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;
206 * writethrough fields. These MUST remain at the end of this
207 * structure and the 'cache' member must be the first as it
208 * is used to determine the offset of the writethrough fields.
212 bio_end_io_t *saved_bi_end_io;
213 struct dm_bio_details bio_details;
216 struct dm_cache_migration {
217 struct list_head list;
220 unsigned long start_jiffies;
221 dm_oblock_t old_oblock;
222 dm_oblock_t new_oblock;
230 struct dm_bio_prison_cell *old_ocell;
231 struct dm_bio_prison_cell *new_ocell;
235 * Processing a bio in the worker thread may require these memory
236 * allocations. We prealloc to avoid deadlocks (the same worker thread
237 * frees them back to the mempool).
240 struct dm_cache_migration *mg;
241 struct dm_bio_prison_cell *cell1;
242 struct dm_bio_prison_cell *cell2;
245 static void wake_worker(struct cache *cache)
247 queue_work(cache->wq, &cache->worker);
250 /*----------------------------------------------------------------*/
252 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
254 /* FIXME: change to use a local slab. */
255 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
258 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
260 dm_bio_prison_free_cell(cache->prison, cell);
263 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
266 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
272 p->cell1 = alloc_prison_cell(cache);
278 p->cell2 = alloc_prison_cell(cache);
286 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
289 free_prison_cell(cache, p->cell2);
292 free_prison_cell(cache, p->cell1);
295 mempool_free(p->mg, cache->migration_pool);
298 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
300 struct dm_cache_migration *mg = p->mg;
309 * You must have a cell within the prealloc struct to return. If not this
310 * function will BUG() rather than returning NULL.
312 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
314 struct dm_bio_prison_cell *r = NULL;
320 } else if (p->cell2) {
330 * You can't have more than two cells in a prealloc struct. BUG() will be
331 * called if you try and overfill.
333 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
345 /*----------------------------------------------------------------*/
347 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
351 key->block = from_oblock(oblock);
355 * The caller hands in a preallocated cell, and a free function for it.
356 * The cell will be freed if there's an error, or if it wasn't used because
357 * a cell with that key already exists.
359 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
361 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
362 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
363 cell_free_fn free_fn, void *free_context,
364 struct dm_bio_prison_cell **cell_result)
367 struct dm_cell_key key;
369 build_key(oblock, &key);
370 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
372 free_fn(free_context, cell_prealloc);
377 static int get_cell(struct cache *cache,
379 struct prealloc *structs,
380 struct dm_bio_prison_cell **cell_result)
383 struct dm_cell_key key;
384 struct dm_bio_prison_cell *cell_prealloc;
386 cell_prealloc = prealloc_get_cell(structs);
388 build_key(oblock, &key);
389 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
391 prealloc_put_cell(structs, cell_prealloc);
396 /*----------------------------------------------------------------*/
398 static bool is_dirty(struct cache *cache, dm_cblock_t b)
400 return test_bit(from_cblock(b), cache->dirty_bitset);
403 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
405 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
406 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
407 policy_set_dirty(cache->policy, oblock);
411 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
413 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
414 policy_clear_dirty(cache->policy, oblock);
415 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
416 if (!from_cblock(cache->nr_dirty))
417 dm_table_event(cache->ti->table);
421 /*----------------------------------------------------------------*/
423 static bool block_size_is_power_of_two(struct cache *cache)
425 return cache->sectors_per_block_shift >= 0;
428 static dm_block_t block_div(dm_block_t b, uint32_t n)
435 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
437 uint32_t discard_blocks = cache->discard_block_size;
438 dm_block_t b = from_oblock(oblock);
440 if (!block_size_is_power_of_two(cache))
441 discard_blocks = discard_blocks / cache->sectors_per_block;
443 discard_blocks >>= cache->sectors_per_block_shift;
445 b = block_div(b, discard_blocks);
450 static void set_discard(struct cache *cache, dm_dblock_t b)
454 atomic_inc(&cache->stats.discard_count);
456 spin_lock_irqsave(&cache->lock, flags);
457 set_bit(from_dblock(b), cache->discard_bitset);
458 spin_unlock_irqrestore(&cache->lock, flags);
461 static void clear_discard(struct cache *cache, dm_dblock_t b)
465 spin_lock_irqsave(&cache->lock, flags);
466 clear_bit(from_dblock(b), cache->discard_bitset);
467 spin_unlock_irqrestore(&cache->lock, flags);
470 static bool is_discarded(struct cache *cache, dm_dblock_t b)
475 spin_lock_irqsave(&cache->lock, flags);
476 r = test_bit(from_dblock(b), cache->discard_bitset);
477 spin_unlock_irqrestore(&cache->lock, flags);
482 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
487 spin_lock_irqsave(&cache->lock, flags);
488 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
489 cache->discard_bitset);
490 spin_unlock_irqrestore(&cache->lock, flags);
495 /*----------------------------------------------------------------*/
497 static void load_stats(struct cache *cache)
499 struct dm_cache_statistics stats;
501 dm_cache_metadata_get_stats(cache->cmd, &stats);
502 atomic_set(&cache->stats.read_hit, stats.read_hits);
503 atomic_set(&cache->stats.read_miss, stats.read_misses);
504 atomic_set(&cache->stats.write_hit, stats.write_hits);
505 atomic_set(&cache->stats.write_miss, stats.write_misses);
508 static void save_stats(struct cache *cache)
510 struct dm_cache_statistics stats;
512 stats.read_hits = atomic_read(&cache->stats.read_hit);
513 stats.read_misses = atomic_read(&cache->stats.read_miss);
514 stats.write_hits = atomic_read(&cache->stats.write_hit);
515 stats.write_misses = atomic_read(&cache->stats.write_miss);
517 dm_cache_metadata_set_stats(cache->cmd, &stats);
520 /*----------------------------------------------------------------
522 *--------------------------------------------------------------*/
525 * If using writeback, leave out struct per_bio_data's writethrough fields.
527 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
528 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
530 static size_t get_per_bio_data_size(struct cache *cache)
532 return cache->features.write_through ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
535 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
537 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
542 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
544 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
547 pb->req_nr = dm_bio_get_target_bio_nr(bio);
548 pb->all_io_entry = NULL;
553 /*----------------------------------------------------------------
555 *--------------------------------------------------------------*/
556 static void remap_to_origin(struct cache *cache, struct bio *bio)
558 bio->bi_bdev = cache->origin_dev->bdev;
561 static void remap_to_cache(struct cache *cache, struct bio *bio,
564 sector_t bi_sector = bio->bi_sector;
566 bio->bi_bdev = cache->cache_dev->bdev;
567 if (!block_size_is_power_of_two(cache))
568 bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
569 sector_div(bi_sector, cache->sectors_per_block);
571 bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
572 (bi_sector & (cache->sectors_per_block - 1));
575 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
578 size_t pb_data_size = get_per_bio_data_size(cache);
579 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
581 spin_lock_irqsave(&cache->lock, flags);
582 if (cache->need_tick_bio &&
583 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
585 cache->need_tick_bio = false;
587 spin_unlock_irqrestore(&cache->lock, flags);
590 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
593 check_if_tick_bio_needed(cache, bio);
594 remap_to_origin(cache, bio);
595 if (bio_data_dir(bio) == WRITE)
596 clear_discard(cache, oblock_to_dblock(cache, oblock));
599 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
600 dm_oblock_t oblock, dm_cblock_t cblock)
602 remap_to_cache(cache, bio, cblock);
603 if (bio_data_dir(bio) == WRITE) {
604 set_dirty(cache, oblock, cblock);
605 clear_discard(cache, oblock_to_dblock(cache, oblock));
609 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
611 sector_t block_nr = bio->bi_sector;
613 if (!block_size_is_power_of_two(cache))
614 (void) sector_div(block_nr, cache->sectors_per_block);
616 block_nr >>= cache->sectors_per_block_shift;
618 return to_oblock(block_nr);
621 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
623 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
626 static void issue(struct cache *cache, struct bio *bio)
630 if (!bio_triggers_commit(cache, bio)) {
631 generic_make_request(bio);
636 * Batch together any bios that trigger commits and then issue a
637 * single commit for them in do_worker().
639 spin_lock_irqsave(&cache->lock, flags);
640 cache->commit_requested = true;
641 bio_list_add(&cache->deferred_flush_bios, bio);
642 spin_unlock_irqrestore(&cache->lock, flags);
645 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
649 spin_lock_irqsave(&cache->lock, flags);
650 bio_list_add(&cache->deferred_writethrough_bios, bio);
651 spin_unlock_irqrestore(&cache->lock, flags);
656 static void writethrough_endio(struct bio *bio, int err)
658 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
659 bio->bi_end_io = pb->saved_bi_end_io;
666 dm_bio_restore(&pb->bio_details, bio);
667 remap_to_cache(pb->cache, bio, pb->cblock);
670 * We can't issue this bio directly, since we're in interrupt
671 * context. So it gets put on a bio list for processing by the
674 defer_writethrough_bio(pb->cache, bio);
678 * When running in writethrough mode we need to send writes to clean blocks
679 * to both the cache and origin devices. In future we'd like to clone the
680 * bio and send them in parallel, but for now we're doing them in
681 * series as this is easier.
683 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
684 dm_oblock_t oblock, dm_cblock_t cblock)
686 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
690 pb->saved_bi_end_io = bio->bi_end_io;
691 dm_bio_record(&pb->bio_details, bio);
692 bio->bi_end_io = writethrough_endio;
694 remap_to_origin_clear_discard(pb->cache, bio, oblock);
697 /*----------------------------------------------------------------
698 * Migration processing
700 * Migration covers moving data from the origin device to the cache, or
702 *--------------------------------------------------------------*/
703 static void free_migration(struct dm_cache_migration *mg)
705 mempool_free(mg, mg->cache->migration_pool);
708 static void inc_nr_migrations(struct cache *cache)
710 atomic_inc(&cache->nr_migrations);
713 static void dec_nr_migrations(struct cache *cache)
715 atomic_dec(&cache->nr_migrations);
718 * Wake the worker in case we're suspending the target.
720 wake_up(&cache->migration_wait);
723 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
726 (holder ? dm_cell_release : dm_cell_release_no_holder)
727 (cache->prison, cell, &cache->deferred_bios);
728 free_prison_cell(cache, cell);
731 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
736 spin_lock_irqsave(&cache->lock, flags);
737 __cell_defer(cache, cell, holder);
738 spin_unlock_irqrestore(&cache->lock, flags);
743 static void cleanup_migration(struct dm_cache_migration *mg)
745 dec_nr_migrations(mg->cache);
749 static void migration_failure(struct dm_cache_migration *mg)
751 struct cache *cache = mg->cache;
754 DMWARN_LIMIT("writeback failed; couldn't copy block");
755 set_dirty(cache, mg->old_oblock, mg->cblock);
756 cell_defer(cache, mg->old_ocell, false);
758 } else if (mg->demote) {
759 DMWARN_LIMIT("demotion failed; couldn't copy block");
760 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
762 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
764 cell_defer(cache, mg->new_ocell, 1);
766 DMWARN_LIMIT("promotion failed; couldn't copy block");
767 policy_remove_mapping(cache->policy, mg->new_oblock);
768 cell_defer(cache, mg->new_ocell, 1);
771 cleanup_migration(mg);
774 static void migration_success_pre_commit(struct dm_cache_migration *mg)
777 struct cache *cache = mg->cache;
780 cell_defer(cache, mg->old_ocell, false);
781 clear_dirty(cache, mg->old_oblock, mg->cblock);
782 cleanup_migration(mg);
785 } else if (mg->demote) {
786 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
787 DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
788 policy_force_mapping(cache->policy, mg->new_oblock,
791 cell_defer(cache, mg->new_ocell, true);
792 cleanup_migration(mg);
796 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
797 DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
798 policy_remove_mapping(cache->policy, mg->new_oblock);
799 cleanup_migration(mg);
804 spin_lock_irqsave(&cache->lock, flags);
805 list_add_tail(&mg->list, &cache->need_commit_migrations);
806 cache->commit_requested = true;
807 spin_unlock_irqrestore(&cache->lock, flags);
810 static void migration_success_post_commit(struct dm_cache_migration *mg)
813 struct cache *cache = mg->cache;
816 DMWARN("writeback unexpectedly triggered commit");
819 } else if (mg->demote) {
820 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
825 spin_lock_irqsave(&cache->lock, flags);
826 list_add_tail(&mg->list, &cache->quiesced_migrations);
827 spin_unlock_irqrestore(&cache->lock, flags);
830 cleanup_migration(mg);
833 cell_defer(cache, mg->new_ocell, true);
834 clear_dirty(cache, mg->new_oblock, mg->cblock);
835 cleanup_migration(mg);
839 static void copy_complete(int read_err, unsigned long write_err, void *context)
842 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
843 struct cache *cache = mg->cache;
845 if (read_err || write_err)
848 spin_lock_irqsave(&cache->lock, flags);
849 list_add_tail(&mg->list, &cache->completed_migrations);
850 spin_unlock_irqrestore(&cache->lock, flags);
855 static void issue_copy_real(struct dm_cache_migration *mg)
858 struct dm_io_region o_region, c_region;
859 struct cache *cache = mg->cache;
861 o_region.bdev = cache->origin_dev->bdev;
862 o_region.count = cache->sectors_per_block;
864 c_region.bdev = cache->cache_dev->bdev;
865 c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
866 c_region.count = cache->sectors_per_block;
868 if (mg->writeback || mg->demote) {
870 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
871 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
874 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
875 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
879 migration_failure(mg);
882 static void avoid_copy(struct dm_cache_migration *mg)
884 atomic_inc(&mg->cache->stats.copies_avoided);
885 migration_success_pre_commit(mg);
888 static void issue_copy(struct dm_cache_migration *mg)
891 struct cache *cache = mg->cache;
893 if (mg->writeback || mg->demote)
894 avoid = !is_dirty(cache, mg->cblock) ||
895 is_discarded_oblock(cache, mg->old_oblock);
897 avoid = is_discarded_oblock(cache, mg->new_oblock);
899 avoid ? avoid_copy(mg) : issue_copy_real(mg);
902 static void complete_migration(struct dm_cache_migration *mg)
905 migration_failure(mg);
907 migration_success_pre_commit(mg);
910 static void process_migrations(struct cache *cache, struct list_head *head,
911 void (*fn)(struct dm_cache_migration *))
914 struct list_head list;
915 struct dm_cache_migration *mg, *tmp;
917 INIT_LIST_HEAD(&list);
918 spin_lock_irqsave(&cache->lock, flags);
919 list_splice_init(head, &list);
920 spin_unlock_irqrestore(&cache->lock, flags);
922 list_for_each_entry_safe(mg, tmp, &list, list)
926 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
928 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
931 static void queue_quiesced_migration(struct dm_cache_migration *mg)
934 struct cache *cache = mg->cache;
936 spin_lock_irqsave(&cache->lock, flags);
937 __queue_quiesced_migration(mg);
938 spin_unlock_irqrestore(&cache->lock, flags);
943 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
946 struct dm_cache_migration *mg, *tmp;
948 spin_lock_irqsave(&cache->lock, flags);
949 list_for_each_entry_safe(mg, tmp, work, list)
950 __queue_quiesced_migration(mg);
951 spin_unlock_irqrestore(&cache->lock, flags);
956 static void check_for_quiesced_migrations(struct cache *cache,
957 struct per_bio_data *pb)
959 struct list_head work;
961 if (!pb->all_io_entry)
964 INIT_LIST_HEAD(&work);
965 if (pb->all_io_entry)
966 dm_deferred_entry_dec(pb->all_io_entry, &work);
968 if (!list_empty(&work))
969 queue_quiesced_migrations(cache, &work);
972 static void quiesce_migration(struct dm_cache_migration *mg)
974 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
975 queue_quiesced_migration(mg);
978 static void promote(struct cache *cache, struct prealloc *structs,
979 dm_oblock_t oblock, dm_cblock_t cblock,
980 struct dm_bio_prison_cell *cell)
982 struct dm_cache_migration *mg = prealloc_get_migration(structs);
985 mg->writeback = false;
989 mg->new_oblock = oblock;
991 mg->old_ocell = NULL;
992 mg->new_ocell = cell;
993 mg->start_jiffies = jiffies;
995 inc_nr_migrations(cache);
996 quiesce_migration(mg);
999 static void writeback(struct cache *cache, struct prealloc *structs,
1000 dm_oblock_t oblock, dm_cblock_t cblock,
1001 struct dm_bio_prison_cell *cell)
1003 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1006 mg->writeback = true;
1008 mg->promote = false;
1010 mg->old_oblock = oblock;
1011 mg->cblock = cblock;
1012 mg->old_ocell = cell;
1013 mg->new_ocell = NULL;
1014 mg->start_jiffies = jiffies;
1016 inc_nr_migrations(cache);
1017 quiesce_migration(mg);
1020 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1021 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1023 struct dm_bio_prison_cell *old_ocell,
1024 struct dm_bio_prison_cell *new_ocell)
1026 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1029 mg->writeback = false;
1033 mg->old_oblock = old_oblock;
1034 mg->new_oblock = new_oblock;
1035 mg->cblock = cblock;
1036 mg->old_ocell = old_ocell;
1037 mg->new_ocell = new_ocell;
1038 mg->start_jiffies = jiffies;
1040 inc_nr_migrations(cache);
1041 quiesce_migration(mg);
1044 /*----------------------------------------------------------------
1046 *--------------------------------------------------------------*/
1047 static void defer_bio(struct cache *cache, struct bio *bio)
1049 unsigned long flags;
1051 spin_lock_irqsave(&cache->lock, flags);
1052 bio_list_add(&cache->deferred_bios, bio);
1053 spin_unlock_irqrestore(&cache->lock, flags);
1058 static void process_flush_bio(struct cache *cache, struct bio *bio)
1060 size_t pb_data_size = get_per_bio_data_size(cache);
1061 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1063 BUG_ON(bio->bi_size);
1065 remap_to_origin(cache, bio);
1067 remap_to_cache(cache, bio, 0);
1073 * People generally discard large parts of a device, eg, the whole device
1074 * when formatting. Splitting these large discards up into cache block
1075 * sized ios and then quiescing (always neccessary for discard) takes too
1078 * We keep it simple, and allow any size of discard to come in, and just
1079 * mark off blocks on the discard bitset. No passdown occurs!
1081 * To implement passdown we need to change the bio_prison such that a cell
1082 * can have a key that spans many blocks.
1084 static void process_discard_bio(struct cache *cache, struct bio *bio)
1086 dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
1087 cache->discard_block_size);
1088 dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
1091 end_block = block_div(end_block, cache->discard_block_size);
1093 for (b = start_block; b < end_block; b++)
1094 set_discard(cache, to_dblock(b));
1099 static bool spare_migration_bandwidth(struct cache *cache)
1101 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1102 cache->sectors_per_block;
1103 return current_volume < cache->migration_threshold;
1106 static bool is_writethrough_io(struct cache *cache, struct bio *bio,
1109 return bio_data_dir(bio) == WRITE &&
1110 cache->features.write_through && !is_dirty(cache, cblock);
1113 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1115 atomic_inc(bio_data_dir(bio) == READ ?
1116 &cache->stats.read_hit : &cache->stats.write_hit);
1119 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1121 atomic_inc(bio_data_dir(bio) == READ ?
1122 &cache->stats.read_miss : &cache->stats.write_miss);
1125 static void process_bio(struct cache *cache, struct prealloc *structs,
1129 bool release_cell = true;
1130 dm_oblock_t block = get_bio_block(cache, bio);
1131 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1132 struct policy_result lookup_result;
1133 size_t pb_data_size = get_per_bio_data_size(cache);
1134 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1135 bool discarded_block = is_discarded_oblock(cache, block);
1136 bool can_migrate = discarded_block || spare_migration_bandwidth(cache);
1139 * Check to see if that block is currently migrating.
1141 cell_prealloc = prealloc_get_cell(structs);
1142 r = bio_detain(cache, block, bio, cell_prealloc,
1143 (cell_free_fn) prealloc_put_cell,
1144 structs, &new_ocell);
1148 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1149 bio, &lookup_result);
1151 if (r == -EWOULDBLOCK)
1152 /* migration has been denied */
1153 lookup_result.op = POLICY_MISS;
1155 switch (lookup_result.op) {
1157 inc_hit_counter(cache, bio);
1158 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1160 if (is_writethrough_io(cache, bio, lookup_result.cblock))
1161 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1163 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
1169 inc_miss_counter(cache, bio);
1170 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1171 remap_to_origin_clear_discard(cache, bio, block);
1176 atomic_inc(&cache->stats.promotion);
1177 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1178 release_cell = false;
1181 case POLICY_REPLACE:
1182 cell_prealloc = prealloc_get_cell(structs);
1183 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1184 (cell_free_fn) prealloc_put_cell,
1185 structs, &old_ocell);
1188 * We have to be careful to avoid lock inversion of
1189 * the cells. So we back off, and wait for the
1190 * old_ocell to become free.
1192 policy_force_mapping(cache->policy, block,
1193 lookup_result.old_oblock);
1194 atomic_inc(&cache->stats.cache_cell_clash);
1197 atomic_inc(&cache->stats.demotion);
1198 atomic_inc(&cache->stats.promotion);
1200 demote_then_promote(cache, structs, lookup_result.old_oblock,
1201 block, lookup_result.cblock,
1202 old_ocell, new_ocell);
1203 release_cell = false;
1207 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1208 (unsigned) lookup_result.op);
1213 cell_defer(cache, new_ocell, false);
1216 static int need_commit_due_to_time(struct cache *cache)
1218 return jiffies < cache->last_commit_jiffies ||
1219 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1222 static int commit_if_needed(struct cache *cache)
1224 if (dm_cache_changed_this_transaction(cache->cmd) &&
1225 (cache->commit_requested || need_commit_due_to_time(cache))) {
1226 atomic_inc(&cache->stats.commit_count);
1227 cache->last_commit_jiffies = jiffies;
1228 cache->commit_requested = false;
1229 return dm_cache_commit(cache->cmd, false);
1235 static void process_deferred_bios(struct cache *cache)
1237 unsigned long flags;
1238 struct bio_list bios;
1240 struct prealloc structs;
1242 memset(&structs, 0, sizeof(structs));
1243 bio_list_init(&bios);
1245 spin_lock_irqsave(&cache->lock, flags);
1246 bio_list_merge(&bios, &cache->deferred_bios);
1247 bio_list_init(&cache->deferred_bios);
1248 spin_unlock_irqrestore(&cache->lock, flags);
1250 while (!bio_list_empty(&bios)) {
1252 * If we've got no free migration structs, and processing
1253 * this bio might require one, we pause until there are some
1254 * prepared mappings to process.
1256 if (prealloc_data_structs(cache, &structs)) {
1257 spin_lock_irqsave(&cache->lock, flags);
1258 bio_list_merge(&cache->deferred_bios, &bios);
1259 spin_unlock_irqrestore(&cache->lock, flags);
1263 bio = bio_list_pop(&bios);
1265 if (bio->bi_rw & REQ_FLUSH)
1266 process_flush_bio(cache, bio);
1267 else if (bio->bi_rw & REQ_DISCARD)
1268 process_discard_bio(cache, bio);
1270 process_bio(cache, &structs, bio);
1273 prealloc_free_structs(cache, &structs);
1276 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1278 unsigned long flags;
1279 struct bio_list bios;
1282 bio_list_init(&bios);
1284 spin_lock_irqsave(&cache->lock, flags);
1285 bio_list_merge(&bios, &cache->deferred_flush_bios);
1286 bio_list_init(&cache->deferred_flush_bios);
1287 spin_unlock_irqrestore(&cache->lock, flags);
1289 while ((bio = bio_list_pop(&bios)))
1290 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1293 static void process_deferred_writethrough_bios(struct cache *cache)
1295 unsigned long flags;
1296 struct bio_list bios;
1299 bio_list_init(&bios);
1301 spin_lock_irqsave(&cache->lock, flags);
1302 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1303 bio_list_init(&cache->deferred_writethrough_bios);
1304 spin_unlock_irqrestore(&cache->lock, flags);
1306 while ((bio = bio_list_pop(&bios)))
1307 generic_make_request(bio);
1310 static void writeback_some_dirty_blocks(struct cache *cache)
1315 struct prealloc structs;
1316 struct dm_bio_prison_cell *old_ocell;
1318 memset(&structs, 0, sizeof(structs));
1320 while (spare_migration_bandwidth(cache)) {
1321 if (prealloc_data_structs(cache, &structs))
1324 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1328 r = get_cell(cache, oblock, &structs, &old_ocell);
1330 policy_set_dirty(cache->policy, oblock);
1334 writeback(cache, &structs, oblock, cblock, old_ocell);
1337 prealloc_free_structs(cache, &structs);
1340 /*----------------------------------------------------------------
1342 *--------------------------------------------------------------*/
1343 static void start_quiescing(struct cache *cache)
1345 unsigned long flags;
1347 spin_lock_irqsave(&cache->lock, flags);
1348 cache->quiescing = 1;
1349 spin_unlock_irqrestore(&cache->lock, flags);
1352 static void stop_quiescing(struct cache *cache)
1354 unsigned long flags;
1356 spin_lock_irqsave(&cache->lock, flags);
1357 cache->quiescing = 0;
1358 spin_unlock_irqrestore(&cache->lock, flags);
1361 static bool is_quiescing(struct cache *cache)
1364 unsigned long flags;
1366 spin_lock_irqsave(&cache->lock, flags);
1367 r = cache->quiescing;
1368 spin_unlock_irqrestore(&cache->lock, flags);
1373 static void wait_for_migrations(struct cache *cache)
1375 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1378 static void stop_worker(struct cache *cache)
1380 cancel_delayed_work(&cache->waker);
1381 flush_workqueue(cache->wq);
1384 static void requeue_deferred_io(struct cache *cache)
1387 struct bio_list bios;
1389 bio_list_init(&bios);
1390 bio_list_merge(&bios, &cache->deferred_bios);
1391 bio_list_init(&cache->deferred_bios);
1393 while ((bio = bio_list_pop(&bios)))
1394 bio_endio(bio, DM_ENDIO_REQUEUE);
1397 static int more_work(struct cache *cache)
1399 if (is_quiescing(cache))
1400 return !list_empty(&cache->quiesced_migrations) ||
1401 !list_empty(&cache->completed_migrations) ||
1402 !list_empty(&cache->need_commit_migrations);
1404 return !bio_list_empty(&cache->deferred_bios) ||
1405 !bio_list_empty(&cache->deferred_flush_bios) ||
1406 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1407 !list_empty(&cache->quiesced_migrations) ||
1408 !list_empty(&cache->completed_migrations) ||
1409 !list_empty(&cache->need_commit_migrations);
1412 static void do_worker(struct work_struct *ws)
1414 struct cache *cache = container_of(ws, struct cache, worker);
1417 if (!is_quiescing(cache))
1418 process_deferred_bios(cache);
1420 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1421 process_migrations(cache, &cache->completed_migrations, complete_migration);
1423 writeback_some_dirty_blocks(cache);
1425 process_deferred_writethrough_bios(cache);
1427 if (commit_if_needed(cache)) {
1428 process_deferred_flush_bios(cache, false);
1431 * FIXME: rollback metadata or just go into a
1432 * failure mode and error everything
1435 process_deferred_flush_bios(cache, true);
1436 process_migrations(cache, &cache->need_commit_migrations,
1437 migration_success_post_commit);
1439 } while (more_work(cache));
1443 * We want to commit periodically so that not too much
1444 * unwritten metadata builds up.
1446 static void do_waker(struct work_struct *ws)
1448 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1449 policy_tick(cache->policy);
1451 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1454 /*----------------------------------------------------------------*/
1456 static int is_congested(struct dm_dev *dev, int bdi_bits)
1458 struct request_queue *q = bdev_get_queue(dev->bdev);
1459 return bdi_congested(&q->backing_dev_info, bdi_bits);
1462 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1464 struct cache *cache = container_of(cb, struct cache, callbacks);
1466 return is_congested(cache->origin_dev, bdi_bits) ||
1467 is_congested(cache->cache_dev, bdi_bits);
1470 /*----------------------------------------------------------------
1472 *--------------------------------------------------------------*/
1475 * This function gets called on the error paths of the constructor, so we
1476 * have to cope with a partially initialised struct.
1478 static void destroy(struct cache *cache)
1482 if (cache->next_migration)
1483 mempool_free(cache->next_migration, cache->migration_pool);
1485 if (cache->migration_pool)
1486 mempool_destroy(cache->migration_pool);
1488 if (cache->all_io_ds)
1489 dm_deferred_set_destroy(cache->all_io_ds);
1492 dm_bio_prison_destroy(cache->prison);
1495 destroy_workqueue(cache->wq);
1497 if (cache->dirty_bitset)
1498 free_bitset(cache->dirty_bitset);
1500 if (cache->discard_bitset)
1501 free_bitset(cache->discard_bitset);
1504 dm_kcopyd_client_destroy(cache->copier);
1507 dm_cache_metadata_close(cache->cmd);
1509 if (cache->metadata_dev)
1510 dm_put_device(cache->ti, cache->metadata_dev);
1512 if (cache->origin_dev)
1513 dm_put_device(cache->ti, cache->origin_dev);
1515 if (cache->cache_dev)
1516 dm_put_device(cache->ti, cache->cache_dev);
1519 dm_cache_policy_destroy(cache->policy);
1521 for (i = 0; i < cache->nr_ctr_args ; i++)
1522 kfree(cache->ctr_args[i]);
1523 kfree(cache->ctr_args);
1528 static void cache_dtr(struct dm_target *ti)
1530 struct cache *cache = ti->private;
1535 static sector_t get_dev_size(struct dm_dev *dev)
1537 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1540 /*----------------------------------------------------------------*/
1543 * Construct a cache device mapping.
1545 * cache <metadata dev> <cache dev> <origin dev> <block size>
1546 * <#feature args> [<feature arg>]*
1547 * <policy> <#policy args> [<policy arg>]*
1549 * metadata dev : fast device holding the persistent metadata
1550 * cache dev : fast device holding cached data blocks
1551 * origin dev : slow device holding original data blocks
1552 * block size : cache unit size in sectors
1554 * #feature args : number of feature arguments passed
1555 * feature args : writethrough. (The default is writeback.)
1557 * policy : the replacement policy to use
1558 * #policy args : an even number of policy arguments corresponding
1559 * to key/value pairs passed to the policy
1560 * policy args : key/value pairs passed to the policy
1561 * E.g. 'sequential_threshold 1024'
1562 * See cache-policies.txt for details.
1564 * Optional feature arguments are:
1565 * writethrough : write through caching that prohibits cache block
1566 * content from being different from origin block content.
1567 * Without this argument, the default behaviour is to write
1568 * back cache block contents later for performance reasons,
1569 * so they may differ from the corresponding origin blocks.
1572 struct dm_target *ti;
1574 struct dm_dev *metadata_dev;
1576 struct dm_dev *cache_dev;
1577 sector_t cache_sectors;
1579 struct dm_dev *origin_dev;
1580 sector_t origin_sectors;
1582 uint32_t block_size;
1584 const char *policy_name;
1586 const char **policy_argv;
1588 struct cache_features features;
1591 static void destroy_cache_args(struct cache_args *ca)
1593 if (ca->metadata_dev)
1594 dm_put_device(ca->ti, ca->metadata_dev);
1597 dm_put_device(ca->ti, ca->cache_dev);
1600 dm_put_device(ca->ti, ca->origin_dev);
1605 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1608 *error = "Insufficient args";
1615 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1619 sector_t metadata_dev_size;
1620 char b[BDEVNAME_SIZE];
1622 if (!at_least_one_arg(as, error))
1625 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1628 *error = "Error opening metadata device";
1632 metadata_dev_size = get_dev_size(ca->metadata_dev);
1633 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1634 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1635 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1640 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1645 if (!at_least_one_arg(as, error))
1648 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1651 *error = "Error opening cache device";
1654 ca->cache_sectors = get_dev_size(ca->cache_dev);
1659 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1664 if (!at_least_one_arg(as, error))
1667 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1670 *error = "Error opening origin device";
1674 ca->origin_sectors = get_dev_size(ca->origin_dev);
1675 if (ca->ti->len > ca->origin_sectors) {
1676 *error = "Device size larger than cached device";
1683 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1688 if (!at_least_one_arg(as, error))
1691 if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
1692 tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1693 tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1694 *error = "Invalid data block size";
1698 if (tmp > ca->cache_sectors) {
1699 *error = "Data block size is larger than the cache device";
1703 ca->block_size = tmp;
1708 static void init_features(struct cache_features *cf)
1710 cf->mode = CM_WRITE;
1711 cf->write_through = false;
1714 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1717 static struct dm_arg _args[] = {
1718 {0, 1, "Invalid number of cache feature arguments"},
1724 struct cache_features *cf = &ca->features;
1728 r = dm_read_arg_group(_args, as, &argc, error);
1733 arg = dm_shift_arg(as);
1735 if (!strcasecmp(arg, "writeback"))
1736 cf->write_through = false;
1738 else if (!strcasecmp(arg, "writethrough"))
1739 cf->write_through = true;
1742 *error = "Unrecognised cache feature requested";
1750 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
1753 static struct dm_arg _args[] = {
1754 {0, 1024, "Invalid number of policy arguments"},
1759 if (!at_least_one_arg(as, error))
1762 ca->policy_name = dm_shift_arg(as);
1764 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
1768 ca->policy_argv = (const char **)as->argv;
1769 dm_consume_args(as, ca->policy_argc);
1774 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
1778 struct dm_arg_set as;
1783 r = parse_metadata_dev(ca, &as, error);
1787 r = parse_cache_dev(ca, &as, error);
1791 r = parse_origin_dev(ca, &as, error);
1795 r = parse_block_size(ca, &as, error);
1799 r = parse_features(ca, &as, error);
1803 r = parse_policy(ca, &as, error);
1810 /*----------------------------------------------------------------*/
1812 static struct kmem_cache *migration_cache;
1814 #define NOT_CORE_OPTION 1
1816 static int process_config_option(struct cache *cache, const char *key, const char *value)
1820 if (!strcasecmp(key, "migration_threshold")) {
1821 if (kstrtoul(value, 10, &tmp))
1824 cache->migration_threshold = tmp;
1828 return NOT_CORE_OPTION;
1831 static int set_config_value(struct cache *cache, const char *key, const char *value)
1833 int r = process_config_option(cache, key, value);
1835 if (r == NOT_CORE_OPTION)
1836 r = policy_set_config_value(cache->policy, key, value);
1839 DMWARN("bad config value for %s: %s", key, value);
1844 static int set_config_values(struct cache *cache, int argc, const char **argv)
1849 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
1854 r = set_config_value(cache, argv[0], argv[1]);
1865 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
1868 cache->policy = dm_cache_policy_create(ca->policy_name,
1870 cache->origin_sectors,
1871 cache->sectors_per_block);
1872 if (!cache->policy) {
1873 *error = "Error creating cache's policy";
1881 * We want the discard block size to be a power of two, at least the size
1882 * of the cache block size, and have no more than 2^14 discard blocks
1883 * across the origin.
1885 #define MAX_DISCARD_BLOCKS (1 << 14)
1887 static bool too_many_discard_blocks(sector_t discard_block_size,
1888 sector_t origin_size)
1890 (void) sector_div(origin_size, discard_block_size);
1892 return origin_size > MAX_DISCARD_BLOCKS;
1895 static sector_t calculate_discard_block_size(sector_t cache_block_size,
1896 sector_t origin_size)
1898 sector_t discard_block_size;
1900 discard_block_size = roundup_pow_of_two(cache_block_size);
1903 while (too_many_discard_blocks(discard_block_size, origin_size))
1904 discard_block_size *= 2;
1906 return discard_block_size;
1909 #define DEFAULT_MIGRATION_THRESHOLD 2048
1911 static int cache_create(struct cache_args *ca, struct cache **result)
1914 char **error = &ca->ti->error;
1915 struct cache *cache;
1916 struct dm_target *ti = ca->ti;
1917 dm_block_t origin_blocks;
1918 struct dm_cache_metadata *cmd;
1919 bool may_format = ca->features.mode == CM_WRITE;
1921 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
1926 ti->private = cache;
1927 ti->num_flush_bios = 2;
1928 ti->flush_supported = true;
1930 ti->num_discard_bios = 1;
1931 ti->discards_supported = true;
1932 ti->discard_zeroes_data_unsupported = true;
1934 cache->features = ca->features;
1935 ti->per_bio_data_size = get_per_bio_data_size(cache);
1937 cache->callbacks.congested_fn = cache_is_congested;
1938 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
1940 cache->metadata_dev = ca->metadata_dev;
1941 cache->origin_dev = ca->origin_dev;
1942 cache->cache_dev = ca->cache_dev;
1944 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
1946 /* FIXME: factor out this whole section */
1947 origin_blocks = cache->origin_sectors = ca->origin_sectors;
1948 origin_blocks = block_div(origin_blocks, ca->block_size);
1949 cache->origin_blocks = to_oblock(origin_blocks);
1951 cache->sectors_per_block = ca->block_size;
1952 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
1957 if (ca->block_size & (ca->block_size - 1)) {
1958 dm_block_t cache_size = ca->cache_sectors;
1960 cache->sectors_per_block_shift = -1;
1961 cache_size = block_div(cache_size, ca->block_size);
1962 cache->cache_size = to_cblock(cache_size);
1964 cache->sectors_per_block_shift = __ffs(ca->block_size);
1965 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
1968 r = create_cache_policy(cache, ca, error);
1972 cache->policy_nr_args = ca->policy_argc;
1973 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
1975 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
1977 *error = "Error setting cache policy's config values";
1981 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
1982 ca->block_size, may_format,
1983 dm_cache_policy_get_hint_size(cache->policy));
1985 *error = "Error creating metadata object";
1991 spin_lock_init(&cache->lock);
1992 bio_list_init(&cache->deferred_bios);
1993 bio_list_init(&cache->deferred_flush_bios);
1994 bio_list_init(&cache->deferred_writethrough_bios);
1995 INIT_LIST_HEAD(&cache->quiesced_migrations);
1996 INIT_LIST_HEAD(&cache->completed_migrations);
1997 INIT_LIST_HEAD(&cache->need_commit_migrations);
1998 atomic_set(&cache->nr_migrations, 0);
1999 init_waitqueue_head(&cache->migration_wait);
2002 cache->nr_dirty = 0;
2003 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2004 if (!cache->dirty_bitset) {
2005 *error = "could not allocate dirty bitset";
2008 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2010 cache->discard_block_size =
2011 calculate_discard_block_size(cache->sectors_per_block,
2012 cache->origin_sectors);
2013 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2014 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2015 if (!cache->discard_bitset) {
2016 *error = "could not allocate discard bitset";
2019 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2021 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2022 if (IS_ERR(cache->copier)) {
2023 *error = "could not create kcopyd client";
2024 r = PTR_ERR(cache->copier);
2028 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2030 *error = "could not create workqueue for metadata object";
2033 INIT_WORK(&cache->worker, do_worker);
2034 INIT_DELAYED_WORK(&cache->waker, do_waker);
2035 cache->last_commit_jiffies = jiffies;
2037 cache->prison = dm_bio_prison_create(PRISON_CELLS);
2038 if (!cache->prison) {
2039 *error = "could not create bio prison";
2043 cache->all_io_ds = dm_deferred_set_create();
2044 if (!cache->all_io_ds) {
2045 *error = "could not create all_io deferred set";
2049 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2051 if (!cache->migration_pool) {
2052 *error = "Error creating cache's migration mempool";
2056 cache->next_migration = NULL;
2058 cache->need_tick_bio = true;
2059 cache->sized = false;
2060 cache->quiescing = false;
2061 cache->commit_requested = false;
2062 cache->loaded_mappings = false;
2063 cache->loaded_discards = false;
2067 atomic_set(&cache->stats.demotion, 0);
2068 atomic_set(&cache->stats.promotion, 0);
2069 atomic_set(&cache->stats.copies_avoided, 0);
2070 atomic_set(&cache->stats.cache_cell_clash, 0);
2071 atomic_set(&cache->stats.commit_count, 0);
2072 atomic_set(&cache->stats.discard_count, 0);
2082 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2087 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2090 for (i = 0; i < argc; i++) {
2091 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2100 cache->nr_ctr_args = argc;
2101 cache->ctr_args = copy;
2106 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2109 struct cache_args *ca;
2110 struct cache *cache = NULL;
2112 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2114 ti->error = "Error allocating memory for cache";
2119 r = parse_cache_args(ca, argc, argv, &ti->error);
2123 r = cache_create(ca, &cache);
2127 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2133 ti->private = cache;
2136 destroy_cache_args(ca);
2140 static int cache_map(struct dm_target *ti, struct bio *bio)
2142 struct cache *cache = ti->private;
2145 dm_oblock_t block = get_bio_block(cache, bio);
2146 size_t pb_data_size = get_per_bio_data_size(cache);
2147 bool can_migrate = false;
2148 bool discarded_block;
2149 struct dm_bio_prison_cell *cell;
2150 struct policy_result lookup_result;
2151 struct per_bio_data *pb;
2153 if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2155 * This can only occur if the io goes to a partial block at
2156 * the end of the origin device. We don't cache these.
2157 * Just remap to the origin and carry on.
2159 remap_to_origin_clear_discard(cache, bio, block);
2160 return DM_MAPIO_REMAPPED;
2163 pb = init_per_bio_data(bio, pb_data_size);
2165 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2166 defer_bio(cache, bio);
2167 return DM_MAPIO_SUBMITTED;
2171 * Check to see if that block is currently migrating.
2173 cell = alloc_prison_cell(cache);
2175 defer_bio(cache, bio);
2176 return DM_MAPIO_SUBMITTED;
2179 r = bio_detain(cache, block, bio, cell,
2180 (cell_free_fn) free_prison_cell,
2184 defer_bio(cache, bio);
2186 return DM_MAPIO_SUBMITTED;
2189 discarded_block = is_discarded_oblock(cache, block);
2191 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2192 bio, &lookup_result);
2193 if (r == -EWOULDBLOCK) {
2194 cell_defer(cache, cell, true);
2195 return DM_MAPIO_SUBMITTED;
2198 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2200 return DM_MAPIO_SUBMITTED;
2203 switch (lookup_result.op) {
2205 inc_hit_counter(cache, bio);
2206 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2208 if (is_writethrough_io(cache, bio, lookup_result.cblock))
2209 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2211 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2213 cell_defer(cache, cell, false);
2217 inc_miss_counter(cache, bio);
2218 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2220 if (pb->req_nr != 0) {
2222 * This is a duplicate writethrough io that is no
2223 * longer needed because the block has been demoted.
2226 cell_defer(cache, cell, false);
2227 return DM_MAPIO_SUBMITTED;
2229 remap_to_origin_clear_discard(cache, bio, block);
2230 cell_defer(cache, cell, false);
2235 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2236 (unsigned) lookup_result.op);
2238 return DM_MAPIO_SUBMITTED;
2241 return DM_MAPIO_REMAPPED;
2244 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2246 struct cache *cache = ti->private;
2247 unsigned long flags;
2248 size_t pb_data_size = get_per_bio_data_size(cache);
2249 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2252 policy_tick(cache->policy);
2254 spin_lock_irqsave(&cache->lock, flags);
2255 cache->need_tick_bio = true;
2256 spin_unlock_irqrestore(&cache->lock, flags);
2259 check_for_quiesced_migrations(cache, pb);
2264 static int write_dirty_bitset(struct cache *cache)
2268 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2269 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2270 is_dirty(cache, to_cblock(i)));
2278 static int write_discard_bitset(struct cache *cache)
2282 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2283 cache->discard_nr_blocks);
2285 DMERR("could not resize on-disk discard bitset");
2289 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2290 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2291 is_discarded(cache, to_dblock(i)));
2299 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2302 struct cache *cache = context;
2303 return dm_cache_save_hint(cache->cmd, cblock, hint);
2306 static int write_hints(struct cache *cache)
2310 r = dm_cache_begin_hints(cache->cmd, cache->policy);
2312 DMERR("dm_cache_begin_hints failed");
2316 r = policy_walk_mappings(cache->policy, save_hint, cache);
2318 DMERR("policy_walk_mappings failed");
2324 * returns true on success
2326 static bool sync_metadata(struct cache *cache)
2330 r1 = write_dirty_bitset(cache);
2332 DMERR("could not write dirty bitset");
2334 r2 = write_discard_bitset(cache);
2336 DMERR("could not write discard bitset");
2340 r3 = write_hints(cache);
2342 DMERR("could not write hints");
2345 * If writing the above metadata failed, we still commit, but don't
2346 * set the clean shutdown flag. This will effectively force every
2347 * dirty bit to be set on reload.
2349 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2351 DMERR("could not write cache metadata. Data loss may occur.");
2353 return !r1 && !r2 && !r3 && !r4;
2356 static void cache_postsuspend(struct dm_target *ti)
2358 struct cache *cache = ti->private;
2360 start_quiescing(cache);
2361 wait_for_migrations(cache);
2363 requeue_deferred_io(cache);
2364 stop_quiescing(cache);
2366 (void) sync_metadata(cache);
2369 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2370 bool dirty, uint32_t hint, bool hint_valid)
2373 struct cache *cache = context;
2375 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2380 set_dirty(cache, oblock, cblock);
2382 clear_dirty(cache, oblock, cblock);
2387 static int load_discard(void *context, sector_t discard_block_size,
2388 dm_dblock_t dblock, bool discard)
2390 struct cache *cache = context;
2392 /* FIXME: handle mis-matched block size */
2395 set_discard(cache, dblock);
2397 clear_discard(cache, dblock);
2402 static int cache_preresume(struct dm_target *ti)
2405 struct cache *cache = ti->private;
2406 sector_t actual_cache_size = get_dev_size(cache->cache_dev);
2407 (void) sector_div(actual_cache_size, cache->sectors_per_block);
2410 * Check to see if the cache has resized.
2412 if (from_cblock(cache->cache_size) != actual_cache_size || !cache->sized) {
2413 cache->cache_size = to_cblock(actual_cache_size);
2415 r = dm_cache_resize(cache->cmd, cache->cache_size);
2417 DMERR("could not resize cache metadata");
2421 cache->sized = true;
2424 if (!cache->loaded_mappings) {
2425 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2426 load_mapping, cache);
2428 DMERR("could not load cache mappings");
2432 cache->loaded_mappings = true;
2435 if (!cache->loaded_discards) {
2436 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2438 DMERR("could not load origin discards");
2442 cache->loaded_discards = true;
2448 static void cache_resume(struct dm_target *ti)
2450 struct cache *cache = ti->private;
2452 cache->need_tick_bio = true;
2453 do_waker(&cache->waker.work);
2459 * <#used metadata blocks>/<#total metadata blocks>
2460 * <#read hits> <#read misses> <#write hits> <#write misses>
2461 * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2462 * <#features> <features>*
2463 * <#core args> <core args>
2464 * <#policy args> <policy args>*
2466 static void cache_status(struct dm_target *ti, status_type_t type,
2467 unsigned status_flags, char *result, unsigned maxlen)
2472 dm_block_t nr_free_blocks_metadata = 0;
2473 dm_block_t nr_blocks_metadata = 0;
2474 char buf[BDEVNAME_SIZE];
2475 struct cache *cache = ti->private;
2476 dm_cblock_t residency;
2479 case STATUSTYPE_INFO:
2480 /* Commit to ensure statistics aren't out-of-date */
2481 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2482 r = dm_cache_commit(cache->cmd, false);
2484 DMERR("could not commit metadata for accurate status");
2487 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2488 &nr_free_blocks_metadata);
2490 DMERR("could not get metadata free block count");
2494 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2496 DMERR("could not get metadata device size");
2500 residency = policy_residency(cache->policy);
2502 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2503 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2504 (unsigned long long)nr_blocks_metadata,
2505 (unsigned) atomic_read(&cache->stats.read_hit),
2506 (unsigned) atomic_read(&cache->stats.read_miss),
2507 (unsigned) atomic_read(&cache->stats.write_hit),
2508 (unsigned) atomic_read(&cache->stats.write_miss),
2509 (unsigned) atomic_read(&cache->stats.demotion),
2510 (unsigned) atomic_read(&cache->stats.promotion),
2511 (unsigned long long) from_cblock(residency),
2514 if (cache->features.write_through)
2515 DMEMIT("1 writethrough ");
2519 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2521 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2523 DMERR("policy_emit_config_values returned %d", r);
2528 case STATUSTYPE_TABLE:
2529 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2531 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2533 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2536 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2537 DMEMIT(" %s", cache->ctr_args[i]);
2538 if (cache->nr_ctr_args)
2539 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2549 * Supports <key> <value>.
2551 * The key migration_threshold is supported by the cache target core.
2553 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2555 struct cache *cache = ti->private;
2560 return set_config_value(cache, argv[0], argv[1]);
2563 static int cache_iterate_devices(struct dm_target *ti,
2564 iterate_devices_callout_fn fn, void *data)
2567 struct cache *cache = ti->private;
2569 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
2571 r = fn(ti, cache->origin_dev, 0, ti->len, data);
2577 * We assume I/O is going to the origin (which is the volume
2578 * more likely to have restrictions e.g. by being striped).
2579 * (Looking up the exact location of the data would be expensive
2580 * and could always be out of date by the time the bio is submitted.)
2582 static int cache_bvec_merge(struct dm_target *ti,
2583 struct bvec_merge_data *bvm,
2584 struct bio_vec *biovec, int max_size)
2586 struct cache *cache = ti->private;
2587 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
2589 if (!q->merge_bvec_fn)
2592 bvm->bi_bdev = cache->origin_dev->bdev;
2593 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2596 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
2599 * FIXME: these limits may be incompatible with the cache device
2601 limits->max_discard_sectors = cache->discard_block_size * 1024;
2602 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
2605 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
2607 struct cache *cache = ti->private;
2609 blk_limits_io_min(limits, 0);
2610 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
2611 set_discard_limits(cache, limits);
2614 /*----------------------------------------------------------------*/
2616 static struct target_type cache_target = {
2618 .version = {1, 1, 1},
2619 .module = THIS_MODULE,
2623 .end_io = cache_end_io,
2624 .postsuspend = cache_postsuspend,
2625 .preresume = cache_preresume,
2626 .resume = cache_resume,
2627 .status = cache_status,
2628 .message = cache_message,
2629 .iterate_devices = cache_iterate_devices,
2630 .merge = cache_bvec_merge,
2631 .io_hints = cache_io_hints,
2634 static int __init dm_cache_init(void)
2638 r = dm_register_target(&cache_target);
2640 DMERR("cache target registration failed: %d", r);
2644 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
2645 if (!migration_cache) {
2646 dm_unregister_target(&cache_target);
2653 static void __exit dm_cache_exit(void)
2655 dm_unregister_target(&cache_target);
2656 kmem_cache_destroy(migration_cache);
2659 module_init(dm_cache_init);
2660 module_exit(dm_cache_exit);
2662 MODULE_DESCRIPTION(DM_NAME " cache target");
2663 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
2664 MODULE_LICENSE("GPL");