Loading...
Note: File does not exist in v3.5.6.
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#include "dm-cache-background-tracker.h"
14
15#include <linux/dm-io.h>
16#include <linux/dm-kcopyd.h>
17#include <linux/jiffies.h>
18#include <linux/init.h>
19#include <linux/mempool.h>
20#include <linux/module.h>
21#include <linux/rwsem.h>
22#include <linux/slab.h>
23#include <linux/vmalloc.h>
24
25#define DM_MSG_PREFIX "cache"
26
27DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
28 "A percentage of time allocated for copying to and/or from cache");
29
30/*----------------------------------------------------------------*/
31
32/*
33 * Glossary:
34 *
35 * oblock: index of an origin block
36 * cblock: index of a cache block
37 * promotion: movement of a block from origin to cache
38 * demotion: movement of a block from cache to origin
39 * migration: movement of a block between the origin and cache device,
40 * either direction
41 */
42
43/*----------------------------------------------------------------*/
44
45/*
46 * Represents a chunk of future work. 'input' allows continuations to pass
47 * values between themselves, typically error values.
48 */
49struct continuation {
50 struct work_struct ws;
51 blk_status_t input;
52};
53
54static inline void init_continuation(struct continuation *k,
55 void (*fn)(struct work_struct *))
56{
57 INIT_WORK(&k->ws, fn);
58 k->input = 0;
59}
60
61static inline void queue_continuation(struct workqueue_struct *wq,
62 struct continuation *k)
63{
64 queue_work(wq, &k->ws);
65}
66
67/*----------------------------------------------------------------*/
68
69/*
70 * The batcher collects together pieces of work that need a particular
71 * operation to occur before they can proceed (typically a commit).
72 */
73struct batcher {
74 /*
75 * The operation that everyone is waiting for.
76 */
77 blk_status_t (*commit_op)(void *context);
78 void *commit_context;
79
80 /*
81 * This is how bios should be issued once the commit op is complete
82 * (accounted_request).
83 */
84 void (*issue_op)(struct bio *bio, void *context);
85 void *issue_context;
86
87 /*
88 * Queued work gets put on here after commit.
89 */
90 struct workqueue_struct *wq;
91
92 spinlock_t lock;
93 struct list_head work_items;
94 struct bio_list bios;
95 struct work_struct commit_work;
96
97 bool commit_scheduled;
98};
99
100static void __commit(struct work_struct *_ws)
101{
102 struct batcher *b = container_of(_ws, struct batcher, commit_work);
103 blk_status_t r;
104 struct list_head work_items;
105 struct work_struct *ws, *tmp;
106 struct continuation *k;
107 struct bio *bio;
108 struct bio_list bios;
109
110 INIT_LIST_HEAD(&work_items);
111 bio_list_init(&bios);
112
113 /*
114 * We have to grab these before the commit_op to avoid a race
115 * condition.
116 */
117 spin_lock_irq(&b->lock);
118 list_splice_init(&b->work_items, &work_items);
119 bio_list_merge_init(&bios, &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_init(&cache->deferred_bios, bios);
569 spin_unlock_irq(&cache->lock);
570
571 wake_deferred_bio_worker(cache);
572}
573
574/*----------------------------------------------------------------*/
575
576static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
577{
578 bool r;
579 struct per_bio_data *pb;
580 struct dm_cell_key_v2 key;
581 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
582 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
583
584 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
585
586 build_key(oblock, end, &key);
587 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
588 if (!r) {
589 /*
590 * Failed to get the lock.
591 */
592 free_prison_cell(cache, cell_prealloc);
593 return r;
594 }
595
596 if (cell != cell_prealloc)
597 free_prison_cell(cache, cell_prealloc);
598
599 pb = get_per_bio_data(bio);
600 pb->cell = cell;
601
602 return r;
603}
604
605/*----------------------------------------------------------------*/
606
607static bool is_dirty(struct cache *cache, dm_cblock_t b)
608{
609 return test_bit(from_cblock(b), cache->dirty_bitset);
610}
611
612static void set_dirty(struct cache *cache, dm_cblock_t cblock)
613{
614 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
615 atomic_inc(&cache->nr_dirty);
616 policy_set_dirty(cache->policy, cblock);
617 }
618}
619
620/*
621 * These two are called when setting after migrations to force the policy
622 * and dirty bitset to be in sync.
623 */
624static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
625{
626 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
627 atomic_inc(&cache->nr_dirty);
628 policy_set_dirty(cache->policy, cblock);
629}
630
631static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
632{
633 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
634 if (atomic_dec_return(&cache->nr_dirty) == 0)
635 dm_table_event(cache->ti->table);
636 }
637
638 policy_clear_dirty(cache->policy, cblock);
639}
640
641/*----------------------------------------------------------------*/
642
643static bool block_size_is_power_of_two(struct cache *cache)
644{
645 return cache->sectors_per_block_shift >= 0;
646}
647
648static dm_block_t block_div(dm_block_t b, uint32_t n)
649{
650 do_div(b, n);
651
652 return b;
653}
654
655static dm_block_t oblocks_per_dblock(struct cache *cache)
656{
657 dm_block_t oblocks = cache->discard_block_size;
658
659 if (block_size_is_power_of_two(cache))
660 oblocks >>= cache->sectors_per_block_shift;
661 else
662 oblocks = block_div(oblocks, cache->sectors_per_block);
663
664 return oblocks;
665}
666
667static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
668{
669 return to_dblock(block_div(from_oblock(oblock),
670 oblocks_per_dblock(cache)));
671}
672
673static void set_discard(struct cache *cache, dm_dblock_t b)
674{
675 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
676 atomic_inc(&cache->stats.discard_count);
677
678 spin_lock_irq(&cache->lock);
679 set_bit(from_dblock(b), cache->discard_bitset);
680 spin_unlock_irq(&cache->lock);
681}
682
683static void clear_discard(struct cache *cache, dm_dblock_t b)
684{
685 spin_lock_irq(&cache->lock);
686 clear_bit(from_dblock(b), cache->discard_bitset);
687 spin_unlock_irq(&cache->lock);
688}
689
690static bool is_discarded(struct cache *cache, dm_dblock_t b)
691{
692 int r;
693
694 spin_lock_irq(&cache->lock);
695 r = test_bit(from_dblock(b), cache->discard_bitset);
696 spin_unlock_irq(&cache->lock);
697
698 return r;
699}
700
701static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
702{
703 int r;
704
705 spin_lock_irq(&cache->lock);
706 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
707 cache->discard_bitset);
708 spin_unlock_irq(&cache->lock);
709
710 return r;
711}
712
713/*
714 * -------------------------------------------------------------
715 * Remapping
716 *--------------------------------------------------------------
717 */
718static void remap_to_origin(struct cache *cache, struct bio *bio)
719{
720 bio_set_dev(bio, cache->origin_dev->bdev);
721}
722
723static void remap_to_cache(struct cache *cache, struct bio *bio,
724 dm_cblock_t cblock)
725{
726 sector_t bi_sector = bio->bi_iter.bi_sector;
727 sector_t block = from_cblock(cblock);
728
729 bio_set_dev(bio, cache->cache_dev->bdev);
730 if (!block_size_is_power_of_two(cache))
731 bio->bi_iter.bi_sector =
732 (block * cache->sectors_per_block) +
733 sector_div(bi_sector, cache->sectors_per_block);
734 else
735 bio->bi_iter.bi_sector =
736 (block << cache->sectors_per_block_shift) |
737 (bi_sector & (cache->sectors_per_block - 1));
738}
739
740static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
741{
742 struct per_bio_data *pb;
743
744 spin_lock_irq(&cache->lock);
745 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
746 bio_op(bio) != REQ_OP_DISCARD) {
747 pb = get_per_bio_data(bio);
748 pb->tick = true;
749 cache->need_tick_bio = false;
750 }
751 spin_unlock_irq(&cache->lock);
752}
753
754static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
755 dm_oblock_t oblock)
756{
757 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
758 check_if_tick_bio_needed(cache, bio);
759 remap_to_origin(cache, bio);
760 if (bio_data_dir(bio) == WRITE)
761 clear_discard(cache, oblock_to_dblock(cache, oblock));
762}
763
764static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
765 dm_oblock_t oblock, dm_cblock_t cblock)
766{
767 check_if_tick_bio_needed(cache, bio);
768 remap_to_cache(cache, bio, cblock);
769 if (bio_data_dir(bio) == WRITE) {
770 set_dirty(cache, cblock);
771 clear_discard(cache, oblock_to_dblock(cache, oblock));
772 }
773}
774
775static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
776{
777 sector_t block_nr = bio->bi_iter.bi_sector;
778
779 if (!block_size_is_power_of_two(cache))
780 (void) sector_div(block_nr, cache->sectors_per_block);
781 else
782 block_nr >>= cache->sectors_per_block_shift;
783
784 return to_oblock(block_nr);
785}
786
787static bool accountable_bio(struct cache *cache, struct bio *bio)
788{
789 return bio_op(bio) != REQ_OP_DISCARD;
790}
791
792static void accounted_begin(struct cache *cache, struct bio *bio)
793{
794 struct per_bio_data *pb;
795
796 if (accountable_bio(cache, bio)) {
797 pb = get_per_bio_data(bio);
798 pb->len = bio_sectors(bio);
799 dm_iot_io_begin(&cache->tracker, pb->len);
800 }
801}
802
803static void accounted_complete(struct cache *cache, struct bio *bio)
804{
805 struct per_bio_data *pb = get_per_bio_data(bio);
806
807 dm_iot_io_end(&cache->tracker, pb->len);
808}
809
810static void accounted_request(struct cache *cache, struct bio *bio)
811{
812 accounted_begin(cache, bio);
813 dm_submit_bio_remap(bio, NULL);
814}
815
816static void issue_op(struct bio *bio, void *context)
817{
818 struct cache *cache = context;
819
820 accounted_request(cache, bio);
821}
822
823/*
824 * When running in writethrough mode we need to send writes to clean blocks
825 * to both the cache and origin devices. Clone the bio and send them in parallel.
826 */
827static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
828 dm_oblock_t oblock, dm_cblock_t cblock)
829{
830 struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
831 GFP_NOIO, &cache->bs);
832
833 BUG_ON(!origin_bio);
834
835 bio_chain(origin_bio, bio);
836
837 if (bio_data_dir(origin_bio) == WRITE)
838 clear_discard(cache, oblock_to_dblock(cache, oblock));
839 submit_bio(origin_bio);
840
841 remap_to_cache(cache, bio, cblock);
842}
843
844/*
845 *--------------------------------------------------------------
846 * Failure modes
847 *--------------------------------------------------------------
848 */
849static enum cache_metadata_mode get_cache_mode(struct cache *cache)
850{
851 return cache->features.mode;
852}
853
854static const char *cache_device_name(struct cache *cache)
855{
856 return dm_table_device_name(cache->ti->table);
857}
858
859static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
860{
861 static const char *descs[] = {
862 "write",
863 "read-only",
864 "fail"
865 };
866
867 dm_table_event(cache->ti->table);
868 DMINFO("%s: switching cache to %s mode",
869 cache_device_name(cache), descs[(int)mode]);
870}
871
872static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
873{
874 bool needs_check;
875 enum cache_metadata_mode old_mode = get_cache_mode(cache);
876
877 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
878 DMERR("%s: unable to read needs_check flag, setting failure mode.",
879 cache_device_name(cache));
880 new_mode = CM_FAIL;
881 }
882
883 if (new_mode == CM_WRITE && needs_check) {
884 DMERR("%s: unable to switch cache to write mode until repaired.",
885 cache_device_name(cache));
886 if (old_mode != new_mode)
887 new_mode = old_mode;
888 else
889 new_mode = CM_READ_ONLY;
890 }
891
892 /* Never move out of fail mode */
893 if (old_mode == CM_FAIL)
894 new_mode = CM_FAIL;
895
896 switch (new_mode) {
897 case CM_FAIL:
898 case CM_READ_ONLY:
899 dm_cache_metadata_set_read_only(cache->cmd);
900 break;
901
902 case CM_WRITE:
903 dm_cache_metadata_set_read_write(cache->cmd);
904 break;
905 }
906
907 cache->features.mode = new_mode;
908
909 if (new_mode != old_mode)
910 notify_mode_switch(cache, new_mode);
911}
912
913static void abort_transaction(struct cache *cache)
914{
915 const char *dev_name = cache_device_name(cache);
916
917 if (get_cache_mode(cache) >= CM_READ_ONLY)
918 return;
919
920 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
921 if (dm_cache_metadata_abort(cache->cmd)) {
922 DMERR("%s: failed to abort metadata transaction", dev_name);
923 set_cache_mode(cache, CM_FAIL);
924 }
925
926 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
927 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
928 set_cache_mode(cache, CM_FAIL);
929 }
930}
931
932static void metadata_operation_failed(struct cache *cache, const char *op, int r)
933{
934 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
935 cache_device_name(cache), op, r);
936 abort_transaction(cache);
937 set_cache_mode(cache, CM_READ_ONLY);
938}
939
940/*----------------------------------------------------------------*/
941
942static void load_stats(struct cache *cache)
943{
944 struct dm_cache_statistics stats;
945
946 dm_cache_metadata_get_stats(cache->cmd, &stats);
947 atomic_set(&cache->stats.read_hit, stats.read_hits);
948 atomic_set(&cache->stats.read_miss, stats.read_misses);
949 atomic_set(&cache->stats.write_hit, stats.write_hits);
950 atomic_set(&cache->stats.write_miss, stats.write_misses);
951}
952
953static void save_stats(struct cache *cache)
954{
955 struct dm_cache_statistics stats;
956
957 if (get_cache_mode(cache) >= CM_READ_ONLY)
958 return;
959
960 stats.read_hits = atomic_read(&cache->stats.read_hit);
961 stats.read_misses = atomic_read(&cache->stats.read_miss);
962 stats.write_hits = atomic_read(&cache->stats.write_hit);
963 stats.write_misses = atomic_read(&cache->stats.write_miss);
964
965 dm_cache_metadata_set_stats(cache->cmd, &stats);
966}
967
968static void update_stats(struct cache_stats *stats, enum policy_operation op)
969{
970 switch (op) {
971 case POLICY_PROMOTE:
972 atomic_inc(&stats->promotion);
973 break;
974
975 case POLICY_DEMOTE:
976 atomic_inc(&stats->demotion);
977 break;
978
979 case POLICY_WRITEBACK:
980 atomic_inc(&stats->writeback);
981 break;
982 }
983}
984
985/*
986 *---------------------------------------------------------------------
987 * Migration processing
988 *
989 * Migration covers moving data from the origin device to the cache, or
990 * vice versa.
991 *---------------------------------------------------------------------
992 */
993static void inc_io_migrations(struct cache *cache)
994{
995 atomic_inc(&cache->nr_io_migrations);
996}
997
998static void dec_io_migrations(struct cache *cache)
999{
1000 atomic_dec(&cache->nr_io_migrations);
1001}
1002
1003static bool discard_or_flush(struct bio *bio)
1004{
1005 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1006}
1007
1008static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1009 dm_dblock_t *b, dm_dblock_t *e)
1010{
1011 sector_t sb = bio->bi_iter.bi_sector;
1012 sector_t se = bio_end_sector(bio);
1013
1014 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1015
1016 if (se - sb < cache->discard_block_size)
1017 *e = *b;
1018 else
1019 *e = to_dblock(block_div(se, cache->discard_block_size));
1020}
1021
1022/*----------------------------------------------------------------*/
1023
1024static void prevent_background_work(struct cache *cache)
1025{
1026 lockdep_off();
1027 down_write(&cache->background_work_lock);
1028 lockdep_on();
1029}
1030
1031static void allow_background_work(struct cache *cache)
1032{
1033 lockdep_off();
1034 up_write(&cache->background_work_lock);
1035 lockdep_on();
1036}
1037
1038static bool background_work_begin(struct cache *cache)
1039{
1040 bool r;
1041
1042 lockdep_off();
1043 r = down_read_trylock(&cache->background_work_lock);
1044 lockdep_on();
1045
1046 return r;
1047}
1048
1049static void background_work_end(struct cache *cache)
1050{
1051 lockdep_off();
1052 up_read(&cache->background_work_lock);
1053 lockdep_on();
1054}
1055
1056/*----------------------------------------------------------------*/
1057
1058static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1059{
1060 return (bio_data_dir(bio) == WRITE) &&
1061 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1062}
1063
1064static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1065{
1066 return writeback_mode(cache) &&
1067 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1068}
1069
1070static void quiesce(struct dm_cache_migration *mg,
1071 void (*continuation)(struct work_struct *))
1072{
1073 init_continuation(&mg->k, continuation);
1074 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1075}
1076
1077static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1078{
1079 struct continuation *k = container_of(ws, struct continuation, ws);
1080
1081 return container_of(k, struct dm_cache_migration, k);
1082}
1083
1084static void copy_complete(int read_err, unsigned long write_err, void *context)
1085{
1086 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1087
1088 if (read_err || write_err)
1089 mg->k.input = BLK_STS_IOERR;
1090
1091 queue_continuation(mg->cache->wq, &mg->k);
1092}
1093
1094static void copy(struct dm_cache_migration *mg, bool promote)
1095{
1096 struct dm_io_region o_region, c_region;
1097 struct cache *cache = mg->cache;
1098
1099 o_region.bdev = cache->origin_dev->bdev;
1100 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1101 o_region.count = cache->sectors_per_block;
1102
1103 c_region.bdev = cache->cache_dev->bdev;
1104 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1105 c_region.count = cache->sectors_per_block;
1106
1107 if (promote)
1108 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1109 else
1110 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1111}
1112
1113static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1114{
1115 struct per_bio_data *pb = get_per_bio_data(bio);
1116
1117 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1118 free_prison_cell(cache, pb->cell);
1119 pb->cell = NULL;
1120}
1121
1122static void overwrite_endio(struct bio *bio)
1123{
1124 struct dm_cache_migration *mg = bio->bi_private;
1125 struct cache *cache = mg->cache;
1126 struct per_bio_data *pb = get_per_bio_data(bio);
1127
1128 dm_unhook_bio(&pb->hook_info, bio);
1129
1130 if (bio->bi_status)
1131 mg->k.input = bio->bi_status;
1132
1133 queue_continuation(cache->wq, &mg->k);
1134}
1135
1136static void overwrite(struct dm_cache_migration *mg,
1137 void (*continuation)(struct work_struct *))
1138{
1139 struct bio *bio = mg->overwrite_bio;
1140 struct per_bio_data *pb = get_per_bio_data(bio);
1141
1142 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1143
1144 /*
1145 * The overwrite bio is part of the copy operation, as such it does
1146 * not set/clear discard or dirty flags.
1147 */
1148 if (mg->op->op == POLICY_PROMOTE)
1149 remap_to_cache(mg->cache, bio, mg->op->cblock);
1150 else
1151 remap_to_origin(mg->cache, bio);
1152
1153 init_continuation(&mg->k, continuation);
1154 accounted_request(mg->cache, bio);
1155}
1156
1157/*
1158 * Migration steps:
1159 *
1160 * 1) exclusive lock preventing WRITEs
1161 * 2) quiesce
1162 * 3) copy or issue overwrite bio
1163 * 4) upgrade to exclusive lock preventing READs and WRITEs
1164 * 5) quiesce
1165 * 6) update metadata and commit
1166 * 7) unlock
1167 */
1168static void mg_complete(struct dm_cache_migration *mg, bool success)
1169{
1170 struct bio_list bios;
1171 struct cache *cache = mg->cache;
1172 struct policy_work *op = mg->op;
1173 dm_cblock_t cblock = op->cblock;
1174
1175 if (success)
1176 update_stats(&cache->stats, op->op);
1177
1178 switch (op->op) {
1179 case POLICY_PROMOTE:
1180 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1181 policy_complete_background_work(cache->policy, op, success);
1182
1183 if (mg->overwrite_bio) {
1184 if (success)
1185 force_set_dirty(cache, cblock);
1186 else if (mg->k.input)
1187 mg->overwrite_bio->bi_status = mg->k.input;
1188 else
1189 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1190 bio_endio(mg->overwrite_bio);
1191 } else {
1192 if (success)
1193 force_clear_dirty(cache, cblock);
1194 dec_io_migrations(cache);
1195 }
1196 break;
1197
1198 case POLICY_DEMOTE:
1199 /*
1200 * We clear dirty here to update the nr_dirty counter.
1201 */
1202 if (success)
1203 force_clear_dirty(cache, cblock);
1204 policy_complete_background_work(cache->policy, op, success);
1205 dec_io_migrations(cache);
1206 break;
1207
1208 case POLICY_WRITEBACK:
1209 if (success)
1210 force_clear_dirty(cache, cblock);
1211 policy_complete_background_work(cache->policy, op, success);
1212 dec_io_migrations(cache);
1213 break;
1214 }
1215
1216 bio_list_init(&bios);
1217 if (mg->cell) {
1218 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1219 free_prison_cell(cache, mg->cell);
1220 }
1221
1222 free_migration(mg);
1223 defer_bios(cache, &bios);
1224 wake_migration_worker(cache);
1225
1226 background_work_end(cache);
1227}
1228
1229static void mg_success(struct work_struct *ws)
1230{
1231 struct dm_cache_migration *mg = ws_to_mg(ws);
1232
1233 mg_complete(mg, mg->k.input == 0);
1234}
1235
1236static void mg_update_metadata(struct work_struct *ws)
1237{
1238 int r;
1239 struct dm_cache_migration *mg = ws_to_mg(ws);
1240 struct cache *cache = mg->cache;
1241 struct policy_work *op = mg->op;
1242
1243 switch (op->op) {
1244 case POLICY_PROMOTE:
1245 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1246 if (r) {
1247 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1248 cache_device_name(cache));
1249 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1250
1251 mg_complete(mg, false);
1252 return;
1253 }
1254 mg_complete(mg, true);
1255 break;
1256
1257 case POLICY_DEMOTE:
1258 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1259 if (r) {
1260 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1261 cache_device_name(cache));
1262 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1263
1264 mg_complete(mg, false);
1265 return;
1266 }
1267
1268 /*
1269 * It would be nice if we only had to commit when a REQ_FLUSH
1270 * comes through. But there's one scenario that we have to
1271 * look out for:
1272 *
1273 * - vblock x in a cache block
1274 * - domotion occurs
1275 * - cache block gets reallocated and over written
1276 * - crash
1277 *
1278 * When we recover, because there was no commit the cache will
1279 * rollback to having the data for vblock x in the cache block.
1280 * But the cache block has since been overwritten, so it'll end
1281 * up pointing to data that was never in 'x' during the history
1282 * of the device.
1283 *
1284 * To avoid this issue we require a commit as part of the
1285 * demotion operation.
1286 */
1287 init_continuation(&mg->k, mg_success);
1288 continue_after_commit(&cache->committer, &mg->k);
1289 schedule_commit(&cache->committer);
1290 break;
1291
1292 case POLICY_WRITEBACK:
1293 mg_complete(mg, true);
1294 break;
1295 }
1296}
1297
1298static void mg_update_metadata_after_copy(struct work_struct *ws)
1299{
1300 struct dm_cache_migration *mg = ws_to_mg(ws);
1301
1302 /*
1303 * Did the copy succeed?
1304 */
1305 if (mg->k.input)
1306 mg_complete(mg, false);
1307 else
1308 mg_update_metadata(ws);
1309}
1310
1311static void mg_upgrade_lock(struct work_struct *ws)
1312{
1313 int r;
1314 struct dm_cache_migration *mg = ws_to_mg(ws);
1315
1316 /*
1317 * Did the copy succeed?
1318 */
1319 if (mg->k.input)
1320 mg_complete(mg, false);
1321
1322 else {
1323 /*
1324 * Now we want the lock to prevent both reads and writes.
1325 */
1326 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1327 READ_WRITE_LOCK_LEVEL);
1328 if (r < 0)
1329 mg_complete(mg, false);
1330
1331 else if (r)
1332 quiesce(mg, mg_update_metadata);
1333
1334 else
1335 mg_update_metadata(ws);
1336 }
1337}
1338
1339static void mg_full_copy(struct work_struct *ws)
1340{
1341 struct dm_cache_migration *mg = ws_to_mg(ws);
1342 struct cache *cache = mg->cache;
1343 struct policy_work *op = mg->op;
1344 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1345
1346 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1347 is_discarded_oblock(cache, op->oblock)) {
1348 mg_upgrade_lock(ws);
1349 return;
1350 }
1351
1352 init_continuation(&mg->k, mg_upgrade_lock);
1353 copy(mg, is_policy_promote);
1354}
1355
1356static void mg_copy(struct work_struct *ws)
1357{
1358 struct dm_cache_migration *mg = ws_to_mg(ws);
1359
1360 if (mg->overwrite_bio) {
1361 /*
1362 * No exclusive lock was held when we last checked if the bio
1363 * was optimisable. So we have to check again in case things
1364 * have changed (eg, the block may no longer be discarded).
1365 */
1366 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1367 /*
1368 * Fallback to a real full copy after doing some tidying up.
1369 */
1370 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1371
1372 BUG_ON(rb); /* An exclusive lock must _not_ be held for this block */
1373 mg->overwrite_bio = NULL;
1374 inc_io_migrations(mg->cache);
1375 mg_full_copy(ws);
1376 return;
1377 }
1378
1379 /*
1380 * It's safe to do this here, even though it's new data
1381 * because all IO has been locked out of the block.
1382 *
1383 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1384 * so _not_ using mg_upgrade_lock() as continutation.
1385 */
1386 overwrite(mg, mg_update_metadata_after_copy);
1387
1388 } else
1389 mg_full_copy(ws);
1390}
1391
1392static int mg_lock_writes(struct dm_cache_migration *mg)
1393{
1394 int r;
1395 struct dm_cell_key_v2 key;
1396 struct cache *cache = mg->cache;
1397 struct dm_bio_prison_cell_v2 *prealloc;
1398
1399 prealloc = alloc_prison_cell(cache);
1400
1401 /*
1402 * Prevent writes to the block, but allow reads to continue.
1403 * Unless we're using an overwrite bio, in which case we lock
1404 * everything.
1405 */
1406 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1407 r = dm_cell_lock_v2(cache->prison, &key,
1408 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1409 prealloc, &mg->cell);
1410 if (r < 0) {
1411 free_prison_cell(cache, prealloc);
1412 mg_complete(mg, false);
1413 return r;
1414 }
1415
1416 if (mg->cell != prealloc)
1417 free_prison_cell(cache, prealloc);
1418
1419 if (r == 0)
1420 mg_copy(&mg->k.ws);
1421 else
1422 quiesce(mg, mg_copy);
1423
1424 return 0;
1425}
1426
1427static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1428{
1429 struct dm_cache_migration *mg;
1430
1431 if (!background_work_begin(cache)) {
1432 policy_complete_background_work(cache->policy, op, false);
1433 return -EPERM;
1434 }
1435
1436 mg = alloc_migration(cache);
1437
1438 mg->op = op;
1439 mg->overwrite_bio = bio;
1440
1441 if (!bio)
1442 inc_io_migrations(cache);
1443
1444 return mg_lock_writes(mg);
1445}
1446
1447/*
1448 *--------------------------------------------------------------
1449 * invalidation processing
1450 *--------------------------------------------------------------
1451 */
1452
1453static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1454{
1455 struct bio_list bios;
1456 struct cache *cache = mg->cache;
1457
1458 bio_list_init(&bios);
1459 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1460 free_prison_cell(cache, mg->cell);
1461
1462 if (!success && mg->overwrite_bio)
1463 bio_io_error(mg->overwrite_bio);
1464
1465 free_migration(mg);
1466 defer_bios(cache, &bios);
1467
1468 background_work_end(cache);
1469}
1470
1471static void invalidate_completed(struct work_struct *ws)
1472{
1473 struct dm_cache_migration *mg = ws_to_mg(ws);
1474
1475 invalidate_complete(mg, !mg->k.input);
1476}
1477
1478static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1479{
1480 int r;
1481
1482 r = policy_invalidate_mapping(cache->policy, cblock);
1483 if (!r) {
1484 r = dm_cache_remove_mapping(cache->cmd, cblock);
1485 if (r) {
1486 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1487 cache_device_name(cache));
1488 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1489 }
1490
1491 } else if (r == -ENODATA) {
1492 /*
1493 * Harmless, already unmapped.
1494 */
1495 r = 0;
1496
1497 } else
1498 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1499
1500 return r;
1501}
1502
1503static void invalidate_remove(struct work_struct *ws)
1504{
1505 int r;
1506 struct dm_cache_migration *mg = ws_to_mg(ws);
1507 struct cache *cache = mg->cache;
1508
1509 r = invalidate_cblock(cache, mg->invalidate_cblock);
1510 if (r) {
1511 invalidate_complete(mg, false);
1512 return;
1513 }
1514
1515 init_continuation(&mg->k, invalidate_completed);
1516 continue_after_commit(&cache->committer, &mg->k);
1517 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1518 mg->overwrite_bio = NULL;
1519 schedule_commit(&cache->committer);
1520}
1521
1522static int invalidate_lock(struct dm_cache_migration *mg)
1523{
1524 int r;
1525 struct dm_cell_key_v2 key;
1526 struct cache *cache = mg->cache;
1527 struct dm_bio_prison_cell_v2 *prealloc;
1528
1529 prealloc = alloc_prison_cell(cache);
1530
1531 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1532 r = dm_cell_lock_v2(cache->prison, &key,
1533 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1534 if (r < 0) {
1535 free_prison_cell(cache, prealloc);
1536 invalidate_complete(mg, false);
1537 return r;
1538 }
1539
1540 if (mg->cell != prealloc)
1541 free_prison_cell(cache, prealloc);
1542
1543 if (r)
1544 quiesce(mg, invalidate_remove);
1545
1546 else {
1547 /*
1548 * We can't call invalidate_remove() directly here because we
1549 * might still be in request context.
1550 */
1551 init_continuation(&mg->k, invalidate_remove);
1552 queue_work(cache->wq, &mg->k.ws);
1553 }
1554
1555 return 0;
1556}
1557
1558static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1559 dm_oblock_t oblock, struct bio *bio)
1560{
1561 struct dm_cache_migration *mg;
1562
1563 if (!background_work_begin(cache))
1564 return -EPERM;
1565
1566 mg = alloc_migration(cache);
1567
1568 mg->overwrite_bio = bio;
1569 mg->invalidate_cblock = cblock;
1570 mg->invalidate_oblock = oblock;
1571
1572 return invalidate_lock(mg);
1573}
1574
1575/*
1576 *--------------------------------------------------------------
1577 * bio processing
1578 *--------------------------------------------------------------
1579 */
1580
1581enum busy {
1582 IDLE,
1583 BUSY
1584};
1585
1586static enum busy spare_migration_bandwidth(struct cache *cache)
1587{
1588 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1589 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1590 cache->sectors_per_block;
1591
1592 if (idle && current_volume <= cache->migration_threshold)
1593 return IDLE;
1594 else
1595 return BUSY;
1596}
1597
1598static void inc_hit_counter(struct cache *cache, struct bio *bio)
1599{
1600 atomic_inc(bio_data_dir(bio) == READ ?
1601 &cache->stats.read_hit : &cache->stats.write_hit);
1602}
1603
1604static void inc_miss_counter(struct cache *cache, struct bio *bio)
1605{
1606 atomic_inc(bio_data_dir(bio) == READ ?
1607 &cache->stats.read_miss : &cache->stats.write_miss);
1608}
1609
1610/*----------------------------------------------------------------*/
1611
1612static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1613 bool *commit_needed)
1614{
1615 int r, data_dir;
1616 bool rb, background_queued;
1617 dm_cblock_t cblock;
1618
1619 *commit_needed = false;
1620
1621 rb = bio_detain_shared(cache, block, bio);
1622 if (!rb) {
1623 /*
1624 * An exclusive lock is held for this block, so we have to
1625 * wait. We set the commit_needed flag so the current
1626 * transaction will be committed asap, allowing this lock
1627 * to be dropped.
1628 */
1629 *commit_needed = true;
1630 return DM_MAPIO_SUBMITTED;
1631 }
1632
1633 data_dir = bio_data_dir(bio);
1634
1635 if (optimisable_bio(cache, bio, block)) {
1636 struct policy_work *op = NULL;
1637
1638 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1639 if (unlikely(r && r != -ENOENT)) {
1640 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1641 cache_device_name(cache), r);
1642 bio_io_error(bio);
1643 return DM_MAPIO_SUBMITTED;
1644 }
1645
1646 if (r == -ENOENT && op) {
1647 bio_drop_shared_lock(cache, bio);
1648 BUG_ON(op->op != POLICY_PROMOTE);
1649 mg_start(cache, op, bio);
1650 return DM_MAPIO_SUBMITTED;
1651 }
1652 } else {
1653 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1654 if (unlikely(r && r != -ENOENT)) {
1655 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1656 cache_device_name(cache), r);
1657 bio_io_error(bio);
1658 return DM_MAPIO_SUBMITTED;
1659 }
1660
1661 if (background_queued)
1662 wake_migration_worker(cache);
1663 }
1664
1665 if (r == -ENOENT) {
1666 struct per_bio_data *pb = get_per_bio_data(bio);
1667
1668 /*
1669 * Miss.
1670 */
1671 inc_miss_counter(cache, bio);
1672 if (pb->req_nr == 0) {
1673 accounted_begin(cache, bio);
1674 remap_to_origin_clear_discard(cache, bio, block);
1675 } else {
1676 /*
1677 * This is a duplicate writethrough io that is no
1678 * longer needed because the block has been demoted.
1679 */
1680 bio_endio(bio);
1681 return DM_MAPIO_SUBMITTED;
1682 }
1683 } else {
1684 /*
1685 * Hit.
1686 */
1687 inc_hit_counter(cache, bio);
1688
1689 /*
1690 * Passthrough always maps to the origin, invalidating any
1691 * cache blocks that are written to.
1692 */
1693 if (passthrough_mode(cache)) {
1694 if (bio_data_dir(bio) == WRITE) {
1695 bio_drop_shared_lock(cache, bio);
1696 atomic_inc(&cache->stats.demotion);
1697 invalidate_start(cache, cblock, block, bio);
1698 } else
1699 remap_to_origin_clear_discard(cache, bio, block);
1700 } else {
1701 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1702 !is_dirty(cache, cblock)) {
1703 remap_to_origin_and_cache(cache, bio, block, cblock);
1704 accounted_begin(cache, bio);
1705 } else
1706 remap_to_cache_dirty(cache, bio, block, cblock);
1707 }
1708 }
1709
1710 /*
1711 * dm core turns FUA requests into a separate payload and FLUSH req.
1712 */
1713 if (bio->bi_opf & REQ_FUA) {
1714 /*
1715 * issue_after_commit will call accounted_begin a second time. So
1716 * we call accounted_complete() to avoid double accounting.
1717 */
1718 accounted_complete(cache, bio);
1719 issue_after_commit(&cache->committer, bio);
1720 *commit_needed = true;
1721 return DM_MAPIO_SUBMITTED;
1722 }
1723
1724 return DM_MAPIO_REMAPPED;
1725}
1726
1727static bool process_bio(struct cache *cache, struct bio *bio)
1728{
1729 bool commit_needed;
1730
1731 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1732 dm_submit_bio_remap(bio, NULL);
1733
1734 return commit_needed;
1735}
1736
1737/*
1738 * A non-zero return indicates read_only or fail_io mode.
1739 */
1740static int commit(struct cache *cache, bool clean_shutdown)
1741{
1742 int r;
1743
1744 if (get_cache_mode(cache) >= CM_READ_ONLY)
1745 return -EINVAL;
1746
1747 atomic_inc(&cache->stats.commit_count);
1748 r = dm_cache_commit(cache->cmd, clean_shutdown);
1749 if (r)
1750 metadata_operation_failed(cache, "dm_cache_commit", r);
1751
1752 return r;
1753}
1754
1755/*
1756 * Used by the batcher.
1757 */
1758static blk_status_t commit_op(void *context)
1759{
1760 struct cache *cache = context;
1761
1762 if (dm_cache_changed_this_transaction(cache->cmd))
1763 return errno_to_blk_status(commit(cache, false));
1764
1765 return 0;
1766}
1767
1768/*----------------------------------------------------------------*/
1769
1770static bool process_flush_bio(struct cache *cache, struct bio *bio)
1771{
1772 struct per_bio_data *pb = get_per_bio_data(bio);
1773
1774 if (!pb->req_nr)
1775 remap_to_origin(cache, bio);
1776 else
1777 remap_to_cache(cache, bio, 0);
1778
1779 issue_after_commit(&cache->committer, bio);
1780 return true;
1781}
1782
1783static bool process_discard_bio(struct cache *cache, struct bio *bio)
1784{
1785 dm_dblock_t b, e;
1786
1787 /*
1788 * FIXME: do we need to lock the region? Or can we just assume the
1789 * user wont be so foolish as to issue discard concurrently with
1790 * other IO?
1791 */
1792 calc_discard_block_range(cache, bio, &b, &e);
1793 while (b != e) {
1794 set_discard(cache, b);
1795 b = to_dblock(from_dblock(b) + 1);
1796 }
1797
1798 if (cache->features.discard_passdown) {
1799 remap_to_origin(cache, bio);
1800 dm_submit_bio_remap(bio, NULL);
1801 } else
1802 bio_endio(bio);
1803
1804 return false;
1805}
1806
1807static void process_deferred_bios(struct work_struct *ws)
1808{
1809 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1810
1811 bool commit_needed = false;
1812 struct bio_list bios;
1813 struct bio *bio;
1814
1815 bio_list_init(&bios);
1816
1817 spin_lock_irq(&cache->lock);
1818 bio_list_merge_init(&bios, &cache->deferred_bios);
1819 spin_unlock_irq(&cache->lock);
1820
1821 while ((bio = bio_list_pop(&bios))) {
1822 if (bio->bi_opf & REQ_PREFLUSH)
1823 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1824
1825 else if (bio_op(bio) == REQ_OP_DISCARD)
1826 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1827
1828 else
1829 commit_needed = process_bio(cache, bio) || commit_needed;
1830 cond_resched();
1831 }
1832
1833 if (commit_needed)
1834 schedule_commit(&cache->committer);
1835}
1836
1837/*
1838 *--------------------------------------------------------------
1839 * Main worker loop
1840 *--------------------------------------------------------------
1841 */
1842static void requeue_deferred_bios(struct cache *cache)
1843{
1844 struct bio *bio;
1845 struct bio_list bios;
1846
1847 bio_list_init(&bios);
1848 bio_list_merge_init(&bios, &cache->deferred_bios);
1849
1850 while ((bio = bio_list_pop(&bios))) {
1851 bio->bi_status = BLK_STS_DM_REQUEUE;
1852 bio_endio(bio);
1853 cond_resched();
1854 }
1855}
1856
1857/*
1858 * We want to commit periodically so that not too much
1859 * unwritten metadata builds up.
1860 */
1861static void do_waker(struct work_struct *ws)
1862{
1863 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1864
1865 policy_tick(cache->policy, true);
1866 wake_migration_worker(cache);
1867 schedule_commit(&cache->committer);
1868 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1869}
1870
1871static void check_migrations(struct work_struct *ws)
1872{
1873 int r;
1874 struct policy_work *op;
1875 struct cache *cache = container_of(ws, struct cache, migration_worker);
1876 enum busy b;
1877
1878 for (;;) {
1879 b = spare_migration_bandwidth(cache);
1880
1881 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1882 if (r == -ENODATA)
1883 break;
1884
1885 if (r) {
1886 DMERR_LIMIT("%s: policy_background_work failed",
1887 cache_device_name(cache));
1888 break;
1889 }
1890
1891 r = mg_start(cache, op, NULL);
1892 if (r)
1893 break;
1894
1895 cond_resched();
1896 }
1897}
1898
1899/*
1900 *--------------------------------------------------------------
1901 * Target methods
1902 *--------------------------------------------------------------
1903 */
1904
1905/*
1906 * This function gets called on the error paths of the constructor, so we
1907 * have to cope with a partially initialised struct.
1908 */
1909static void __destroy(struct cache *cache)
1910{
1911 mempool_exit(&cache->migration_pool);
1912
1913 if (cache->prison)
1914 dm_bio_prison_destroy_v2(cache->prison);
1915
1916 if (cache->wq)
1917 destroy_workqueue(cache->wq);
1918
1919 if (cache->dirty_bitset)
1920 free_bitset(cache->dirty_bitset);
1921
1922 if (cache->discard_bitset)
1923 free_bitset(cache->discard_bitset);
1924
1925 if (cache->copier)
1926 dm_kcopyd_client_destroy(cache->copier);
1927
1928 if (cache->cmd)
1929 dm_cache_metadata_close(cache->cmd);
1930
1931 if (cache->metadata_dev)
1932 dm_put_device(cache->ti, cache->metadata_dev);
1933
1934 if (cache->origin_dev)
1935 dm_put_device(cache->ti, cache->origin_dev);
1936
1937 if (cache->cache_dev)
1938 dm_put_device(cache->ti, cache->cache_dev);
1939
1940 if (cache->policy)
1941 dm_cache_policy_destroy(cache->policy);
1942
1943 bioset_exit(&cache->bs);
1944
1945 kfree(cache);
1946}
1947
1948static void destroy(struct cache *cache)
1949{
1950 unsigned int i;
1951
1952 cancel_delayed_work_sync(&cache->waker);
1953
1954 for (i = 0; i < cache->nr_ctr_args ; i++)
1955 kfree(cache->ctr_args[i]);
1956 kfree(cache->ctr_args);
1957
1958 __destroy(cache);
1959}
1960
1961static void cache_dtr(struct dm_target *ti)
1962{
1963 struct cache *cache = ti->private;
1964
1965 destroy(cache);
1966}
1967
1968static sector_t get_dev_size(struct dm_dev *dev)
1969{
1970 return bdev_nr_sectors(dev->bdev);
1971}
1972
1973/*----------------------------------------------------------------*/
1974
1975/*
1976 * Construct a cache device mapping.
1977 *
1978 * cache <metadata dev> <cache dev> <origin dev> <block size>
1979 * <#feature args> [<feature arg>]*
1980 * <policy> <#policy args> [<policy arg>]*
1981 *
1982 * metadata dev : fast device holding the persistent metadata
1983 * cache dev : fast device holding cached data blocks
1984 * origin dev : slow device holding original data blocks
1985 * block size : cache unit size in sectors
1986 *
1987 * #feature args : number of feature arguments passed
1988 * feature args : writethrough. (The default is writeback.)
1989 *
1990 * policy : the replacement policy to use
1991 * #policy args : an even number of policy arguments corresponding
1992 * to key/value pairs passed to the policy
1993 * policy args : key/value pairs passed to the policy
1994 * E.g. 'sequential_threshold 1024'
1995 * See cache-policies.txt for details.
1996 *
1997 * Optional feature arguments are:
1998 * writethrough : write through caching that prohibits cache block
1999 * content from being different from origin block content.
2000 * Without this argument, the default behaviour is to write
2001 * back cache block contents later for performance reasons,
2002 * so they may differ from the corresponding origin blocks.
2003 */
2004struct cache_args {
2005 struct dm_target *ti;
2006
2007 struct dm_dev *metadata_dev;
2008
2009 struct dm_dev *cache_dev;
2010 sector_t cache_sectors;
2011
2012 struct dm_dev *origin_dev;
2013
2014 uint32_t block_size;
2015
2016 const char *policy_name;
2017 int policy_argc;
2018 const char **policy_argv;
2019
2020 struct cache_features features;
2021};
2022
2023static void destroy_cache_args(struct cache_args *ca)
2024{
2025 if (ca->metadata_dev)
2026 dm_put_device(ca->ti, ca->metadata_dev);
2027
2028 if (ca->cache_dev)
2029 dm_put_device(ca->ti, ca->cache_dev);
2030
2031 if (ca->origin_dev)
2032 dm_put_device(ca->ti, ca->origin_dev);
2033
2034 kfree(ca);
2035}
2036
2037static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2038{
2039 if (!as->argc) {
2040 *error = "Insufficient args";
2041 return false;
2042 }
2043
2044 return true;
2045}
2046
2047static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2048 char **error)
2049{
2050 int r;
2051 sector_t metadata_dev_size;
2052
2053 if (!at_least_one_arg(as, error))
2054 return -EINVAL;
2055
2056 r = dm_get_device(ca->ti, dm_shift_arg(as),
2057 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->metadata_dev);
2058 if (r) {
2059 *error = "Error opening metadata device";
2060 return r;
2061 }
2062
2063 metadata_dev_size = get_dev_size(ca->metadata_dev);
2064 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2065 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2066 ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2067
2068 return 0;
2069}
2070
2071static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2072 char **error)
2073{
2074 int r;
2075
2076 if (!at_least_one_arg(as, error))
2077 return -EINVAL;
2078
2079 r = dm_get_device(ca->ti, dm_shift_arg(as),
2080 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->cache_dev);
2081 if (r) {
2082 *error = "Error opening cache device";
2083 return r;
2084 }
2085 ca->cache_sectors = get_dev_size(ca->cache_dev);
2086
2087 return 0;
2088}
2089
2090static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2091 char **error)
2092{
2093 sector_t origin_sectors;
2094 int r;
2095
2096 if (!at_least_one_arg(as, error))
2097 return -EINVAL;
2098
2099 r = dm_get_device(ca->ti, dm_shift_arg(as),
2100 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->origin_dev);
2101 if (r) {
2102 *error = "Error opening origin device";
2103 return r;
2104 }
2105
2106 origin_sectors = get_dev_size(ca->origin_dev);
2107 if (ca->ti->len > origin_sectors) {
2108 *error = "Device size larger than cached device";
2109 return -EINVAL;
2110 }
2111
2112 return 0;
2113}
2114
2115static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2116 char **error)
2117{
2118 unsigned long block_size;
2119
2120 if (!at_least_one_arg(as, error))
2121 return -EINVAL;
2122
2123 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2124 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2125 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2126 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2127 *error = "Invalid data block size";
2128 return -EINVAL;
2129 }
2130
2131 if (block_size > ca->cache_sectors) {
2132 *error = "Data block size is larger than the cache device";
2133 return -EINVAL;
2134 }
2135
2136 ca->block_size = block_size;
2137
2138 return 0;
2139}
2140
2141static void init_features(struct cache_features *cf)
2142{
2143 cf->mode = CM_WRITE;
2144 cf->io_mode = CM_IO_WRITEBACK;
2145 cf->metadata_version = 1;
2146 cf->discard_passdown = true;
2147}
2148
2149static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2150 char **error)
2151{
2152 static const struct dm_arg _args[] = {
2153 {0, 3, "Invalid number of cache feature arguments"},
2154 };
2155
2156 int r, mode_ctr = 0;
2157 unsigned int argc;
2158 const char *arg;
2159 struct cache_features *cf = &ca->features;
2160
2161 init_features(cf);
2162
2163 r = dm_read_arg_group(_args, as, &argc, error);
2164 if (r)
2165 return -EINVAL;
2166
2167 while (argc--) {
2168 arg = dm_shift_arg(as);
2169
2170 if (!strcasecmp(arg, "writeback")) {
2171 cf->io_mode = CM_IO_WRITEBACK;
2172 mode_ctr++;
2173 }
2174
2175 else if (!strcasecmp(arg, "writethrough")) {
2176 cf->io_mode = CM_IO_WRITETHROUGH;
2177 mode_ctr++;
2178 }
2179
2180 else if (!strcasecmp(arg, "passthrough")) {
2181 cf->io_mode = CM_IO_PASSTHROUGH;
2182 mode_ctr++;
2183 }
2184
2185 else if (!strcasecmp(arg, "metadata2"))
2186 cf->metadata_version = 2;
2187
2188 else if (!strcasecmp(arg, "no_discard_passdown"))
2189 cf->discard_passdown = false;
2190
2191 else {
2192 *error = "Unrecognised cache feature requested";
2193 return -EINVAL;
2194 }
2195 }
2196
2197 if (mode_ctr > 1) {
2198 *error = "Duplicate cache io_mode features requested";
2199 return -EINVAL;
2200 }
2201
2202 return 0;
2203}
2204
2205static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2206 char **error)
2207{
2208 static const struct dm_arg _args[] = {
2209 {0, 1024, "Invalid number of policy arguments"},
2210 };
2211
2212 int r;
2213
2214 if (!at_least_one_arg(as, error))
2215 return -EINVAL;
2216
2217 ca->policy_name = dm_shift_arg(as);
2218
2219 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2220 if (r)
2221 return -EINVAL;
2222
2223 ca->policy_argv = (const char **)as->argv;
2224 dm_consume_args(as, ca->policy_argc);
2225
2226 return 0;
2227}
2228
2229static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2230 char **error)
2231{
2232 int r;
2233 struct dm_arg_set as;
2234
2235 as.argc = argc;
2236 as.argv = argv;
2237
2238 r = parse_metadata_dev(ca, &as, error);
2239 if (r)
2240 return r;
2241
2242 r = parse_cache_dev(ca, &as, error);
2243 if (r)
2244 return r;
2245
2246 r = parse_origin_dev(ca, &as, error);
2247 if (r)
2248 return r;
2249
2250 r = parse_block_size(ca, &as, error);
2251 if (r)
2252 return r;
2253
2254 r = parse_features(ca, &as, error);
2255 if (r)
2256 return r;
2257
2258 r = parse_policy(ca, &as, error);
2259 if (r)
2260 return r;
2261
2262 return 0;
2263}
2264
2265/*----------------------------------------------------------------*/
2266
2267static struct kmem_cache *migration_cache = NULL;
2268
2269#define NOT_CORE_OPTION 1
2270
2271static int process_config_option(struct cache *cache, const char *key, const char *value)
2272{
2273 unsigned long tmp;
2274
2275 if (!strcasecmp(key, "migration_threshold")) {
2276 if (kstrtoul(value, 10, &tmp))
2277 return -EINVAL;
2278
2279 cache->migration_threshold = tmp;
2280 return 0;
2281 }
2282
2283 return NOT_CORE_OPTION;
2284}
2285
2286static int set_config_value(struct cache *cache, const char *key, const char *value)
2287{
2288 int r = process_config_option(cache, key, value);
2289
2290 if (r == NOT_CORE_OPTION)
2291 r = policy_set_config_value(cache->policy, key, value);
2292
2293 if (r)
2294 DMWARN("bad config value for %s: %s", key, value);
2295
2296 return r;
2297}
2298
2299static int set_config_values(struct cache *cache, int argc, const char **argv)
2300{
2301 int r = 0;
2302
2303 if (argc & 1) {
2304 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2305 return -EINVAL;
2306 }
2307
2308 while (argc) {
2309 r = set_config_value(cache, argv[0], argv[1]);
2310 if (r)
2311 break;
2312
2313 argc -= 2;
2314 argv += 2;
2315 }
2316
2317 return r;
2318}
2319
2320static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2321 char **error)
2322{
2323 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2324 cache->cache_size,
2325 cache->origin_sectors,
2326 cache->sectors_per_block);
2327 if (IS_ERR(p)) {
2328 *error = "Error creating cache's policy";
2329 return PTR_ERR(p);
2330 }
2331 cache->policy = p;
2332 BUG_ON(!cache->policy);
2333
2334 return 0;
2335}
2336
2337/*
2338 * We want the discard block size to be at least the size of the cache
2339 * block size and have no more than 2^14 discard blocks across the origin.
2340 */
2341#define MAX_DISCARD_BLOCKS (1 << 14)
2342
2343static bool too_many_discard_blocks(sector_t discard_block_size,
2344 sector_t origin_size)
2345{
2346 (void) sector_div(origin_size, discard_block_size);
2347
2348 return origin_size > MAX_DISCARD_BLOCKS;
2349}
2350
2351static sector_t calculate_discard_block_size(sector_t cache_block_size,
2352 sector_t origin_size)
2353{
2354 sector_t discard_block_size = cache_block_size;
2355
2356 if (origin_size)
2357 while (too_many_discard_blocks(discard_block_size, origin_size))
2358 discard_block_size *= 2;
2359
2360 return discard_block_size;
2361}
2362
2363static void set_cache_size(struct cache *cache, dm_cblock_t size)
2364{
2365 dm_block_t nr_blocks = from_cblock(size);
2366
2367 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2368 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2369 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2370 "Please consider increasing the cache block size to reduce the overall cache block count.",
2371 (unsigned long long) nr_blocks);
2372
2373 cache->cache_size = size;
2374}
2375
2376#define DEFAULT_MIGRATION_THRESHOLD 2048
2377
2378static int cache_create(struct cache_args *ca, struct cache **result)
2379{
2380 int r = 0;
2381 char **error = &ca->ti->error;
2382 struct cache *cache;
2383 struct dm_target *ti = ca->ti;
2384 dm_block_t origin_blocks;
2385 struct dm_cache_metadata *cmd;
2386 bool may_format = ca->features.mode == CM_WRITE;
2387
2388 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2389 if (!cache)
2390 return -ENOMEM;
2391
2392 cache->ti = ca->ti;
2393 ti->private = cache;
2394 ti->accounts_remapped_io = true;
2395 ti->num_flush_bios = 2;
2396 ti->flush_supported = true;
2397
2398 ti->num_discard_bios = 1;
2399 ti->discards_supported = true;
2400
2401 ti->per_io_data_size = sizeof(struct per_bio_data);
2402
2403 cache->features = ca->features;
2404 if (writethrough_mode(cache)) {
2405 /* Create bioset for writethrough bios issued to origin */
2406 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2407 if (r)
2408 goto bad;
2409 }
2410
2411 cache->metadata_dev = ca->metadata_dev;
2412 cache->origin_dev = ca->origin_dev;
2413 cache->cache_dev = ca->cache_dev;
2414
2415 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2416
2417 origin_blocks = cache->origin_sectors = ti->len;
2418 origin_blocks = block_div(origin_blocks, ca->block_size);
2419 cache->origin_blocks = to_oblock(origin_blocks);
2420
2421 cache->sectors_per_block = ca->block_size;
2422 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2423 r = -EINVAL;
2424 goto bad;
2425 }
2426
2427 if (ca->block_size & (ca->block_size - 1)) {
2428 dm_block_t cache_size = ca->cache_sectors;
2429
2430 cache->sectors_per_block_shift = -1;
2431 cache_size = block_div(cache_size, ca->block_size);
2432 set_cache_size(cache, to_cblock(cache_size));
2433 } else {
2434 cache->sectors_per_block_shift = __ffs(ca->block_size);
2435 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2436 }
2437
2438 r = create_cache_policy(cache, ca, error);
2439 if (r)
2440 goto bad;
2441
2442 cache->policy_nr_args = ca->policy_argc;
2443 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2444
2445 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2446 if (r) {
2447 *error = "Error setting cache policy's config values";
2448 goto bad;
2449 }
2450
2451 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2452 ca->block_size, may_format,
2453 dm_cache_policy_get_hint_size(cache->policy),
2454 ca->features.metadata_version);
2455 if (IS_ERR(cmd)) {
2456 *error = "Error creating metadata object";
2457 r = PTR_ERR(cmd);
2458 goto bad;
2459 }
2460 cache->cmd = cmd;
2461 set_cache_mode(cache, CM_WRITE);
2462 if (get_cache_mode(cache) != CM_WRITE) {
2463 *error = "Unable to get write access to metadata, please check/repair metadata.";
2464 r = -EINVAL;
2465 goto bad;
2466 }
2467
2468 if (passthrough_mode(cache)) {
2469 bool all_clean;
2470
2471 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2472 if (r) {
2473 *error = "dm_cache_metadata_all_clean() failed";
2474 goto bad;
2475 }
2476
2477 if (!all_clean) {
2478 *error = "Cannot enter passthrough mode unless all blocks are clean";
2479 r = -EINVAL;
2480 goto bad;
2481 }
2482
2483 policy_allow_migrations(cache->policy, false);
2484 }
2485
2486 spin_lock_init(&cache->lock);
2487 bio_list_init(&cache->deferred_bios);
2488 atomic_set(&cache->nr_allocated_migrations, 0);
2489 atomic_set(&cache->nr_io_migrations, 0);
2490 init_waitqueue_head(&cache->migration_wait);
2491
2492 r = -ENOMEM;
2493 atomic_set(&cache->nr_dirty, 0);
2494 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2495 if (!cache->dirty_bitset) {
2496 *error = "could not allocate dirty bitset";
2497 goto bad;
2498 }
2499 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2500
2501 cache->discard_block_size =
2502 calculate_discard_block_size(cache->sectors_per_block,
2503 cache->origin_sectors);
2504 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2505 cache->discard_block_size));
2506 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2507 if (!cache->discard_bitset) {
2508 *error = "could not allocate discard bitset";
2509 goto bad;
2510 }
2511 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2512
2513 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2514 if (IS_ERR(cache->copier)) {
2515 *error = "could not create kcopyd client";
2516 r = PTR_ERR(cache->copier);
2517 goto bad;
2518 }
2519
2520 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2521 if (!cache->wq) {
2522 *error = "could not create workqueue for metadata object";
2523 goto bad;
2524 }
2525 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2526 INIT_WORK(&cache->migration_worker, check_migrations);
2527 INIT_DELAYED_WORK(&cache->waker, do_waker);
2528
2529 cache->prison = dm_bio_prison_create_v2(cache->wq);
2530 if (!cache->prison) {
2531 *error = "could not create bio prison";
2532 goto bad;
2533 }
2534
2535 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2536 migration_cache);
2537 if (r) {
2538 *error = "Error creating cache's migration mempool";
2539 goto bad;
2540 }
2541
2542 cache->need_tick_bio = true;
2543 cache->sized = false;
2544 cache->invalidate = false;
2545 cache->commit_requested = false;
2546 cache->loaded_mappings = false;
2547 cache->loaded_discards = false;
2548
2549 load_stats(cache);
2550
2551 atomic_set(&cache->stats.demotion, 0);
2552 atomic_set(&cache->stats.promotion, 0);
2553 atomic_set(&cache->stats.copies_avoided, 0);
2554 atomic_set(&cache->stats.cache_cell_clash, 0);
2555 atomic_set(&cache->stats.commit_count, 0);
2556 atomic_set(&cache->stats.discard_count, 0);
2557
2558 spin_lock_init(&cache->invalidation_lock);
2559 INIT_LIST_HEAD(&cache->invalidation_requests);
2560
2561 batcher_init(&cache->committer, commit_op, cache,
2562 issue_op, cache, cache->wq);
2563 dm_iot_init(&cache->tracker);
2564
2565 init_rwsem(&cache->background_work_lock);
2566 prevent_background_work(cache);
2567
2568 *result = cache;
2569 return 0;
2570bad:
2571 __destroy(cache);
2572 return r;
2573}
2574
2575static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2576{
2577 unsigned int i;
2578 const char **copy;
2579
2580 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2581 if (!copy)
2582 return -ENOMEM;
2583 for (i = 0; i < argc; i++) {
2584 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2585 if (!copy[i]) {
2586 while (i--)
2587 kfree(copy[i]);
2588 kfree(copy);
2589 return -ENOMEM;
2590 }
2591 }
2592
2593 cache->nr_ctr_args = argc;
2594 cache->ctr_args = copy;
2595
2596 return 0;
2597}
2598
2599static int cache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2600{
2601 int r = -EINVAL;
2602 struct cache_args *ca;
2603 struct cache *cache = NULL;
2604
2605 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2606 if (!ca) {
2607 ti->error = "Error allocating memory for cache";
2608 return -ENOMEM;
2609 }
2610 ca->ti = ti;
2611
2612 r = parse_cache_args(ca, argc, argv, &ti->error);
2613 if (r)
2614 goto out;
2615
2616 r = cache_create(ca, &cache);
2617 if (r)
2618 goto out;
2619
2620 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2621 if (r) {
2622 __destroy(cache);
2623 goto out;
2624 }
2625
2626 ti->private = cache;
2627out:
2628 destroy_cache_args(ca);
2629 return r;
2630}
2631
2632/*----------------------------------------------------------------*/
2633
2634static int cache_map(struct dm_target *ti, struct bio *bio)
2635{
2636 struct cache *cache = ti->private;
2637
2638 int r;
2639 bool commit_needed;
2640 dm_oblock_t block = get_bio_block(cache, bio);
2641
2642 init_per_bio_data(bio);
2643 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2644 /*
2645 * This can only occur if the io goes to a partial block at
2646 * the end of the origin device. We don't cache these.
2647 * Just remap to the origin and carry on.
2648 */
2649 remap_to_origin(cache, bio);
2650 accounted_begin(cache, bio);
2651 return DM_MAPIO_REMAPPED;
2652 }
2653
2654 if (discard_or_flush(bio)) {
2655 defer_bio(cache, bio);
2656 return DM_MAPIO_SUBMITTED;
2657 }
2658
2659 r = map_bio(cache, bio, block, &commit_needed);
2660 if (commit_needed)
2661 schedule_commit(&cache->committer);
2662
2663 return r;
2664}
2665
2666static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2667{
2668 struct cache *cache = ti->private;
2669 unsigned long flags;
2670 struct per_bio_data *pb = get_per_bio_data(bio);
2671
2672 if (pb->tick) {
2673 policy_tick(cache->policy, false);
2674
2675 spin_lock_irqsave(&cache->lock, flags);
2676 cache->need_tick_bio = true;
2677 spin_unlock_irqrestore(&cache->lock, flags);
2678 }
2679
2680 bio_drop_shared_lock(cache, bio);
2681 accounted_complete(cache, bio);
2682
2683 return DM_ENDIO_DONE;
2684}
2685
2686static int write_dirty_bitset(struct cache *cache)
2687{
2688 int r;
2689
2690 if (get_cache_mode(cache) >= CM_READ_ONLY)
2691 return -EINVAL;
2692
2693 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2694 if (r)
2695 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2696
2697 return r;
2698}
2699
2700static int write_discard_bitset(struct cache *cache)
2701{
2702 unsigned int i, r;
2703
2704 if (get_cache_mode(cache) >= CM_READ_ONLY)
2705 return -EINVAL;
2706
2707 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2708 cache->discard_nr_blocks);
2709 if (r) {
2710 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2711 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2712 return r;
2713 }
2714
2715 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2716 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2717 is_discarded(cache, to_dblock(i)));
2718 if (r) {
2719 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2720 return r;
2721 }
2722 }
2723
2724 return 0;
2725}
2726
2727static int write_hints(struct cache *cache)
2728{
2729 int r;
2730
2731 if (get_cache_mode(cache) >= CM_READ_ONLY)
2732 return -EINVAL;
2733
2734 r = dm_cache_write_hints(cache->cmd, cache->policy);
2735 if (r) {
2736 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2737 return r;
2738 }
2739
2740 return 0;
2741}
2742
2743/*
2744 * returns true on success
2745 */
2746static bool sync_metadata(struct cache *cache)
2747{
2748 int r1, r2, r3, r4;
2749
2750 r1 = write_dirty_bitset(cache);
2751 if (r1)
2752 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2753
2754 r2 = write_discard_bitset(cache);
2755 if (r2)
2756 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2757
2758 save_stats(cache);
2759
2760 r3 = write_hints(cache);
2761 if (r3)
2762 DMERR("%s: could not write hints", cache_device_name(cache));
2763
2764 /*
2765 * If writing the above metadata failed, we still commit, but don't
2766 * set the clean shutdown flag. This will effectively force every
2767 * dirty bit to be set on reload.
2768 */
2769 r4 = commit(cache, !r1 && !r2 && !r3);
2770 if (r4)
2771 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2772
2773 return !r1 && !r2 && !r3 && !r4;
2774}
2775
2776static void cache_postsuspend(struct dm_target *ti)
2777{
2778 struct cache *cache = ti->private;
2779
2780 prevent_background_work(cache);
2781 BUG_ON(atomic_read(&cache->nr_io_migrations));
2782
2783 cancel_delayed_work_sync(&cache->waker);
2784 drain_workqueue(cache->wq);
2785 WARN_ON(cache->tracker.in_flight);
2786
2787 /*
2788 * If it's a flush suspend there won't be any deferred bios, so this
2789 * call is harmless.
2790 */
2791 requeue_deferred_bios(cache);
2792
2793 if (get_cache_mode(cache) == CM_WRITE)
2794 (void) sync_metadata(cache);
2795}
2796
2797static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2798 bool dirty, uint32_t hint, bool hint_valid)
2799{
2800 struct cache *cache = context;
2801
2802 if (dirty) {
2803 set_bit(from_cblock(cblock), cache->dirty_bitset);
2804 atomic_inc(&cache->nr_dirty);
2805 } else
2806 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2807
2808 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2809}
2810
2811/*
2812 * The discard block size in the on disk metadata is not
2813 * necessarily the same as we're currently using. So we have to
2814 * be careful to only set the discarded attribute if we know it
2815 * covers a complete block of the new size.
2816 */
2817struct discard_load_info {
2818 struct cache *cache;
2819
2820 /*
2821 * These blocks are sized using the on disk dblock size, rather
2822 * than the current one.
2823 */
2824 dm_block_t block_size;
2825 dm_block_t discard_begin, discard_end;
2826};
2827
2828static void discard_load_info_init(struct cache *cache,
2829 struct discard_load_info *li)
2830{
2831 li->cache = cache;
2832 li->discard_begin = li->discard_end = 0;
2833}
2834
2835static void set_discard_range(struct discard_load_info *li)
2836{
2837 sector_t b, e;
2838
2839 if (li->discard_begin == li->discard_end)
2840 return;
2841
2842 /*
2843 * Convert to sectors.
2844 */
2845 b = li->discard_begin * li->block_size;
2846 e = li->discard_end * li->block_size;
2847
2848 /*
2849 * Then convert back to the current dblock size.
2850 */
2851 b = dm_sector_div_up(b, li->cache->discard_block_size);
2852 sector_div(e, li->cache->discard_block_size);
2853
2854 /*
2855 * The origin may have shrunk, so we need to check we're still in
2856 * bounds.
2857 */
2858 if (e > from_dblock(li->cache->discard_nr_blocks))
2859 e = from_dblock(li->cache->discard_nr_blocks);
2860
2861 for (; b < e; b++)
2862 set_discard(li->cache, to_dblock(b));
2863}
2864
2865static int load_discard(void *context, sector_t discard_block_size,
2866 dm_dblock_t dblock, bool discard)
2867{
2868 struct discard_load_info *li = context;
2869
2870 li->block_size = discard_block_size;
2871
2872 if (discard) {
2873 if (from_dblock(dblock) == li->discard_end)
2874 /*
2875 * We're already in a discard range, just extend it.
2876 */
2877 li->discard_end = li->discard_end + 1ULL;
2878
2879 else {
2880 /*
2881 * Emit the old range and start a new one.
2882 */
2883 set_discard_range(li);
2884 li->discard_begin = from_dblock(dblock);
2885 li->discard_end = li->discard_begin + 1ULL;
2886 }
2887 } else {
2888 set_discard_range(li);
2889 li->discard_begin = li->discard_end = 0;
2890 }
2891
2892 return 0;
2893}
2894
2895static dm_cblock_t get_cache_dev_size(struct cache *cache)
2896{
2897 sector_t size = get_dev_size(cache->cache_dev);
2898 (void) sector_div(size, cache->sectors_per_block);
2899 return to_cblock(size);
2900}
2901
2902static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2903{
2904 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2905 DMERR("%s: unable to extend cache due to missing cache table reload",
2906 cache_device_name(cache));
2907 return false;
2908 }
2909
2910 /*
2911 * We can't drop a dirty block when shrinking the cache.
2912 */
2913 if (cache->loaded_mappings) {
2914 new_size = to_cblock(find_next_bit(cache->dirty_bitset,
2915 from_cblock(cache->cache_size),
2916 from_cblock(new_size)));
2917 if (new_size != cache->cache_size) {
2918 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2919 cache_device_name(cache),
2920 (unsigned long long) from_cblock(new_size));
2921 return false;
2922 }
2923 }
2924
2925 return true;
2926}
2927
2928static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2929{
2930 int r;
2931
2932 r = dm_cache_resize(cache->cmd, new_size);
2933 if (r) {
2934 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2935 metadata_operation_failed(cache, "dm_cache_resize", r);
2936 return r;
2937 }
2938
2939 set_cache_size(cache, new_size);
2940
2941 return 0;
2942}
2943
2944static int cache_preresume(struct dm_target *ti)
2945{
2946 int r = 0;
2947 struct cache *cache = ti->private;
2948 dm_cblock_t csize = get_cache_dev_size(cache);
2949
2950 /*
2951 * Check to see if the cache has resized.
2952 */
2953 if (!cache->sized || csize != cache->cache_size) {
2954 if (!can_resize(cache, csize))
2955 return -EINVAL;
2956
2957 r = resize_cache_dev(cache, csize);
2958 if (r)
2959 return r;
2960
2961 cache->sized = true;
2962 }
2963
2964 if (!cache->loaded_mappings) {
2965 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2966 load_mapping, cache);
2967 if (r) {
2968 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2969 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2970 return r;
2971 }
2972
2973 cache->loaded_mappings = true;
2974 }
2975
2976 if (!cache->loaded_discards) {
2977 struct discard_load_info li;
2978
2979 /*
2980 * The discard bitset could have been resized, or the
2981 * discard block size changed. To be safe we start by
2982 * setting every dblock to not discarded.
2983 */
2984 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2985
2986 discard_load_info_init(cache, &li);
2987 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2988 if (r) {
2989 DMERR("%s: could not load origin discards", cache_device_name(cache));
2990 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2991 return r;
2992 }
2993 set_discard_range(&li);
2994
2995 cache->loaded_discards = true;
2996 }
2997
2998 return r;
2999}
3000
3001static void cache_resume(struct dm_target *ti)
3002{
3003 struct cache *cache = ti->private;
3004
3005 cache->need_tick_bio = true;
3006 allow_background_work(cache);
3007 do_waker(&cache->waker.work);
3008}
3009
3010static void emit_flags(struct cache *cache, char *result,
3011 unsigned int maxlen, ssize_t *sz_ptr)
3012{
3013 ssize_t sz = *sz_ptr;
3014 struct cache_features *cf = &cache->features;
3015 unsigned int count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3016
3017 DMEMIT("%u ", count);
3018
3019 if (cf->metadata_version == 2)
3020 DMEMIT("metadata2 ");
3021
3022 if (writethrough_mode(cache))
3023 DMEMIT("writethrough ");
3024
3025 else if (passthrough_mode(cache))
3026 DMEMIT("passthrough ");
3027
3028 else if (writeback_mode(cache))
3029 DMEMIT("writeback ");
3030
3031 else {
3032 DMEMIT("unknown ");
3033 DMERR("%s: internal error: unknown io mode: %d",
3034 cache_device_name(cache), (int) cf->io_mode);
3035 }
3036
3037 if (!cf->discard_passdown)
3038 DMEMIT("no_discard_passdown ");
3039
3040 *sz_ptr = sz;
3041}
3042
3043/*
3044 * Status format:
3045 *
3046 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3047 * <cache block size> <#used cache blocks>/<#total cache blocks>
3048 * <#read hits> <#read misses> <#write hits> <#write misses>
3049 * <#demotions> <#promotions> <#dirty>
3050 * <#features> <features>*
3051 * <#core args> <core args>
3052 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3053 */
3054static void cache_status(struct dm_target *ti, status_type_t type,
3055 unsigned int status_flags, char *result, unsigned int maxlen)
3056{
3057 int r = 0;
3058 unsigned int i;
3059 ssize_t sz = 0;
3060 dm_block_t nr_free_blocks_metadata = 0;
3061 dm_block_t nr_blocks_metadata = 0;
3062 char buf[BDEVNAME_SIZE];
3063 struct cache *cache = ti->private;
3064 dm_cblock_t residency;
3065 bool needs_check;
3066
3067 switch (type) {
3068 case STATUSTYPE_INFO:
3069 if (get_cache_mode(cache) == CM_FAIL) {
3070 DMEMIT("Fail");
3071 break;
3072 }
3073
3074 /* Commit to ensure statistics aren't out-of-date */
3075 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3076 (void) commit(cache, false);
3077
3078 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3079 if (r) {
3080 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3081 cache_device_name(cache), r);
3082 goto err;
3083 }
3084
3085 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3086 if (r) {
3087 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3088 cache_device_name(cache), r);
3089 goto err;
3090 }
3091
3092 residency = policy_residency(cache->policy);
3093
3094 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3095 (unsigned int)DM_CACHE_METADATA_BLOCK_SIZE,
3096 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3097 (unsigned long long)nr_blocks_metadata,
3098 (unsigned long long)cache->sectors_per_block,
3099 (unsigned long long) from_cblock(residency),
3100 (unsigned long long) from_cblock(cache->cache_size),
3101 (unsigned int) atomic_read(&cache->stats.read_hit),
3102 (unsigned int) atomic_read(&cache->stats.read_miss),
3103 (unsigned int) atomic_read(&cache->stats.write_hit),
3104 (unsigned int) atomic_read(&cache->stats.write_miss),
3105 (unsigned int) atomic_read(&cache->stats.demotion),
3106 (unsigned int) atomic_read(&cache->stats.promotion),
3107 (unsigned long) atomic_read(&cache->nr_dirty));
3108
3109 emit_flags(cache, result, maxlen, &sz);
3110
3111 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3112
3113 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3114 if (sz < maxlen) {
3115 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3116 if (r)
3117 DMERR("%s: policy_emit_config_values returned %d",
3118 cache_device_name(cache), r);
3119 }
3120
3121 if (get_cache_mode(cache) == CM_READ_ONLY)
3122 DMEMIT("ro ");
3123 else
3124 DMEMIT("rw ");
3125
3126 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3127
3128 if (r || needs_check)
3129 DMEMIT("needs_check ");
3130 else
3131 DMEMIT("- ");
3132
3133 break;
3134
3135 case STATUSTYPE_TABLE:
3136 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3137 DMEMIT("%s ", buf);
3138 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3139 DMEMIT("%s ", buf);
3140 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3141 DMEMIT("%s", buf);
3142
3143 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3144 DMEMIT(" %s", cache->ctr_args[i]);
3145 if (cache->nr_ctr_args)
3146 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3147 break;
3148
3149 case STATUSTYPE_IMA:
3150 DMEMIT_TARGET_NAME_VERSION(ti->type);
3151 if (get_cache_mode(cache) == CM_FAIL)
3152 DMEMIT(",metadata_mode=fail");
3153 else if (get_cache_mode(cache) == CM_READ_ONLY)
3154 DMEMIT(",metadata_mode=ro");
3155 else
3156 DMEMIT(",metadata_mode=rw");
3157
3158 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3159 DMEMIT(",cache_metadata_device=%s", buf);
3160 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3161 DMEMIT(",cache_device=%s", buf);
3162 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3163 DMEMIT(",cache_origin_device=%s", buf);
3164 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3165 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3166 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3167 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3168 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3169 DMEMIT(";");
3170 break;
3171 }
3172
3173 return;
3174
3175err:
3176 DMEMIT("Error");
3177}
3178
3179/*
3180 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3181 * the one-past-the-end value.
3182 */
3183struct cblock_range {
3184 dm_cblock_t begin;
3185 dm_cblock_t end;
3186};
3187
3188/*
3189 * A cache block range can take two forms:
3190 *
3191 * i) A single cblock, eg. '3456'
3192 * ii) A begin and end cblock with a dash between, eg. 123-234
3193 */
3194static int parse_cblock_range(struct cache *cache, const char *str,
3195 struct cblock_range *result)
3196{
3197 char dummy;
3198 uint64_t b, e;
3199 int r;
3200
3201 /*
3202 * Try and parse form (ii) first.
3203 */
3204 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3205
3206 if (r == 2) {
3207 result->begin = to_cblock(b);
3208 result->end = to_cblock(e);
3209 return 0;
3210 }
3211
3212 /*
3213 * That didn't work, try form (i).
3214 */
3215 r = sscanf(str, "%llu%c", &b, &dummy);
3216
3217 if (r == 1) {
3218 result->begin = to_cblock(b);
3219 result->end = to_cblock(from_cblock(result->begin) + 1u);
3220 return 0;
3221 }
3222
3223 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3224 return -EINVAL;
3225}
3226
3227static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3228{
3229 uint64_t b = from_cblock(range->begin);
3230 uint64_t e = from_cblock(range->end);
3231 uint64_t n = from_cblock(cache->cache_size);
3232
3233 if (b >= n) {
3234 DMERR("%s: begin cblock out of range: %llu >= %llu",
3235 cache_device_name(cache), b, n);
3236 return -EINVAL;
3237 }
3238
3239 if (e > n) {
3240 DMERR("%s: end cblock out of range: %llu > %llu",
3241 cache_device_name(cache), e, n);
3242 return -EINVAL;
3243 }
3244
3245 if (b >= e) {
3246 DMERR("%s: invalid cblock range: %llu >= %llu",
3247 cache_device_name(cache), b, e);
3248 return -EINVAL;
3249 }
3250
3251 return 0;
3252}
3253
3254static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3255{
3256 return to_cblock(from_cblock(b) + 1);
3257}
3258
3259static int request_invalidation(struct cache *cache, struct cblock_range *range)
3260{
3261 int r = 0;
3262
3263 /*
3264 * We don't need to do any locking here because we know we're in
3265 * passthrough mode. There's is potential for a race between an
3266 * invalidation triggered by an io and an invalidation message. This
3267 * is harmless, we must not worry if the policy call fails.
3268 */
3269 while (range->begin != range->end) {
3270 r = invalidate_cblock(cache, range->begin);
3271 if (r)
3272 return r;
3273
3274 range->begin = cblock_succ(range->begin);
3275 }
3276
3277 cache->commit_requested = true;
3278 return r;
3279}
3280
3281static int process_invalidate_cblocks_message(struct cache *cache, unsigned int count,
3282 const char **cblock_ranges)
3283{
3284 int r = 0;
3285 unsigned int i;
3286 struct cblock_range range;
3287
3288 if (!passthrough_mode(cache)) {
3289 DMERR("%s: cache has to be in passthrough mode for invalidation",
3290 cache_device_name(cache));
3291 return -EPERM;
3292 }
3293
3294 for (i = 0; i < count; i++) {
3295 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3296 if (r)
3297 break;
3298
3299 r = validate_cblock_range(cache, &range);
3300 if (r)
3301 break;
3302
3303 /*
3304 * Pass begin and end origin blocks to the worker and wake it.
3305 */
3306 r = request_invalidation(cache, &range);
3307 if (r)
3308 break;
3309 }
3310
3311 return r;
3312}
3313
3314/*
3315 * Supports
3316 * "<key> <value>"
3317 * and
3318 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3319 *
3320 * The key migration_threshold is supported by the cache target core.
3321 */
3322static int cache_message(struct dm_target *ti, unsigned int argc, char **argv,
3323 char *result, unsigned int maxlen)
3324{
3325 struct cache *cache = ti->private;
3326
3327 if (!argc)
3328 return -EINVAL;
3329
3330 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3331 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3332 cache_device_name(cache));
3333 return -EOPNOTSUPP;
3334 }
3335
3336 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3337 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3338
3339 if (argc != 2)
3340 return -EINVAL;
3341
3342 return set_config_value(cache, argv[0], argv[1]);
3343}
3344
3345static int cache_iterate_devices(struct dm_target *ti,
3346 iterate_devices_callout_fn fn, void *data)
3347{
3348 int r = 0;
3349 struct cache *cache = ti->private;
3350
3351 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3352 if (!r)
3353 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3354
3355 return r;
3356}
3357
3358/*
3359 * If discard_passdown was enabled verify that the origin device
3360 * supports discards. Disable discard_passdown if not.
3361 */
3362static void disable_passdown_if_not_supported(struct cache *cache)
3363{
3364 struct block_device *origin_bdev = cache->origin_dev->bdev;
3365 struct queue_limits *origin_limits = bdev_limits(origin_bdev);
3366 const char *reason = NULL;
3367
3368 if (!cache->features.discard_passdown)
3369 return;
3370
3371 if (!bdev_max_discard_sectors(origin_bdev))
3372 reason = "discard unsupported";
3373
3374 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3375 reason = "max discard sectors smaller than a block";
3376
3377 if (reason) {
3378 DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3379 origin_bdev, reason);
3380 cache->features.discard_passdown = false;
3381 }
3382}
3383
3384static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3385{
3386 struct block_device *origin_bdev = cache->origin_dev->bdev;
3387 struct queue_limits *origin_limits = bdev_limits(origin_bdev);
3388
3389 if (!cache->features.discard_passdown) {
3390 /* No passdown is done so setting own virtual limits */
3391 limits->max_hw_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3392 cache->origin_sectors);
3393 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3394 return;
3395 }
3396
3397 /*
3398 * cache_iterate_devices() is stacking both origin and fast device limits
3399 * but discards aren't passed to fast device, so inherit origin's limits.
3400 */
3401 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3402 limits->discard_granularity = origin_limits->discard_granularity;
3403 limits->discard_alignment = origin_limits->discard_alignment;
3404}
3405
3406static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3407{
3408 struct cache *cache = ti->private;
3409 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3410
3411 /*
3412 * If the system-determined stacked limits are compatible with the
3413 * cache's blocksize (io_opt is a factor) do not override them.
3414 */
3415 if (io_opt_sectors < cache->sectors_per_block ||
3416 do_div(io_opt_sectors, cache->sectors_per_block)) {
3417 limits->io_min = cache->sectors_per_block << SECTOR_SHIFT;
3418 limits->io_opt = cache->sectors_per_block << SECTOR_SHIFT;
3419 }
3420
3421 disable_passdown_if_not_supported(cache);
3422 set_discard_limits(cache, limits);
3423}
3424
3425/*----------------------------------------------------------------*/
3426
3427static struct target_type cache_target = {
3428 .name = "cache",
3429 .version = {2, 2, 0},
3430 .module = THIS_MODULE,
3431 .ctr = cache_ctr,
3432 .dtr = cache_dtr,
3433 .map = cache_map,
3434 .end_io = cache_end_io,
3435 .postsuspend = cache_postsuspend,
3436 .preresume = cache_preresume,
3437 .resume = cache_resume,
3438 .status = cache_status,
3439 .message = cache_message,
3440 .iterate_devices = cache_iterate_devices,
3441 .io_hints = cache_io_hints,
3442};
3443
3444static int __init dm_cache_init(void)
3445{
3446 int r;
3447
3448 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3449 if (!migration_cache) {
3450 r = -ENOMEM;
3451 goto err;
3452 }
3453
3454 btracker_work_cache = kmem_cache_create("dm_cache_bt_work",
3455 sizeof(struct bt_work), __alignof__(struct bt_work), 0, NULL);
3456 if (!btracker_work_cache) {
3457 r = -ENOMEM;
3458 goto err;
3459 }
3460
3461 r = dm_register_target(&cache_target);
3462 if (r) {
3463 goto err;
3464 }
3465
3466 return 0;
3467
3468err:
3469 kmem_cache_destroy(migration_cache);
3470 kmem_cache_destroy(btracker_work_cache);
3471 return r;
3472}
3473
3474static void __exit dm_cache_exit(void)
3475{
3476 dm_unregister_target(&cache_target);
3477 kmem_cache_destroy(migration_cache);
3478 kmem_cache_destroy(btracker_work_cache);
3479}
3480
3481module_init(dm_cache_init);
3482module_exit(dm_cache_exit);
3483
3484MODULE_DESCRIPTION(DM_NAME " cache target");
3485MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3486MODULE_LICENSE("GPL");