2 * Copyright (C) 2012 Red Hat. All rights reserved.
4 * This file is released under the GPL.
8 #include "dm-bio-prison.h"
9 #include "dm-cache-metadata.h"
11 #include <linux/dm-io.h>
12 #include <linux/dm-kcopyd.h>
13 #include <linux/init.h>
14 #include <linux/mempool.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/vmalloc.h>
19 #define DM_MSG_PREFIX "cache"
21 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
22 "A percentage of time allocated for copying to and/or from cache");
24 /*----------------------------------------------------------------*/
29 * oblock: index of an origin block
30 * cblock: index of a cache block
31 * promotion: movement of a block from origin to cache
32 * demotion: movement of a block from cache to origin
33 * migration: movement of a block between the origin and cache device,
37 /*----------------------------------------------------------------*/
39 static size_t bitset_size_in_bytes(unsigned nr_entries)
41 return sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
44 static unsigned long *alloc_bitset(unsigned nr_entries)
46 size_t s = bitset_size_in_bytes(nr_entries);
50 static void clear_bitset(void *bitset, unsigned nr_entries)
52 size_t s = bitset_size_in_bytes(nr_entries);
56 static void free_bitset(unsigned long *bits)
61 /*----------------------------------------------------------------*/
63 #define PRISON_CELLS 1024
64 #define MIGRATION_POOL_SIZE 128
65 #define COMMIT_PERIOD HZ
66 #define MIGRATION_COUNT_WINDOW 10
69 * The block size of the device holding cache data must be >= 32KB
71 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
74 * FIXME: the cache is read/write for the time being.
77 CM_WRITE, /* metadata may be changed */
78 CM_READ_ONLY, /* metadata may not be changed */
81 struct cache_features {
93 atomic_t copies_avoided;
94 atomic_t cache_cell_clash;
95 atomic_t commit_count;
96 atomic_t discard_count;
100 struct dm_target *ti;
101 struct dm_target_callbacks callbacks;
104 * Metadata is written to this device.
106 struct dm_dev *metadata_dev;
109 * The slower of the two data devices. Typically a spindle.
111 struct dm_dev *origin_dev;
114 * The faster of the two data devices. Typically an SSD.
116 struct dm_dev *cache_dev;
119 * Cache features such as write-through.
121 struct cache_features features;
124 * Size of the origin device in _complete_ blocks and native sectors.
126 dm_oblock_t origin_blocks;
127 sector_t origin_sectors;
130 * Size of the cache device in blocks.
132 dm_cblock_t cache_size;
135 * Fields for converting from sectors to blocks.
137 uint32_t sectors_per_block;
138 int sectors_per_block_shift;
140 struct dm_cache_metadata *cmd;
143 struct bio_list deferred_bios;
144 struct bio_list deferred_flush_bios;
145 struct list_head quiesced_migrations;
146 struct list_head completed_migrations;
147 struct list_head need_commit_migrations;
148 sector_t migration_threshold;
149 atomic_t nr_migrations;
150 wait_queue_head_t migration_wait;
153 * cache_size entries, dirty if set
155 dm_cblock_t nr_dirty;
156 unsigned long *dirty_bitset;
159 * origin_blocks entries, discarded if set.
161 uint32_t discard_block_size; /* a power of 2 times sectors per block */
162 dm_dblock_t discard_nr_blocks;
163 unsigned long *discard_bitset;
165 struct dm_kcopyd_client *copier;
166 struct workqueue_struct *wq;
167 struct work_struct worker;
169 struct delayed_work waker;
170 unsigned long last_commit_jiffies;
172 struct dm_bio_prison *prison;
173 struct dm_deferred_set *all_io_ds;
175 mempool_t *migration_pool;
176 struct dm_cache_migration *next_migration;
178 struct dm_cache_policy *policy;
179 unsigned policy_nr_args;
181 bool need_tick_bio:1;
184 bool commit_requested:1;
185 bool loaded_mappings:1;
186 bool loaded_discards:1;
188 struct cache_stats stats;
191 * Rather than reconstructing the table line for the status we just
192 * save it and regurgitate.
194 unsigned nr_ctr_args;
195 const char **ctr_args;
198 struct per_bio_data {
201 struct dm_deferred_entry *all_io_entry;
204 struct dm_cache_migration {
205 struct list_head list;
208 unsigned long start_jiffies;
209 dm_oblock_t old_oblock;
210 dm_oblock_t new_oblock;
218 struct dm_bio_prison_cell *old_ocell;
219 struct dm_bio_prison_cell *new_ocell;
223 * Processing a bio in the worker thread may require these memory
224 * allocations. We prealloc to avoid deadlocks (the same worker thread
225 * frees them back to the mempool).
228 struct dm_cache_migration *mg;
229 struct dm_bio_prison_cell *cell1;
230 struct dm_bio_prison_cell *cell2;
233 static void wake_worker(struct cache *cache)
235 queue_work(cache->wq, &cache->worker);
238 /*----------------------------------------------------------------*/
240 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
242 /* FIXME: change to use a local slab. */
243 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
246 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
248 dm_bio_prison_free_cell(cache->prison, cell);
251 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
254 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
260 p->cell1 = alloc_prison_cell(cache);
266 p->cell2 = alloc_prison_cell(cache);
274 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
277 free_prison_cell(cache, p->cell2);
280 free_prison_cell(cache, p->cell1);
283 mempool_free(p->mg, cache->migration_pool);
286 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
288 struct dm_cache_migration *mg = p->mg;
297 * You must have a cell within the prealloc struct to return. If not this
298 * function will BUG() rather than returning NULL.
300 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
302 struct dm_bio_prison_cell *r = NULL;
308 } else if (p->cell2) {
318 * You can't have more than two cells in a prealloc struct. BUG() will be
319 * called if you try and overfill.
321 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
333 /*----------------------------------------------------------------*/
335 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
339 key->block = from_oblock(oblock);
343 * The caller hands in a preallocated cell, and a free function for it.
344 * The cell will be freed if there's an error, or if it wasn't used because
345 * a cell with that key already exists.
347 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
349 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
350 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
351 cell_free_fn free_fn, void *free_context,
352 struct dm_bio_prison_cell **cell_result)
355 struct dm_cell_key key;
357 build_key(oblock, &key);
358 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
360 free_fn(free_context, cell_prealloc);
365 static int get_cell(struct cache *cache,
367 struct prealloc *structs,
368 struct dm_bio_prison_cell **cell_result)
371 struct dm_cell_key key;
372 struct dm_bio_prison_cell *cell_prealloc;
374 cell_prealloc = prealloc_get_cell(structs);
376 build_key(oblock, &key);
377 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
379 prealloc_put_cell(structs, cell_prealloc);
384 /*----------------------------------------------------------------*/
386 static bool is_dirty(struct cache *cache, dm_cblock_t b)
388 return test_bit(from_cblock(b), cache->dirty_bitset);
391 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
393 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
394 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
395 policy_set_dirty(cache->policy, oblock);
399 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
401 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
402 policy_clear_dirty(cache->policy, oblock);
403 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
404 if (!from_cblock(cache->nr_dirty))
405 dm_table_event(cache->ti->table);
409 /*----------------------------------------------------------------*/
410 static bool block_size_is_power_of_two(struct cache *cache)
412 return cache->sectors_per_block_shift >= 0;
415 static dm_block_t block_div(dm_block_t b, uint32_t n)
422 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
424 uint32_t discard_blocks = cache->discard_block_size;
425 dm_block_t b = from_oblock(oblock);
427 if (!block_size_is_power_of_two(cache))
428 discard_blocks = discard_blocks / cache->sectors_per_block;
430 discard_blocks >>= cache->sectors_per_block_shift;
432 b = block_div(b, discard_blocks);
437 static void set_discard(struct cache *cache, dm_dblock_t b)
441 atomic_inc(&cache->stats.discard_count);
443 spin_lock_irqsave(&cache->lock, flags);
444 set_bit(from_dblock(b), cache->discard_bitset);
445 spin_unlock_irqrestore(&cache->lock, flags);
448 static void clear_discard(struct cache *cache, dm_dblock_t b)
452 spin_lock_irqsave(&cache->lock, flags);
453 clear_bit(from_dblock(b), cache->discard_bitset);
454 spin_unlock_irqrestore(&cache->lock, flags);
457 static bool is_discarded(struct cache *cache, dm_dblock_t b)
462 spin_lock_irqsave(&cache->lock, flags);
463 r = test_bit(from_dblock(b), cache->discard_bitset);
464 spin_unlock_irqrestore(&cache->lock, flags);
469 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
474 spin_lock_irqsave(&cache->lock, flags);
475 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
476 cache->discard_bitset);
477 spin_unlock_irqrestore(&cache->lock, flags);
482 /*----------------------------------------------------------------*/
484 static void load_stats(struct cache *cache)
486 struct dm_cache_statistics stats;
488 dm_cache_metadata_get_stats(cache->cmd, &stats);
489 atomic_set(&cache->stats.read_hit, stats.read_hits);
490 atomic_set(&cache->stats.read_miss, stats.read_misses);
491 atomic_set(&cache->stats.write_hit, stats.write_hits);
492 atomic_set(&cache->stats.write_miss, stats.write_misses);
495 static void save_stats(struct cache *cache)
497 struct dm_cache_statistics stats;
499 stats.read_hits = atomic_read(&cache->stats.read_hit);
500 stats.read_misses = atomic_read(&cache->stats.read_miss);
501 stats.write_hits = atomic_read(&cache->stats.write_hit);
502 stats.write_misses = atomic_read(&cache->stats.write_miss);
504 dm_cache_metadata_set_stats(cache->cmd, &stats);
507 /*----------------------------------------------------------------
509 *--------------------------------------------------------------*/
510 static struct per_bio_data *get_per_bio_data(struct bio *bio)
512 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
517 static struct per_bio_data *init_per_bio_data(struct bio *bio)
519 struct per_bio_data *pb = get_per_bio_data(bio);
522 pb->req_nr = dm_bio_get_target_bio_nr(bio);
523 pb->all_io_entry = NULL;
528 /*----------------------------------------------------------------
530 *--------------------------------------------------------------*/
531 static void remap_to_origin(struct cache *cache, struct bio *bio)
533 bio->bi_bdev = cache->origin_dev->bdev;
536 static void remap_to_cache(struct cache *cache, struct bio *bio,
539 sector_t bi_sector = bio->bi_sector;
541 bio->bi_bdev = cache->cache_dev->bdev;
542 if (!block_size_is_power_of_two(cache))
543 bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
544 sector_div(bi_sector, cache->sectors_per_block);
546 bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
547 (bi_sector & (cache->sectors_per_block - 1));
550 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
553 struct per_bio_data *pb = get_per_bio_data(bio);
555 spin_lock_irqsave(&cache->lock, flags);
556 if (cache->need_tick_bio &&
557 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
559 cache->need_tick_bio = false;
561 spin_unlock_irqrestore(&cache->lock, flags);
564 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
567 check_if_tick_bio_needed(cache, bio);
568 remap_to_origin(cache, bio);
569 if (bio_data_dir(bio) == WRITE)
570 clear_discard(cache, oblock_to_dblock(cache, oblock));
573 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
574 dm_oblock_t oblock, dm_cblock_t cblock)
576 remap_to_cache(cache, bio, cblock);
577 if (bio_data_dir(bio) == WRITE) {
578 set_dirty(cache, oblock, cblock);
579 clear_discard(cache, oblock_to_dblock(cache, oblock));
583 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
585 sector_t block_nr = bio->bi_sector;
587 if (!block_size_is_power_of_two(cache))
588 (void) sector_div(block_nr, cache->sectors_per_block);
590 block_nr >>= cache->sectors_per_block_shift;
592 return to_oblock(block_nr);
595 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
597 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
600 static void issue(struct cache *cache, struct bio *bio)
604 if (!bio_triggers_commit(cache, bio)) {
605 generic_make_request(bio);
610 * Batch together any bios that trigger commits and then issue a
611 * single commit for them in do_worker().
613 spin_lock_irqsave(&cache->lock, flags);
614 cache->commit_requested = true;
615 bio_list_add(&cache->deferred_flush_bios, bio);
616 spin_unlock_irqrestore(&cache->lock, flags);
619 /*----------------------------------------------------------------
620 * Migration processing
622 * Migration covers moving data from the origin device to the cache, or
624 *--------------------------------------------------------------*/
625 static void free_migration(struct dm_cache_migration *mg)
627 mempool_free(mg, mg->cache->migration_pool);
630 static void inc_nr_migrations(struct cache *cache)
632 atomic_inc(&cache->nr_migrations);
635 static void dec_nr_migrations(struct cache *cache)
637 atomic_dec(&cache->nr_migrations);
640 * Wake the worker in case we're suspending the target.
642 wake_up(&cache->migration_wait);
645 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
648 (holder ? dm_cell_release : dm_cell_release_no_holder)
649 (cache->prison, cell, &cache->deferred_bios);
650 free_prison_cell(cache, cell);
653 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
658 spin_lock_irqsave(&cache->lock, flags);
659 __cell_defer(cache, cell, holder);
660 spin_unlock_irqrestore(&cache->lock, flags);
665 static void cleanup_migration(struct dm_cache_migration *mg)
667 dec_nr_migrations(mg->cache);
671 static void migration_failure(struct dm_cache_migration *mg)
673 struct cache *cache = mg->cache;
676 DMWARN_LIMIT("writeback failed; couldn't copy block");
677 set_dirty(cache, mg->old_oblock, mg->cblock);
678 cell_defer(cache, mg->old_ocell, false);
680 } else if (mg->demote) {
681 DMWARN_LIMIT("demotion failed; couldn't copy block");
682 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
684 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
686 cell_defer(cache, mg->new_ocell, 1);
688 DMWARN_LIMIT("promotion failed; couldn't copy block");
689 policy_remove_mapping(cache->policy, mg->new_oblock);
690 cell_defer(cache, mg->new_ocell, 1);
693 cleanup_migration(mg);
696 static void migration_success_pre_commit(struct dm_cache_migration *mg)
699 struct cache *cache = mg->cache;
702 cell_defer(cache, mg->old_ocell, false);
703 clear_dirty(cache, mg->old_oblock, mg->cblock);
704 cleanup_migration(mg);
707 } else if (mg->demote) {
708 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
709 DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
710 policy_force_mapping(cache->policy, mg->new_oblock,
713 cell_defer(cache, mg->new_ocell, true);
714 cleanup_migration(mg);
718 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
719 DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
720 policy_remove_mapping(cache->policy, mg->new_oblock);
721 cleanup_migration(mg);
726 spin_lock_irqsave(&cache->lock, flags);
727 list_add_tail(&mg->list, &cache->need_commit_migrations);
728 cache->commit_requested = true;
729 spin_unlock_irqrestore(&cache->lock, flags);
732 static void migration_success_post_commit(struct dm_cache_migration *mg)
735 struct cache *cache = mg->cache;
738 DMWARN("writeback unexpectedly triggered commit");
741 } else if (mg->demote) {
742 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
747 spin_lock_irqsave(&cache->lock, flags);
748 list_add_tail(&mg->list, &cache->quiesced_migrations);
749 spin_unlock_irqrestore(&cache->lock, flags);
752 cleanup_migration(mg);
755 cell_defer(cache, mg->new_ocell, true);
756 clear_dirty(cache, mg->new_oblock, mg->cblock);
757 cleanup_migration(mg);
761 static void copy_complete(int read_err, unsigned long write_err, void *context)
764 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
765 struct cache *cache = mg->cache;
767 if (read_err || write_err)
770 spin_lock_irqsave(&cache->lock, flags);
771 list_add_tail(&mg->list, &cache->completed_migrations);
772 spin_unlock_irqrestore(&cache->lock, flags);
777 static void issue_copy_real(struct dm_cache_migration *mg)
780 struct dm_io_region o_region, c_region;
781 struct cache *cache = mg->cache;
783 o_region.bdev = cache->origin_dev->bdev;
784 o_region.count = cache->sectors_per_block;
786 c_region.bdev = cache->cache_dev->bdev;
787 c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
788 c_region.count = cache->sectors_per_block;
790 if (mg->writeback || mg->demote) {
792 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
793 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
796 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
797 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
801 migration_failure(mg);
804 static void avoid_copy(struct dm_cache_migration *mg)
806 atomic_inc(&mg->cache->stats.copies_avoided);
807 migration_success_pre_commit(mg);
810 static void issue_copy(struct dm_cache_migration *mg)
813 struct cache *cache = mg->cache;
815 if (mg->writeback || mg->demote)
816 avoid = !is_dirty(cache, mg->cblock) ||
817 is_discarded_oblock(cache, mg->old_oblock);
819 avoid = is_discarded_oblock(cache, mg->new_oblock);
821 avoid ? avoid_copy(mg) : issue_copy_real(mg);
824 static void complete_migration(struct dm_cache_migration *mg)
827 migration_failure(mg);
829 migration_success_pre_commit(mg);
832 static void process_migrations(struct cache *cache, struct list_head *head,
833 void (*fn)(struct dm_cache_migration *))
836 struct list_head list;
837 struct dm_cache_migration *mg, *tmp;
839 INIT_LIST_HEAD(&list);
840 spin_lock_irqsave(&cache->lock, flags);
841 list_splice_init(head, &list);
842 spin_unlock_irqrestore(&cache->lock, flags);
844 list_for_each_entry_safe(mg, tmp, &list, list)
848 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
850 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
853 static void queue_quiesced_migration(struct dm_cache_migration *mg)
856 struct cache *cache = mg->cache;
858 spin_lock_irqsave(&cache->lock, flags);
859 __queue_quiesced_migration(mg);
860 spin_unlock_irqrestore(&cache->lock, flags);
865 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
868 struct dm_cache_migration *mg, *tmp;
870 spin_lock_irqsave(&cache->lock, flags);
871 list_for_each_entry_safe(mg, tmp, work, list)
872 __queue_quiesced_migration(mg);
873 spin_unlock_irqrestore(&cache->lock, flags);
878 static void check_for_quiesced_migrations(struct cache *cache,
879 struct per_bio_data *pb)
881 struct list_head work;
883 if (!pb->all_io_entry)
886 INIT_LIST_HEAD(&work);
887 if (pb->all_io_entry)
888 dm_deferred_entry_dec(pb->all_io_entry, &work);
890 if (!list_empty(&work))
891 queue_quiesced_migrations(cache, &work);
894 static void quiesce_migration(struct dm_cache_migration *mg)
896 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
897 queue_quiesced_migration(mg);
900 static void promote(struct cache *cache, struct prealloc *structs,
901 dm_oblock_t oblock, dm_cblock_t cblock,
902 struct dm_bio_prison_cell *cell)
904 struct dm_cache_migration *mg = prealloc_get_migration(structs);
907 mg->writeback = false;
911 mg->new_oblock = oblock;
913 mg->old_ocell = NULL;
914 mg->new_ocell = cell;
915 mg->start_jiffies = jiffies;
917 inc_nr_migrations(cache);
918 quiesce_migration(mg);
921 static void writeback(struct cache *cache, struct prealloc *structs,
922 dm_oblock_t oblock, dm_cblock_t cblock,
923 struct dm_bio_prison_cell *cell)
925 struct dm_cache_migration *mg = prealloc_get_migration(structs);
928 mg->writeback = true;
932 mg->old_oblock = oblock;
934 mg->old_ocell = cell;
935 mg->new_ocell = NULL;
936 mg->start_jiffies = jiffies;
938 inc_nr_migrations(cache);
939 quiesce_migration(mg);
942 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
943 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
945 struct dm_bio_prison_cell *old_ocell,
946 struct dm_bio_prison_cell *new_ocell)
948 struct dm_cache_migration *mg = prealloc_get_migration(structs);
951 mg->writeback = false;
955 mg->old_oblock = old_oblock;
956 mg->new_oblock = new_oblock;
958 mg->old_ocell = old_ocell;
959 mg->new_ocell = new_ocell;
960 mg->start_jiffies = jiffies;
962 inc_nr_migrations(cache);
963 quiesce_migration(mg);
966 /*----------------------------------------------------------------
968 *--------------------------------------------------------------*/
969 static void defer_bio(struct cache *cache, struct bio *bio)
973 spin_lock_irqsave(&cache->lock, flags);
974 bio_list_add(&cache->deferred_bios, bio);
975 spin_unlock_irqrestore(&cache->lock, flags);
980 static void process_flush_bio(struct cache *cache, struct bio *bio)
982 struct per_bio_data *pb = get_per_bio_data(bio);
984 BUG_ON(bio->bi_size);
986 remap_to_origin(cache, bio);
988 remap_to_cache(cache, bio, 0);
994 * People generally discard large parts of a device, eg, the whole device
995 * when formatting. Splitting these large discards up into cache block
996 * sized ios and then quiescing (always neccessary for discard) takes too
999 * We keep it simple, and allow any size of discard to come in, and just
1000 * mark off blocks on the discard bitset. No passdown occurs!
1002 * To implement passdown we need to change the bio_prison such that a cell
1003 * can have a key that spans many blocks.
1005 static void process_discard_bio(struct cache *cache, struct bio *bio)
1007 dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
1008 cache->discard_block_size);
1009 dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
1012 end_block = block_div(end_block, cache->discard_block_size);
1014 for (b = start_block; b < end_block; b++)
1015 set_discard(cache, to_dblock(b));
1020 static bool spare_migration_bandwidth(struct cache *cache)
1022 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1023 cache->sectors_per_block;
1024 return current_volume < cache->migration_threshold;
1027 static bool is_writethrough_io(struct cache *cache, struct bio *bio,
1030 return bio_data_dir(bio) == WRITE &&
1031 cache->features.write_through && !is_dirty(cache, cblock);
1034 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1036 atomic_inc(bio_data_dir(bio) == READ ?
1037 &cache->stats.read_hit : &cache->stats.write_hit);
1040 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1042 atomic_inc(bio_data_dir(bio) == READ ?
1043 &cache->stats.read_miss : &cache->stats.write_miss);
1046 static void process_bio(struct cache *cache, struct prealloc *structs,
1050 bool release_cell = true;
1051 dm_oblock_t block = get_bio_block(cache, bio);
1052 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1053 struct policy_result lookup_result;
1054 struct per_bio_data *pb = get_per_bio_data(bio);
1055 bool discarded_block = is_discarded_oblock(cache, block);
1056 bool can_migrate = discarded_block || spare_migration_bandwidth(cache);
1059 * Check to see if that block is currently migrating.
1061 cell_prealloc = prealloc_get_cell(structs);
1062 r = bio_detain(cache, block, bio, cell_prealloc,
1063 (cell_free_fn) prealloc_put_cell,
1064 structs, &new_ocell);
1068 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1069 bio, &lookup_result);
1071 if (r == -EWOULDBLOCK)
1072 /* migration has been denied */
1073 lookup_result.op = POLICY_MISS;
1075 switch (lookup_result.op) {
1077 inc_hit_counter(cache, bio);
1078 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1080 if (is_writethrough_io(cache, bio, lookup_result.cblock)) {
1082 * No need to mark anything dirty in write through mode.
1085 remap_to_cache(cache, bio, lookup_result.cblock) :
1086 remap_to_origin_clear_discard(cache, bio, block);
1088 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
1094 inc_miss_counter(cache, bio);
1095 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1097 if (pb->req_nr != 0) {
1099 * This is a duplicate writethrough io that is no
1100 * longer needed because the block has been demoted.
1104 remap_to_origin_clear_discard(cache, bio, block);
1110 atomic_inc(&cache->stats.promotion);
1111 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1112 release_cell = false;
1115 case POLICY_REPLACE:
1116 cell_prealloc = prealloc_get_cell(structs);
1117 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1118 (cell_free_fn) prealloc_put_cell,
1119 structs, &old_ocell);
1122 * We have to be careful to avoid lock inversion of
1123 * the cells. So we back off, and wait for the
1124 * old_ocell to become free.
1126 policy_force_mapping(cache->policy, block,
1127 lookup_result.old_oblock);
1128 atomic_inc(&cache->stats.cache_cell_clash);
1131 atomic_inc(&cache->stats.demotion);
1132 atomic_inc(&cache->stats.promotion);
1134 demote_then_promote(cache, structs, lookup_result.old_oblock,
1135 block, lookup_result.cblock,
1136 old_ocell, new_ocell);
1137 release_cell = false;
1141 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1142 (unsigned) lookup_result.op);
1147 cell_defer(cache, new_ocell, false);
1150 static int need_commit_due_to_time(struct cache *cache)
1152 return jiffies < cache->last_commit_jiffies ||
1153 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1156 static int commit_if_needed(struct cache *cache)
1158 if (dm_cache_changed_this_transaction(cache->cmd) &&
1159 (cache->commit_requested || need_commit_due_to_time(cache))) {
1160 atomic_inc(&cache->stats.commit_count);
1161 cache->last_commit_jiffies = jiffies;
1162 cache->commit_requested = false;
1163 return dm_cache_commit(cache->cmd, false);
1169 static void process_deferred_bios(struct cache *cache)
1171 unsigned long flags;
1172 struct bio_list bios;
1174 struct prealloc structs;
1176 memset(&structs, 0, sizeof(structs));
1177 bio_list_init(&bios);
1179 spin_lock_irqsave(&cache->lock, flags);
1180 bio_list_merge(&bios, &cache->deferred_bios);
1181 bio_list_init(&cache->deferred_bios);
1182 spin_unlock_irqrestore(&cache->lock, flags);
1184 while (!bio_list_empty(&bios)) {
1186 * If we've got no free migration structs, and processing
1187 * this bio might require one, we pause until there are some
1188 * prepared mappings to process.
1190 if (prealloc_data_structs(cache, &structs)) {
1191 spin_lock_irqsave(&cache->lock, flags);
1192 bio_list_merge(&cache->deferred_bios, &bios);
1193 spin_unlock_irqrestore(&cache->lock, flags);
1197 bio = bio_list_pop(&bios);
1199 if (bio->bi_rw & REQ_FLUSH)
1200 process_flush_bio(cache, bio);
1201 else if (bio->bi_rw & REQ_DISCARD)
1202 process_discard_bio(cache, bio);
1204 process_bio(cache, &structs, bio);
1207 prealloc_free_structs(cache, &structs);
1210 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1212 unsigned long flags;
1213 struct bio_list bios;
1216 bio_list_init(&bios);
1218 spin_lock_irqsave(&cache->lock, flags);
1219 bio_list_merge(&bios, &cache->deferred_flush_bios);
1220 bio_list_init(&cache->deferred_flush_bios);
1221 spin_unlock_irqrestore(&cache->lock, flags);
1223 while ((bio = bio_list_pop(&bios)))
1224 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1227 static void writeback_some_dirty_blocks(struct cache *cache)
1232 struct prealloc structs;
1233 struct dm_bio_prison_cell *old_ocell;
1235 memset(&structs, 0, sizeof(structs));
1237 while (spare_migration_bandwidth(cache)) {
1238 if (prealloc_data_structs(cache, &structs))
1241 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1245 r = get_cell(cache, oblock, &structs, &old_ocell);
1247 policy_set_dirty(cache->policy, oblock);
1251 writeback(cache, &structs, oblock, cblock, old_ocell);
1254 prealloc_free_structs(cache, &structs);
1257 /*----------------------------------------------------------------
1259 *--------------------------------------------------------------*/
1260 static void start_quiescing(struct cache *cache)
1262 unsigned long flags;
1264 spin_lock_irqsave(&cache->lock, flags);
1265 cache->quiescing = 1;
1266 spin_unlock_irqrestore(&cache->lock, flags);
1269 static void stop_quiescing(struct cache *cache)
1271 unsigned long flags;
1273 spin_lock_irqsave(&cache->lock, flags);
1274 cache->quiescing = 0;
1275 spin_unlock_irqrestore(&cache->lock, flags);
1278 static bool is_quiescing(struct cache *cache)
1281 unsigned long flags;
1283 spin_lock_irqsave(&cache->lock, flags);
1284 r = cache->quiescing;
1285 spin_unlock_irqrestore(&cache->lock, flags);
1290 static void wait_for_migrations(struct cache *cache)
1292 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1295 static void stop_worker(struct cache *cache)
1297 cancel_delayed_work(&cache->waker);
1298 flush_workqueue(cache->wq);
1301 static void requeue_deferred_io(struct cache *cache)
1304 struct bio_list bios;
1306 bio_list_init(&bios);
1307 bio_list_merge(&bios, &cache->deferred_bios);
1308 bio_list_init(&cache->deferred_bios);
1310 while ((bio = bio_list_pop(&bios)))
1311 bio_endio(bio, DM_ENDIO_REQUEUE);
1314 static int more_work(struct cache *cache)
1316 if (is_quiescing(cache))
1317 return !list_empty(&cache->quiesced_migrations) ||
1318 !list_empty(&cache->completed_migrations) ||
1319 !list_empty(&cache->need_commit_migrations);
1321 return !bio_list_empty(&cache->deferred_bios) ||
1322 !bio_list_empty(&cache->deferred_flush_bios) ||
1323 !list_empty(&cache->quiesced_migrations) ||
1324 !list_empty(&cache->completed_migrations) ||
1325 !list_empty(&cache->need_commit_migrations);
1328 static void do_worker(struct work_struct *ws)
1330 struct cache *cache = container_of(ws, struct cache, worker);
1333 if (!is_quiescing(cache))
1334 process_deferred_bios(cache);
1336 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1337 process_migrations(cache, &cache->completed_migrations, complete_migration);
1339 writeback_some_dirty_blocks(cache);
1341 if (commit_if_needed(cache)) {
1342 process_deferred_flush_bios(cache, false);
1345 * FIXME: rollback metadata or just go into a
1346 * failure mode and error everything
1349 process_deferred_flush_bios(cache, true);
1350 process_migrations(cache, &cache->need_commit_migrations,
1351 migration_success_post_commit);
1353 } while (more_work(cache));
1357 * We want to commit periodically so that not too much
1358 * unwritten metadata builds up.
1360 static void do_waker(struct work_struct *ws)
1362 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1364 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1367 /*----------------------------------------------------------------*/
1369 static int is_congested(struct dm_dev *dev, int bdi_bits)
1371 struct request_queue *q = bdev_get_queue(dev->bdev);
1372 return bdi_congested(&q->backing_dev_info, bdi_bits);
1375 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1377 struct cache *cache = container_of(cb, struct cache, callbacks);
1379 return is_congested(cache->origin_dev, bdi_bits) ||
1380 is_congested(cache->cache_dev, bdi_bits);
1383 /*----------------------------------------------------------------
1385 *--------------------------------------------------------------*/
1388 * This function gets called on the error paths of the constructor, so we
1389 * have to cope with a partially initialised struct.
1391 static void destroy(struct cache *cache)
1395 if (cache->next_migration)
1396 mempool_free(cache->next_migration, cache->migration_pool);
1398 if (cache->migration_pool)
1399 mempool_destroy(cache->migration_pool);
1401 if (cache->all_io_ds)
1402 dm_deferred_set_destroy(cache->all_io_ds);
1405 dm_bio_prison_destroy(cache->prison);
1408 destroy_workqueue(cache->wq);
1410 if (cache->dirty_bitset)
1411 free_bitset(cache->dirty_bitset);
1413 if (cache->discard_bitset)
1414 free_bitset(cache->discard_bitset);
1417 dm_kcopyd_client_destroy(cache->copier);
1420 dm_cache_metadata_close(cache->cmd);
1422 if (cache->metadata_dev)
1423 dm_put_device(cache->ti, cache->metadata_dev);
1425 if (cache->origin_dev)
1426 dm_put_device(cache->ti, cache->origin_dev);
1428 if (cache->cache_dev)
1429 dm_put_device(cache->ti, cache->cache_dev);
1432 dm_cache_policy_destroy(cache->policy);
1434 for (i = 0; i < cache->nr_ctr_args ; i++)
1435 kfree(cache->ctr_args[i]);
1436 kfree(cache->ctr_args);
1441 static void cache_dtr(struct dm_target *ti)
1443 struct cache *cache = ti->private;
1448 static sector_t get_dev_size(struct dm_dev *dev)
1450 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1453 /*----------------------------------------------------------------*/
1456 * Construct a cache device mapping.
1458 * cache <metadata dev> <cache dev> <origin dev> <block size>
1459 * <#feature args> [<feature arg>]*
1460 * <policy> <#policy args> [<policy arg>]*
1462 * metadata dev : fast device holding the persistent metadata
1463 * cache dev : fast device holding cached data blocks
1464 * origin dev : slow device holding original data blocks
1465 * block size : cache unit size in sectors
1467 * #feature args : number of feature arguments passed
1468 * feature args : writethrough. (The default is writeback.)
1470 * policy : the replacement policy to use
1471 * #policy args : an even number of policy arguments corresponding
1472 * to key/value pairs passed to the policy
1473 * policy args : key/value pairs passed to the policy
1474 * E.g. 'sequential_threshold 1024'
1475 * See cache-policies.txt for details.
1477 * Optional feature arguments are:
1478 * writethrough : write through caching that prohibits cache block
1479 * content from being different from origin block content.
1480 * Without this argument, the default behaviour is to write
1481 * back cache block contents later for performance reasons,
1482 * so they may differ from the corresponding origin blocks.
1485 struct dm_target *ti;
1487 struct dm_dev *metadata_dev;
1489 struct dm_dev *cache_dev;
1490 sector_t cache_sectors;
1492 struct dm_dev *origin_dev;
1493 sector_t origin_sectors;
1495 uint32_t block_size;
1497 const char *policy_name;
1499 const char **policy_argv;
1501 struct cache_features features;
1504 static void destroy_cache_args(struct cache_args *ca)
1506 if (ca->metadata_dev)
1507 dm_put_device(ca->ti, ca->metadata_dev);
1510 dm_put_device(ca->ti, ca->cache_dev);
1513 dm_put_device(ca->ti, ca->origin_dev);
1518 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1521 *error = "Insufficient args";
1528 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1532 sector_t metadata_dev_size;
1533 char b[BDEVNAME_SIZE];
1535 if (!at_least_one_arg(as, error))
1538 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1541 *error = "Error opening metadata device";
1545 metadata_dev_size = get_dev_size(ca->metadata_dev);
1546 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1547 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1548 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1553 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1558 if (!at_least_one_arg(as, error))
1561 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1564 *error = "Error opening cache device";
1567 ca->cache_sectors = get_dev_size(ca->cache_dev);
1572 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1577 if (!at_least_one_arg(as, error))
1580 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1583 *error = "Error opening origin device";
1587 ca->origin_sectors = get_dev_size(ca->origin_dev);
1588 if (ca->ti->len > ca->origin_sectors) {
1589 *error = "Device size larger than cached device";
1596 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1601 if (!at_least_one_arg(as, error))
1604 if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
1605 tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1606 tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1607 *error = "Invalid data block size";
1611 if (tmp > ca->cache_sectors) {
1612 *error = "Data block size is larger than the cache device";
1616 ca->block_size = tmp;
1621 static void init_features(struct cache_features *cf)
1623 cf->mode = CM_WRITE;
1624 cf->write_through = false;
1627 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1630 static struct dm_arg _args[] = {
1631 {0, 1, "Invalid number of cache feature arguments"},
1637 struct cache_features *cf = &ca->features;
1641 r = dm_read_arg_group(_args, as, &argc, error);
1646 arg = dm_shift_arg(as);
1648 if (!strcasecmp(arg, "writeback"))
1649 cf->write_through = false;
1651 else if (!strcasecmp(arg, "writethrough"))
1652 cf->write_through = true;
1655 *error = "Unrecognised cache feature requested";
1663 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
1666 static struct dm_arg _args[] = {
1667 {0, 1024, "Invalid number of policy arguments"},
1672 if (!at_least_one_arg(as, error))
1675 ca->policy_name = dm_shift_arg(as);
1677 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
1681 ca->policy_argv = (const char **)as->argv;
1682 dm_consume_args(as, ca->policy_argc);
1687 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
1691 struct dm_arg_set as;
1696 r = parse_metadata_dev(ca, &as, error);
1700 r = parse_cache_dev(ca, &as, error);
1704 r = parse_origin_dev(ca, &as, error);
1708 r = parse_block_size(ca, &as, error);
1712 r = parse_features(ca, &as, error);
1716 r = parse_policy(ca, &as, error);
1723 /*----------------------------------------------------------------*/
1725 static struct kmem_cache *migration_cache;
1727 static int set_config_values(struct dm_cache_policy *p, int argc, const char **argv)
1732 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
1737 r = policy_set_config_value(p, argv[0], argv[1]);
1739 DMWARN("policy_set_config_value failed: key = '%s', value = '%s'",
1751 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
1756 cache->policy = dm_cache_policy_create(ca->policy_name,
1758 cache->origin_sectors,
1759 cache->sectors_per_block);
1760 if (!cache->policy) {
1761 *error = "Error creating cache's policy";
1765 r = set_config_values(cache->policy, ca->policy_argc, ca->policy_argv);
1767 dm_cache_policy_destroy(cache->policy);
1773 * We want the discard block size to be a power of two, at least the size
1774 * of the cache block size, and have no more than 2^14 discard blocks
1775 * across the origin.
1777 #define MAX_DISCARD_BLOCKS (1 << 14)
1779 static bool too_many_discard_blocks(sector_t discard_block_size,
1780 sector_t origin_size)
1782 (void) sector_div(origin_size, discard_block_size);
1784 return origin_size > MAX_DISCARD_BLOCKS;
1787 static sector_t calculate_discard_block_size(sector_t cache_block_size,
1788 sector_t origin_size)
1790 sector_t discard_block_size;
1792 discard_block_size = roundup_pow_of_two(cache_block_size);
1795 while (too_many_discard_blocks(discard_block_size, origin_size))
1796 discard_block_size *= 2;
1798 return discard_block_size;
1801 #define DEFAULT_MIGRATION_THRESHOLD (2048 * 100)
1803 static unsigned cache_num_write_bios(struct dm_target *ti, struct bio *bio);
1805 static int cache_create(struct cache_args *ca, struct cache **result)
1808 char **error = &ca->ti->error;
1809 struct cache *cache;
1810 struct dm_target *ti = ca->ti;
1811 dm_block_t origin_blocks;
1812 struct dm_cache_metadata *cmd;
1813 bool may_format = ca->features.mode == CM_WRITE;
1815 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
1820 ti->private = cache;
1821 ti->per_bio_data_size = sizeof(struct per_bio_data);
1822 ti->num_flush_bios = 2;
1823 ti->flush_supported = true;
1825 ti->num_discard_bios = 1;
1826 ti->discards_supported = true;
1827 ti->discard_zeroes_data_unsupported = true;
1829 memcpy(&cache->features, &ca->features, sizeof(cache->features));
1831 if (cache->features.write_through)
1832 ti->num_write_bios = cache_num_write_bios;
1834 cache->callbacks.congested_fn = cache_is_congested;
1835 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
1837 cache->metadata_dev = ca->metadata_dev;
1838 cache->origin_dev = ca->origin_dev;
1839 cache->cache_dev = ca->cache_dev;
1841 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
1843 /* FIXME: factor out this whole section */
1844 origin_blocks = cache->origin_sectors = ca->origin_sectors;
1845 origin_blocks = block_div(origin_blocks, ca->block_size);
1846 cache->origin_blocks = to_oblock(origin_blocks);
1848 cache->sectors_per_block = ca->block_size;
1849 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
1854 if (ca->block_size & (ca->block_size - 1)) {
1855 dm_block_t cache_size = ca->cache_sectors;
1857 cache->sectors_per_block_shift = -1;
1858 cache_size = block_div(cache_size, ca->block_size);
1859 cache->cache_size = to_cblock(cache_size);
1861 cache->sectors_per_block_shift = __ffs(ca->block_size);
1862 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
1865 r = create_cache_policy(cache, ca, error);
1868 cache->policy_nr_args = ca->policy_argc;
1870 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
1871 ca->block_size, may_format,
1872 dm_cache_policy_get_hint_size(cache->policy));
1874 *error = "Error creating metadata object";
1880 spin_lock_init(&cache->lock);
1881 bio_list_init(&cache->deferred_bios);
1882 bio_list_init(&cache->deferred_flush_bios);
1883 INIT_LIST_HEAD(&cache->quiesced_migrations);
1884 INIT_LIST_HEAD(&cache->completed_migrations);
1885 INIT_LIST_HEAD(&cache->need_commit_migrations);
1886 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
1887 atomic_set(&cache->nr_migrations, 0);
1888 init_waitqueue_head(&cache->migration_wait);
1890 cache->nr_dirty = 0;
1891 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
1892 if (!cache->dirty_bitset) {
1893 *error = "could not allocate dirty bitset";
1896 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
1898 cache->discard_block_size =
1899 calculate_discard_block_size(cache->sectors_per_block,
1900 cache->origin_sectors);
1901 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
1902 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
1903 if (!cache->discard_bitset) {
1904 *error = "could not allocate discard bitset";
1907 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
1909 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1910 if (IS_ERR(cache->copier)) {
1911 *error = "could not create kcopyd client";
1912 r = PTR_ERR(cache->copier);
1916 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1918 *error = "could not create workqueue for metadata object";
1921 INIT_WORK(&cache->worker, do_worker);
1922 INIT_DELAYED_WORK(&cache->waker, do_waker);
1923 cache->last_commit_jiffies = jiffies;
1925 cache->prison = dm_bio_prison_create(PRISON_CELLS);
1926 if (!cache->prison) {
1927 *error = "could not create bio prison";
1931 cache->all_io_ds = dm_deferred_set_create();
1932 if (!cache->all_io_ds) {
1933 *error = "could not create all_io deferred set";
1937 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
1939 if (!cache->migration_pool) {
1940 *error = "Error creating cache's migration mempool";
1944 cache->next_migration = NULL;
1946 cache->need_tick_bio = true;
1947 cache->sized = false;
1948 cache->quiescing = false;
1949 cache->commit_requested = false;
1950 cache->loaded_mappings = false;
1951 cache->loaded_discards = false;
1955 atomic_set(&cache->stats.demotion, 0);
1956 atomic_set(&cache->stats.promotion, 0);
1957 atomic_set(&cache->stats.copies_avoided, 0);
1958 atomic_set(&cache->stats.cache_cell_clash, 0);
1959 atomic_set(&cache->stats.commit_count, 0);
1960 atomic_set(&cache->stats.discard_count, 0);
1970 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
1975 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
1978 for (i = 0; i < argc; i++) {
1979 copy[i] = kstrdup(argv[i], GFP_KERNEL);
1988 cache->nr_ctr_args = argc;
1989 cache->ctr_args = copy;
1994 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
1997 struct cache_args *ca;
1998 struct cache *cache = NULL;
2000 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2002 ti->error = "Error allocating memory for cache";
2007 r = parse_cache_args(ca, argc, argv, &ti->error);
2011 r = cache_create(ca, &cache);
2015 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2021 ti->private = cache;
2024 destroy_cache_args(ca);
2028 static unsigned cache_num_write_bios(struct dm_target *ti, struct bio *bio)
2031 struct cache *cache = ti->private;
2032 dm_oblock_t block = get_bio_block(cache, bio);
2035 r = policy_lookup(cache->policy, block, &cblock);
2037 return 2; /* assume the worst */
2039 return (!r && !is_dirty(cache, cblock)) ? 2 : 1;
2042 static int cache_map(struct dm_target *ti, struct bio *bio)
2044 struct cache *cache = ti->private;
2047 dm_oblock_t block = get_bio_block(cache, bio);
2048 bool can_migrate = false;
2049 bool discarded_block;
2050 struct dm_bio_prison_cell *cell;
2051 struct policy_result lookup_result;
2052 struct per_bio_data *pb;
2054 if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2056 * This can only occur if the io goes to a partial block at
2057 * the end of the origin device. We don't cache these.
2058 * Just remap to the origin and carry on.
2060 remap_to_origin_clear_discard(cache, bio, block);
2061 return DM_MAPIO_REMAPPED;
2064 pb = init_per_bio_data(bio);
2066 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2067 defer_bio(cache, bio);
2068 return DM_MAPIO_SUBMITTED;
2072 * Check to see if that block is currently migrating.
2074 cell = alloc_prison_cell(cache);
2076 defer_bio(cache, bio);
2077 return DM_MAPIO_SUBMITTED;
2080 r = bio_detain(cache, block, bio, cell,
2081 (cell_free_fn) free_prison_cell,
2085 defer_bio(cache, bio);
2087 return DM_MAPIO_SUBMITTED;
2090 discarded_block = is_discarded_oblock(cache, block);
2092 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2093 bio, &lookup_result);
2094 if (r == -EWOULDBLOCK) {
2095 cell_defer(cache, cell, true);
2096 return DM_MAPIO_SUBMITTED;
2099 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2101 return DM_MAPIO_SUBMITTED;
2104 switch (lookup_result.op) {
2106 inc_hit_counter(cache, bio);
2107 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2109 if (is_writethrough_io(cache, bio, lookup_result.cblock)) {
2111 * No need to mark anything dirty in write through mode.
2114 remap_to_cache(cache, bio, lookup_result.cblock) :
2115 remap_to_origin_clear_discard(cache, bio, block);
2116 cell_defer(cache, cell, false);
2118 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2119 cell_defer(cache, cell, false);
2124 inc_miss_counter(cache, bio);
2125 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2127 if (pb->req_nr != 0) {
2129 * This is a duplicate writethrough io that is no
2130 * longer needed because the block has been demoted.
2133 cell_defer(cache, cell, false);
2134 return DM_MAPIO_SUBMITTED;
2136 remap_to_origin_clear_discard(cache, bio, block);
2137 cell_defer(cache, cell, false);
2142 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2143 (unsigned) lookup_result.op);
2145 return DM_MAPIO_SUBMITTED;
2148 return DM_MAPIO_REMAPPED;
2151 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2153 struct cache *cache = ti->private;
2154 unsigned long flags;
2155 struct per_bio_data *pb = get_per_bio_data(bio);
2158 policy_tick(cache->policy);
2160 spin_lock_irqsave(&cache->lock, flags);
2161 cache->need_tick_bio = true;
2162 spin_unlock_irqrestore(&cache->lock, flags);
2165 check_for_quiesced_migrations(cache, pb);
2170 static int write_dirty_bitset(struct cache *cache)
2174 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2175 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2176 is_dirty(cache, to_cblock(i)));
2184 static int write_discard_bitset(struct cache *cache)
2188 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2189 cache->discard_nr_blocks);
2191 DMERR("could not resize on-disk discard bitset");
2195 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2196 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2197 is_discarded(cache, to_dblock(i)));
2205 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2208 struct cache *cache = context;
2209 return dm_cache_save_hint(cache->cmd, cblock, hint);
2212 static int write_hints(struct cache *cache)
2216 r = dm_cache_begin_hints(cache->cmd, cache->policy);
2218 DMERR("dm_cache_begin_hints failed");
2222 r = policy_walk_mappings(cache->policy, save_hint, cache);
2224 DMERR("policy_walk_mappings failed");
2230 * returns true on success
2232 static bool sync_metadata(struct cache *cache)
2236 r1 = write_dirty_bitset(cache);
2238 DMERR("could not write dirty bitset");
2240 r2 = write_discard_bitset(cache);
2242 DMERR("could not write discard bitset");
2246 r3 = write_hints(cache);
2248 DMERR("could not write hints");
2251 * If writing the above metadata failed, we still commit, but don't
2252 * set the clean shutdown flag. This will effectively force every
2253 * dirty bit to be set on reload.
2255 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2257 DMERR("could not write cache metadata. Data loss may occur.");
2259 return !r1 && !r2 && !r3 && !r4;
2262 static void cache_postsuspend(struct dm_target *ti)
2264 struct cache *cache = ti->private;
2266 start_quiescing(cache);
2267 wait_for_migrations(cache);
2269 requeue_deferred_io(cache);
2270 stop_quiescing(cache);
2272 (void) sync_metadata(cache);
2275 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2276 bool dirty, uint32_t hint, bool hint_valid)
2279 struct cache *cache = context;
2281 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2286 set_dirty(cache, oblock, cblock);
2288 clear_dirty(cache, oblock, cblock);
2293 static int load_discard(void *context, sector_t discard_block_size,
2294 dm_dblock_t dblock, bool discard)
2296 struct cache *cache = context;
2298 /* FIXME: handle mis-matched block size */
2301 set_discard(cache, dblock);
2303 clear_discard(cache, dblock);
2308 static int cache_preresume(struct dm_target *ti)
2311 struct cache *cache = ti->private;
2312 sector_t actual_cache_size = get_dev_size(cache->cache_dev);
2313 (void) sector_div(actual_cache_size, cache->sectors_per_block);
2316 * Check to see if the cache has resized.
2318 if (from_cblock(cache->cache_size) != actual_cache_size || !cache->sized) {
2319 cache->cache_size = to_cblock(actual_cache_size);
2321 r = dm_cache_resize(cache->cmd, cache->cache_size);
2323 DMERR("could not resize cache metadata");
2327 cache->sized = true;
2330 if (!cache->loaded_mappings) {
2331 r = dm_cache_load_mappings(cache->cmd,
2332 dm_cache_policy_get_name(cache->policy),
2333 load_mapping, cache);
2335 DMERR("could not load cache mappings");
2339 cache->loaded_mappings = true;
2342 if (!cache->loaded_discards) {
2343 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2345 DMERR("could not load origin discards");
2349 cache->loaded_discards = true;
2355 static void cache_resume(struct dm_target *ti)
2357 struct cache *cache = ti->private;
2359 cache->need_tick_bio = true;
2360 do_waker(&cache->waker.work);
2366 * <#used metadata blocks>/<#total metadata blocks>
2367 * <#read hits> <#read misses> <#write hits> <#write misses>
2368 * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2369 * <#features> <features>*
2370 * <#core args> <core args>
2371 * <#policy args> <policy args>*
2373 static void cache_status(struct dm_target *ti, status_type_t type,
2374 unsigned status_flags, char *result, unsigned maxlen)
2379 dm_block_t nr_free_blocks_metadata = 0;
2380 dm_block_t nr_blocks_metadata = 0;
2381 char buf[BDEVNAME_SIZE];
2382 struct cache *cache = ti->private;
2383 dm_cblock_t residency;
2386 case STATUSTYPE_INFO:
2387 /* Commit to ensure statistics aren't out-of-date */
2388 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2389 r = dm_cache_commit(cache->cmd, false);
2391 DMERR("could not commit metadata for accurate status");
2394 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2395 &nr_free_blocks_metadata);
2397 DMERR("could not get metadata free block count");
2401 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2403 DMERR("could not get metadata device size");
2407 residency = policy_residency(cache->policy);
2409 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2410 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2411 (unsigned long long)nr_blocks_metadata,
2412 (unsigned) atomic_read(&cache->stats.read_hit),
2413 (unsigned) atomic_read(&cache->stats.read_miss),
2414 (unsigned) atomic_read(&cache->stats.write_hit),
2415 (unsigned) atomic_read(&cache->stats.write_miss),
2416 (unsigned) atomic_read(&cache->stats.demotion),
2417 (unsigned) atomic_read(&cache->stats.promotion),
2418 (unsigned long long) from_cblock(residency),
2421 if (cache->features.write_through)
2422 DMEMIT("1 writethrough ");
2426 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2428 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2430 DMERR("policy_emit_config_values returned %d", r);
2435 case STATUSTYPE_TABLE:
2436 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2438 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2440 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2443 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2444 DMEMIT(" %s", cache->ctr_args[i]);
2445 if (cache->nr_ctr_args)
2446 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2455 #define NOT_CORE_OPTION 1
2457 static int process_config_option(struct cache *cache, char **argv)
2461 if (!strcasecmp(argv[0], "migration_threshold")) {
2462 if (kstrtoul(argv[1], 10, &tmp))
2465 cache->migration_threshold = tmp;
2469 return NOT_CORE_OPTION;
2473 * Supports <key> <value>.
2475 * The key migration_threshold is supported by the cache target core.
2477 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2480 struct cache *cache = ti->private;
2485 r = process_config_option(cache, argv);
2486 if (r == NOT_CORE_OPTION)
2487 return policy_set_config_value(cache->policy, argv[0], argv[1]);
2492 static int cache_iterate_devices(struct dm_target *ti,
2493 iterate_devices_callout_fn fn, void *data)
2496 struct cache *cache = ti->private;
2498 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
2500 r = fn(ti, cache->origin_dev, 0, ti->len, data);
2506 * We assume I/O is going to the origin (which is the volume
2507 * more likely to have restrictions e.g. by being striped).
2508 * (Looking up the exact location of the data would be expensive
2509 * and could always be out of date by the time the bio is submitted.)
2511 static int cache_bvec_merge(struct dm_target *ti,
2512 struct bvec_merge_data *bvm,
2513 struct bio_vec *biovec, int max_size)
2515 struct cache *cache = ti->private;
2516 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
2518 if (!q->merge_bvec_fn)
2521 bvm->bi_bdev = cache->origin_dev->bdev;
2522 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2525 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
2528 * FIXME: these limits may be incompatible with the cache device
2530 limits->max_discard_sectors = cache->discard_block_size * 1024;
2531 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
2534 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
2536 struct cache *cache = ti->private;
2538 blk_limits_io_min(limits, 0);
2539 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
2540 set_discard_limits(cache, limits);
2543 /*----------------------------------------------------------------*/
2545 static struct target_type cache_target = {
2547 .version = {1, 0, 0},
2548 .module = THIS_MODULE,
2552 .end_io = cache_end_io,
2553 .postsuspend = cache_postsuspend,
2554 .preresume = cache_preresume,
2555 .resume = cache_resume,
2556 .status = cache_status,
2557 .message = cache_message,
2558 .iterate_devices = cache_iterate_devices,
2559 .merge = cache_bvec_merge,
2560 .io_hints = cache_io_hints,
2563 static int __init dm_cache_init(void)
2567 r = dm_register_target(&cache_target);
2569 DMERR("cache target registration failed: %d", r);
2573 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
2574 if (!migration_cache) {
2575 dm_unregister_target(&cache_target);
2582 static void __exit dm_cache_exit(void)
2584 dm_unregister_target(&cache_target);
2585 kmem_cache_destroy(migration_cache);
2588 module_init(dm_cache_init);
2589 module_exit(dm_cache_exit);
2591 MODULE_DESCRIPTION(DM_NAME " cache target");
2592 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
2593 MODULE_LICENSE("GPL");