2 * Copyright (C) 2012 Red Hat. All rights reserved.
4 * This file is released under the GPL.
8 #include "dm-bio-prison-v2.h"
9 #include "dm-bio-record.h"
10 #include "dm-cache-metadata.h"
11 #include "dm-io-tracker.h"
13 #include <linux/dm-io.h>
14 #include <linux/dm-kcopyd.h>
15 #include <linux/jiffies.h>
16 #include <linux/init.h>
17 #include <linux/mempool.h>
18 #include <linux/module.h>
19 #include <linux/rwsem.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
23 #define DM_MSG_PREFIX "cache"
25 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
26 "A percentage of time allocated for copying to and/or from cache");
28 /*----------------------------------------------------------------*/
33 * oblock: index of an origin block
34 * cblock: index of a cache block
35 * promotion: movement of a block from origin to cache
36 * demotion: movement of a block from cache to origin
37 * migration: movement of a block between the origin and cache device,
41 /*----------------------------------------------------------------*/
44 * Represents a chunk of future work. 'input' allows continuations to pass
45 * values between themselves, typically error values.
48 struct work_struct ws;
52 static inline void init_continuation(struct continuation *k,
53 void (*fn)(struct work_struct *))
55 INIT_WORK(&k->ws, fn);
59 static inline void queue_continuation(struct workqueue_struct *wq,
60 struct continuation *k)
62 queue_work(wq, &k->ws);
65 /*----------------------------------------------------------------*/
68 * The batcher collects together pieces of work that need a particular
69 * operation to occur before they can proceed (typically a commit).
73 * The operation that everyone is waiting for.
75 blk_status_t (*commit_op)(void *context);
79 * This is how bios should be issued once the commit op is complete
80 * (accounted_request).
82 void (*issue_op)(struct bio *bio, void *context);
86 * Queued work gets put on here after commit.
88 struct workqueue_struct *wq;
91 struct list_head work_items;
93 struct work_struct commit_work;
95 bool commit_scheduled;
98 static void __commit(struct work_struct *_ws)
100 struct batcher *b = container_of(_ws, struct batcher, commit_work);
102 struct list_head work_items;
103 struct work_struct *ws, *tmp;
104 struct continuation *k;
106 struct bio_list bios;
108 INIT_LIST_HEAD(&work_items);
109 bio_list_init(&bios);
112 * We have to grab these before the commit_op to avoid a race
115 spin_lock_irq(&b->lock);
116 list_splice_init(&b->work_items, &work_items);
117 bio_list_merge(&bios, &b->bios);
118 bio_list_init(&b->bios);
119 b->commit_scheduled = false;
120 spin_unlock_irq(&b->lock);
122 r = b->commit_op(b->commit_context);
124 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
125 k = container_of(ws, struct continuation, ws);
127 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
128 queue_work(b->wq, ws);
131 while ((bio = bio_list_pop(&bios))) {
136 b->issue_op(bio, b->issue_context);
140 static void batcher_init(struct batcher *b,
141 blk_status_t (*commit_op)(void *),
142 void *commit_context,
143 void (*issue_op)(struct bio *bio, void *),
145 struct workqueue_struct *wq)
147 b->commit_op = commit_op;
148 b->commit_context = commit_context;
149 b->issue_op = issue_op;
150 b->issue_context = issue_context;
153 spin_lock_init(&b->lock);
154 INIT_LIST_HEAD(&b->work_items);
155 bio_list_init(&b->bios);
156 INIT_WORK(&b->commit_work, __commit);
157 b->commit_scheduled = false;
160 static void async_commit(struct batcher *b)
162 queue_work(b->wq, &b->commit_work);
165 static void continue_after_commit(struct batcher *b, struct continuation *k)
167 bool commit_scheduled;
169 spin_lock_irq(&b->lock);
170 commit_scheduled = b->commit_scheduled;
171 list_add_tail(&k->ws.entry, &b->work_items);
172 spin_unlock_irq(&b->lock);
174 if (commit_scheduled)
179 * Bios are errored if commit failed.
181 static void issue_after_commit(struct batcher *b, struct bio *bio)
183 bool commit_scheduled;
185 spin_lock_irq(&b->lock);
186 commit_scheduled = b->commit_scheduled;
187 bio_list_add(&b->bios, bio);
188 spin_unlock_irq(&b->lock);
190 if (commit_scheduled)
195 * Call this if some urgent work is waiting for the commit to complete.
197 static void schedule_commit(struct batcher *b)
201 spin_lock_irq(&b->lock);
202 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
203 b->commit_scheduled = true;
204 spin_unlock_irq(&b->lock);
211 * There are a couple of places where we let a bio run, but want to do some
212 * work before calling its endio function. We do this by temporarily
213 * changing the endio fn.
215 struct dm_hook_info {
216 bio_end_io_t *bi_end_io;
219 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
220 bio_end_io_t *bi_end_io, void *bi_private)
222 h->bi_end_io = bio->bi_end_io;
224 bio->bi_end_io = bi_end_io;
225 bio->bi_private = bi_private;
228 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
230 bio->bi_end_io = h->bi_end_io;
233 /*----------------------------------------------------------------*/
235 #define MIGRATION_POOL_SIZE 128
236 #define COMMIT_PERIOD HZ
237 #define MIGRATION_COUNT_WINDOW 10
240 * The block size of the device holding cache data must be
241 * between 32KB and 1GB.
243 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
244 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
246 enum cache_metadata_mode {
247 CM_WRITE, /* metadata may be changed */
248 CM_READ_ONLY, /* metadata may not be changed */
254 * Data is written to cached blocks only. These blocks are marked
255 * dirty. If you lose the cache device you will lose data.
256 * Potential performance increase for both reads and writes.
261 * Data is written to both cache and origin. Blocks are never
262 * dirty. Potential performance benfit for reads only.
267 * A degraded mode useful for various cache coherency situations
268 * (eg, rolling back snapshots). Reads and writes always go to the
269 * origin. If a write goes to a cached oblock, then the cache
270 * block is invalidated.
275 struct cache_features {
276 enum cache_metadata_mode mode;
277 enum cache_io_mode io_mode;
278 unsigned metadata_version;
279 bool discard_passdown:1;
290 atomic_t copies_avoided;
291 atomic_t cache_cell_clash;
292 atomic_t commit_count;
293 atomic_t discard_count;
297 struct dm_target *ti;
301 * Fields for converting from sectors to blocks.
303 int sectors_per_block_shift;
304 sector_t sectors_per_block;
306 struct dm_cache_metadata *cmd;
309 * Metadata is written to this device.
311 struct dm_dev *metadata_dev;
314 * The slower of the two data devices. Typically a spindle.
316 struct dm_dev *origin_dev;
319 * The faster of the two data devices. Typically an SSD.
321 struct dm_dev *cache_dev;
324 * Size of the origin device in _complete_ blocks and native sectors.
326 dm_oblock_t origin_blocks;
327 sector_t origin_sectors;
330 * Size of the cache device in blocks.
332 dm_cblock_t cache_size;
335 * Invalidation fields.
337 spinlock_t invalidation_lock;
338 struct list_head invalidation_requests;
340 sector_t migration_threshold;
341 wait_queue_head_t migration_wait;
342 atomic_t nr_allocated_migrations;
345 * The number of in flight migrations that are performing
346 * background io. eg, promotion, writeback.
348 atomic_t nr_io_migrations;
350 struct bio_list deferred_bios;
352 struct rw_semaphore quiesce_lock;
355 * origin_blocks entries, discarded if set.
357 dm_dblock_t discard_nr_blocks;
358 unsigned long *discard_bitset;
359 uint32_t discard_block_size; /* a power of 2 times sectors per block */
362 * Rather than reconstructing the table line for the status we just
363 * save it and regurgitate.
365 unsigned nr_ctr_args;
366 const char **ctr_args;
368 struct dm_kcopyd_client *copier;
369 struct work_struct deferred_bio_worker;
370 struct work_struct migration_worker;
371 struct workqueue_struct *wq;
372 struct delayed_work waker;
373 struct dm_bio_prison_v2 *prison;
376 * cache_size entries, dirty if set
378 unsigned long *dirty_bitset;
381 unsigned policy_nr_args;
382 struct dm_cache_policy *policy;
385 * Cache features such as write-through.
387 struct cache_features features;
389 struct cache_stats stats;
391 bool need_tick_bio:1;
394 bool commit_requested:1;
395 bool loaded_mappings:1;
396 bool loaded_discards:1;
398 struct rw_semaphore background_work_lock;
400 struct batcher committer;
401 struct work_struct commit_ws;
403 struct dm_io_tracker tracker;
405 mempool_t migration_pool;
410 struct per_bio_data {
413 struct dm_bio_prison_cell_v2 *cell;
414 struct dm_hook_info hook_info;
418 struct dm_cache_migration {
419 struct continuation k;
422 struct policy_work *op;
423 struct bio *overwrite_bio;
424 struct dm_bio_prison_cell_v2 *cell;
426 dm_cblock_t invalidate_cblock;
427 dm_oblock_t invalidate_oblock;
430 /*----------------------------------------------------------------*/
432 static bool writethrough_mode(struct cache *cache)
434 return cache->features.io_mode == CM_IO_WRITETHROUGH;
437 static bool writeback_mode(struct cache *cache)
439 return cache->features.io_mode == CM_IO_WRITEBACK;
442 static inline bool passthrough_mode(struct cache *cache)
444 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
447 /*----------------------------------------------------------------*/
449 static void wake_deferred_bio_worker(struct cache *cache)
451 queue_work(cache->wq, &cache->deferred_bio_worker);
454 static void wake_migration_worker(struct cache *cache)
456 if (passthrough_mode(cache))
459 queue_work(cache->wq, &cache->migration_worker);
462 /*----------------------------------------------------------------*/
464 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
466 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
469 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
471 dm_bio_prison_free_cell_v2(cache->prison, cell);
474 static struct dm_cache_migration *alloc_migration(struct cache *cache)
476 struct dm_cache_migration *mg;
478 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
480 memset(mg, 0, sizeof(*mg));
483 atomic_inc(&cache->nr_allocated_migrations);
488 static void free_migration(struct dm_cache_migration *mg)
490 struct cache *cache = mg->cache;
492 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
493 wake_up(&cache->migration_wait);
495 mempool_free(mg, &cache->migration_pool);
498 /*----------------------------------------------------------------*/
500 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
502 return to_oblock(from_oblock(b) + 1ull);
505 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
509 key->block_begin = from_oblock(begin);
510 key->block_end = from_oblock(end);
514 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
515 * level 1 which prevents *both* READs and WRITEs.
517 #define WRITE_LOCK_LEVEL 0
518 #define READ_WRITE_LOCK_LEVEL 1
520 static unsigned lock_level(struct bio *bio)
522 return bio_data_dir(bio) == WRITE ?
524 READ_WRITE_LOCK_LEVEL;
527 /*----------------------------------------------------------------
529 *--------------------------------------------------------------*/
531 static struct per_bio_data *get_per_bio_data(struct bio *bio)
533 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
538 static struct per_bio_data *init_per_bio_data(struct bio *bio)
540 struct per_bio_data *pb = get_per_bio_data(bio);
543 pb->req_nr = dm_bio_get_target_bio_nr(bio);
550 /*----------------------------------------------------------------*/
552 static void defer_bio(struct cache *cache, struct bio *bio)
554 spin_lock_irq(&cache->lock);
555 bio_list_add(&cache->deferred_bios, bio);
556 spin_unlock_irq(&cache->lock);
558 wake_deferred_bio_worker(cache);
561 static void defer_bios(struct cache *cache, struct bio_list *bios)
563 spin_lock_irq(&cache->lock);
564 bio_list_merge(&cache->deferred_bios, bios);
566 spin_unlock_irq(&cache->lock);
568 wake_deferred_bio_worker(cache);
571 /*----------------------------------------------------------------*/
573 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
576 struct per_bio_data *pb;
577 struct dm_cell_key_v2 key;
578 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
579 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
581 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
583 build_key(oblock, end, &key);
584 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
587 * Failed to get the lock.
589 free_prison_cell(cache, cell_prealloc);
593 if (cell != cell_prealloc)
594 free_prison_cell(cache, cell_prealloc);
596 pb = get_per_bio_data(bio);
602 /*----------------------------------------------------------------*/
604 static bool is_dirty(struct cache *cache, dm_cblock_t b)
606 return test_bit(from_cblock(b), cache->dirty_bitset);
609 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
611 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
612 atomic_inc(&cache->nr_dirty);
613 policy_set_dirty(cache->policy, cblock);
618 * These two are called when setting after migrations to force the policy
619 * and dirty bitset to be in sync.
621 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
623 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
624 atomic_inc(&cache->nr_dirty);
625 policy_set_dirty(cache->policy, cblock);
628 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
630 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
631 if (atomic_dec_return(&cache->nr_dirty) == 0)
632 dm_table_event(cache->ti->table);
635 policy_clear_dirty(cache->policy, cblock);
638 /*----------------------------------------------------------------*/
640 static bool block_size_is_power_of_two(struct cache *cache)
642 return cache->sectors_per_block_shift >= 0;
645 static dm_block_t block_div(dm_block_t b, uint32_t n)
652 static dm_block_t oblocks_per_dblock(struct cache *cache)
654 dm_block_t oblocks = cache->discard_block_size;
656 if (block_size_is_power_of_two(cache))
657 oblocks >>= cache->sectors_per_block_shift;
659 oblocks = block_div(oblocks, cache->sectors_per_block);
664 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
666 return to_dblock(block_div(from_oblock(oblock),
667 oblocks_per_dblock(cache)));
670 static void set_discard(struct cache *cache, dm_dblock_t b)
672 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
673 atomic_inc(&cache->stats.discard_count);
675 spin_lock_irq(&cache->lock);
676 set_bit(from_dblock(b), cache->discard_bitset);
677 spin_unlock_irq(&cache->lock);
680 static void clear_discard(struct cache *cache, dm_dblock_t b)
682 spin_lock_irq(&cache->lock);
683 clear_bit(from_dblock(b), cache->discard_bitset);
684 spin_unlock_irq(&cache->lock);
687 static bool is_discarded(struct cache *cache, dm_dblock_t b)
690 spin_lock_irq(&cache->lock);
691 r = test_bit(from_dblock(b), cache->discard_bitset);
692 spin_unlock_irq(&cache->lock);
697 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
700 spin_lock_irq(&cache->lock);
701 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
702 cache->discard_bitset);
703 spin_unlock_irq(&cache->lock);
708 /*----------------------------------------------------------------
710 *--------------------------------------------------------------*/
711 static void remap_to_origin(struct cache *cache, struct bio *bio)
713 bio_set_dev(bio, cache->origin_dev->bdev);
716 static void remap_to_cache(struct cache *cache, struct bio *bio,
719 sector_t bi_sector = bio->bi_iter.bi_sector;
720 sector_t block = from_cblock(cblock);
722 bio_set_dev(bio, cache->cache_dev->bdev);
723 if (!block_size_is_power_of_two(cache))
724 bio->bi_iter.bi_sector =
725 (block * cache->sectors_per_block) +
726 sector_div(bi_sector, cache->sectors_per_block);
728 bio->bi_iter.bi_sector =
729 (block << cache->sectors_per_block_shift) |
730 (bi_sector & (cache->sectors_per_block - 1));
733 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
735 struct per_bio_data *pb;
737 spin_lock_irq(&cache->lock);
738 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
739 bio_op(bio) != REQ_OP_DISCARD) {
740 pb = get_per_bio_data(bio);
742 cache->need_tick_bio = false;
744 spin_unlock_irq(&cache->lock);
747 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
750 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
751 check_if_tick_bio_needed(cache, bio);
752 remap_to_origin(cache, bio);
753 if (bio_data_dir(bio) == WRITE)
754 clear_discard(cache, oblock_to_dblock(cache, oblock));
757 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
758 dm_oblock_t oblock, dm_cblock_t cblock)
760 check_if_tick_bio_needed(cache, bio);
761 remap_to_cache(cache, bio, cblock);
762 if (bio_data_dir(bio) == WRITE) {
763 set_dirty(cache, cblock);
764 clear_discard(cache, oblock_to_dblock(cache, oblock));
768 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
770 sector_t block_nr = bio->bi_iter.bi_sector;
772 if (!block_size_is_power_of_two(cache))
773 (void) sector_div(block_nr, cache->sectors_per_block);
775 block_nr >>= cache->sectors_per_block_shift;
777 return to_oblock(block_nr);
780 static bool accountable_bio(struct cache *cache, struct bio *bio)
782 return bio_op(bio) != REQ_OP_DISCARD;
785 static void accounted_begin(struct cache *cache, struct bio *bio)
787 struct per_bio_data *pb;
789 if (accountable_bio(cache, bio)) {
790 pb = get_per_bio_data(bio);
791 pb->len = bio_sectors(bio);
792 dm_iot_io_begin(&cache->tracker, pb->len);
796 static void accounted_complete(struct cache *cache, struct bio *bio)
798 struct per_bio_data *pb = get_per_bio_data(bio);
800 dm_iot_io_end(&cache->tracker, pb->len);
803 static void accounted_request(struct cache *cache, struct bio *bio)
805 accounted_begin(cache, bio);
806 dm_submit_bio_remap(bio, NULL);
809 static void issue_op(struct bio *bio, void *context)
811 struct cache *cache = context;
812 accounted_request(cache, bio);
816 * When running in writethrough mode we need to send writes to clean blocks
817 * to both the cache and origin devices. Clone the bio and send them in parallel.
819 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
820 dm_oblock_t oblock, dm_cblock_t cblock)
822 struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
823 GFP_NOIO, &cache->bs);
827 bio_chain(origin_bio, bio);
829 if (bio_data_dir(origin_bio) == WRITE)
830 clear_discard(cache, oblock_to_dblock(cache, oblock));
831 submit_bio(origin_bio);
833 remap_to_cache(cache, bio, cblock);
836 /*----------------------------------------------------------------
838 *--------------------------------------------------------------*/
839 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
841 return cache->features.mode;
844 static const char *cache_device_name(struct cache *cache)
846 return dm_table_device_name(cache->ti->table);
849 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
851 const char *descs[] = {
857 dm_table_event(cache->ti->table);
858 DMINFO("%s: switching cache to %s mode",
859 cache_device_name(cache), descs[(int)mode]);
862 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
865 enum cache_metadata_mode old_mode = get_cache_mode(cache);
867 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
868 DMERR("%s: unable to read needs_check flag, setting failure mode.",
869 cache_device_name(cache));
873 if (new_mode == CM_WRITE && needs_check) {
874 DMERR("%s: unable to switch cache to write mode until repaired.",
875 cache_device_name(cache));
876 if (old_mode != new_mode)
879 new_mode = CM_READ_ONLY;
882 /* Never move out of fail mode */
883 if (old_mode == CM_FAIL)
889 dm_cache_metadata_set_read_only(cache->cmd);
893 dm_cache_metadata_set_read_write(cache->cmd);
897 cache->features.mode = new_mode;
899 if (new_mode != old_mode)
900 notify_mode_switch(cache, new_mode);
903 static void abort_transaction(struct cache *cache)
905 const char *dev_name = cache_device_name(cache);
907 if (get_cache_mode(cache) >= CM_READ_ONLY)
910 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
911 if (dm_cache_metadata_abort(cache->cmd)) {
912 DMERR("%s: failed to abort metadata transaction", dev_name);
913 set_cache_mode(cache, CM_FAIL);
916 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
917 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
918 set_cache_mode(cache, CM_FAIL);
922 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
924 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
925 cache_device_name(cache), op, r);
926 abort_transaction(cache);
927 set_cache_mode(cache, CM_READ_ONLY);
930 /*----------------------------------------------------------------*/
932 static void load_stats(struct cache *cache)
934 struct dm_cache_statistics stats;
936 dm_cache_metadata_get_stats(cache->cmd, &stats);
937 atomic_set(&cache->stats.read_hit, stats.read_hits);
938 atomic_set(&cache->stats.read_miss, stats.read_misses);
939 atomic_set(&cache->stats.write_hit, stats.write_hits);
940 atomic_set(&cache->stats.write_miss, stats.write_misses);
943 static void save_stats(struct cache *cache)
945 struct dm_cache_statistics stats;
947 if (get_cache_mode(cache) >= CM_READ_ONLY)
950 stats.read_hits = atomic_read(&cache->stats.read_hit);
951 stats.read_misses = atomic_read(&cache->stats.read_miss);
952 stats.write_hits = atomic_read(&cache->stats.write_hit);
953 stats.write_misses = atomic_read(&cache->stats.write_miss);
955 dm_cache_metadata_set_stats(cache->cmd, &stats);
958 static void update_stats(struct cache_stats *stats, enum policy_operation op)
962 atomic_inc(&stats->promotion);
966 atomic_inc(&stats->demotion);
969 case POLICY_WRITEBACK:
970 atomic_inc(&stats->writeback);
975 /*----------------------------------------------------------------
976 * Migration processing
978 * Migration covers moving data from the origin device to the cache, or
980 *--------------------------------------------------------------*/
982 static void inc_io_migrations(struct cache *cache)
984 atomic_inc(&cache->nr_io_migrations);
987 static void dec_io_migrations(struct cache *cache)
989 atomic_dec(&cache->nr_io_migrations);
992 static bool discard_or_flush(struct bio *bio)
994 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
997 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
998 dm_dblock_t *b, dm_dblock_t *e)
1000 sector_t sb = bio->bi_iter.bi_sector;
1001 sector_t se = bio_end_sector(bio);
1003 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1005 if (se - sb < cache->discard_block_size)
1008 *e = to_dblock(block_div(se, cache->discard_block_size));
1011 /*----------------------------------------------------------------*/
1013 static void prevent_background_work(struct cache *cache)
1016 down_write(&cache->background_work_lock);
1020 static void allow_background_work(struct cache *cache)
1023 up_write(&cache->background_work_lock);
1027 static bool background_work_begin(struct cache *cache)
1032 r = down_read_trylock(&cache->background_work_lock);
1038 static void background_work_end(struct cache *cache)
1041 up_read(&cache->background_work_lock);
1045 /*----------------------------------------------------------------*/
1047 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1049 return (bio_data_dir(bio) == WRITE) &&
1050 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1053 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1055 return writeback_mode(cache) &&
1056 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1059 static void quiesce(struct dm_cache_migration *mg,
1060 void (*continuation)(struct work_struct *))
1062 init_continuation(&mg->k, continuation);
1063 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1066 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1068 struct continuation *k = container_of(ws, struct continuation, ws);
1069 return container_of(k, struct dm_cache_migration, k);
1072 static void copy_complete(int read_err, unsigned long write_err, void *context)
1074 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1076 if (read_err || write_err)
1077 mg->k.input = BLK_STS_IOERR;
1079 queue_continuation(mg->cache->wq, &mg->k);
1082 static void copy(struct dm_cache_migration *mg, bool promote)
1084 struct dm_io_region o_region, c_region;
1085 struct cache *cache = mg->cache;
1087 o_region.bdev = cache->origin_dev->bdev;
1088 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1089 o_region.count = cache->sectors_per_block;
1091 c_region.bdev = cache->cache_dev->bdev;
1092 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1093 c_region.count = cache->sectors_per_block;
1096 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1098 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1101 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1103 struct per_bio_data *pb = get_per_bio_data(bio);
1105 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1106 free_prison_cell(cache, pb->cell);
1110 static void overwrite_endio(struct bio *bio)
1112 struct dm_cache_migration *mg = bio->bi_private;
1113 struct cache *cache = mg->cache;
1114 struct per_bio_data *pb = get_per_bio_data(bio);
1116 dm_unhook_bio(&pb->hook_info, bio);
1119 mg->k.input = bio->bi_status;
1121 queue_continuation(cache->wq, &mg->k);
1124 static void overwrite(struct dm_cache_migration *mg,
1125 void (*continuation)(struct work_struct *))
1127 struct bio *bio = mg->overwrite_bio;
1128 struct per_bio_data *pb = get_per_bio_data(bio);
1130 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1133 * The overwrite bio is part of the copy operation, as such it does
1134 * not set/clear discard or dirty flags.
1136 if (mg->op->op == POLICY_PROMOTE)
1137 remap_to_cache(mg->cache, bio, mg->op->cblock);
1139 remap_to_origin(mg->cache, bio);
1141 init_continuation(&mg->k, continuation);
1142 accounted_request(mg->cache, bio);
1148 * 1) exclusive lock preventing WRITEs
1150 * 3) copy or issue overwrite bio
1151 * 4) upgrade to exclusive lock preventing READs and WRITEs
1153 * 6) update metadata and commit
1156 static void mg_complete(struct dm_cache_migration *mg, bool success)
1158 struct bio_list bios;
1159 struct cache *cache = mg->cache;
1160 struct policy_work *op = mg->op;
1161 dm_cblock_t cblock = op->cblock;
1164 update_stats(&cache->stats, op->op);
1167 case POLICY_PROMOTE:
1168 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1169 policy_complete_background_work(cache->policy, op, success);
1171 if (mg->overwrite_bio) {
1173 force_set_dirty(cache, cblock);
1174 else if (mg->k.input)
1175 mg->overwrite_bio->bi_status = mg->k.input;
1177 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1178 bio_endio(mg->overwrite_bio);
1181 force_clear_dirty(cache, cblock);
1182 dec_io_migrations(cache);
1188 * We clear dirty here to update the nr_dirty counter.
1191 force_clear_dirty(cache, cblock);
1192 policy_complete_background_work(cache->policy, op, success);
1193 dec_io_migrations(cache);
1196 case POLICY_WRITEBACK:
1198 force_clear_dirty(cache, cblock);
1199 policy_complete_background_work(cache->policy, op, success);
1200 dec_io_migrations(cache);
1204 bio_list_init(&bios);
1206 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1207 free_prison_cell(cache, mg->cell);
1211 defer_bios(cache, &bios);
1212 wake_migration_worker(cache);
1214 background_work_end(cache);
1217 static void mg_success(struct work_struct *ws)
1219 struct dm_cache_migration *mg = ws_to_mg(ws);
1220 mg_complete(mg, mg->k.input == 0);
1223 static void mg_update_metadata(struct work_struct *ws)
1226 struct dm_cache_migration *mg = ws_to_mg(ws);
1227 struct cache *cache = mg->cache;
1228 struct policy_work *op = mg->op;
1231 case POLICY_PROMOTE:
1232 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1234 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1235 cache_device_name(cache));
1236 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1238 mg_complete(mg, false);
1241 mg_complete(mg, true);
1245 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1247 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1248 cache_device_name(cache));
1249 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1251 mg_complete(mg, false);
1256 * It would be nice if we only had to commit when a REQ_FLUSH
1257 * comes through. But there's one scenario that we have to
1260 * - vblock x in a cache block
1262 * - cache block gets reallocated and over written
1265 * When we recover, because there was no commit the cache will
1266 * rollback to having the data for vblock x in the cache block.
1267 * But the cache block has since been overwritten, so it'll end
1268 * up pointing to data that was never in 'x' during the history
1271 * To avoid this issue we require a commit as part of the
1272 * demotion operation.
1274 init_continuation(&mg->k, mg_success);
1275 continue_after_commit(&cache->committer, &mg->k);
1276 schedule_commit(&cache->committer);
1279 case POLICY_WRITEBACK:
1280 mg_complete(mg, true);
1285 static void mg_update_metadata_after_copy(struct work_struct *ws)
1287 struct dm_cache_migration *mg = ws_to_mg(ws);
1290 * Did the copy succeed?
1293 mg_complete(mg, false);
1295 mg_update_metadata(ws);
1298 static void mg_upgrade_lock(struct work_struct *ws)
1301 struct dm_cache_migration *mg = ws_to_mg(ws);
1304 * Did the copy succeed?
1307 mg_complete(mg, false);
1311 * Now we want the lock to prevent both reads and writes.
1313 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1314 READ_WRITE_LOCK_LEVEL);
1316 mg_complete(mg, false);
1319 quiesce(mg, mg_update_metadata);
1322 mg_update_metadata(ws);
1326 static void mg_full_copy(struct work_struct *ws)
1328 struct dm_cache_migration *mg = ws_to_mg(ws);
1329 struct cache *cache = mg->cache;
1330 struct policy_work *op = mg->op;
1331 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1333 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1334 is_discarded_oblock(cache, op->oblock)) {
1335 mg_upgrade_lock(ws);
1339 init_continuation(&mg->k, mg_upgrade_lock);
1340 copy(mg, is_policy_promote);
1343 static void mg_copy(struct work_struct *ws)
1345 struct dm_cache_migration *mg = ws_to_mg(ws);
1347 if (mg->overwrite_bio) {
1349 * No exclusive lock was held when we last checked if the bio
1350 * was optimisable. So we have to check again in case things
1351 * have changed (eg, the block may no longer be discarded).
1353 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1355 * Fallback to a real full copy after doing some tidying up.
1357 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1358 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1359 mg->overwrite_bio = NULL;
1360 inc_io_migrations(mg->cache);
1366 * It's safe to do this here, even though it's new data
1367 * because all IO has been locked out of the block.
1369 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1370 * so _not_ using mg_upgrade_lock() as continutation.
1372 overwrite(mg, mg_update_metadata_after_copy);
1378 static int mg_lock_writes(struct dm_cache_migration *mg)
1381 struct dm_cell_key_v2 key;
1382 struct cache *cache = mg->cache;
1383 struct dm_bio_prison_cell_v2 *prealloc;
1385 prealloc = alloc_prison_cell(cache);
1388 * Prevent writes to the block, but allow reads to continue.
1389 * Unless we're using an overwrite bio, in which case we lock
1392 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1393 r = dm_cell_lock_v2(cache->prison, &key,
1394 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1395 prealloc, &mg->cell);
1397 free_prison_cell(cache, prealloc);
1398 mg_complete(mg, false);
1402 if (mg->cell != prealloc)
1403 free_prison_cell(cache, prealloc);
1408 quiesce(mg, mg_copy);
1413 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1415 struct dm_cache_migration *mg;
1417 if (!background_work_begin(cache)) {
1418 policy_complete_background_work(cache->policy, op, false);
1422 mg = alloc_migration(cache);
1425 mg->overwrite_bio = bio;
1428 inc_io_migrations(cache);
1430 return mg_lock_writes(mg);
1433 /*----------------------------------------------------------------
1434 * invalidation processing
1435 *--------------------------------------------------------------*/
1437 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1439 struct bio_list bios;
1440 struct cache *cache = mg->cache;
1442 bio_list_init(&bios);
1443 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1444 free_prison_cell(cache, mg->cell);
1446 if (!success && mg->overwrite_bio)
1447 bio_io_error(mg->overwrite_bio);
1450 defer_bios(cache, &bios);
1452 background_work_end(cache);
1455 static void invalidate_completed(struct work_struct *ws)
1457 struct dm_cache_migration *mg = ws_to_mg(ws);
1458 invalidate_complete(mg, !mg->k.input);
1461 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1463 int r = policy_invalidate_mapping(cache->policy, cblock);
1465 r = dm_cache_remove_mapping(cache->cmd, cblock);
1467 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1468 cache_device_name(cache));
1469 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1472 } else if (r == -ENODATA) {
1474 * Harmless, already unmapped.
1479 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1484 static void invalidate_remove(struct work_struct *ws)
1487 struct dm_cache_migration *mg = ws_to_mg(ws);
1488 struct cache *cache = mg->cache;
1490 r = invalidate_cblock(cache, mg->invalidate_cblock);
1492 invalidate_complete(mg, false);
1496 init_continuation(&mg->k, invalidate_completed);
1497 continue_after_commit(&cache->committer, &mg->k);
1498 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1499 mg->overwrite_bio = NULL;
1500 schedule_commit(&cache->committer);
1503 static int invalidate_lock(struct dm_cache_migration *mg)
1506 struct dm_cell_key_v2 key;
1507 struct cache *cache = mg->cache;
1508 struct dm_bio_prison_cell_v2 *prealloc;
1510 prealloc = alloc_prison_cell(cache);
1512 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1513 r = dm_cell_lock_v2(cache->prison, &key,
1514 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1516 free_prison_cell(cache, prealloc);
1517 invalidate_complete(mg, false);
1521 if (mg->cell != prealloc)
1522 free_prison_cell(cache, prealloc);
1525 quiesce(mg, invalidate_remove);
1529 * We can't call invalidate_remove() directly here because we
1530 * might still be in request context.
1532 init_continuation(&mg->k, invalidate_remove);
1533 queue_work(cache->wq, &mg->k.ws);
1539 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1540 dm_oblock_t oblock, struct bio *bio)
1542 struct dm_cache_migration *mg;
1544 if (!background_work_begin(cache))
1547 mg = alloc_migration(cache);
1549 mg->overwrite_bio = bio;
1550 mg->invalidate_cblock = cblock;
1551 mg->invalidate_oblock = oblock;
1553 return invalidate_lock(mg);
1556 /*----------------------------------------------------------------
1558 *--------------------------------------------------------------*/
1565 static enum busy spare_migration_bandwidth(struct cache *cache)
1567 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1568 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1569 cache->sectors_per_block;
1571 if (idle && current_volume <= cache->migration_threshold)
1577 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1579 atomic_inc(bio_data_dir(bio) == READ ?
1580 &cache->stats.read_hit : &cache->stats.write_hit);
1583 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1585 atomic_inc(bio_data_dir(bio) == READ ?
1586 &cache->stats.read_miss : &cache->stats.write_miss);
1589 /*----------------------------------------------------------------*/
1591 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1592 bool *commit_needed)
1595 bool rb, background_queued;
1598 *commit_needed = false;
1600 rb = bio_detain_shared(cache, block, bio);
1603 * An exclusive lock is held for this block, so we have to
1604 * wait. We set the commit_needed flag so the current
1605 * transaction will be committed asap, allowing this lock
1608 *commit_needed = true;
1609 return DM_MAPIO_SUBMITTED;
1612 data_dir = bio_data_dir(bio);
1614 if (optimisable_bio(cache, bio, block)) {
1615 struct policy_work *op = NULL;
1617 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1618 if (unlikely(r && r != -ENOENT)) {
1619 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1620 cache_device_name(cache), r);
1622 return DM_MAPIO_SUBMITTED;
1625 if (r == -ENOENT && op) {
1626 bio_drop_shared_lock(cache, bio);
1627 BUG_ON(op->op != POLICY_PROMOTE);
1628 mg_start(cache, op, bio);
1629 return DM_MAPIO_SUBMITTED;
1632 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1633 if (unlikely(r && r != -ENOENT)) {
1634 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1635 cache_device_name(cache), r);
1637 return DM_MAPIO_SUBMITTED;
1640 if (background_queued)
1641 wake_migration_worker(cache);
1645 struct per_bio_data *pb = get_per_bio_data(bio);
1650 inc_miss_counter(cache, bio);
1651 if (pb->req_nr == 0) {
1652 accounted_begin(cache, bio);
1653 remap_to_origin_clear_discard(cache, bio, block);
1656 * This is a duplicate writethrough io that is no
1657 * longer needed because the block has been demoted.
1660 return DM_MAPIO_SUBMITTED;
1666 inc_hit_counter(cache, bio);
1669 * Passthrough always maps to the origin, invalidating any
1670 * cache blocks that are written to.
1672 if (passthrough_mode(cache)) {
1673 if (bio_data_dir(bio) == WRITE) {
1674 bio_drop_shared_lock(cache, bio);
1675 atomic_inc(&cache->stats.demotion);
1676 invalidate_start(cache, cblock, block, bio);
1678 remap_to_origin_clear_discard(cache, bio, block);
1680 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1681 !is_dirty(cache, cblock)) {
1682 remap_to_origin_and_cache(cache, bio, block, cblock);
1683 accounted_begin(cache, bio);
1685 remap_to_cache_dirty(cache, bio, block, cblock);
1690 * dm core turns FUA requests into a separate payload and FLUSH req.
1692 if (bio->bi_opf & REQ_FUA) {
1694 * issue_after_commit will call accounted_begin a second time. So
1695 * we call accounted_complete() to avoid double accounting.
1697 accounted_complete(cache, bio);
1698 issue_after_commit(&cache->committer, bio);
1699 *commit_needed = true;
1700 return DM_MAPIO_SUBMITTED;
1703 return DM_MAPIO_REMAPPED;
1706 static bool process_bio(struct cache *cache, struct bio *bio)
1710 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1711 dm_submit_bio_remap(bio, NULL);
1713 return commit_needed;
1717 * A non-zero return indicates read_only or fail_io mode.
1719 static int commit(struct cache *cache, bool clean_shutdown)
1723 if (get_cache_mode(cache) >= CM_READ_ONLY)
1726 atomic_inc(&cache->stats.commit_count);
1727 r = dm_cache_commit(cache->cmd, clean_shutdown);
1729 metadata_operation_failed(cache, "dm_cache_commit", r);
1735 * Used by the batcher.
1737 static blk_status_t commit_op(void *context)
1739 struct cache *cache = context;
1741 if (dm_cache_changed_this_transaction(cache->cmd))
1742 return errno_to_blk_status(commit(cache, false));
1747 /*----------------------------------------------------------------*/
1749 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1751 struct per_bio_data *pb = get_per_bio_data(bio);
1754 remap_to_origin(cache, bio);
1756 remap_to_cache(cache, bio, 0);
1758 issue_after_commit(&cache->committer, bio);
1762 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1766 // FIXME: do we need to lock the region? Or can we just assume the
1767 // user wont be so foolish as to issue discard concurrently with
1769 calc_discard_block_range(cache, bio, &b, &e);
1771 set_discard(cache, b);
1772 b = to_dblock(from_dblock(b) + 1);
1775 if (cache->features.discard_passdown) {
1776 remap_to_origin(cache, bio);
1777 dm_submit_bio_remap(bio, NULL);
1784 static void process_deferred_bios(struct work_struct *ws)
1786 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1788 bool commit_needed = false;
1789 struct bio_list bios;
1792 bio_list_init(&bios);
1794 spin_lock_irq(&cache->lock);
1795 bio_list_merge(&bios, &cache->deferred_bios);
1796 bio_list_init(&cache->deferred_bios);
1797 spin_unlock_irq(&cache->lock);
1799 while ((bio = bio_list_pop(&bios))) {
1800 if (bio->bi_opf & REQ_PREFLUSH)
1801 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1803 else if (bio_op(bio) == REQ_OP_DISCARD)
1804 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1807 commit_needed = process_bio(cache, bio) || commit_needed;
1811 schedule_commit(&cache->committer);
1814 /*----------------------------------------------------------------
1816 *--------------------------------------------------------------*/
1818 static void requeue_deferred_bios(struct cache *cache)
1821 struct bio_list bios;
1823 bio_list_init(&bios);
1824 bio_list_merge(&bios, &cache->deferred_bios);
1825 bio_list_init(&cache->deferred_bios);
1827 while ((bio = bio_list_pop(&bios))) {
1828 bio->bi_status = BLK_STS_DM_REQUEUE;
1834 * We want to commit periodically so that not too much
1835 * unwritten metadata builds up.
1837 static void do_waker(struct work_struct *ws)
1839 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1841 policy_tick(cache->policy, true);
1842 wake_migration_worker(cache);
1843 schedule_commit(&cache->committer);
1844 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1847 static void check_migrations(struct work_struct *ws)
1850 struct policy_work *op;
1851 struct cache *cache = container_of(ws, struct cache, migration_worker);
1855 b = spare_migration_bandwidth(cache);
1857 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1862 DMERR_LIMIT("%s: policy_background_work failed",
1863 cache_device_name(cache));
1867 r = mg_start(cache, op, NULL);
1873 /*----------------------------------------------------------------
1875 *--------------------------------------------------------------*/
1878 * This function gets called on the error paths of the constructor, so we
1879 * have to cope with a partially initialised struct.
1881 static void destroy(struct cache *cache)
1885 mempool_exit(&cache->migration_pool);
1888 dm_bio_prison_destroy_v2(cache->prison);
1890 cancel_delayed_work_sync(&cache->waker);
1892 destroy_workqueue(cache->wq);
1894 if (cache->dirty_bitset)
1895 free_bitset(cache->dirty_bitset);
1897 if (cache->discard_bitset)
1898 free_bitset(cache->discard_bitset);
1901 dm_kcopyd_client_destroy(cache->copier);
1904 dm_cache_metadata_close(cache->cmd);
1906 if (cache->metadata_dev)
1907 dm_put_device(cache->ti, cache->metadata_dev);
1909 if (cache->origin_dev)
1910 dm_put_device(cache->ti, cache->origin_dev);
1912 if (cache->cache_dev)
1913 dm_put_device(cache->ti, cache->cache_dev);
1916 dm_cache_policy_destroy(cache->policy);
1918 for (i = 0; i < cache->nr_ctr_args ; i++)
1919 kfree(cache->ctr_args[i]);
1920 kfree(cache->ctr_args);
1922 bioset_exit(&cache->bs);
1927 static void cache_dtr(struct dm_target *ti)
1929 struct cache *cache = ti->private;
1934 static sector_t get_dev_size(struct dm_dev *dev)
1936 return bdev_nr_sectors(dev->bdev);
1939 /*----------------------------------------------------------------*/
1942 * Construct a cache device mapping.
1944 * cache <metadata dev> <cache dev> <origin dev> <block size>
1945 * <#feature args> [<feature arg>]*
1946 * <policy> <#policy args> [<policy arg>]*
1948 * metadata dev : fast device holding the persistent metadata
1949 * cache dev : fast device holding cached data blocks
1950 * origin dev : slow device holding original data blocks
1951 * block size : cache unit size in sectors
1953 * #feature args : number of feature arguments passed
1954 * feature args : writethrough. (The default is writeback.)
1956 * policy : the replacement policy to use
1957 * #policy args : an even number of policy arguments corresponding
1958 * to key/value pairs passed to the policy
1959 * policy args : key/value pairs passed to the policy
1960 * E.g. 'sequential_threshold 1024'
1961 * See cache-policies.txt for details.
1963 * Optional feature arguments are:
1964 * writethrough : write through caching that prohibits cache block
1965 * content from being different from origin block content.
1966 * Without this argument, the default behaviour is to write
1967 * back cache block contents later for performance reasons,
1968 * so they may differ from the corresponding origin blocks.
1971 struct dm_target *ti;
1973 struct dm_dev *metadata_dev;
1975 struct dm_dev *cache_dev;
1976 sector_t cache_sectors;
1978 struct dm_dev *origin_dev;
1979 sector_t origin_sectors;
1981 uint32_t block_size;
1983 const char *policy_name;
1985 const char **policy_argv;
1987 struct cache_features features;
1990 static void destroy_cache_args(struct cache_args *ca)
1992 if (ca->metadata_dev)
1993 dm_put_device(ca->ti, ca->metadata_dev);
1996 dm_put_device(ca->ti, ca->cache_dev);
1999 dm_put_device(ca->ti, ca->origin_dev);
2004 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2007 *error = "Insufficient args";
2014 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2018 sector_t metadata_dev_size;
2020 if (!at_least_one_arg(as, error))
2023 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2026 *error = "Error opening metadata device";
2030 metadata_dev_size = get_dev_size(ca->metadata_dev);
2031 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2032 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2033 ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2038 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2043 if (!at_least_one_arg(as, error))
2046 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2049 *error = "Error opening cache device";
2052 ca->cache_sectors = get_dev_size(ca->cache_dev);
2057 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2062 if (!at_least_one_arg(as, error))
2065 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2068 *error = "Error opening origin device";
2072 ca->origin_sectors = get_dev_size(ca->origin_dev);
2073 if (ca->ti->len > ca->origin_sectors) {
2074 *error = "Device size larger than cached device";
2081 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2084 unsigned long block_size;
2086 if (!at_least_one_arg(as, error))
2089 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2090 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2091 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2092 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2093 *error = "Invalid data block size";
2097 if (block_size > ca->cache_sectors) {
2098 *error = "Data block size is larger than the cache device";
2102 ca->block_size = block_size;
2107 static void init_features(struct cache_features *cf)
2109 cf->mode = CM_WRITE;
2110 cf->io_mode = CM_IO_WRITEBACK;
2111 cf->metadata_version = 1;
2112 cf->discard_passdown = true;
2115 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2118 static const struct dm_arg _args[] = {
2119 {0, 3, "Invalid number of cache feature arguments"},
2122 int r, mode_ctr = 0;
2125 struct cache_features *cf = &ca->features;
2129 r = dm_read_arg_group(_args, as, &argc, error);
2134 arg = dm_shift_arg(as);
2136 if (!strcasecmp(arg, "writeback")) {
2137 cf->io_mode = CM_IO_WRITEBACK;
2141 else if (!strcasecmp(arg, "writethrough")) {
2142 cf->io_mode = CM_IO_WRITETHROUGH;
2146 else if (!strcasecmp(arg, "passthrough")) {
2147 cf->io_mode = CM_IO_PASSTHROUGH;
2151 else if (!strcasecmp(arg, "metadata2"))
2152 cf->metadata_version = 2;
2154 else if (!strcasecmp(arg, "no_discard_passdown"))
2155 cf->discard_passdown = false;
2158 *error = "Unrecognised cache feature requested";
2164 *error = "Duplicate cache io_mode features requested";
2171 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2174 static const struct dm_arg _args[] = {
2175 {0, 1024, "Invalid number of policy arguments"},
2180 if (!at_least_one_arg(as, error))
2183 ca->policy_name = dm_shift_arg(as);
2185 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2189 ca->policy_argv = (const char **)as->argv;
2190 dm_consume_args(as, ca->policy_argc);
2195 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2199 struct dm_arg_set as;
2204 r = parse_metadata_dev(ca, &as, error);
2208 r = parse_cache_dev(ca, &as, error);
2212 r = parse_origin_dev(ca, &as, error);
2216 r = parse_block_size(ca, &as, error);
2220 r = parse_features(ca, &as, error);
2224 r = parse_policy(ca, &as, error);
2231 /*----------------------------------------------------------------*/
2233 static struct kmem_cache *migration_cache;
2235 #define NOT_CORE_OPTION 1
2237 static int process_config_option(struct cache *cache, const char *key, const char *value)
2241 if (!strcasecmp(key, "migration_threshold")) {
2242 if (kstrtoul(value, 10, &tmp))
2245 cache->migration_threshold = tmp;
2249 return NOT_CORE_OPTION;
2252 static int set_config_value(struct cache *cache, const char *key, const char *value)
2254 int r = process_config_option(cache, key, value);
2256 if (r == NOT_CORE_OPTION)
2257 r = policy_set_config_value(cache->policy, key, value);
2260 DMWARN("bad config value for %s: %s", key, value);
2265 static int set_config_values(struct cache *cache, int argc, const char **argv)
2270 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2275 r = set_config_value(cache, argv[0], argv[1]);
2286 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2289 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2291 cache->origin_sectors,
2292 cache->sectors_per_block);
2294 *error = "Error creating cache's policy";
2298 BUG_ON(!cache->policy);
2304 * We want the discard block size to be at least the size of the cache
2305 * block size and have no more than 2^14 discard blocks across the origin.
2307 #define MAX_DISCARD_BLOCKS (1 << 14)
2309 static bool too_many_discard_blocks(sector_t discard_block_size,
2310 sector_t origin_size)
2312 (void) sector_div(origin_size, discard_block_size);
2314 return origin_size > MAX_DISCARD_BLOCKS;
2317 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2318 sector_t origin_size)
2320 sector_t discard_block_size = cache_block_size;
2323 while (too_many_discard_blocks(discard_block_size, origin_size))
2324 discard_block_size *= 2;
2326 return discard_block_size;
2329 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2331 dm_block_t nr_blocks = from_cblock(size);
2333 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2334 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2335 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2336 "Please consider increasing the cache block size to reduce the overall cache block count.",
2337 (unsigned long long) nr_blocks);
2339 cache->cache_size = size;
2342 #define DEFAULT_MIGRATION_THRESHOLD 2048
2344 static int cache_create(struct cache_args *ca, struct cache **result)
2347 char **error = &ca->ti->error;
2348 struct cache *cache;
2349 struct dm_target *ti = ca->ti;
2350 dm_block_t origin_blocks;
2351 struct dm_cache_metadata *cmd;
2352 bool may_format = ca->features.mode == CM_WRITE;
2354 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2359 ti->private = cache;
2360 ti->accounts_remapped_io = true;
2361 ti->num_flush_bios = 2;
2362 ti->flush_supported = true;
2364 ti->num_discard_bios = 1;
2365 ti->discards_supported = true;
2367 ti->per_io_data_size = sizeof(struct per_bio_data);
2369 cache->features = ca->features;
2370 if (writethrough_mode(cache)) {
2371 /* Create bioset for writethrough bios issued to origin */
2372 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2377 cache->metadata_dev = ca->metadata_dev;
2378 cache->origin_dev = ca->origin_dev;
2379 cache->cache_dev = ca->cache_dev;
2381 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2383 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2384 origin_blocks = block_div(origin_blocks, ca->block_size);
2385 cache->origin_blocks = to_oblock(origin_blocks);
2387 cache->sectors_per_block = ca->block_size;
2388 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2393 if (ca->block_size & (ca->block_size - 1)) {
2394 dm_block_t cache_size = ca->cache_sectors;
2396 cache->sectors_per_block_shift = -1;
2397 cache_size = block_div(cache_size, ca->block_size);
2398 set_cache_size(cache, to_cblock(cache_size));
2400 cache->sectors_per_block_shift = __ffs(ca->block_size);
2401 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2404 r = create_cache_policy(cache, ca, error);
2408 cache->policy_nr_args = ca->policy_argc;
2409 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2411 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2413 *error = "Error setting cache policy's config values";
2417 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2418 ca->block_size, may_format,
2419 dm_cache_policy_get_hint_size(cache->policy),
2420 ca->features.metadata_version);
2422 *error = "Error creating metadata object";
2427 set_cache_mode(cache, CM_WRITE);
2428 if (get_cache_mode(cache) != CM_WRITE) {
2429 *error = "Unable to get write access to metadata, please check/repair metadata.";
2434 if (passthrough_mode(cache)) {
2437 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2439 *error = "dm_cache_metadata_all_clean() failed";
2444 *error = "Cannot enter passthrough mode unless all blocks are clean";
2449 policy_allow_migrations(cache->policy, false);
2452 spin_lock_init(&cache->lock);
2453 bio_list_init(&cache->deferred_bios);
2454 atomic_set(&cache->nr_allocated_migrations, 0);
2455 atomic_set(&cache->nr_io_migrations, 0);
2456 init_waitqueue_head(&cache->migration_wait);
2459 atomic_set(&cache->nr_dirty, 0);
2460 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2461 if (!cache->dirty_bitset) {
2462 *error = "could not allocate dirty bitset";
2465 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2467 cache->discard_block_size =
2468 calculate_discard_block_size(cache->sectors_per_block,
2469 cache->origin_sectors);
2470 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2471 cache->discard_block_size));
2472 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2473 if (!cache->discard_bitset) {
2474 *error = "could not allocate discard bitset";
2477 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2479 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2480 if (IS_ERR(cache->copier)) {
2481 *error = "could not create kcopyd client";
2482 r = PTR_ERR(cache->copier);
2486 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2488 *error = "could not create workqueue for metadata object";
2491 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2492 INIT_WORK(&cache->migration_worker, check_migrations);
2493 INIT_DELAYED_WORK(&cache->waker, do_waker);
2495 cache->prison = dm_bio_prison_create_v2(cache->wq);
2496 if (!cache->prison) {
2497 *error = "could not create bio prison";
2501 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2504 *error = "Error creating cache's migration mempool";
2508 cache->need_tick_bio = true;
2509 cache->sized = false;
2510 cache->invalidate = false;
2511 cache->commit_requested = false;
2512 cache->loaded_mappings = false;
2513 cache->loaded_discards = false;
2517 atomic_set(&cache->stats.demotion, 0);
2518 atomic_set(&cache->stats.promotion, 0);
2519 atomic_set(&cache->stats.copies_avoided, 0);
2520 atomic_set(&cache->stats.cache_cell_clash, 0);
2521 atomic_set(&cache->stats.commit_count, 0);
2522 atomic_set(&cache->stats.discard_count, 0);
2524 spin_lock_init(&cache->invalidation_lock);
2525 INIT_LIST_HEAD(&cache->invalidation_requests);
2527 batcher_init(&cache->committer, commit_op, cache,
2528 issue_op, cache, cache->wq);
2529 dm_iot_init(&cache->tracker);
2531 init_rwsem(&cache->background_work_lock);
2532 prevent_background_work(cache);
2541 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2546 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2549 for (i = 0; i < argc; i++) {
2550 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2559 cache->nr_ctr_args = argc;
2560 cache->ctr_args = copy;
2565 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2568 struct cache_args *ca;
2569 struct cache *cache = NULL;
2571 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2573 ti->error = "Error allocating memory for cache";
2578 r = parse_cache_args(ca, argc, argv, &ti->error);
2582 r = cache_create(ca, &cache);
2586 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2592 ti->private = cache;
2594 destroy_cache_args(ca);
2598 /*----------------------------------------------------------------*/
2600 static int cache_map(struct dm_target *ti, struct bio *bio)
2602 struct cache *cache = ti->private;
2606 dm_oblock_t block = get_bio_block(cache, bio);
2608 init_per_bio_data(bio);
2609 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2611 * This can only occur if the io goes to a partial block at
2612 * the end of the origin device. We don't cache these.
2613 * Just remap to the origin and carry on.
2615 remap_to_origin(cache, bio);
2616 accounted_begin(cache, bio);
2617 return DM_MAPIO_REMAPPED;
2620 if (discard_or_flush(bio)) {
2621 defer_bio(cache, bio);
2622 return DM_MAPIO_SUBMITTED;
2625 r = map_bio(cache, bio, block, &commit_needed);
2627 schedule_commit(&cache->committer);
2632 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2634 struct cache *cache = ti->private;
2635 unsigned long flags;
2636 struct per_bio_data *pb = get_per_bio_data(bio);
2639 policy_tick(cache->policy, false);
2641 spin_lock_irqsave(&cache->lock, flags);
2642 cache->need_tick_bio = true;
2643 spin_unlock_irqrestore(&cache->lock, flags);
2646 bio_drop_shared_lock(cache, bio);
2647 accounted_complete(cache, bio);
2649 return DM_ENDIO_DONE;
2652 static int write_dirty_bitset(struct cache *cache)
2656 if (get_cache_mode(cache) >= CM_READ_ONLY)
2659 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2661 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2666 static int write_discard_bitset(struct cache *cache)
2670 if (get_cache_mode(cache) >= CM_READ_ONLY)
2673 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2674 cache->discard_nr_blocks);
2676 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2677 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2681 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2682 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2683 is_discarded(cache, to_dblock(i)));
2685 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2693 static int write_hints(struct cache *cache)
2697 if (get_cache_mode(cache) >= CM_READ_ONLY)
2700 r = dm_cache_write_hints(cache->cmd, cache->policy);
2702 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2710 * returns true on success
2712 static bool sync_metadata(struct cache *cache)
2716 r1 = write_dirty_bitset(cache);
2718 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2720 r2 = write_discard_bitset(cache);
2722 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2726 r3 = write_hints(cache);
2728 DMERR("%s: could not write hints", cache_device_name(cache));
2731 * If writing the above metadata failed, we still commit, but don't
2732 * set the clean shutdown flag. This will effectively force every
2733 * dirty bit to be set on reload.
2735 r4 = commit(cache, !r1 && !r2 && !r3);
2737 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2739 return !r1 && !r2 && !r3 && !r4;
2742 static void cache_postsuspend(struct dm_target *ti)
2744 struct cache *cache = ti->private;
2746 prevent_background_work(cache);
2747 BUG_ON(atomic_read(&cache->nr_io_migrations));
2749 cancel_delayed_work_sync(&cache->waker);
2750 drain_workqueue(cache->wq);
2751 WARN_ON(cache->tracker.in_flight);
2754 * If it's a flush suspend there won't be any deferred bios, so this
2757 requeue_deferred_bios(cache);
2759 if (get_cache_mode(cache) == CM_WRITE)
2760 (void) sync_metadata(cache);
2763 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2764 bool dirty, uint32_t hint, bool hint_valid)
2766 struct cache *cache = context;
2769 set_bit(from_cblock(cblock), cache->dirty_bitset);
2770 atomic_inc(&cache->nr_dirty);
2772 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2774 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2778 * The discard block size in the on disk metadata is not
2779 * necessarily the same as we're currently using. So we have to
2780 * be careful to only set the discarded attribute if we know it
2781 * covers a complete block of the new size.
2783 struct discard_load_info {
2784 struct cache *cache;
2787 * These blocks are sized using the on disk dblock size, rather
2788 * than the current one.
2790 dm_block_t block_size;
2791 dm_block_t discard_begin, discard_end;
2794 static void discard_load_info_init(struct cache *cache,
2795 struct discard_load_info *li)
2798 li->discard_begin = li->discard_end = 0;
2801 static void set_discard_range(struct discard_load_info *li)
2805 if (li->discard_begin == li->discard_end)
2809 * Convert to sectors.
2811 b = li->discard_begin * li->block_size;
2812 e = li->discard_end * li->block_size;
2815 * Then convert back to the current dblock size.
2817 b = dm_sector_div_up(b, li->cache->discard_block_size);
2818 sector_div(e, li->cache->discard_block_size);
2821 * The origin may have shrunk, so we need to check we're still in
2824 if (e > from_dblock(li->cache->discard_nr_blocks))
2825 e = from_dblock(li->cache->discard_nr_blocks);
2828 set_discard(li->cache, to_dblock(b));
2831 static int load_discard(void *context, sector_t discard_block_size,
2832 dm_dblock_t dblock, bool discard)
2834 struct discard_load_info *li = context;
2836 li->block_size = discard_block_size;
2839 if (from_dblock(dblock) == li->discard_end)
2841 * We're already in a discard range, just extend it.
2843 li->discard_end = li->discard_end + 1ULL;
2847 * Emit the old range and start a new one.
2849 set_discard_range(li);
2850 li->discard_begin = from_dblock(dblock);
2851 li->discard_end = li->discard_begin + 1ULL;
2854 set_discard_range(li);
2855 li->discard_begin = li->discard_end = 0;
2861 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2863 sector_t size = get_dev_size(cache->cache_dev);
2864 (void) sector_div(size, cache->sectors_per_block);
2865 return to_cblock(size);
2868 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2870 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2872 DMERR("%s: unable to extend cache due to missing cache table reload",
2873 cache_device_name(cache));
2879 * We can't drop a dirty block when shrinking the cache.
2881 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2882 new_size = to_cblock(from_cblock(new_size) + 1);
2883 if (is_dirty(cache, new_size)) {
2884 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2885 cache_device_name(cache),
2886 (unsigned long long) from_cblock(new_size));
2894 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2898 r = dm_cache_resize(cache->cmd, new_size);
2900 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2901 metadata_operation_failed(cache, "dm_cache_resize", r);
2905 set_cache_size(cache, new_size);
2910 static int cache_preresume(struct dm_target *ti)
2913 struct cache *cache = ti->private;
2914 dm_cblock_t csize = get_cache_dev_size(cache);
2917 * Check to see if the cache has resized.
2919 if (!cache->sized) {
2920 r = resize_cache_dev(cache, csize);
2924 cache->sized = true;
2926 } else if (csize != cache->cache_size) {
2927 if (!can_resize(cache, csize))
2930 r = resize_cache_dev(cache, csize);
2935 if (!cache->loaded_mappings) {
2936 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2937 load_mapping, cache);
2939 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2940 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2944 cache->loaded_mappings = true;
2947 if (!cache->loaded_discards) {
2948 struct discard_load_info li;
2951 * The discard bitset could have been resized, or the
2952 * discard block size changed. To be safe we start by
2953 * setting every dblock to not discarded.
2955 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2957 discard_load_info_init(cache, &li);
2958 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2960 DMERR("%s: could not load origin discards", cache_device_name(cache));
2961 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2964 set_discard_range(&li);
2966 cache->loaded_discards = true;
2972 static void cache_resume(struct dm_target *ti)
2974 struct cache *cache = ti->private;
2976 cache->need_tick_bio = true;
2977 allow_background_work(cache);
2978 do_waker(&cache->waker.work);
2981 static void emit_flags(struct cache *cache, char *result,
2982 unsigned maxlen, ssize_t *sz_ptr)
2984 ssize_t sz = *sz_ptr;
2985 struct cache_features *cf = &cache->features;
2986 unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
2988 DMEMIT("%u ", count);
2990 if (cf->metadata_version == 2)
2991 DMEMIT("metadata2 ");
2993 if (writethrough_mode(cache))
2994 DMEMIT("writethrough ");
2996 else if (passthrough_mode(cache))
2997 DMEMIT("passthrough ");
2999 else if (writeback_mode(cache))
3000 DMEMIT("writeback ");
3004 DMERR("%s: internal error: unknown io mode: %d",
3005 cache_device_name(cache), (int) cf->io_mode);
3008 if (!cf->discard_passdown)
3009 DMEMIT("no_discard_passdown ");
3017 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3018 * <cache block size> <#used cache blocks>/<#total cache blocks>
3019 * <#read hits> <#read misses> <#write hits> <#write misses>
3020 * <#demotions> <#promotions> <#dirty>
3021 * <#features> <features>*
3022 * <#core args> <core args>
3023 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3025 static void cache_status(struct dm_target *ti, status_type_t type,
3026 unsigned status_flags, char *result, unsigned maxlen)
3031 dm_block_t nr_free_blocks_metadata = 0;
3032 dm_block_t nr_blocks_metadata = 0;
3033 char buf[BDEVNAME_SIZE];
3034 struct cache *cache = ti->private;
3035 dm_cblock_t residency;
3039 case STATUSTYPE_INFO:
3040 if (get_cache_mode(cache) == CM_FAIL) {
3045 /* Commit to ensure statistics aren't out-of-date */
3046 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3047 (void) commit(cache, false);
3049 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3051 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3052 cache_device_name(cache), r);
3056 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3058 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3059 cache_device_name(cache), r);
3063 residency = policy_residency(cache->policy);
3065 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3066 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3067 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3068 (unsigned long long)nr_blocks_metadata,
3069 (unsigned long long)cache->sectors_per_block,
3070 (unsigned long long) from_cblock(residency),
3071 (unsigned long long) from_cblock(cache->cache_size),
3072 (unsigned) atomic_read(&cache->stats.read_hit),
3073 (unsigned) atomic_read(&cache->stats.read_miss),
3074 (unsigned) atomic_read(&cache->stats.write_hit),
3075 (unsigned) atomic_read(&cache->stats.write_miss),
3076 (unsigned) atomic_read(&cache->stats.demotion),
3077 (unsigned) atomic_read(&cache->stats.promotion),
3078 (unsigned long) atomic_read(&cache->nr_dirty));
3080 emit_flags(cache, result, maxlen, &sz);
3082 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3084 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3086 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3088 DMERR("%s: policy_emit_config_values returned %d",
3089 cache_device_name(cache), r);
3092 if (get_cache_mode(cache) == CM_READ_ONLY)
3097 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3099 if (r || needs_check)
3100 DMEMIT("needs_check ");
3106 case STATUSTYPE_TABLE:
3107 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3109 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3111 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3114 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3115 DMEMIT(" %s", cache->ctr_args[i]);
3116 if (cache->nr_ctr_args)
3117 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3120 case STATUSTYPE_IMA:
3121 DMEMIT_TARGET_NAME_VERSION(ti->type);
3122 if (get_cache_mode(cache) == CM_FAIL)
3123 DMEMIT(",metadata_mode=fail");
3124 else if (get_cache_mode(cache) == CM_READ_ONLY)
3125 DMEMIT(",metadata_mode=ro");
3127 DMEMIT(",metadata_mode=rw");
3129 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3130 DMEMIT(",cache_metadata_device=%s", buf);
3131 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3132 DMEMIT(",cache_device=%s", buf);
3133 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3134 DMEMIT(",cache_origin_device=%s", buf);
3135 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3136 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3137 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3138 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3139 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3151 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3152 * the one-past-the-end value.
3154 struct cblock_range {
3160 * A cache block range can take two forms:
3162 * i) A single cblock, eg. '3456'
3163 * ii) A begin and end cblock with a dash between, eg. 123-234
3165 static int parse_cblock_range(struct cache *cache, const char *str,
3166 struct cblock_range *result)
3173 * Try and parse form (ii) first.
3175 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3180 result->begin = to_cblock(b);
3181 result->end = to_cblock(e);
3186 * That didn't work, try form (i).
3188 r = sscanf(str, "%llu%c", &b, &dummy);
3193 result->begin = to_cblock(b);
3194 result->end = to_cblock(from_cblock(result->begin) + 1u);
3198 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3202 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3204 uint64_t b = from_cblock(range->begin);
3205 uint64_t e = from_cblock(range->end);
3206 uint64_t n = from_cblock(cache->cache_size);
3209 DMERR("%s: begin cblock out of range: %llu >= %llu",
3210 cache_device_name(cache), b, n);
3215 DMERR("%s: end cblock out of range: %llu > %llu",
3216 cache_device_name(cache), e, n);
3221 DMERR("%s: invalid cblock range: %llu >= %llu",
3222 cache_device_name(cache), b, e);
3229 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3231 return to_cblock(from_cblock(b) + 1);
3234 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3239 * We don't need to do any locking here because we know we're in
3240 * passthrough mode. There's is potential for a race between an
3241 * invalidation triggered by an io and an invalidation message. This
3242 * is harmless, we must not worry if the policy call fails.
3244 while (range->begin != range->end) {
3245 r = invalidate_cblock(cache, range->begin);
3249 range->begin = cblock_succ(range->begin);
3252 cache->commit_requested = true;
3256 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3257 const char **cblock_ranges)
3261 struct cblock_range range;
3263 if (!passthrough_mode(cache)) {
3264 DMERR("%s: cache has to be in passthrough mode for invalidation",
3265 cache_device_name(cache));
3269 for (i = 0; i < count; i++) {
3270 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3274 r = validate_cblock_range(cache, &range);
3279 * Pass begin and end origin blocks to the worker and wake it.
3281 r = request_invalidation(cache, &range);
3293 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3295 * The key migration_threshold is supported by the cache target core.
3297 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3298 char *result, unsigned maxlen)
3300 struct cache *cache = ti->private;
3305 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3306 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3307 cache_device_name(cache));
3311 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3312 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3317 return set_config_value(cache, argv[0], argv[1]);
3320 static int cache_iterate_devices(struct dm_target *ti,
3321 iterate_devices_callout_fn fn, void *data)
3324 struct cache *cache = ti->private;
3326 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3328 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3334 * If discard_passdown was enabled verify that the origin device
3335 * supports discards. Disable discard_passdown if not.
3337 static void disable_passdown_if_not_supported(struct cache *cache)
3339 struct block_device *origin_bdev = cache->origin_dev->bdev;
3340 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3341 const char *reason = NULL;
3343 if (!cache->features.discard_passdown)
3346 if (!bdev_max_discard_sectors(origin_bdev))
3347 reason = "discard unsupported";
3349 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3350 reason = "max discard sectors smaller than a block";
3353 DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3354 origin_bdev, reason);
3355 cache->features.discard_passdown = false;
3359 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3361 struct block_device *origin_bdev = cache->origin_dev->bdev;
3362 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3364 if (!cache->features.discard_passdown) {
3365 /* No passdown is done so setting own virtual limits */
3366 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3367 cache->origin_sectors);
3368 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3373 * cache_iterate_devices() is stacking both origin and fast device limits
3374 * but discards aren't passed to fast device, so inherit origin's limits.
3376 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3377 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3378 limits->discard_granularity = origin_limits->discard_granularity;
3379 limits->discard_alignment = origin_limits->discard_alignment;
3380 limits->discard_misaligned = origin_limits->discard_misaligned;
3383 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3385 struct cache *cache = ti->private;
3386 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3389 * If the system-determined stacked limits are compatible with the
3390 * cache's blocksize (io_opt is a factor) do not override them.
3392 if (io_opt_sectors < cache->sectors_per_block ||
3393 do_div(io_opt_sectors, cache->sectors_per_block)) {
3394 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3395 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3398 disable_passdown_if_not_supported(cache);
3399 set_discard_limits(cache, limits);
3402 /*----------------------------------------------------------------*/
3404 static struct target_type cache_target = {
3406 .version = {2, 2, 0},
3407 .module = THIS_MODULE,
3411 .end_io = cache_end_io,
3412 .postsuspend = cache_postsuspend,
3413 .preresume = cache_preresume,
3414 .resume = cache_resume,
3415 .status = cache_status,
3416 .message = cache_message,
3417 .iterate_devices = cache_iterate_devices,
3418 .io_hints = cache_io_hints,
3421 static int __init dm_cache_init(void)
3425 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3426 if (!migration_cache)
3429 r = dm_register_target(&cache_target);
3431 DMERR("cache target registration failed: %d", r);
3432 kmem_cache_destroy(migration_cache);
3439 static void __exit dm_cache_exit(void)
3441 dm_unregister_target(&cache_target);
3442 kmem_cache_destroy(migration_cache);
3445 module_init(dm_cache_init);
3446 module_exit(dm_cache_exit);
3448 MODULE_DESCRIPTION(DM_NAME " cache target");
3449 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3450 MODULE_LICENSE("GPL");