Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2012 Red Hat. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.h"
9#include "dm-bio-prison-v2.h"
10#include "dm-bio-record.h"
11#include "dm-cache-metadata.h"
12#include "dm-io-tracker.h"
13
14#include <linux/dm-io.h>
15#include <linux/dm-kcopyd.h>
16#include <linux/jiffies.h>
17#include <linux/init.h>
18#include <linux/mempool.h>
19#include <linux/module.h>
20#include <linux/rwsem.h>
21#include <linux/slab.h>
22#include <linux/vmalloc.h>
23
24#define DM_MSG_PREFIX "cache"
25
26DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
27 "A percentage of time allocated for copying to and/or from cache");
28
29/*----------------------------------------------------------------*/
30
31/*
32 * Glossary:
33 *
34 * oblock: index of an origin block
35 * cblock: index of a cache block
36 * promotion: movement of a block from origin to cache
37 * demotion: movement of a block from cache to origin
38 * migration: movement of a block between the origin and cache device,
39 * either direction
40 */
41
42/*----------------------------------------------------------------*/
43
44/*
45 * Represents a chunk of future work. 'input' allows continuations to pass
46 * values between themselves, typically error values.
47 */
48struct continuation {
49 struct work_struct ws;
50 blk_status_t input;
51};
52
53static inline void init_continuation(struct continuation *k,
54 void (*fn)(struct work_struct *))
55{
56 INIT_WORK(&k->ws, fn);
57 k->input = 0;
58}
59
60static inline void queue_continuation(struct workqueue_struct *wq,
61 struct continuation *k)
62{
63 queue_work(wq, &k->ws);
64}
65
66/*----------------------------------------------------------------*/
67
68/*
69 * The batcher collects together pieces of work that need a particular
70 * operation to occur before they can proceed (typically a commit).
71 */
72struct batcher {
73 /*
74 * The operation that everyone is waiting for.
75 */
76 blk_status_t (*commit_op)(void *context);
77 void *commit_context;
78
79 /*
80 * This is how bios should be issued once the commit op is complete
81 * (accounted_request).
82 */
83 void (*issue_op)(struct bio *bio, void *context);
84 void *issue_context;
85
86 /*
87 * Queued work gets put on here after commit.
88 */
89 struct workqueue_struct *wq;
90
91 spinlock_t lock;
92 struct list_head work_items;
93 struct bio_list bios;
94 struct work_struct commit_work;
95
96 bool commit_scheduled;
97};
98
99static void __commit(struct work_struct *_ws)
100{
101 struct batcher *b = container_of(_ws, struct batcher, commit_work);
102 blk_status_t r;
103 struct list_head work_items;
104 struct work_struct *ws, *tmp;
105 struct continuation *k;
106 struct bio *bio;
107 struct bio_list bios;
108
109 INIT_LIST_HEAD(&work_items);
110 bio_list_init(&bios);
111
112 /*
113 * We have to grab these before the commit_op to avoid a race
114 * condition.
115 */
116 spin_lock_irq(&b->lock);
117 list_splice_init(&b->work_items, &work_items);
118 bio_list_merge(&bios, &b->bios);
119 bio_list_init(&b->bios);
120 b->commit_scheduled = false;
121 spin_unlock_irq(&b->lock);
122
123 r = b->commit_op(b->commit_context);
124
125 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
126 k = container_of(ws, struct continuation, ws);
127 k->input = r;
128 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
129 queue_work(b->wq, ws);
130 }
131
132 while ((bio = bio_list_pop(&bios))) {
133 if (r) {
134 bio->bi_status = r;
135 bio_endio(bio);
136 } else
137 b->issue_op(bio, b->issue_context);
138 }
139}
140
141static void batcher_init(struct batcher *b,
142 blk_status_t (*commit_op)(void *),
143 void *commit_context,
144 void (*issue_op)(struct bio *bio, void *),
145 void *issue_context,
146 struct workqueue_struct *wq)
147{
148 b->commit_op = commit_op;
149 b->commit_context = commit_context;
150 b->issue_op = issue_op;
151 b->issue_context = issue_context;
152 b->wq = wq;
153
154 spin_lock_init(&b->lock);
155 INIT_LIST_HEAD(&b->work_items);
156 bio_list_init(&b->bios);
157 INIT_WORK(&b->commit_work, __commit);
158 b->commit_scheduled = false;
159}
160
161static void async_commit(struct batcher *b)
162{
163 queue_work(b->wq, &b->commit_work);
164}
165
166static void continue_after_commit(struct batcher *b, struct continuation *k)
167{
168 bool commit_scheduled;
169
170 spin_lock_irq(&b->lock);
171 commit_scheduled = b->commit_scheduled;
172 list_add_tail(&k->ws.entry, &b->work_items);
173 spin_unlock_irq(&b->lock);
174
175 if (commit_scheduled)
176 async_commit(b);
177}
178
179/*
180 * Bios are errored if commit failed.
181 */
182static void issue_after_commit(struct batcher *b, struct bio *bio)
183{
184 bool commit_scheduled;
185
186 spin_lock_irq(&b->lock);
187 commit_scheduled = b->commit_scheduled;
188 bio_list_add(&b->bios, bio);
189 spin_unlock_irq(&b->lock);
190
191 if (commit_scheduled)
192 async_commit(b);
193}
194
195/*
196 * Call this if some urgent work is waiting for the commit to complete.
197 */
198static void schedule_commit(struct batcher *b)
199{
200 bool immediate;
201
202 spin_lock_irq(&b->lock);
203 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
204 b->commit_scheduled = true;
205 spin_unlock_irq(&b->lock);
206
207 if (immediate)
208 async_commit(b);
209}
210
211/*
212 * There are a couple of places where we let a bio run, but want to do some
213 * work before calling its endio function. We do this by temporarily
214 * changing the endio fn.
215 */
216struct dm_hook_info {
217 bio_end_io_t *bi_end_io;
218};
219
220static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
221 bio_end_io_t *bi_end_io, void *bi_private)
222{
223 h->bi_end_io = bio->bi_end_io;
224
225 bio->bi_end_io = bi_end_io;
226 bio->bi_private = bi_private;
227}
228
229static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
230{
231 bio->bi_end_io = h->bi_end_io;
232}
233
234/*----------------------------------------------------------------*/
235
236#define MIGRATION_POOL_SIZE 128
237#define COMMIT_PERIOD HZ
238#define MIGRATION_COUNT_WINDOW 10
239
240/*
241 * The block size of the device holding cache data must be
242 * between 32KB and 1GB.
243 */
244#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
245#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
246
247enum cache_metadata_mode {
248 CM_WRITE, /* metadata may be changed */
249 CM_READ_ONLY, /* metadata may not be changed */
250 CM_FAIL
251};
252
253enum cache_io_mode {
254 /*
255 * Data is written to cached blocks only. These blocks are marked
256 * dirty. If you lose the cache device you will lose data.
257 * Potential performance increase for both reads and writes.
258 */
259 CM_IO_WRITEBACK,
260
261 /*
262 * Data is written to both cache and origin. Blocks are never
263 * dirty. Potential performance benfit for reads only.
264 */
265 CM_IO_WRITETHROUGH,
266
267 /*
268 * A degraded mode useful for various cache coherency situations
269 * (eg, rolling back snapshots). Reads and writes always go to the
270 * origin. If a write goes to a cached oblock, then the cache
271 * block is invalidated.
272 */
273 CM_IO_PASSTHROUGH
274};
275
276struct cache_features {
277 enum cache_metadata_mode mode;
278 enum cache_io_mode io_mode;
279 unsigned int metadata_version;
280 bool discard_passdown:1;
281};
282
283struct cache_stats {
284 atomic_t read_hit;
285 atomic_t read_miss;
286 atomic_t write_hit;
287 atomic_t write_miss;
288 atomic_t demotion;
289 atomic_t promotion;
290 atomic_t writeback;
291 atomic_t copies_avoided;
292 atomic_t cache_cell_clash;
293 atomic_t commit_count;
294 atomic_t discard_count;
295};
296
297struct cache {
298 struct dm_target *ti;
299 spinlock_t lock;
300
301 /*
302 * Fields for converting from sectors to blocks.
303 */
304 int sectors_per_block_shift;
305 sector_t sectors_per_block;
306
307 struct dm_cache_metadata *cmd;
308
309 /*
310 * Metadata is written to this device.
311 */
312 struct dm_dev *metadata_dev;
313
314 /*
315 * The slower of the two data devices. Typically a spindle.
316 */
317 struct dm_dev *origin_dev;
318
319 /*
320 * The faster of the two data devices. Typically an SSD.
321 */
322 struct dm_dev *cache_dev;
323
324 /*
325 * Size of the origin device in _complete_ blocks and native sectors.
326 */
327 dm_oblock_t origin_blocks;
328 sector_t origin_sectors;
329
330 /*
331 * Size of the cache device in blocks.
332 */
333 dm_cblock_t cache_size;
334
335 /*
336 * Invalidation fields.
337 */
338 spinlock_t invalidation_lock;
339 struct list_head invalidation_requests;
340
341 sector_t migration_threshold;
342 wait_queue_head_t migration_wait;
343 atomic_t nr_allocated_migrations;
344
345 /*
346 * The number of in flight migrations that are performing
347 * background io. eg, promotion, writeback.
348 */
349 atomic_t nr_io_migrations;
350
351 struct bio_list deferred_bios;
352
353 struct rw_semaphore quiesce_lock;
354
355 /*
356 * origin_blocks entries, discarded if set.
357 */
358 dm_dblock_t discard_nr_blocks;
359 unsigned long *discard_bitset;
360 uint32_t discard_block_size; /* a power of 2 times sectors per block */
361
362 /*
363 * Rather than reconstructing the table line for the status we just
364 * save it and regurgitate.
365 */
366 unsigned int nr_ctr_args;
367 const char **ctr_args;
368
369 struct dm_kcopyd_client *copier;
370 struct work_struct deferred_bio_worker;
371 struct work_struct migration_worker;
372 struct workqueue_struct *wq;
373 struct delayed_work waker;
374 struct dm_bio_prison_v2 *prison;
375
376 /*
377 * cache_size entries, dirty if set
378 */
379 unsigned long *dirty_bitset;
380 atomic_t nr_dirty;
381
382 unsigned int policy_nr_args;
383 struct dm_cache_policy *policy;
384
385 /*
386 * Cache features such as write-through.
387 */
388 struct cache_features features;
389
390 struct cache_stats stats;
391
392 bool need_tick_bio:1;
393 bool sized:1;
394 bool invalidate:1;
395 bool commit_requested:1;
396 bool loaded_mappings:1;
397 bool loaded_discards:1;
398
399 struct rw_semaphore background_work_lock;
400
401 struct batcher committer;
402 struct work_struct commit_ws;
403
404 struct dm_io_tracker tracker;
405
406 mempool_t migration_pool;
407
408 struct bio_set bs;
409};
410
411struct per_bio_data {
412 bool tick:1;
413 unsigned int req_nr:2;
414 struct dm_bio_prison_cell_v2 *cell;
415 struct dm_hook_info hook_info;
416 sector_t len;
417};
418
419struct dm_cache_migration {
420 struct continuation k;
421 struct cache *cache;
422
423 struct policy_work *op;
424 struct bio *overwrite_bio;
425 struct dm_bio_prison_cell_v2 *cell;
426
427 dm_cblock_t invalidate_cblock;
428 dm_oblock_t invalidate_oblock;
429};
430
431/*----------------------------------------------------------------*/
432
433static bool writethrough_mode(struct cache *cache)
434{
435 return cache->features.io_mode == CM_IO_WRITETHROUGH;
436}
437
438static bool writeback_mode(struct cache *cache)
439{
440 return cache->features.io_mode == CM_IO_WRITEBACK;
441}
442
443static inline bool passthrough_mode(struct cache *cache)
444{
445 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
446}
447
448/*----------------------------------------------------------------*/
449
450static void wake_deferred_bio_worker(struct cache *cache)
451{
452 queue_work(cache->wq, &cache->deferred_bio_worker);
453}
454
455static void wake_migration_worker(struct cache *cache)
456{
457 if (passthrough_mode(cache))
458 return;
459
460 queue_work(cache->wq, &cache->migration_worker);
461}
462
463/*----------------------------------------------------------------*/
464
465static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
466{
467 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
468}
469
470static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
471{
472 dm_bio_prison_free_cell_v2(cache->prison, cell);
473}
474
475static struct dm_cache_migration *alloc_migration(struct cache *cache)
476{
477 struct dm_cache_migration *mg;
478
479 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
480
481 memset(mg, 0, sizeof(*mg));
482
483 mg->cache = cache;
484 atomic_inc(&cache->nr_allocated_migrations);
485
486 return mg;
487}
488
489static void free_migration(struct dm_cache_migration *mg)
490{
491 struct cache *cache = mg->cache;
492
493 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
494 wake_up(&cache->migration_wait);
495
496 mempool_free(mg, &cache->migration_pool);
497}
498
499/*----------------------------------------------------------------*/
500
501static inline dm_oblock_t oblock_succ(dm_oblock_t b)
502{
503 return to_oblock(from_oblock(b) + 1ull);
504}
505
506static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
507{
508 key->virtual = 0;
509 key->dev = 0;
510 key->block_begin = from_oblock(begin);
511 key->block_end = from_oblock(end);
512}
513
514/*
515 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
516 * level 1 which prevents *both* READs and WRITEs.
517 */
518#define WRITE_LOCK_LEVEL 0
519#define READ_WRITE_LOCK_LEVEL 1
520
521static unsigned int lock_level(struct bio *bio)
522{
523 return bio_data_dir(bio) == WRITE ?
524 WRITE_LOCK_LEVEL :
525 READ_WRITE_LOCK_LEVEL;
526}
527
528/*
529 *--------------------------------------------------------------
530 * Per bio data
531 *--------------------------------------------------------------
532 */
533
534static struct per_bio_data *get_per_bio_data(struct bio *bio)
535{
536 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
537
538 BUG_ON(!pb);
539 return pb;
540}
541
542static struct per_bio_data *init_per_bio_data(struct bio *bio)
543{
544 struct per_bio_data *pb = get_per_bio_data(bio);
545
546 pb->tick = false;
547 pb->req_nr = dm_bio_get_target_bio_nr(bio);
548 pb->cell = NULL;
549 pb->len = 0;
550
551 return pb;
552}
553
554/*----------------------------------------------------------------*/
555
556static void defer_bio(struct cache *cache, struct bio *bio)
557{
558 spin_lock_irq(&cache->lock);
559 bio_list_add(&cache->deferred_bios, bio);
560 spin_unlock_irq(&cache->lock);
561
562 wake_deferred_bio_worker(cache);
563}
564
565static void defer_bios(struct cache *cache, struct bio_list *bios)
566{
567 spin_lock_irq(&cache->lock);
568 bio_list_merge(&cache->deferred_bios, bios);
569 bio_list_init(bios);
570 spin_unlock_irq(&cache->lock);
571
572 wake_deferred_bio_worker(cache);
573}
574
575/*----------------------------------------------------------------*/
576
577static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
578{
579 bool r;
580 struct per_bio_data *pb;
581 struct dm_cell_key_v2 key;
582 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
583 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
584
585 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
586
587 build_key(oblock, end, &key);
588 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
589 if (!r) {
590 /*
591 * Failed to get the lock.
592 */
593 free_prison_cell(cache, cell_prealloc);
594 return r;
595 }
596
597 if (cell != cell_prealloc)
598 free_prison_cell(cache, cell_prealloc);
599
600 pb = get_per_bio_data(bio);
601 pb->cell = cell;
602
603 return r;
604}
605
606/*----------------------------------------------------------------*/
607
608static bool is_dirty(struct cache *cache, dm_cblock_t b)
609{
610 return test_bit(from_cblock(b), cache->dirty_bitset);
611}
612
613static void set_dirty(struct cache *cache, dm_cblock_t cblock)
614{
615 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
616 atomic_inc(&cache->nr_dirty);
617 policy_set_dirty(cache->policy, cblock);
618 }
619}
620
621/*
622 * These two are called when setting after migrations to force the policy
623 * and dirty bitset to be in sync.
624 */
625static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
626{
627 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
628 atomic_inc(&cache->nr_dirty);
629 policy_set_dirty(cache->policy, cblock);
630}
631
632static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
633{
634 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
635 if (atomic_dec_return(&cache->nr_dirty) == 0)
636 dm_table_event(cache->ti->table);
637 }
638
639 policy_clear_dirty(cache->policy, cblock);
640}
641
642/*----------------------------------------------------------------*/
643
644static bool block_size_is_power_of_two(struct cache *cache)
645{
646 return cache->sectors_per_block_shift >= 0;
647}
648
649static dm_block_t block_div(dm_block_t b, uint32_t n)
650{
651 do_div(b, n);
652
653 return b;
654}
655
656static dm_block_t oblocks_per_dblock(struct cache *cache)
657{
658 dm_block_t oblocks = cache->discard_block_size;
659
660 if (block_size_is_power_of_two(cache))
661 oblocks >>= cache->sectors_per_block_shift;
662 else
663 oblocks = block_div(oblocks, cache->sectors_per_block);
664
665 return oblocks;
666}
667
668static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
669{
670 return to_dblock(block_div(from_oblock(oblock),
671 oblocks_per_dblock(cache)));
672}
673
674static void set_discard(struct cache *cache, dm_dblock_t b)
675{
676 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
677 atomic_inc(&cache->stats.discard_count);
678
679 spin_lock_irq(&cache->lock);
680 set_bit(from_dblock(b), cache->discard_bitset);
681 spin_unlock_irq(&cache->lock);
682}
683
684static void clear_discard(struct cache *cache, dm_dblock_t b)
685{
686 spin_lock_irq(&cache->lock);
687 clear_bit(from_dblock(b), cache->discard_bitset);
688 spin_unlock_irq(&cache->lock);
689}
690
691static bool is_discarded(struct cache *cache, dm_dblock_t b)
692{
693 int r;
694
695 spin_lock_irq(&cache->lock);
696 r = test_bit(from_dblock(b), cache->discard_bitset);
697 spin_unlock_irq(&cache->lock);
698
699 return r;
700}
701
702static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
703{
704 int r;
705
706 spin_lock_irq(&cache->lock);
707 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
708 cache->discard_bitset);
709 spin_unlock_irq(&cache->lock);
710
711 return r;
712}
713
714/*
715 * -------------------------------------------------------------
716 * Remapping
717 *--------------------------------------------------------------
718 */
719static void remap_to_origin(struct cache *cache, struct bio *bio)
720{
721 bio_set_dev(bio, cache->origin_dev->bdev);
722}
723
724static void remap_to_cache(struct cache *cache, struct bio *bio,
725 dm_cblock_t cblock)
726{
727 sector_t bi_sector = bio->bi_iter.bi_sector;
728 sector_t block = from_cblock(cblock);
729
730 bio_set_dev(bio, cache->cache_dev->bdev);
731 if (!block_size_is_power_of_two(cache))
732 bio->bi_iter.bi_sector =
733 (block * cache->sectors_per_block) +
734 sector_div(bi_sector, cache->sectors_per_block);
735 else
736 bio->bi_iter.bi_sector =
737 (block << cache->sectors_per_block_shift) |
738 (bi_sector & (cache->sectors_per_block - 1));
739}
740
741static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
742{
743 struct per_bio_data *pb;
744
745 spin_lock_irq(&cache->lock);
746 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
747 bio_op(bio) != REQ_OP_DISCARD) {
748 pb = get_per_bio_data(bio);
749 pb->tick = true;
750 cache->need_tick_bio = false;
751 }
752 spin_unlock_irq(&cache->lock);
753}
754
755static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
756 dm_oblock_t oblock)
757{
758 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
759 check_if_tick_bio_needed(cache, bio);
760 remap_to_origin(cache, bio);
761 if (bio_data_dir(bio) == WRITE)
762 clear_discard(cache, oblock_to_dblock(cache, oblock));
763}
764
765static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
766 dm_oblock_t oblock, dm_cblock_t cblock)
767{
768 check_if_tick_bio_needed(cache, bio);
769 remap_to_cache(cache, bio, cblock);
770 if (bio_data_dir(bio) == WRITE) {
771 set_dirty(cache, cblock);
772 clear_discard(cache, oblock_to_dblock(cache, oblock));
773 }
774}
775
776static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
777{
778 sector_t block_nr = bio->bi_iter.bi_sector;
779
780 if (!block_size_is_power_of_two(cache))
781 (void) sector_div(block_nr, cache->sectors_per_block);
782 else
783 block_nr >>= cache->sectors_per_block_shift;
784
785 return to_oblock(block_nr);
786}
787
788static bool accountable_bio(struct cache *cache, struct bio *bio)
789{
790 return bio_op(bio) != REQ_OP_DISCARD;
791}
792
793static void accounted_begin(struct cache *cache, struct bio *bio)
794{
795 struct per_bio_data *pb;
796
797 if (accountable_bio(cache, bio)) {
798 pb = get_per_bio_data(bio);
799 pb->len = bio_sectors(bio);
800 dm_iot_io_begin(&cache->tracker, pb->len);
801 }
802}
803
804static void accounted_complete(struct cache *cache, struct bio *bio)
805{
806 struct per_bio_data *pb = get_per_bio_data(bio);
807
808 dm_iot_io_end(&cache->tracker, pb->len);
809}
810
811static void accounted_request(struct cache *cache, struct bio *bio)
812{
813 accounted_begin(cache, bio);
814 dm_submit_bio_remap(bio, NULL);
815}
816
817static void issue_op(struct bio *bio, void *context)
818{
819 struct cache *cache = context;
820
821 accounted_request(cache, bio);
822}
823
824/*
825 * When running in writethrough mode we need to send writes to clean blocks
826 * to both the cache and origin devices. Clone the bio and send them in parallel.
827 */
828static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
829 dm_oblock_t oblock, dm_cblock_t cblock)
830{
831 struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
832 GFP_NOIO, &cache->bs);
833
834 BUG_ON(!origin_bio);
835
836 bio_chain(origin_bio, bio);
837
838 if (bio_data_dir(origin_bio) == WRITE)
839 clear_discard(cache, oblock_to_dblock(cache, oblock));
840 submit_bio(origin_bio);
841
842 remap_to_cache(cache, bio, cblock);
843}
844
845/*
846 *--------------------------------------------------------------
847 * Failure modes
848 *--------------------------------------------------------------
849 */
850static enum cache_metadata_mode get_cache_mode(struct cache *cache)
851{
852 return cache->features.mode;
853}
854
855static const char *cache_device_name(struct cache *cache)
856{
857 return dm_table_device_name(cache->ti->table);
858}
859
860static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
861{
862 static const char *descs[] = {
863 "write",
864 "read-only",
865 "fail"
866 };
867
868 dm_table_event(cache->ti->table);
869 DMINFO("%s: switching cache to %s mode",
870 cache_device_name(cache), descs[(int)mode]);
871}
872
873static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
874{
875 bool needs_check;
876 enum cache_metadata_mode old_mode = get_cache_mode(cache);
877
878 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
879 DMERR("%s: unable to read needs_check flag, setting failure mode.",
880 cache_device_name(cache));
881 new_mode = CM_FAIL;
882 }
883
884 if (new_mode == CM_WRITE && needs_check) {
885 DMERR("%s: unable to switch cache to write mode until repaired.",
886 cache_device_name(cache));
887 if (old_mode != new_mode)
888 new_mode = old_mode;
889 else
890 new_mode = CM_READ_ONLY;
891 }
892
893 /* Never move out of fail mode */
894 if (old_mode == CM_FAIL)
895 new_mode = CM_FAIL;
896
897 switch (new_mode) {
898 case CM_FAIL:
899 case CM_READ_ONLY:
900 dm_cache_metadata_set_read_only(cache->cmd);
901 break;
902
903 case CM_WRITE:
904 dm_cache_metadata_set_read_write(cache->cmd);
905 break;
906 }
907
908 cache->features.mode = new_mode;
909
910 if (new_mode != old_mode)
911 notify_mode_switch(cache, new_mode);
912}
913
914static void abort_transaction(struct cache *cache)
915{
916 const char *dev_name = cache_device_name(cache);
917
918 if (get_cache_mode(cache) >= CM_READ_ONLY)
919 return;
920
921 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
922 if (dm_cache_metadata_abort(cache->cmd)) {
923 DMERR("%s: failed to abort metadata transaction", dev_name);
924 set_cache_mode(cache, CM_FAIL);
925 }
926
927 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
928 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
929 set_cache_mode(cache, CM_FAIL);
930 }
931}
932
933static void metadata_operation_failed(struct cache *cache, const char *op, int r)
934{
935 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
936 cache_device_name(cache), op, r);
937 abort_transaction(cache);
938 set_cache_mode(cache, CM_READ_ONLY);
939}
940
941/*----------------------------------------------------------------*/
942
943static void load_stats(struct cache *cache)
944{
945 struct dm_cache_statistics stats;
946
947 dm_cache_metadata_get_stats(cache->cmd, &stats);
948 atomic_set(&cache->stats.read_hit, stats.read_hits);
949 atomic_set(&cache->stats.read_miss, stats.read_misses);
950 atomic_set(&cache->stats.write_hit, stats.write_hits);
951 atomic_set(&cache->stats.write_miss, stats.write_misses);
952}
953
954static void save_stats(struct cache *cache)
955{
956 struct dm_cache_statistics stats;
957
958 if (get_cache_mode(cache) >= CM_READ_ONLY)
959 return;
960
961 stats.read_hits = atomic_read(&cache->stats.read_hit);
962 stats.read_misses = atomic_read(&cache->stats.read_miss);
963 stats.write_hits = atomic_read(&cache->stats.write_hit);
964 stats.write_misses = atomic_read(&cache->stats.write_miss);
965
966 dm_cache_metadata_set_stats(cache->cmd, &stats);
967}
968
969static void update_stats(struct cache_stats *stats, enum policy_operation op)
970{
971 switch (op) {
972 case POLICY_PROMOTE:
973 atomic_inc(&stats->promotion);
974 break;
975
976 case POLICY_DEMOTE:
977 atomic_inc(&stats->demotion);
978 break;
979
980 case POLICY_WRITEBACK:
981 atomic_inc(&stats->writeback);
982 break;
983 }
984}
985
986/*
987 *---------------------------------------------------------------------
988 * Migration processing
989 *
990 * Migration covers moving data from the origin device to the cache, or
991 * vice versa.
992 *---------------------------------------------------------------------
993 */
994static void inc_io_migrations(struct cache *cache)
995{
996 atomic_inc(&cache->nr_io_migrations);
997}
998
999static void dec_io_migrations(struct cache *cache)
1000{
1001 atomic_dec(&cache->nr_io_migrations);
1002}
1003
1004static bool discard_or_flush(struct bio *bio)
1005{
1006 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1007}
1008
1009static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1010 dm_dblock_t *b, dm_dblock_t *e)
1011{
1012 sector_t sb = bio->bi_iter.bi_sector;
1013 sector_t se = bio_end_sector(bio);
1014
1015 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1016
1017 if (se - sb < cache->discard_block_size)
1018 *e = *b;
1019 else
1020 *e = to_dblock(block_div(se, cache->discard_block_size));
1021}
1022
1023/*----------------------------------------------------------------*/
1024
1025static void prevent_background_work(struct cache *cache)
1026{
1027 lockdep_off();
1028 down_write(&cache->background_work_lock);
1029 lockdep_on();
1030}
1031
1032static void allow_background_work(struct cache *cache)
1033{
1034 lockdep_off();
1035 up_write(&cache->background_work_lock);
1036 lockdep_on();
1037}
1038
1039static bool background_work_begin(struct cache *cache)
1040{
1041 bool r;
1042
1043 lockdep_off();
1044 r = down_read_trylock(&cache->background_work_lock);
1045 lockdep_on();
1046
1047 return r;
1048}
1049
1050static void background_work_end(struct cache *cache)
1051{
1052 lockdep_off();
1053 up_read(&cache->background_work_lock);
1054 lockdep_on();
1055}
1056
1057/*----------------------------------------------------------------*/
1058
1059static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1060{
1061 return (bio_data_dir(bio) == WRITE) &&
1062 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1063}
1064
1065static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1066{
1067 return writeback_mode(cache) &&
1068 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1069}
1070
1071static void quiesce(struct dm_cache_migration *mg,
1072 void (*continuation)(struct work_struct *))
1073{
1074 init_continuation(&mg->k, continuation);
1075 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1076}
1077
1078static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1079{
1080 struct continuation *k = container_of(ws, struct continuation, ws);
1081
1082 return container_of(k, struct dm_cache_migration, k);
1083}
1084
1085static void copy_complete(int read_err, unsigned long write_err, void *context)
1086{
1087 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1088
1089 if (read_err || write_err)
1090 mg->k.input = BLK_STS_IOERR;
1091
1092 queue_continuation(mg->cache->wq, &mg->k);
1093}
1094
1095static void copy(struct dm_cache_migration *mg, bool promote)
1096{
1097 struct dm_io_region o_region, c_region;
1098 struct cache *cache = mg->cache;
1099
1100 o_region.bdev = cache->origin_dev->bdev;
1101 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1102 o_region.count = cache->sectors_per_block;
1103
1104 c_region.bdev = cache->cache_dev->bdev;
1105 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1106 c_region.count = cache->sectors_per_block;
1107
1108 if (promote)
1109 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1110 else
1111 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1112}
1113
1114static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1115{
1116 struct per_bio_data *pb = get_per_bio_data(bio);
1117
1118 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1119 free_prison_cell(cache, pb->cell);
1120 pb->cell = NULL;
1121}
1122
1123static void overwrite_endio(struct bio *bio)
1124{
1125 struct dm_cache_migration *mg = bio->bi_private;
1126 struct cache *cache = mg->cache;
1127 struct per_bio_data *pb = get_per_bio_data(bio);
1128
1129 dm_unhook_bio(&pb->hook_info, bio);
1130
1131 if (bio->bi_status)
1132 mg->k.input = bio->bi_status;
1133
1134 queue_continuation(cache->wq, &mg->k);
1135}
1136
1137static void overwrite(struct dm_cache_migration *mg,
1138 void (*continuation)(struct work_struct *))
1139{
1140 struct bio *bio = mg->overwrite_bio;
1141 struct per_bio_data *pb = get_per_bio_data(bio);
1142
1143 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1144
1145 /*
1146 * The overwrite bio is part of the copy operation, as such it does
1147 * not set/clear discard or dirty flags.
1148 */
1149 if (mg->op->op == POLICY_PROMOTE)
1150 remap_to_cache(mg->cache, bio, mg->op->cblock);
1151 else
1152 remap_to_origin(mg->cache, bio);
1153
1154 init_continuation(&mg->k, continuation);
1155 accounted_request(mg->cache, bio);
1156}
1157
1158/*
1159 * Migration steps:
1160 *
1161 * 1) exclusive lock preventing WRITEs
1162 * 2) quiesce
1163 * 3) copy or issue overwrite bio
1164 * 4) upgrade to exclusive lock preventing READs and WRITEs
1165 * 5) quiesce
1166 * 6) update metadata and commit
1167 * 7) unlock
1168 */
1169static void mg_complete(struct dm_cache_migration *mg, bool success)
1170{
1171 struct bio_list bios;
1172 struct cache *cache = mg->cache;
1173 struct policy_work *op = mg->op;
1174 dm_cblock_t cblock = op->cblock;
1175
1176 if (success)
1177 update_stats(&cache->stats, op->op);
1178
1179 switch (op->op) {
1180 case POLICY_PROMOTE:
1181 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1182 policy_complete_background_work(cache->policy, op, success);
1183
1184 if (mg->overwrite_bio) {
1185 if (success)
1186 force_set_dirty(cache, cblock);
1187 else if (mg->k.input)
1188 mg->overwrite_bio->bi_status = mg->k.input;
1189 else
1190 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1191 bio_endio(mg->overwrite_bio);
1192 } else {
1193 if (success)
1194 force_clear_dirty(cache, cblock);
1195 dec_io_migrations(cache);
1196 }
1197 break;
1198
1199 case POLICY_DEMOTE:
1200 /*
1201 * We clear dirty here to update the nr_dirty counter.
1202 */
1203 if (success)
1204 force_clear_dirty(cache, cblock);
1205 policy_complete_background_work(cache->policy, op, success);
1206 dec_io_migrations(cache);
1207 break;
1208
1209 case POLICY_WRITEBACK:
1210 if (success)
1211 force_clear_dirty(cache, cblock);
1212 policy_complete_background_work(cache->policy, op, success);
1213 dec_io_migrations(cache);
1214 break;
1215 }
1216
1217 bio_list_init(&bios);
1218 if (mg->cell) {
1219 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1220 free_prison_cell(cache, mg->cell);
1221 }
1222
1223 free_migration(mg);
1224 defer_bios(cache, &bios);
1225 wake_migration_worker(cache);
1226
1227 background_work_end(cache);
1228}
1229
1230static void mg_success(struct work_struct *ws)
1231{
1232 struct dm_cache_migration *mg = ws_to_mg(ws);
1233
1234 mg_complete(mg, mg->k.input == 0);
1235}
1236
1237static void mg_update_metadata(struct work_struct *ws)
1238{
1239 int r;
1240 struct dm_cache_migration *mg = ws_to_mg(ws);
1241 struct cache *cache = mg->cache;
1242 struct policy_work *op = mg->op;
1243
1244 switch (op->op) {
1245 case POLICY_PROMOTE:
1246 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1247 if (r) {
1248 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1249 cache_device_name(cache));
1250 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1251
1252 mg_complete(mg, false);
1253 return;
1254 }
1255 mg_complete(mg, true);
1256 break;
1257
1258 case POLICY_DEMOTE:
1259 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1260 if (r) {
1261 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1262 cache_device_name(cache));
1263 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1264
1265 mg_complete(mg, false);
1266 return;
1267 }
1268
1269 /*
1270 * It would be nice if we only had to commit when a REQ_FLUSH
1271 * comes through. But there's one scenario that we have to
1272 * look out for:
1273 *
1274 * - vblock x in a cache block
1275 * - domotion occurs
1276 * - cache block gets reallocated and over written
1277 * - crash
1278 *
1279 * When we recover, because there was no commit the cache will
1280 * rollback to having the data for vblock x in the cache block.
1281 * But the cache block has since been overwritten, so it'll end
1282 * up pointing to data that was never in 'x' during the history
1283 * of the device.
1284 *
1285 * To avoid this issue we require a commit as part of the
1286 * demotion operation.
1287 */
1288 init_continuation(&mg->k, mg_success);
1289 continue_after_commit(&cache->committer, &mg->k);
1290 schedule_commit(&cache->committer);
1291 break;
1292
1293 case POLICY_WRITEBACK:
1294 mg_complete(mg, true);
1295 break;
1296 }
1297}
1298
1299static void mg_update_metadata_after_copy(struct work_struct *ws)
1300{
1301 struct dm_cache_migration *mg = ws_to_mg(ws);
1302
1303 /*
1304 * Did the copy succeed?
1305 */
1306 if (mg->k.input)
1307 mg_complete(mg, false);
1308 else
1309 mg_update_metadata(ws);
1310}
1311
1312static void mg_upgrade_lock(struct work_struct *ws)
1313{
1314 int r;
1315 struct dm_cache_migration *mg = ws_to_mg(ws);
1316
1317 /*
1318 * Did the copy succeed?
1319 */
1320 if (mg->k.input)
1321 mg_complete(mg, false);
1322
1323 else {
1324 /*
1325 * Now we want the lock to prevent both reads and writes.
1326 */
1327 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1328 READ_WRITE_LOCK_LEVEL);
1329 if (r < 0)
1330 mg_complete(mg, false);
1331
1332 else if (r)
1333 quiesce(mg, mg_update_metadata);
1334
1335 else
1336 mg_update_metadata(ws);
1337 }
1338}
1339
1340static void mg_full_copy(struct work_struct *ws)
1341{
1342 struct dm_cache_migration *mg = ws_to_mg(ws);
1343 struct cache *cache = mg->cache;
1344 struct policy_work *op = mg->op;
1345 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1346
1347 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1348 is_discarded_oblock(cache, op->oblock)) {
1349 mg_upgrade_lock(ws);
1350 return;
1351 }
1352
1353 init_continuation(&mg->k, mg_upgrade_lock);
1354 copy(mg, is_policy_promote);
1355}
1356
1357static void mg_copy(struct work_struct *ws)
1358{
1359 struct dm_cache_migration *mg = ws_to_mg(ws);
1360
1361 if (mg->overwrite_bio) {
1362 /*
1363 * No exclusive lock was held when we last checked if the bio
1364 * was optimisable. So we have to check again in case things
1365 * have changed (eg, the block may no longer be discarded).
1366 */
1367 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1368 /*
1369 * Fallback to a real full copy after doing some tidying up.
1370 */
1371 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1372
1373 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1374 mg->overwrite_bio = NULL;
1375 inc_io_migrations(mg->cache);
1376 mg_full_copy(ws);
1377 return;
1378 }
1379
1380 /*
1381 * It's safe to do this here, even though it's new data
1382 * because all IO has been locked out of the block.
1383 *
1384 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1385 * so _not_ using mg_upgrade_lock() as continutation.
1386 */
1387 overwrite(mg, mg_update_metadata_after_copy);
1388
1389 } else
1390 mg_full_copy(ws);
1391}
1392
1393static int mg_lock_writes(struct dm_cache_migration *mg)
1394{
1395 int r;
1396 struct dm_cell_key_v2 key;
1397 struct cache *cache = mg->cache;
1398 struct dm_bio_prison_cell_v2 *prealloc;
1399
1400 prealloc = alloc_prison_cell(cache);
1401
1402 /*
1403 * Prevent writes to the block, but allow reads to continue.
1404 * Unless we're using an overwrite bio, in which case we lock
1405 * everything.
1406 */
1407 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1408 r = dm_cell_lock_v2(cache->prison, &key,
1409 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1410 prealloc, &mg->cell);
1411 if (r < 0) {
1412 free_prison_cell(cache, prealloc);
1413 mg_complete(mg, false);
1414 return r;
1415 }
1416
1417 if (mg->cell != prealloc)
1418 free_prison_cell(cache, prealloc);
1419
1420 if (r == 0)
1421 mg_copy(&mg->k.ws);
1422 else
1423 quiesce(mg, mg_copy);
1424
1425 return 0;
1426}
1427
1428static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1429{
1430 struct dm_cache_migration *mg;
1431
1432 if (!background_work_begin(cache)) {
1433 policy_complete_background_work(cache->policy, op, false);
1434 return -EPERM;
1435 }
1436
1437 mg = alloc_migration(cache);
1438
1439 mg->op = op;
1440 mg->overwrite_bio = bio;
1441
1442 if (!bio)
1443 inc_io_migrations(cache);
1444
1445 return mg_lock_writes(mg);
1446}
1447
1448/*
1449 *--------------------------------------------------------------
1450 * invalidation processing
1451 *--------------------------------------------------------------
1452 */
1453
1454static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1455{
1456 struct bio_list bios;
1457 struct cache *cache = mg->cache;
1458
1459 bio_list_init(&bios);
1460 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1461 free_prison_cell(cache, mg->cell);
1462
1463 if (!success && mg->overwrite_bio)
1464 bio_io_error(mg->overwrite_bio);
1465
1466 free_migration(mg);
1467 defer_bios(cache, &bios);
1468
1469 background_work_end(cache);
1470}
1471
1472static void invalidate_completed(struct work_struct *ws)
1473{
1474 struct dm_cache_migration *mg = ws_to_mg(ws);
1475
1476 invalidate_complete(mg, !mg->k.input);
1477}
1478
1479static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1480{
1481 int r;
1482
1483 r = policy_invalidate_mapping(cache->policy, cblock);
1484 if (!r) {
1485 r = dm_cache_remove_mapping(cache->cmd, cblock);
1486 if (r) {
1487 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1488 cache_device_name(cache));
1489 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1490 }
1491
1492 } else if (r == -ENODATA) {
1493 /*
1494 * Harmless, already unmapped.
1495 */
1496 r = 0;
1497
1498 } else
1499 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1500
1501 return r;
1502}
1503
1504static void invalidate_remove(struct work_struct *ws)
1505{
1506 int r;
1507 struct dm_cache_migration *mg = ws_to_mg(ws);
1508 struct cache *cache = mg->cache;
1509
1510 r = invalidate_cblock(cache, mg->invalidate_cblock);
1511 if (r) {
1512 invalidate_complete(mg, false);
1513 return;
1514 }
1515
1516 init_continuation(&mg->k, invalidate_completed);
1517 continue_after_commit(&cache->committer, &mg->k);
1518 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1519 mg->overwrite_bio = NULL;
1520 schedule_commit(&cache->committer);
1521}
1522
1523static int invalidate_lock(struct dm_cache_migration *mg)
1524{
1525 int r;
1526 struct dm_cell_key_v2 key;
1527 struct cache *cache = mg->cache;
1528 struct dm_bio_prison_cell_v2 *prealloc;
1529
1530 prealloc = alloc_prison_cell(cache);
1531
1532 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1533 r = dm_cell_lock_v2(cache->prison, &key,
1534 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1535 if (r < 0) {
1536 free_prison_cell(cache, prealloc);
1537 invalidate_complete(mg, false);
1538 return r;
1539 }
1540
1541 if (mg->cell != prealloc)
1542 free_prison_cell(cache, prealloc);
1543
1544 if (r)
1545 quiesce(mg, invalidate_remove);
1546
1547 else {
1548 /*
1549 * We can't call invalidate_remove() directly here because we
1550 * might still be in request context.
1551 */
1552 init_continuation(&mg->k, invalidate_remove);
1553 queue_work(cache->wq, &mg->k.ws);
1554 }
1555
1556 return 0;
1557}
1558
1559static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1560 dm_oblock_t oblock, struct bio *bio)
1561{
1562 struct dm_cache_migration *mg;
1563
1564 if (!background_work_begin(cache))
1565 return -EPERM;
1566
1567 mg = alloc_migration(cache);
1568
1569 mg->overwrite_bio = bio;
1570 mg->invalidate_cblock = cblock;
1571 mg->invalidate_oblock = oblock;
1572
1573 return invalidate_lock(mg);
1574}
1575
1576/*
1577 *--------------------------------------------------------------
1578 * bio processing
1579 *--------------------------------------------------------------
1580 */
1581
1582enum busy {
1583 IDLE,
1584 BUSY
1585};
1586
1587static enum busy spare_migration_bandwidth(struct cache *cache)
1588{
1589 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1590 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1591 cache->sectors_per_block;
1592
1593 if (idle && current_volume <= cache->migration_threshold)
1594 return IDLE;
1595 else
1596 return BUSY;
1597}
1598
1599static void inc_hit_counter(struct cache *cache, struct bio *bio)
1600{
1601 atomic_inc(bio_data_dir(bio) == READ ?
1602 &cache->stats.read_hit : &cache->stats.write_hit);
1603}
1604
1605static void inc_miss_counter(struct cache *cache, struct bio *bio)
1606{
1607 atomic_inc(bio_data_dir(bio) == READ ?
1608 &cache->stats.read_miss : &cache->stats.write_miss);
1609}
1610
1611/*----------------------------------------------------------------*/
1612
1613static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1614 bool *commit_needed)
1615{
1616 int r, data_dir;
1617 bool rb, background_queued;
1618 dm_cblock_t cblock;
1619
1620 *commit_needed = false;
1621
1622 rb = bio_detain_shared(cache, block, bio);
1623 if (!rb) {
1624 /*
1625 * An exclusive lock is held for this block, so we have to
1626 * wait. We set the commit_needed flag so the current
1627 * transaction will be committed asap, allowing this lock
1628 * to be dropped.
1629 */
1630 *commit_needed = true;
1631 return DM_MAPIO_SUBMITTED;
1632 }
1633
1634 data_dir = bio_data_dir(bio);
1635
1636 if (optimisable_bio(cache, bio, block)) {
1637 struct policy_work *op = NULL;
1638
1639 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1640 if (unlikely(r && r != -ENOENT)) {
1641 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1642 cache_device_name(cache), r);
1643 bio_io_error(bio);
1644 return DM_MAPIO_SUBMITTED;
1645 }
1646
1647 if (r == -ENOENT && op) {
1648 bio_drop_shared_lock(cache, bio);
1649 BUG_ON(op->op != POLICY_PROMOTE);
1650 mg_start(cache, op, bio);
1651 return DM_MAPIO_SUBMITTED;
1652 }
1653 } else {
1654 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1655 if (unlikely(r && r != -ENOENT)) {
1656 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1657 cache_device_name(cache), r);
1658 bio_io_error(bio);
1659 return DM_MAPIO_SUBMITTED;
1660 }
1661
1662 if (background_queued)
1663 wake_migration_worker(cache);
1664 }
1665
1666 if (r == -ENOENT) {
1667 struct per_bio_data *pb = get_per_bio_data(bio);
1668
1669 /*
1670 * Miss.
1671 */
1672 inc_miss_counter(cache, bio);
1673 if (pb->req_nr == 0) {
1674 accounted_begin(cache, bio);
1675 remap_to_origin_clear_discard(cache, bio, block);
1676 } else {
1677 /*
1678 * This is a duplicate writethrough io that is no
1679 * longer needed because the block has been demoted.
1680 */
1681 bio_endio(bio);
1682 return DM_MAPIO_SUBMITTED;
1683 }
1684 } else {
1685 /*
1686 * Hit.
1687 */
1688 inc_hit_counter(cache, bio);
1689
1690 /*
1691 * Passthrough always maps to the origin, invalidating any
1692 * cache blocks that are written to.
1693 */
1694 if (passthrough_mode(cache)) {
1695 if (bio_data_dir(bio) == WRITE) {
1696 bio_drop_shared_lock(cache, bio);
1697 atomic_inc(&cache->stats.demotion);
1698 invalidate_start(cache, cblock, block, bio);
1699 } else
1700 remap_to_origin_clear_discard(cache, bio, block);
1701 } else {
1702 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1703 !is_dirty(cache, cblock)) {
1704 remap_to_origin_and_cache(cache, bio, block, cblock);
1705 accounted_begin(cache, bio);
1706 } else
1707 remap_to_cache_dirty(cache, bio, block, cblock);
1708 }
1709 }
1710
1711 /*
1712 * dm core turns FUA requests into a separate payload and FLUSH req.
1713 */
1714 if (bio->bi_opf & REQ_FUA) {
1715 /*
1716 * issue_after_commit will call accounted_begin a second time. So
1717 * we call accounted_complete() to avoid double accounting.
1718 */
1719 accounted_complete(cache, bio);
1720 issue_after_commit(&cache->committer, bio);
1721 *commit_needed = true;
1722 return DM_MAPIO_SUBMITTED;
1723 }
1724
1725 return DM_MAPIO_REMAPPED;
1726}
1727
1728static bool process_bio(struct cache *cache, struct bio *bio)
1729{
1730 bool commit_needed;
1731
1732 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1733 dm_submit_bio_remap(bio, NULL);
1734
1735 return commit_needed;
1736}
1737
1738/*
1739 * A non-zero return indicates read_only or fail_io mode.
1740 */
1741static int commit(struct cache *cache, bool clean_shutdown)
1742{
1743 int r;
1744
1745 if (get_cache_mode(cache) >= CM_READ_ONLY)
1746 return -EINVAL;
1747
1748 atomic_inc(&cache->stats.commit_count);
1749 r = dm_cache_commit(cache->cmd, clean_shutdown);
1750 if (r)
1751 metadata_operation_failed(cache, "dm_cache_commit", r);
1752
1753 return r;
1754}
1755
1756/*
1757 * Used by the batcher.
1758 */
1759static blk_status_t commit_op(void *context)
1760{
1761 struct cache *cache = context;
1762
1763 if (dm_cache_changed_this_transaction(cache->cmd))
1764 return errno_to_blk_status(commit(cache, false));
1765
1766 return 0;
1767}
1768
1769/*----------------------------------------------------------------*/
1770
1771static bool process_flush_bio(struct cache *cache, struct bio *bio)
1772{
1773 struct per_bio_data *pb = get_per_bio_data(bio);
1774
1775 if (!pb->req_nr)
1776 remap_to_origin(cache, bio);
1777 else
1778 remap_to_cache(cache, bio, 0);
1779
1780 issue_after_commit(&cache->committer, bio);
1781 return true;
1782}
1783
1784static bool process_discard_bio(struct cache *cache, struct bio *bio)
1785{
1786 dm_dblock_t b, e;
1787
1788 /*
1789 * FIXME: do we need to lock the region? Or can we just assume the
1790 * user wont be so foolish as to issue discard concurrently with
1791 * other IO?
1792 */
1793 calc_discard_block_range(cache, bio, &b, &e);
1794 while (b != e) {
1795 set_discard(cache, b);
1796 b = to_dblock(from_dblock(b) + 1);
1797 }
1798
1799 if (cache->features.discard_passdown) {
1800 remap_to_origin(cache, bio);
1801 dm_submit_bio_remap(bio, NULL);
1802 } else
1803 bio_endio(bio);
1804
1805 return false;
1806}
1807
1808static void process_deferred_bios(struct work_struct *ws)
1809{
1810 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1811
1812 bool commit_needed = false;
1813 struct bio_list bios;
1814 struct bio *bio;
1815
1816 bio_list_init(&bios);
1817
1818 spin_lock_irq(&cache->lock);
1819 bio_list_merge(&bios, &cache->deferred_bios);
1820 bio_list_init(&cache->deferred_bios);
1821 spin_unlock_irq(&cache->lock);
1822
1823 while ((bio = bio_list_pop(&bios))) {
1824 if (bio->bi_opf & REQ_PREFLUSH)
1825 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1826
1827 else if (bio_op(bio) == REQ_OP_DISCARD)
1828 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1829
1830 else
1831 commit_needed = process_bio(cache, bio) || commit_needed;
1832 cond_resched();
1833 }
1834
1835 if (commit_needed)
1836 schedule_commit(&cache->committer);
1837}
1838
1839/*
1840 *--------------------------------------------------------------
1841 * Main worker loop
1842 *--------------------------------------------------------------
1843 */
1844static void requeue_deferred_bios(struct cache *cache)
1845{
1846 struct bio *bio;
1847 struct bio_list bios;
1848
1849 bio_list_init(&bios);
1850 bio_list_merge(&bios, &cache->deferred_bios);
1851 bio_list_init(&cache->deferred_bios);
1852
1853 while ((bio = bio_list_pop(&bios))) {
1854 bio->bi_status = BLK_STS_DM_REQUEUE;
1855 bio_endio(bio);
1856 cond_resched();
1857 }
1858}
1859
1860/*
1861 * We want to commit periodically so that not too much
1862 * unwritten metadata builds up.
1863 */
1864static void do_waker(struct work_struct *ws)
1865{
1866 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1867
1868 policy_tick(cache->policy, true);
1869 wake_migration_worker(cache);
1870 schedule_commit(&cache->committer);
1871 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1872}
1873
1874static void check_migrations(struct work_struct *ws)
1875{
1876 int r;
1877 struct policy_work *op;
1878 struct cache *cache = container_of(ws, struct cache, migration_worker);
1879 enum busy b;
1880
1881 for (;;) {
1882 b = spare_migration_bandwidth(cache);
1883
1884 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1885 if (r == -ENODATA)
1886 break;
1887
1888 if (r) {
1889 DMERR_LIMIT("%s: policy_background_work failed",
1890 cache_device_name(cache));
1891 break;
1892 }
1893
1894 r = mg_start(cache, op, NULL);
1895 if (r)
1896 break;
1897
1898 cond_resched();
1899 }
1900}
1901
1902/*
1903 *--------------------------------------------------------------
1904 * Target methods
1905 *--------------------------------------------------------------
1906 */
1907
1908/*
1909 * This function gets called on the error paths of the constructor, so we
1910 * have to cope with a partially initialised struct.
1911 */
1912static void destroy(struct cache *cache)
1913{
1914 unsigned int i;
1915
1916 mempool_exit(&cache->migration_pool);
1917
1918 if (cache->prison)
1919 dm_bio_prison_destroy_v2(cache->prison);
1920
1921 cancel_delayed_work_sync(&cache->waker);
1922 if (cache->wq)
1923 destroy_workqueue(cache->wq);
1924
1925 if (cache->dirty_bitset)
1926 free_bitset(cache->dirty_bitset);
1927
1928 if (cache->discard_bitset)
1929 free_bitset(cache->discard_bitset);
1930
1931 if (cache->copier)
1932 dm_kcopyd_client_destroy(cache->copier);
1933
1934 if (cache->cmd)
1935 dm_cache_metadata_close(cache->cmd);
1936
1937 if (cache->metadata_dev)
1938 dm_put_device(cache->ti, cache->metadata_dev);
1939
1940 if (cache->origin_dev)
1941 dm_put_device(cache->ti, cache->origin_dev);
1942
1943 if (cache->cache_dev)
1944 dm_put_device(cache->ti, cache->cache_dev);
1945
1946 if (cache->policy)
1947 dm_cache_policy_destroy(cache->policy);
1948
1949 for (i = 0; i < cache->nr_ctr_args ; i++)
1950 kfree(cache->ctr_args[i]);
1951 kfree(cache->ctr_args);
1952
1953 bioset_exit(&cache->bs);
1954
1955 kfree(cache);
1956}
1957
1958static void cache_dtr(struct dm_target *ti)
1959{
1960 struct cache *cache = ti->private;
1961
1962 destroy(cache);
1963}
1964
1965static sector_t get_dev_size(struct dm_dev *dev)
1966{
1967 return bdev_nr_sectors(dev->bdev);
1968}
1969
1970/*----------------------------------------------------------------*/
1971
1972/*
1973 * Construct a cache device mapping.
1974 *
1975 * cache <metadata dev> <cache dev> <origin dev> <block size>
1976 * <#feature args> [<feature arg>]*
1977 * <policy> <#policy args> [<policy arg>]*
1978 *
1979 * metadata dev : fast device holding the persistent metadata
1980 * cache dev : fast device holding cached data blocks
1981 * origin dev : slow device holding original data blocks
1982 * block size : cache unit size in sectors
1983 *
1984 * #feature args : number of feature arguments passed
1985 * feature args : writethrough. (The default is writeback.)
1986 *
1987 * policy : the replacement policy to use
1988 * #policy args : an even number of policy arguments corresponding
1989 * to key/value pairs passed to the policy
1990 * policy args : key/value pairs passed to the policy
1991 * E.g. 'sequential_threshold 1024'
1992 * See cache-policies.txt for details.
1993 *
1994 * Optional feature arguments are:
1995 * writethrough : write through caching that prohibits cache block
1996 * content from being different from origin block content.
1997 * Without this argument, the default behaviour is to write
1998 * back cache block contents later for performance reasons,
1999 * so they may differ from the corresponding origin blocks.
2000 */
2001struct cache_args {
2002 struct dm_target *ti;
2003
2004 struct dm_dev *metadata_dev;
2005
2006 struct dm_dev *cache_dev;
2007 sector_t cache_sectors;
2008
2009 struct dm_dev *origin_dev;
2010 sector_t origin_sectors;
2011
2012 uint32_t block_size;
2013
2014 const char *policy_name;
2015 int policy_argc;
2016 const char **policy_argv;
2017
2018 struct cache_features features;
2019};
2020
2021static void destroy_cache_args(struct cache_args *ca)
2022{
2023 if (ca->metadata_dev)
2024 dm_put_device(ca->ti, ca->metadata_dev);
2025
2026 if (ca->cache_dev)
2027 dm_put_device(ca->ti, ca->cache_dev);
2028
2029 if (ca->origin_dev)
2030 dm_put_device(ca->ti, ca->origin_dev);
2031
2032 kfree(ca);
2033}
2034
2035static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2036{
2037 if (!as->argc) {
2038 *error = "Insufficient args";
2039 return false;
2040 }
2041
2042 return true;
2043}
2044
2045static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2046 char **error)
2047{
2048 int r;
2049 sector_t metadata_dev_size;
2050
2051 if (!at_least_one_arg(as, error))
2052 return -EINVAL;
2053
2054 r = dm_get_device(ca->ti, dm_shift_arg(as),
2055 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->metadata_dev);
2056 if (r) {
2057 *error = "Error opening metadata device";
2058 return r;
2059 }
2060
2061 metadata_dev_size = get_dev_size(ca->metadata_dev);
2062 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2063 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2064 ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2065
2066 return 0;
2067}
2068
2069static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2070 char **error)
2071{
2072 int r;
2073
2074 if (!at_least_one_arg(as, error))
2075 return -EINVAL;
2076
2077 r = dm_get_device(ca->ti, dm_shift_arg(as),
2078 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->cache_dev);
2079 if (r) {
2080 *error = "Error opening cache device";
2081 return r;
2082 }
2083 ca->cache_sectors = get_dev_size(ca->cache_dev);
2084
2085 return 0;
2086}
2087
2088static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2089 char **error)
2090{
2091 int r;
2092
2093 if (!at_least_one_arg(as, error))
2094 return -EINVAL;
2095
2096 r = dm_get_device(ca->ti, dm_shift_arg(as),
2097 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->origin_dev);
2098 if (r) {
2099 *error = "Error opening origin device";
2100 return r;
2101 }
2102
2103 ca->origin_sectors = get_dev_size(ca->origin_dev);
2104 if (ca->ti->len > ca->origin_sectors) {
2105 *error = "Device size larger than cached device";
2106 return -EINVAL;
2107 }
2108
2109 return 0;
2110}
2111
2112static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2113 char **error)
2114{
2115 unsigned long block_size;
2116
2117 if (!at_least_one_arg(as, error))
2118 return -EINVAL;
2119
2120 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2121 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2122 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2123 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2124 *error = "Invalid data block size";
2125 return -EINVAL;
2126 }
2127
2128 if (block_size > ca->cache_sectors) {
2129 *error = "Data block size is larger than the cache device";
2130 return -EINVAL;
2131 }
2132
2133 ca->block_size = block_size;
2134
2135 return 0;
2136}
2137
2138static void init_features(struct cache_features *cf)
2139{
2140 cf->mode = CM_WRITE;
2141 cf->io_mode = CM_IO_WRITEBACK;
2142 cf->metadata_version = 1;
2143 cf->discard_passdown = true;
2144}
2145
2146static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2147 char **error)
2148{
2149 static const struct dm_arg _args[] = {
2150 {0, 3, "Invalid number of cache feature arguments"},
2151 };
2152
2153 int r, mode_ctr = 0;
2154 unsigned int argc;
2155 const char *arg;
2156 struct cache_features *cf = &ca->features;
2157
2158 init_features(cf);
2159
2160 r = dm_read_arg_group(_args, as, &argc, error);
2161 if (r)
2162 return -EINVAL;
2163
2164 while (argc--) {
2165 arg = dm_shift_arg(as);
2166
2167 if (!strcasecmp(arg, "writeback")) {
2168 cf->io_mode = CM_IO_WRITEBACK;
2169 mode_ctr++;
2170 }
2171
2172 else if (!strcasecmp(arg, "writethrough")) {
2173 cf->io_mode = CM_IO_WRITETHROUGH;
2174 mode_ctr++;
2175 }
2176
2177 else if (!strcasecmp(arg, "passthrough")) {
2178 cf->io_mode = CM_IO_PASSTHROUGH;
2179 mode_ctr++;
2180 }
2181
2182 else if (!strcasecmp(arg, "metadata2"))
2183 cf->metadata_version = 2;
2184
2185 else if (!strcasecmp(arg, "no_discard_passdown"))
2186 cf->discard_passdown = false;
2187
2188 else {
2189 *error = "Unrecognised cache feature requested";
2190 return -EINVAL;
2191 }
2192 }
2193
2194 if (mode_ctr > 1) {
2195 *error = "Duplicate cache io_mode features requested";
2196 return -EINVAL;
2197 }
2198
2199 return 0;
2200}
2201
2202static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2203 char **error)
2204{
2205 static const struct dm_arg _args[] = {
2206 {0, 1024, "Invalid number of policy arguments"},
2207 };
2208
2209 int r;
2210
2211 if (!at_least_one_arg(as, error))
2212 return -EINVAL;
2213
2214 ca->policy_name = dm_shift_arg(as);
2215
2216 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2217 if (r)
2218 return -EINVAL;
2219
2220 ca->policy_argv = (const char **)as->argv;
2221 dm_consume_args(as, ca->policy_argc);
2222
2223 return 0;
2224}
2225
2226static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2227 char **error)
2228{
2229 int r;
2230 struct dm_arg_set as;
2231
2232 as.argc = argc;
2233 as.argv = argv;
2234
2235 r = parse_metadata_dev(ca, &as, error);
2236 if (r)
2237 return r;
2238
2239 r = parse_cache_dev(ca, &as, error);
2240 if (r)
2241 return r;
2242
2243 r = parse_origin_dev(ca, &as, error);
2244 if (r)
2245 return r;
2246
2247 r = parse_block_size(ca, &as, error);
2248 if (r)
2249 return r;
2250
2251 r = parse_features(ca, &as, error);
2252 if (r)
2253 return r;
2254
2255 r = parse_policy(ca, &as, error);
2256 if (r)
2257 return r;
2258
2259 return 0;
2260}
2261
2262/*----------------------------------------------------------------*/
2263
2264static struct kmem_cache *migration_cache;
2265
2266#define NOT_CORE_OPTION 1
2267
2268static int process_config_option(struct cache *cache, const char *key, const char *value)
2269{
2270 unsigned long tmp;
2271
2272 if (!strcasecmp(key, "migration_threshold")) {
2273 if (kstrtoul(value, 10, &tmp))
2274 return -EINVAL;
2275
2276 cache->migration_threshold = tmp;
2277 return 0;
2278 }
2279
2280 return NOT_CORE_OPTION;
2281}
2282
2283static int set_config_value(struct cache *cache, const char *key, const char *value)
2284{
2285 int r = process_config_option(cache, key, value);
2286
2287 if (r == NOT_CORE_OPTION)
2288 r = policy_set_config_value(cache->policy, key, value);
2289
2290 if (r)
2291 DMWARN("bad config value for %s: %s", key, value);
2292
2293 return r;
2294}
2295
2296static int set_config_values(struct cache *cache, int argc, const char **argv)
2297{
2298 int r = 0;
2299
2300 if (argc & 1) {
2301 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2302 return -EINVAL;
2303 }
2304
2305 while (argc) {
2306 r = set_config_value(cache, argv[0], argv[1]);
2307 if (r)
2308 break;
2309
2310 argc -= 2;
2311 argv += 2;
2312 }
2313
2314 return r;
2315}
2316
2317static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2318 char **error)
2319{
2320 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2321 cache->cache_size,
2322 cache->origin_sectors,
2323 cache->sectors_per_block);
2324 if (IS_ERR(p)) {
2325 *error = "Error creating cache's policy";
2326 return PTR_ERR(p);
2327 }
2328 cache->policy = p;
2329 BUG_ON(!cache->policy);
2330
2331 return 0;
2332}
2333
2334/*
2335 * We want the discard block size to be at least the size of the cache
2336 * block size and have no more than 2^14 discard blocks across the origin.
2337 */
2338#define MAX_DISCARD_BLOCKS (1 << 14)
2339
2340static bool too_many_discard_blocks(sector_t discard_block_size,
2341 sector_t origin_size)
2342{
2343 (void) sector_div(origin_size, discard_block_size);
2344
2345 return origin_size > MAX_DISCARD_BLOCKS;
2346}
2347
2348static sector_t calculate_discard_block_size(sector_t cache_block_size,
2349 sector_t origin_size)
2350{
2351 sector_t discard_block_size = cache_block_size;
2352
2353 if (origin_size)
2354 while (too_many_discard_blocks(discard_block_size, origin_size))
2355 discard_block_size *= 2;
2356
2357 return discard_block_size;
2358}
2359
2360static void set_cache_size(struct cache *cache, dm_cblock_t size)
2361{
2362 dm_block_t nr_blocks = from_cblock(size);
2363
2364 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2365 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2366 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2367 "Please consider increasing the cache block size to reduce the overall cache block count.",
2368 (unsigned long long) nr_blocks);
2369
2370 cache->cache_size = size;
2371}
2372
2373#define DEFAULT_MIGRATION_THRESHOLD 2048
2374
2375static int cache_create(struct cache_args *ca, struct cache **result)
2376{
2377 int r = 0;
2378 char **error = &ca->ti->error;
2379 struct cache *cache;
2380 struct dm_target *ti = ca->ti;
2381 dm_block_t origin_blocks;
2382 struct dm_cache_metadata *cmd;
2383 bool may_format = ca->features.mode == CM_WRITE;
2384
2385 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2386 if (!cache)
2387 return -ENOMEM;
2388
2389 cache->ti = ca->ti;
2390 ti->private = cache;
2391 ti->accounts_remapped_io = true;
2392 ti->num_flush_bios = 2;
2393 ti->flush_supported = true;
2394
2395 ti->num_discard_bios = 1;
2396 ti->discards_supported = true;
2397
2398 ti->per_io_data_size = sizeof(struct per_bio_data);
2399
2400 cache->features = ca->features;
2401 if (writethrough_mode(cache)) {
2402 /* Create bioset for writethrough bios issued to origin */
2403 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2404 if (r)
2405 goto bad;
2406 }
2407
2408 cache->metadata_dev = ca->metadata_dev;
2409 cache->origin_dev = ca->origin_dev;
2410 cache->cache_dev = ca->cache_dev;
2411
2412 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2413
2414 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2415 origin_blocks = block_div(origin_blocks, ca->block_size);
2416 cache->origin_blocks = to_oblock(origin_blocks);
2417
2418 cache->sectors_per_block = ca->block_size;
2419 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2420 r = -EINVAL;
2421 goto bad;
2422 }
2423
2424 if (ca->block_size & (ca->block_size - 1)) {
2425 dm_block_t cache_size = ca->cache_sectors;
2426
2427 cache->sectors_per_block_shift = -1;
2428 cache_size = block_div(cache_size, ca->block_size);
2429 set_cache_size(cache, to_cblock(cache_size));
2430 } else {
2431 cache->sectors_per_block_shift = __ffs(ca->block_size);
2432 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2433 }
2434
2435 r = create_cache_policy(cache, ca, error);
2436 if (r)
2437 goto bad;
2438
2439 cache->policy_nr_args = ca->policy_argc;
2440 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2441
2442 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2443 if (r) {
2444 *error = "Error setting cache policy's config values";
2445 goto bad;
2446 }
2447
2448 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2449 ca->block_size, may_format,
2450 dm_cache_policy_get_hint_size(cache->policy),
2451 ca->features.metadata_version);
2452 if (IS_ERR(cmd)) {
2453 *error = "Error creating metadata object";
2454 r = PTR_ERR(cmd);
2455 goto bad;
2456 }
2457 cache->cmd = cmd;
2458 set_cache_mode(cache, CM_WRITE);
2459 if (get_cache_mode(cache) != CM_WRITE) {
2460 *error = "Unable to get write access to metadata, please check/repair metadata.";
2461 r = -EINVAL;
2462 goto bad;
2463 }
2464
2465 if (passthrough_mode(cache)) {
2466 bool all_clean;
2467
2468 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2469 if (r) {
2470 *error = "dm_cache_metadata_all_clean() failed";
2471 goto bad;
2472 }
2473
2474 if (!all_clean) {
2475 *error = "Cannot enter passthrough mode unless all blocks are clean";
2476 r = -EINVAL;
2477 goto bad;
2478 }
2479
2480 policy_allow_migrations(cache->policy, false);
2481 }
2482
2483 spin_lock_init(&cache->lock);
2484 bio_list_init(&cache->deferred_bios);
2485 atomic_set(&cache->nr_allocated_migrations, 0);
2486 atomic_set(&cache->nr_io_migrations, 0);
2487 init_waitqueue_head(&cache->migration_wait);
2488
2489 r = -ENOMEM;
2490 atomic_set(&cache->nr_dirty, 0);
2491 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2492 if (!cache->dirty_bitset) {
2493 *error = "could not allocate dirty bitset";
2494 goto bad;
2495 }
2496 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2497
2498 cache->discard_block_size =
2499 calculate_discard_block_size(cache->sectors_per_block,
2500 cache->origin_sectors);
2501 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2502 cache->discard_block_size));
2503 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2504 if (!cache->discard_bitset) {
2505 *error = "could not allocate discard bitset";
2506 goto bad;
2507 }
2508 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2509
2510 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2511 if (IS_ERR(cache->copier)) {
2512 *error = "could not create kcopyd client";
2513 r = PTR_ERR(cache->copier);
2514 goto bad;
2515 }
2516
2517 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2518 if (!cache->wq) {
2519 *error = "could not create workqueue for metadata object";
2520 goto bad;
2521 }
2522 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2523 INIT_WORK(&cache->migration_worker, check_migrations);
2524 INIT_DELAYED_WORK(&cache->waker, do_waker);
2525
2526 cache->prison = dm_bio_prison_create_v2(cache->wq);
2527 if (!cache->prison) {
2528 *error = "could not create bio prison";
2529 goto bad;
2530 }
2531
2532 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2533 migration_cache);
2534 if (r) {
2535 *error = "Error creating cache's migration mempool";
2536 goto bad;
2537 }
2538
2539 cache->need_tick_bio = true;
2540 cache->sized = false;
2541 cache->invalidate = false;
2542 cache->commit_requested = false;
2543 cache->loaded_mappings = false;
2544 cache->loaded_discards = false;
2545
2546 load_stats(cache);
2547
2548 atomic_set(&cache->stats.demotion, 0);
2549 atomic_set(&cache->stats.promotion, 0);
2550 atomic_set(&cache->stats.copies_avoided, 0);
2551 atomic_set(&cache->stats.cache_cell_clash, 0);
2552 atomic_set(&cache->stats.commit_count, 0);
2553 atomic_set(&cache->stats.discard_count, 0);
2554
2555 spin_lock_init(&cache->invalidation_lock);
2556 INIT_LIST_HEAD(&cache->invalidation_requests);
2557
2558 batcher_init(&cache->committer, commit_op, cache,
2559 issue_op, cache, cache->wq);
2560 dm_iot_init(&cache->tracker);
2561
2562 init_rwsem(&cache->background_work_lock);
2563 prevent_background_work(cache);
2564
2565 *result = cache;
2566 return 0;
2567bad:
2568 destroy(cache);
2569 return r;
2570}
2571
2572static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2573{
2574 unsigned int i;
2575 const char **copy;
2576
2577 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2578 if (!copy)
2579 return -ENOMEM;
2580 for (i = 0; i < argc; i++) {
2581 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2582 if (!copy[i]) {
2583 while (i--)
2584 kfree(copy[i]);
2585 kfree(copy);
2586 return -ENOMEM;
2587 }
2588 }
2589
2590 cache->nr_ctr_args = argc;
2591 cache->ctr_args = copy;
2592
2593 return 0;
2594}
2595
2596static int cache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2597{
2598 int r = -EINVAL;
2599 struct cache_args *ca;
2600 struct cache *cache = NULL;
2601
2602 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2603 if (!ca) {
2604 ti->error = "Error allocating memory for cache";
2605 return -ENOMEM;
2606 }
2607 ca->ti = ti;
2608
2609 r = parse_cache_args(ca, argc, argv, &ti->error);
2610 if (r)
2611 goto out;
2612
2613 r = cache_create(ca, &cache);
2614 if (r)
2615 goto out;
2616
2617 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2618 if (r) {
2619 destroy(cache);
2620 goto out;
2621 }
2622
2623 ti->private = cache;
2624out:
2625 destroy_cache_args(ca);
2626 return r;
2627}
2628
2629/*----------------------------------------------------------------*/
2630
2631static int cache_map(struct dm_target *ti, struct bio *bio)
2632{
2633 struct cache *cache = ti->private;
2634
2635 int r;
2636 bool commit_needed;
2637 dm_oblock_t block = get_bio_block(cache, bio);
2638
2639 init_per_bio_data(bio);
2640 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2641 /*
2642 * This can only occur if the io goes to a partial block at
2643 * the end of the origin device. We don't cache these.
2644 * Just remap to the origin and carry on.
2645 */
2646 remap_to_origin(cache, bio);
2647 accounted_begin(cache, bio);
2648 return DM_MAPIO_REMAPPED;
2649 }
2650
2651 if (discard_or_flush(bio)) {
2652 defer_bio(cache, bio);
2653 return DM_MAPIO_SUBMITTED;
2654 }
2655
2656 r = map_bio(cache, bio, block, &commit_needed);
2657 if (commit_needed)
2658 schedule_commit(&cache->committer);
2659
2660 return r;
2661}
2662
2663static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2664{
2665 struct cache *cache = ti->private;
2666 unsigned long flags;
2667 struct per_bio_data *pb = get_per_bio_data(bio);
2668
2669 if (pb->tick) {
2670 policy_tick(cache->policy, false);
2671
2672 spin_lock_irqsave(&cache->lock, flags);
2673 cache->need_tick_bio = true;
2674 spin_unlock_irqrestore(&cache->lock, flags);
2675 }
2676
2677 bio_drop_shared_lock(cache, bio);
2678 accounted_complete(cache, bio);
2679
2680 return DM_ENDIO_DONE;
2681}
2682
2683static int write_dirty_bitset(struct cache *cache)
2684{
2685 int r;
2686
2687 if (get_cache_mode(cache) >= CM_READ_ONLY)
2688 return -EINVAL;
2689
2690 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2691 if (r)
2692 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2693
2694 return r;
2695}
2696
2697static int write_discard_bitset(struct cache *cache)
2698{
2699 unsigned int i, r;
2700
2701 if (get_cache_mode(cache) >= CM_READ_ONLY)
2702 return -EINVAL;
2703
2704 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2705 cache->discard_nr_blocks);
2706 if (r) {
2707 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2708 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2709 return r;
2710 }
2711
2712 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2713 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2714 is_discarded(cache, to_dblock(i)));
2715 if (r) {
2716 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2717 return r;
2718 }
2719 }
2720
2721 return 0;
2722}
2723
2724static int write_hints(struct cache *cache)
2725{
2726 int r;
2727
2728 if (get_cache_mode(cache) >= CM_READ_ONLY)
2729 return -EINVAL;
2730
2731 r = dm_cache_write_hints(cache->cmd, cache->policy);
2732 if (r) {
2733 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2734 return r;
2735 }
2736
2737 return 0;
2738}
2739
2740/*
2741 * returns true on success
2742 */
2743static bool sync_metadata(struct cache *cache)
2744{
2745 int r1, r2, r3, r4;
2746
2747 r1 = write_dirty_bitset(cache);
2748 if (r1)
2749 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2750
2751 r2 = write_discard_bitset(cache);
2752 if (r2)
2753 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2754
2755 save_stats(cache);
2756
2757 r3 = write_hints(cache);
2758 if (r3)
2759 DMERR("%s: could not write hints", cache_device_name(cache));
2760
2761 /*
2762 * If writing the above metadata failed, we still commit, but don't
2763 * set the clean shutdown flag. This will effectively force every
2764 * dirty bit to be set on reload.
2765 */
2766 r4 = commit(cache, !r1 && !r2 && !r3);
2767 if (r4)
2768 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2769
2770 return !r1 && !r2 && !r3 && !r4;
2771}
2772
2773static void cache_postsuspend(struct dm_target *ti)
2774{
2775 struct cache *cache = ti->private;
2776
2777 prevent_background_work(cache);
2778 BUG_ON(atomic_read(&cache->nr_io_migrations));
2779
2780 cancel_delayed_work_sync(&cache->waker);
2781 drain_workqueue(cache->wq);
2782 WARN_ON(cache->tracker.in_flight);
2783
2784 /*
2785 * If it's a flush suspend there won't be any deferred bios, so this
2786 * call is harmless.
2787 */
2788 requeue_deferred_bios(cache);
2789
2790 if (get_cache_mode(cache) == CM_WRITE)
2791 (void) sync_metadata(cache);
2792}
2793
2794static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2795 bool dirty, uint32_t hint, bool hint_valid)
2796{
2797 struct cache *cache = context;
2798
2799 if (dirty) {
2800 set_bit(from_cblock(cblock), cache->dirty_bitset);
2801 atomic_inc(&cache->nr_dirty);
2802 } else
2803 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2804
2805 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2806}
2807
2808/*
2809 * The discard block size in the on disk metadata is not
2810 * necessarily the same as we're currently using. So we have to
2811 * be careful to only set the discarded attribute if we know it
2812 * covers a complete block of the new size.
2813 */
2814struct discard_load_info {
2815 struct cache *cache;
2816
2817 /*
2818 * These blocks are sized using the on disk dblock size, rather
2819 * than the current one.
2820 */
2821 dm_block_t block_size;
2822 dm_block_t discard_begin, discard_end;
2823};
2824
2825static void discard_load_info_init(struct cache *cache,
2826 struct discard_load_info *li)
2827{
2828 li->cache = cache;
2829 li->discard_begin = li->discard_end = 0;
2830}
2831
2832static void set_discard_range(struct discard_load_info *li)
2833{
2834 sector_t b, e;
2835
2836 if (li->discard_begin == li->discard_end)
2837 return;
2838
2839 /*
2840 * Convert to sectors.
2841 */
2842 b = li->discard_begin * li->block_size;
2843 e = li->discard_end * li->block_size;
2844
2845 /*
2846 * Then convert back to the current dblock size.
2847 */
2848 b = dm_sector_div_up(b, li->cache->discard_block_size);
2849 sector_div(e, li->cache->discard_block_size);
2850
2851 /*
2852 * The origin may have shrunk, so we need to check we're still in
2853 * bounds.
2854 */
2855 if (e > from_dblock(li->cache->discard_nr_blocks))
2856 e = from_dblock(li->cache->discard_nr_blocks);
2857
2858 for (; b < e; b++)
2859 set_discard(li->cache, to_dblock(b));
2860}
2861
2862static int load_discard(void *context, sector_t discard_block_size,
2863 dm_dblock_t dblock, bool discard)
2864{
2865 struct discard_load_info *li = context;
2866
2867 li->block_size = discard_block_size;
2868
2869 if (discard) {
2870 if (from_dblock(dblock) == li->discard_end)
2871 /*
2872 * We're already in a discard range, just extend it.
2873 */
2874 li->discard_end = li->discard_end + 1ULL;
2875
2876 else {
2877 /*
2878 * Emit the old range and start a new one.
2879 */
2880 set_discard_range(li);
2881 li->discard_begin = from_dblock(dblock);
2882 li->discard_end = li->discard_begin + 1ULL;
2883 }
2884 } else {
2885 set_discard_range(li);
2886 li->discard_begin = li->discard_end = 0;
2887 }
2888
2889 return 0;
2890}
2891
2892static dm_cblock_t get_cache_dev_size(struct cache *cache)
2893{
2894 sector_t size = get_dev_size(cache->cache_dev);
2895 (void) sector_div(size, cache->sectors_per_block);
2896 return to_cblock(size);
2897}
2898
2899static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2900{
2901 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2902 if (cache->sized) {
2903 DMERR("%s: unable to extend cache due to missing cache table reload",
2904 cache_device_name(cache));
2905 return false;
2906 }
2907 }
2908
2909 /*
2910 * We can't drop a dirty block when shrinking the cache.
2911 */
2912 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2913 new_size = to_cblock(from_cblock(new_size) + 1);
2914 if (is_dirty(cache, new_size)) {
2915 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2916 cache_device_name(cache),
2917 (unsigned long long) from_cblock(new_size));
2918 return false;
2919 }
2920 }
2921
2922 return true;
2923}
2924
2925static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2926{
2927 int r;
2928
2929 r = dm_cache_resize(cache->cmd, new_size);
2930 if (r) {
2931 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2932 metadata_operation_failed(cache, "dm_cache_resize", r);
2933 return r;
2934 }
2935
2936 set_cache_size(cache, new_size);
2937
2938 return 0;
2939}
2940
2941static int cache_preresume(struct dm_target *ti)
2942{
2943 int r = 0;
2944 struct cache *cache = ti->private;
2945 dm_cblock_t csize = get_cache_dev_size(cache);
2946
2947 /*
2948 * Check to see if the cache has resized.
2949 */
2950 if (!cache->sized) {
2951 r = resize_cache_dev(cache, csize);
2952 if (r)
2953 return r;
2954
2955 cache->sized = true;
2956
2957 } else if (csize != cache->cache_size) {
2958 if (!can_resize(cache, csize))
2959 return -EINVAL;
2960
2961 r = resize_cache_dev(cache, csize);
2962 if (r)
2963 return r;
2964 }
2965
2966 if (!cache->loaded_mappings) {
2967 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2968 load_mapping, cache);
2969 if (r) {
2970 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2971 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2972 return r;
2973 }
2974
2975 cache->loaded_mappings = true;
2976 }
2977
2978 if (!cache->loaded_discards) {
2979 struct discard_load_info li;
2980
2981 /*
2982 * The discard bitset could have been resized, or the
2983 * discard block size changed. To be safe we start by
2984 * setting every dblock to not discarded.
2985 */
2986 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2987
2988 discard_load_info_init(cache, &li);
2989 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2990 if (r) {
2991 DMERR("%s: could not load origin discards", cache_device_name(cache));
2992 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2993 return r;
2994 }
2995 set_discard_range(&li);
2996
2997 cache->loaded_discards = true;
2998 }
2999
3000 return r;
3001}
3002
3003static void cache_resume(struct dm_target *ti)
3004{
3005 struct cache *cache = ti->private;
3006
3007 cache->need_tick_bio = true;
3008 allow_background_work(cache);
3009 do_waker(&cache->waker.work);
3010}
3011
3012static void emit_flags(struct cache *cache, char *result,
3013 unsigned int maxlen, ssize_t *sz_ptr)
3014{
3015 ssize_t sz = *sz_ptr;
3016 struct cache_features *cf = &cache->features;
3017 unsigned int count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3018
3019 DMEMIT("%u ", count);
3020
3021 if (cf->metadata_version == 2)
3022 DMEMIT("metadata2 ");
3023
3024 if (writethrough_mode(cache))
3025 DMEMIT("writethrough ");
3026
3027 else if (passthrough_mode(cache))
3028 DMEMIT("passthrough ");
3029
3030 else if (writeback_mode(cache))
3031 DMEMIT("writeback ");
3032
3033 else {
3034 DMEMIT("unknown ");
3035 DMERR("%s: internal error: unknown io mode: %d",
3036 cache_device_name(cache), (int) cf->io_mode);
3037 }
3038
3039 if (!cf->discard_passdown)
3040 DMEMIT("no_discard_passdown ");
3041
3042 *sz_ptr = sz;
3043}
3044
3045/*
3046 * Status format:
3047 *
3048 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3049 * <cache block size> <#used cache blocks>/<#total cache blocks>
3050 * <#read hits> <#read misses> <#write hits> <#write misses>
3051 * <#demotions> <#promotions> <#dirty>
3052 * <#features> <features>*
3053 * <#core args> <core args>
3054 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3055 */
3056static void cache_status(struct dm_target *ti, status_type_t type,
3057 unsigned int status_flags, char *result, unsigned int maxlen)
3058{
3059 int r = 0;
3060 unsigned int i;
3061 ssize_t sz = 0;
3062 dm_block_t nr_free_blocks_metadata = 0;
3063 dm_block_t nr_blocks_metadata = 0;
3064 char buf[BDEVNAME_SIZE];
3065 struct cache *cache = ti->private;
3066 dm_cblock_t residency;
3067 bool needs_check;
3068
3069 switch (type) {
3070 case STATUSTYPE_INFO:
3071 if (get_cache_mode(cache) == CM_FAIL) {
3072 DMEMIT("Fail");
3073 break;
3074 }
3075
3076 /* Commit to ensure statistics aren't out-of-date */
3077 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3078 (void) commit(cache, false);
3079
3080 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3081 if (r) {
3082 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3083 cache_device_name(cache), r);
3084 goto err;
3085 }
3086
3087 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3088 if (r) {
3089 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3090 cache_device_name(cache), r);
3091 goto err;
3092 }
3093
3094 residency = policy_residency(cache->policy);
3095
3096 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3097 (unsigned int)DM_CACHE_METADATA_BLOCK_SIZE,
3098 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3099 (unsigned long long)nr_blocks_metadata,
3100 (unsigned long long)cache->sectors_per_block,
3101 (unsigned long long) from_cblock(residency),
3102 (unsigned long long) from_cblock(cache->cache_size),
3103 (unsigned int) atomic_read(&cache->stats.read_hit),
3104 (unsigned int) atomic_read(&cache->stats.read_miss),
3105 (unsigned int) atomic_read(&cache->stats.write_hit),
3106 (unsigned int) atomic_read(&cache->stats.write_miss),
3107 (unsigned int) atomic_read(&cache->stats.demotion),
3108 (unsigned int) atomic_read(&cache->stats.promotion),
3109 (unsigned long) atomic_read(&cache->nr_dirty));
3110
3111 emit_flags(cache, result, maxlen, &sz);
3112
3113 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3114
3115 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3116 if (sz < maxlen) {
3117 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3118 if (r)
3119 DMERR("%s: policy_emit_config_values returned %d",
3120 cache_device_name(cache), r);
3121 }
3122
3123 if (get_cache_mode(cache) == CM_READ_ONLY)
3124 DMEMIT("ro ");
3125 else
3126 DMEMIT("rw ");
3127
3128 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3129
3130 if (r || needs_check)
3131 DMEMIT("needs_check ");
3132 else
3133 DMEMIT("- ");
3134
3135 break;
3136
3137 case STATUSTYPE_TABLE:
3138 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3139 DMEMIT("%s ", buf);
3140 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3141 DMEMIT("%s ", buf);
3142 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3143 DMEMIT("%s", buf);
3144
3145 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3146 DMEMIT(" %s", cache->ctr_args[i]);
3147 if (cache->nr_ctr_args)
3148 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3149 break;
3150
3151 case STATUSTYPE_IMA:
3152 DMEMIT_TARGET_NAME_VERSION(ti->type);
3153 if (get_cache_mode(cache) == CM_FAIL)
3154 DMEMIT(",metadata_mode=fail");
3155 else if (get_cache_mode(cache) == CM_READ_ONLY)
3156 DMEMIT(",metadata_mode=ro");
3157 else
3158 DMEMIT(",metadata_mode=rw");
3159
3160 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3161 DMEMIT(",cache_metadata_device=%s", buf);
3162 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3163 DMEMIT(",cache_device=%s", buf);
3164 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3165 DMEMIT(",cache_origin_device=%s", buf);
3166 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3167 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3168 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3169 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3170 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3171 DMEMIT(";");
3172 break;
3173 }
3174
3175 return;
3176
3177err:
3178 DMEMIT("Error");
3179}
3180
3181/*
3182 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3183 * the one-past-the-end value.
3184 */
3185struct cblock_range {
3186 dm_cblock_t begin;
3187 dm_cblock_t end;
3188};
3189
3190/*
3191 * A cache block range can take two forms:
3192 *
3193 * i) A single cblock, eg. '3456'
3194 * ii) A begin and end cblock with a dash between, eg. 123-234
3195 */
3196static int parse_cblock_range(struct cache *cache, const char *str,
3197 struct cblock_range *result)
3198{
3199 char dummy;
3200 uint64_t b, e;
3201 int r;
3202
3203 /*
3204 * Try and parse form (ii) first.
3205 */
3206 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3207 if (r < 0)
3208 return r;
3209
3210 if (r == 2) {
3211 result->begin = to_cblock(b);
3212 result->end = to_cblock(e);
3213 return 0;
3214 }
3215
3216 /*
3217 * That didn't work, try form (i).
3218 */
3219 r = sscanf(str, "%llu%c", &b, &dummy);
3220 if (r < 0)
3221 return r;
3222
3223 if (r == 1) {
3224 result->begin = to_cblock(b);
3225 result->end = to_cblock(from_cblock(result->begin) + 1u);
3226 return 0;
3227 }
3228
3229 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3230 return -EINVAL;
3231}
3232
3233static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3234{
3235 uint64_t b = from_cblock(range->begin);
3236 uint64_t e = from_cblock(range->end);
3237 uint64_t n = from_cblock(cache->cache_size);
3238
3239 if (b >= n) {
3240 DMERR("%s: begin cblock out of range: %llu >= %llu",
3241 cache_device_name(cache), b, n);
3242 return -EINVAL;
3243 }
3244
3245 if (e > n) {
3246 DMERR("%s: end cblock out of range: %llu > %llu",
3247 cache_device_name(cache), e, n);
3248 return -EINVAL;
3249 }
3250
3251 if (b >= e) {
3252 DMERR("%s: invalid cblock range: %llu >= %llu",
3253 cache_device_name(cache), b, e);
3254 return -EINVAL;
3255 }
3256
3257 return 0;
3258}
3259
3260static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3261{
3262 return to_cblock(from_cblock(b) + 1);
3263}
3264
3265static int request_invalidation(struct cache *cache, struct cblock_range *range)
3266{
3267 int r = 0;
3268
3269 /*
3270 * We don't need to do any locking here because we know we're in
3271 * passthrough mode. There's is potential for a race between an
3272 * invalidation triggered by an io and an invalidation message. This
3273 * is harmless, we must not worry if the policy call fails.
3274 */
3275 while (range->begin != range->end) {
3276 r = invalidate_cblock(cache, range->begin);
3277 if (r)
3278 return r;
3279
3280 range->begin = cblock_succ(range->begin);
3281 }
3282
3283 cache->commit_requested = true;
3284 return r;
3285}
3286
3287static int process_invalidate_cblocks_message(struct cache *cache, unsigned int count,
3288 const char **cblock_ranges)
3289{
3290 int r = 0;
3291 unsigned int i;
3292 struct cblock_range range;
3293
3294 if (!passthrough_mode(cache)) {
3295 DMERR("%s: cache has to be in passthrough mode for invalidation",
3296 cache_device_name(cache));
3297 return -EPERM;
3298 }
3299
3300 for (i = 0; i < count; i++) {
3301 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3302 if (r)
3303 break;
3304
3305 r = validate_cblock_range(cache, &range);
3306 if (r)
3307 break;
3308
3309 /*
3310 * Pass begin and end origin blocks to the worker and wake it.
3311 */
3312 r = request_invalidation(cache, &range);
3313 if (r)
3314 break;
3315 }
3316
3317 return r;
3318}
3319
3320/*
3321 * Supports
3322 * "<key> <value>"
3323 * and
3324 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3325 *
3326 * The key migration_threshold is supported by the cache target core.
3327 */
3328static int cache_message(struct dm_target *ti, unsigned int argc, char **argv,
3329 char *result, unsigned int maxlen)
3330{
3331 struct cache *cache = ti->private;
3332
3333 if (!argc)
3334 return -EINVAL;
3335
3336 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3337 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3338 cache_device_name(cache));
3339 return -EOPNOTSUPP;
3340 }
3341
3342 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3343 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3344
3345 if (argc != 2)
3346 return -EINVAL;
3347
3348 return set_config_value(cache, argv[0], argv[1]);
3349}
3350
3351static int cache_iterate_devices(struct dm_target *ti,
3352 iterate_devices_callout_fn fn, void *data)
3353{
3354 int r = 0;
3355 struct cache *cache = ti->private;
3356
3357 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3358 if (!r)
3359 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3360
3361 return r;
3362}
3363
3364/*
3365 * If discard_passdown was enabled verify that the origin device
3366 * supports discards. Disable discard_passdown if not.
3367 */
3368static void disable_passdown_if_not_supported(struct cache *cache)
3369{
3370 struct block_device *origin_bdev = cache->origin_dev->bdev;
3371 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3372 const char *reason = NULL;
3373
3374 if (!cache->features.discard_passdown)
3375 return;
3376
3377 if (!bdev_max_discard_sectors(origin_bdev))
3378 reason = "discard unsupported";
3379
3380 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3381 reason = "max discard sectors smaller than a block";
3382
3383 if (reason) {
3384 DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3385 origin_bdev, reason);
3386 cache->features.discard_passdown = false;
3387 }
3388}
3389
3390static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3391{
3392 struct block_device *origin_bdev = cache->origin_dev->bdev;
3393 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3394
3395 if (!cache->features.discard_passdown) {
3396 /* No passdown is done so setting own virtual limits */
3397 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3398 cache->origin_sectors);
3399 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3400 return;
3401 }
3402
3403 /*
3404 * cache_iterate_devices() is stacking both origin and fast device limits
3405 * but discards aren't passed to fast device, so inherit origin's limits.
3406 */
3407 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3408 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3409 limits->discard_granularity = origin_limits->discard_granularity;
3410 limits->discard_alignment = origin_limits->discard_alignment;
3411 limits->discard_misaligned = origin_limits->discard_misaligned;
3412}
3413
3414static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3415{
3416 struct cache *cache = ti->private;
3417 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3418
3419 /*
3420 * If the system-determined stacked limits are compatible with the
3421 * cache's blocksize (io_opt is a factor) do not override them.
3422 */
3423 if (io_opt_sectors < cache->sectors_per_block ||
3424 do_div(io_opt_sectors, cache->sectors_per_block)) {
3425 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3426 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3427 }
3428
3429 disable_passdown_if_not_supported(cache);
3430 set_discard_limits(cache, limits);
3431}
3432
3433/*----------------------------------------------------------------*/
3434
3435static struct target_type cache_target = {
3436 .name = "cache",
3437 .version = {2, 2, 0},
3438 .module = THIS_MODULE,
3439 .ctr = cache_ctr,
3440 .dtr = cache_dtr,
3441 .map = cache_map,
3442 .end_io = cache_end_io,
3443 .postsuspend = cache_postsuspend,
3444 .preresume = cache_preresume,
3445 .resume = cache_resume,
3446 .status = cache_status,
3447 .message = cache_message,
3448 .iterate_devices = cache_iterate_devices,
3449 .io_hints = cache_io_hints,
3450};
3451
3452static int __init dm_cache_init(void)
3453{
3454 int r;
3455
3456 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3457 if (!migration_cache)
3458 return -ENOMEM;
3459
3460 r = dm_register_target(&cache_target);
3461 if (r) {
3462 kmem_cache_destroy(migration_cache);
3463 return r;
3464 }
3465
3466 return 0;
3467}
3468
3469static void __exit dm_cache_exit(void)
3470{
3471 dm_unregister_target(&cache_target);
3472 kmem_cache_destroy(migration_cache);
3473}
3474
3475module_init(dm_cache_init);
3476module_exit(dm_cache_exit);
3477
3478MODULE_DESCRIPTION(DM_NAME " cache target");
3479MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3480MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2012 Red Hat. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include "dm-bio-prison.h"
9#include "dm-bio-record.h"
10#include "dm-cache-metadata.h"
11
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
14#include <linux/jiffies.h>
15#include <linux/init.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/vmalloc.h>
20
21#define DM_MSG_PREFIX "cache"
22
23DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
24 "A percentage of time allocated for copying to and/or from cache");
25
26/*----------------------------------------------------------------*/
27
28#define IOT_RESOLUTION 4
29
30struct io_tracker {
31 spinlock_t lock;
32
33 /*
34 * Sectors of in-flight IO.
35 */
36 sector_t in_flight;
37
38 /*
39 * The time, in jiffies, when this device became idle (if it is
40 * indeed idle).
41 */
42 unsigned long idle_time;
43 unsigned long last_update_time;
44};
45
46static void iot_init(struct io_tracker *iot)
47{
48 spin_lock_init(&iot->lock);
49 iot->in_flight = 0ul;
50 iot->idle_time = 0ul;
51 iot->last_update_time = jiffies;
52}
53
54static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
55{
56 if (iot->in_flight)
57 return false;
58
59 return time_after(jiffies, iot->idle_time + jifs);
60}
61
62static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
63{
64 bool r;
65 unsigned long flags;
66
67 spin_lock_irqsave(&iot->lock, flags);
68 r = __iot_idle_for(iot, jifs);
69 spin_unlock_irqrestore(&iot->lock, flags);
70
71 return r;
72}
73
74static void iot_io_begin(struct io_tracker *iot, sector_t len)
75{
76 unsigned long flags;
77
78 spin_lock_irqsave(&iot->lock, flags);
79 iot->in_flight += len;
80 spin_unlock_irqrestore(&iot->lock, flags);
81}
82
83static void __iot_io_end(struct io_tracker *iot, sector_t len)
84{
85 iot->in_flight -= len;
86 if (!iot->in_flight)
87 iot->idle_time = jiffies;
88}
89
90static void iot_io_end(struct io_tracker *iot, sector_t len)
91{
92 unsigned long flags;
93
94 spin_lock_irqsave(&iot->lock, flags);
95 __iot_io_end(iot, len);
96 spin_unlock_irqrestore(&iot->lock, flags);
97}
98
99/*----------------------------------------------------------------*/
100
101/*
102 * Glossary:
103 *
104 * oblock: index of an origin block
105 * cblock: index of a cache block
106 * promotion: movement of a block from origin to cache
107 * demotion: movement of a block from cache to origin
108 * migration: movement of a block between the origin and cache device,
109 * either direction
110 */
111
112/*----------------------------------------------------------------*/
113
114/*
115 * There are a couple of places where we let a bio run, but want to do some
116 * work before calling its endio function. We do this by temporarily
117 * changing the endio fn.
118 */
119struct dm_hook_info {
120 bio_end_io_t *bi_end_io;
121};
122
123static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
124 bio_end_io_t *bi_end_io, void *bi_private)
125{
126 h->bi_end_io = bio->bi_end_io;
127
128 bio->bi_end_io = bi_end_io;
129 bio->bi_private = bi_private;
130}
131
132static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
133{
134 bio->bi_end_io = h->bi_end_io;
135}
136
137/*----------------------------------------------------------------*/
138
139#define MIGRATION_POOL_SIZE 128
140#define COMMIT_PERIOD HZ
141#define MIGRATION_COUNT_WINDOW 10
142
143/*
144 * The block size of the device holding cache data must be
145 * between 32KB and 1GB.
146 */
147#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
148#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
149
150enum cache_metadata_mode {
151 CM_WRITE, /* metadata may be changed */
152 CM_READ_ONLY, /* metadata may not be changed */
153 CM_FAIL
154};
155
156enum cache_io_mode {
157 /*
158 * Data is written to cached blocks only. These blocks are marked
159 * dirty. If you lose the cache device you will lose data.
160 * Potential performance increase for both reads and writes.
161 */
162 CM_IO_WRITEBACK,
163
164 /*
165 * Data is written to both cache and origin. Blocks are never
166 * dirty. Potential performance benfit for reads only.
167 */
168 CM_IO_WRITETHROUGH,
169
170 /*
171 * A degraded mode useful for various cache coherency situations
172 * (eg, rolling back snapshots). Reads and writes always go to the
173 * origin. If a write goes to a cached oblock, then the cache
174 * block is invalidated.
175 */
176 CM_IO_PASSTHROUGH
177};
178
179struct cache_features {
180 enum cache_metadata_mode mode;
181 enum cache_io_mode io_mode;
182};
183
184struct cache_stats {
185 atomic_t read_hit;
186 atomic_t read_miss;
187 atomic_t write_hit;
188 atomic_t write_miss;
189 atomic_t demotion;
190 atomic_t promotion;
191 atomic_t copies_avoided;
192 atomic_t cache_cell_clash;
193 atomic_t commit_count;
194 atomic_t discard_count;
195};
196
197/*
198 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
199 * the one-past-the-end value.
200 */
201struct cblock_range {
202 dm_cblock_t begin;
203 dm_cblock_t end;
204};
205
206struct invalidation_request {
207 struct list_head list;
208 struct cblock_range *cblocks;
209
210 atomic_t complete;
211 int err;
212
213 wait_queue_head_t result_wait;
214};
215
216struct cache {
217 struct dm_target *ti;
218 struct dm_target_callbacks callbacks;
219
220 struct dm_cache_metadata *cmd;
221
222 /*
223 * Metadata is written to this device.
224 */
225 struct dm_dev *metadata_dev;
226
227 /*
228 * The slower of the two data devices. Typically a spindle.
229 */
230 struct dm_dev *origin_dev;
231
232 /*
233 * The faster of the two data devices. Typically an SSD.
234 */
235 struct dm_dev *cache_dev;
236
237 /*
238 * Size of the origin device in _complete_ blocks and native sectors.
239 */
240 dm_oblock_t origin_blocks;
241 sector_t origin_sectors;
242
243 /*
244 * Size of the cache device in blocks.
245 */
246 dm_cblock_t cache_size;
247
248 /*
249 * Fields for converting from sectors to blocks.
250 */
251 sector_t sectors_per_block;
252 int sectors_per_block_shift;
253
254 spinlock_t lock;
255 struct list_head deferred_cells;
256 struct bio_list deferred_bios;
257 struct bio_list deferred_flush_bios;
258 struct bio_list deferred_writethrough_bios;
259 struct list_head quiesced_migrations;
260 struct list_head completed_migrations;
261 struct list_head need_commit_migrations;
262 sector_t migration_threshold;
263 wait_queue_head_t migration_wait;
264 atomic_t nr_allocated_migrations;
265
266 /*
267 * The number of in flight migrations that are performing
268 * background io. eg, promotion, writeback.
269 */
270 atomic_t nr_io_migrations;
271
272 wait_queue_head_t quiescing_wait;
273 atomic_t quiescing;
274 atomic_t quiescing_ack;
275
276 /*
277 * cache_size entries, dirty if set
278 */
279 atomic_t nr_dirty;
280 unsigned long *dirty_bitset;
281
282 /*
283 * origin_blocks entries, discarded if set.
284 */
285 dm_dblock_t discard_nr_blocks;
286 unsigned long *discard_bitset;
287 uint32_t discard_block_size; /* a power of 2 times sectors per block */
288
289 /*
290 * Rather than reconstructing the table line for the status we just
291 * save it and regurgitate.
292 */
293 unsigned nr_ctr_args;
294 const char **ctr_args;
295
296 struct dm_kcopyd_client *copier;
297 struct workqueue_struct *wq;
298 struct work_struct worker;
299
300 struct delayed_work waker;
301 unsigned long last_commit_jiffies;
302
303 struct dm_bio_prison *prison;
304 struct dm_deferred_set *all_io_ds;
305
306 mempool_t *migration_pool;
307
308 struct dm_cache_policy *policy;
309 unsigned policy_nr_args;
310
311 bool need_tick_bio:1;
312 bool sized:1;
313 bool invalidate:1;
314 bool commit_requested:1;
315 bool loaded_mappings:1;
316 bool loaded_discards:1;
317
318 /*
319 * Cache features such as write-through.
320 */
321 struct cache_features features;
322
323 struct cache_stats stats;
324
325 /*
326 * Invalidation fields.
327 */
328 spinlock_t invalidation_lock;
329 struct list_head invalidation_requests;
330
331 struct io_tracker origin_tracker;
332};
333
334struct per_bio_data {
335 bool tick:1;
336 unsigned req_nr:2;
337 struct dm_deferred_entry *all_io_entry;
338 struct dm_hook_info hook_info;
339 sector_t len;
340
341 /*
342 * writethrough fields. These MUST remain at the end of this
343 * structure and the 'cache' member must be the first as it
344 * is used to determine the offset of the writethrough fields.
345 */
346 struct cache *cache;
347 dm_cblock_t cblock;
348 struct dm_bio_details bio_details;
349};
350
351struct dm_cache_migration {
352 struct list_head list;
353 struct cache *cache;
354
355 unsigned long start_jiffies;
356 dm_oblock_t old_oblock;
357 dm_oblock_t new_oblock;
358 dm_cblock_t cblock;
359
360 bool err:1;
361 bool discard:1;
362 bool writeback:1;
363 bool demote:1;
364 bool promote:1;
365 bool requeue_holder:1;
366 bool invalidate:1;
367
368 struct dm_bio_prison_cell *old_ocell;
369 struct dm_bio_prison_cell *new_ocell;
370};
371
372/*
373 * Processing a bio in the worker thread may require these memory
374 * allocations. We prealloc to avoid deadlocks (the same worker thread
375 * frees them back to the mempool).
376 */
377struct prealloc {
378 struct dm_cache_migration *mg;
379 struct dm_bio_prison_cell *cell1;
380 struct dm_bio_prison_cell *cell2;
381};
382
383static enum cache_metadata_mode get_cache_mode(struct cache *cache);
384
385static void wake_worker(struct cache *cache)
386{
387 queue_work(cache->wq, &cache->worker);
388}
389
390/*----------------------------------------------------------------*/
391
392static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
393{
394 /* FIXME: change to use a local slab. */
395 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
396}
397
398static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
399{
400 dm_bio_prison_free_cell(cache->prison, cell);
401}
402
403static struct dm_cache_migration *alloc_migration(struct cache *cache)
404{
405 struct dm_cache_migration *mg;
406
407 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
408 if (mg) {
409 mg->cache = cache;
410 atomic_inc(&mg->cache->nr_allocated_migrations);
411 }
412
413 return mg;
414}
415
416static void free_migration(struct dm_cache_migration *mg)
417{
418 struct cache *cache = mg->cache;
419
420 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
421 wake_up(&cache->migration_wait);
422
423 mempool_free(mg, cache->migration_pool);
424}
425
426static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
427{
428 if (!p->mg) {
429 p->mg = alloc_migration(cache);
430 if (!p->mg)
431 return -ENOMEM;
432 }
433
434 if (!p->cell1) {
435 p->cell1 = alloc_prison_cell(cache);
436 if (!p->cell1)
437 return -ENOMEM;
438 }
439
440 if (!p->cell2) {
441 p->cell2 = alloc_prison_cell(cache);
442 if (!p->cell2)
443 return -ENOMEM;
444 }
445
446 return 0;
447}
448
449static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
450{
451 if (p->cell2)
452 free_prison_cell(cache, p->cell2);
453
454 if (p->cell1)
455 free_prison_cell(cache, p->cell1);
456
457 if (p->mg)
458 free_migration(p->mg);
459}
460
461static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
462{
463 struct dm_cache_migration *mg = p->mg;
464
465 BUG_ON(!mg);
466 p->mg = NULL;
467
468 return mg;
469}
470
471/*
472 * You must have a cell within the prealloc struct to return. If not this
473 * function will BUG() rather than returning NULL.
474 */
475static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
476{
477 struct dm_bio_prison_cell *r = NULL;
478
479 if (p->cell1) {
480 r = p->cell1;
481 p->cell1 = NULL;
482
483 } else if (p->cell2) {
484 r = p->cell2;
485 p->cell2 = NULL;
486 } else
487 BUG();
488
489 return r;
490}
491
492/*
493 * You can't have more than two cells in a prealloc struct. BUG() will be
494 * called if you try and overfill.
495 */
496static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
497{
498 if (!p->cell2)
499 p->cell2 = cell;
500
501 else if (!p->cell1)
502 p->cell1 = cell;
503
504 else
505 BUG();
506}
507
508/*----------------------------------------------------------------*/
509
510static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key *key)
511{
512 key->virtual = 0;
513 key->dev = 0;
514 key->block_begin = from_oblock(begin);
515 key->block_end = from_oblock(end);
516}
517
518/*
519 * The caller hands in a preallocated cell, and a free function for it.
520 * The cell will be freed if there's an error, or if it wasn't used because
521 * a cell with that key already exists.
522 */
523typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
524
525static int bio_detain_range(struct cache *cache, dm_oblock_t oblock_begin, dm_oblock_t oblock_end,
526 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
527 cell_free_fn free_fn, void *free_context,
528 struct dm_bio_prison_cell **cell_result)
529{
530 int r;
531 struct dm_cell_key key;
532
533 build_key(oblock_begin, oblock_end, &key);
534 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
535 if (r)
536 free_fn(free_context, cell_prealloc);
537
538 return r;
539}
540
541static int bio_detain(struct cache *cache, dm_oblock_t oblock,
542 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
543 cell_free_fn free_fn, void *free_context,
544 struct dm_bio_prison_cell **cell_result)
545{
546 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
547 return bio_detain_range(cache, oblock, end, bio,
548 cell_prealloc, free_fn, free_context, cell_result);
549}
550
551static int get_cell(struct cache *cache,
552 dm_oblock_t oblock,
553 struct prealloc *structs,
554 struct dm_bio_prison_cell **cell_result)
555{
556 int r;
557 struct dm_cell_key key;
558 struct dm_bio_prison_cell *cell_prealloc;
559
560 cell_prealloc = prealloc_get_cell(structs);
561
562 build_key(oblock, to_oblock(from_oblock(oblock) + 1ULL), &key);
563 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
564 if (r)
565 prealloc_put_cell(structs, cell_prealloc);
566
567 return r;
568}
569
570/*----------------------------------------------------------------*/
571
572static bool is_dirty(struct cache *cache, dm_cblock_t b)
573{
574 return test_bit(from_cblock(b), cache->dirty_bitset);
575}
576
577static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
578{
579 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
580 atomic_inc(&cache->nr_dirty);
581 policy_set_dirty(cache->policy, oblock);
582 }
583}
584
585static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
586{
587 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
588 policy_clear_dirty(cache->policy, oblock);
589 if (atomic_dec_return(&cache->nr_dirty) == 0)
590 dm_table_event(cache->ti->table);
591 }
592}
593
594/*----------------------------------------------------------------*/
595
596static bool block_size_is_power_of_two(struct cache *cache)
597{
598 return cache->sectors_per_block_shift >= 0;
599}
600
601/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
602#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
603__always_inline
604#endif
605static dm_block_t block_div(dm_block_t b, uint32_t n)
606{
607 do_div(b, n);
608
609 return b;
610}
611
612static dm_block_t oblocks_per_dblock(struct cache *cache)
613{
614 dm_block_t oblocks = cache->discard_block_size;
615
616 if (block_size_is_power_of_two(cache))
617 oblocks >>= cache->sectors_per_block_shift;
618 else
619 oblocks = block_div(oblocks, cache->sectors_per_block);
620
621 return oblocks;
622}
623
624static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
625{
626 return to_dblock(block_div(from_oblock(oblock),
627 oblocks_per_dblock(cache)));
628}
629
630static dm_oblock_t dblock_to_oblock(struct cache *cache, dm_dblock_t dblock)
631{
632 return to_oblock(from_dblock(dblock) * oblocks_per_dblock(cache));
633}
634
635static void set_discard(struct cache *cache, dm_dblock_t b)
636{
637 unsigned long flags;
638
639 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
640 atomic_inc(&cache->stats.discard_count);
641
642 spin_lock_irqsave(&cache->lock, flags);
643 set_bit(from_dblock(b), cache->discard_bitset);
644 spin_unlock_irqrestore(&cache->lock, flags);
645}
646
647static void clear_discard(struct cache *cache, dm_dblock_t b)
648{
649 unsigned long flags;
650
651 spin_lock_irqsave(&cache->lock, flags);
652 clear_bit(from_dblock(b), cache->discard_bitset);
653 spin_unlock_irqrestore(&cache->lock, flags);
654}
655
656static bool is_discarded(struct cache *cache, dm_dblock_t b)
657{
658 int r;
659 unsigned long flags;
660
661 spin_lock_irqsave(&cache->lock, flags);
662 r = test_bit(from_dblock(b), cache->discard_bitset);
663 spin_unlock_irqrestore(&cache->lock, flags);
664
665 return r;
666}
667
668static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
669{
670 int r;
671 unsigned long flags;
672
673 spin_lock_irqsave(&cache->lock, flags);
674 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
675 cache->discard_bitset);
676 spin_unlock_irqrestore(&cache->lock, flags);
677
678 return r;
679}
680
681/*----------------------------------------------------------------*/
682
683static void load_stats(struct cache *cache)
684{
685 struct dm_cache_statistics stats;
686
687 dm_cache_metadata_get_stats(cache->cmd, &stats);
688 atomic_set(&cache->stats.read_hit, stats.read_hits);
689 atomic_set(&cache->stats.read_miss, stats.read_misses);
690 atomic_set(&cache->stats.write_hit, stats.write_hits);
691 atomic_set(&cache->stats.write_miss, stats.write_misses);
692}
693
694static void save_stats(struct cache *cache)
695{
696 struct dm_cache_statistics stats;
697
698 if (get_cache_mode(cache) >= CM_READ_ONLY)
699 return;
700
701 stats.read_hits = atomic_read(&cache->stats.read_hit);
702 stats.read_misses = atomic_read(&cache->stats.read_miss);
703 stats.write_hits = atomic_read(&cache->stats.write_hit);
704 stats.write_misses = atomic_read(&cache->stats.write_miss);
705
706 dm_cache_metadata_set_stats(cache->cmd, &stats);
707}
708
709/*----------------------------------------------------------------
710 * Per bio data
711 *--------------------------------------------------------------*/
712
713/*
714 * If using writeback, leave out struct per_bio_data's writethrough fields.
715 */
716#define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
717#define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
718
719static bool writethrough_mode(struct cache_features *f)
720{
721 return f->io_mode == CM_IO_WRITETHROUGH;
722}
723
724static bool writeback_mode(struct cache_features *f)
725{
726 return f->io_mode == CM_IO_WRITEBACK;
727}
728
729static bool passthrough_mode(struct cache_features *f)
730{
731 return f->io_mode == CM_IO_PASSTHROUGH;
732}
733
734static size_t get_per_bio_data_size(struct cache *cache)
735{
736 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
737}
738
739static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
740{
741 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
742 BUG_ON(!pb);
743 return pb;
744}
745
746static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
747{
748 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
749
750 pb->tick = false;
751 pb->req_nr = dm_bio_get_target_bio_nr(bio);
752 pb->all_io_entry = NULL;
753 pb->len = 0;
754
755 return pb;
756}
757
758/*----------------------------------------------------------------
759 * Remapping
760 *--------------------------------------------------------------*/
761static void remap_to_origin(struct cache *cache, struct bio *bio)
762{
763 bio->bi_bdev = cache->origin_dev->bdev;
764}
765
766static void remap_to_cache(struct cache *cache, struct bio *bio,
767 dm_cblock_t cblock)
768{
769 sector_t bi_sector = bio->bi_iter.bi_sector;
770 sector_t block = from_cblock(cblock);
771
772 bio->bi_bdev = cache->cache_dev->bdev;
773 if (!block_size_is_power_of_two(cache))
774 bio->bi_iter.bi_sector =
775 (block * cache->sectors_per_block) +
776 sector_div(bi_sector, cache->sectors_per_block);
777 else
778 bio->bi_iter.bi_sector =
779 (block << cache->sectors_per_block_shift) |
780 (bi_sector & (cache->sectors_per_block - 1));
781}
782
783static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
784{
785 unsigned long flags;
786 size_t pb_data_size = get_per_bio_data_size(cache);
787 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
788
789 spin_lock_irqsave(&cache->lock, flags);
790 if (cache->need_tick_bio &&
791 !(bio->bi_opf & (REQ_FUA | REQ_PREFLUSH)) &&
792 bio_op(bio) != REQ_OP_DISCARD) {
793 pb->tick = true;
794 cache->need_tick_bio = false;
795 }
796 spin_unlock_irqrestore(&cache->lock, flags);
797}
798
799static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
800 dm_oblock_t oblock)
801{
802 check_if_tick_bio_needed(cache, bio);
803 remap_to_origin(cache, bio);
804 if (bio_data_dir(bio) == WRITE)
805 clear_discard(cache, oblock_to_dblock(cache, oblock));
806}
807
808static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
809 dm_oblock_t oblock, dm_cblock_t cblock)
810{
811 check_if_tick_bio_needed(cache, bio);
812 remap_to_cache(cache, bio, cblock);
813 if (bio_data_dir(bio) == WRITE) {
814 set_dirty(cache, oblock, cblock);
815 clear_discard(cache, oblock_to_dblock(cache, oblock));
816 }
817}
818
819static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
820{
821 sector_t block_nr = bio->bi_iter.bi_sector;
822
823 if (!block_size_is_power_of_two(cache))
824 (void) sector_div(block_nr, cache->sectors_per_block);
825 else
826 block_nr >>= cache->sectors_per_block_shift;
827
828 return to_oblock(block_nr);
829}
830
831static int bio_triggers_commit(struct cache *cache, struct bio *bio)
832{
833 return bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
834}
835
836/*
837 * You must increment the deferred set whilst the prison cell is held. To
838 * encourage this, we ask for 'cell' to be passed in.
839 */
840static void inc_ds(struct cache *cache, struct bio *bio,
841 struct dm_bio_prison_cell *cell)
842{
843 size_t pb_data_size = get_per_bio_data_size(cache);
844 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
845
846 BUG_ON(!cell);
847 BUG_ON(pb->all_io_entry);
848
849 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
850}
851
852static bool accountable_bio(struct cache *cache, struct bio *bio)
853{
854 return ((bio->bi_bdev == cache->origin_dev->bdev) &&
855 bio_op(bio) != REQ_OP_DISCARD);
856}
857
858static void accounted_begin(struct cache *cache, struct bio *bio)
859{
860 size_t pb_data_size = get_per_bio_data_size(cache);
861 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
862
863 if (accountable_bio(cache, bio)) {
864 pb->len = bio_sectors(bio);
865 iot_io_begin(&cache->origin_tracker, pb->len);
866 }
867}
868
869static void accounted_complete(struct cache *cache, struct bio *bio)
870{
871 size_t pb_data_size = get_per_bio_data_size(cache);
872 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
873
874 iot_io_end(&cache->origin_tracker, pb->len);
875}
876
877static void accounted_request(struct cache *cache, struct bio *bio)
878{
879 accounted_begin(cache, bio);
880 generic_make_request(bio);
881}
882
883static void issue(struct cache *cache, struct bio *bio)
884{
885 unsigned long flags;
886
887 if (!bio_triggers_commit(cache, bio)) {
888 accounted_request(cache, bio);
889 return;
890 }
891
892 /*
893 * Batch together any bios that trigger commits and then issue a
894 * single commit for them in do_worker().
895 */
896 spin_lock_irqsave(&cache->lock, flags);
897 cache->commit_requested = true;
898 bio_list_add(&cache->deferred_flush_bios, bio);
899 spin_unlock_irqrestore(&cache->lock, flags);
900}
901
902static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
903{
904 inc_ds(cache, bio, cell);
905 issue(cache, bio);
906}
907
908static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
909{
910 unsigned long flags;
911
912 spin_lock_irqsave(&cache->lock, flags);
913 bio_list_add(&cache->deferred_writethrough_bios, bio);
914 spin_unlock_irqrestore(&cache->lock, flags);
915
916 wake_worker(cache);
917}
918
919static void writethrough_endio(struct bio *bio)
920{
921 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
922
923 dm_unhook_bio(&pb->hook_info, bio);
924
925 if (bio->bi_error) {
926 bio_endio(bio);
927 return;
928 }
929
930 dm_bio_restore(&pb->bio_details, bio);
931 remap_to_cache(pb->cache, bio, pb->cblock);
932
933 /*
934 * We can't issue this bio directly, since we're in interrupt
935 * context. So it gets put on a bio list for processing by the
936 * worker thread.
937 */
938 defer_writethrough_bio(pb->cache, bio);
939}
940
941/*
942 * When running in writethrough mode we need to send writes to clean blocks
943 * to both the cache and origin devices. In future we'd like to clone the
944 * bio and send them in parallel, but for now we're doing them in
945 * series as this is easier.
946 */
947static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
948 dm_oblock_t oblock, dm_cblock_t cblock)
949{
950 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
951
952 pb->cache = cache;
953 pb->cblock = cblock;
954 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
955 dm_bio_record(&pb->bio_details, bio);
956
957 remap_to_origin_clear_discard(pb->cache, bio, oblock);
958}
959
960/*----------------------------------------------------------------
961 * Failure modes
962 *--------------------------------------------------------------*/
963static enum cache_metadata_mode get_cache_mode(struct cache *cache)
964{
965 return cache->features.mode;
966}
967
968static const char *cache_device_name(struct cache *cache)
969{
970 return dm_device_name(dm_table_get_md(cache->ti->table));
971}
972
973static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
974{
975 const char *descs[] = {
976 "write",
977 "read-only",
978 "fail"
979 };
980
981 dm_table_event(cache->ti->table);
982 DMINFO("%s: switching cache to %s mode",
983 cache_device_name(cache), descs[(int)mode]);
984}
985
986static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
987{
988 bool needs_check;
989 enum cache_metadata_mode old_mode = get_cache_mode(cache);
990
991 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
992 DMERR("%s: unable to read needs_check flag, setting failure mode.",
993 cache_device_name(cache));
994 new_mode = CM_FAIL;
995 }
996
997 if (new_mode == CM_WRITE && needs_check) {
998 DMERR("%s: unable to switch cache to write mode until repaired.",
999 cache_device_name(cache));
1000 if (old_mode != new_mode)
1001 new_mode = old_mode;
1002 else
1003 new_mode = CM_READ_ONLY;
1004 }
1005
1006 /* Never move out of fail mode */
1007 if (old_mode == CM_FAIL)
1008 new_mode = CM_FAIL;
1009
1010 switch (new_mode) {
1011 case CM_FAIL:
1012 case CM_READ_ONLY:
1013 dm_cache_metadata_set_read_only(cache->cmd);
1014 break;
1015
1016 case CM_WRITE:
1017 dm_cache_metadata_set_read_write(cache->cmd);
1018 break;
1019 }
1020
1021 cache->features.mode = new_mode;
1022
1023 if (new_mode != old_mode)
1024 notify_mode_switch(cache, new_mode);
1025}
1026
1027static void abort_transaction(struct cache *cache)
1028{
1029 const char *dev_name = cache_device_name(cache);
1030
1031 if (get_cache_mode(cache) >= CM_READ_ONLY)
1032 return;
1033
1034 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1035 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1036 set_cache_mode(cache, CM_FAIL);
1037 }
1038
1039 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1040 if (dm_cache_metadata_abort(cache->cmd)) {
1041 DMERR("%s: failed to abort metadata transaction", dev_name);
1042 set_cache_mode(cache, CM_FAIL);
1043 }
1044}
1045
1046static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1047{
1048 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1049 cache_device_name(cache), op, r);
1050 abort_transaction(cache);
1051 set_cache_mode(cache, CM_READ_ONLY);
1052}
1053
1054/*----------------------------------------------------------------
1055 * Migration processing
1056 *
1057 * Migration covers moving data from the origin device to the cache, or
1058 * vice versa.
1059 *--------------------------------------------------------------*/
1060static void inc_io_migrations(struct cache *cache)
1061{
1062 atomic_inc(&cache->nr_io_migrations);
1063}
1064
1065static void dec_io_migrations(struct cache *cache)
1066{
1067 atomic_dec(&cache->nr_io_migrations);
1068}
1069
1070static bool discard_or_flush(struct bio *bio)
1071{
1072 return bio_op(bio) == REQ_OP_DISCARD ||
1073 bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1074}
1075
1076static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1077{
1078 if (discard_or_flush(cell->holder)) {
1079 /*
1080 * We have to handle these bios individually.
1081 */
1082 dm_cell_release(cache->prison, cell, &cache->deferred_bios);
1083 free_prison_cell(cache, cell);
1084 } else
1085 list_add_tail(&cell->user_list, &cache->deferred_cells);
1086}
1087
1088static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1089{
1090 unsigned long flags;
1091
1092 if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1093 /*
1094 * There was no prisoner to promote to holder, the
1095 * cell has been released.
1096 */
1097 free_prison_cell(cache, cell);
1098 return;
1099 }
1100
1101 spin_lock_irqsave(&cache->lock, flags);
1102 __cell_defer(cache, cell);
1103 spin_unlock_irqrestore(&cache->lock, flags);
1104
1105 wake_worker(cache);
1106}
1107
1108static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1109{
1110 dm_cell_error(cache->prison, cell, err);
1111 free_prison_cell(cache, cell);
1112}
1113
1114static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1115{
1116 cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1117}
1118
1119static void free_io_migration(struct dm_cache_migration *mg)
1120{
1121 struct cache *cache = mg->cache;
1122
1123 dec_io_migrations(cache);
1124 free_migration(mg);
1125 wake_worker(cache);
1126}
1127
1128static void migration_failure(struct dm_cache_migration *mg)
1129{
1130 struct cache *cache = mg->cache;
1131 const char *dev_name = cache_device_name(cache);
1132
1133 if (mg->writeback) {
1134 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1135 set_dirty(cache, mg->old_oblock, mg->cblock);
1136 cell_defer(cache, mg->old_ocell, false);
1137
1138 } else if (mg->demote) {
1139 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1140 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1141
1142 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1143 if (mg->promote)
1144 cell_defer(cache, mg->new_ocell, true);
1145 } else {
1146 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1147 policy_remove_mapping(cache->policy, mg->new_oblock);
1148 cell_defer(cache, mg->new_ocell, true);
1149 }
1150
1151 free_io_migration(mg);
1152}
1153
1154static void migration_success_pre_commit(struct dm_cache_migration *mg)
1155{
1156 int r;
1157 unsigned long flags;
1158 struct cache *cache = mg->cache;
1159
1160 if (mg->writeback) {
1161 clear_dirty(cache, mg->old_oblock, mg->cblock);
1162 cell_defer(cache, mg->old_ocell, false);
1163 free_io_migration(mg);
1164 return;
1165
1166 } else if (mg->demote) {
1167 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1168 if (r) {
1169 DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1170 cache_device_name(cache));
1171 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1172 policy_force_mapping(cache->policy, mg->new_oblock,
1173 mg->old_oblock);
1174 if (mg->promote)
1175 cell_defer(cache, mg->new_ocell, true);
1176 free_io_migration(mg);
1177 return;
1178 }
1179 } else {
1180 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1181 if (r) {
1182 DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1183 cache_device_name(cache));
1184 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1185 policy_remove_mapping(cache->policy, mg->new_oblock);
1186 free_io_migration(mg);
1187 return;
1188 }
1189 }
1190
1191 spin_lock_irqsave(&cache->lock, flags);
1192 list_add_tail(&mg->list, &cache->need_commit_migrations);
1193 cache->commit_requested = true;
1194 spin_unlock_irqrestore(&cache->lock, flags);
1195}
1196
1197static void migration_success_post_commit(struct dm_cache_migration *mg)
1198{
1199 unsigned long flags;
1200 struct cache *cache = mg->cache;
1201
1202 if (mg->writeback) {
1203 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1204 cache_device_name(cache));
1205 return;
1206
1207 } else if (mg->demote) {
1208 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1209
1210 if (mg->promote) {
1211 mg->demote = false;
1212
1213 spin_lock_irqsave(&cache->lock, flags);
1214 list_add_tail(&mg->list, &cache->quiesced_migrations);
1215 spin_unlock_irqrestore(&cache->lock, flags);
1216
1217 } else {
1218 if (mg->invalidate)
1219 policy_remove_mapping(cache->policy, mg->old_oblock);
1220 free_io_migration(mg);
1221 }
1222
1223 } else {
1224 if (mg->requeue_holder) {
1225 clear_dirty(cache, mg->new_oblock, mg->cblock);
1226 cell_defer(cache, mg->new_ocell, true);
1227 } else {
1228 /*
1229 * The block was promoted via an overwrite, so it's dirty.
1230 */
1231 set_dirty(cache, mg->new_oblock, mg->cblock);
1232 bio_endio(mg->new_ocell->holder);
1233 cell_defer(cache, mg->new_ocell, false);
1234 }
1235 free_io_migration(mg);
1236 }
1237}
1238
1239static void copy_complete(int read_err, unsigned long write_err, void *context)
1240{
1241 unsigned long flags;
1242 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1243 struct cache *cache = mg->cache;
1244
1245 if (read_err || write_err)
1246 mg->err = true;
1247
1248 spin_lock_irqsave(&cache->lock, flags);
1249 list_add_tail(&mg->list, &cache->completed_migrations);
1250 spin_unlock_irqrestore(&cache->lock, flags);
1251
1252 wake_worker(cache);
1253}
1254
1255static void issue_copy(struct dm_cache_migration *mg)
1256{
1257 int r;
1258 struct dm_io_region o_region, c_region;
1259 struct cache *cache = mg->cache;
1260 sector_t cblock = from_cblock(mg->cblock);
1261
1262 o_region.bdev = cache->origin_dev->bdev;
1263 o_region.count = cache->sectors_per_block;
1264
1265 c_region.bdev = cache->cache_dev->bdev;
1266 c_region.sector = cblock * cache->sectors_per_block;
1267 c_region.count = cache->sectors_per_block;
1268
1269 if (mg->writeback || mg->demote) {
1270 /* demote */
1271 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1272 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1273 } else {
1274 /* promote */
1275 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1276 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1277 }
1278
1279 if (r < 0) {
1280 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1281 migration_failure(mg);
1282 }
1283}
1284
1285static void overwrite_endio(struct bio *bio)
1286{
1287 struct dm_cache_migration *mg = bio->bi_private;
1288 struct cache *cache = mg->cache;
1289 size_t pb_data_size = get_per_bio_data_size(cache);
1290 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1291 unsigned long flags;
1292
1293 dm_unhook_bio(&pb->hook_info, bio);
1294
1295 if (bio->bi_error)
1296 mg->err = true;
1297
1298 mg->requeue_holder = false;
1299
1300 spin_lock_irqsave(&cache->lock, flags);
1301 list_add_tail(&mg->list, &cache->completed_migrations);
1302 spin_unlock_irqrestore(&cache->lock, flags);
1303
1304 wake_worker(cache);
1305}
1306
1307static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1308{
1309 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1310 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1311
1312 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1313 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1314
1315 /*
1316 * No need to inc_ds() here, since the cell will be held for the
1317 * duration of the io.
1318 */
1319 accounted_request(mg->cache, bio);
1320}
1321
1322static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1323{
1324 return (bio_data_dir(bio) == WRITE) &&
1325 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1326}
1327
1328static void avoid_copy(struct dm_cache_migration *mg)
1329{
1330 atomic_inc(&mg->cache->stats.copies_avoided);
1331 migration_success_pre_commit(mg);
1332}
1333
1334static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1335 dm_dblock_t *b, dm_dblock_t *e)
1336{
1337 sector_t sb = bio->bi_iter.bi_sector;
1338 sector_t se = bio_end_sector(bio);
1339
1340 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1341
1342 if (se - sb < cache->discard_block_size)
1343 *e = *b;
1344 else
1345 *e = to_dblock(block_div(se, cache->discard_block_size));
1346}
1347
1348static void issue_discard(struct dm_cache_migration *mg)
1349{
1350 dm_dblock_t b, e;
1351 struct bio *bio = mg->new_ocell->holder;
1352 struct cache *cache = mg->cache;
1353
1354 calc_discard_block_range(cache, bio, &b, &e);
1355 while (b != e) {
1356 set_discard(cache, b);
1357 b = to_dblock(from_dblock(b) + 1);
1358 }
1359
1360 bio_endio(bio);
1361 cell_defer(cache, mg->new_ocell, false);
1362 free_migration(mg);
1363 wake_worker(cache);
1364}
1365
1366static void issue_copy_or_discard(struct dm_cache_migration *mg)
1367{
1368 bool avoid;
1369 struct cache *cache = mg->cache;
1370
1371 if (mg->discard) {
1372 issue_discard(mg);
1373 return;
1374 }
1375
1376 if (mg->writeback || mg->demote)
1377 avoid = !is_dirty(cache, mg->cblock) ||
1378 is_discarded_oblock(cache, mg->old_oblock);
1379 else {
1380 struct bio *bio = mg->new_ocell->holder;
1381
1382 avoid = is_discarded_oblock(cache, mg->new_oblock);
1383
1384 if (writeback_mode(&cache->features) &&
1385 !avoid && bio_writes_complete_block(cache, bio)) {
1386 issue_overwrite(mg, bio);
1387 return;
1388 }
1389 }
1390
1391 avoid ? avoid_copy(mg) : issue_copy(mg);
1392}
1393
1394static void complete_migration(struct dm_cache_migration *mg)
1395{
1396 if (mg->err)
1397 migration_failure(mg);
1398 else
1399 migration_success_pre_commit(mg);
1400}
1401
1402static void process_migrations(struct cache *cache, struct list_head *head,
1403 void (*fn)(struct dm_cache_migration *))
1404{
1405 unsigned long flags;
1406 struct list_head list;
1407 struct dm_cache_migration *mg, *tmp;
1408
1409 INIT_LIST_HEAD(&list);
1410 spin_lock_irqsave(&cache->lock, flags);
1411 list_splice_init(head, &list);
1412 spin_unlock_irqrestore(&cache->lock, flags);
1413
1414 list_for_each_entry_safe(mg, tmp, &list, list)
1415 fn(mg);
1416}
1417
1418static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1419{
1420 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1421}
1422
1423static void queue_quiesced_migration(struct dm_cache_migration *mg)
1424{
1425 unsigned long flags;
1426 struct cache *cache = mg->cache;
1427
1428 spin_lock_irqsave(&cache->lock, flags);
1429 __queue_quiesced_migration(mg);
1430 spin_unlock_irqrestore(&cache->lock, flags);
1431
1432 wake_worker(cache);
1433}
1434
1435static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1436{
1437 unsigned long flags;
1438 struct dm_cache_migration *mg, *tmp;
1439
1440 spin_lock_irqsave(&cache->lock, flags);
1441 list_for_each_entry_safe(mg, tmp, work, list)
1442 __queue_quiesced_migration(mg);
1443 spin_unlock_irqrestore(&cache->lock, flags);
1444
1445 wake_worker(cache);
1446}
1447
1448static void check_for_quiesced_migrations(struct cache *cache,
1449 struct per_bio_data *pb)
1450{
1451 struct list_head work;
1452
1453 if (!pb->all_io_entry)
1454 return;
1455
1456 INIT_LIST_HEAD(&work);
1457 dm_deferred_entry_dec(pb->all_io_entry, &work);
1458
1459 if (!list_empty(&work))
1460 queue_quiesced_migrations(cache, &work);
1461}
1462
1463static void quiesce_migration(struct dm_cache_migration *mg)
1464{
1465 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1466 queue_quiesced_migration(mg);
1467}
1468
1469static void promote(struct cache *cache, struct prealloc *structs,
1470 dm_oblock_t oblock, dm_cblock_t cblock,
1471 struct dm_bio_prison_cell *cell)
1472{
1473 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1474
1475 mg->err = false;
1476 mg->discard = false;
1477 mg->writeback = false;
1478 mg->demote = false;
1479 mg->promote = true;
1480 mg->requeue_holder = true;
1481 mg->invalidate = false;
1482 mg->cache = cache;
1483 mg->new_oblock = oblock;
1484 mg->cblock = cblock;
1485 mg->old_ocell = NULL;
1486 mg->new_ocell = cell;
1487 mg->start_jiffies = jiffies;
1488
1489 inc_io_migrations(cache);
1490 quiesce_migration(mg);
1491}
1492
1493static void writeback(struct cache *cache, struct prealloc *structs,
1494 dm_oblock_t oblock, dm_cblock_t cblock,
1495 struct dm_bio_prison_cell *cell)
1496{
1497 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1498
1499 mg->err = false;
1500 mg->discard = false;
1501 mg->writeback = true;
1502 mg->demote = false;
1503 mg->promote = false;
1504 mg->requeue_holder = true;
1505 mg->invalidate = false;
1506 mg->cache = cache;
1507 mg->old_oblock = oblock;
1508 mg->cblock = cblock;
1509 mg->old_ocell = cell;
1510 mg->new_ocell = NULL;
1511 mg->start_jiffies = jiffies;
1512
1513 inc_io_migrations(cache);
1514 quiesce_migration(mg);
1515}
1516
1517static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1518 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1519 dm_cblock_t cblock,
1520 struct dm_bio_prison_cell *old_ocell,
1521 struct dm_bio_prison_cell *new_ocell)
1522{
1523 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1524
1525 mg->err = false;
1526 mg->discard = false;
1527 mg->writeback = false;
1528 mg->demote = true;
1529 mg->promote = true;
1530 mg->requeue_holder = true;
1531 mg->invalidate = false;
1532 mg->cache = cache;
1533 mg->old_oblock = old_oblock;
1534 mg->new_oblock = new_oblock;
1535 mg->cblock = cblock;
1536 mg->old_ocell = old_ocell;
1537 mg->new_ocell = new_ocell;
1538 mg->start_jiffies = jiffies;
1539
1540 inc_io_migrations(cache);
1541 quiesce_migration(mg);
1542}
1543
1544/*
1545 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1546 * block are thrown away.
1547 */
1548static void invalidate(struct cache *cache, struct prealloc *structs,
1549 dm_oblock_t oblock, dm_cblock_t cblock,
1550 struct dm_bio_prison_cell *cell)
1551{
1552 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1553
1554 mg->err = false;
1555 mg->discard = false;
1556 mg->writeback = false;
1557 mg->demote = true;
1558 mg->promote = false;
1559 mg->requeue_holder = true;
1560 mg->invalidate = true;
1561 mg->cache = cache;
1562 mg->old_oblock = oblock;
1563 mg->cblock = cblock;
1564 mg->old_ocell = cell;
1565 mg->new_ocell = NULL;
1566 mg->start_jiffies = jiffies;
1567
1568 inc_io_migrations(cache);
1569 quiesce_migration(mg);
1570}
1571
1572static void discard(struct cache *cache, struct prealloc *structs,
1573 struct dm_bio_prison_cell *cell)
1574{
1575 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1576
1577 mg->err = false;
1578 mg->discard = true;
1579 mg->writeback = false;
1580 mg->demote = false;
1581 mg->promote = false;
1582 mg->requeue_holder = false;
1583 mg->invalidate = false;
1584 mg->cache = cache;
1585 mg->old_ocell = NULL;
1586 mg->new_ocell = cell;
1587 mg->start_jiffies = jiffies;
1588
1589 quiesce_migration(mg);
1590}
1591
1592/*----------------------------------------------------------------
1593 * bio processing
1594 *--------------------------------------------------------------*/
1595static void defer_bio(struct cache *cache, struct bio *bio)
1596{
1597 unsigned long flags;
1598
1599 spin_lock_irqsave(&cache->lock, flags);
1600 bio_list_add(&cache->deferred_bios, bio);
1601 spin_unlock_irqrestore(&cache->lock, flags);
1602
1603 wake_worker(cache);
1604}
1605
1606static void process_flush_bio(struct cache *cache, struct bio *bio)
1607{
1608 size_t pb_data_size = get_per_bio_data_size(cache);
1609 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1610
1611 BUG_ON(bio->bi_iter.bi_size);
1612 if (!pb->req_nr)
1613 remap_to_origin(cache, bio);
1614 else
1615 remap_to_cache(cache, bio, 0);
1616
1617 /*
1618 * REQ_PREFLUSH is not directed at any particular block so we don't
1619 * need to inc_ds(). REQ_FUA's are split into a write + REQ_PREFLUSH
1620 * by dm-core.
1621 */
1622 issue(cache, bio);
1623}
1624
1625static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1626 struct bio *bio)
1627{
1628 int r;
1629 dm_dblock_t b, e;
1630 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1631
1632 calc_discard_block_range(cache, bio, &b, &e);
1633 if (b == e) {
1634 bio_endio(bio);
1635 return;
1636 }
1637
1638 cell_prealloc = prealloc_get_cell(structs);
1639 r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1640 (cell_free_fn) prealloc_put_cell,
1641 structs, &new_ocell);
1642 if (r > 0)
1643 return;
1644
1645 discard(cache, structs, new_ocell);
1646}
1647
1648static bool spare_migration_bandwidth(struct cache *cache)
1649{
1650 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1651 cache->sectors_per_block;
1652 return current_volume < cache->migration_threshold;
1653}
1654
1655static void inc_hit_counter(struct cache *cache, struct bio *bio)
1656{
1657 atomic_inc(bio_data_dir(bio) == READ ?
1658 &cache->stats.read_hit : &cache->stats.write_hit);
1659}
1660
1661static void inc_miss_counter(struct cache *cache, struct bio *bio)
1662{
1663 atomic_inc(bio_data_dir(bio) == READ ?
1664 &cache->stats.read_miss : &cache->stats.write_miss);
1665}
1666
1667/*----------------------------------------------------------------*/
1668
1669struct inc_detail {
1670 struct cache *cache;
1671 struct bio_list bios_for_issue;
1672 struct bio_list unhandled_bios;
1673 bool any_writes;
1674};
1675
1676static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1677{
1678 struct bio *bio;
1679 struct inc_detail *detail = context;
1680 struct cache *cache = detail->cache;
1681
1682 inc_ds(cache, cell->holder, cell);
1683 if (bio_data_dir(cell->holder) == WRITE)
1684 detail->any_writes = true;
1685
1686 while ((bio = bio_list_pop(&cell->bios))) {
1687 if (discard_or_flush(bio)) {
1688 bio_list_add(&detail->unhandled_bios, bio);
1689 continue;
1690 }
1691
1692 if (bio_data_dir(bio) == WRITE)
1693 detail->any_writes = true;
1694
1695 bio_list_add(&detail->bios_for_issue, bio);
1696 inc_ds(cache, bio, cell);
1697 }
1698}
1699
1700// FIXME: refactor these two
1701static void remap_cell_to_origin_clear_discard(struct cache *cache,
1702 struct dm_bio_prison_cell *cell,
1703 dm_oblock_t oblock, bool issue_holder)
1704{
1705 struct bio *bio;
1706 unsigned long flags;
1707 struct inc_detail detail;
1708
1709 detail.cache = cache;
1710 bio_list_init(&detail.bios_for_issue);
1711 bio_list_init(&detail.unhandled_bios);
1712 detail.any_writes = false;
1713
1714 spin_lock_irqsave(&cache->lock, flags);
1715 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1716 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1717 spin_unlock_irqrestore(&cache->lock, flags);
1718
1719 remap_to_origin(cache, cell->holder);
1720 if (issue_holder)
1721 issue(cache, cell->holder);
1722 else
1723 accounted_begin(cache, cell->holder);
1724
1725 if (detail.any_writes)
1726 clear_discard(cache, oblock_to_dblock(cache, oblock));
1727
1728 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1729 remap_to_origin(cache, bio);
1730 issue(cache, bio);
1731 }
1732
1733 free_prison_cell(cache, cell);
1734}
1735
1736static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1737 dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1738{
1739 struct bio *bio;
1740 unsigned long flags;
1741 struct inc_detail detail;
1742
1743 detail.cache = cache;
1744 bio_list_init(&detail.bios_for_issue);
1745 bio_list_init(&detail.unhandled_bios);
1746 detail.any_writes = false;
1747
1748 spin_lock_irqsave(&cache->lock, flags);
1749 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1750 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1751 spin_unlock_irqrestore(&cache->lock, flags);
1752
1753 remap_to_cache(cache, cell->holder, cblock);
1754 if (issue_holder)
1755 issue(cache, cell->holder);
1756 else
1757 accounted_begin(cache, cell->holder);
1758
1759 if (detail.any_writes) {
1760 set_dirty(cache, oblock, cblock);
1761 clear_discard(cache, oblock_to_dblock(cache, oblock));
1762 }
1763
1764 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1765 remap_to_cache(cache, bio, cblock);
1766 issue(cache, bio);
1767 }
1768
1769 free_prison_cell(cache, cell);
1770}
1771
1772/*----------------------------------------------------------------*/
1773
1774struct old_oblock_lock {
1775 struct policy_locker locker;
1776 struct cache *cache;
1777 struct prealloc *structs;
1778 struct dm_bio_prison_cell *cell;
1779};
1780
1781static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1782{
1783 /* This should never be called */
1784 BUG();
1785 return 0;
1786}
1787
1788static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1789{
1790 struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1791 struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1792
1793 return bio_detain(l->cache, b, NULL, cell_prealloc,
1794 (cell_free_fn) prealloc_put_cell,
1795 l->structs, &l->cell);
1796}
1797
1798static void process_cell(struct cache *cache, struct prealloc *structs,
1799 struct dm_bio_prison_cell *new_ocell)
1800{
1801 int r;
1802 bool release_cell = true;
1803 struct bio *bio = new_ocell->holder;
1804 dm_oblock_t block = get_bio_block(cache, bio);
1805 struct policy_result lookup_result;
1806 bool passthrough = passthrough_mode(&cache->features);
1807 bool fast_promotion, can_migrate;
1808 struct old_oblock_lock ool;
1809
1810 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1811 can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1812
1813 ool.locker.fn = cell_locker;
1814 ool.cache = cache;
1815 ool.structs = structs;
1816 ool.cell = NULL;
1817 r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1818 bio, &ool.locker, &lookup_result);
1819
1820 if (r == -EWOULDBLOCK)
1821 /* migration has been denied */
1822 lookup_result.op = POLICY_MISS;
1823
1824 switch (lookup_result.op) {
1825 case POLICY_HIT:
1826 if (passthrough) {
1827 inc_miss_counter(cache, bio);
1828
1829 /*
1830 * Passthrough always maps to the origin,
1831 * invalidating any cache blocks that are written
1832 * to.
1833 */
1834
1835 if (bio_data_dir(bio) == WRITE) {
1836 atomic_inc(&cache->stats.demotion);
1837 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1838 release_cell = false;
1839
1840 } else {
1841 /* FIXME: factor out issue_origin() */
1842 remap_to_origin_clear_discard(cache, bio, block);
1843 inc_and_issue(cache, bio, new_ocell);
1844 }
1845 } else {
1846 inc_hit_counter(cache, bio);
1847
1848 if (bio_data_dir(bio) == WRITE &&
1849 writethrough_mode(&cache->features) &&
1850 !is_dirty(cache, lookup_result.cblock)) {
1851 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1852 inc_and_issue(cache, bio, new_ocell);
1853
1854 } else {
1855 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1856 release_cell = false;
1857 }
1858 }
1859
1860 break;
1861
1862 case POLICY_MISS:
1863 inc_miss_counter(cache, bio);
1864 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1865 release_cell = false;
1866 break;
1867
1868 case POLICY_NEW:
1869 atomic_inc(&cache->stats.promotion);
1870 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1871 release_cell = false;
1872 break;
1873
1874 case POLICY_REPLACE:
1875 atomic_inc(&cache->stats.demotion);
1876 atomic_inc(&cache->stats.promotion);
1877 demote_then_promote(cache, structs, lookup_result.old_oblock,
1878 block, lookup_result.cblock,
1879 ool.cell, new_ocell);
1880 release_cell = false;
1881 break;
1882
1883 default:
1884 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1885 cache_device_name(cache), __func__,
1886 (unsigned) lookup_result.op);
1887 bio_io_error(bio);
1888 }
1889
1890 if (release_cell)
1891 cell_defer(cache, new_ocell, false);
1892}
1893
1894static void process_bio(struct cache *cache, struct prealloc *structs,
1895 struct bio *bio)
1896{
1897 int r;
1898 dm_oblock_t block = get_bio_block(cache, bio);
1899 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1900
1901 /*
1902 * Check to see if that block is currently migrating.
1903 */
1904 cell_prealloc = prealloc_get_cell(structs);
1905 r = bio_detain(cache, block, bio, cell_prealloc,
1906 (cell_free_fn) prealloc_put_cell,
1907 structs, &new_ocell);
1908 if (r > 0)
1909 return;
1910
1911 process_cell(cache, structs, new_ocell);
1912}
1913
1914static int need_commit_due_to_time(struct cache *cache)
1915{
1916 return jiffies < cache->last_commit_jiffies ||
1917 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1918}
1919
1920/*
1921 * A non-zero return indicates read_only or fail_io mode.
1922 */
1923static int commit(struct cache *cache, bool clean_shutdown)
1924{
1925 int r;
1926
1927 if (get_cache_mode(cache) >= CM_READ_ONLY)
1928 return -EINVAL;
1929
1930 atomic_inc(&cache->stats.commit_count);
1931 r = dm_cache_commit(cache->cmd, clean_shutdown);
1932 if (r)
1933 metadata_operation_failed(cache, "dm_cache_commit", r);
1934
1935 return r;
1936}
1937
1938static int commit_if_needed(struct cache *cache)
1939{
1940 int r = 0;
1941
1942 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1943 dm_cache_changed_this_transaction(cache->cmd)) {
1944 r = commit(cache, false);
1945 cache->commit_requested = false;
1946 cache->last_commit_jiffies = jiffies;
1947 }
1948
1949 return r;
1950}
1951
1952static void process_deferred_bios(struct cache *cache)
1953{
1954 bool prealloc_used = false;
1955 unsigned long flags;
1956 struct bio_list bios;
1957 struct bio *bio;
1958 struct prealloc structs;
1959
1960 memset(&structs, 0, sizeof(structs));
1961 bio_list_init(&bios);
1962
1963 spin_lock_irqsave(&cache->lock, flags);
1964 bio_list_merge(&bios, &cache->deferred_bios);
1965 bio_list_init(&cache->deferred_bios);
1966 spin_unlock_irqrestore(&cache->lock, flags);
1967
1968 while (!bio_list_empty(&bios)) {
1969 /*
1970 * If we've got no free migration structs, and processing
1971 * this bio might require one, we pause until there are some
1972 * prepared mappings to process.
1973 */
1974 prealloc_used = true;
1975 if (prealloc_data_structs(cache, &structs)) {
1976 spin_lock_irqsave(&cache->lock, flags);
1977 bio_list_merge(&cache->deferred_bios, &bios);
1978 spin_unlock_irqrestore(&cache->lock, flags);
1979 break;
1980 }
1981
1982 bio = bio_list_pop(&bios);
1983
1984 if (bio->bi_opf & REQ_PREFLUSH)
1985 process_flush_bio(cache, bio);
1986 else if (bio_op(bio) == REQ_OP_DISCARD)
1987 process_discard_bio(cache, &structs, bio);
1988 else
1989 process_bio(cache, &structs, bio);
1990 }
1991
1992 if (prealloc_used)
1993 prealloc_free_structs(cache, &structs);
1994}
1995
1996static void process_deferred_cells(struct cache *cache)
1997{
1998 bool prealloc_used = false;
1999 unsigned long flags;
2000 struct dm_bio_prison_cell *cell, *tmp;
2001 struct list_head cells;
2002 struct prealloc structs;
2003
2004 memset(&structs, 0, sizeof(structs));
2005
2006 INIT_LIST_HEAD(&cells);
2007
2008 spin_lock_irqsave(&cache->lock, flags);
2009 list_splice_init(&cache->deferred_cells, &cells);
2010 spin_unlock_irqrestore(&cache->lock, flags);
2011
2012 list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2013 /*
2014 * If we've got no free migration structs, and processing
2015 * this bio might require one, we pause until there are some
2016 * prepared mappings to process.
2017 */
2018 prealloc_used = true;
2019 if (prealloc_data_structs(cache, &structs)) {
2020 spin_lock_irqsave(&cache->lock, flags);
2021 list_splice(&cells, &cache->deferred_cells);
2022 spin_unlock_irqrestore(&cache->lock, flags);
2023 break;
2024 }
2025
2026 process_cell(cache, &structs, cell);
2027 }
2028
2029 if (prealloc_used)
2030 prealloc_free_structs(cache, &structs);
2031}
2032
2033static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2034{
2035 unsigned long flags;
2036 struct bio_list bios;
2037 struct bio *bio;
2038
2039 bio_list_init(&bios);
2040
2041 spin_lock_irqsave(&cache->lock, flags);
2042 bio_list_merge(&bios, &cache->deferred_flush_bios);
2043 bio_list_init(&cache->deferred_flush_bios);
2044 spin_unlock_irqrestore(&cache->lock, flags);
2045
2046 /*
2047 * These bios have already been through inc_ds()
2048 */
2049 while ((bio = bio_list_pop(&bios)))
2050 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2051}
2052
2053static void process_deferred_writethrough_bios(struct cache *cache)
2054{
2055 unsigned long flags;
2056 struct bio_list bios;
2057 struct bio *bio;
2058
2059 bio_list_init(&bios);
2060
2061 spin_lock_irqsave(&cache->lock, flags);
2062 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2063 bio_list_init(&cache->deferred_writethrough_bios);
2064 spin_unlock_irqrestore(&cache->lock, flags);
2065
2066 /*
2067 * These bios have already been through inc_ds()
2068 */
2069 while ((bio = bio_list_pop(&bios)))
2070 accounted_request(cache, bio);
2071}
2072
2073static void writeback_some_dirty_blocks(struct cache *cache)
2074{
2075 bool prealloc_used = false;
2076 dm_oblock_t oblock;
2077 dm_cblock_t cblock;
2078 struct prealloc structs;
2079 struct dm_bio_prison_cell *old_ocell;
2080 bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2081
2082 memset(&structs, 0, sizeof(structs));
2083
2084 while (spare_migration_bandwidth(cache)) {
2085 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2086 break; /* no work to do */
2087
2088 prealloc_used = true;
2089 if (prealloc_data_structs(cache, &structs) ||
2090 get_cell(cache, oblock, &structs, &old_ocell)) {
2091 policy_set_dirty(cache->policy, oblock);
2092 break;
2093 }
2094
2095 writeback(cache, &structs, oblock, cblock, old_ocell);
2096 }
2097
2098 if (prealloc_used)
2099 prealloc_free_structs(cache, &structs);
2100}
2101
2102/*----------------------------------------------------------------
2103 * Invalidations.
2104 * Dropping something from the cache *without* writing back.
2105 *--------------------------------------------------------------*/
2106
2107static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2108{
2109 int r = 0;
2110 uint64_t begin = from_cblock(req->cblocks->begin);
2111 uint64_t end = from_cblock(req->cblocks->end);
2112
2113 while (begin != end) {
2114 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2115 if (!r) {
2116 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2117 if (r) {
2118 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2119 break;
2120 }
2121
2122 } else if (r == -ENODATA) {
2123 /* harmless, already unmapped */
2124 r = 0;
2125
2126 } else {
2127 DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2128 break;
2129 }
2130
2131 begin++;
2132 }
2133
2134 cache->commit_requested = true;
2135
2136 req->err = r;
2137 atomic_set(&req->complete, 1);
2138
2139 wake_up(&req->result_wait);
2140}
2141
2142static void process_invalidation_requests(struct cache *cache)
2143{
2144 struct list_head list;
2145 struct invalidation_request *req, *tmp;
2146
2147 INIT_LIST_HEAD(&list);
2148 spin_lock(&cache->invalidation_lock);
2149 list_splice_init(&cache->invalidation_requests, &list);
2150 spin_unlock(&cache->invalidation_lock);
2151
2152 list_for_each_entry_safe (req, tmp, &list, list)
2153 process_invalidation_request(cache, req);
2154}
2155
2156/*----------------------------------------------------------------
2157 * Main worker loop
2158 *--------------------------------------------------------------*/
2159static bool is_quiescing(struct cache *cache)
2160{
2161 return atomic_read(&cache->quiescing);
2162}
2163
2164static void ack_quiescing(struct cache *cache)
2165{
2166 if (is_quiescing(cache)) {
2167 atomic_inc(&cache->quiescing_ack);
2168 wake_up(&cache->quiescing_wait);
2169 }
2170}
2171
2172static void wait_for_quiescing_ack(struct cache *cache)
2173{
2174 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2175}
2176
2177static void start_quiescing(struct cache *cache)
2178{
2179 atomic_inc(&cache->quiescing);
2180 wait_for_quiescing_ack(cache);
2181}
2182
2183static void stop_quiescing(struct cache *cache)
2184{
2185 atomic_set(&cache->quiescing, 0);
2186 atomic_set(&cache->quiescing_ack, 0);
2187}
2188
2189static void wait_for_migrations(struct cache *cache)
2190{
2191 wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2192}
2193
2194static void stop_worker(struct cache *cache)
2195{
2196 cancel_delayed_work(&cache->waker);
2197 flush_workqueue(cache->wq);
2198}
2199
2200static void requeue_deferred_cells(struct cache *cache)
2201{
2202 unsigned long flags;
2203 struct list_head cells;
2204 struct dm_bio_prison_cell *cell, *tmp;
2205
2206 INIT_LIST_HEAD(&cells);
2207 spin_lock_irqsave(&cache->lock, flags);
2208 list_splice_init(&cache->deferred_cells, &cells);
2209 spin_unlock_irqrestore(&cache->lock, flags);
2210
2211 list_for_each_entry_safe(cell, tmp, &cells, user_list)
2212 cell_requeue(cache, cell);
2213}
2214
2215static void requeue_deferred_bios(struct cache *cache)
2216{
2217 struct bio *bio;
2218 struct bio_list bios;
2219
2220 bio_list_init(&bios);
2221 bio_list_merge(&bios, &cache->deferred_bios);
2222 bio_list_init(&cache->deferred_bios);
2223
2224 while ((bio = bio_list_pop(&bios))) {
2225 bio->bi_error = DM_ENDIO_REQUEUE;
2226 bio_endio(bio);
2227 }
2228}
2229
2230static int more_work(struct cache *cache)
2231{
2232 if (is_quiescing(cache))
2233 return !list_empty(&cache->quiesced_migrations) ||
2234 !list_empty(&cache->completed_migrations) ||
2235 !list_empty(&cache->need_commit_migrations);
2236 else
2237 return !bio_list_empty(&cache->deferred_bios) ||
2238 !list_empty(&cache->deferred_cells) ||
2239 !bio_list_empty(&cache->deferred_flush_bios) ||
2240 !bio_list_empty(&cache->deferred_writethrough_bios) ||
2241 !list_empty(&cache->quiesced_migrations) ||
2242 !list_empty(&cache->completed_migrations) ||
2243 !list_empty(&cache->need_commit_migrations) ||
2244 cache->invalidate;
2245}
2246
2247static void do_worker(struct work_struct *ws)
2248{
2249 struct cache *cache = container_of(ws, struct cache, worker);
2250
2251 do {
2252 if (!is_quiescing(cache)) {
2253 writeback_some_dirty_blocks(cache);
2254 process_deferred_writethrough_bios(cache);
2255 process_deferred_bios(cache);
2256 process_deferred_cells(cache);
2257 process_invalidation_requests(cache);
2258 }
2259
2260 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2261 process_migrations(cache, &cache->completed_migrations, complete_migration);
2262
2263 if (commit_if_needed(cache)) {
2264 process_deferred_flush_bios(cache, false);
2265 process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2266 } else {
2267 process_deferred_flush_bios(cache, true);
2268 process_migrations(cache, &cache->need_commit_migrations,
2269 migration_success_post_commit);
2270 }
2271
2272 ack_quiescing(cache);
2273
2274 } while (more_work(cache));
2275}
2276
2277/*
2278 * We want to commit periodically so that not too much
2279 * unwritten metadata builds up.
2280 */
2281static void do_waker(struct work_struct *ws)
2282{
2283 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2284 policy_tick(cache->policy, true);
2285 wake_worker(cache);
2286 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2287}
2288
2289/*----------------------------------------------------------------*/
2290
2291static int is_congested(struct dm_dev *dev, int bdi_bits)
2292{
2293 struct request_queue *q = bdev_get_queue(dev->bdev);
2294 return bdi_congested(&q->backing_dev_info, bdi_bits);
2295}
2296
2297static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2298{
2299 struct cache *cache = container_of(cb, struct cache, callbacks);
2300
2301 return is_congested(cache->origin_dev, bdi_bits) ||
2302 is_congested(cache->cache_dev, bdi_bits);
2303}
2304
2305/*----------------------------------------------------------------
2306 * Target methods
2307 *--------------------------------------------------------------*/
2308
2309/*
2310 * This function gets called on the error paths of the constructor, so we
2311 * have to cope with a partially initialised struct.
2312 */
2313static void destroy(struct cache *cache)
2314{
2315 unsigned i;
2316
2317 mempool_destroy(cache->migration_pool);
2318
2319 if (cache->all_io_ds)
2320 dm_deferred_set_destroy(cache->all_io_ds);
2321
2322 if (cache->prison)
2323 dm_bio_prison_destroy(cache->prison);
2324
2325 if (cache->wq)
2326 destroy_workqueue(cache->wq);
2327
2328 if (cache->dirty_bitset)
2329 free_bitset(cache->dirty_bitset);
2330
2331 if (cache->discard_bitset)
2332 free_bitset(cache->discard_bitset);
2333
2334 if (cache->copier)
2335 dm_kcopyd_client_destroy(cache->copier);
2336
2337 if (cache->cmd)
2338 dm_cache_metadata_close(cache->cmd);
2339
2340 if (cache->metadata_dev)
2341 dm_put_device(cache->ti, cache->metadata_dev);
2342
2343 if (cache->origin_dev)
2344 dm_put_device(cache->ti, cache->origin_dev);
2345
2346 if (cache->cache_dev)
2347 dm_put_device(cache->ti, cache->cache_dev);
2348
2349 if (cache->policy)
2350 dm_cache_policy_destroy(cache->policy);
2351
2352 for (i = 0; i < cache->nr_ctr_args ; i++)
2353 kfree(cache->ctr_args[i]);
2354 kfree(cache->ctr_args);
2355
2356 kfree(cache);
2357}
2358
2359static void cache_dtr(struct dm_target *ti)
2360{
2361 struct cache *cache = ti->private;
2362
2363 destroy(cache);
2364}
2365
2366static sector_t get_dev_size(struct dm_dev *dev)
2367{
2368 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2369}
2370
2371/*----------------------------------------------------------------*/
2372
2373/*
2374 * Construct a cache device mapping.
2375 *
2376 * cache <metadata dev> <cache dev> <origin dev> <block size>
2377 * <#feature args> [<feature arg>]*
2378 * <policy> <#policy args> [<policy arg>]*
2379 *
2380 * metadata dev : fast device holding the persistent metadata
2381 * cache dev : fast device holding cached data blocks
2382 * origin dev : slow device holding original data blocks
2383 * block size : cache unit size in sectors
2384 *
2385 * #feature args : number of feature arguments passed
2386 * feature args : writethrough. (The default is writeback.)
2387 *
2388 * policy : the replacement policy to use
2389 * #policy args : an even number of policy arguments corresponding
2390 * to key/value pairs passed to the policy
2391 * policy args : key/value pairs passed to the policy
2392 * E.g. 'sequential_threshold 1024'
2393 * See cache-policies.txt for details.
2394 *
2395 * Optional feature arguments are:
2396 * writethrough : write through caching that prohibits cache block
2397 * content from being different from origin block content.
2398 * Without this argument, the default behaviour is to write
2399 * back cache block contents later for performance reasons,
2400 * so they may differ from the corresponding origin blocks.
2401 */
2402struct cache_args {
2403 struct dm_target *ti;
2404
2405 struct dm_dev *metadata_dev;
2406
2407 struct dm_dev *cache_dev;
2408 sector_t cache_sectors;
2409
2410 struct dm_dev *origin_dev;
2411 sector_t origin_sectors;
2412
2413 uint32_t block_size;
2414
2415 const char *policy_name;
2416 int policy_argc;
2417 const char **policy_argv;
2418
2419 struct cache_features features;
2420};
2421
2422static void destroy_cache_args(struct cache_args *ca)
2423{
2424 if (ca->metadata_dev)
2425 dm_put_device(ca->ti, ca->metadata_dev);
2426
2427 if (ca->cache_dev)
2428 dm_put_device(ca->ti, ca->cache_dev);
2429
2430 if (ca->origin_dev)
2431 dm_put_device(ca->ti, ca->origin_dev);
2432
2433 kfree(ca);
2434}
2435
2436static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2437{
2438 if (!as->argc) {
2439 *error = "Insufficient args";
2440 return false;
2441 }
2442
2443 return true;
2444}
2445
2446static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2447 char **error)
2448{
2449 int r;
2450 sector_t metadata_dev_size;
2451 char b[BDEVNAME_SIZE];
2452
2453 if (!at_least_one_arg(as, error))
2454 return -EINVAL;
2455
2456 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2457 &ca->metadata_dev);
2458 if (r) {
2459 *error = "Error opening metadata device";
2460 return r;
2461 }
2462
2463 metadata_dev_size = get_dev_size(ca->metadata_dev);
2464 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2465 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2466 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2467
2468 return 0;
2469}
2470
2471static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2472 char **error)
2473{
2474 int r;
2475
2476 if (!at_least_one_arg(as, error))
2477 return -EINVAL;
2478
2479 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2480 &ca->cache_dev);
2481 if (r) {
2482 *error = "Error opening cache device";
2483 return r;
2484 }
2485 ca->cache_sectors = get_dev_size(ca->cache_dev);
2486
2487 return 0;
2488}
2489
2490static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2491 char **error)
2492{
2493 int r;
2494
2495 if (!at_least_one_arg(as, error))
2496 return -EINVAL;
2497
2498 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2499 &ca->origin_dev);
2500 if (r) {
2501 *error = "Error opening origin device";
2502 return r;
2503 }
2504
2505 ca->origin_sectors = get_dev_size(ca->origin_dev);
2506 if (ca->ti->len > ca->origin_sectors) {
2507 *error = "Device size larger than cached device";
2508 return -EINVAL;
2509 }
2510
2511 return 0;
2512}
2513
2514static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2515 char **error)
2516{
2517 unsigned long block_size;
2518
2519 if (!at_least_one_arg(as, error))
2520 return -EINVAL;
2521
2522 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2523 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2524 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2525 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2526 *error = "Invalid data block size";
2527 return -EINVAL;
2528 }
2529
2530 if (block_size > ca->cache_sectors) {
2531 *error = "Data block size is larger than the cache device";
2532 return -EINVAL;
2533 }
2534
2535 ca->block_size = block_size;
2536
2537 return 0;
2538}
2539
2540static void init_features(struct cache_features *cf)
2541{
2542 cf->mode = CM_WRITE;
2543 cf->io_mode = CM_IO_WRITEBACK;
2544}
2545
2546static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2547 char **error)
2548{
2549 static struct dm_arg _args[] = {
2550 {0, 1, "Invalid number of cache feature arguments"},
2551 };
2552
2553 int r;
2554 unsigned argc;
2555 const char *arg;
2556 struct cache_features *cf = &ca->features;
2557
2558 init_features(cf);
2559
2560 r = dm_read_arg_group(_args, as, &argc, error);
2561 if (r)
2562 return -EINVAL;
2563
2564 while (argc--) {
2565 arg = dm_shift_arg(as);
2566
2567 if (!strcasecmp(arg, "writeback"))
2568 cf->io_mode = CM_IO_WRITEBACK;
2569
2570 else if (!strcasecmp(arg, "writethrough"))
2571 cf->io_mode = CM_IO_WRITETHROUGH;
2572
2573 else if (!strcasecmp(arg, "passthrough"))
2574 cf->io_mode = CM_IO_PASSTHROUGH;
2575
2576 else {
2577 *error = "Unrecognised cache feature requested";
2578 return -EINVAL;
2579 }
2580 }
2581
2582 return 0;
2583}
2584
2585static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2586 char **error)
2587{
2588 static struct dm_arg _args[] = {
2589 {0, 1024, "Invalid number of policy arguments"},
2590 };
2591
2592 int r;
2593
2594 if (!at_least_one_arg(as, error))
2595 return -EINVAL;
2596
2597 ca->policy_name = dm_shift_arg(as);
2598
2599 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2600 if (r)
2601 return -EINVAL;
2602
2603 ca->policy_argv = (const char **)as->argv;
2604 dm_consume_args(as, ca->policy_argc);
2605
2606 return 0;
2607}
2608
2609static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2610 char **error)
2611{
2612 int r;
2613 struct dm_arg_set as;
2614
2615 as.argc = argc;
2616 as.argv = argv;
2617
2618 r = parse_metadata_dev(ca, &as, error);
2619 if (r)
2620 return r;
2621
2622 r = parse_cache_dev(ca, &as, error);
2623 if (r)
2624 return r;
2625
2626 r = parse_origin_dev(ca, &as, error);
2627 if (r)
2628 return r;
2629
2630 r = parse_block_size(ca, &as, error);
2631 if (r)
2632 return r;
2633
2634 r = parse_features(ca, &as, error);
2635 if (r)
2636 return r;
2637
2638 r = parse_policy(ca, &as, error);
2639 if (r)
2640 return r;
2641
2642 return 0;
2643}
2644
2645/*----------------------------------------------------------------*/
2646
2647static struct kmem_cache *migration_cache;
2648
2649#define NOT_CORE_OPTION 1
2650
2651static int process_config_option(struct cache *cache, const char *key, const char *value)
2652{
2653 unsigned long tmp;
2654
2655 if (!strcasecmp(key, "migration_threshold")) {
2656 if (kstrtoul(value, 10, &tmp))
2657 return -EINVAL;
2658
2659 cache->migration_threshold = tmp;
2660 return 0;
2661 }
2662
2663 return NOT_CORE_OPTION;
2664}
2665
2666static int set_config_value(struct cache *cache, const char *key, const char *value)
2667{
2668 int r = process_config_option(cache, key, value);
2669
2670 if (r == NOT_CORE_OPTION)
2671 r = policy_set_config_value(cache->policy, key, value);
2672
2673 if (r)
2674 DMWARN("bad config value for %s: %s", key, value);
2675
2676 return r;
2677}
2678
2679static int set_config_values(struct cache *cache, int argc, const char **argv)
2680{
2681 int r = 0;
2682
2683 if (argc & 1) {
2684 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2685 return -EINVAL;
2686 }
2687
2688 while (argc) {
2689 r = set_config_value(cache, argv[0], argv[1]);
2690 if (r)
2691 break;
2692
2693 argc -= 2;
2694 argv += 2;
2695 }
2696
2697 return r;
2698}
2699
2700static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2701 char **error)
2702{
2703 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2704 cache->cache_size,
2705 cache->origin_sectors,
2706 cache->sectors_per_block);
2707 if (IS_ERR(p)) {
2708 *error = "Error creating cache's policy";
2709 return PTR_ERR(p);
2710 }
2711 cache->policy = p;
2712
2713 return 0;
2714}
2715
2716/*
2717 * We want the discard block size to be at least the size of the cache
2718 * block size and have no more than 2^14 discard blocks across the origin.
2719 */
2720#define MAX_DISCARD_BLOCKS (1 << 14)
2721
2722static bool too_many_discard_blocks(sector_t discard_block_size,
2723 sector_t origin_size)
2724{
2725 (void) sector_div(origin_size, discard_block_size);
2726
2727 return origin_size > MAX_DISCARD_BLOCKS;
2728}
2729
2730static sector_t calculate_discard_block_size(sector_t cache_block_size,
2731 sector_t origin_size)
2732{
2733 sector_t discard_block_size = cache_block_size;
2734
2735 if (origin_size)
2736 while (too_many_discard_blocks(discard_block_size, origin_size))
2737 discard_block_size *= 2;
2738
2739 return discard_block_size;
2740}
2741
2742static void set_cache_size(struct cache *cache, dm_cblock_t size)
2743{
2744 dm_block_t nr_blocks = from_cblock(size);
2745
2746 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2747 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2748 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2749 "Please consider increasing the cache block size to reduce the overall cache block count.",
2750 (unsigned long long) nr_blocks);
2751
2752 cache->cache_size = size;
2753}
2754
2755#define DEFAULT_MIGRATION_THRESHOLD 2048
2756
2757static int cache_create(struct cache_args *ca, struct cache **result)
2758{
2759 int r = 0;
2760 char **error = &ca->ti->error;
2761 struct cache *cache;
2762 struct dm_target *ti = ca->ti;
2763 dm_block_t origin_blocks;
2764 struct dm_cache_metadata *cmd;
2765 bool may_format = ca->features.mode == CM_WRITE;
2766
2767 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2768 if (!cache)
2769 return -ENOMEM;
2770
2771 cache->ti = ca->ti;
2772 ti->private = cache;
2773 ti->num_flush_bios = 2;
2774 ti->flush_supported = true;
2775
2776 ti->num_discard_bios = 1;
2777 ti->discards_supported = true;
2778 ti->discard_zeroes_data_unsupported = true;
2779 ti->split_discard_bios = false;
2780
2781 cache->features = ca->features;
2782 ti->per_io_data_size = get_per_bio_data_size(cache);
2783
2784 cache->callbacks.congested_fn = cache_is_congested;
2785 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2786
2787 cache->metadata_dev = ca->metadata_dev;
2788 cache->origin_dev = ca->origin_dev;
2789 cache->cache_dev = ca->cache_dev;
2790
2791 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2792
2793 /* FIXME: factor out this whole section */
2794 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2795 origin_blocks = block_div(origin_blocks, ca->block_size);
2796 cache->origin_blocks = to_oblock(origin_blocks);
2797
2798 cache->sectors_per_block = ca->block_size;
2799 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2800 r = -EINVAL;
2801 goto bad;
2802 }
2803
2804 if (ca->block_size & (ca->block_size - 1)) {
2805 dm_block_t cache_size = ca->cache_sectors;
2806
2807 cache->sectors_per_block_shift = -1;
2808 cache_size = block_div(cache_size, ca->block_size);
2809 set_cache_size(cache, to_cblock(cache_size));
2810 } else {
2811 cache->sectors_per_block_shift = __ffs(ca->block_size);
2812 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2813 }
2814
2815 r = create_cache_policy(cache, ca, error);
2816 if (r)
2817 goto bad;
2818
2819 cache->policy_nr_args = ca->policy_argc;
2820 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2821
2822 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2823 if (r) {
2824 *error = "Error setting cache policy's config values";
2825 goto bad;
2826 }
2827
2828 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2829 ca->block_size, may_format,
2830 dm_cache_policy_get_hint_size(cache->policy));
2831 if (IS_ERR(cmd)) {
2832 *error = "Error creating metadata object";
2833 r = PTR_ERR(cmd);
2834 goto bad;
2835 }
2836 cache->cmd = cmd;
2837 set_cache_mode(cache, CM_WRITE);
2838 if (get_cache_mode(cache) != CM_WRITE) {
2839 *error = "Unable to get write access to metadata, please check/repair metadata.";
2840 r = -EINVAL;
2841 goto bad;
2842 }
2843
2844 if (passthrough_mode(&cache->features)) {
2845 bool all_clean;
2846
2847 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2848 if (r) {
2849 *error = "dm_cache_metadata_all_clean() failed";
2850 goto bad;
2851 }
2852
2853 if (!all_clean) {
2854 *error = "Cannot enter passthrough mode unless all blocks are clean";
2855 r = -EINVAL;
2856 goto bad;
2857 }
2858 }
2859
2860 spin_lock_init(&cache->lock);
2861 INIT_LIST_HEAD(&cache->deferred_cells);
2862 bio_list_init(&cache->deferred_bios);
2863 bio_list_init(&cache->deferred_flush_bios);
2864 bio_list_init(&cache->deferred_writethrough_bios);
2865 INIT_LIST_HEAD(&cache->quiesced_migrations);
2866 INIT_LIST_HEAD(&cache->completed_migrations);
2867 INIT_LIST_HEAD(&cache->need_commit_migrations);
2868 atomic_set(&cache->nr_allocated_migrations, 0);
2869 atomic_set(&cache->nr_io_migrations, 0);
2870 init_waitqueue_head(&cache->migration_wait);
2871
2872 init_waitqueue_head(&cache->quiescing_wait);
2873 atomic_set(&cache->quiescing, 0);
2874 atomic_set(&cache->quiescing_ack, 0);
2875
2876 r = -ENOMEM;
2877 atomic_set(&cache->nr_dirty, 0);
2878 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2879 if (!cache->dirty_bitset) {
2880 *error = "could not allocate dirty bitset";
2881 goto bad;
2882 }
2883 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2884
2885 cache->discard_block_size =
2886 calculate_discard_block_size(cache->sectors_per_block,
2887 cache->origin_sectors);
2888 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2889 cache->discard_block_size));
2890 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2891 if (!cache->discard_bitset) {
2892 *error = "could not allocate discard bitset";
2893 goto bad;
2894 }
2895 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2896
2897 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2898 if (IS_ERR(cache->copier)) {
2899 *error = "could not create kcopyd client";
2900 r = PTR_ERR(cache->copier);
2901 goto bad;
2902 }
2903
2904 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2905 if (!cache->wq) {
2906 *error = "could not create workqueue for metadata object";
2907 goto bad;
2908 }
2909 INIT_WORK(&cache->worker, do_worker);
2910 INIT_DELAYED_WORK(&cache->waker, do_waker);
2911 cache->last_commit_jiffies = jiffies;
2912
2913 cache->prison = dm_bio_prison_create();
2914 if (!cache->prison) {
2915 *error = "could not create bio prison";
2916 goto bad;
2917 }
2918
2919 cache->all_io_ds = dm_deferred_set_create();
2920 if (!cache->all_io_ds) {
2921 *error = "could not create all_io deferred set";
2922 goto bad;
2923 }
2924
2925 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2926 migration_cache);
2927 if (!cache->migration_pool) {
2928 *error = "Error creating cache's migration mempool";
2929 goto bad;
2930 }
2931
2932 cache->need_tick_bio = true;
2933 cache->sized = false;
2934 cache->invalidate = false;
2935 cache->commit_requested = false;
2936 cache->loaded_mappings = false;
2937 cache->loaded_discards = false;
2938
2939 load_stats(cache);
2940
2941 atomic_set(&cache->stats.demotion, 0);
2942 atomic_set(&cache->stats.promotion, 0);
2943 atomic_set(&cache->stats.copies_avoided, 0);
2944 atomic_set(&cache->stats.cache_cell_clash, 0);
2945 atomic_set(&cache->stats.commit_count, 0);
2946 atomic_set(&cache->stats.discard_count, 0);
2947
2948 spin_lock_init(&cache->invalidation_lock);
2949 INIT_LIST_HEAD(&cache->invalidation_requests);
2950
2951 iot_init(&cache->origin_tracker);
2952
2953 *result = cache;
2954 return 0;
2955
2956bad:
2957 destroy(cache);
2958 return r;
2959}
2960
2961static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2962{
2963 unsigned i;
2964 const char **copy;
2965
2966 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2967 if (!copy)
2968 return -ENOMEM;
2969 for (i = 0; i < argc; i++) {
2970 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2971 if (!copy[i]) {
2972 while (i--)
2973 kfree(copy[i]);
2974 kfree(copy);
2975 return -ENOMEM;
2976 }
2977 }
2978
2979 cache->nr_ctr_args = argc;
2980 cache->ctr_args = copy;
2981
2982 return 0;
2983}
2984
2985static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2986{
2987 int r = -EINVAL;
2988 struct cache_args *ca;
2989 struct cache *cache = NULL;
2990
2991 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2992 if (!ca) {
2993 ti->error = "Error allocating memory for cache";
2994 return -ENOMEM;
2995 }
2996 ca->ti = ti;
2997
2998 r = parse_cache_args(ca, argc, argv, &ti->error);
2999 if (r)
3000 goto out;
3001
3002 r = cache_create(ca, &cache);
3003 if (r)
3004 goto out;
3005
3006 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
3007 if (r) {
3008 destroy(cache);
3009 goto out;
3010 }
3011
3012 ti->private = cache;
3013
3014out:
3015 destroy_cache_args(ca);
3016 return r;
3017}
3018
3019/*----------------------------------------------------------------*/
3020
3021static int cache_map(struct dm_target *ti, struct bio *bio)
3022{
3023 struct cache *cache = ti->private;
3024
3025 int r;
3026 struct dm_bio_prison_cell *cell = NULL;
3027 dm_oblock_t block = get_bio_block(cache, bio);
3028 size_t pb_data_size = get_per_bio_data_size(cache);
3029 bool can_migrate = false;
3030 bool fast_promotion;
3031 struct policy_result lookup_result;
3032 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3033 struct old_oblock_lock ool;
3034
3035 ool.locker.fn = null_locker;
3036
3037 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3038 /*
3039 * This can only occur if the io goes to a partial block at
3040 * the end of the origin device. We don't cache these.
3041 * Just remap to the origin and carry on.
3042 */
3043 remap_to_origin(cache, bio);
3044 accounted_begin(cache, bio);
3045 return DM_MAPIO_REMAPPED;
3046 }
3047
3048 if (discard_or_flush(bio)) {
3049 defer_bio(cache, bio);
3050 return DM_MAPIO_SUBMITTED;
3051 }
3052
3053 /*
3054 * Check to see if that block is currently migrating.
3055 */
3056 cell = alloc_prison_cell(cache);
3057 if (!cell) {
3058 defer_bio(cache, bio);
3059 return DM_MAPIO_SUBMITTED;
3060 }
3061
3062 r = bio_detain(cache, block, bio, cell,
3063 (cell_free_fn) free_prison_cell,
3064 cache, &cell);
3065 if (r) {
3066 if (r < 0)
3067 defer_bio(cache, bio);
3068
3069 return DM_MAPIO_SUBMITTED;
3070 }
3071
3072 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3073
3074 r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3075 bio, &ool.locker, &lookup_result);
3076 if (r == -EWOULDBLOCK) {
3077 cell_defer(cache, cell, true);
3078 return DM_MAPIO_SUBMITTED;
3079
3080 } else if (r) {
3081 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3082 cache_device_name(cache), r);
3083 cell_defer(cache, cell, false);
3084 bio_io_error(bio);
3085 return DM_MAPIO_SUBMITTED;
3086 }
3087
3088 r = DM_MAPIO_REMAPPED;
3089 switch (lookup_result.op) {
3090 case POLICY_HIT:
3091 if (passthrough_mode(&cache->features)) {
3092 if (bio_data_dir(bio) == WRITE) {
3093 /*
3094 * We need to invalidate this block, so
3095 * defer for the worker thread.
3096 */
3097 cell_defer(cache, cell, true);
3098 r = DM_MAPIO_SUBMITTED;
3099
3100 } else {
3101 inc_miss_counter(cache, bio);
3102 remap_to_origin_clear_discard(cache, bio, block);
3103 accounted_begin(cache, bio);
3104 inc_ds(cache, bio, cell);
3105 // FIXME: we want to remap hits or misses straight
3106 // away rather than passing over to the worker.
3107 cell_defer(cache, cell, false);
3108 }
3109
3110 } else {
3111 inc_hit_counter(cache, bio);
3112 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3113 !is_dirty(cache, lookup_result.cblock)) {
3114 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3115 accounted_begin(cache, bio);
3116 inc_ds(cache, bio, cell);
3117 cell_defer(cache, cell, false);
3118
3119 } else
3120 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3121 }
3122 break;
3123
3124 case POLICY_MISS:
3125 inc_miss_counter(cache, bio);
3126 if (pb->req_nr != 0) {
3127 /*
3128 * This is a duplicate writethrough io that is no
3129 * longer needed because the block has been demoted.
3130 */
3131 bio_endio(bio);
3132 // FIXME: remap everything as a miss
3133 cell_defer(cache, cell, false);
3134 r = DM_MAPIO_SUBMITTED;
3135
3136 } else
3137 remap_cell_to_origin_clear_discard(cache, cell, block, false);
3138 break;
3139
3140 default:
3141 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3142 cache_device_name(cache), __func__,
3143 (unsigned) lookup_result.op);
3144 cell_defer(cache, cell, false);
3145 bio_io_error(bio);
3146 r = DM_MAPIO_SUBMITTED;
3147 }
3148
3149 return r;
3150}
3151
3152static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3153{
3154 struct cache *cache = ti->private;
3155 unsigned long flags;
3156 size_t pb_data_size = get_per_bio_data_size(cache);
3157 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3158
3159 if (pb->tick) {
3160 policy_tick(cache->policy, false);
3161
3162 spin_lock_irqsave(&cache->lock, flags);
3163 cache->need_tick_bio = true;
3164 spin_unlock_irqrestore(&cache->lock, flags);
3165 }
3166
3167 check_for_quiesced_migrations(cache, pb);
3168 accounted_complete(cache, bio);
3169
3170 return 0;
3171}
3172
3173static int write_dirty_bitset(struct cache *cache)
3174{
3175 unsigned i, r;
3176
3177 if (get_cache_mode(cache) >= CM_READ_ONLY)
3178 return -EINVAL;
3179
3180 for (i = 0; i < from_cblock(cache->cache_size); i++) {
3181 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3182 is_dirty(cache, to_cblock(i)));
3183 if (r) {
3184 metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3185 return r;
3186 }
3187 }
3188
3189 return 0;
3190}
3191
3192static int write_discard_bitset(struct cache *cache)
3193{
3194 unsigned i, r;
3195
3196 if (get_cache_mode(cache) >= CM_READ_ONLY)
3197 return -EINVAL;
3198
3199 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3200 cache->discard_nr_blocks);
3201 if (r) {
3202 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3203 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3204 return r;
3205 }
3206
3207 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3208 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3209 is_discarded(cache, to_dblock(i)));
3210 if (r) {
3211 metadata_operation_failed(cache, "dm_cache_set_discard", r);
3212 return r;
3213 }
3214 }
3215
3216 return 0;
3217}
3218
3219static int write_hints(struct cache *cache)
3220{
3221 int r;
3222
3223 if (get_cache_mode(cache) >= CM_READ_ONLY)
3224 return -EINVAL;
3225
3226 r = dm_cache_write_hints(cache->cmd, cache->policy);
3227 if (r) {
3228 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3229 return r;
3230 }
3231
3232 return 0;
3233}
3234
3235/*
3236 * returns true on success
3237 */
3238static bool sync_metadata(struct cache *cache)
3239{
3240 int r1, r2, r3, r4;
3241
3242 r1 = write_dirty_bitset(cache);
3243 if (r1)
3244 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3245
3246 r2 = write_discard_bitset(cache);
3247 if (r2)
3248 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3249
3250 save_stats(cache);
3251
3252 r3 = write_hints(cache);
3253 if (r3)
3254 DMERR("%s: could not write hints", cache_device_name(cache));
3255
3256 /*
3257 * If writing the above metadata failed, we still commit, but don't
3258 * set the clean shutdown flag. This will effectively force every
3259 * dirty bit to be set on reload.
3260 */
3261 r4 = commit(cache, !r1 && !r2 && !r3);
3262 if (r4)
3263 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3264
3265 return !r1 && !r2 && !r3 && !r4;
3266}
3267
3268static void cache_postsuspend(struct dm_target *ti)
3269{
3270 struct cache *cache = ti->private;
3271
3272 start_quiescing(cache);
3273 wait_for_migrations(cache);
3274 stop_worker(cache);
3275 requeue_deferred_bios(cache);
3276 requeue_deferred_cells(cache);
3277 stop_quiescing(cache);
3278
3279 if (get_cache_mode(cache) == CM_WRITE)
3280 (void) sync_metadata(cache);
3281}
3282
3283static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3284 bool dirty, uint32_t hint, bool hint_valid)
3285{
3286 int r;
3287 struct cache *cache = context;
3288
3289 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3290 if (r)
3291 return r;
3292
3293 if (dirty)
3294 set_dirty(cache, oblock, cblock);
3295 else
3296 clear_dirty(cache, oblock, cblock);
3297
3298 return 0;
3299}
3300
3301/*
3302 * The discard block size in the on disk metadata is not
3303 * neccessarily the same as we're currently using. So we have to
3304 * be careful to only set the discarded attribute if we know it
3305 * covers a complete block of the new size.
3306 */
3307struct discard_load_info {
3308 struct cache *cache;
3309
3310 /*
3311 * These blocks are sized using the on disk dblock size, rather
3312 * than the current one.
3313 */
3314 dm_block_t block_size;
3315 dm_block_t discard_begin, discard_end;
3316};
3317
3318static void discard_load_info_init(struct cache *cache,
3319 struct discard_load_info *li)
3320{
3321 li->cache = cache;
3322 li->discard_begin = li->discard_end = 0;
3323}
3324
3325static void set_discard_range(struct discard_load_info *li)
3326{
3327 sector_t b, e;
3328
3329 if (li->discard_begin == li->discard_end)
3330 return;
3331
3332 /*
3333 * Convert to sectors.
3334 */
3335 b = li->discard_begin * li->block_size;
3336 e = li->discard_end * li->block_size;
3337
3338 /*
3339 * Then convert back to the current dblock size.
3340 */
3341 b = dm_sector_div_up(b, li->cache->discard_block_size);
3342 sector_div(e, li->cache->discard_block_size);
3343
3344 /*
3345 * The origin may have shrunk, so we need to check we're still in
3346 * bounds.
3347 */
3348 if (e > from_dblock(li->cache->discard_nr_blocks))
3349 e = from_dblock(li->cache->discard_nr_blocks);
3350
3351 for (; b < e; b++)
3352 set_discard(li->cache, to_dblock(b));
3353}
3354
3355static int load_discard(void *context, sector_t discard_block_size,
3356 dm_dblock_t dblock, bool discard)
3357{
3358 struct discard_load_info *li = context;
3359
3360 li->block_size = discard_block_size;
3361
3362 if (discard) {
3363 if (from_dblock(dblock) == li->discard_end)
3364 /*
3365 * We're already in a discard range, just extend it.
3366 */
3367 li->discard_end = li->discard_end + 1ULL;
3368
3369 else {
3370 /*
3371 * Emit the old range and start a new one.
3372 */
3373 set_discard_range(li);
3374 li->discard_begin = from_dblock(dblock);
3375 li->discard_end = li->discard_begin + 1ULL;
3376 }
3377 } else {
3378 set_discard_range(li);
3379 li->discard_begin = li->discard_end = 0;
3380 }
3381
3382 return 0;
3383}
3384
3385static dm_cblock_t get_cache_dev_size(struct cache *cache)
3386{
3387 sector_t size = get_dev_size(cache->cache_dev);
3388 (void) sector_div(size, cache->sectors_per_block);
3389 return to_cblock(size);
3390}
3391
3392static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3393{
3394 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3395 return true;
3396
3397 /*
3398 * We can't drop a dirty block when shrinking the cache.
3399 */
3400 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3401 new_size = to_cblock(from_cblock(new_size) + 1);
3402 if (is_dirty(cache, new_size)) {
3403 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3404 cache_device_name(cache),
3405 (unsigned long long) from_cblock(new_size));
3406 return false;
3407 }
3408 }
3409
3410 return true;
3411}
3412
3413static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3414{
3415 int r;
3416
3417 r = dm_cache_resize(cache->cmd, new_size);
3418 if (r) {
3419 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3420 metadata_operation_failed(cache, "dm_cache_resize", r);
3421 return r;
3422 }
3423
3424 set_cache_size(cache, new_size);
3425
3426 return 0;
3427}
3428
3429static int cache_preresume(struct dm_target *ti)
3430{
3431 int r = 0;
3432 struct cache *cache = ti->private;
3433 dm_cblock_t csize = get_cache_dev_size(cache);
3434
3435 /*
3436 * Check to see if the cache has resized.
3437 */
3438 if (!cache->sized) {
3439 r = resize_cache_dev(cache, csize);
3440 if (r)
3441 return r;
3442
3443 cache->sized = true;
3444
3445 } else if (csize != cache->cache_size) {
3446 if (!can_resize(cache, csize))
3447 return -EINVAL;
3448
3449 r = resize_cache_dev(cache, csize);
3450 if (r)
3451 return r;
3452 }
3453
3454 if (!cache->loaded_mappings) {
3455 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3456 load_mapping, cache);
3457 if (r) {
3458 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3459 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3460 return r;
3461 }
3462
3463 cache->loaded_mappings = true;
3464 }
3465
3466 if (!cache->loaded_discards) {
3467 struct discard_load_info li;
3468
3469 /*
3470 * The discard bitset could have been resized, or the
3471 * discard block size changed. To be safe we start by
3472 * setting every dblock to not discarded.
3473 */
3474 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3475
3476 discard_load_info_init(cache, &li);
3477 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3478 if (r) {
3479 DMERR("%s: could not load origin discards", cache_device_name(cache));
3480 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3481 return r;
3482 }
3483 set_discard_range(&li);
3484
3485 cache->loaded_discards = true;
3486 }
3487
3488 return r;
3489}
3490
3491static void cache_resume(struct dm_target *ti)
3492{
3493 struct cache *cache = ti->private;
3494
3495 cache->need_tick_bio = true;
3496 do_waker(&cache->waker.work);
3497}
3498
3499/*
3500 * Status format:
3501 *
3502 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3503 * <cache block size> <#used cache blocks>/<#total cache blocks>
3504 * <#read hits> <#read misses> <#write hits> <#write misses>
3505 * <#demotions> <#promotions> <#dirty>
3506 * <#features> <features>*
3507 * <#core args> <core args>
3508 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3509 */
3510static void cache_status(struct dm_target *ti, status_type_t type,
3511 unsigned status_flags, char *result, unsigned maxlen)
3512{
3513 int r = 0;
3514 unsigned i;
3515 ssize_t sz = 0;
3516 dm_block_t nr_free_blocks_metadata = 0;
3517 dm_block_t nr_blocks_metadata = 0;
3518 char buf[BDEVNAME_SIZE];
3519 struct cache *cache = ti->private;
3520 dm_cblock_t residency;
3521 bool needs_check;
3522
3523 switch (type) {
3524 case STATUSTYPE_INFO:
3525 if (get_cache_mode(cache) == CM_FAIL) {
3526 DMEMIT("Fail");
3527 break;
3528 }
3529
3530 /* Commit to ensure statistics aren't out-of-date */
3531 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3532 (void) commit(cache, false);
3533
3534 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3535 if (r) {
3536 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3537 cache_device_name(cache), r);
3538 goto err;
3539 }
3540
3541 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3542 if (r) {
3543 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3544 cache_device_name(cache), r);
3545 goto err;
3546 }
3547
3548 residency = policy_residency(cache->policy);
3549
3550 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3551 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3552 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3553 (unsigned long long)nr_blocks_metadata,
3554 (unsigned long long)cache->sectors_per_block,
3555 (unsigned long long) from_cblock(residency),
3556 (unsigned long long) from_cblock(cache->cache_size),
3557 (unsigned) atomic_read(&cache->stats.read_hit),
3558 (unsigned) atomic_read(&cache->stats.read_miss),
3559 (unsigned) atomic_read(&cache->stats.write_hit),
3560 (unsigned) atomic_read(&cache->stats.write_miss),
3561 (unsigned) atomic_read(&cache->stats.demotion),
3562 (unsigned) atomic_read(&cache->stats.promotion),
3563 (unsigned long) atomic_read(&cache->nr_dirty));
3564
3565 if (writethrough_mode(&cache->features))
3566 DMEMIT("1 writethrough ");
3567
3568 else if (passthrough_mode(&cache->features))
3569 DMEMIT("1 passthrough ");
3570
3571 else if (writeback_mode(&cache->features))
3572 DMEMIT("1 writeback ");
3573
3574 else {
3575 DMERR("%s: internal error: unknown io mode: %d",
3576 cache_device_name(cache), (int) cache->features.io_mode);
3577 goto err;
3578 }
3579
3580 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3581
3582 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3583 if (sz < maxlen) {
3584 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3585 if (r)
3586 DMERR("%s: policy_emit_config_values returned %d",
3587 cache_device_name(cache), r);
3588 }
3589
3590 if (get_cache_mode(cache) == CM_READ_ONLY)
3591 DMEMIT("ro ");
3592 else
3593 DMEMIT("rw ");
3594
3595 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3596
3597 if (r || needs_check)
3598 DMEMIT("needs_check ");
3599 else
3600 DMEMIT("- ");
3601
3602 break;
3603
3604 case STATUSTYPE_TABLE:
3605 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3606 DMEMIT("%s ", buf);
3607 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3608 DMEMIT("%s ", buf);
3609 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3610 DMEMIT("%s", buf);
3611
3612 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3613 DMEMIT(" %s", cache->ctr_args[i]);
3614 if (cache->nr_ctr_args)
3615 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3616 }
3617
3618 return;
3619
3620err:
3621 DMEMIT("Error");
3622}
3623
3624/*
3625 * A cache block range can take two forms:
3626 *
3627 * i) A single cblock, eg. '3456'
3628 * ii) A begin and end cblock with dots between, eg. 123-234
3629 */
3630static int parse_cblock_range(struct cache *cache, const char *str,
3631 struct cblock_range *result)
3632{
3633 char dummy;
3634 uint64_t b, e;
3635 int r;
3636
3637 /*
3638 * Try and parse form (ii) first.
3639 */
3640 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3641 if (r < 0)
3642 return r;
3643
3644 if (r == 2) {
3645 result->begin = to_cblock(b);
3646 result->end = to_cblock(e);
3647 return 0;
3648 }
3649
3650 /*
3651 * That didn't work, try form (i).
3652 */
3653 r = sscanf(str, "%llu%c", &b, &dummy);
3654 if (r < 0)
3655 return r;
3656
3657 if (r == 1) {
3658 result->begin = to_cblock(b);
3659 result->end = to_cblock(from_cblock(result->begin) + 1u);
3660 return 0;
3661 }
3662
3663 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3664 return -EINVAL;
3665}
3666
3667static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3668{
3669 uint64_t b = from_cblock(range->begin);
3670 uint64_t e = from_cblock(range->end);
3671 uint64_t n = from_cblock(cache->cache_size);
3672
3673 if (b >= n) {
3674 DMERR("%s: begin cblock out of range: %llu >= %llu",
3675 cache_device_name(cache), b, n);
3676 return -EINVAL;
3677 }
3678
3679 if (e > n) {
3680 DMERR("%s: end cblock out of range: %llu > %llu",
3681 cache_device_name(cache), e, n);
3682 return -EINVAL;
3683 }
3684
3685 if (b >= e) {
3686 DMERR("%s: invalid cblock range: %llu >= %llu",
3687 cache_device_name(cache), b, e);
3688 return -EINVAL;
3689 }
3690
3691 return 0;
3692}
3693
3694static int request_invalidation(struct cache *cache, struct cblock_range *range)
3695{
3696 struct invalidation_request req;
3697
3698 INIT_LIST_HEAD(&req.list);
3699 req.cblocks = range;
3700 atomic_set(&req.complete, 0);
3701 req.err = 0;
3702 init_waitqueue_head(&req.result_wait);
3703
3704 spin_lock(&cache->invalidation_lock);
3705 list_add(&req.list, &cache->invalidation_requests);
3706 spin_unlock(&cache->invalidation_lock);
3707 wake_worker(cache);
3708
3709 wait_event(req.result_wait, atomic_read(&req.complete));
3710 return req.err;
3711}
3712
3713static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3714 const char **cblock_ranges)
3715{
3716 int r = 0;
3717 unsigned i;
3718 struct cblock_range range;
3719
3720 if (!passthrough_mode(&cache->features)) {
3721 DMERR("%s: cache has to be in passthrough mode for invalidation",
3722 cache_device_name(cache));
3723 return -EPERM;
3724 }
3725
3726 for (i = 0; i < count; i++) {
3727 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3728 if (r)
3729 break;
3730
3731 r = validate_cblock_range(cache, &range);
3732 if (r)
3733 break;
3734
3735 /*
3736 * Pass begin and end origin blocks to the worker and wake it.
3737 */
3738 r = request_invalidation(cache, &range);
3739 if (r)
3740 break;
3741 }
3742
3743 return r;
3744}
3745
3746/*
3747 * Supports
3748 * "<key> <value>"
3749 * and
3750 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3751 *
3752 * The key migration_threshold is supported by the cache target core.
3753 */
3754static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3755{
3756 struct cache *cache = ti->private;
3757
3758 if (!argc)
3759 return -EINVAL;
3760
3761 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3762 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3763 cache_device_name(cache));
3764 return -EOPNOTSUPP;
3765 }
3766
3767 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3768 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3769
3770 if (argc != 2)
3771 return -EINVAL;
3772
3773 return set_config_value(cache, argv[0], argv[1]);
3774}
3775
3776static int cache_iterate_devices(struct dm_target *ti,
3777 iterate_devices_callout_fn fn, void *data)
3778{
3779 int r = 0;
3780 struct cache *cache = ti->private;
3781
3782 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3783 if (!r)
3784 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3785
3786 return r;
3787}
3788
3789static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3790{
3791 /*
3792 * FIXME: these limits may be incompatible with the cache device
3793 */
3794 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3795 cache->origin_sectors);
3796 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3797}
3798
3799static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3800{
3801 struct cache *cache = ti->private;
3802 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3803
3804 /*
3805 * If the system-determined stacked limits are compatible with the
3806 * cache's blocksize (io_opt is a factor) do not override them.
3807 */
3808 if (io_opt_sectors < cache->sectors_per_block ||
3809 do_div(io_opt_sectors, cache->sectors_per_block)) {
3810 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3811 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3812 }
3813 set_discard_limits(cache, limits);
3814}
3815
3816/*----------------------------------------------------------------*/
3817
3818static struct target_type cache_target = {
3819 .name = "cache",
3820 .version = {1, 9, 0},
3821 .module = THIS_MODULE,
3822 .ctr = cache_ctr,
3823 .dtr = cache_dtr,
3824 .map = cache_map,
3825 .end_io = cache_end_io,
3826 .postsuspend = cache_postsuspend,
3827 .preresume = cache_preresume,
3828 .resume = cache_resume,
3829 .status = cache_status,
3830 .message = cache_message,
3831 .iterate_devices = cache_iterate_devices,
3832 .io_hints = cache_io_hints,
3833};
3834
3835static int __init dm_cache_init(void)
3836{
3837 int r;
3838
3839 r = dm_register_target(&cache_target);
3840 if (r) {
3841 DMERR("cache target registration failed: %d", r);
3842 return r;
3843 }
3844
3845 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3846 if (!migration_cache) {
3847 dm_unregister_target(&cache_target);
3848 return -ENOMEM;
3849 }
3850
3851 return 0;
3852}
3853
3854static void __exit dm_cache_exit(void)
3855{
3856 dm_unregister_target(&cache_target);
3857 kmem_cache_destroy(migration_cache);
3858}
3859
3860module_init(dm_cache_init);
3861module_exit(dm_cache_exit);
3862
3863MODULE_DESCRIPTION(DM_NAME " cache target");
3864MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3865MODULE_LICENSE("GPL");