 dm_iot_io_end(&cache->tracker, pb->len);
809}
810
811static void accounted_request(struct cache *cache, struct bio *bio)
812{
813 accounted_begin(cache, bio);
814 dm_submit_bio_remap(bio, NULL);
815}
816
817static void issue_op(struct bio *bio, void *context)
818{
819 struct cache *cache = context;
820
821 accounted_request(cache, bio);
822}
823
824/*
825 * When running in writethrough mode we need to send writes to clean blocks
826 * to both the cache and origin devices. Clone the bio and send them in parallel.
827 */
828static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
829 dm_oblock_t oblock, dm_cblock_t cblock)
830{
831 struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
832 GFP_NOIO, &cache->bs);
833
834 BUG_ON(!origin_bio);
835
836 bio_chain(origin_bio, bio);
837
838 if (bio_data_dir(origin_bio) == WRITE)
839 clear_discard(cache, oblock_to_dblock(cache, oblock));
840 submit_bio(origin_bio);
841
842 remap_to_cache(cache, bio, cblock);
843}
844
845/*
846 *--------------------------------------------------------------
847 * Failure modes
848 *--------------------------------------------------------------
849 */
850static enum cache_metadata_mode get_cache_mode(struct cache *cache)
851{
852 return cache->features.mode;
853}
854
855static const char *cache_device_name(struct cache *cache)
856{
857 return dm_table_device_name(cache->ti->table);
858}
859
860static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
861{
862 static const char *descs[] = {
863 "write",
864 "read-only",
865 "fail"
866 };
867
868 dm_table_event(cache->ti->table);
869 DMINFO("%s: switching cache to %s mode",
870 cache_device_name(cache), descs[(int)mode]);
871}
872
873static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
874{
875 bool needs_check;
876 enum cache_metadata_mode old_mode = get_cache_mode(cache);
877
878 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
879 DMERR("%s: unable to read needs_check flag, setting failure mode.",
880 cache_device_name(cache));
881 new_mode = CM_FAIL;
882 }
883
884 if (new_mode == CM_WRITE && needs_check) {
885 DMERR("%s: unable to switch cache to write mode until repaired.",
886 cache_device_name(cache));
887 if (old_mode != new_mode)
888 new_mode = old_mode;
889 else
890 new_mode = CM_READ_ONLY;
891 }
892
893 /* Never move out of fail mode */
894 if (old_mode == CM_FAIL)
895 new_mode = CM_FAIL;
896
897 switch (new_mode) {
898 case CM_FAIL:
899 case CM_READ_ONLY:
900 dm_cache_metadata_set_read_only(cache->cmd);
901 break;
902
903 case CM_WRITE:
904 dm_cache_metadata_set_read_write(cache->cmd);
905 break;
906 }
907
908 cache->features.mode = new_mode;
909
910 if (new_mode != old_mode)
911 notify_mode_switch(cache, new_mode);
912}
913
914static void abort_transaction(struct cache *cache)
915{
916 const char *dev_name = cache_device_name(cache);
917
918 if (get_cache_mode(cache) >= CM_READ_ONLY)
919 return;
920
921 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
922 if (dm_cache_metadata_abort(cache->cmd)) {
923 DMERR("%s: failed to abort metadata transaction", dev_name);
924 set_cache_mode(cache, CM_FAIL);
925 }
926
927 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
928 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
929 set_cache_mode(cache, CM_FAIL);
930 }
931}
932
933static void metadata_operation_failed(struct cache *cache, const char *op, int r)
934{
935 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
936 cache_device_name(cache), op, r);
937 abort_transaction(cache);
938 set_cache_mode(cache, CM_READ_ONLY);
939}
940
941/*----------------------------------------------------------------*/
942
943static void load_stats(struct cache *cache)
944{
945 struct dm_cache_statistics stats;
946
947 dm_cache_metadata_get_stats(cache->cmd, &stats);
948 atomic_set(&cache->stats.read_hit, stats.read_hits);
949 atomic_set(&cache->stats.read_miss, stats.read_misses);
950 atomic_set(&cache->stats.write_hit, stats.write_hits);
951 atomic_set(&cache->stats.write_miss, stats.write_misses);
952}
953
954static void save_stats(struct cache *cache)
955{
956 struct dm_cache_statistics stats;
957
958 if (get_cache_mode(cache) >= CM_READ_ONLY)
959 return;
960
961 stats.read_hits = atomic_read(&cache->stats.read_hit);
962 stats.read_misses = atomic_read(&cache->stats.read_miss);
963 stats.write_hits = atomic_read(&cache->stats.write_hit);
964 stats.write_misses = atomic_read(&cache->stats.write_miss);
965
966 dm_cache_metadata_set_stats(cache->cmd, &stats);
967}
968
969static void update_stats(struct cache_stats *stats, enum policy_operation op)
970{
971 switch (op) {
972 case POLICY_PROMOTE:
973 atomic_inc(&stats->promotion);
974 break;
975
976 case POLICY_DEMOTE:
977 atomic_inc(&stats->demotion);
978 break;
979
980 case POLICY_WRITEBACK:
981 atomic_inc(&stats->writeback);
982 break;
983 }
984}
985
986/*
987 *---------------------------------------------------------------------
988 * Migration processing
989 *
990 * Migration covers moving data from the origin device to the cache, or
991 * vice versa.
992 *---------------------------------------------------------------------
993 */
994static void inc_io_migrations(struct cache *cache)
995{
996 atomic_inc(&cache->nr_io_migrations);
997}
998
999static void dec_io_migrations(struct cache *cache)
1000{
1001 atomic_dec(&cache->nr_io_migrations);
1002}
1003
1004static bool discard_or_flush(struct bio *bio)
1005{
1006 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1007}
1008
1009static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1010 dm_dblock_t *b, dm_dblock_t *e)
1011{
1012 sector_t sb = bio->bi_iter.bi_sector;
1013 sector_t se = bio_end_sector(bio);
1014
1015 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1016
1017 if (se - sb < cache->discard_block_size)
1018 *e = *b;
1019 else
1020 *e = to_dblock(block_div(se, cache->discard_block_size));
1021}
1022
1023/*----------------------------------------------------------------*/
1024
1025static void prevent_background_work(struct cache *cache)
1026{
1027 lockdep_off();
1028 down_write(&cache->background_work_lock);
1029 lockdep_on();
1030}
1031
1032static void allow_background_work(struct cache *cache)
1033{
1034 lockdep_off();
1035 up_write(&cache->background_work_lock);
1036 lockdep_on();
1037}
1038
1039static bool background_work_begin(struct cache *cache)
1040{
1041 bool r;
1042
1043 lockdep_off();
1044 r = down_read_trylock(&cache->background_work_lock);
1045 lockdep_on();
1046
1047 return r;
1048}
1049
1050static void background_work_end(struct cache *cache)
1051{
1052 lockdep_off();
1053 up_read(&cache->background_work_lock);
1054 lockdep_on();
1055}
1056
1057/*----------------------------------------------------------------*/
1058
1059static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1060{
1061 return (bio_data_dir(bio) == WRITE) &&
1062 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1063}
1064
1065static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1066{
1067 return writeback_mode(cache) &&
1068 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1069}
1070
1071static void quiesce(struct dm_cache_migration *mg,
1072 void (*continuation)(struct work_struct *))
1073{
1074 init_continuation(&mg->k, continuation);
1075 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1076}
1077
1078static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1079{
1080 struct continuation *k = container_of(ws, struct continuation, ws);
1081
1082 return container_of(k, struct dm_cache_migration, k);
1083}
1084
1085static void copy_complete(int read_err, unsigned long write_err, void *context)
1086{
1087 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1088
1089 if (read_err || write_err)
1090 mg->k.input = BLK_STS_IOERR;
1091
1092 queue_continuation(mg->cache->wq, &mg->k);
1093}
1094
1095static void copy(struct dm_cache_migration *mg, bool promote)
1096{
1097 struct dm_io_region o_region, c_region;
1098 struct cache *cache = mg->cache;
1099
1100 o_region.bdev = cache->origin_dev->bdev;
1101 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1102 o_region.count = cache->sectors_per_block;
1103
1104 c_region.bdev = cache->cache_dev->bdev;
1105 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1106 c_region.count = cache->sectors_per_block;
1107
1108 if (promote)
1109 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1110 else
1111 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1112}
1113
1114static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1115{
1116 struct per_bio_data *pb = get_per_bio_data(bio);
1117
1118 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1119 free_prison_cell(cache, pb->cell);
1120 pb->cell = NULL;
1121}
1122
1123static void overwrite_endio(struct bio *bio)
1124{
1125 struct dm_cache_migration *mg = bio->bi_private;
1126 struct cache *cache = mg->cache;
1127 struct per_bio_data *pb = get_per_bio_data(bio);
1128
1129 dm_unhook_bio(&pb->hook_info, bio);
1130
1131 if (bio->bi_status)
1132 mg->k.input = bio->bi_status;
1133
1134 queue_continuation(cache->wq, &mg->k);
1135}
1136
1137static void overwrite(struct dm_cache_migration *mg,
1138 void (*continuation)(struct work_struct *))
1139{
1140 struct bio *bio = mg->overwrite_bio;
1141 struct per_bio_data *pb = get_per_bio_data(bio);
1142
1143 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1144
1145 /*
1146 * The overwrite bio is part of the copy operation, as such it does
1147 * not set/clear discard or dirty flags.
1148 */
1149 if (mg->op->op == POLICY_PROMOTE)
1150 remap_to_cache(mg->cache, bio, mg->op->cblock);
1151 else
1152 remap_to_origin(mg->cache, bio);
1153
1154 init_continuation(&mg->k, continuation);
1155 accounted_request(mg->cache, bio);
1156}
1157
1158/*
1159 * Migration steps:
1160 *
1161 * 1) exclusive lock preventing WRITEs
1162 * 2) quiesce
1163 * 3) copy or issue overwrite bio
1164 * 4) upgrade to exclusive lock preventing READs and WRITEs
1165 * 5) quiesce
1166 * 6) update metadata and commit
1167 * 7) unlock
1168 */
1169static void mg_complete(struct dm_cache_migration *mg, bool success)
1170{
1171 struct bio_list bios;
1172 struct cache *cache = mg->cache;
1173 struct policy_work *op = mg->op;
1174 dm_cblock_t cblock = op->cblock;
1175
1176 if (success)
1177 update_stats(&cache->stats, op->op);
1178
1179 switch (op->op) {
1180 case POLICY_PROMOTE:
1181 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1182 policy_complete_background_work(cache->policy, op, success);
1183
1184 if (mg->overwrite_bio) {
1185 if (success)
1186 force_set_dirty(cache, cblock);
1187 else if (mg->k.input)
1188 mg->overwrite_bio->bi_status = mg->k.input;
1189 else
1190 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1191 bio_endio(mg->overwrite_bio);
1192 } else {
1193 if (success)
1194 force_clear_dirty(cache, cblock);
1195 dec_io_migrations(cache);
1196 }
1197 break;
1198
1199 case POLICY_DEMOTE:
1200 /*
1201 * We clear dirty here to update the nr_dirty counter.
1202 */
1203 if (success)
1204 force_clear_dirty(cache, cblock);
1205 policy_complete_background_work(cache->policy, op, success);
1206 dec_io_migrations(cache);
1207 break;
1208
1209 case POLICY_WRITEBACK:
1210 if (success)
1211 force_clear_dirty(cache, cblock);
1212 policy_complete_background_work(cache->policy, op, success);
1213 dec_io_migrations(cache);
1214 break;
1215 }
1216
1217 bio_list_init(&bios);
1218 if (mg->cell) {
1219 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1220 free_prison_cell(cache, mg->cell);
1221 }
1222
1223 free_migration(mg);
1224 defer_bios(cache, &bios);
1225 wake_migration_worker(cache);
1226
1227 background_work_end(cache);
1228}
1229
1230static void mg_success(struct work_struct *ws)
1231{
1232 struct dm_cache_migration *mg = ws_to_mg(ws);
1233
1234 mg_complete(mg, mg->k.input == 0);
1235}
1236
1237static void mg_update_metadata(struct work_struct *ws)
1238{
1239 int r;
1240 struct dm_cache_migration *mg = ws_to_mg(ws);
1241 struct cache *cache = mg->cache;
1242 struct policy_work *op = mg->op;
1243
1244 switch (op->op) {
1245 case POLICY_PROMOTE:
1246 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1247 if (r) {
1248 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1249 cache_device_name(cache));
1250 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1251
1252 mg_complete(mg, false);
1253 return;
1254 }
1255 mg_complete(mg, true);
1256 break;
1257
1258 case POLICY_DEMOTE:
1259 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1260 if (r) {
1261 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1262 cache_device_name(cache));
1263 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1264
1265 mg_complete(mg, false);
1266 return;
1267 }
1268
1269 /*
1270 * It would be nice if we only had to commit when a REQ_FLUSH
1271 * comes through. But there's one scenario that we have to
1272 * look out for:
1273 *
1274 * - vblock x in a cache block
1275 * - domotion occurs
1276 * - cache block gets reallocated and over written
1277 * - crash
1278 *
1279 * When we recover, because there was no commit the cache will
1280 * rollback to having the data for vblock x in the cache block.
1281 * But the cache block has since been overwritten, so it'll end
1282 * up pointing to data that was never in 'x' during the history
1283 * of the device.
1284 *
1285 * To avoid this issue we require a commit as part of the
1286 * demotion operation.
1287 */
1288 init_continuation(&mg->k, mg_success);
1289 continue_after_commit(&cache->committer, &mg->k);
1290 schedule_commit(&cache->committer);
1291 break;
1292
1293 case POLICY_WRITEBACK:
1294 mg_complete(mg, true);
1295 break;
1296 }
1297}
1298
1299static void mg_update_metadata_after_copy(struct work_struct *ws)
1300{
1301 struct dm_cache_migration *mg = ws_to_mg(ws);
1302
1303 /*
1304 * Did the copy succeed?
1305 */
1306 if (mg->k.input)
1307 mg_complete(mg, false);
1308 else
1309 mg_update_metadata(ws);
1310}
1311
1312static void mg_upgrade_lock(struct work_struct *ws)
1313{
1314 int r;
1315 struct dm_cache_migration *mg = ws_to_mg(ws);
1316
1317 /*
1318 * Did the copy succeed?
1319 */
1320 if (mg->k.input)
1321 mg_complete(mg, false);
1322
1323 else {
1324 /*
1325 * Now we want the lock to prevent both reads and writes.
1326 */
1327 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1328 READ_WRITE_LOCK_LEVEL);
1329 if (r < 0)
1330 mg_complete(mg, false);
1331
1332 else if (r)
1333 quiesce(mg, mg_update_metadata);
1334
1335 else
1336 mg_update_metadata(ws);
1337 }
1338}
1339
1340static void mg_full_copy(struct work_struct *ws)
1341{
1342 struct dm_cache_migration *mg = ws_to_mg(ws);
1343 struct cache *cache = mg->cache;
1344 struct policy_work *op = mg->op;
1345 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1346
1347 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1348 is_discarded_oblock(cache, op->oblock)) {
1349 mg_upgrade_lock(ws);
1350 return;
1351 }
1352
1353 init_continuation(&mg->k, mg_upgrade_lock);
1354 copy(mg, is_policy_promote);
1355}
1356
1357static void mg_copy(struct work_struct *ws)
1358{
1359 struct dm_cache_migration *mg = ws_to_mg(ws);
1360
1361 if (mg->overwrite_bio) {
1362 /*
1363 * No exclusive lock was held when we last checked if the bio
1364 * was optimisable. So we have to check again in case things
1365 * have changed (eg, the block may no longer be discarded).
1366 */
1367 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1368 /*
1369 * Fallback to a real full copy after doing some tidying up.
1370 */
1371 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1372
1373 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1374 mg->overwrite_bio = NULL;
1375 inc_io_migrations(mg->cache);
1376 mg_full_copy(ws);
1377 return;
1378 }
1379
1380 /*
1381 * It's safe to do this here, even though it's new data
1382 * because all IO has been locked out of the block.
1383 *
1384 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1385 * so _not_ using mg_upgrade_lock() as continutation.
1386 */
1387 overwrite(mg, mg_update_metadata_after_copy);
1388
1389 } else
1390 mg_full_copy(ws);
1391}
1392
1393static int mg_lock_writes(struct dm_cache_migration *mg)
1394{
1395 int r;
1396 struct dm_cell_key_v2 key;
1397 struct cache *cache = mg->cache;
1398 struct dm_bio_prison_cell_v2 *prealloc;
1399
1400 prealloc = alloc_prison_cell(cache);
1401
1402 /*
1403 * Prevent writes to the block, but allow reads to continue.
1404 * Unless we're using an overwrite bio, in which case we lock
1405 * everything.
1406 */
1407 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1408 r = dm_cell_lock_v2(cache->prison, &key,
1409 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1410 prealloc, &mg->cell);
1411 if (r < 0) {
1412 free_prison_cell(cache, prealloc);
1413 mg_complete(mg, false);
1414 return r;
1415 }
1416
1417 if (mg->cell != prealloc)
1418 free_prison_cell(cache, prealloc);
1419
1420 if (r == 0)
1421 mg_copy(&mg->k.ws);
1422 else
1423 quiesce(mg, mg_copy);
1424
1425 return 0;
1426}
1427
1428static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1429{
1430 struct dm_cache_migration *mg;
1431
1432 if (!background_work_begin(cache)) {
1433 policy_complete_background_work(cache->policy, op, false);
1434 return -EPERM;
1435 }
1436
1437 mg = alloc_migration(cache);
1438
1439 mg->op = op;
1440 mg->overwrite_bio = bio;
1441
1442 if (!bio)
1443 inc_io_migrations(cache);
1444
1445 return mg_lock_writes(mg);
1446}
1447
1448/*
1449 *--------------------------------------------------------------
1450 * invalidation processing
1451 *--------------------------------------------------------------
1452 */
1453
1454static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1455{
1456 struct bio_list bios;
1457 struct cache *cache = mg->cache;
1458
1459 bio_list_init(&bios);
1460 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1461 free_prison_cell(cache, mg->cell);
1462
1463 if (!success && mg->overwrite_bio)
1464 bio_io_error(mg->overwrite_bio);
1465
1466 free_migration(mg);
1467 defer_bios(cache, &bios);
1468
1469 background_work_end(cache);
1470}
1471
1472static void invalidate_completed(struct work_struct *ws)
1473{
1474 struct dm_cache_migration *mg = ws_to_mg(ws);
1475
1476 invalidate_complete(mg, !mg->k.input);
1477}
1478
1479static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1480{
1481 int r;
1482
1483 r = policy_invalidate_mapping(cache->policy, cblock);
1484 if (!r) {
1485 r = dm_cache_remove_mapping(cache->cmd, cblock);
1486 if (r) {
1487 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1488 cache_device_name(cache));
1489 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1490 }
1491
1492 } else if (r == -ENODATA) {
1493 /*
1494 * Harmless, already unmapped.
1495 */
1496 r = 0;
1497
1498 } else
1499 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1500
1501 return r;
1502}
1503
1504static void invalidate_remove(struct work_struct *ws)
1505{
1506 int r;
1507 struct dm_cache_migration *mg = ws_to_mg(ws);
1508 struct cache *cache = mg->cache;
1509
1510 r = invalidate_cblock(cache, mg->invalidate_cblock);
1511 if (r) {
1512 invalidate_complete(mg, false);
1513 return;
1514 }
1515
1516 init_continuation(&mg->k, invalidate_completed);
1517 continue_after_commit(&cache->committer, &mg->k);
1518 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1519 mg->overwrite_bio = NULL;
1520 schedule_commit(&cache->committer);
1521}
1522
1523static int invalidate_lock(struct dm_cache_migration *mg)
1524{
1525 int r;
1526 struct dm_cell_key_v2 key;
1527 struct cache *cache = mg->cache;
1528 struct dm_bio_prison_cell_v2 *prealloc;
1529
1530 prealloc = alloc_prison_cell(cache);
1531
1532 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1533 r = dm_cell_lock_v2(cache->prison, &key,
1534 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1535 if (r < 0) {
1536 free_prison_cell(cache, prealloc);
1537 invalidate_complete(mg, false);
1538 return r;
1539 }
1540
1541 if (mg->cell != prealloc)
1542 free_prison_cell(cache, prealloc);
1543
1544 if (r)
1545 quiesce(mg, invalidate_remove);
1546
1547 else {
1548 /*
1549 * We can't call invalidate_remove() directly here because we
1550 * might still be in request context.
1551 */
1552 init_continuation(&mg->k, invalidate_remove);
1553 queue_work(cache->wq, &mg->k.ws);
1554 }
1555
1556 return 0;
1557}
1558
1559static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1560 dm_oblock_t oblock, struct bio *bio)
1561{
1562 struct dm_cache_migration *mg;
1563
1564 if (!background_work_begin(cache))
1565 return -EPERM;
1566
1567 mg = alloc_migration(cache);
1568
1569 mg->overwrite_bio = bio;
1570 mg->invalidate_cblock = cblock;
1571 mg->invalidate_oblock = oblock;
1572
1573 return invalidate_lock(mg);
1574}
1575
1576/*
1577 *--------------------------------------------------------------
1578 * bio processing
1579 *--------------------------------------------------------------
1580 */
1581
1582enum busy {
1583 IDLE,
1584 BUSY
1585};
1586
1587static enum busy spare_migration_bandwidth(struct cache *cache)
1588{
1589 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1590 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1591 cache->sectors_per_block;
1592
1593 if (idle && current_volume <= cache->migration_threshold)
1594 return IDLE;
1595 else
1596 return BUSY;
1597}
1598
1599static void inc_hit_counter(struct cache *cache, struct bio *bio)
1600{
1601 atomic_inc(bio_data_dir(bio) == READ ?
1602 &cache->stats.read_hit : &cache->stats.write_hit);
1603}
1604
1605static void inc_miss_counter(struct cache *cache, struct bio *bio)
1606{
1607 atomic_inc(bio_data_dir(bio) == READ ?
1608 &cache->stats.read_miss : &cache->stats.write_miss);
1609}
1610
1611/*----------------------------------------------------------------*/
1612
1613static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1614 bool *commit_needed)
1615{
1616 int r, data_dir;
1617 bool rb, background_queued;
1618 dm_cblock_t cblock;
1619
1620 *commit_needed = false;
1621
1622 rb = bio_detain_shared(cache, block, bio);
1623 if (!rb) {
1624 /*
1625 * An exclusive lock is held for this block, so we have to
1626 * wait. We set the commit_needed flag so the current
1627 * transaction will be committed asap, allowing this lock
1628 * to be dropped.
1629 */
1630 *commit_needed = true;
1631 return DM_MAPIO_SUBMITTED;
1632 }
1633
1634 data_dir = bio_data_dir(bio);
1635
1636 if (optimisable_bio(cache, bio, block)) {
1637 struct policy_work *op = NULL;
1638
1639 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1640 if (unlikely(r && r != -ENOENT)) {
1641 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1642 cache_device_name(cache), r);
1643 bio_io_error(bio);
1644 return DM_MAPIO_SUBMITTED;
1645 }
1646
1647 if (r == -ENOENT && op) {
1648 bio_drop_shared_lock(cache, bio);
1649 BUG_ON(op->op != POLICY_PROMOTE);
1650 mg_start(cache, op, bio);
1651 return DM_MAPIO_SUBMITTED;
1652 }
1653 } else {
1654 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1655 if (unlikely(r && r != -ENOENT)) {
1656 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1657 cache_device_name(cache), r);
1658 bio_io_error(bio);
1659 return DM_MAPIO_SUBMITTED;
1660 }
1661
1662 if (background_queued)
1663 wake_migration_worker(cache);
1664 }
1665
1666 if (r == -ENOENT) {
1667 struct per_bio_data *pb = get_per_bio_data(bio);
1668
1669 /*
1670 * Miss.
1671 */
1672 inc_miss_counter(cache, bio);
1673 if (pb->req_nr == 0) {
1674 accounted_begin(cache, bio);
1675 remap_to_origin_clear_discard(cache, bio, block);
1676 } else {
1677 /*
1678 * This is a duplicate writethrough io that is no
1679 * longer needed because the block has been demoted.
1680 */
1681 bio_endio(bio);
1682 return DM_MAPIO_SUBMITTED;
1683 }
1684 } else {
1685 /*
1686 * Hit.
1687 */
1688 inc_hit_counter(cache, bio);
1689
1690 /*
1691 * Passthrough always maps to the origin, invalidating any
1692 * cache blocks that are written to.
1693 */
1694 if (passthrough_mode(cache)) {
1695 if (bio_data_dir(bio) == WRITE) {
1696 bio_drop_shared_lock(cache, bio);
1697 atomic_inc(&cache->stats.demotion);
1698 invalidate_start(cache, cblock, block, bio);
1699 } else
1700 remap_to_origin_clear_discard(cache, bio, block);
1701 } else {
1702 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1703 !is_dirty(cache, cblock)) {
1704 remap_to_origin_and_cache(cache, bio, block, cblock);
1705 accounted_begin(cache, bio);
1706 } else
1707 remap_to_cache_dirty(cache, bio, block, cblock);
1708 }
1709 }
1710
1711 /*
1712 * dm core turns FUA requests into a separate payload and FLUSH req.
1713 */
1714 if (bio->bi_opf & REQ_FUA) {
1715 /*
1716 * issue_after_commit will call accounted_begin a second time. So
1717 * we call accounted_complete() to avoid double accounting.
1718 */
1719 accounted_complete(cache, bio);
1720 issue_after_commit(&cache->committer, bio);
1721 *commit_needed = true;
1722 return DM_MAPIO_SUBMITTED;
1723 }
1724
1725 return DM_MAPIO_REMAPPED;
1726}
1727
1728static bool process_bio(struct cache *cache, struct bio *bio)
1729{
1730 bool commit_needed;
1731
1732 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1733 dm_submit_bio_remap(bio, NULL);
1734
1735 return commit_needed;
1736}
1737
1738/*
1739 * A non-zero return indicates read_only or fail_io mode.
1740 */
1741static int commit(struct cache *cache, bool clean_shutdown)
1742{
1743 int r;
1744
1745 if (get_cache_mode(cache) >= CM_READ_ONLY)
1746 return -EINVAL;
1747
1748 atomic_inc(&cache->stats.commit_count);
1749 r = dm_cache_commit(cache->cmd, clean_shutdown);
1750 if (r)
1751 metadata_operation_failed(cache, "dm_cache_commit", r);
1752
1753 return r;
1754}
1755
1756/*
1757 * Used by the batcher.
1758 */
1759static blk_status_t commit_op(void *context)
1760{
1761 struct cache *cache = context;
1762
1763 if (dm_cache_changed_this_transaction(cache->cmd))
1764 return errno_to_blk_status(commit(cache, false));
1765
1766 return 0;
1767}
1768
1769/*----------------------------------------------------------------*/
1770
1771static bool process_flush_bio(struct cache *cache, struct bio *bio)
1772{
1773 struct per_bio_data *pb = get_per_bio_data(bio);
1774
1775 if (!pb->req_nr)
1776 remap_to_origin(cache, bio);
1777 else
1778 remap_to_cache(cache, bio, 0);
1779
1780 issue_after_commit(&cache->committer, bio);
1781 return true;
1782}
1783
1784static bool process_discard_bio(struct cache *cache, struct bio *bio)
1785{
1786 dm_dblock_t b, e;
1787
1788 /*
1789 * FIXME: do we need to lock the region? Or can we just assume the
1790 * user wont be so foolish as to issue discard concurrently with
1791 * other IO?
1792 */
1793 calc_discard_block_range(cache, bio, &b, &e);
1794 while (b != e) {
1795 set_discard(cache, b);
1796 b = to_dblock(from_dblock(b) + 1);
1797 }
1798
1799 if (cache->features.discard_passdown) {
1800 remap_to_origin(cache, bio);
1801 dm_submit_bio_remap(bio, NULL);
1802 } else
1803 bio_endio(bio);
1804
1805 return false;
1806}
1807
1808static void process_deferred_bios(struct work_struct *ws)
1809{
1810 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1811
1812 bool commit_needed = false;
1813 struct bio_list bios;
1814 struct bio *bio;
1815
1816 bio_list_init(&bios);
1817
1818 spin_lock_irq(&cache->lock);
1819 bio_list_merge(&bios, &cache->deferred_bios);
1820 bio_list_init(&cache->deferred_bios);
1821 spin_unlock_irq(&cache->lock);
1822
1823 while ((bio = bio_list_pop(&bios))) {
1824 if (bio->bi_opf & REQ_PREFLUSH)
1825 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1826
1827 else if (bio_op(bio) == REQ_OP_DISCARD)
1828 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1829
1830 else
1831 commit_needed = process_bio(cache, bio) || commit_needed;
1832 cond_resched();
1833 }
1834
1835 if (commit_needed)
1836 schedule_commit(&cache->committer);
1837}
1838
1839/*
1840 *--------------------------------------------------------------
1841 * Main worker loop
1842 *--------------------------------------------------------------
1843 */
1844static void requeue_deferred_bios(struct cache *cache)
1845{
1846 struct bio *bio;
1847 struct bio_list bios;
1848
1849 bio_list_init(&bios);
1850 bio_list_merge(&bios, &cache->deferred_bios);
1851 bio_list_init(&cache->deferred_bios);
1852
1853 while ((bio = bio_list_pop(&bios))) {
1854 bio->bi_status = BLK_STS_DM_REQUEUE;
1855 bio_endio(bio);
1856 cond_resched();
1857 }
1858}
1859
1860/*
1861 * We want to commit periodically so that not too much
1862 * unwritten metadata builds up.
1863 */
1864static void do_waker(struct work_struct *ws)
1865{
1866 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1867
1868 policy_tick(cache->policy, true);
1869 wake_migration_worker(cache);
1870 schedule_commit(&cache->committer);
1871 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1872}
1873
1874static void check_migrations(struct work_struct *ws)
1875{
1876 int r;
1877 struct policy_work *op;
1878 struct cache *cache = container_of(ws, struct cache, migration_worker);
1879 enum busy b;
1880
1881 for (;;) {
1882 b = spare_migration_bandwidth(cache);
1883
1884 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1885 if (r == -ENODATA)
1886 break;
1887
1888 if (r) {
1889 DMERR_LIMIT("%s: policy_background_work failed",
1890 cache_device_name(cache));
1891 break;
1892 }
1893
1894 r = mg_start(cache, op, NULL);
1895 if (r)
1896 break;
1897
1898 cond_resched();
1899 }
1900}
1901
1902/*
1903 *--------------------------------------------------------------
1904 * Target methods
1905 *--------------------------------------------------------------
1906 */
1907
1908/*
1909 * This function gets called on the error paths of the constructor, so we
1910 * have to cope with a partially initialised struct.
1911 */
1912static void destroy(struct cache *cache)
1913{
1914 unsigned int i;
1915
1916 mempool_exit(&cache->migration_pool);
1917
1918 if (cache->prison)
1919 dm_bio_prison_destroy_v2(cache->prison);
1920
1921 cancel_delayed_work_sync(&cache->waker);
1922 if (cache->wq)
1923 destroy_workqueue(cache->wq);
1924
1925 if (cache->dirty_bitset)
1926 free_bitset(cache->dirty_bitset);
1927
1928 if (cache->discard_bitset)
1929 free_bitset(cache->discard_bitset);
1930
1931 if (cache->copier)
1932 dm_kcopyd_client_destroy(cache->copier);
1933
1934 if (cache->cmd)
1935 dm_cache_metadata_close(cache->cmd);
1936
1937 if (cache->metadata_dev)
1938 dm_put_device(cache->ti, cache->metadata_dev);
1939
1940 if (cache->origin_dev)
1941 dm_put_device(cache->ti, cache->origin_dev);
1942
1943 if (cache->cache_dev)
1944 dm_put_device(cache->ti, cache->cache_dev);
1945
1946 if (cache->policy)
1947 dm_cache_policy_destroy(cache->policy);
1948
1949 for (i = 0; i < cache->nr_ctr_args ; i++)
1950 kfree(cache->ctr_args[i]);
1951 kfree(cache->ctr_args);
1952
1953 bioset_exit(&cache->bs);
1954
1955 kfree(cache);
1956}
1957
1958static void cache_dtr(struct dm_target *ti)
1959{
1960 struct cache *cache = ti->private;
1961
1962 destroy(cache);
1963}
1964
1965static sector_t get_dev_size(struct dm_dev *dev)
1966{
1967 return bdev_nr_sectors(dev->bdev);
1968}
1969
1970/*----------------------------------------------------------------*/
1971
1972/*
1973 * Construct a cache device mapping.
1974 *
1975 * cache <metadata dev> <cache dev> <origin dev> <block size>
1976 * <#feature args> [<feature arg>]*
1977 * <policy> <#policy args> [<policy arg>]*
1978 *
1979 * metadata dev : fast device holding the persistent metadata
1980 * cache dev : fast device holding cached data blocks
1981 * origin dev : slow device holding original data blocks
1982 * block size : cache unit size in sectors
1983 *
1984 * #feature args : number of feature arguments passed
1985 * feature args : writethrough. (The default is writeback.)
1986 *
1987 * policy : the replacement policy to use
1988 * #policy args : an even number of policy arguments corresponding
1989 * to key/value pairs passed to the policy
1990 * policy args : key/value pairs passed to the policy
1991 * E.g. 'sequential_threshold 1024'
1992 * See cache-policies.txt for details.
1993 *
1994 * Optional feature arguments are:
1995 * writethrough : write through caching that prohibits cache block
1996 * content from being different from origin block content.
1997 * Without this argument, the default behaviour is to write
1998 * back cache block contents later for performance reasons,
1999 * so they may differ from the corresponding origin blocks.
2000 */
2001struct cache_args {
2002 struct dm_target *ti;
2003
2004 struct dm_dev *metadata_dev;
2005
2006 struct dm_dev *cache_dev;
2007 sector_t cache_sectors;
2008
2009 struct dm_dev *origin_dev;
2010 sector_t origin_sectors;
2011
2012 uint32_t block_size;
2013
2014 const char *policy_name;
2015 int policy_argc;
2016 const char **policy_argv;
2017
2018 struct cache_features features;
2019};
2020
2021static void destroy_cache_args(struct cache_args *ca)
2022{
2023 if (ca->metadata_dev)
2024 dm_put_device(ca->ti, ca->metadata_dev);
2025
2026 if (ca->cache_dev)
2027 dm_put_device(ca->ti, ca->cache_dev);
2028
2029 if (ca->origin_dev)
2030 dm_put_device(ca->ti, ca->origin_dev);
2031
2032 kfree(ca);
2033}
2034
2035static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2036{
2037 if (!as->argc) {
2038 *error = "Insufficient args";
2039 return false;
2040 }
2041
2042 return true;
2043}
2044
2045static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2046 char **error)
2047{
2048 int r;
2049 sector_t metadata_dev_size;
2050
2051 if (!at_least_one_arg(as, error))
2052 return -EINVAL;
2053
2054 r = dm_get_device(ca->ti, dm_shift_arg(as),
2055 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->metadata_dev);
2056 if (r) {
2057 *error = "Error opening metadata device";
2058 return r;
2059 }
2060
2061 metadata_dev_size = get_dev_size(ca->metadata_dev);
2062 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2063 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2064 ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2065
2066 return 0;
2067}
2068
2069static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2070 char **error)
2071{
2072 int r;
2073
2074 if (!at_least_one_arg(as, error))
2075 return -EINVAL;
2076
2077 r = dm_get_device(ca->ti, dm_shift_arg(as),
2078 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->cache_dev);
2079 if (r) {
2080 *error = "Error opening cache device";
2081 return r;
2082 }
2083 ca->cache_sectors = get_dev_size(ca->cache_dev);
2084
2085 return 0;
2086}
2087
2088static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2089 char **error)
2090{
2091 int r;
2092
2093 if (!at_least_one_arg(as, error))
2094 return -EINVAL;
2095
2096 r = dm_get_device(ca->ti, dm_shift_arg(as),
2097 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->origin_dev);
2098 if (r) {
2099 *error = "Error opening origin device";
2100 return r;
2101 }
2102
2103 ca->origin_sectors = get_dev_size(ca->origin_dev);
2104 if (ca->ti->len > ca->origin_sectors) {
2105 *error = "Device size larger than cached device";
2106 return -EINVAL;
2107 }
2108
2109 return 0;
2110}
2111
2112static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2113 char **error)
2114{
2115 unsigned long block_size;
2116
2117 if (!at_least_one_arg(as, error))
2118 return -EINVAL;
2119
2120 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2121 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2122 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2123 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2124 *error = "Invalid data block size";
2125 return -EINVAL;
2126 }
2127
2128 if (block_size > ca->cache_sectors) {
2129 *error = "Data block size is larger than the cache device";
2130 return -EINVAL;
2131 }
2132
2133 ca->block_size = block_size;
2134
2135 return 0;
2136}
2137
2138static void init_features(struct cache_features *cf)
2139{
2140 cf->mode = CM_WRITE;
2141 cf->io_mode = CM_IO_WRITEBACK;
2142 cf->metadata_version = 1;
2143 cf->discard_passdown = true;
2144}
2145
2146static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2147 char **error)
2148{
2149 static const struct dm_arg _args[] = {
2150 {0, 3, "Invalid number of cache feature arguments"},
2151 };
2152
2153 int r, mode_ctr = 0;
2154 unsigned int argc;
2155 const char *arg;
2156 struct cache_features *cf = &ca->features;
2157
2158 init_features(cf);
2159
2160 r = dm_read_arg_group(_args, as, &argc, error);
2161 if (r)
2162 return -EINVAL;
2163
2164 while (argc--) {
2165 arg = dm_shift_arg(as);
2166
2167 if (!strcasecmp(arg, "writeback")) {
2168 cf->io_mode = CM_IO_WRITEBACK;
2169 mode_ctr++;
2170 }
2171
2172 else if (!strcasecmp(arg, "writethrough")) {
2173 cf->io_mode = CM_IO_WRITETHROUGH;
2174 mode_ctr++;
2175 }
2176
2177 else if (!strcasecmp(arg, "passthrough")) {
2178 cf->io_mode = CM_IO_PASSTHROUGH;
2179 mode_ctr++;
2180 }
2181
2182 else if (!strcasecmp(arg, "metadata2"))
2183 cf->metadata_version = 2;
2184
2185 else if (!strcasecmp(arg, "no_discard_passdown"))
2186 cf->discard_passdown = false;
2187
2188 else {
2189 *error = "Unrecognised cache feature requested";
2190 return -EINVAL;
2191 }
2192 }
2193
2194 if (mode_ctr > 1) {
2195 *error = "Duplicate cache io_mode features requested";
2196 return -EINVAL;
2197 }
2198
2199 return 0;
2200}
2201
2202static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2203 char **error)
2204{
2205 static const struct dm_arg _args[] = {
2206 {0, 1024, "Invalid number of policy arguments"},
2207 };
2208
2209 int r;
2210
2211 if (!at_least_one_arg(as, error))
2212 return -EINVAL;
2213
2214 ca->policy_name = dm_shift_arg(as);
2215
2216 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2217 if (r)
2218 return -EINVAL;
2219
2220 ca->policy_argv = (const char **)as->argv;
2221 dm_consume_args(as, ca->policy_argc);
2222
2223 return 0;
2224}
2225
2226static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2227 char **error)
2228{
2229 int r;
2230 struct dm_arg_set as;
2231
2232 as.argc = argc;
2233 as.argv = argv;
2234
2235 r = parse_metadata_dev(ca, &as, error);
2236 if (r)
2237 return r;
2238
2239 r = parse_cache_dev(ca, &as, error);
2240 if (r)
2241 return r;
2242
2243 r = parse_origin_dev(ca, &as, error);
2244 if (r)
2245 return r;
2246
2247 r = parse_block_size(ca, &as, error);
2248 if (r)
2249 return r;
2250
2251 r = parse_features(ca, &as, error);
2252 if (r)
2253 return r;
2254
2255 r = parse_policy(ca, &as, error);
2256 if (r)
2257 return r;
2258
2259 return 0;
2260}
2261
2262/*----------------------------------------------------------------*/
2263
2264static struct kmem_cache *migration_cache;
2265
2266#define NOT_CORE_OPTION 1
2267
2268static int process_config_option(struct cache *cache, const char *key, const char *value)
2269{
2270 unsigned long tmp;
2271
2272 if (!strcasecmp(key, "migration_threshold")) {
2273 if (kstrtoul(value, 10, &tmp))
2274 return -EINVAL;
2275
2276 cache->migration_threshold = tmp;
2277 return 0;
2278 }
2279
2280 return NOT_CORE_OPTION;
2281}
2282
2283static int set_config_value(struct cache *cache, const char *key, const char *value)
2284{
2285 int r = process_config_option(cache, key, value);
2286
2287 if (r == NOT_CORE_OPTION)
2288 r = policy_set_config_value(cache->policy, key, value);
2289
2290 if (r)
2291 DMWARN("bad config value for %s: %s", key, value);
2292
2293 return r;
2294}
2295
2296static int set_config_values(struct cache *cache, int argc, const char **argv)
2297{
2298 int r = 0;
2299
2300 if (argc & 1) {
2301 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2302 return -EINVAL;
2303 }
2304
2305 while (argc) {
2306 r = set_config_value(cache, argv[0], argv[1]);
2307 if (r)
2308 break;
2309
2310 argc -= 2;
2311 argv += 2;
2312 }
2313
2314 return r;
2315}
2316
2317static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2318 char **error)
2319{
2320 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2321 cache->cache_size,
2322 cache->origin_sectors,
2323 cache->sectors_per_block);
2324 if (IS_ERR(p)) {
2325 *error = "Error creating cache's policy";
2326 return PTR_ERR(p);
2327 }
2328 cache->policy = p;
2329 BUG_ON(!cache->policy);
2330
2331 return 0;
2332}
2333
2334/*
2335 * We want the discard block size to be at least the size of the cache
2336 * block size and have no more than 2^14 discard blocks across the origin.
2337 */
2338#define MAX_DISCARD_BLOCKS (1 << 14)
2339
2340static bool too_many_discard_blocks(sector_t discard_block_size,
2341 sector_t origin_size)
2342{
2343 (void) sector_div(origin_size, discard_block_size);
2344
2345 return origin_size > MAX_DISCARD_BLOCKS;
2346}
2347
2348static sector_t calculate_discard_block_size(sector_t cache_block_size,
2349 sector_t origin_size)
2350{
2351 sector_t discard_block_size = cache_block_size;
2352
2353 if (origin_size)
2354 while (too_many_discard_blocks(discard_block_size, origin_size))
2355 discard_block_size *= 2;
2356
2357 return discard_block_size;
2358}
2359
2360static void set_cache_size(struct cache *cache, dm_cblock_t size)
2361{
2362 dm_block_t nr_blocks = from_cblock(size);
2363
2364 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2365 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2366 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2367 "Please consider increasing the cache block size to reduce the overall cache block count.",
2368 (unsigned long long) nr_blocks);
2369
2370 cache->cache_size = size;
2371}
2372
2373#define DEFAULT_MIGRATION_THRESHOLD 2048
2374
2375static int cache_create(struct cache_args *ca, struct cache **result)
2376{
2377 int r = 0;
2378 char **error = &ca->ti->error;
2379 struct cache *cache;
2380 struct dm_target *ti = ca->ti;
2381 dm_block_t origin_blocks;
2382 struct dm_cache_metadata *cmd;
2383 bool may_format = ca->features.mode == CM_WRITE;
2384
2385 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2386 if (!cache)
2387 return -ENOMEM;
2388
2389 cache->ti = ca->ti;
2390 ti->private = cache;
2391 ti->accounts_remapped_io = true;
2392 ti->num_flush_bios = 2;
2393 ti->flush_supported = true;
2394
2395 ti->num_discard_bios = 1;
2396 ti->discards_supported = true;
2397
2398 ti->per_io_data_size = sizeof(struct per_bio_data);
2399
2400 cache->features = ca->features;
2401 if (writethrough_mode(cache)) {
2402 /* Create bioset for writethrough bios issued to origin */
2403 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2404 if (r)
2405 goto bad;
2406 }
2407
2408 cache->metadata_dev = ca->metadata_dev;
2409 cache->origin_dev = ca->origin_dev;
2410 cache->cache_dev = ca->cache_dev;
2411
2412 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2413
2414 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2415 origin_blocks = block_div(origin_blocks, ca->block_size);
2416 cache->origin_blocks = to_oblock(origin_blocks);
2417
2418 cache->sectors_per_block = ca->block_size;
2419 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2420 r = -EINVAL;
2421 goto bad;
2422 }
2423
2424 if (ca->block_size & (ca->block_size - 1)) {
2425 dm_block_t cache_size = ca->cache_sectors;
2426
2427 cache->sectors_per_block_shift = -1;
2428 cache_size = block_div(cache_size, ca->block_size);
2429 set_cache_size(cache, to_cblock(cache_size));
2430 } else {
2431 cache->sectors_per_block_shift = __ffs(ca->block_size);
2432 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2433 }
2434
2435 r = create_cache_policy(cache, ca, error);
2436 if (r)
2437 goto bad;
2438
2439 cache->policy_nr_args = ca->policy_argc;
2440 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2441
2442 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2443 if (r) {
2444 *error = "Error setting cache policy's config values";
2445 goto bad;
2446 }
2447
2448 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2449 ca->block_size, may_format,
2450 dm_cache_policy_get_hint_size(cache->policy),
2451 ca->features.metadata_version);
2452 if (IS_ERR(cmd)) {
2453 *error = "Error creating metadata object";
2454 r = PTR_ERR(cmd);
2455 goto bad;
2456 }
2457 cache->cmd = cmd;
2458 set_cache_mode(cache, CM_WRITE);
2459 if (get_cache_mode(cache) != CM_WRITE) {
2460 *error = "Unable to get write access to metadata, please check/repair metadata.";
2461 r = -EINVAL;
2462 goto bad;
2463 }
2464
2465 if (passthrough_mode(cache)) {
2466 bool all_clean;
2467
2468 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2469 if (r) {
2470 *error = "dm_cache_metadata_all_clean() failed";
2471 goto bad;
2472 }
2473
2474 if (!all_clean) {
2475 *error = "Cannot enter passthrough mode unless all blocks are clean";
2476 r = -EINVAL;
2477 goto bad;
2478 }
2479
2480 policy_allow_migrations(cache->policy, false);
2481 }
2482
2483 spin_lock_init(&cache->lock);
2484 bio_list_init(&cache->deferred_bios);
2485 atomic_set(&cache->nr_allocated_migrations, 0);
2486 atomic_set(&cache->nr_io_migrations, 0);
2487 init_waitqueue_head(&cache->migration_wait);
2488
2489 r = -ENOMEM;
2490 atomic_set(&cache->nr_dirty, 0);
2491 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2492 if (!cache->dirty_bitset) {
2493 *error = "could not allocate dirty bitset";
2494 goto bad;
2495 }
2496 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2497
2498 cache->discard_block_size =
2499 calculate_discard_block_size(cache->sectors_per_block,
2500 cache->origin_sectors);
2501 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2502 cache->discard_block_size));
2503 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2504 if (!cache->discard_bitset) {
2505 *error = "could not allocate discard bitset";
2506 goto bad;
2507 }
2508 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2509
2510 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2511 if (IS_ERR(cache->copier)) {
2512 *error = "could not create kcopyd client";
2513 r = PTR_ERR(cache->copier);
2514 goto bad;
2515 }
2516
2517 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2518 if (!cache->wq) {
2519 *error = "could not create workqueue for metadata object";
2520 goto bad;
2521 }
2522 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2523 INIT_WORK(&cache->migration_worker, check_migrations);
2524 INIT_DELAYED_WORK(&cache->waker, do_waker);
2525
2526 cache->prison = dm_bio_prison_create_v2(cache->wq);
2527 if (!cache->prison) {
2528 *error = "could not create bio prison";
2529 goto bad;
2530 }
2531
2532 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2533 migration_cache);
2534 if (r) {
2535 *error = "Error creating cache's migration mempool";
2536 goto bad;
2537 }
2538
2539 cache->need_tick_bio = true;
2540 cache->sized = false;
2541 cache->invalidate = false;
2542 cache->commit_requested = false;
2543 cache->loaded_mappings = false;
2544 cache->loaded_discards = false;
2545
2546 load_stats(cache);
2547
2548 atomic_set(&cache->stats.demotion, 0);
2549 atomic_set(&cache->stats.promotion, 0);
2550 atomic_set(&cache->stats.copies_avoided, 0);
2551 atomic_set(&cache->stats.cache_cell_clash, 0);
2552 atomic_set(&cache->stats.commit_count, 0);
2553 atomic_set(&cache->stats.discard_count, 0);
2554
2555 spin_lock_init(&cache->invalidation_lock);
2556 INIT_LIST_HEAD(&cache->invalidation_requests);
2557
2558 batcher_init(&cache->committer, commit_op, cache,
2559 issue_op, cache, cache->wq);
2560 dm_iot_init(&cache->tracker);
2561
2562 init_rwsem(&cache->background_work_lock);
2563 prevent_background_work(cache);
2564
2565 *result = cache;
2566 return 0;
2567bad:
2568 destroy(cache);
2569 return r;
2570}
2571
2572static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2573{
2574 unsigned int i;
2575 const char **copy;
2576
2577 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2578 if (!copy)
2579 return -ENOMEM;
2580 for (i = 0; i < argc; i++) {
2581 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2582 if (!copy[i]) {
2583 while (i--)
2584 kfree(copy[i]);
2585 kfree(copy);
2586 return -ENOMEM;
2587 }
2588 }
2589
2590 cache->nr_ctr_args = argc;
2591 cache->ctr_args = copy;
2592
2593 return 0;
2594}
2595
2596static int cache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2597{
2598 int r = -EINVAL;
2599 struct cache_args *ca;
2600 struct cache *cache = NULL;
2601
2602 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2603 if (!ca) {
2604 ti->error = "Error allocating memory for cache";
2605 return -ENOMEM;
2606 }
2607 ca->ti = ti;
2608
2609 r = parse_cache_args(ca, argc, argv, &ti->error);
2610 if (r)
2611 goto out;
2612
2613 r = cache_create(ca, &cache);
2614 if (r)
2615 goto out;
2616
2617 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2618 if (r) {
2619 destroy(cache);
2620 goto out;
2621 }
2622
2623 ti->private = cache;
2624out:
2625 destroy_cache_args(ca);
2626 return r;
2627}
2628
2629/*----------------------------------------------------------------*/
2630
2631static int cache_map(struct dm_target *ti, struct bio *bio)
2632{
2633 struct cache *cache = ti->private;
2634
2635 int r;
2636 bool commit_needed;
2637 dm_oblock_t block = get_bio_block(cache, bio);
2638
2639 init_per_bio_data(bio);
2640 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2641 /*
2642 * This can only occur if the io goes to a partial block at
2643 * the end of the origin device. We don't cache these.
2644 * Just remap to the origin and carry on.
2645 */
2646 remap_to_origin(cache, bio);
2647 accounted_begin(cache, bio);
2648 return DM_MAPIO_REMAPPED;
2649 }
2650
2651 if (discard_or_flush(bio)) {
2652 defer_bio(cache, bio);
2653 return DM_MAPIO_SUBMITTED;
2654 }
2655
2656 r = map_bio(cache, bio, block, &commit_needed);
2657 if (commit_needed)
2658 schedule_commit(&cache->committer);
2659
2660 return r;
2661}
2662
2663static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2664{
2665 struct cache *cache = ti->private;
2666 unsigned long flags;
2667 struct per_bio_data *pb = get_per_bio_data(bio);
2668
2669 if (pb->tick) {
2670 policy_tick(cache->policy, false);
2671
2672 spin_lock_irqsave(&cache->lock, flags);
2673 cache->need_tick_bio = true;
2674 spin_unlock_irqrestore(&cache->lock, flags);
2675 }
2676
2677 bio_drop_shared_lock(cache, bio);
2678 accounted_complete(cache, bio);
2679
2680 return DM_ENDIO_DONE;
2681}
2682
2683static int write_dirty_bitset(struct cache *cache)
2684{
2685 int r;
2686
2687 if (get_cache_mode(cache) >= CM_READ_ONLY)
2688 return -EINVAL;
2689
2690 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2691 if (r)
2692 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2693
2694 return r;
2695}
2696
2697static int write_discard_bitset(struct cache *cache)
2698{
2699 unsigned int i, r;
2700
2701 if (get_cache_mode(cache) >= CM_READ_ONLY)
2702 return -EINVAL;
2703
2704 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2705 cache->discard_nr_blocks);
2706 if (r) {
2707 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2708 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2709 return r;
2710 }
2711
2712 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2713 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2714 is_discarded(cache, to_dblock(i)));
2715 if (r) {
2716 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2717 return r;
2718 }
2719 }
2720
2721 return 0;
2722}
2723
2724static int write_hints(struct cache *cache)
2725{
2726 int r;
2727
2728 if (get_cache_mode(cache) >= CM_READ_ONLY)
2729 return -EINVAL;
2730
2731 r = dm_cache_write_hints(cache->cmd, cache->policy);
2732 if (r) {
2733 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2734 return r;
2735 }
2736
2737 return 0;
2738}
2739
2740/*
2741 * returns true on success
2742 */
2743static bool sync_metadata(struct cache *cache)
2744{
2745 int r1, r2, r3, r4;
2746
2747 r1 = write_dirty_bitset(cache);
2748 if (r1)
2749 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2750
2751 r2 = write_discard_bitset(cache);
2752 if (r2)
2753 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2754
2755 save_stats(cache);
2756
2757 r3 = write_hints(cache);
2758 if (r3)
2759 DMERR("%s: could not write hints", cache_device_name(cache));
2760
2761 /*
2762 * If writing the above metadata failed, we still commit, but don't
2763 * set the clean shutdown flag. This will effectively force every
2764 * dirty bit to be set on reload.
2765 */
2766 r4 = commit(cache, !r1 && !r2 && !r3);
2767 if (r4)
2768 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2769
2770 return !r1 && !r2 && !r3 && !r4;
2771}
2772
2773static void cache_postsuspend(struct dm_target *ti)
2774{
2775 struct cache *cache = ti->private;
2776
2777 prevent_background_work(cache);
2778 BUG_ON(atomic_read(&cache->nr_io_migrations));
2779
2780 cancel_delayed_work_sync(&cache->waker);
2781 drain_workqueue(cache->wq);
2782 WARN_ON(cache->tracker.in_flight);
2783
2784 /*
2785 * If it's a flush suspend there won't be any deferred bios, so this
2786 * call is harmless.
2787 */
2788 requeue_deferred_bios(cache);
2789
2790 if (get_cache_mode(cache) == CM_WRITE)
2791 (void) sync_metadata(cache);
2792}
2793
2794static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2795 bool dirty, uint32_t hint, bool hint_valid)
2796{
2797 struct cache *cache = context;
2798
2799 if (dirty) {
2800 set_bit(from_cblock(cblock), cache->dirty_bitset);
2801 atomic_inc(&cache->nr_dirty);
2802 } else
2803 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2804
2805 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2806}
2807
2808/*
2809 * The discard block size in the on disk metadata is not
2810 * necessarily the same as we're currently using. So we have to
2811 * be careful to only set the discarded attribute if we know it
2812 * covers a complete block of the new size.
2813 */
2814struct discard_load_info {
2815 struct cache *cache;
2816
2817 /*
2818 * These blocks are sized using the on disk dblock size, rather
2819 * than the current one.
2820 */
2821 dm_block_t block_size;
2822 dm_block_t discard_begin, discard_end;
2823};
2824
2825static void discard_load_info_init(struct cache *cache,
2826 struct discard_load_info *li)
2827{
2828 li->cache = cache;
2829 li->discard_begin = li->discard_end = 0;
2830}
2831
2832static void set_discard_range(struct discard_load_info *li)
2833{
2834 sector_t b, e;
2835
2836 if (li->discard_begin == li->discard_end)
2837 return;
2838
2839 /*
2840 * Convert to sectors.
2841 */
2842 b = li->discard_begin * li->block_size;
2843 e = li->discard_end * li->block_size;
2844
2845 /*
2846 * Then convert back to the current dblock size.
2847 */
2848 b = dm_sector_div_up(b, li->cache->discard_block_size);
2849 sector_div(e, li->cache->discard_block_size);
2850
2851 /*
2852 * The origin may have shrunk, so we need to check we're still in
2853 * bounds.
2854 */
2855 if (e > from_dblock(li->cache->discard_nr_blocks))
2856 e = from_dblock(li->cache->discard_nr_blocks);
2857
2858 for (; b < e; b++)
2859 set_discard(li->cache, to_dblock(b));
2860}
2861
2862static int load_discard(void *context, sector_t discard_block_size,
2863 dm_dblock_t dblock, bool discard)
2864{
2865 struct discard_load_info *li = context;
2866
2867 li->block_size = discard_block_size;
2868
2869 if (discard) {
2870 if (from_dblock(dblock) == li->discard_end)
2871 /*
2872 * We're already in a discard range, just extend it.
2873 */
2874 li->discard_end = li->discard_end + 1ULL;
2875
2876 else {
2877 /*
2878 * Emit the old range and start a new one.
2879 */
2880 set_discard_range(li);
2881 li->discard_begin = from_dblock(dblock);
2882 li->discard_end = li->discard_begin + 1ULL;
2883 }
2884 } else {
2885 set_discard_range(li);
2886 li->discard_begin = li->discard_end = 0;
2887 }
2888
2889 return 0;
2890}
2891
2892static dm_cblock_t get_cache_dev_size(struct cache *cache)
2893{
2894 sector_t size = get_dev_size(cache->cache_dev);
2895 (void) sector_div(size, cache->sectors_per_block);
2896 return to_cblock(size);
2897}
2898
2899static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2900{
2901 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2902 if (cache->sized) {
2903 DMERR("%s: unable to extend cache due to missing cache table reload",
2904 cache_device_name(cache));
2905 return false;
2906 }
2907 }
2908
2909 /*
2910 * We can't drop a dirty block when shrinking the cache.
2911 */
2912 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2913 new_size = to_cblock(from_cblock(new_size) + 1);
2914 if (is_dirty(cache, new_size)) {
2915 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2916 cache_device_name(cache),
2917 (unsigned long long) from_cblock(new_size));
2918 return false;
2919 }
2920 }
2921
2922 return true;
2923}
2924
2925static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2926{
2927 int r;
2928
2929 r = dm_cache_resize(cache->cmd, new_size);
2930 if (r) {
2931 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2932 metadata_operation_failed(cache, "dm_cache_resize", r);
2933 return r;
2934 }
2935
2936 set_cache_size(cache, new_size);
2937
2938 return 0;
2939}
2940
2941static int cache_preresume(struct dm_target *ti)
2942{
2943 int r = 0;
2944 struct cache *cache = ti->private;
2945 dm_cblock_t csize = get_cache_dev_size(cache);
2946
2947 /*
2948 * Check to see if the cache has resized.
2949 */
2950 if (!cache->sized) {
2951 r = resize_cache_dev(cache, csize);
2952 if (r)
2953 return r;
2954
2955 cache->sized = true;
2956
2957 } else if (csize != cache->cache_size) {
2958 if (!can_resize(cache, csize))
2959 return -EINVAL;
2960
2961 r = resize_cache_dev(cache, csize);
2962 if (r)
2963 return r;
2964 }
2965
2966 if (!cache->loaded_mappings) {
2967 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2968 load_mapping, cache);
2969 if (r) {
2970 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2971 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2972 return r;
2973 }
2974
2975 cache->loaded_mappings = true;
2976 }
2977
2978 if (!cache->loaded_discards) {
2979 struct discard_load_info li;
2980
2981 /*
2982 * The discard bitset could have been resized, or the
2983 * discard block size changed. To be safe we start by
2984 * setting every dblock to not discarded.
2985 */
2986 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2987
2988 discard_load_info_init(cache, &li);
2989 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2990 if (r) {
2991 DMERR("%s: could not load origin discards", cache_device_name(cache));
2992 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2993 return r;
2994 }
2995 set_discard_range(&li);
2996
2997 cache->loaded_discards = true;
2998 }
2999
3000 return r;
3001}
3002
3003static void cache_resume(struct dm_target *ti)
3004{
3005 struct cache *cache = ti->private;
3006
3007 cache->need_tick_bio = true;
3008 allow_background_work(cache);
3009 do_waker(&cache->waker.work);
3010}
3011
3012static void emit_flags(struct cache *cache, char *result,
3013 unsigned int maxlen, ssize_t *sz_ptr)
3014{
3015 ssize_t sz = *sz_ptr;
3016 struct cache_features *cf = &cache->features;
3017 unsigned int count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3018
3019 DMEMIT("%u ", count);
3020
3021 if (cf->metadata_version == 2)
3022 DMEMIT("metadata2 ");
3023
3024 if (writethrough_mode(cache))
3025 DMEMIT("writethrough ");
3026
3027 else if (passthrough_mode(cache))
3028 DMEMIT("passthrough ");
3029
3030 else if (writeback_mode(cache))
3031 DMEMIT("writeback ");
3032
3033 else {
3034 DMEMIT("unknown ");
3035 DMERR("%s: internal error: unknown io mode: %d",
3036 cache_device_name(cache), (int) cf->io_mode);
3037 }
3038
3039 if (!cf->discard_passdown)
3040 DMEMIT("no_discard_passdown ");
3041
3042 *sz_ptr = sz;
3043}
3044
3045/*
3046 * Status format:
3047 *
3048 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3049 * <cache block size> <#used cache blocks>/<#total cache blocks>
3050 * <#read hits> <#read misses> <#write hits> <#write misses>
3051 * <#demotions> <#promotions> <#dirty>
3052 * <#features> <features>*
3053 * <#core args> <core args>
3054 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3055 */
3056static void cache_status(struct dm_target *ti, status_type_t type,
3057 unsigned int status_flags, char *result, unsigned int maxlen)
3058{
3059 int r = 0;
3060 unsigned int i;
3061 ssize_t sz = 0;
3062 dm_block_t nr_free_blocks_metadata = 0;
3063 dm_block_t nr_blocks_metadata = 0;
3064 char buf[BDEVNAME_SIZE];
3065 struct cache *cache = ti->private;
3066 dm_cblock_t residency;
3067 bool needs_check;
3068
3069 switch (type) {
3070 case STATUSTYPE_INFO:
3071 if (get_cache_mode(cache) == CM_FAIL) {
3072 DMEMIT("Fail");
3073 break;
3074 }
3075
3076 /* Commit to ensure statistics aren't out-of-date */
3077 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3078 (void) commit(cache, false);
3079
3080 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3081 if (r) {
3082 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3083 cache_device_name(cache), r);
3084 goto err;
3085 }
3086
3087 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3088 if (r) {
3089 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3090 cache_device_name(cache), r);
3091 goto err;
3092 }
3093
3094 residency = policy_residency(cache->policy);
3095
3096 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3097 (unsigned int)DM_CACHE_METADATA_BLOCK_SIZE,
3098 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3099 (unsigned long long)nr_blocks_metadata,
3100 (unsigned long long)cache->sectors_per_block,
3101 (unsigned long long) from_cblock(residency),
3102 (unsigned long long) from_cblock(cache->cache_size),
3103 (unsigned int) atomic_read(&cache->stats.read_hit),
3104 (unsigned int) atomic_read(&cache->stats.read_miss),
3105 (unsigned int) atomic_read(&cache->stats.write_hit),
3106 (unsigned int) atomic_read(&cache->stats.write_miss),
3107 (unsigned int) atomic_read(&cache->stats.demotion),
3108 (unsigned int) atomic_read(&cache->stats.promotion),
3109 (unsigned long) atomic_read(&cache->nr_dirty));
3110
3111 emit_flags(cache, result, maxlen, &sz);
3112
3113 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3114
3115 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3116 if (sz < maxlen) {
3117 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3118 if (r)
3119 DMERR("%s: policy_emit_config_values returned %d",
3120 cache_device_name(cache), r);
3121 }
3122
3123 if (get_cache_mode(cache) == CM_READ_ONLY)
3124 DMEMIT("ro ");
3125 else
3126 DMEMIT("rw ");
3127
3128 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3129
3130 if (r || needs_check)
3131 DMEMIT("needs_check ");
3132 else
3133 DMEMIT("- ");
3134
3135 break;
3136
3137 case STATUSTYPE_TABLE:
3138 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3139 DMEMIT("%s ", buf);
3140 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3141 DMEMIT("%s ", buf);
3142 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3143 DMEMIT("%s", buf);
3144
3145 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3146 DMEMIT(" %s", cache->ctr_args[i]);
3147 if (cache->nr_ctr_args)
3148 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3149 break;
3150
3151 case STATUSTYPE_IMA:
3152 DMEMIT_TARGET_NAME_VERSION(ti->type);
3153 if (get_cache_mode(cache) == CM_FAIL)
3154 DMEMIT(",metadata_mode=fail");
3155 else if (get_cache_mode(cache) == CM_READ_ONLY)
3156 DMEMIT(",metadata_mode=ro");
3157 else
3158 DMEMIT(",metadata_mode=rw");
3159
3160 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3161 DMEMIT(",cache_metadata_device=%s", buf);
3162 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3163 DMEMIT(",cache_device=%s", buf);
3164 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3165 DMEMIT(",cache_origin_device=%s", buf);
3166 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3167 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3168 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3169 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3170 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3171 DMEMIT(";");
3172 break;
3173 }
3174
3175 return;
3176
3177err:
3178 DMEMIT("Error");
3179}
3180
3181/*
3182 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3183 * the one-past-the-end value.
3184 */
3185struct cblock_range {
3186 dm_cblock_t begin;
3187 dm_cblock_t end;
3188};
3189
3190/*
3191 * A cache block range can take two forms:
3192 *
3193 * i) A single cblock, eg. '3456'
3194 * ii) A begin and end cblock with a dash between, eg. 123-234
3195 */
3196static int parse_cblock_range(struct cache *cache, const char *str,
3197 struct cblock_range *result)
3198{
3199 char dummy;
3200 uint64_t b, e;
3201 int r;
3202
3203 /*
3204 * Try and parse form (ii) first.
3205 */
3206 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3207 if (r < 0)
3208 return r;
3209
3210 if (r == 2) {
3211 result->begin = to_cblock(b);
3212 result->end = to_cblock(e);
3213 return 0;
3214 }
3215
3216 /*
3217 * That didn't work, try form (i).
3218 */
3219 r = sscanf(str, "%llu%c", &b, &dummy);
3220 if (r < 0)
3221 return r;
3222
3223 if (r == 1) {
3224 result->begin = to_cblock(b);
3225 result->end = to_cblock(from_cblock(result->begin) + 1u);
3226 return 0;
3227 }
3228
3229 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3230 return -EINVAL;
3231}
3232
3233static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3234{
3235 uint64_t b = from_cblock(range->begin);
3236 uint64_t e = from_cblock(range->end);
3237 uint64_t n = from_cblock(cache->cache_size);
3238
3239 if (b >= n) {
3240 DMERR("%s: begin cblock out of range: %llu >= %llu",
3241 cache_device_name(cache), b, n);
3242 return -EINVAL;
3243 }
3244
3245 if (e > n) {
3246 DMERR("%s: end cblock out of range: %llu > %llu",
3247 cache_device_name(cache), e, n);
3248 return -EINVAL;
3249 }
3250
3251 if (b >= e) {
3252 DMERR("%s: invalid cblock range: %llu >= %llu",
3253 cache_device_name(cache), b, e);
3254 return -EINVAL;
3255 }
3256
3257 return 0;
3258}
3259
3260static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3261{
3262 return to_cblock(from_cblock(b) + 1);
3263}
3264
3265static int request_invalidation(struct cache *cache, struct cblock_range *range)
3266{
3267 int r = 0;
3268
3269 /*
3270 * We don't need to do any locking here because we know we're in
3271 * passthrough mode. There's is potential for a race between an
3272 * invalidation triggered by an io and an invalidation message. This
3273 * is harmless, we must not worry if the policy call fails.
3274 */
3275 while (range->begin != range->end) {
3276 r = invalidate_cblock(cache, range->begin);
3277 if (r)
3278 return r;
3279
3280 range->begin = cblock_succ(range->begin);
3281 }
3282
3283 cache->commit_requested = true;
3284 return r;
3285}
3286
3287static int process_invalidate_cblocks_message(struct cache *cache, unsigned int count,
3288 const char **cblock_ranges)
3289{
3290 int r = 0;
3291 unsigned int i;
3292 struct cblock_range range;
3293
3294 if (!passthrough_mode(cache)) {
3295 DMERR("%s: cache has to be in passthrough mode for invalidation",
3296 cache_device_name(cache));
3297 return -EPERM;
3298 }
3299
3300 for (i = 0; i < count; i++) {
3301 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3302 if (r)
3303 break;
3304
3305 r = validate_cblock_range(cache, &range);
3306 if (r)
3307 break;
3308
3309 /*
3310 * Pass begin and end origin blocks to the worker and wake it.
3311 */
3312 r = request_invalidation(cache, &range);
3313 if (r)
3314 break;
3315 }
3316
3317 return r;
3318}
3319
3320/*
3321 * Supports
3322 * "<key> <value>"
3323 * and
3324 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3325 *
3326 * The key migration_threshold is supported by the cache target core.
3327 */
3328static int cache_message(struct dm_target *ti, unsigned int argc, char **argv,
3329 char *result, unsigned int maxlen)
3330{
3331 struct cache *cache = ti->private;
3332
3333 if (!argc)
3334 return -EINVAL;
3335
3336 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3337 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3338 cache_device_name(cache));
3339 return -EOPNOTSUPP;
3340 }
3341
3342 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3343 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3344
3345 if (argc != 2)
3346 return -EINVAL;
3347
3348 return set_config_value(cache, argv[0], argv[1]);
3349}
3350
3351static int cache_iterate_devices(struct dm_target *ti,
3352 iterate_devices_callout_fn fn, void *data)
3353{
3354 int r = 0;
3355 struct cache *cache = ti->private;
3356
3357 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3358 if (!r)
3359 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3360
3361 return r;
3362}
3363
3364/*
3365 * If discard_passdown was enabled verify that the origin device
3366 * supports discards. Disable discard_passdown if not.
3367 */
3368static void disable_passdown_if_not_supported(struct cache *cache)
3369{
3370 struct block_device *origin_bdev = cache->origin_dev->bdev;
3371 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3372 const char *reason = NULL;
3373
3374 if (!cache->features.discard_passdown)
3375 return;
3376
3377 if (!bdev_max_discard_sectors(origin_bdev))
3378 reason = "discard unsupported";
3379
3380 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3381 reason = "max discard sectors smaller than a block";
3382
3383 if (reason) {
3384 DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3385 origin_bdev, reason);
3386 cache->features.discard_passdown = false;
3387 }
3388}
3389
3390static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3391{
3392 struct block_device *origin_bdev = cache->origin_dev->bdev;
3393 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3394
3395 if (!cache->features.discard_passdown) {
3396 /* No passdown is done so setting own virtual limits */
3397 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3398 cache->origin_sectors);
3399 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3400 return;
3401 }
3402
3403 /*
3404 * cache_iterate_devices() is stacking both origin and fast device limits
3405 * but discards aren't passed to fast device, so inherit origin's limits.
3406 */
3407 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3408 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3409 limits->discard_granularity = origin_limits->discard_granularity;
3410 limits->discard_alignment = origin_limits->discard_alignment;
3411 limits->discard_misaligned = origin_limits->discard_misaligned;
3412}
3413
3414static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3415{
3416 struct cache *cache = ti->private;
3417 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3418
3419 /*
3420 * If the system-determined stacked limits are compatible with the
3421 * cache's blocksize (io_opt is a factor) do not override them.
3422 */
3423 if (io_opt_sectors < cache->sectors_per_block ||
3424 do_div(io_opt_sectors, cache->sectors_per_block)) {
3425 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3426 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3427 }
3428
3429 disable_passdown_if_not_supported(cache);
3430 set_discard_limits(cache, limits);
3431}
3432
3433/*----------------------------------------------------------------*/
3434
3435static struct target_type cache_target = {
3436 .name = "cache",
3437 .version = {2, 2, 0},
3438 .module = THIS_MODULE,
3439 .ctr = cache_ctr,
3440 .dtr = cache_dtr,
3441 .map = cache_map,
3442 .end_io = cache_end_io,
3443 .postsuspend = cache_postsuspend,
3444 .preresume = cache_preresume,
3445 .resume = cache_resume,
3446 .status = cache_status,
3447 .message = cache_message,
3448 .iterate_devices = cache_iterate_devices,
3449 .io_hints = cache_io_hints,
3450};
3451
3452static int __init dm_cache_init(void)
3453{
3454 int r;
3455
3456 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3457 if (!migration_cache)
3458 return -ENOMEM;
3459
3460 r = dm_register_target(&cache_target);
3461 if (r) {
3462 kmem_cache_destroy(migration_cache);
3463 return r;
3464 }
3465
3466 return 0;
3467}
3468
3469static void __exit dm_cache_exit(void)
3470{
3471 dm_unregister_target(&cache_target);
3472 kmem_cache_destroy(migration_cache);
3473}
3474
3475module_init(dm_cache_init);
3476module_exit(dm_cache_exit);
3477
3478MODULE_DESCRIPTION(DM_NAME " cache target");
3479MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3480MODULE_LICENSE("GPL");
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-v2.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/rwsem.h>
19#include <linux/slab.h>
20#include <linux/vmalloc.h>
21
22#define DM_MSG_PREFIX "cache"
23
24DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
25 "A percentage of time allocated for copying to and/or from cache");
26
27/*----------------------------------------------------------------*/
28
29/*
30 * Glossary:
31 *
32 * oblock: index of an origin block
33 * cblock: index of a cache block
34 * promotion: movement of a block from origin to cache
35 * demotion: movement of a block from cache to origin
36 * migration: movement of a block between the origin and cache device,
37 * either direction
38 */
39
40/*----------------------------------------------------------------*/
41
42struct io_tracker {
43 spinlock_t lock;
44
45 /*
46 * Sectors of in-flight IO.
47 */
48 sector_t in_flight;
49
50 /*
51 * The time, in jiffies, when this device became idle (if it is
52 * indeed idle).
53 */
54 unsigned long idle_time;
55 unsigned long last_update_time;
56};
57
58static void iot_init(struct io_tracker *iot)
59{
60 spin_lock_init(&iot->lock);
61 iot->in_flight = 0ul;
62 iot->idle_time = 0ul;
63 iot->last_update_time = jiffies;
64}
65
66static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
67{
68 if (iot->in_flight)
69 return false;
70
71 return time_after(jiffies, iot->idle_time + jifs);
72}
73
74static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
75{
76 bool r;
77
78 spin_lock_irq(&iot->lock);
79 r = __iot_idle_for(iot, jifs);
80 spin_unlock_irq(&iot->lock);
81
82 return r;
83}
84
85static void iot_io_begin(struct io_tracker *iot, sector_t len)
86{
87 spin_lock_irq(&iot->lock);
88 iot->in_flight += len;
89 spin_unlock_irq(&iot->lock);
90}
91
92static void __iot_io_end(struct io_tracker *iot, sector_t len)
93{
94 if (!len)
95 return;
96
97 iot->in_flight -= len;
98 if (!iot->in_flight)
99 iot->idle_time = jiffies;
100}
101
102static void iot_io_end(struct io_tracker *iot, sector_t len)
103{
104 unsigned long flags;
105
106 spin_lock_irqsave(&iot->lock, flags);
107 __iot_io_end(iot, len);
108 spin_unlock_irqrestore(&iot->lock, flags);
109}
110
111/*----------------------------------------------------------------*/
112
113/*
114 * Represents a chunk of future work. 'input' allows continuations to pass
115 * values between themselves, typically error values.
116 */
117struct continuation {
118 struct work_struct ws;
119 blk_status_t input;
120};
121
122static inline void init_continuation(struct continuation *k,
123 void (*fn)(struct work_struct *))
124{
125 INIT_WORK(&k->ws, fn);
126 k->input = 0;
127}
128
129static inline void queue_continuation(struct workqueue_struct *wq,
130 struct continuation *k)
131{
132 queue_work(wq, &k->ws);
133}
134
135/*----------------------------------------------------------------*/
136
137/*
138 * The batcher collects together pieces of work that need a particular
139 * operation to occur before they can proceed (typically a commit).
140 */
141struct batcher {
142 /*
143 * The operation that everyone is waiting for.
144 */
145 blk_status_t (*commit_op)(void *context);
146 void *commit_context;
147
148 /*
149 * This is how bios should be issued once the commit op is complete
150 * (accounted_request).
151 */
152 void (*issue_op)(struct bio *bio, void *context);
153 void *issue_context;
154
155 /*
156 * Queued work gets put on here after commit.
157 */
158 struct workqueue_struct *wq;
159
160 spinlock_t lock;
161 struct list_head work_items;
162 struct bio_list bios;
163 struct work_struct commit_work;
164
165 bool commit_scheduled;
166};
167
168static void __commit(struct work_struct *_ws)
169{
170 struct batcher *b = container_of(_ws, struct batcher, commit_work);
171 blk_status_t r;
172 struct list_head work_items;
173 struct work_struct *ws, *tmp;
174 struct continuation *k;
175 struct bio *bio;
176 struct bio_list bios;
177
178 INIT_LIST_HEAD(&work_items);
179 bio_list_init(&bios);
180
181 /*
182 * We have to grab these before the commit_op to avoid a race
183 * condition.
184 */
185 spin_lock_irq(&b->lock);
186 list_splice_init(&b->work_items, &work_items);
187 bio_list_merge(&bios, &b->bios);
188 bio_list_init(&b->bios);
189 b->commit_scheduled = false;
190 spin_unlock_irq(&b->lock);
191
192 r = b->commit_op(b->commit_context);
193
194 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
195 k = container_of(ws, struct continuation, ws);
196 k->input = r;
197 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
198 queue_work(b->wq, ws);
199 }
200
201 while ((bio = bio_list_pop(&bios))) {
202 if (r) {
203 bio->bi_status = r;
204 bio_endio(bio);
205 } else
206 b->issue_op(bio, b->issue_context);
207 }
208}
209
210static void batcher_init(struct batcher *b,
211 blk_status_t (*commit_op)(void *),
212 void *commit_context,
213 void (*issue_op)(struct bio *bio, void *),
214 void *issue_context,
215 struct workqueue_struct *wq)
216{
217 b->commit_op = commit_op;
218 b->commit_context = commit_context;
219 b->issue_op = issue_op;
220 b->issue_context = issue_context;
221 b->wq = wq;
222
223 spin_lock_init(&b->lock);
224 INIT_LIST_HEAD(&b->work_items);
225 bio_list_init(&b->bios);
226 INIT_WORK(&b->commit_work, __commit);
227 b->commit_scheduled = false;
228}
229
230static void async_commit(struct batcher *b)
231{
232 queue_work(b->wq, &b->commit_work);
233}
234
235static void continue_after_commit(struct batcher *b, struct continuation *k)
236{
237 bool commit_scheduled;
238
239 spin_lock_irq(&b->lock);
240 commit_scheduled = b->commit_scheduled;
241 list_add_tail(&k->ws.entry, &b->work_items);
242 spin_unlock_irq(&b->lock);
243
244 if (commit_scheduled)
245 async_commit(b);
246}
247
248/*
249 * Bios are errored if commit failed.
250 */
251static void issue_after_commit(struct batcher *b, struct bio *bio)
252{
253 bool commit_scheduled;
254
255 spin_lock_irq(&b->lock);
256 commit_scheduled = b->commit_scheduled;
257 bio_list_add(&b->bios, bio);
258 spin_unlock_irq(&b->lock);
259
260 if (commit_scheduled)
261 async_commit(b);
262}
263
264/*
265 * Call this if some urgent work is waiting for the commit to complete.
266 */
267static void schedule_commit(struct batcher *b)
268{
269 bool immediate;
270
271 spin_lock_irq(&b->lock);
272 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
273 b->commit_scheduled = true;
274 spin_unlock_irq(&b->lock);
275
276 if (immediate)
277 async_commit(b);
278}
279
280/*
281 * There are a couple of places where we let a bio run, but want to do some
282 * work before calling its endio function. We do this by temporarily
283 * changing the endio fn.
284 */
285struct dm_hook_info {
286 bio_end_io_t *bi_end_io;
287};
288
289static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
290 bio_end_io_t *bi_end_io, void *bi_private)
291{
292 h->bi_end_io = bio->bi_end_io;
293
294 bio->bi_end_io = bi_end_io;
295 bio->bi_private = bi_private;
296}
297
298static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
299{
300 bio->bi_end_io = h->bi_end_io;
301}
302
303/*----------------------------------------------------------------*/
304
305#define MIGRATION_POOL_SIZE 128
306#define COMMIT_PERIOD HZ
307#define MIGRATION_COUNT_WINDOW 10
308
309/*
310 * The block size of the device holding cache data must be
311 * between 32KB and 1GB.
312 */
313#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
314#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
315
316enum cache_metadata_mode {
317 CM_WRITE, /* metadata may be changed */
318 CM_READ_ONLY, /* metadata may not be changed */
319 CM_FAIL
320};
321
322enum cache_io_mode {
323 /*
324 * Data is written to cached blocks only. These blocks are marked
325 * dirty. If you lose the cache device you will lose data.
326 * Potential performance increase for both reads and writes.
327 */
328 CM_IO_WRITEBACK,
329
330 /*
331 * Data is written to both cache and origin. Blocks are never
332 * dirty. Potential performance benfit for reads only.
333 */
334 CM_IO_WRITETHROUGH,
335
336 /*
337 * A degraded mode useful for various cache coherency situations
338 * (eg, rolling back snapshots). Reads and writes always go to the
339 * origin. If a write goes to a cached oblock, then the cache
340 * block is invalidated.
341 */
342 CM_IO_PASSTHROUGH
343};
344
345struct cache_features {
346 enum cache_metadata_mode mode;
347 enum cache_io_mode io_mode;
348 unsigned metadata_version;
349 bool discard_passdown:1;
350};
351
352struct cache_stats {
353 atomic_t read_hit;
354 atomic_t read_miss;
355 atomic_t write_hit;
356 atomic_t write_miss;
357 atomic_t demotion;
358 atomic_t promotion;
359 atomic_t writeback;
360 atomic_t copies_avoided;
361 atomic_t cache_cell_clash;
362 atomic_t commit_count;
363 atomic_t discard_count;
364};
365
366struct cache {
367 struct dm_target *ti;
368 spinlock_t lock;
369
370 /*
371 * Fields for converting from sectors to blocks.
372 */
373 int sectors_per_block_shift;
374 sector_t sectors_per_block;
375
376 struct dm_cache_metadata *cmd;
377
378 /*
379 * Metadata is written to this device.
380 */
381 struct dm_dev *metadata_dev;
382
383 /*
384 * The slower of the two data devices. Typically a spindle.
385 */
386 struct dm_dev *origin_dev;
387
388 /*
389 * The faster of the two data devices. Typically an SSD.
390 */
391 struct dm_dev *cache_dev;
392
393 /*
394 * Size of the origin device in _complete_ blocks and native sectors.
395 */
396 dm_oblock_t origin_blocks;
397 sector_t origin_sectors;
398
399 /*
400 * Size of the cache device in blocks.
401 */
402 dm_cblock_t cache_size;
403
404 /*
405 * Invalidation fields.
406 */
407 spinlock_t invalidation_lock;
408 struct list_head invalidation_requests;
409
410 sector_t migration_threshold;
411 wait_queue_head_t migration_wait;
412 atomic_t nr_allocated_migrations;
413
414 /*
415 * The number of in flight migrations that are performing
416 * background io. eg, promotion, writeback.
417 */
418 atomic_t nr_io_migrations;
419
420 struct bio_list deferred_bios;
421
422 struct rw_semaphore quiesce_lock;
423
424 /*
425 * origin_blocks entries, discarded if set.
426 */
427 dm_dblock_t discard_nr_blocks;
428 unsigned long *discard_bitset;
429 uint32_t discard_block_size; /* a power of 2 times sectors per block */
430
431 /*
432 * Rather than reconstructing the table line for the status we just
433 * save it and regurgitate.
434 */
435 unsigned nr_ctr_args;
436 const char **ctr_args;
437
438 struct dm_kcopyd_client *copier;
439 struct work_struct deferred_bio_worker;
440 struct work_struct migration_worker;
441 struct workqueue_struct *wq;
442 struct delayed_work waker;
443 struct dm_bio_prison_v2 *prison;
444
445 /*
446 * cache_size entries, dirty if set
447 */
448 unsigned long *dirty_bitset;
449 atomic_t nr_dirty;
450
451 unsigned policy_nr_args;
452 struct dm_cache_policy *policy;
453
454 /*
455 * Cache features such as write-through.
456 */
457 struct cache_features features;
458
459 struct cache_stats stats;
460
461 bool need_tick_bio:1;
462 bool sized:1;
463 bool invalidate:1;
464 bool commit_requested:1;
465 bool loaded_mappings:1;
466 bool loaded_discards:1;
467
468 struct rw_semaphore background_work_lock;
469
470 struct batcher committer;
471 struct work_struct commit_ws;
472
473 struct io_tracker tracker;
474
475 mempool_t migration_pool;
476
477 struct bio_set bs;
478};
479
480struct per_bio_data {
481 bool tick:1;
482 unsigned req_nr:2;
483 struct dm_bio_prison_cell_v2 *cell;
484 struct dm_hook_info hook_info;
485 sector_t len;
486};
487
488struct dm_cache_migration {
489 struct continuation k;
490 struct cache *cache;
491
492 struct policy_work *op;
493 struct bio *overwrite_bio;
494 struct dm_bio_prison_cell_v2 *cell;
495
496 dm_cblock_t invalidate_cblock;
497 dm_oblock_t invalidate_oblock;
498};
499
500/*----------------------------------------------------------------*/
501
502static bool writethrough_mode(struct cache *cache)
503{
504 return cache->features.io_mode == CM_IO_WRITETHROUGH;
505}
506
507static bool writeback_mode(struct cache *cache)
508{
509 return cache->features.io_mode == CM_IO_WRITEBACK;
510}
511
512static inline bool passthrough_mode(struct cache *cache)
513{
514 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
515}
516
517/*----------------------------------------------------------------*/
518
519static void wake_deferred_bio_worker(struct cache *cache)
520{
521 queue_work(cache->wq, &cache->deferred_bio_worker);
522}
523
524static void wake_migration_worker(struct cache *cache)
525{
526 if (passthrough_mode(cache))
527 return;
528
529 queue_work(cache->wq, &cache->migration_worker);
530}
531
532/*----------------------------------------------------------------*/
533
534static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
535{
536 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
537}
538
539static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
540{
541 dm_bio_prison_free_cell_v2(cache->prison, cell);
542}
543
544static struct dm_cache_migration *alloc_migration(struct cache *cache)
545{
546 struct dm_cache_migration *mg;
547
548 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
549
550 memset(mg, 0, sizeof(*mg));
551
552 mg->cache = cache;
553 atomic_inc(&cache->nr_allocated_migrations);
554
555 return mg;
556}
557
558static void free_migration(struct dm_cache_migration *mg)
559{
560 struct cache *cache = mg->cache;
561
562 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
563 wake_up(&cache->migration_wait);
564
565 mempool_free(mg, &cache->migration_pool);
566}
567
568/*----------------------------------------------------------------*/
569
570static inline dm_oblock_t oblock_succ(dm_oblock_t b)
571{
572 return to_oblock(from_oblock(b) + 1ull);
573}
574
575static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
576{
577 key->virtual = 0;
578 key->dev = 0;
579 key->block_begin = from_oblock(begin);
580 key->block_end = from_oblock(end);
581}
582
583/*
584 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
585 * level 1 which prevents *both* READs and WRITEs.
586 */
587#define WRITE_LOCK_LEVEL 0
588#define READ_WRITE_LOCK_LEVEL 1
589
590static unsigned lock_level(struct bio *bio)
591{
592 return bio_data_dir(bio) == WRITE ?
593 WRITE_LOCK_LEVEL :
594 READ_WRITE_LOCK_LEVEL;
595}
596
597/*----------------------------------------------------------------
598 * Per bio data
599 *--------------------------------------------------------------*/
600
601static struct per_bio_data *get_per_bio_data(struct bio *bio)
602{
603 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
604 BUG_ON(!pb);
605 return pb;
606}
607
608static struct per_bio_data *init_per_bio_data(struct bio *bio)
609{
610 struct per_bio_data *pb = get_per_bio_data(bio);
611
612 pb->tick = false;
613 pb->req_nr = dm_bio_get_target_bio_nr(bio);
614 pb->cell = NULL;
615 pb->len = 0;
616
617 return pb;
618}
619
620/*----------------------------------------------------------------*/
621
622static void defer_bio(struct cache *cache, struct bio *bio)
623{
624 spin_lock_irq(&cache->lock);
625 bio_list_add(&cache->deferred_bios, bio);
626 spin_unlock_irq(&cache->lock);
627
628 wake_deferred_bio_worker(cache);
629}
630
631static void defer_bios(struct cache *cache, struct bio_list *bios)
632{
633 spin_lock_irq(&cache->lock);
634 bio_list_merge(&cache->deferred_bios, bios);
635 bio_list_init(bios);
636 spin_unlock_irq(&cache->lock);
637
638 wake_deferred_bio_worker(cache);
639}
640
641/*----------------------------------------------------------------*/
642
643static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
644{
645 bool r;
646 struct per_bio_data *pb;
647 struct dm_cell_key_v2 key;
648 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
649 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
650
651 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
652
653 build_key(oblock, end, &key);
654 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
655 if (!r) {
656 /*
657 * Failed to get the lock.
658 */
659 free_prison_cell(cache, cell_prealloc);
660 return r;
661 }
662
663 if (cell != cell_prealloc)
664 free_prison_cell(cache, cell_prealloc);
665
666 pb = get_per_bio_data(bio);
667 pb->cell = cell;
668
669 return r;
670}
671
672/*----------------------------------------------------------------*/
673
674static bool is_dirty(struct cache *cache, dm_cblock_t b)
675{
676 return test_bit(from_cblock(b), cache->dirty_bitset);
677}
678
679static void set_dirty(struct cache *cache, dm_cblock_t cblock)
680{
681 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
682 atomic_inc(&cache->nr_dirty);
683 policy_set_dirty(cache->policy, cblock);
684 }
685}
686
687/*
688 * These two are called when setting after migrations to force the policy
689 * and dirty bitset to be in sync.
690 */
691static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
692{
693 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
694 atomic_inc(&cache->nr_dirty);
695 policy_set_dirty(cache->policy, cblock);
696}
697
698static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
699{
700 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
701 if (atomic_dec_return(&cache->nr_dirty) == 0)
702 dm_table_event(cache->ti->table);
703 }
704
705 policy_clear_dirty(cache->policy, cblock);
706}
707
708/*----------------------------------------------------------------*/
709
710static bool block_size_is_power_of_two(struct cache *cache)
711{
712 return cache->sectors_per_block_shift >= 0;
713}
714
715/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
716#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
717__always_inline
718#endif
719static dm_block_t block_div(dm_block_t b, uint32_t n)
720{
721 do_div(b, n);
722
723 return b;
724}
725
726static dm_block_t oblocks_per_dblock(struct cache *cache)
727{
728 dm_block_t oblocks = cache->discard_block_size;
729
730 if (block_size_is_power_of_two(cache))
731 oblocks >>= cache->sectors_per_block_shift;
732 else
733 oblocks = block_div(oblocks, cache->sectors_per_block);
734
735 return oblocks;
736}
737
738static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
739{
740 return to_dblock(block_div(from_oblock(oblock),
741 oblocks_per_dblock(cache)));
742}
743
744static void set_discard(struct cache *cache, dm_dblock_t b)
745{
746 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
747 atomic_inc(&cache->stats.discard_count);
748
749 spin_lock_irq(&cache->lock);
750 set_bit(from_dblock(b), cache->discard_bitset);
751 spin_unlock_irq(&cache->lock);
752}
753
754static void clear_discard(struct cache *cache, dm_dblock_t b)
755{
756 spin_lock_irq(&cache->lock);
757 clear_bit(from_dblock(b), cache->discard_bitset);
758 spin_unlock_irq(&cache->lock);
759}
760
761static bool is_discarded(struct cache *cache, dm_dblock_t b)
762{
763 int r;
764 spin_lock_irq(&cache->lock);
765 r = test_bit(from_dblock(b), cache->discard_bitset);
766 spin_unlock_irq(&cache->lock);
767
768 return r;
769}
770
771static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
772{
773 int r;
774 spin_lock_irq(&cache->lock);
775 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
776 cache->discard_bitset);
777 spin_unlock_irq(&cache->lock);
778
779 return r;
780}
781
782/*----------------------------------------------------------------
783 * Remapping
784 *--------------------------------------------------------------*/
785static void remap_to_origin(struct cache *cache, struct bio *bio)
786{
787 bio_set_dev(bio, cache->origin_dev->bdev);
788}
789
790static void remap_to_cache(struct cache *cache, struct bio *bio,
791 dm_cblock_t cblock)
792{
793 sector_t bi_sector = bio->bi_iter.bi_sector;
794 sector_t block = from_cblock(cblock);
795
796 bio_set_dev(bio, cache->cache_dev->bdev);
797 if (!block_size_is_power_of_two(cache))
798 bio->bi_iter.bi_sector =
799 (block * cache->sectors_per_block) +
800 sector_div(bi_sector, cache->sectors_per_block);
801 else
802 bio->bi_iter.bi_sector =
803 (block << cache->sectors_per_block_shift) |
804 (bi_sector & (cache->sectors_per_block - 1));
805}
806
807static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
808{
809 struct per_bio_data *pb;
810
811 spin_lock_irq(&cache->lock);
812 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
813 bio_op(bio) != REQ_OP_DISCARD) {
814 pb = get_per_bio_data(bio);
815 pb->tick = true;
816 cache->need_tick_bio = false;
817 }
818 spin_unlock_irq(&cache->lock);
819}
820
821static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
822 dm_oblock_t oblock, bool bio_has_pbd)
823{
824 if (bio_has_pbd)
825 check_if_tick_bio_needed(cache, bio);
826 remap_to_origin(cache, bio);
827 if (bio_data_dir(bio) == WRITE)
828 clear_discard(cache, oblock_to_dblock(cache, oblock));
829}
830
831static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
832 dm_oblock_t oblock)
833{
834 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
835 __remap_to_origin_clear_discard(cache, bio, oblock, true);
836}
837
838static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
839 dm_oblock_t oblock, dm_cblock_t cblock)
840{
841 check_if_tick_bio_needed(cache, bio);
842 remap_to_cache(cache, bio, cblock);
843 if (bio_data_dir(bio) == WRITE) {
844 set_dirty(cache, cblock);
845 clear_discard(cache, oblock_to_dblock(cache, oblock));
846 }
847}
848
849static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
850{
851 sector_t block_nr = bio->bi_iter.bi_sector;
852
853 if (!block_size_is_power_of_two(cache))
854 (void) sector_div(block_nr, cache->sectors_per_block);
855 else
856 block_nr >>= cache->sectors_per_block_shift;
857
858 return to_oblock(block_nr);
859}
860
861static bool accountable_bio(struct cache *cache, struct bio *bio)
862{
863 return bio_op(bio) != REQ_OP_DISCARD;
864}
865
866static void accounted_begin(struct cache *cache, struct bio *bio)
867{
868 struct per_bio_data *pb;
869
870 if (accountable_bio(cache, bio)) {
871 pb = get_per_bio_data(bio);
872 pb->len = bio_sectors(bio);
873 iot_io_begin(&cache->tracker, pb->len);
874 }
875}
876
877static void accounted_complete(struct cache *cache, struct bio *bio)
878{
879 struct per_bio_data *pb = get_per_bio_data(bio);
880
881 iot_io_end(&cache->tracker, pb->len);
882}
883
884static void accounted_request(struct cache *cache, struct bio *bio)
885{
886 accounted_begin(cache, bio);
887 submit_bio_noacct(bio);
888}
889
890static void issue_op(struct bio *bio, void *context)
891{
892 struct cache *cache = context;
893 accounted_request(cache, bio);
894}
895
896/*
897 * When running in writethrough mode we need to send writes to clean blocks
898 * to both the cache and origin devices. Clone the bio and send them in parallel.
899 */
900static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
901 dm_oblock_t oblock, dm_cblock_t cblock)
902{
903 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
904
905 BUG_ON(!origin_bio);
906
907 bio_chain(origin_bio, bio);
908 /*
909 * Passing false to __remap_to_origin_clear_discard() skips
910 * all code that might use per_bio_data (since clone doesn't have it)
911 */
912 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
913 submit_bio(origin_bio);
914
915 remap_to_cache(cache, bio, cblock);
916}
917
918/*----------------------------------------------------------------
919 * Failure modes
920 *--------------------------------------------------------------*/
921static enum cache_metadata_mode get_cache_mode(struct cache *cache)
922{
923 return cache->features.mode;
924}
925
926static const char *cache_device_name(struct cache *cache)
927{
928 return dm_device_name(dm_table_get_md(cache->ti->table));
929}
930
931static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
932{
933 const char *descs[] = {
934 "write",
935 "read-only",
936 "fail"
937 };
938
939 dm_table_event(cache->ti->table);
940 DMINFO("%s: switching cache to %s mode",
941 cache_device_name(cache), descs[(int)mode]);
942}
943
944static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
945{
946 bool needs_check;
947 enum cache_metadata_mode old_mode = get_cache_mode(cache);
948
949 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
950 DMERR("%s: unable to read needs_check flag, setting failure mode.",
951 cache_device_name(cache));
952 new_mode = CM_FAIL;
953 }
954
955 if (new_mode == CM_WRITE && needs_check) {
956 DMERR("%s: unable to switch cache to write mode until repaired.",
957 cache_device_name(cache));
958 if (old_mode != new_mode)
959 new_mode = old_mode;
960 else
961 new_mode = CM_READ_ONLY;
962 }
963
964 /* Never move out of fail mode */
965 if (old_mode == CM_FAIL)
966 new_mode = CM_FAIL;
967
968 switch (new_mode) {
969 case CM_FAIL:
970 case CM_READ_ONLY:
971 dm_cache_metadata_set_read_only(cache->cmd);
972 break;
973
974 case CM_WRITE:
975 dm_cache_metadata_set_read_write(cache->cmd);
976 break;
977 }
978
979 cache->features.mode = new_mode;
980
981 if (new_mode != old_mode)
982 notify_mode_switch(cache, new_mode);
983}
984
985static void abort_transaction(struct cache *cache)
986{
987 const char *dev_name = cache_device_name(cache);
988
989 if (get_cache_mode(cache) >= CM_READ_ONLY)
990 return;
991
992 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
993 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
994 set_cache_mode(cache, CM_FAIL);
995 }
996
997 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
998 if (dm_cache_metadata_abort(cache->cmd)) {
999 DMERR("%s: failed to abort metadata transaction", dev_name);
1000 set_cache_mode(cache, CM_FAIL);
1001 }
1002}
1003
1004static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1005{
1006 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1007 cache_device_name(cache), op, r);
1008 abort_transaction(cache);
1009 set_cache_mode(cache, CM_READ_ONLY);
1010}
1011
1012/*----------------------------------------------------------------*/
1013
1014static void load_stats(struct cache *cache)
1015{
1016 struct dm_cache_statistics stats;
1017
1018 dm_cache_metadata_get_stats(cache->cmd, &stats);
1019 atomic_set(&cache->stats.read_hit, stats.read_hits);
1020 atomic_set(&cache->stats.read_miss, stats.read_misses);
1021 atomic_set(&cache->stats.write_hit, stats.write_hits);
1022 atomic_set(&cache->stats.write_miss, stats.write_misses);
1023}
1024
1025static void save_stats(struct cache *cache)
1026{
1027 struct dm_cache_statistics stats;
1028
1029 if (get_cache_mode(cache) >= CM_READ_ONLY)
1030 return;
1031
1032 stats.read_hits = atomic_read(&cache->stats.read_hit);
1033 stats.read_misses = atomic_read(&cache->stats.read_miss);
1034 stats.write_hits = atomic_read(&cache->stats.write_hit);
1035 stats.write_misses = atomic_read(&cache->stats.write_miss);
1036
1037 dm_cache_metadata_set_stats(cache->cmd, &stats);
1038}
1039
1040static void update_stats(struct cache_stats *stats, enum policy_operation op)
1041{
1042 switch (op) {
1043 case POLICY_PROMOTE:
1044 atomic_inc(&stats->promotion);
1045 break;
1046
1047 case POLICY_DEMOTE:
1048 atomic_inc(&stats->demotion);
1049 break;
1050
1051 case POLICY_WRITEBACK:
1052 atomic_inc(&stats->writeback);
1053 break;
1054 }
1055}
1056
1057/*----------------------------------------------------------------
1058 * Migration processing
1059 *
1060 * Migration covers moving data from the origin device to the cache, or
1061 * vice versa.
1062 *--------------------------------------------------------------*/
1063
1064static void inc_io_migrations(struct cache *cache)
1065{
1066 atomic_inc(&cache->nr_io_migrations);
1067}
1068
1069static void dec_io_migrations(struct cache *cache)
1070{
1071 atomic_dec(&cache->nr_io_migrations);
1072}
1073
1074static bool discard_or_flush(struct bio *bio)
1075{
1076 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1077}
1078
1079static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1080 dm_dblock_t *b, dm_dblock_t *e)
1081{
1082 sector_t sb = bio->bi_iter.bi_sector;
1083 sector_t se = bio_end_sector(bio);
1084
1085 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1086
1087 if (se - sb < cache->discard_block_size)
1088 *e = *b;
1089 else
1090 *e = to_dblock(block_div(se, cache->discard_block_size));
1091}
1092
1093/*----------------------------------------------------------------*/
1094
1095static void prevent_background_work(struct cache *cache)
1096{
1097 lockdep_off();
1098 down_write(&cache->background_work_lock);
1099 lockdep_on();
1100}
1101
1102static void allow_background_work(struct cache *cache)
1103{
1104 lockdep_off();
1105 up_write(&cache->background_work_lock);
1106 lockdep_on();
1107}
1108
1109static bool background_work_begin(struct cache *cache)
1110{
1111 bool r;
1112
1113 lockdep_off();
1114 r = down_read_trylock(&cache->background_work_lock);
1115 lockdep_on();
1116
1117 return r;
1118}
1119
1120static void background_work_end(struct cache *cache)
1121{
1122 lockdep_off();
1123 up_read(&cache->background_work_lock);
1124 lockdep_on();
1125}
1126
1127/*----------------------------------------------------------------*/
1128
1129static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1130{
1131 return (bio_data_dir(bio) == WRITE) &&
1132 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1133}
1134
1135static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1136{
1137 return writeback_mode(cache) &&
1138 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1139}
1140
1141static void quiesce(struct dm_cache_migration *mg,
1142 void (*continuation)(struct work_struct *))
1143{
1144 init_continuation(&mg->k, continuation);
1145 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1146}
1147
1148static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1149{
1150 struct continuation *k = container_of(ws, struct continuation, ws);
1151 return container_of(k, struct dm_cache_migration, k);
1152}
1153
1154static void copy_complete(int read_err, unsigned long write_err, void *context)
1155{
1156 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1157
1158 if (read_err || write_err)
1159 mg->k.input = BLK_STS_IOERR;
1160
1161 queue_continuation(mg->cache->wq, &mg->k);
1162}
1163
1164static void copy(struct dm_cache_migration *mg, bool promote)
1165{
1166 struct dm_io_region o_region, c_region;
1167 struct cache *cache = mg->cache;
1168
1169 o_region.bdev = cache->origin_dev->bdev;
1170 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1171 o_region.count = cache->sectors_per_block;
1172
1173 c_region.bdev = cache->cache_dev->bdev;
1174 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1175 c_region.count = cache->sectors_per_block;
1176
1177 if (promote)
1178 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1179 else
1180 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1181}
1182
1183static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1184{
1185 struct per_bio_data *pb = get_per_bio_data(bio);
1186
1187 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1188 free_prison_cell(cache, pb->cell);
1189 pb->cell = NULL;
1190}
1191
1192static void overwrite_endio(struct bio *bio)
1193{
1194 struct dm_cache_migration *mg = bio->bi_private;
1195 struct cache *cache = mg->cache;
1196 struct per_bio_data *pb = get_per_bio_data(bio);
1197
1198 dm_unhook_bio(&pb->hook_info, bio);
1199
1200 if (bio->bi_status)
1201 mg->k.input = bio->bi_status;
1202
1203 queue_continuation(cache->wq, &mg->k);
1204}
1205
1206static void overwrite(struct dm_cache_migration *mg,
1207 void (*continuation)(struct work_struct *))
1208{
1209 struct bio *bio = mg->overwrite_bio;
1210 struct per_bio_data *pb = get_per_bio_data(bio);
1211
1212 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1213
1214 /*
1215 * The overwrite bio is part of the copy operation, as such it does
1216 * not set/clear discard or dirty flags.
1217 */
1218 if (mg->op->op == POLICY_PROMOTE)
1219 remap_to_cache(mg->cache, bio, mg->op->cblock);
1220 else
1221 remap_to_origin(mg->cache, bio);
1222
1223 init_continuation(&mg->k, continuation);
1224 accounted_request(mg->cache, bio);
1225}
1226
1227/*
1228 * Migration steps:
1229 *
1230 * 1) exclusive lock preventing WRITEs
1231 * 2) quiesce
1232 * 3) copy or issue overwrite bio
1233 * 4) upgrade to exclusive lock preventing READs and WRITEs
1234 * 5) quiesce
1235 * 6) update metadata and commit
1236 * 7) unlock
1237 */
1238static void mg_complete(struct dm_cache_migration *mg, bool success)
1239{
1240 struct bio_list bios;
1241 struct cache *cache = mg->cache;
1242 struct policy_work *op = mg->op;
1243 dm_cblock_t cblock = op->cblock;
1244
1245 if (success)
1246 update_stats(&cache->stats, op->op);
1247
1248 switch (op->op) {
1249 case POLICY_PROMOTE:
1250 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1251 policy_complete_background_work(cache->policy, op, success);
1252
1253 if (mg->overwrite_bio) {
1254 if (success)
1255 force_set_dirty(cache, cblock);
1256 else if (mg->k.input)
1257 mg->overwrite_bio->bi_status = mg->k.input;
1258 else
1259 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1260 bio_endio(mg->overwrite_bio);
1261 } else {
1262 if (success)
1263 force_clear_dirty(cache, cblock);
1264 dec_io_migrations(cache);
1265 }
1266 break;
1267
1268 case POLICY_DEMOTE:
1269 /*
1270 * We clear dirty here to update the nr_dirty counter.
1271 */
1272 if (success)
1273 force_clear_dirty(cache, cblock);
1274 policy_complete_background_work(cache->policy, op, success);
1275 dec_io_migrations(cache);
1276 break;
1277
1278 case POLICY_WRITEBACK:
1279 if (success)
1280 force_clear_dirty(cache, cblock);
1281 policy_complete_background_work(cache->policy, op, success);
1282 dec_io_migrations(cache);
1283 break;
1284 }
1285
1286 bio_list_init(&bios);
1287 if (mg->cell) {
1288 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1289 free_prison_cell(cache, mg->cell);
1290 }
1291
1292 free_migration(mg);
1293 defer_bios(cache, &bios);
1294 wake_migration_worker(cache);
1295
1296 background_work_end(cache);
1297}
1298
1299static void mg_success(struct work_struct *ws)
1300{
1301 struct dm_cache_migration *mg = ws_to_mg(ws);
1302 mg_complete(mg, mg->k.input == 0);
1303}
1304
1305static void mg_update_metadata(struct work_struct *ws)
1306{
1307 int r;
1308 struct dm_cache_migration *mg = ws_to_mg(ws);
1309 struct cache *cache = mg->cache;
1310 struct policy_work *op = mg->op;
1311
1312 switch (op->op) {
1313 case POLICY_PROMOTE:
1314 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1315 if (r) {
1316 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1317 cache_device_name(cache));
1318 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1319
1320 mg_complete(mg, false);
1321 return;
1322 }
1323 mg_complete(mg, true);
1324 break;
1325
1326 case POLICY_DEMOTE:
1327 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1328 if (r) {
1329 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1330 cache_device_name(cache));
1331 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1332
1333 mg_complete(mg, false);
1334 return;
1335 }
1336
1337 /*
1338 * It would be nice if we only had to commit when a REQ_FLUSH
1339 * comes through. But there's one scenario that we have to
1340 * look out for:
1341 *
1342 * - vblock x in a cache block
1343 * - domotion occurs
1344 * - cache block gets reallocated and over written
1345 * - crash
1346 *
1347 * When we recover, because there was no commit the cache will
1348 * rollback to having the data for vblock x in the cache block.
1349 * But the cache block has since been overwritten, so it'll end
1350 * up pointing to data that was never in 'x' during the history
1351 * of the device.
1352 *
1353 * To avoid this issue we require a commit as part of the
1354 * demotion operation.
1355 */
1356 init_continuation(&mg->k, mg_success);
1357 continue_after_commit(&cache->committer, &mg->k);
1358 schedule_commit(&cache->committer);
1359 break;
1360
1361 case POLICY_WRITEBACK:
1362 mg_complete(mg, true);
1363 break;
1364 }
1365}
1366
1367static void mg_update_metadata_after_copy(struct work_struct *ws)
1368{
1369 struct dm_cache_migration *mg = ws_to_mg(ws);
1370
1371 /*
1372 * Did the copy succeed?
1373 */
1374 if (mg->k.input)
1375 mg_complete(mg, false);
1376 else
1377 mg_update_metadata(ws);
1378}
1379
1380static void mg_upgrade_lock(struct work_struct *ws)
1381{
1382 int r;
1383 struct dm_cache_migration *mg = ws_to_mg(ws);
1384
1385 /*
1386 * Did the copy succeed?
1387 */
1388 if (mg->k.input)
1389 mg_complete(mg, false);
1390
1391 else {
1392 /*
1393 * Now we want the lock to prevent both reads and writes.
1394 */
1395 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1396 READ_WRITE_LOCK_LEVEL);
1397 if (r < 0)
1398 mg_complete(mg, false);
1399
1400 else if (r)
1401 quiesce(mg, mg_update_metadata);
1402
1403 else
1404 mg_update_metadata(ws);
1405 }
1406}
1407
1408static void mg_full_copy(struct work_struct *ws)
1409{
1410 struct dm_cache_migration *mg = ws_to_mg(ws);
1411 struct cache *cache = mg->cache;
1412 struct policy_work *op = mg->op;
1413 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1414
1415 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1416 is_discarded_oblock(cache, op->oblock)) {
1417 mg_upgrade_lock(ws);
1418 return;
1419 }
1420
1421 init_continuation(&mg->k, mg_upgrade_lock);
1422 copy(mg, is_policy_promote);
1423}
1424
1425static void mg_copy(struct work_struct *ws)
1426{
1427 struct dm_cache_migration *mg = ws_to_mg(ws);
1428
1429 if (mg->overwrite_bio) {
1430 /*
1431 * No exclusive lock was held when we last checked if the bio
1432 * was optimisable. So we have to check again in case things
1433 * have changed (eg, the block may no longer be discarded).
1434 */
1435 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1436 /*
1437 * Fallback to a real full copy after doing some tidying up.
1438 */
1439 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1440 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1441 mg->overwrite_bio = NULL;
1442 inc_io_migrations(mg->cache);
1443 mg_full_copy(ws);
1444 return;
1445 }
1446
1447 /*
1448 * It's safe to do this here, even though it's new data
1449 * because all IO has been locked out of the block.
1450 *
1451 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1452 * so _not_ using mg_upgrade_lock() as continutation.
1453 */
1454 overwrite(mg, mg_update_metadata_after_copy);
1455
1456 } else
1457 mg_full_copy(ws);
1458}
1459
1460static int mg_lock_writes(struct dm_cache_migration *mg)
1461{
1462 int r;
1463 struct dm_cell_key_v2 key;
1464 struct cache *cache = mg->cache;
1465 struct dm_bio_prison_cell_v2 *prealloc;
1466
1467 prealloc = alloc_prison_cell(cache);
1468
1469 /*
1470 * Prevent writes to the block, but allow reads to continue.
1471 * Unless we're using an overwrite bio, in which case we lock
1472 * everything.
1473 */
1474 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1475 r = dm_cell_lock_v2(cache->prison, &key,
1476 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1477 prealloc, &mg->cell);
1478 if (r < 0) {
1479 free_prison_cell(cache, prealloc);
1480 mg_complete(mg, false);
1481 return r;
1482 }
1483
1484 if (mg->cell != prealloc)
1485 free_prison_cell(cache, prealloc);
1486
1487 if (r == 0)
1488 mg_copy(&mg->k.ws);
1489 else
1490 quiesce(mg, mg_copy);
1491
1492 return 0;
1493}
1494
1495static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1496{
1497 struct dm_cache_migration *mg;
1498
1499 if (!background_work_begin(cache)) {
1500 policy_complete_background_work(cache->policy, op, false);
1501 return -EPERM;
1502 }
1503
1504 mg = alloc_migration(cache);
1505
1506 mg->op = op;
1507 mg->overwrite_bio = bio;
1508
1509 if (!bio)
1510 inc_io_migrations(cache);
1511
1512 return mg_lock_writes(mg);
1513}
1514
1515/*----------------------------------------------------------------
1516 * invalidation processing
1517 *--------------------------------------------------------------*/
1518
1519static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1520{
1521 struct bio_list bios;
1522 struct cache *cache = mg->cache;
1523
1524 bio_list_init(&bios);
1525 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1526 free_prison_cell(cache, mg->cell);
1527
1528 if (!success && mg->overwrite_bio)
1529 bio_io_error(mg->overwrite_bio);
1530
1531 free_migration(mg);
1532 defer_bios(cache, &bios);
1533
1534 background_work_end(cache);
1535}
1536
1537static void invalidate_completed(struct work_struct *ws)
1538{
1539 struct dm_cache_migration *mg = ws_to_mg(ws);
1540 invalidate_complete(mg, !mg->k.input);
1541}
1542
1543static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1544{
1545 int r = policy_invalidate_mapping(cache->policy, cblock);
1546 if (!r) {
1547 r = dm_cache_remove_mapping(cache->cmd, cblock);
1548 if (r) {
1549 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1550 cache_device_name(cache));
1551 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1552 }
1553
1554 } else if (r == -ENODATA) {
1555 /*
1556 * Harmless, already unmapped.
1557 */
1558 r = 0;
1559
1560 } else
1561 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1562
1563 return r;
1564}
1565
1566static void invalidate_remove(struct work_struct *ws)
1567{
1568 int r;
1569 struct dm_cache_migration *mg = ws_to_mg(ws);
1570 struct cache *cache = mg->cache;
1571
1572 r = invalidate_cblock(cache, mg->invalidate_cblock);
1573 if (r) {
1574 invalidate_complete(mg, false);
1575 return;
1576 }
1577
1578 init_continuation(&mg->k, invalidate_completed);
1579 continue_after_commit(&cache->committer, &mg->k);
1580 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1581 mg->overwrite_bio = NULL;
1582 schedule_commit(&cache->committer);
1583}
1584
1585static int invalidate_lock(struct dm_cache_migration *mg)
1586{
1587 int r;
1588 struct dm_cell_key_v2 key;
1589 struct cache *cache = mg->cache;
1590 struct dm_bio_prison_cell_v2 *prealloc;
1591
1592 prealloc = alloc_prison_cell(cache);
1593
1594 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1595 r = dm_cell_lock_v2(cache->prison, &key,
1596 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1597 if (r < 0) {
1598 free_prison_cell(cache, prealloc);
1599 invalidate_complete(mg, false);
1600 return r;
1601 }
1602
1603 if (mg->cell != prealloc)
1604 free_prison_cell(cache, prealloc);
1605
1606 if (r)
1607 quiesce(mg, invalidate_remove);
1608
1609 else {
1610 /*
1611 * We can't call invalidate_remove() directly here because we
1612 * might still be in request context.
1613 */
1614 init_continuation(&mg->k, invalidate_remove);
1615 queue_work(cache->wq, &mg->k.ws);
1616 }
1617
1618 return 0;
1619}
1620
1621static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1622 dm_oblock_t oblock, struct bio *bio)
1623{
1624 struct dm_cache_migration *mg;
1625
1626 if (!background_work_begin(cache))
1627 return -EPERM;
1628
1629 mg = alloc_migration(cache);
1630
1631 mg->overwrite_bio = bio;
1632 mg->invalidate_cblock = cblock;
1633 mg->invalidate_oblock = oblock;
1634
1635 return invalidate_lock(mg);
1636}
1637
1638/*----------------------------------------------------------------
1639 * bio processing
1640 *--------------------------------------------------------------*/
1641
1642enum busy {
1643 IDLE,
1644 BUSY
1645};
1646
1647static enum busy spare_migration_bandwidth(struct cache *cache)
1648{
1649 bool idle = iot_idle_for(&cache->tracker, HZ);
1650 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1651 cache->sectors_per_block;
1652
1653 if (idle && current_volume <= cache->migration_threshold)
1654 return IDLE;
1655 else
1656 return BUSY;
1657}
1658
1659static void inc_hit_counter(struct cache *cache, struct bio *bio)
1660{
1661 atomic_inc(bio_data_dir(bio) == READ ?
1662 &cache->stats.read_hit : &cache->stats.write_hit);
1663}
1664
1665static void inc_miss_counter(struct cache *cache, struct bio *bio)
1666{
1667 atomic_inc(bio_data_dir(bio) == READ ?
1668 &cache->stats.read_miss : &cache->stats.write_miss);
1669}
1670
1671/*----------------------------------------------------------------*/
1672
1673static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1674 bool *commit_needed)
1675{
1676 int r, data_dir;
1677 bool rb, background_queued;
1678 dm_cblock_t cblock;
1679
1680 *commit_needed = false;
1681
1682 rb = bio_detain_shared(cache, block, bio);
1683 if (!rb) {
1684 /*
1685 * An exclusive lock is held for this block, so we have to
1686 * wait. We set the commit_needed flag so the current
1687 * transaction will be committed asap, allowing this lock
1688 * to be dropped.
1689 */
1690 *commit_needed = true;
1691 return DM_MAPIO_SUBMITTED;
1692 }
1693
1694 data_dir = bio_data_dir(bio);
1695
1696 if (optimisable_bio(cache, bio, block)) {
1697 struct policy_work *op = NULL;
1698
1699 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1700 if (unlikely(r && r != -ENOENT)) {
1701 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1702 cache_device_name(cache), r);
1703 bio_io_error(bio);
1704 return DM_MAPIO_SUBMITTED;
1705 }
1706
1707 if (r == -ENOENT && op) {
1708 bio_drop_shared_lock(cache, bio);
1709 BUG_ON(op->op != POLICY_PROMOTE);
1710 mg_start(cache, op, bio);
1711 return DM_MAPIO_SUBMITTED;
1712 }
1713 } else {
1714 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1715 if (unlikely(r && r != -ENOENT)) {
1716 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1717 cache_device_name(cache), r);
1718 bio_io_error(bio);
1719 return DM_MAPIO_SUBMITTED;
1720 }
1721
1722 if (background_queued)
1723 wake_migration_worker(cache);
1724 }
1725
1726 if (r == -ENOENT) {
1727 struct per_bio_data *pb = get_per_bio_data(bio);
1728
1729 /*
1730 * Miss.
1731 */
1732 inc_miss_counter(cache, bio);
1733 if (pb->req_nr == 0) {
1734 accounted_begin(cache, bio);
1735 remap_to_origin_clear_discard(cache, bio, block);
1736 } else {
1737 /*
1738 * This is a duplicate writethrough io that is no
1739 * longer needed because the block has been demoted.
1740 */
1741 bio_endio(bio);
1742 return DM_MAPIO_SUBMITTED;
1743 }
1744 } else {
1745 /*
1746 * Hit.
1747 */
1748 inc_hit_counter(cache, bio);
1749
1750 /*
1751 * Passthrough always maps to the origin, invalidating any
1752 * cache blocks that are written to.
1753 */
1754 if (passthrough_mode(cache)) {
1755 if (bio_data_dir(bio) == WRITE) {
1756 bio_drop_shared_lock(cache, bio);
1757 atomic_inc(&cache->stats.demotion);
1758 invalidate_start(cache, cblock, block, bio);
1759 } else
1760 remap_to_origin_clear_discard(cache, bio, block);
1761 } else {
1762 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1763 !is_dirty(cache, cblock)) {
1764 remap_to_origin_and_cache(cache, bio, block, cblock);
1765 accounted_begin(cache, bio);
1766 } else
1767 remap_to_cache_dirty(cache, bio, block, cblock);
1768 }
1769 }
1770
1771 /*
1772 * dm core turns FUA requests into a separate payload and FLUSH req.
1773 */
1774 if (bio->bi_opf & REQ_FUA) {
1775 /*
1776 * issue_after_commit will call accounted_begin a second time. So
1777 * we call accounted_complete() to avoid double accounting.
1778 */
1779 accounted_complete(cache, bio);
1780 issue_after_commit(&cache->committer, bio);
1781 *commit_needed = true;
1782 return DM_MAPIO_SUBMITTED;
1783 }
1784
1785 return DM_MAPIO_REMAPPED;
1786}
1787
1788static bool process_bio(struct cache *cache, struct bio *bio)
1789{
1790 bool commit_needed;
1791
1792 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1793 submit_bio_noacct(bio);
1794
1795 return commit_needed;
1796}
1797
1798/*
1799 * A non-zero return indicates read_only or fail_io mode.
1800 */
1801static int commit(struct cache *cache, bool clean_shutdown)
1802{
1803 int r;
1804
1805 if (get_cache_mode(cache) >= CM_READ_ONLY)
1806 return -EINVAL;
1807
1808 atomic_inc(&cache->stats.commit_count);
1809 r = dm_cache_commit(cache->cmd, clean_shutdown);
1810 if (r)
1811 metadata_operation_failed(cache, "dm_cache_commit", r);
1812
1813 return r;
1814}
1815
1816/*
1817 * Used by the batcher.
1818 */
1819static blk_status_t commit_op(void *context)
1820{
1821 struct cache *cache = context;
1822
1823 if (dm_cache_changed_this_transaction(cache->cmd))
1824 return errno_to_blk_status(commit(cache, false));
1825
1826 return 0;
1827}
1828
1829/*----------------------------------------------------------------*/
1830
1831static bool process_flush_bio(struct cache *cache, struct bio *bio)
1832{
1833 struct per_bio_data *pb = get_per_bio_data(bio);
1834
1835 if (!pb->req_nr)
1836 remap_to_origin(cache, bio);
1837 else
1838 remap_to_cache(cache, bio, 0);
1839
1840 issue_after_commit(&cache->committer, bio);
1841 return true;
1842}
1843
1844static bool process_discard_bio(struct cache *cache, struct bio *bio)
1845{
1846 dm_dblock_t b, e;
1847
1848 // FIXME: do we need to lock the region? Or can we just assume the
1849 // user wont be so foolish as to issue discard concurrently with
1850 // other IO?
1851 calc_discard_block_range(cache, bio, &b, &e);
1852 while (b != e) {
1853 set_discard(cache, b);
1854 b = to_dblock(from_dblock(b) + 1);
1855 }
1856
1857 if (cache->features.discard_passdown) {
1858 remap_to_origin(cache, bio);
1859 submit_bio_noacct(bio);
1860 } else
1861 bio_endio(bio);
1862
1863 return false;
1864}
1865
1866static void process_deferred_bios(struct work_struct *ws)
1867{
1868 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1869
1870 bool commit_needed = false;
1871 struct bio_list bios;
1872 struct bio *bio;
1873
1874 bio_list_init(&bios);
1875
1876 spin_lock_irq(&cache->lock);
1877 bio_list_merge(&bios, &cache->deferred_bios);
1878 bio_list_init(&cache->deferred_bios);
1879 spin_unlock_irq(&cache->lock);
1880
1881 while ((bio = bio_list_pop(&bios))) {
1882 if (bio->bi_opf & REQ_PREFLUSH)
1883 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1884
1885 else if (bio_op(bio) == REQ_OP_DISCARD)
1886 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1887
1888 else
1889 commit_needed = process_bio(cache, bio) || commit_needed;
1890 }
1891
1892 if (commit_needed)
1893 schedule_commit(&cache->committer);
1894}
1895
1896/*----------------------------------------------------------------
1897 * Main worker loop
1898 *--------------------------------------------------------------*/
1899
1900static void requeue_deferred_bios(struct cache *cache)
1901{
1902 struct bio *bio;
1903 struct bio_list bios;
1904
1905 bio_list_init(&bios);
1906 bio_list_merge(&bios, &cache->deferred_bios);
1907 bio_list_init(&cache->deferred_bios);
1908
1909 while ((bio = bio_list_pop(&bios))) {
1910 bio->bi_status = BLK_STS_DM_REQUEUE;
1911 bio_endio(bio);
1912 }
1913}
1914
1915/*
1916 * We want to commit periodically so that not too much
1917 * unwritten metadata builds up.
1918 */
1919static void do_waker(struct work_struct *ws)
1920{
1921 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1922
1923 policy_tick(cache->policy, true);
1924 wake_migration_worker(cache);
1925 schedule_commit(&cache->committer);
1926 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1927}
1928
1929static void check_migrations(struct work_struct *ws)
1930{
1931 int r;
1932 struct policy_work *op;
1933 struct cache *cache = container_of(ws, struct cache, migration_worker);
1934 enum busy b;
1935
1936 for (;;) {
1937 b = spare_migration_bandwidth(cache);
1938
1939 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1940 if (r == -ENODATA)
1941 break;
1942
1943 if (r) {
1944 DMERR_LIMIT("%s: policy_background_work failed",
1945 cache_device_name(cache));
1946 break;
1947 }
1948
1949 r = mg_start(cache, op, NULL);
1950 if (r)
1951 break;
1952 }
1953}
1954
1955/*----------------------------------------------------------------
1956 * Target methods
1957 *--------------------------------------------------------------*/
1958
1959/*
1960 * This function gets called on the error paths of the constructor, so we
1961 * have to cope with a partially initialised struct.
1962 */
1963static void destroy(struct cache *cache)
1964{
1965 unsigned i;
1966
1967 mempool_exit(&cache->migration_pool);
1968
1969 if (cache->prison)
1970 dm_bio_prison_destroy_v2(cache->prison);
1971
1972 if (cache->wq)
1973 destroy_workqueue(cache->wq);
1974
1975 if (cache->dirty_bitset)
1976 free_bitset(cache->dirty_bitset);
1977
1978 if (cache->discard_bitset)
1979 free_bitset(cache->discard_bitset);
1980
1981 if (cache->copier)
1982 dm_kcopyd_client_destroy(cache->copier);
1983
1984 if (cache->cmd)
1985 dm_cache_metadata_close(cache->cmd);
1986
1987 if (cache->metadata_dev)
1988 dm_put_device(cache->ti, cache->metadata_dev);
1989
1990 if (cache->origin_dev)
1991 dm_put_device(cache->ti, cache->origin_dev);
1992
1993 if (cache->cache_dev)
1994 dm_put_device(cache->ti, cache->cache_dev);
1995
1996 if (cache->policy)
1997 dm_cache_policy_destroy(cache->policy);
1998
1999 for (i = 0; i < cache->nr_ctr_args ; i++)
2000 kfree(cache->ctr_args[i]);
2001 kfree(cache->ctr_args);
2002
2003 bioset_exit(&cache->bs);
2004
2005 kfree(cache);
2006}
2007
2008static void cache_dtr(struct dm_target *ti)
2009{
2010 struct cache *cache = ti->private;
2011
2012 destroy(cache);
2013}
2014
2015static sector_t get_dev_size(struct dm_dev *dev)
2016{
2017 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2018}
2019
2020/*----------------------------------------------------------------*/
2021
2022/*
2023 * Construct a cache device mapping.
2024 *
2025 * cache <metadata dev> <cache dev> <origin dev> <block size>
2026 * <#feature args> [<feature arg>]*
2027 * <policy> <#policy args> [<policy arg>]*
2028 *
2029 * metadata dev : fast device holding the persistent metadata
2030 * cache dev : fast device holding cached data blocks
2031 * origin dev : slow device holding original data blocks
2032 * block size : cache unit size in sectors
2033 *
2034 * #feature args : number of feature arguments passed
2035 * feature args : writethrough. (The default is writeback.)
2036 *
2037 * policy : the replacement policy to use
2038 * #policy args : an even number of policy arguments corresponding
2039 * to key/value pairs passed to the policy
2040 * policy args : key/value pairs passed to the policy
2041 * E.g. 'sequential_threshold 1024'
2042 * See cache-policies.txt for details.
2043 *
2044 * Optional feature arguments are:
2045 * writethrough : write through caching that prohibits cache block
2046 * content from being different from origin block content.
2047 * Without this argument, the default behaviour is to write
2048 * back cache block contents later for performance reasons,
2049 * so they may differ from the corresponding origin blocks.
2050 */
2051struct cache_args {
2052 struct dm_target *ti;
2053
2054 struct dm_dev *metadata_dev;
2055
2056 struct dm_dev *cache_dev;
2057 sector_t cache_sectors;
2058
2059 struct dm_dev *origin_dev;
2060 sector_t origin_sectors;
2061
2062 uint32_t block_size;
2063
2064 const char *policy_name;
2065 int policy_argc;
2066 const char **policy_argv;
2067
2068 struct cache_features features;
2069};
2070
2071static void destroy_cache_args(struct cache_args *ca)
2072{
2073 if (ca->metadata_dev)
2074 dm_put_device(ca->ti, ca->metadata_dev);
2075
2076 if (ca->cache_dev)
2077 dm_put_device(ca->ti, ca->cache_dev);
2078
2079 if (ca->origin_dev)
2080 dm_put_device(ca->ti, ca->origin_dev);
2081
2082 kfree(ca);
2083}
2084
2085static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2086{
2087 if (!as->argc) {
2088 *error = "Insufficient args";
2089 return false;
2090 }
2091
2092 return true;
2093}
2094
2095static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2096 char **error)
2097{
2098 int r;
2099 sector_t metadata_dev_size;
2100 char b[BDEVNAME_SIZE];
2101
2102 if (!at_least_one_arg(as, error))
2103 return -EINVAL;
2104
2105 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2106 &ca->metadata_dev);
2107 if (r) {
2108 *error = "Error opening metadata device";
2109 return r;
2110 }
2111
2112 metadata_dev_size = get_dev_size(ca->metadata_dev);
2113 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2114 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2115 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2116
2117 return 0;
2118}
2119
2120static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2121 char **error)
2122{
2123 int r;
2124
2125 if (!at_least_one_arg(as, error))
2126 return -EINVAL;
2127
2128 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2129 &ca->cache_dev);
2130 if (r) {
2131 *error = "Error opening cache device";
2132 return r;
2133 }
2134 ca->cache_sectors = get_dev_size(ca->cache_dev);
2135
2136 return 0;
2137}
2138
2139static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2140 char **error)
2141{
2142 int r;
2143
2144 if (!at_least_one_arg(as, error))
2145 return -EINVAL;
2146
2147 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2148 &ca->origin_dev);
2149 if (r) {
2150 *error = "Error opening origin device";
2151 return r;
2152 }
2153
2154 ca->origin_sectors = get_dev_size(ca->origin_dev);
2155 if (ca->ti->len > ca->origin_sectors) {
2156 *error = "Device size larger than cached device";
2157 return -EINVAL;
2158 }
2159
2160 return 0;
2161}
2162
2163static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2164 char **error)
2165{
2166 unsigned long block_size;
2167
2168 if (!at_least_one_arg(as, error))
2169 return -EINVAL;
2170
2171 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2172 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2173 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2174 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2175 *error = "Invalid data block size";
2176 return -EINVAL;
2177 }
2178
2179 if (block_size > ca->cache_sectors) {
2180 *error = "Data block size is larger than the cache device";
2181 return -EINVAL;
2182 }
2183
2184 ca->block_size = block_size;
2185
2186 return 0;
2187}
2188
2189static void init_features(struct cache_features *cf)
2190{
2191 cf->mode = CM_WRITE;
2192 cf->io_mode = CM_IO_WRITEBACK;
2193 cf->metadata_version = 1;
2194 cf->discard_passdown = true;
2195}
2196
2197static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2198 char **error)
2199{
2200 static const struct dm_arg _args[] = {
2201 {0, 3, "Invalid number of cache feature arguments"},
2202 };
2203
2204 int r, mode_ctr = 0;
2205 unsigned argc;
2206 const char *arg;
2207 struct cache_features *cf = &ca->features;
2208
2209 init_features(cf);
2210
2211 r = dm_read_arg_group(_args, as, &argc, error);
2212 if (r)
2213 return -EINVAL;
2214
2215 while (argc--) {
2216 arg = dm_shift_arg(as);
2217
2218 if (!strcasecmp(arg, "writeback")) {
2219 cf->io_mode = CM_IO_WRITEBACK;
2220 mode_ctr++;
2221 }
2222
2223 else if (!strcasecmp(arg, "writethrough")) {
2224 cf->io_mode = CM_IO_WRITETHROUGH;
2225 mode_ctr++;
2226 }
2227
2228 else if (!strcasecmp(arg, "passthrough")) {
2229 cf->io_mode = CM_IO_PASSTHROUGH;
2230 mode_ctr++;
2231 }
2232
2233 else if (!strcasecmp(arg, "metadata2"))
2234 cf->metadata_version = 2;
2235
2236 else if (!strcasecmp(arg, "no_discard_passdown"))
2237 cf->discard_passdown = false;
2238
2239 else {
2240 *error = "Unrecognised cache feature requested";
2241 return -EINVAL;
2242 }
2243 }
2244
2245 if (mode_ctr > 1) {
2246 *error = "Duplicate cache io_mode features requested";
2247 return -EINVAL;
2248 }
2249
2250 return 0;
2251}
2252
2253static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2254 char **error)
2255{
2256 static const struct dm_arg _args[] = {
2257 {0, 1024, "Invalid number of policy arguments"},
2258 };
2259
2260 int r;
2261
2262 if (!at_least_one_arg(as, error))
2263 return -EINVAL;
2264
2265 ca->policy_name = dm_shift_arg(as);
2266
2267 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2268 if (r)
2269 return -EINVAL;
2270
2271 ca->policy_argv = (const char **)as->argv;
2272 dm_consume_args(as, ca->policy_argc);
2273
2274 return 0;
2275}
2276
2277static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2278 char **error)
2279{
2280 int r;
2281 struct dm_arg_set as;
2282
2283 as.argc = argc;
2284 as.argv = argv;
2285
2286 r = parse_metadata_dev(ca, &as, error);
2287 if (r)
2288 return r;
2289
2290 r = parse_cache_dev(ca, &as, error);
2291 if (r)
2292 return r;
2293
2294 r = parse_origin_dev(ca, &as, error);
2295 if (r)
2296 return r;
2297
2298 r = parse_block_size(ca, &as, error);
2299 if (r)
2300 return r;
2301
2302 r = parse_features(ca, &as, error);
2303 if (r)
2304 return r;
2305
2306 r = parse_policy(ca, &as, error);
2307 if (r)
2308 return r;
2309
2310 return 0;
2311}
2312
2313/*----------------------------------------------------------------*/
2314
2315static struct kmem_cache *migration_cache;
2316
2317#define NOT_CORE_OPTION 1
2318
2319static int process_config_option(struct cache *cache, const char *key, const char *value)
2320{
2321 unsigned long tmp;
2322
2323 if (!strcasecmp(key, "migration_threshold")) {
2324 if (kstrtoul(value, 10, &tmp))
2325 return -EINVAL;
2326
2327 cache->migration_threshold = tmp;
2328 return 0;
2329 }
2330
2331 return NOT_CORE_OPTION;
2332}
2333
2334static int set_config_value(struct cache *cache, const char *key, const char *value)
2335{
2336 int r = process_config_option(cache, key, value);
2337
2338 if (r == NOT_CORE_OPTION)
2339 r = policy_set_config_value(cache->policy, key, value);
2340
2341 if (r)
2342 DMWARN("bad config value for %s: %s", key, value);
2343
2344 return r;
2345}
2346
2347static int set_config_values(struct cache *cache, int argc, const char **argv)
2348{
2349 int r = 0;
2350
2351 if (argc & 1) {
2352 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2353 return -EINVAL;
2354 }
2355
2356 while (argc) {
2357 r = set_config_value(cache, argv[0], argv[1]);
2358 if (r)
2359 break;
2360
2361 argc -= 2;
2362 argv += 2;
2363 }
2364
2365 return r;
2366}
2367
2368static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2369 char **error)
2370{
2371 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2372 cache->cache_size,
2373 cache->origin_sectors,
2374 cache->sectors_per_block);
2375 if (IS_ERR(p)) {
2376 *error = "Error creating cache's policy";
2377 return PTR_ERR(p);
2378 }
2379 cache->policy = p;
2380 BUG_ON(!cache->policy);
2381
2382 return 0;
2383}
2384
2385/*
2386 * We want the discard block size to be at least the size of the cache
2387 * block size and have no more than 2^14 discard blocks across the origin.
2388 */
2389#define MAX_DISCARD_BLOCKS (1 << 14)
2390
2391static bool too_many_discard_blocks(sector_t discard_block_size,
2392 sector_t origin_size)
2393{
2394 (void) sector_div(origin_size, discard_block_size);
2395
2396 return origin_size > MAX_DISCARD_BLOCKS;
2397}
2398
2399static sector_t calculate_discard_block_size(sector_t cache_block_size,
2400 sector_t origin_size)
2401{
2402 sector_t discard_block_size = cache_block_size;
2403
2404 if (origin_size)
2405 while (too_many_discard_blocks(discard_block_size, origin_size))
2406 discard_block_size *= 2;
2407
2408 return discard_block_size;
2409}
2410
2411static void set_cache_size(struct cache *cache, dm_cblock_t size)
2412{
2413 dm_block_t nr_blocks = from_cblock(size);
2414
2415 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2416 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2417 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2418 "Please consider increasing the cache block size to reduce the overall cache block count.",
2419 (unsigned long long) nr_blocks);
2420
2421 cache->cache_size = size;
2422}
2423
2424#define DEFAULT_MIGRATION_THRESHOLD 2048
2425
2426static int cache_create(struct cache_args *ca, struct cache **result)
2427{
2428 int r = 0;
2429 char **error = &ca->ti->error;
2430 struct cache *cache;
2431 struct dm_target *ti = ca->ti;
2432 dm_block_t origin_blocks;
2433 struct dm_cache_metadata *cmd;
2434 bool may_format = ca->features.mode == CM_WRITE;
2435
2436 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2437 if (!cache)
2438 return -ENOMEM;
2439
2440 cache->ti = ca->ti;
2441 ti->private = cache;
2442 ti->num_flush_bios = 2;
2443 ti->flush_supported = true;
2444
2445 ti->num_discard_bios = 1;
2446 ti->discards_supported = true;
2447
2448 ti->per_io_data_size = sizeof(struct per_bio_data);
2449
2450 cache->features = ca->features;
2451 if (writethrough_mode(cache)) {
2452 /* Create bioset for writethrough bios issued to origin */
2453 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2454 if (r)
2455 goto bad;
2456 }
2457
2458 cache->metadata_dev = ca->metadata_dev;
2459 cache->origin_dev = ca->origin_dev;
2460 cache->cache_dev = ca->cache_dev;
2461
2462 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2463
2464 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2465 origin_blocks = block_div(origin_blocks, ca->block_size);
2466 cache->origin_blocks = to_oblock(origin_blocks);
2467
2468 cache->sectors_per_block = ca->block_size;
2469 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2470 r = -EINVAL;
2471 goto bad;
2472 }
2473
2474 if (ca->block_size & (ca->block_size - 1)) {
2475 dm_block_t cache_size = ca->cache_sectors;
2476
2477 cache->sectors_per_block_shift = -1;
2478 cache_size = block_div(cache_size, ca->block_size);
2479 set_cache_size(cache, to_cblock(cache_size));
2480 } else {
2481 cache->sectors_per_block_shift = __ffs(ca->block_size);
2482 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2483 }
2484
2485 r = create_cache_policy(cache, ca, error);
2486 if (r)
2487 goto bad;
2488
2489 cache->policy_nr_args = ca->policy_argc;
2490 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2491
2492 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2493 if (r) {
2494 *error = "Error setting cache policy's config values";
2495 goto bad;
2496 }
2497
2498 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2499 ca->block_size, may_format,
2500 dm_cache_policy_get_hint_size(cache->policy),
2501 ca->features.metadata_version);
2502 if (IS_ERR(cmd)) {
2503 *error = "Error creating metadata object";
2504 r = PTR_ERR(cmd);
2505 goto bad;
2506 }
2507 cache->cmd = cmd;
2508 set_cache_mode(cache, CM_WRITE);
2509 if (get_cache_mode(cache) != CM_WRITE) {
2510 *error = "Unable to get write access to metadata, please check/repair metadata.";
2511 r = -EINVAL;
2512 goto bad;
2513 }
2514
2515 if (passthrough_mode(cache)) {
2516 bool all_clean;
2517
2518 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2519 if (r) {
2520 *error = "dm_cache_metadata_all_clean() failed";
2521 goto bad;
2522 }
2523
2524 if (!all_clean) {
2525 *error = "Cannot enter passthrough mode unless all blocks are clean";
2526 r = -EINVAL;
2527 goto bad;
2528 }
2529
2530 policy_allow_migrations(cache->policy, false);
2531 }
2532
2533 spin_lock_init(&cache->lock);
2534 bio_list_init(&cache->deferred_bios);
2535 atomic_set(&cache->nr_allocated_migrations, 0);
2536 atomic_set(&cache->nr_io_migrations, 0);
2537 init_waitqueue_head(&cache->migration_wait);
2538
2539 r = -ENOMEM;
2540 atomic_set(&cache->nr_dirty, 0);
2541 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2542 if (!cache->dirty_bitset) {
2543 *error = "could not allocate dirty bitset";
2544 goto bad;
2545 }
2546 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2547
2548 cache->discard_block_size =
2549 calculate_discard_block_size(cache->sectors_per_block,
2550 cache->origin_sectors);
2551 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2552 cache->discard_block_size));
2553 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2554 if (!cache->discard_bitset) {
2555 *error = "could not allocate discard bitset";
2556 goto bad;
2557 }
2558 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2559
2560 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2561 if (IS_ERR(cache->copier)) {
2562 *error = "could not create kcopyd client";
2563 r = PTR_ERR(cache->copier);
2564 goto bad;
2565 }
2566
2567 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2568 if (!cache->wq) {
2569 *error = "could not create workqueue for metadata object";
2570 goto bad;
2571 }
2572 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2573 INIT_WORK(&cache->migration_worker, check_migrations);
2574 INIT_DELAYED_WORK(&cache->waker, do_waker);
2575
2576 cache->prison = dm_bio_prison_create_v2(cache->wq);
2577 if (!cache->prison) {
2578 *error = "could not create bio prison";
2579 goto bad;
2580 }
2581
2582 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2583 migration_cache);
2584 if (r) {
2585 *error = "Error creating cache's migration mempool";
2586 goto bad;
2587 }
2588
2589 cache->need_tick_bio = true;
2590 cache->sized = false;
2591 cache->invalidate = false;
2592 cache->commit_requested = false;
2593 cache->loaded_mappings = false;
2594 cache->loaded_discards = false;
2595
2596 load_stats(cache);
2597
2598 atomic_set(&cache->stats.demotion, 0);
2599 atomic_set(&cache->stats.promotion, 0);
2600 atomic_set(&cache->stats.copies_avoided, 0);
2601 atomic_set(&cache->stats.cache_cell_clash, 0);
2602 atomic_set(&cache->stats.commit_count, 0);
2603 atomic_set(&cache->stats.discard_count, 0);
2604
2605 spin_lock_init(&cache->invalidation_lock);
2606 INIT_LIST_HEAD(&cache->invalidation_requests);
2607
2608 batcher_init(&cache->committer, commit_op, cache,
2609 issue_op, cache, cache->wq);
2610 iot_init(&cache->tracker);
2611
2612 init_rwsem(&cache->background_work_lock);
2613 prevent_background_work(cache);
2614
2615 *result = cache;
2616 return 0;
2617bad:
2618 destroy(cache);
2619 return r;
2620}
2621
2622static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2623{
2624 unsigned i;
2625 const char **copy;
2626
2627 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2628 if (!copy)
2629 return -ENOMEM;
2630 for (i = 0; i < argc; i++) {
2631 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2632 if (!copy[i]) {
2633 while (i--)
2634 kfree(copy[i]);
2635 kfree(copy);
2636 return -ENOMEM;
2637 }
2638 }
2639
2640 cache->nr_ctr_args = argc;
2641 cache->ctr_args = copy;
2642
2643 return 0;
2644}
2645
2646static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2647{
2648 int r = -EINVAL;
2649 struct cache_args *ca;
2650 struct cache *cache = NULL;
2651
2652 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2653 if (!ca) {
2654 ti->error = "Error allocating memory for cache";
2655 return -ENOMEM;
2656 }
2657 ca->ti = ti;
2658
2659 r = parse_cache_args(ca, argc, argv, &ti->error);
2660 if (r)
2661 goto out;
2662
2663 r = cache_create(ca, &cache);
2664 if (r)
2665 goto out;
2666
2667 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2668 if (r) {
2669 destroy(cache);
2670 goto out;
2671 }
2672
2673 ti->private = cache;
2674out:
2675 destroy_cache_args(ca);
2676 return r;
2677}
2678
2679/*----------------------------------------------------------------*/
2680
2681static int cache_map(struct dm_target *ti, struct bio *bio)
2682{
2683 struct cache *cache = ti->private;
2684
2685 int r;
2686 bool commit_needed;
2687 dm_oblock_t block = get_bio_block(cache, bio);
2688
2689 init_per_bio_data(bio);
2690 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2691 /*
2692 * This can only occur if the io goes to a partial block at
2693 * the end of the origin device. We don't cache these.
2694 * Just remap to the origin and carry on.
2695 */
2696 remap_to_origin(cache, bio);
2697 accounted_begin(cache, bio);
2698 return DM_MAPIO_REMAPPED;
2699 }
2700
2701 if (discard_or_flush(bio)) {
2702 defer_bio(cache, bio);
2703 return DM_MAPIO_SUBMITTED;
2704 }
2705
2706 r = map_bio(cache, bio, block, &commit_needed);
2707 if (commit_needed)
2708 schedule_commit(&cache->committer);
2709
2710 return r;
2711}
2712
2713static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2714{
2715 struct cache *cache = ti->private;
2716 unsigned long flags;
2717 struct per_bio_data *pb = get_per_bio_data(bio);
2718
2719 if (pb->tick) {
2720 policy_tick(cache->policy, false);
2721
2722 spin_lock_irqsave(&cache->lock, flags);
2723 cache->need_tick_bio = true;
2724 spin_unlock_irqrestore(&cache->lock, flags);
2725 }
2726
2727 bio_drop_shared_lock(cache, bio);
2728 accounted_complete(cache, bio);
2729
2730 return DM_ENDIO_DONE;
2731}
2732
2733static int write_dirty_bitset(struct cache *cache)
2734{
2735 int r;
2736
2737 if (get_cache_mode(cache) >= CM_READ_ONLY)
2738 return -EINVAL;
2739
2740 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2741 if (r)
2742 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2743
2744 return r;
2745}
2746
2747static int write_discard_bitset(struct cache *cache)
2748{
2749 unsigned i, r;
2750
2751 if (get_cache_mode(cache) >= CM_READ_ONLY)
2752 return -EINVAL;
2753
2754 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2755 cache->discard_nr_blocks);
2756 if (r) {
2757 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2758 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2759 return r;
2760 }
2761
2762 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2763 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2764 is_discarded(cache, to_dblock(i)));
2765 if (r) {
2766 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2767 return r;
2768 }
2769 }
2770
2771 return 0;
2772}
2773
2774static int write_hints(struct cache *cache)
2775{
2776 int r;
2777
2778 if (get_cache_mode(cache) >= CM_READ_ONLY)
2779 return -EINVAL;
2780
2781 r = dm_cache_write_hints(cache->cmd, cache->policy);
2782 if (r) {
2783 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2784 return r;
2785 }
2786
2787 return 0;
2788}
2789
2790/*
2791 * returns true on success
2792 */
2793static bool sync_metadata(struct cache *cache)
2794{
2795 int r1, r2, r3, r4;
2796
2797 r1 = write_dirty_bitset(cache);
2798 if (r1)
2799 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2800
2801 r2 = write_discard_bitset(cache);
2802 if (r2)
2803 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2804
2805 save_stats(cache);
2806
2807 r3 = write_hints(cache);
2808 if (r3)
2809 DMERR("%s: could not write hints", cache_device_name(cache));
2810
2811 /*
2812 * If writing the above metadata failed, we still commit, but don't
2813 * set the clean shutdown flag. This will effectively force every
2814 * dirty bit to be set on reload.
2815 */
2816 r4 = commit(cache, !r1 && !r2 && !r3);
2817 if (r4)
2818 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2819
2820 return !r1 && !r2 && !r3 && !r4;
2821}
2822
2823static void cache_postsuspend(struct dm_target *ti)
2824{
2825 struct cache *cache = ti->private;
2826
2827 prevent_background_work(cache);
2828 BUG_ON(atomic_read(&cache->nr_io_migrations));
2829
2830 cancel_delayed_work_sync(&cache->waker);
2831 drain_workqueue(cache->wq);
2832 WARN_ON(cache->tracker.in_flight);
2833
2834 /*
2835 * If it's a flush suspend there won't be any deferred bios, so this
2836 * call is harmless.
2837 */
2838 requeue_deferred_bios(cache);
2839
2840 if (get_cache_mode(cache) == CM_WRITE)
2841 (void) sync_metadata(cache);
2842}
2843
2844static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2845 bool dirty, uint32_t hint, bool hint_valid)
2846{
2847 int r;
2848 struct cache *cache = context;
2849
2850 if (dirty) {
2851 set_bit(from_cblock(cblock), cache->dirty_bitset);
2852 atomic_inc(&cache->nr_dirty);
2853 } else
2854 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2855
2856 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2857 if (r)
2858 return r;
2859
2860 return 0;
2861}
2862
2863/*
2864 * The discard block size in the on disk metadata is not
2865 * neccessarily the same as we're currently using. So we have to
2866 * be careful to only set the discarded attribute if we know it
2867 * covers a complete block of the new size.
2868 */
2869struct discard_load_info {
2870 struct cache *cache;
2871
2872 /*
2873 * These blocks are sized using the on disk dblock size, rather
2874 * than the current one.
2875 */
2876 dm_block_t block_size;
2877 dm_block_t discard_begin, discard_end;
2878};
2879
2880static void discard_load_info_init(struct cache *cache,
2881 struct discard_load_info *li)
2882{
2883 li->cache = cache;
2884 li->discard_begin = li->discard_end = 0;
2885}
2886
2887static void set_discard_range(struct discard_load_info *li)
2888{
2889 sector_t b, e;
2890
2891 if (li->discard_begin == li->discard_end)
2892 return;
2893
2894 /*
2895 * Convert to sectors.
2896 */
2897 b = li->discard_begin * li->block_size;
2898 e = li->discard_end * li->block_size;
2899
2900 /*
2901 * Then convert back to the current dblock size.
2902 */
2903 b = dm_sector_div_up(b, li->cache->discard_block_size);
2904 sector_div(e, li->cache->discard_block_size);
2905
2906 /*
2907 * The origin may have shrunk, so we need to check we're still in
2908 * bounds.
2909 */
2910 if (e > from_dblock(li->cache->discard_nr_blocks))
2911 e = from_dblock(li->cache->discard_nr_blocks);
2912
2913 for (; b < e; b++)
2914 set_discard(li->cache, to_dblock(b));
2915}
2916
2917static int load_discard(void *context, sector_t discard_block_size,
2918 dm_dblock_t dblock, bool discard)
2919{
2920 struct discard_load_info *li = context;
2921
2922 li->block_size = discard_block_size;
2923
2924 if (discard) {
2925 if (from_dblock(dblock) == li->discard_end)
2926 /*
2927 * We're already in a discard range, just extend it.
2928 */
2929 li->discard_end = li->discard_end + 1ULL;
2930
2931 else {
2932 /*
2933 * Emit the old range and start a new one.
2934 */
2935 set_discard_range(li);
2936 li->discard_begin = from_dblock(dblock);
2937 li->discard_end = li->discard_begin + 1ULL;
2938 }
2939 } else {
2940 set_discard_range(li);
2941 li->discard_begin = li->discard_end = 0;
2942 }
2943
2944 return 0;
2945}
2946
2947static dm_cblock_t get_cache_dev_size(struct cache *cache)
2948{
2949 sector_t size = get_dev_size(cache->cache_dev);
2950 (void) sector_div(size, cache->sectors_per_block);
2951 return to_cblock(size);
2952}
2953
2954static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2955{
2956 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2957 if (cache->sized) {
2958 DMERR("%s: unable to extend cache due to missing cache table reload",
2959 cache_device_name(cache));
2960 return false;
2961 }
2962 }
2963
2964 /*
2965 * We can't drop a dirty block when shrinking the cache.
2966 */
2967 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2968 new_size = to_cblock(from_cblock(new_size) + 1);
2969 if (is_dirty(cache, new_size)) {
2970 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2971 cache_device_name(cache),
2972 (unsigned long long) from_cblock(new_size));
2973 return false;
2974 }
2975 }
2976
2977 return true;
2978}
2979
2980static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2981{
2982 int r;
2983
2984 r = dm_cache_resize(cache->cmd, new_size);
2985 if (r) {
2986 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2987 metadata_operation_failed(cache, "dm_cache_resize", r);
2988 return r;
2989 }
2990
2991 set_cache_size(cache, new_size);
2992
2993 return 0;
2994}
2995
2996static int cache_preresume(struct dm_target *ti)
2997{
2998 int r = 0;
2999 struct cache *cache = ti->private;
3000 dm_cblock_t csize = get_cache_dev_size(cache);
3001
3002 /*
3003 * Check to see if the cache has resized.
3004 */
3005 if (!cache->sized) {
3006 r = resize_cache_dev(cache, csize);
3007 if (r)
3008 return r;
3009
3010 cache->sized = true;
3011
3012 } else if (csize != cache->cache_size) {
3013 if (!can_resize(cache, csize))
3014 return -EINVAL;
3015
3016 r = resize_cache_dev(cache, csize);
3017 if (r)
3018 return r;
3019 }
3020
3021 if (!cache->loaded_mappings) {
3022 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3023 load_mapping, cache);
3024 if (r) {
3025 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3026 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3027 return r;
3028 }
3029
3030 cache->loaded_mappings = true;
3031 }
3032
3033 if (!cache->loaded_discards) {
3034 struct discard_load_info li;
3035
3036 /*
3037 * The discard bitset could have been resized, or the
3038 * discard block size changed. To be safe we start by
3039 * setting every dblock to not discarded.
3040 */
3041 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3042
3043 discard_load_info_init(cache, &li);
3044 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3045 if (r) {
3046 DMERR("%s: could not load origin discards", cache_device_name(cache));
3047 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3048 return r;
3049 }
3050 set_discard_range(&li);
3051
3052 cache->loaded_discards = true;
3053 }
3054
3055 return r;
3056}
3057
3058static void cache_resume(struct dm_target *ti)
3059{
3060 struct cache *cache = ti->private;
3061
3062 cache->need_tick_bio = true;
3063 allow_background_work(cache);
3064 do_waker(&cache->waker.work);
3065}
3066
3067static void emit_flags(struct cache *cache, char *result,
3068 unsigned maxlen, ssize_t *sz_ptr)
3069{
3070 ssize_t sz = *sz_ptr;
3071 struct cache_features *cf = &cache->features;
3072 unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3073
3074 DMEMIT("%u ", count);
3075
3076 if (cf->metadata_version == 2)
3077 DMEMIT("metadata2 ");
3078
3079 if (writethrough_mode(cache))
3080 DMEMIT("writethrough ");
3081
3082 else if (passthrough_mode(cache))
3083 DMEMIT("passthrough ");
3084
3085 else if (writeback_mode(cache))
3086 DMEMIT("writeback ");
3087
3088 else {
3089 DMEMIT("unknown ");
3090 DMERR("%s: internal error: unknown io mode: %d",
3091 cache_device_name(cache), (int) cf->io_mode);
3092 }
3093
3094 if (!cf->discard_passdown)
3095 DMEMIT("no_discard_passdown ");
3096
3097 *sz_ptr = sz;
3098}
3099
3100/*
3101 * Status format:
3102 *
3103 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3104 * <cache block size> <#used cache blocks>/<#total cache blocks>
3105 * <#read hits> <#read misses> <#write hits> <#write misses>
3106 * <#demotions> <#promotions> <#dirty>
3107 * <#features> <features>*
3108 * <#core args> <core args>
3109 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3110 */
3111static void cache_status(struct dm_target *ti, status_type_t type,
3112 unsigned status_flags, char *result, unsigned maxlen)
3113{
3114 int r = 0;
3115 unsigned i;
3116 ssize_t sz = 0;
3117 dm_block_t nr_free_blocks_metadata = 0;
3118 dm_block_t nr_blocks_metadata = 0;
3119 char buf[BDEVNAME_SIZE];
3120 struct cache *cache = ti->private;
3121 dm_cblock_t residency;
3122 bool needs_check;
3123
3124 switch (type) {
3125 case STATUSTYPE_INFO:
3126 if (get_cache_mode(cache) == CM_FAIL) {
3127 DMEMIT("Fail");
3128 break;
3129 }
3130
3131 /* Commit to ensure statistics aren't out-of-date */
3132 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3133 (void) commit(cache, false);
3134
3135 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3136 if (r) {
3137 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3138 cache_device_name(cache), r);
3139 goto err;
3140 }
3141
3142 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3143 if (r) {
3144 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3145 cache_device_name(cache), r);
3146 goto err;
3147 }
3148
3149 residency = policy_residency(cache->policy);
3150
3151 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3152 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3153 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3154 (unsigned long long)nr_blocks_metadata,
3155 (unsigned long long)cache->sectors_per_block,
3156 (unsigned long long) from_cblock(residency),
3157 (unsigned long long) from_cblock(cache->cache_size),
3158 (unsigned) atomic_read(&cache->stats.read_hit),
3159 (unsigned) atomic_read(&cache->stats.read_miss),
3160 (unsigned) atomic_read(&cache->stats.write_hit),
3161 (unsigned) atomic_read(&cache->stats.write_miss),
3162 (unsigned) atomic_read(&cache->stats.demotion),
3163 (unsigned) atomic_read(&cache->stats.promotion),
3164 (unsigned long) atomic_read(&cache->nr_dirty));
3165
3166 emit_flags(cache, result, maxlen, &sz);
3167
3168 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3169
3170 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3171 if (sz < maxlen) {
3172 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3173 if (r)
3174 DMERR("%s: policy_emit_config_values returned %d",
3175 cache_device_name(cache), r);
3176 }
3177
3178 if (get_cache_mode(cache) == CM_READ_ONLY)
3179 DMEMIT("ro ");
3180 else
3181 DMEMIT("rw ");
3182
3183 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3184
3185 if (r || needs_check)
3186 DMEMIT("needs_check ");
3187 else
3188 DMEMIT("- ");
3189
3190 break;
3191
3192 case STATUSTYPE_TABLE:
3193 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3194 DMEMIT("%s ", buf);
3195 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3196 DMEMIT("%s ", buf);
3197 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3198 DMEMIT("%s", buf);
3199
3200 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3201 DMEMIT(" %s", cache->ctr_args[i]);
3202 if (cache->nr_ctr_args)
3203 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3204 }
3205
3206 return;
3207
3208err:
3209 DMEMIT("Error");
3210}
3211
3212/*
3213 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3214 * the one-past-the-end value.
3215 */
3216struct cblock_range {
3217 dm_cblock_t begin;
3218 dm_cblock_t end;
3219};
3220
3221/*
3222 * A cache block range can take two forms:
3223 *
3224 * i) A single cblock, eg. '3456'
3225 * ii) A begin and end cblock with a dash between, eg. 123-234
3226 */
3227static int parse_cblock_range(struct cache *cache, const char *str,
3228 struct cblock_range *result)
3229{
3230 char dummy;
3231 uint64_t b, e;
3232 int r;
3233
3234 /*
3235 * Try and parse form (ii) first.
3236 */
3237 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3238 if (r < 0)
3239 return r;
3240
3241 if (r == 2) {
3242 result->begin = to_cblock(b);
3243 result->end = to_cblock(e);
3244 return 0;
3245 }
3246
3247 /*
3248 * That didn't work, try form (i).
3249 */
3250 r = sscanf(str, "%llu%c", &b, &dummy);
3251 if (r < 0)
3252 return r;
3253
3254 if (r == 1) {
3255 result->begin = to_cblock(b);
3256 result->end = to_cblock(from_cblock(result->begin) + 1u);
3257 return 0;
3258 }
3259
3260 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3261 return -EINVAL;
3262}
3263
3264static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3265{
3266 uint64_t b = from_cblock(range->begin);
3267 uint64_t e = from_cblock(range->end);
3268 uint64_t n = from_cblock(cache->cache_size);
3269
3270 if (b >= n) {
3271 DMERR("%s: begin cblock out of range: %llu >= %llu",
3272 cache_device_name(cache), b, n);
3273 return -EINVAL;
3274 }
3275
3276 if (e > n) {
3277 DMERR("%s: end cblock out of range: %llu > %llu",
3278 cache_device_name(cache), e, n);
3279 return -EINVAL;
3280 }
3281
3282 if (b >= e) {
3283 DMERR("%s: invalid cblock range: %llu >= %llu",
3284 cache_device_name(cache), b, e);
3285 return -EINVAL;
3286 }
3287
3288 return 0;
3289}
3290
3291static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3292{
3293 return to_cblock(from_cblock(b) + 1);
3294}
3295
3296static int request_invalidation(struct cache *cache, struct cblock_range *range)
3297{
3298 int r = 0;
3299
3300 /*
3301 * We don't need to do any locking here because we know we're in
3302 * passthrough mode. There's is potential for a race between an
3303 * invalidation triggered by an io and an invalidation message. This
3304 * is harmless, we must not worry if the policy call fails.
3305 */
3306 while (range->begin != range->end) {
3307 r = invalidate_cblock(cache, range->begin);
3308 if (r)
3309 return r;
3310
3311 range->begin = cblock_succ(range->begin);
3312 }
3313
3314 cache->commit_requested = true;
3315 return r;
3316}
3317
3318static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3319 const char **cblock_ranges)
3320{
3321 int r = 0;
3322 unsigned i;
3323 struct cblock_range range;
3324
3325 if (!passthrough_mode(cache)) {
3326 DMERR("%s: cache has to be in passthrough mode for invalidation",
3327 cache_device_name(cache));
3328 return -EPERM;
3329 }
3330
3331 for (i = 0; i < count; i++) {
3332 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3333 if (r)
3334 break;
3335
3336 r = validate_cblock_range(cache, &range);
3337 if (r)
3338 break;
3339
3340 /*
3341 * Pass begin and end origin blocks to the worker and wake it.
3342 */
3343 r = request_invalidation(cache, &range);
3344 if (r)
3345 break;
3346 }
3347
3348 return r;
3349}
3350
3351/*
3352 * Supports
3353 * "<key> <value>"
3354 * and
3355 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3356 *
3357 * The key migration_threshold is supported by the cache target core.
3358 */
3359static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3360 char *result, unsigned maxlen)
3361{
3362 struct cache *cache = ti->private;
3363
3364 if (!argc)
3365 return -EINVAL;
3366
3367 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3368 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3369 cache_device_name(cache));
3370 return -EOPNOTSUPP;
3371 }
3372
3373 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3374 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3375
3376 if (argc != 2)
3377 return -EINVAL;
3378
3379 return set_config_value(cache, argv[0], argv[1]);
3380}
3381
3382static int cache_iterate_devices(struct dm_target *ti,
3383 iterate_devices_callout_fn fn, void *data)
3384{
3385 int r = 0;
3386 struct cache *cache = ti->private;
3387
3388 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3389 if (!r)
3390 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3391
3392 return r;
3393}
3394
3395static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3396{
3397 struct request_queue *q = bdev_get_queue(origin_bdev);
3398
3399 return q && blk_queue_discard(q);
3400}
3401
3402/*
3403 * If discard_passdown was enabled verify that the origin device
3404 * supports discards. Disable discard_passdown if not.
3405 */
3406static void disable_passdown_if_not_supported(struct cache *cache)
3407{
3408 struct block_device *origin_bdev = cache->origin_dev->bdev;
3409 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3410 const char *reason = NULL;
3411 char buf[BDEVNAME_SIZE];
3412
3413 if (!cache->features.discard_passdown)
3414 return;
3415
3416 if (!origin_dev_supports_discard(origin_bdev))
3417 reason = "discard unsupported";
3418
3419 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3420 reason = "max discard sectors smaller than a block";
3421
3422 if (reason) {
3423 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3424 bdevname(origin_bdev, buf), reason);
3425 cache->features.discard_passdown = false;
3426 }
3427}
3428
3429static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3430{
3431 struct block_device *origin_bdev = cache->origin_dev->bdev;
3432 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3433
3434 if (!cache->features.discard_passdown) {
3435 /* No passdown is done so setting own virtual limits */
3436 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3437 cache->origin_sectors);
3438 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3439 return;
3440 }
3441
3442 /*
3443 * cache_iterate_devices() is stacking both origin and fast device limits
3444 * but discards aren't passed to fast device, so inherit origin's limits.
3445 */
3446 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3447 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3448 limits->discard_granularity = origin_limits->discard_granularity;
3449 limits->discard_alignment = origin_limits->discard_alignment;
3450 limits->discard_misaligned = origin_limits->discard_misaligned;
3451}
3452
3453static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3454{
3455 struct cache *cache = ti->private;
3456 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3457
3458 /*
3459 * If the system-determined stacked limits are compatible with the
3460 * cache's blocksize (io_opt is a factor) do not override them.
3461 */
3462 if (io_opt_sectors < cache->sectors_per_block ||
3463 do_div(io_opt_sectors, cache->sectors_per_block)) {
3464 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3465 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3466 }
3467
3468 disable_passdown_if_not_supported(cache);
3469 set_discard_limits(cache, limits);
3470}
3471
3472/*----------------------------------------------------------------*/
3473
3474static struct target_type cache_target = {
3475 .name = "cache",
3476 .version = {2, 2, 0},
3477 .module = THIS_MODULE,
3478 .ctr = cache_ctr,
3479 .dtr = cache_dtr,
3480 .map = cache_map,
3481 .end_io = cache_end_io,
3482 .postsuspend = cache_postsuspend,
3483 .preresume = cache_preresume,
3484 .resume = cache_resume,
3485 .status = cache_status,
3486 .message = cache_message,
3487 .iterate_devices = cache_iterate_devices,
3488 .io_hints = cache_io_hints,
3489};
3490
3491static int __init dm_cache_init(void)
3492{
3493 int r;
3494
3495 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3496 if (!migration_cache)
3497 return -ENOMEM;
3498
3499 r = dm_register_target(&cache_target);
3500 if (r) {
3501 DMERR("cache target registration failed: %d", r);
3502 kmem_cache_destroy(migration_cache);
3503 return r;
3504 }
3505
3506 return 0;
3507}
3508
3509static void __exit dm_cache_exit(void)
3510{
3511 dm_unregister_target(&cache_target);
3512 kmem_cache_destroy(migration_cache);
3513}
3514
3515module_init(dm_cache_init);
3516module_exit(dm_cache_exit);
3517
3518MODULE_DESCRIPTION(DM_NAME " cache target");
3519MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3520MODULE_LICENSE("GPL");