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-bio-record.h"
10 #include "dm-cache-metadata.h"
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/slab.h>
18 #include <linux/vmalloc.h>
20 #define DM_MSG_PREFIX "cache"
22 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
23 "A percentage of time allocated for copying to and/or from cache");
25 /*----------------------------------------------------------------*/
30 * oblock: index of an origin block
31 * cblock: index of a cache block
32 * promotion: movement of a block from origin to cache
33 * demotion: movement of a block from cache to origin
34 * migration: movement of a block between the origin and cache device,
38 /*----------------------------------------------------------------*/
40 static size_t bitset_size_in_bytes(unsigned nr_entries)
42 return sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
45 static unsigned long *alloc_bitset(unsigned nr_entries)
47 size_t s = bitset_size_in_bytes(nr_entries);
51 static void clear_bitset(void *bitset, unsigned nr_entries)
53 size_t s = bitset_size_in_bytes(nr_entries);
57 static void free_bitset(unsigned long *bits)
62 /*----------------------------------------------------------------*/
64 #define PRISON_CELLS 1024
65 #define MIGRATION_POOL_SIZE 128
66 #define COMMIT_PERIOD HZ
67 #define MIGRATION_COUNT_WINDOW 10
70 * The block size of the device holding cache data must be >= 32KB
72 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
75 * FIXME: the cache is read/write for the time being.
78 CM_WRITE, /* metadata may be changed */
79 CM_READ_ONLY, /* metadata may not be changed */
82 struct cache_features {
94 atomic_t copies_avoided;
95 atomic_t cache_cell_clash;
96 atomic_t commit_count;
97 atomic_t discard_count;
101 struct dm_target *ti;
102 struct dm_target_callbacks callbacks;
105 * Metadata is written to this device.
107 struct dm_dev *metadata_dev;
110 * The slower of the two data devices. Typically a spindle.
112 struct dm_dev *origin_dev;
115 * The faster of the two data devices. Typically an SSD.
117 struct dm_dev *cache_dev;
120 * Cache features such as write-through.
122 struct cache_features features;
125 * Size of the origin device in _complete_ blocks and native sectors.
127 dm_oblock_t origin_blocks;
128 sector_t origin_sectors;
131 * Size of the cache device in blocks.
133 dm_cblock_t cache_size;
136 * Fields for converting from sectors to blocks.
138 uint32_t sectors_per_block;
139 int sectors_per_block_shift;
141 struct dm_cache_metadata *cmd;
144 struct bio_list deferred_bios;
145 struct bio_list deferred_flush_bios;
146 struct bio_list deferred_writethrough_bios;
147 struct list_head quiesced_migrations;
148 struct list_head completed_migrations;
149 struct list_head need_commit_migrations;
150 sector_t migration_threshold;
151 atomic_t nr_migrations;
152 wait_queue_head_t migration_wait;
155 * cache_size entries, dirty if set
157 dm_cblock_t nr_dirty;
158 unsigned long *dirty_bitset;
161 * origin_blocks entries, discarded if set.
163 uint32_t discard_block_size; /* a power of 2 times sectors per block */
164 dm_dblock_t discard_nr_blocks;
165 unsigned long *discard_bitset;
167 struct dm_kcopyd_client *copier;
168 struct workqueue_struct *wq;
169 struct work_struct worker;
171 struct delayed_work waker;
172 unsigned long last_commit_jiffies;
174 struct dm_bio_prison *prison;
175 struct dm_deferred_set *all_io_ds;
177 mempool_t *migration_pool;
178 struct dm_cache_migration *next_migration;
180 struct dm_cache_policy *policy;
181 unsigned policy_nr_args;
183 bool need_tick_bio:1;
186 bool commit_requested:1;
187 bool loaded_mappings:1;
188 bool loaded_discards:1;
190 struct cache_stats stats;
193 * Rather than reconstructing the table line for the status we just
194 * save it and regurgitate.
196 unsigned nr_ctr_args;
197 const char **ctr_args;
200 struct per_bio_data {
203 struct dm_deferred_entry *all_io_entry;
205 /* writethrough fields */
208 bio_end_io_t *saved_bi_end_io;
209 struct dm_bio_details bio_details;
212 struct dm_cache_migration {
213 struct list_head list;
216 unsigned long start_jiffies;
217 dm_oblock_t old_oblock;
218 dm_oblock_t new_oblock;
226 struct dm_bio_prison_cell *old_ocell;
227 struct dm_bio_prison_cell *new_ocell;
231 * Processing a bio in the worker thread may require these memory
232 * allocations. We prealloc to avoid deadlocks (the same worker thread
233 * frees them back to the mempool).
236 struct dm_cache_migration *mg;
237 struct dm_bio_prison_cell *cell1;
238 struct dm_bio_prison_cell *cell2;
241 static void wake_worker(struct cache *cache)
243 queue_work(cache->wq, &cache->worker);
246 /*----------------------------------------------------------------*/
248 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
250 /* FIXME: change to use a local slab. */
251 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
254 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
256 dm_bio_prison_free_cell(cache->prison, cell);
259 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
262 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
268 p->cell1 = alloc_prison_cell(cache);
274 p->cell2 = alloc_prison_cell(cache);
282 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
285 free_prison_cell(cache, p->cell2);
288 free_prison_cell(cache, p->cell1);
291 mempool_free(p->mg, cache->migration_pool);
294 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
296 struct dm_cache_migration *mg = p->mg;
305 * You must have a cell within the prealloc struct to return. If not this
306 * function will BUG() rather than returning NULL.
308 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
310 struct dm_bio_prison_cell *r = NULL;
316 } else if (p->cell2) {
326 * You can't have more than two cells in a prealloc struct. BUG() will be
327 * called if you try and overfill.
329 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
341 /*----------------------------------------------------------------*/
343 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
347 key->block = from_oblock(oblock);
351 * The caller hands in a preallocated cell, and a free function for it.
352 * The cell will be freed if there's an error, or if it wasn't used because
353 * a cell with that key already exists.
355 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
357 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
358 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
359 cell_free_fn free_fn, void *free_context,
360 struct dm_bio_prison_cell **cell_result)
363 struct dm_cell_key key;
365 build_key(oblock, &key);
366 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
368 free_fn(free_context, cell_prealloc);
373 static int get_cell(struct cache *cache,
375 struct prealloc *structs,
376 struct dm_bio_prison_cell **cell_result)
379 struct dm_cell_key key;
380 struct dm_bio_prison_cell *cell_prealloc;
382 cell_prealloc = prealloc_get_cell(structs);
384 build_key(oblock, &key);
385 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
387 prealloc_put_cell(structs, cell_prealloc);
392 /*----------------------------------------------------------------*/
394 static bool is_dirty(struct cache *cache, dm_cblock_t b)
396 return test_bit(from_cblock(b), cache->dirty_bitset);
399 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
401 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
402 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
403 policy_set_dirty(cache->policy, oblock);
407 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
409 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
410 policy_clear_dirty(cache->policy, oblock);
411 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
412 if (!from_cblock(cache->nr_dirty))
413 dm_table_event(cache->ti->table);
417 /*----------------------------------------------------------------*/
418 static bool block_size_is_power_of_two(struct cache *cache)
420 return cache->sectors_per_block_shift >= 0;
423 static dm_block_t block_div(dm_block_t b, uint32_t n)
430 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
432 uint32_t discard_blocks = cache->discard_block_size;
433 dm_block_t b = from_oblock(oblock);
435 if (!block_size_is_power_of_two(cache))
436 discard_blocks = discard_blocks / cache->sectors_per_block;
438 discard_blocks >>= cache->sectors_per_block_shift;
440 b = block_div(b, discard_blocks);
445 static void set_discard(struct cache *cache, dm_dblock_t b)
449 atomic_inc(&cache->stats.discard_count);
451 spin_lock_irqsave(&cache->lock, flags);
452 set_bit(from_dblock(b), cache->discard_bitset);
453 spin_unlock_irqrestore(&cache->lock, flags);
456 static void clear_discard(struct cache *cache, dm_dblock_t b)
460 spin_lock_irqsave(&cache->lock, flags);
461 clear_bit(from_dblock(b), cache->discard_bitset);
462 spin_unlock_irqrestore(&cache->lock, flags);
465 static bool is_discarded(struct cache *cache, dm_dblock_t b)
470 spin_lock_irqsave(&cache->lock, flags);
471 r = test_bit(from_dblock(b), cache->discard_bitset);
472 spin_unlock_irqrestore(&cache->lock, flags);
477 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
482 spin_lock_irqsave(&cache->lock, flags);
483 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
484 cache->discard_bitset);
485 spin_unlock_irqrestore(&cache->lock, flags);
490 /*----------------------------------------------------------------*/
492 static void load_stats(struct cache *cache)
494 struct dm_cache_statistics stats;
496 dm_cache_metadata_get_stats(cache->cmd, &stats);
497 atomic_set(&cache->stats.read_hit, stats.read_hits);
498 atomic_set(&cache->stats.read_miss, stats.read_misses);
499 atomic_set(&cache->stats.write_hit, stats.write_hits);
500 atomic_set(&cache->stats.write_miss, stats.write_misses);
503 static void save_stats(struct cache *cache)
505 struct dm_cache_statistics stats;
507 stats.read_hits = atomic_read(&cache->stats.read_hit);
508 stats.read_misses = atomic_read(&cache->stats.read_miss);
509 stats.write_hits = atomic_read(&cache->stats.write_hit);
510 stats.write_misses = atomic_read(&cache->stats.write_miss);
512 dm_cache_metadata_set_stats(cache->cmd, &stats);
515 /*----------------------------------------------------------------
517 *--------------------------------------------------------------*/
518 static struct per_bio_data *get_per_bio_data(struct bio *bio)
520 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
525 static struct per_bio_data *init_per_bio_data(struct bio *bio)
527 struct per_bio_data *pb = get_per_bio_data(bio);
530 pb->req_nr = dm_bio_get_target_bio_nr(bio);
531 pb->all_io_entry = NULL;
536 /*----------------------------------------------------------------
538 *--------------------------------------------------------------*/
539 static void remap_to_origin(struct cache *cache, struct bio *bio)
541 bio->bi_bdev = cache->origin_dev->bdev;
544 static void remap_to_cache(struct cache *cache, struct bio *bio,
547 sector_t bi_sector = bio->bi_sector;
549 bio->bi_bdev = cache->cache_dev->bdev;
550 if (!block_size_is_power_of_two(cache))
551 bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
552 sector_div(bi_sector, cache->sectors_per_block);
554 bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
555 (bi_sector & (cache->sectors_per_block - 1));
558 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
561 struct per_bio_data *pb = get_per_bio_data(bio);
563 spin_lock_irqsave(&cache->lock, flags);
564 if (cache->need_tick_bio &&
565 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
567 cache->need_tick_bio = false;
569 spin_unlock_irqrestore(&cache->lock, flags);
572 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
575 check_if_tick_bio_needed(cache, bio);
576 remap_to_origin(cache, bio);
577 if (bio_data_dir(bio) == WRITE)
578 clear_discard(cache, oblock_to_dblock(cache, oblock));
581 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
582 dm_oblock_t oblock, dm_cblock_t cblock)
584 remap_to_cache(cache, bio, cblock);
585 if (bio_data_dir(bio) == WRITE) {
586 set_dirty(cache, oblock, cblock);
587 clear_discard(cache, oblock_to_dblock(cache, oblock));
591 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
593 sector_t block_nr = bio->bi_sector;
595 if (!block_size_is_power_of_two(cache))
596 (void) sector_div(block_nr, cache->sectors_per_block);
598 block_nr >>= cache->sectors_per_block_shift;
600 return to_oblock(block_nr);
603 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
605 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
608 static void issue(struct cache *cache, struct bio *bio)
612 if (!bio_triggers_commit(cache, bio)) {
613 generic_make_request(bio);
618 * Batch together any bios that trigger commits and then issue a
619 * single commit for them in do_worker().
621 spin_lock_irqsave(&cache->lock, flags);
622 cache->commit_requested = true;
623 bio_list_add(&cache->deferred_flush_bios, bio);
624 spin_unlock_irqrestore(&cache->lock, flags);
627 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
631 spin_lock_irqsave(&cache->lock, flags);
632 bio_list_add(&cache->deferred_writethrough_bios, bio);
633 spin_unlock_irqrestore(&cache->lock, flags);
638 static void writethrough_endio(struct bio *bio, int err)
640 struct per_bio_data *pb = get_per_bio_data(bio);
641 bio->bi_end_io = pb->saved_bi_end_io;
648 dm_bio_restore(&pb->bio_details, bio);
649 remap_to_cache(pb->cache, bio, pb->cblock);
652 * We can't issue this bio directly, since we're in interrupt
653 * context. So it get's put on a bio list for processing by the
656 defer_writethrough_bio(pb->cache, bio);
660 * When running in writethrough mode we need to send writes to clean blocks
661 * to both the cache and origin devices. In future we'd like to clone the
662 * bio and send them in parallel, but for now we're doing them in
663 * series as this is easier.
665 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
666 dm_oblock_t oblock, dm_cblock_t cblock)
668 struct per_bio_data *pb = get_per_bio_data(bio);
672 pb->saved_bi_end_io = bio->bi_end_io;
673 dm_bio_record(&pb->bio_details, bio);
674 bio->bi_end_io = writethrough_endio;
676 remap_to_origin_clear_discard(pb->cache, bio, oblock);
679 /*----------------------------------------------------------------
680 * Migration processing
682 * Migration covers moving data from the origin device to the cache, or
684 *--------------------------------------------------------------*/
685 static void free_migration(struct dm_cache_migration *mg)
687 mempool_free(mg, mg->cache->migration_pool);
690 static void inc_nr_migrations(struct cache *cache)
692 atomic_inc(&cache->nr_migrations);
695 static void dec_nr_migrations(struct cache *cache)
697 atomic_dec(&cache->nr_migrations);
700 * Wake the worker in case we're suspending the target.
702 wake_up(&cache->migration_wait);
705 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
708 (holder ? dm_cell_release : dm_cell_release_no_holder)
709 (cache->prison, cell, &cache->deferred_bios);
710 free_prison_cell(cache, cell);
713 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
718 spin_lock_irqsave(&cache->lock, flags);
719 __cell_defer(cache, cell, holder);
720 spin_unlock_irqrestore(&cache->lock, flags);
725 static void cleanup_migration(struct dm_cache_migration *mg)
727 dec_nr_migrations(mg->cache);
731 static void migration_failure(struct dm_cache_migration *mg)
733 struct cache *cache = mg->cache;
736 DMWARN_LIMIT("writeback failed; couldn't copy block");
737 set_dirty(cache, mg->old_oblock, mg->cblock);
738 cell_defer(cache, mg->old_ocell, false);
740 } else if (mg->demote) {
741 DMWARN_LIMIT("demotion failed; couldn't copy block");
742 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
744 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
746 cell_defer(cache, mg->new_ocell, 1);
748 DMWARN_LIMIT("promotion failed; couldn't copy block");
749 policy_remove_mapping(cache->policy, mg->new_oblock);
750 cell_defer(cache, mg->new_ocell, 1);
753 cleanup_migration(mg);
756 static void migration_success_pre_commit(struct dm_cache_migration *mg)
759 struct cache *cache = mg->cache;
762 cell_defer(cache, mg->old_ocell, false);
763 clear_dirty(cache, mg->old_oblock, mg->cblock);
764 cleanup_migration(mg);
767 } else if (mg->demote) {
768 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
769 DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
770 policy_force_mapping(cache->policy, mg->new_oblock,
773 cell_defer(cache, mg->new_ocell, true);
774 cleanup_migration(mg);
778 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
779 DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
780 policy_remove_mapping(cache->policy, mg->new_oblock);
781 cleanup_migration(mg);
786 spin_lock_irqsave(&cache->lock, flags);
787 list_add_tail(&mg->list, &cache->need_commit_migrations);
788 cache->commit_requested = true;
789 spin_unlock_irqrestore(&cache->lock, flags);
792 static void migration_success_post_commit(struct dm_cache_migration *mg)
795 struct cache *cache = mg->cache;
798 DMWARN("writeback unexpectedly triggered commit");
801 } else if (mg->demote) {
802 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
807 spin_lock_irqsave(&cache->lock, flags);
808 list_add_tail(&mg->list, &cache->quiesced_migrations);
809 spin_unlock_irqrestore(&cache->lock, flags);
812 cleanup_migration(mg);
815 cell_defer(cache, mg->new_ocell, true);
816 clear_dirty(cache, mg->new_oblock, mg->cblock);
817 cleanup_migration(mg);
821 static void copy_complete(int read_err, unsigned long write_err, void *context)
824 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
825 struct cache *cache = mg->cache;
827 if (read_err || write_err)
830 spin_lock_irqsave(&cache->lock, flags);
831 list_add_tail(&mg->list, &cache->completed_migrations);
832 spin_unlock_irqrestore(&cache->lock, flags);
837 static void issue_copy_real(struct dm_cache_migration *mg)
840 struct dm_io_region o_region, c_region;
841 struct cache *cache = mg->cache;
843 o_region.bdev = cache->origin_dev->bdev;
844 o_region.count = cache->sectors_per_block;
846 c_region.bdev = cache->cache_dev->bdev;
847 c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
848 c_region.count = cache->sectors_per_block;
850 if (mg->writeback || mg->demote) {
852 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
853 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
856 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
857 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
861 migration_failure(mg);
864 static void avoid_copy(struct dm_cache_migration *mg)
866 atomic_inc(&mg->cache->stats.copies_avoided);
867 migration_success_pre_commit(mg);
870 static void issue_copy(struct dm_cache_migration *mg)
873 struct cache *cache = mg->cache;
875 if (mg->writeback || mg->demote)
876 avoid = !is_dirty(cache, mg->cblock) ||
877 is_discarded_oblock(cache, mg->old_oblock);
879 avoid = is_discarded_oblock(cache, mg->new_oblock);
881 avoid ? avoid_copy(mg) : issue_copy_real(mg);
884 static void complete_migration(struct dm_cache_migration *mg)
887 migration_failure(mg);
889 migration_success_pre_commit(mg);
892 static void process_migrations(struct cache *cache, struct list_head *head,
893 void (*fn)(struct dm_cache_migration *))
896 struct list_head list;
897 struct dm_cache_migration *mg, *tmp;
899 INIT_LIST_HEAD(&list);
900 spin_lock_irqsave(&cache->lock, flags);
901 list_splice_init(head, &list);
902 spin_unlock_irqrestore(&cache->lock, flags);
904 list_for_each_entry_safe(mg, tmp, &list, list)
908 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
910 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
913 static void queue_quiesced_migration(struct dm_cache_migration *mg)
916 struct cache *cache = mg->cache;
918 spin_lock_irqsave(&cache->lock, flags);
919 __queue_quiesced_migration(mg);
920 spin_unlock_irqrestore(&cache->lock, flags);
925 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
928 struct dm_cache_migration *mg, *tmp;
930 spin_lock_irqsave(&cache->lock, flags);
931 list_for_each_entry_safe(mg, tmp, work, list)
932 __queue_quiesced_migration(mg);
933 spin_unlock_irqrestore(&cache->lock, flags);
938 static void check_for_quiesced_migrations(struct cache *cache,
939 struct per_bio_data *pb)
941 struct list_head work;
943 if (!pb->all_io_entry)
946 INIT_LIST_HEAD(&work);
947 if (pb->all_io_entry)
948 dm_deferred_entry_dec(pb->all_io_entry, &work);
950 if (!list_empty(&work))
951 queue_quiesced_migrations(cache, &work);
954 static void quiesce_migration(struct dm_cache_migration *mg)
956 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
957 queue_quiesced_migration(mg);
960 static void promote(struct cache *cache, struct prealloc *structs,
961 dm_oblock_t oblock, dm_cblock_t cblock,
962 struct dm_bio_prison_cell *cell)
964 struct dm_cache_migration *mg = prealloc_get_migration(structs);
967 mg->writeback = false;
971 mg->new_oblock = oblock;
973 mg->old_ocell = NULL;
974 mg->new_ocell = cell;
975 mg->start_jiffies = jiffies;
977 inc_nr_migrations(cache);
978 quiesce_migration(mg);
981 static void writeback(struct cache *cache, struct prealloc *structs,
982 dm_oblock_t oblock, dm_cblock_t cblock,
983 struct dm_bio_prison_cell *cell)
985 struct dm_cache_migration *mg = prealloc_get_migration(structs);
988 mg->writeback = true;
992 mg->old_oblock = oblock;
994 mg->old_ocell = cell;
995 mg->new_ocell = NULL;
996 mg->start_jiffies = jiffies;
998 inc_nr_migrations(cache);
999 quiesce_migration(mg);
1002 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1003 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1005 struct dm_bio_prison_cell *old_ocell,
1006 struct dm_bio_prison_cell *new_ocell)
1008 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1011 mg->writeback = false;
1015 mg->old_oblock = old_oblock;
1016 mg->new_oblock = new_oblock;
1017 mg->cblock = cblock;
1018 mg->old_ocell = old_ocell;
1019 mg->new_ocell = new_ocell;
1020 mg->start_jiffies = jiffies;
1022 inc_nr_migrations(cache);
1023 quiesce_migration(mg);
1026 /*----------------------------------------------------------------
1028 *--------------------------------------------------------------*/
1029 static void defer_bio(struct cache *cache, struct bio *bio)
1031 unsigned long flags;
1033 spin_lock_irqsave(&cache->lock, flags);
1034 bio_list_add(&cache->deferred_bios, bio);
1035 spin_unlock_irqrestore(&cache->lock, flags);
1040 static void process_flush_bio(struct cache *cache, struct bio *bio)
1042 struct per_bio_data *pb = get_per_bio_data(bio);
1044 BUG_ON(bio->bi_size);
1046 remap_to_origin(cache, bio);
1048 remap_to_cache(cache, bio, 0);
1054 * People generally discard large parts of a device, eg, the whole device
1055 * when formatting. Splitting these large discards up into cache block
1056 * sized ios and then quiescing (always neccessary for discard) takes too
1059 * We keep it simple, and allow any size of discard to come in, and just
1060 * mark off blocks on the discard bitset. No passdown occurs!
1062 * To implement passdown we need to change the bio_prison such that a cell
1063 * can have a key that spans many blocks.
1065 static void process_discard_bio(struct cache *cache, struct bio *bio)
1067 dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
1068 cache->discard_block_size);
1069 dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
1072 end_block = block_div(end_block, cache->discard_block_size);
1074 for (b = start_block; b < end_block; b++)
1075 set_discard(cache, to_dblock(b));
1080 static bool spare_migration_bandwidth(struct cache *cache)
1082 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1083 cache->sectors_per_block;
1084 return current_volume < cache->migration_threshold;
1087 static bool is_writethrough_io(struct cache *cache, struct bio *bio,
1090 return bio_data_dir(bio) == WRITE &&
1091 cache->features.write_through && !is_dirty(cache, cblock);
1094 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1096 atomic_inc(bio_data_dir(bio) == READ ?
1097 &cache->stats.read_hit : &cache->stats.write_hit);
1100 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1102 atomic_inc(bio_data_dir(bio) == READ ?
1103 &cache->stats.read_miss : &cache->stats.write_miss);
1106 static void process_bio(struct cache *cache, struct prealloc *structs,
1110 bool release_cell = true;
1111 dm_oblock_t block = get_bio_block(cache, bio);
1112 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1113 struct policy_result lookup_result;
1114 struct per_bio_data *pb = get_per_bio_data(bio);
1115 bool discarded_block = is_discarded_oblock(cache, block);
1116 bool can_migrate = discarded_block || spare_migration_bandwidth(cache);
1119 * Check to see if that block is currently migrating.
1121 cell_prealloc = prealloc_get_cell(structs);
1122 r = bio_detain(cache, block, bio, cell_prealloc,
1123 (cell_free_fn) prealloc_put_cell,
1124 structs, &new_ocell);
1128 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1129 bio, &lookup_result);
1131 if (r == -EWOULDBLOCK)
1132 /* migration has been denied */
1133 lookup_result.op = POLICY_MISS;
1135 switch (lookup_result.op) {
1137 inc_hit_counter(cache, bio);
1138 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1140 if (is_writethrough_io(cache, bio, lookup_result.cblock))
1141 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1143 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
1149 inc_miss_counter(cache, bio);
1150 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1151 remap_to_origin_clear_discard(cache, bio, block);
1156 atomic_inc(&cache->stats.promotion);
1157 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1158 release_cell = false;
1161 case POLICY_REPLACE:
1162 cell_prealloc = prealloc_get_cell(structs);
1163 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1164 (cell_free_fn) prealloc_put_cell,
1165 structs, &old_ocell);
1168 * We have to be careful to avoid lock inversion of
1169 * the cells. So we back off, and wait for the
1170 * old_ocell to become free.
1172 policy_force_mapping(cache->policy, block,
1173 lookup_result.old_oblock);
1174 atomic_inc(&cache->stats.cache_cell_clash);
1177 atomic_inc(&cache->stats.demotion);
1178 atomic_inc(&cache->stats.promotion);
1180 demote_then_promote(cache, structs, lookup_result.old_oblock,
1181 block, lookup_result.cblock,
1182 old_ocell, new_ocell);
1183 release_cell = false;
1187 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1188 (unsigned) lookup_result.op);
1193 cell_defer(cache, new_ocell, false);
1196 static int need_commit_due_to_time(struct cache *cache)
1198 return jiffies < cache->last_commit_jiffies ||
1199 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1202 static int commit_if_needed(struct cache *cache)
1204 if (dm_cache_changed_this_transaction(cache->cmd) &&
1205 (cache->commit_requested || need_commit_due_to_time(cache))) {
1206 atomic_inc(&cache->stats.commit_count);
1207 cache->last_commit_jiffies = jiffies;
1208 cache->commit_requested = false;
1209 return dm_cache_commit(cache->cmd, false);
1215 static void process_deferred_bios(struct cache *cache)
1217 unsigned long flags;
1218 struct bio_list bios;
1220 struct prealloc structs;
1222 memset(&structs, 0, sizeof(structs));
1223 bio_list_init(&bios);
1225 spin_lock_irqsave(&cache->lock, flags);
1226 bio_list_merge(&bios, &cache->deferred_bios);
1227 bio_list_init(&cache->deferred_bios);
1228 spin_unlock_irqrestore(&cache->lock, flags);
1230 while (!bio_list_empty(&bios)) {
1232 * If we've got no free migration structs, and processing
1233 * this bio might require one, we pause until there are some
1234 * prepared mappings to process.
1236 if (prealloc_data_structs(cache, &structs)) {
1237 spin_lock_irqsave(&cache->lock, flags);
1238 bio_list_merge(&cache->deferred_bios, &bios);
1239 spin_unlock_irqrestore(&cache->lock, flags);
1243 bio = bio_list_pop(&bios);
1245 if (bio->bi_rw & REQ_FLUSH)
1246 process_flush_bio(cache, bio);
1247 else if (bio->bi_rw & REQ_DISCARD)
1248 process_discard_bio(cache, bio);
1250 process_bio(cache, &structs, bio);
1253 prealloc_free_structs(cache, &structs);
1256 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1258 unsigned long flags;
1259 struct bio_list bios;
1262 bio_list_init(&bios);
1264 spin_lock_irqsave(&cache->lock, flags);
1265 bio_list_merge(&bios, &cache->deferred_flush_bios);
1266 bio_list_init(&cache->deferred_flush_bios);
1267 spin_unlock_irqrestore(&cache->lock, flags);
1269 while ((bio = bio_list_pop(&bios)))
1270 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1273 static void process_deferred_writethrough_bios(struct cache *cache)
1275 unsigned long flags;
1276 struct bio_list bios;
1279 bio_list_init(&bios);
1281 spin_lock_irqsave(&cache->lock, flags);
1282 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1283 bio_list_init(&cache->deferred_writethrough_bios);
1284 spin_unlock_irqrestore(&cache->lock, flags);
1286 while ((bio = bio_list_pop(&bios)))
1287 generic_make_request(bio);
1290 static void writeback_some_dirty_blocks(struct cache *cache)
1295 struct prealloc structs;
1296 struct dm_bio_prison_cell *old_ocell;
1298 memset(&structs, 0, sizeof(structs));
1300 while (spare_migration_bandwidth(cache)) {
1301 if (prealloc_data_structs(cache, &structs))
1304 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1308 r = get_cell(cache, oblock, &structs, &old_ocell);
1310 policy_set_dirty(cache->policy, oblock);
1314 writeback(cache, &structs, oblock, cblock, old_ocell);
1317 prealloc_free_structs(cache, &structs);
1320 /*----------------------------------------------------------------
1322 *--------------------------------------------------------------*/
1323 static void start_quiescing(struct cache *cache)
1325 unsigned long flags;
1327 spin_lock_irqsave(&cache->lock, flags);
1328 cache->quiescing = 1;
1329 spin_unlock_irqrestore(&cache->lock, flags);
1332 static void stop_quiescing(struct cache *cache)
1334 unsigned long flags;
1336 spin_lock_irqsave(&cache->lock, flags);
1337 cache->quiescing = 0;
1338 spin_unlock_irqrestore(&cache->lock, flags);
1341 static bool is_quiescing(struct cache *cache)
1344 unsigned long flags;
1346 spin_lock_irqsave(&cache->lock, flags);
1347 r = cache->quiescing;
1348 spin_unlock_irqrestore(&cache->lock, flags);
1353 static void wait_for_migrations(struct cache *cache)
1355 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1358 static void stop_worker(struct cache *cache)
1360 cancel_delayed_work(&cache->waker);
1361 flush_workqueue(cache->wq);
1364 static void requeue_deferred_io(struct cache *cache)
1367 struct bio_list bios;
1369 bio_list_init(&bios);
1370 bio_list_merge(&bios, &cache->deferred_bios);
1371 bio_list_init(&cache->deferred_bios);
1373 while ((bio = bio_list_pop(&bios)))
1374 bio_endio(bio, DM_ENDIO_REQUEUE);
1377 static int more_work(struct cache *cache)
1379 if (is_quiescing(cache))
1380 return !list_empty(&cache->quiesced_migrations) ||
1381 !list_empty(&cache->completed_migrations) ||
1382 !list_empty(&cache->need_commit_migrations);
1384 return !bio_list_empty(&cache->deferred_bios) ||
1385 !bio_list_empty(&cache->deferred_flush_bios) ||
1386 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1387 !list_empty(&cache->quiesced_migrations) ||
1388 !list_empty(&cache->completed_migrations) ||
1389 !list_empty(&cache->need_commit_migrations);
1392 static void do_worker(struct work_struct *ws)
1394 struct cache *cache = container_of(ws, struct cache, worker);
1397 if (!is_quiescing(cache))
1398 process_deferred_bios(cache);
1400 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1401 process_migrations(cache, &cache->completed_migrations, complete_migration);
1403 writeback_some_dirty_blocks(cache);
1405 process_deferred_writethrough_bios(cache);
1407 if (commit_if_needed(cache)) {
1408 process_deferred_flush_bios(cache, false);
1411 * FIXME: rollback metadata or just go into a
1412 * failure mode and error everything
1415 process_deferred_flush_bios(cache, true);
1416 process_migrations(cache, &cache->need_commit_migrations,
1417 migration_success_post_commit);
1419 } while (more_work(cache));
1423 * We want to commit periodically so that not too much
1424 * unwritten metadata builds up.
1426 static void do_waker(struct work_struct *ws)
1428 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1430 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1433 /*----------------------------------------------------------------*/
1435 static int is_congested(struct dm_dev *dev, int bdi_bits)
1437 struct request_queue *q = bdev_get_queue(dev->bdev);
1438 return bdi_congested(&q->backing_dev_info, bdi_bits);
1441 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1443 struct cache *cache = container_of(cb, struct cache, callbacks);
1445 return is_congested(cache->origin_dev, bdi_bits) ||
1446 is_congested(cache->cache_dev, bdi_bits);
1449 /*----------------------------------------------------------------
1451 *--------------------------------------------------------------*/
1454 * This function gets called on the error paths of the constructor, so we
1455 * have to cope with a partially initialised struct.
1457 static void destroy(struct cache *cache)
1461 if (cache->next_migration)
1462 mempool_free(cache->next_migration, cache->migration_pool);
1464 if (cache->migration_pool)
1465 mempool_destroy(cache->migration_pool);
1467 if (cache->all_io_ds)
1468 dm_deferred_set_destroy(cache->all_io_ds);
1471 dm_bio_prison_destroy(cache->prison);
1474 destroy_workqueue(cache->wq);
1476 if (cache->dirty_bitset)
1477 free_bitset(cache->dirty_bitset);
1479 if (cache->discard_bitset)
1480 free_bitset(cache->discard_bitset);
1483 dm_kcopyd_client_destroy(cache->copier);
1486 dm_cache_metadata_close(cache->cmd);
1488 if (cache->metadata_dev)
1489 dm_put_device(cache->ti, cache->metadata_dev);
1491 if (cache->origin_dev)
1492 dm_put_device(cache->ti, cache->origin_dev);
1494 if (cache->cache_dev)
1495 dm_put_device(cache->ti, cache->cache_dev);
1498 dm_cache_policy_destroy(cache->policy);
1500 for (i = 0; i < cache->nr_ctr_args ; i++)
1501 kfree(cache->ctr_args[i]);
1502 kfree(cache->ctr_args);
1507 static void cache_dtr(struct dm_target *ti)
1509 struct cache *cache = ti->private;
1514 static sector_t get_dev_size(struct dm_dev *dev)
1516 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1519 /*----------------------------------------------------------------*/
1522 * Construct a cache device mapping.
1524 * cache <metadata dev> <cache dev> <origin dev> <block size>
1525 * <#feature args> [<feature arg>]*
1526 * <policy> <#policy args> [<policy arg>]*
1528 * metadata dev : fast device holding the persistent metadata
1529 * cache dev : fast device holding cached data blocks
1530 * origin dev : slow device holding original data blocks
1531 * block size : cache unit size in sectors
1533 * #feature args : number of feature arguments passed
1534 * feature args : writethrough. (The default is writeback.)
1536 * policy : the replacement policy to use
1537 * #policy args : an even number of policy arguments corresponding
1538 * to key/value pairs passed to the policy
1539 * policy args : key/value pairs passed to the policy
1540 * E.g. 'sequential_threshold 1024'
1541 * See cache-policies.txt for details.
1543 * Optional feature arguments are:
1544 * writethrough : write through caching that prohibits cache block
1545 * content from being different from origin block content.
1546 * Without this argument, the default behaviour is to write
1547 * back cache block contents later for performance reasons,
1548 * so they may differ from the corresponding origin blocks.
1551 struct dm_target *ti;
1553 struct dm_dev *metadata_dev;
1555 struct dm_dev *cache_dev;
1556 sector_t cache_sectors;
1558 struct dm_dev *origin_dev;
1559 sector_t origin_sectors;
1561 uint32_t block_size;
1563 const char *policy_name;
1565 const char **policy_argv;
1567 struct cache_features features;
1570 static void destroy_cache_args(struct cache_args *ca)
1572 if (ca->metadata_dev)
1573 dm_put_device(ca->ti, ca->metadata_dev);
1576 dm_put_device(ca->ti, ca->cache_dev);
1579 dm_put_device(ca->ti, ca->origin_dev);
1584 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1587 *error = "Insufficient args";
1594 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1598 sector_t metadata_dev_size;
1599 char b[BDEVNAME_SIZE];
1601 if (!at_least_one_arg(as, error))
1604 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1607 *error = "Error opening metadata device";
1611 metadata_dev_size = get_dev_size(ca->metadata_dev);
1612 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1613 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1614 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1619 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1624 if (!at_least_one_arg(as, error))
1627 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1630 *error = "Error opening cache device";
1633 ca->cache_sectors = get_dev_size(ca->cache_dev);
1638 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1643 if (!at_least_one_arg(as, error))
1646 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1649 *error = "Error opening origin device";
1653 ca->origin_sectors = get_dev_size(ca->origin_dev);
1654 if (ca->ti->len > ca->origin_sectors) {
1655 *error = "Device size larger than cached device";
1662 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1667 if (!at_least_one_arg(as, error))
1670 if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
1671 tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1672 tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1673 *error = "Invalid data block size";
1677 if (tmp > ca->cache_sectors) {
1678 *error = "Data block size is larger than the cache device";
1682 ca->block_size = tmp;
1687 static void init_features(struct cache_features *cf)
1689 cf->mode = CM_WRITE;
1690 cf->write_through = false;
1693 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1696 static struct dm_arg _args[] = {
1697 {0, 1, "Invalid number of cache feature arguments"},
1703 struct cache_features *cf = &ca->features;
1707 r = dm_read_arg_group(_args, as, &argc, error);
1712 arg = dm_shift_arg(as);
1714 if (!strcasecmp(arg, "writeback"))
1715 cf->write_through = false;
1717 else if (!strcasecmp(arg, "writethrough"))
1718 cf->write_through = true;
1721 *error = "Unrecognised cache feature requested";
1729 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
1732 static struct dm_arg _args[] = {
1733 {0, 1024, "Invalid number of policy arguments"},
1738 if (!at_least_one_arg(as, error))
1741 ca->policy_name = dm_shift_arg(as);
1743 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
1747 ca->policy_argv = (const char **)as->argv;
1748 dm_consume_args(as, ca->policy_argc);
1753 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
1757 struct dm_arg_set as;
1762 r = parse_metadata_dev(ca, &as, error);
1766 r = parse_cache_dev(ca, &as, error);
1770 r = parse_origin_dev(ca, &as, error);
1774 r = parse_block_size(ca, &as, error);
1778 r = parse_features(ca, &as, error);
1782 r = parse_policy(ca, &as, error);
1789 /*----------------------------------------------------------------*/
1791 static struct kmem_cache *migration_cache;
1793 static int set_config_values(struct dm_cache_policy *p, int argc, const char **argv)
1798 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
1803 r = policy_set_config_value(p, argv[0], argv[1]);
1805 DMWARN("policy_set_config_value failed: key = '%s', value = '%s'",
1817 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
1822 cache->policy = dm_cache_policy_create(ca->policy_name,
1824 cache->origin_sectors,
1825 cache->sectors_per_block);
1826 if (!cache->policy) {
1827 *error = "Error creating cache's policy";
1831 r = set_config_values(cache->policy, ca->policy_argc, ca->policy_argv);
1833 *error = "Error setting cache policy's config values";
1834 dm_cache_policy_destroy(cache->policy);
1835 cache->policy = NULL;
1842 * We want the discard block size to be a power of two, at least the size
1843 * of the cache block size, and have no more than 2^14 discard blocks
1844 * across the origin.
1846 #define MAX_DISCARD_BLOCKS (1 << 14)
1848 static bool too_many_discard_blocks(sector_t discard_block_size,
1849 sector_t origin_size)
1851 (void) sector_div(origin_size, discard_block_size);
1853 return origin_size > MAX_DISCARD_BLOCKS;
1856 static sector_t calculate_discard_block_size(sector_t cache_block_size,
1857 sector_t origin_size)
1859 sector_t discard_block_size;
1861 discard_block_size = roundup_pow_of_two(cache_block_size);
1864 while (too_many_discard_blocks(discard_block_size, origin_size))
1865 discard_block_size *= 2;
1867 return discard_block_size;
1870 #define DEFAULT_MIGRATION_THRESHOLD (2048 * 100)
1872 static int cache_create(struct cache_args *ca, struct cache **result)
1875 char **error = &ca->ti->error;
1876 struct cache *cache;
1877 struct dm_target *ti = ca->ti;
1878 dm_block_t origin_blocks;
1879 struct dm_cache_metadata *cmd;
1880 bool may_format = ca->features.mode == CM_WRITE;
1882 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
1887 ti->private = cache;
1888 ti->per_bio_data_size = sizeof(struct per_bio_data);
1889 ti->num_flush_bios = 2;
1890 ti->flush_supported = true;
1892 ti->num_discard_bios = 1;
1893 ti->discards_supported = true;
1894 ti->discard_zeroes_data_unsupported = true;
1896 memcpy(&cache->features, &ca->features, sizeof(cache->features));
1898 cache->callbacks.congested_fn = cache_is_congested;
1899 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
1901 cache->metadata_dev = ca->metadata_dev;
1902 cache->origin_dev = ca->origin_dev;
1903 cache->cache_dev = ca->cache_dev;
1905 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
1907 /* FIXME: factor out this whole section */
1908 origin_blocks = cache->origin_sectors = ca->origin_sectors;
1909 origin_blocks = block_div(origin_blocks, ca->block_size);
1910 cache->origin_blocks = to_oblock(origin_blocks);
1912 cache->sectors_per_block = ca->block_size;
1913 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
1918 if (ca->block_size & (ca->block_size - 1)) {
1919 dm_block_t cache_size = ca->cache_sectors;
1921 cache->sectors_per_block_shift = -1;
1922 cache_size = block_div(cache_size, ca->block_size);
1923 cache->cache_size = to_cblock(cache_size);
1925 cache->sectors_per_block_shift = __ffs(ca->block_size);
1926 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
1929 r = create_cache_policy(cache, ca, error);
1932 cache->policy_nr_args = ca->policy_argc;
1934 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
1935 ca->block_size, may_format,
1936 dm_cache_policy_get_hint_size(cache->policy));
1938 *error = "Error creating metadata object";
1944 spin_lock_init(&cache->lock);
1945 bio_list_init(&cache->deferred_bios);
1946 bio_list_init(&cache->deferred_flush_bios);
1947 bio_list_init(&cache->deferred_writethrough_bios);
1948 INIT_LIST_HEAD(&cache->quiesced_migrations);
1949 INIT_LIST_HEAD(&cache->completed_migrations);
1950 INIT_LIST_HEAD(&cache->need_commit_migrations);
1951 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
1952 atomic_set(&cache->nr_migrations, 0);
1953 init_waitqueue_head(&cache->migration_wait);
1955 cache->nr_dirty = 0;
1956 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
1957 if (!cache->dirty_bitset) {
1958 *error = "could not allocate dirty bitset";
1961 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
1963 cache->discard_block_size =
1964 calculate_discard_block_size(cache->sectors_per_block,
1965 cache->origin_sectors);
1966 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
1967 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
1968 if (!cache->discard_bitset) {
1969 *error = "could not allocate discard bitset";
1972 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
1974 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
1975 if (IS_ERR(cache->copier)) {
1976 *error = "could not create kcopyd client";
1977 r = PTR_ERR(cache->copier);
1981 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1983 *error = "could not create workqueue for metadata object";
1986 INIT_WORK(&cache->worker, do_worker);
1987 INIT_DELAYED_WORK(&cache->waker, do_waker);
1988 cache->last_commit_jiffies = jiffies;
1990 cache->prison = dm_bio_prison_create(PRISON_CELLS);
1991 if (!cache->prison) {
1992 *error = "could not create bio prison";
1996 cache->all_io_ds = dm_deferred_set_create();
1997 if (!cache->all_io_ds) {
1998 *error = "could not create all_io deferred set";
2002 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2004 if (!cache->migration_pool) {
2005 *error = "Error creating cache's migration mempool";
2009 cache->next_migration = NULL;
2011 cache->need_tick_bio = true;
2012 cache->sized = false;
2013 cache->quiescing = false;
2014 cache->commit_requested = false;
2015 cache->loaded_mappings = false;
2016 cache->loaded_discards = false;
2020 atomic_set(&cache->stats.demotion, 0);
2021 atomic_set(&cache->stats.promotion, 0);
2022 atomic_set(&cache->stats.copies_avoided, 0);
2023 atomic_set(&cache->stats.cache_cell_clash, 0);
2024 atomic_set(&cache->stats.commit_count, 0);
2025 atomic_set(&cache->stats.discard_count, 0);
2035 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2040 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2043 for (i = 0; i < argc; i++) {
2044 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2053 cache->nr_ctr_args = argc;
2054 cache->ctr_args = copy;
2059 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2062 struct cache_args *ca;
2063 struct cache *cache = NULL;
2065 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2067 ti->error = "Error allocating memory for cache";
2072 r = parse_cache_args(ca, argc, argv, &ti->error);
2076 r = cache_create(ca, &cache);
2080 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2086 ti->private = cache;
2089 destroy_cache_args(ca);
2093 static int cache_map(struct dm_target *ti, struct bio *bio)
2095 struct cache *cache = ti->private;
2098 dm_oblock_t block = get_bio_block(cache, bio);
2099 bool can_migrate = false;
2100 bool discarded_block;
2101 struct dm_bio_prison_cell *cell;
2102 struct policy_result lookup_result;
2103 struct per_bio_data *pb;
2105 if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2107 * This can only occur if the io goes to a partial block at
2108 * the end of the origin device. We don't cache these.
2109 * Just remap to the origin and carry on.
2111 remap_to_origin_clear_discard(cache, bio, block);
2112 return DM_MAPIO_REMAPPED;
2115 pb = init_per_bio_data(bio);
2117 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2118 defer_bio(cache, bio);
2119 return DM_MAPIO_SUBMITTED;
2123 * Check to see if that block is currently migrating.
2125 cell = alloc_prison_cell(cache);
2127 defer_bio(cache, bio);
2128 return DM_MAPIO_SUBMITTED;
2131 r = bio_detain(cache, block, bio, cell,
2132 (cell_free_fn) free_prison_cell,
2136 defer_bio(cache, bio);
2138 return DM_MAPIO_SUBMITTED;
2141 discarded_block = is_discarded_oblock(cache, block);
2143 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2144 bio, &lookup_result);
2145 if (r == -EWOULDBLOCK) {
2146 cell_defer(cache, cell, true);
2147 return DM_MAPIO_SUBMITTED;
2150 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2152 return DM_MAPIO_SUBMITTED;
2155 switch (lookup_result.op) {
2157 inc_hit_counter(cache, bio);
2158 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2160 if (is_writethrough_io(cache, bio, lookup_result.cblock))
2161 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2163 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2165 cell_defer(cache, cell, false);
2169 inc_miss_counter(cache, bio);
2170 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2172 if (pb->req_nr != 0) {
2174 * This is a duplicate writethrough io that is no
2175 * longer needed because the block has been demoted.
2178 cell_defer(cache, cell, false);
2179 return DM_MAPIO_SUBMITTED;
2181 remap_to_origin_clear_discard(cache, bio, block);
2182 cell_defer(cache, cell, false);
2187 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2188 (unsigned) lookup_result.op);
2190 return DM_MAPIO_SUBMITTED;
2193 return DM_MAPIO_REMAPPED;
2196 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2198 struct cache *cache = ti->private;
2199 unsigned long flags;
2200 struct per_bio_data *pb = get_per_bio_data(bio);
2203 policy_tick(cache->policy);
2205 spin_lock_irqsave(&cache->lock, flags);
2206 cache->need_tick_bio = true;
2207 spin_unlock_irqrestore(&cache->lock, flags);
2210 check_for_quiesced_migrations(cache, pb);
2215 static int write_dirty_bitset(struct cache *cache)
2219 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2220 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2221 is_dirty(cache, to_cblock(i)));
2229 static int write_discard_bitset(struct cache *cache)
2233 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2234 cache->discard_nr_blocks);
2236 DMERR("could not resize on-disk discard bitset");
2240 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2241 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2242 is_discarded(cache, to_dblock(i)));
2250 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2253 struct cache *cache = context;
2254 return dm_cache_save_hint(cache->cmd, cblock, hint);
2257 static int write_hints(struct cache *cache)
2261 r = dm_cache_begin_hints(cache->cmd, cache->policy);
2263 DMERR("dm_cache_begin_hints failed");
2267 r = policy_walk_mappings(cache->policy, save_hint, cache);
2269 DMERR("policy_walk_mappings failed");
2275 * returns true on success
2277 static bool sync_metadata(struct cache *cache)
2281 r1 = write_dirty_bitset(cache);
2283 DMERR("could not write dirty bitset");
2285 r2 = write_discard_bitset(cache);
2287 DMERR("could not write discard bitset");
2291 r3 = write_hints(cache);
2293 DMERR("could not write hints");
2296 * If writing the above metadata failed, we still commit, but don't
2297 * set the clean shutdown flag. This will effectively force every
2298 * dirty bit to be set on reload.
2300 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2302 DMERR("could not write cache metadata. Data loss may occur.");
2304 return !r1 && !r2 && !r3 && !r4;
2307 static void cache_postsuspend(struct dm_target *ti)
2309 struct cache *cache = ti->private;
2311 start_quiescing(cache);
2312 wait_for_migrations(cache);
2314 requeue_deferred_io(cache);
2315 stop_quiescing(cache);
2317 (void) sync_metadata(cache);
2320 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2321 bool dirty, uint32_t hint, bool hint_valid)
2324 struct cache *cache = context;
2326 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2331 set_dirty(cache, oblock, cblock);
2333 clear_dirty(cache, oblock, cblock);
2338 static int load_discard(void *context, sector_t discard_block_size,
2339 dm_dblock_t dblock, bool discard)
2341 struct cache *cache = context;
2343 /* FIXME: handle mis-matched block size */
2346 set_discard(cache, dblock);
2348 clear_discard(cache, dblock);
2353 static int cache_preresume(struct dm_target *ti)
2356 struct cache *cache = ti->private;
2357 sector_t actual_cache_size = get_dev_size(cache->cache_dev);
2358 (void) sector_div(actual_cache_size, cache->sectors_per_block);
2361 * Check to see if the cache has resized.
2363 if (from_cblock(cache->cache_size) != actual_cache_size || !cache->sized) {
2364 cache->cache_size = to_cblock(actual_cache_size);
2366 r = dm_cache_resize(cache->cmd, cache->cache_size);
2368 DMERR("could not resize cache metadata");
2372 cache->sized = true;
2375 if (!cache->loaded_mappings) {
2376 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2377 load_mapping, cache);
2379 DMERR("could not load cache mappings");
2383 cache->loaded_mappings = true;
2386 if (!cache->loaded_discards) {
2387 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2389 DMERR("could not load origin discards");
2393 cache->loaded_discards = true;
2399 static void cache_resume(struct dm_target *ti)
2401 struct cache *cache = ti->private;
2403 cache->need_tick_bio = true;
2404 do_waker(&cache->waker.work);
2410 * <#used metadata blocks>/<#total metadata blocks>
2411 * <#read hits> <#read misses> <#write hits> <#write misses>
2412 * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2413 * <#features> <features>*
2414 * <#core args> <core args>
2415 * <#policy args> <policy args>*
2417 static void cache_status(struct dm_target *ti, status_type_t type,
2418 unsigned status_flags, char *result, unsigned maxlen)
2423 dm_block_t nr_free_blocks_metadata = 0;
2424 dm_block_t nr_blocks_metadata = 0;
2425 char buf[BDEVNAME_SIZE];
2426 struct cache *cache = ti->private;
2427 dm_cblock_t residency;
2430 case STATUSTYPE_INFO:
2431 /* Commit to ensure statistics aren't out-of-date */
2432 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2433 r = dm_cache_commit(cache->cmd, false);
2435 DMERR("could not commit metadata for accurate status");
2438 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2439 &nr_free_blocks_metadata);
2441 DMERR("could not get metadata free block count");
2445 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2447 DMERR("could not get metadata device size");
2451 residency = policy_residency(cache->policy);
2453 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2454 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2455 (unsigned long long)nr_blocks_metadata,
2456 (unsigned) atomic_read(&cache->stats.read_hit),
2457 (unsigned) atomic_read(&cache->stats.read_miss),
2458 (unsigned) atomic_read(&cache->stats.write_hit),
2459 (unsigned) atomic_read(&cache->stats.write_miss),
2460 (unsigned) atomic_read(&cache->stats.demotion),
2461 (unsigned) atomic_read(&cache->stats.promotion),
2462 (unsigned long long) from_cblock(residency),
2465 if (cache->features.write_through)
2466 DMEMIT("1 writethrough ");
2470 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2472 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2474 DMERR("policy_emit_config_values returned %d", r);
2479 case STATUSTYPE_TABLE:
2480 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2482 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2484 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2487 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2488 DMEMIT(" %s", cache->ctr_args[i]);
2489 if (cache->nr_ctr_args)
2490 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2499 #define NOT_CORE_OPTION 1
2501 static int process_config_option(struct cache *cache, char **argv)
2505 if (!strcasecmp(argv[0], "migration_threshold")) {
2506 if (kstrtoul(argv[1], 10, &tmp))
2509 cache->migration_threshold = tmp;
2513 return NOT_CORE_OPTION;
2517 * Supports <key> <value>.
2519 * The key migration_threshold is supported by the cache target core.
2521 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2524 struct cache *cache = ti->private;
2529 r = process_config_option(cache, argv);
2530 if (r == NOT_CORE_OPTION)
2531 return policy_set_config_value(cache->policy, argv[0], argv[1]);
2536 static int cache_iterate_devices(struct dm_target *ti,
2537 iterate_devices_callout_fn fn, void *data)
2540 struct cache *cache = ti->private;
2542 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
2544 r = fn(ti, cache->origin_dev, 0, ti->len, data);
2550 * We assume I/O is going to the origin (which is the volume
2551 * more likely to have restrictions e.g. by being striped).
2552 * (Looking up the exact location of the data would be expensive
2553 * and could always be out of date by the time the bio is submitted.)
2555 static int cache_bvec_merge(struct dm_target *ti,
2556 struct bvec_merge_data *bvm,
2557 struct bio_vec *biovec, int max_size)
2559 struct cache *cache = ti->private;
2560 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
2562 if (!q->merge_bvec_fn)
2565 bvm->bi_bdev = cache->origin_dev->bdev;
2566 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2569 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
2572 * FIXME: these limits may be incompatible with the cache device
2574 limits->max_discard_sectors = cache->discard_block_size * 1024;
2575 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
2578 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
2580 struct cache *cache = ti->private;
2582 blk_limits_io_min(limits, 0);
2583 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
2584 set_discard_limits(cache, limits);
2587 /*----------------------------------------------------------------*/
2589 static struct target_type cache_target = {
2591 .version = {1, 1, 0},
2592 .module = THIS_MODULE,
2596 .end_io = cache_end_io,
2597 .postsuspend = cache_postsuspend,
2598 .preresume = cache_preresume,
2599 .resume = cache_resume,
2600 .status = cache_status,
2601 .message = cache_message,
2602 .iterate_devices = cache_iterate_devices,
2603 .merge = cache_bvec_merge,
2604 .io_hints = cache_io_hints,
2607 static int __init dm_cache_init(void)
2611 r = dm_register_target(&cache_target);
2613 DMERR("cache target registration failed: %d", r);
2617 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
2618 if (!migration_cache) {
2619 dm_unregister_target(&cache_target);
2626 static void __exit dm_cache_exit(void)
2628 dm_unregister_target(&cache_target);
2629 kmem_cache_destroy(migration_cache);
2632 module_init(dm_cache_init);
2633 module_exit(dm_cache_exit);
2635 MODULE_DESCRIPTION(DM_NAME " cache target");
2636 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
2637 MODULE_LICENSE("GPL");