2 * Copyright (C) 2012 Red Hat. All rights reserved.
4 * This file is released under the GPL.
8 #include "dm-bio-prison-v2.h"
9 #include "dm-bio-record.h"
10 #include "dm-cache-metadata.h"
11 #include "dm-io-tracker.h"
13 #include <linux/dm-io.h>
14 #include <linux/dm-kcopyd.h>
15 #include <linux/jiffies.h>
16 #include <linux/init.h>
17 #include <linux/mempool.h>
18 #include <linux/module.h>
19 #include <linux/rwsem.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
23 #define DM_MSG_PREFIX "cache"
25 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
26 "A percentage of time allocated for copying to and/or from cache");
28 /*----------------------------------------------------------------*/
33 * oblock: index of an origin block
34 * cblock: index of a cache block
35 * promotion: movement of a block from origin to cache
36 * demotion: movement of a block from cache to origin
37 * migration: movement of a block between the origin and cache device,
41 /*----------------------------------------------------------------*/
44 * Represents a chunk of future work. 'input' allows continuations to pass
45 * values between themselves, typically error values.
48 struct work_struct ws;
52 static inline void init_continuation(struct continuation *k,
53 void (*fn)(struct work_struct *))
55 INIT_WORK(&k->ws, fn);
59 static inline void queue_continuation(struct workqueue_struct *wq,
60 struct continuation *k)
62 queue_work(wq, &k->ws);
65 /*----------------------------------------------------------------*/
68 * The batcher collects together pieces of work that need a particular
69 * operation to occur before they can proceed (typically a commit).
73 * The operation that everyone is waiting for.
75 blk_status_t (*commit_op)(void *context);
79 * This is how bios should be issued once the commit op is complete
80 * (accounted_request).
82 void (*issue_op)(struct bio *bio, void *context);
86 * Queued work gets put on here after commit.
88 struct workqueue_struct *wq;
91 struct list_head work_items;
93 struct work_struct commit_work;
95 bool commit_scheduled;
98 static void __commit(struct work_struct *_ws)
100 struct batcher *b = container_of(_ws, struct batcher, commit_work);
102 struct list_head work_items;
103 struct work_struct *ws, *tmp;
104 struct continuation *k;
106 struct bio_list bios;
108 INIT_LIST_HEAD(&work_items);
109 bio_list_init(&bios);
112 * We have to grab these before the commit_op to avoid a race
115 spin_lock_irq(&b->lock);
116 list_splice_init(&b->work_items, &work_items);
117 bio_list_merge(&bios, &b->bios);
118 bio_list_init(&b->bios);
119 b->commit_scheduled = false;
120 spin_unlock_irq(&b->lock);
122 r = b->commit_op(b->commit_context);
124 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
125 k = container_of(ws, struct continuation, ws);
127 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
128 queue_work(b->wq, ws);
131 while ((bio = bio_list_pop(&bios))) {
136 b->issue_op(bio, b->issue_context);
140 static void batcher_init(struct batcher *b,
141 blk_status_t (*commit_op)(void *),
142 void *commit_context,
143 void (*issue_op)(struct bio *bio, void *),
145 struct workqueue_struct *wq)
147 b->commit_op = commit_op;
148 b->commit_context = commit_context;
149 b->issue_op = issue_op;
150 b->issue_context = issue_context;
153 spin_lock_init(&b->lock);
154 INIT_LIST_HEAD(&b->work_items);
155 bio_list_init(&b->bios);
156 INIT_WORK(&b->commit_work, __commit);
157 b->commit_scheduled = false;
160 static void async_commit(struct batcher *b)
162 queue_work(b->wq, &b->commit_work);
165 static void continue_after_commit(struct batcher *b, struct continuation *k)
167 bool commit_scheduled;
169 spin_lock_irq(&b->lock);
170 commit_scheduled = b->commit_scheduled;
171 list_add_tail(&k->ws.entry, &b->work_items);
172 spin_unlock_irq(&b->lock);
174 if (commit_scheduled)
179 * Bios are errored if commit failed.
181 static void issue_after_commit(struct batcher *b, struct bio *bio)
183 bool commit_scheduled;
185 spin_lock_irq(&b->lock);
186 commit_scheduled = b->commit_scheduled;
187 bio_list_add(&b->bios, bio);
188 spin_unlock_irq(&b->lock);
190 if (commit_scheduled)
195 * Call this if some urgent work is waiting for the commit to complete.
197 static void schedule_commit(struct batcher *b)
201 spin_lock_irq(&b->lock);
202 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
203 b->commit_scheduled = true;
204 spin_unlock_irq(&b->lock);
211 * There are a couple of places where we let a bio run, but want to do some
212 * work before calling its endio function. We do this by temporarily
213 * changing the endio fn.
215 struct dm_hook_info {
216 bio_end_io_t *bi_end_io;
219 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
220 bio_end_io_t *bi_end_io, void *bi_private)
222 h->bi_end_io = bio->bi_end_io;
224 bio->bi_end_io = bi_end_io;
225 bio->bi_private = bi_private;
228 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
230 bio->bi_end_io = h->bi_end_io;
233 /*----------------------------------------------------------------*/
235 #define MIGRATION_POOL_SIZE 128
236 #define COMMIT_PERIOD HZ
237 #define MIGRATION_COUNT_WINDOW 10
240 * The block size of the device holding cache data must be
241 * between 32KB and 1GB.
243 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
244 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
246 enum cache_metadata_mode {
247 CM_WRITE, /* metadata may be changed */
248 CM_READ_ONLY, /* metadata may not be changed */
254 * Data is written to cached blocks only. These blocks are marked
255 * dirty. If you lose the cache device you will lose data.
256 * Potential performance increase for both reads and writes.
261 * Data is written to both cache and origin. Blocks are never
262 * dirty. Potential performance benfit for reads only.
267 * A degraded mode useful for various cache coherency situations
268 * (eg, rolling back snapshots). Reads and writes always go to the
269 * origin. If a write goes to a cached oblock, then the cache
270 * block is invalidated.
275 struct cache_features {
276 enum cache_metadata_mode mode;
277 enum cache_io_mode io_mode;
278 unsigned metadata_version;
279 bool discard_passdown:1;
290 atomic_t copies_avoided;
291 atomic_t cache_cell_clash;
292 atomic_t commit_count;
293 atomic_t discard_count;
297 struct dm_target *ti;
301 * Fields for converting from sectors to blocks.
303 int sectors_per_block_shift;
304 sector_t sectors_per_block;
306 struct dm_cache_metadata *cmd;
309 * Metadata is written to this device.
311 struct dm_dev *metadata_dev;
314 * The slower of the two data devices. Typically a spindle.
316 struct dm_dev *origin_dev;
319 * The faster of the two data devices. Typically an SSD.
321 struct dm_dev *cache_dev;
324 * Size of the origin device in _complete_ blocks and native sectors.
326 dm_oblock_t origin_blocks;
327 sector_t origin_sectors;
330 * Size of the cache device in blocks.
332 dm_cblock_t cache_size;
335 * Invalidation fields.
337 spinlock_t invalidation_lock;
338 struct list_head invalidation_requests;
340 sector_t migration_threshold;
341 wait_queue_head_t migration_wait;
342 atomic_t nr_allocated_migrations;
345 * The number of in flight migrations that are performing
346 * background io. eg, promotion, writeback.
348 atomic_t nr_io_migrations;
350 struct bio_list deferred_bios;
352 struct rw_semaphore quiesce_lock;
355 * origin_blocks entries, discarded if set.
357 dm_dblock_t discard_nr_blocks;
358 unsigned long *discard_bitset;
359 uint32_t discard_block_size; /* a power of 2 times sectors per block */
362 * Rather than reconstructing the table line for the status we just
363 * save it and regurgitate.
365 unsigned nr_ctr_args;
366 const char **ctr_args;
368 struct dm_kcopyd_client *copier;
369 struct work_struct deferred_bio_worker;
370 struct work_struct migration_worker;
371 struct workqueue_struct *wq;
372 struct delayed_work waker;
373 struct dm_bio_prison_v2 *prison;
376 * cache_size entries, dirty if set
378 unsigned long *dirty_bitset;
381 unsigned policy_nr_args;
382 struct dm_cache_policy *policy;
385 * Cache features such as write-through.
387 struct cache_features features;
389 struct cache_stats stats;
391 bool need_tick_bio:1;
394 bool commit_requested:1;
395 bool loaded_mappings:1;
396 bool loaded_discards:1;
398 struct rw_semaphore background_work_lock;
400 struct batcher committer;
401 struct work_struct commit_ws;
403 struct dm_io_tracker tracker;
405 mempool_t migration_pool;
410 struct per_bio_data {
413 struct dm_bio_prison_cell_v2 *cell;
414 struct dm_hook_info hook_info;
418 struct dm_cache_migration {
419 struct continuation k;
422 struct policy_work *op;
423 struct bio *overwrite_bio;
424 struct dm_bio_prison_cell_v2 *cell;
426 dm_cblock_t invalidate_cblock;
427 dm_oblock_t invalidate_oblock;
430 /*----------------------------------------------------------------*/
432 static bool writethrough_mode(struct cache *cache)
434 return cache->features.io_mode == CM_IO_WRITETHROUGH;
437 static bool writeback_mode(struct cache *cache)
439 return cache->features.io_mode == CM_IO_WRITEBACK;
442 static inline bool passthrough_mode(struct cache *cache)
444 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
447 /*----------------------------------------------------------------*/
449 static void wake_deferred_bio_worker(struct cache *cache)
451 queue_work(cache->wq, &cache->deferred_bio_worker);
454 static void wake_migration_worker(struct cache *cache)
456 if (passthrough_mode(cache))
459 queue_work(cache->wq, &cache->migration_worker);
462 /*----------------------------------------------------------------*/
464 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
466 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
469 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
471 dm_bio_prison_free_cell_v2(cache->prison, cell);
474 static struct dm_cache_migration *alloc_migration(struct cache *cache)
476 struct dm_cache_migration *mg;
478 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
480 memset(mg, 0, sizeof(*mg));
483 atomic_inc(&cache->nr_allocated_migrations);
488 static void free_migration(struct dm_cache_migration *mg)
490 struct cache *cache = mg->cache;
492 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
493 wake_up(&cache->migration_wait);
495 mempool_free(mg, &cache->migration_pool);
498 /*----------------------------------------------------------------*/
500 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
502 return to_oblock(from_oblock(b) + 1ull);
505 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
509 key->block_begin = from_oblock(begin);
510 key->block_end = from_oblock(end);
514 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
515 * level 1 which prevents *both* READs and WRITEs.
517 #define WRITE_LOCK_LEVEL 0
518 #define READ_WRITE_LOCK_LEVEL 1
520 static unsigned lock_level(struct bio *bio)
522 return bio_data_dir(bio) == WRITE ?
524 READ_WRITE_LOCK_LEVEL;
527 /*----------------------------------------------------------------
529 *--------------------------------------------------------------*/
531 static struct per_bio_data *get_per_bio_data(struct bio *bio)
533 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
538 static struct per_bio_data *init_per_bio_data(struct bio *bio)
540 struct per_bio_data *pb = get_per_bio_data(bio);
543 pb->req_nr = dm_bio_get_target_bio_nr(bio);
550 /*----------------------------------------------------------------*/
552 static void defer_bio(struct cache *cache, struct bio *bio)
554 spin_lock_irq(&cache->lock);
555 bio_list_add(&cache->deferred_bios, bio);
556 spin_unlock_irq(&cache->lock);
558 wake_deferred_bio_worker(cache);
561 static void defer_bios(struct cache *cache, struct bio_list *bios)
563 spin_lock_irq(&cache->lock);
564 bio_list_merge(&cache->deferred_bios, bios);
566 spin_unlock_irq(&cache->lock);
568 wake_deferred_bio_worker(cache);
571 /*----------------------------------------------------------------*/
573 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
576 struct per_bio_data *pb;
577 struct dm_cell_key_v2 key;
578 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
579 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
581 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
583 build_key(oblock, end, &key);
584 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
587 * Failed to get the lock.
589 free_prison_cell(cache, cell_prealloc);
593 if (cell != cell_prealloc)
594 free_prison_cell(cache, cell_prealloc);
596 pb = get_per_bio_data(bio);
602 /*----------------------------------------------------------------*/
604 static bool is_dirty(struct cache *cache, dm_cblock_t b)
606 return test_bit(from_cblock(b), cache->dirty_bitset);
609 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
611 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
612 atomic_inc(&cache->nr_dirty);
613 policy_set_dirty(cache->policy, cblock);
618 * These two are called when setting after migrations to force the policy
619 * and dirty bitset to be in sync.
621 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
623 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
624 atomic_inc(&cache->nr_dirty);
625 policy_set_dirty(cache->policy, cblock);
628 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
630 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
631 if (atomic_dec_return(&cache->nr_dirty) == 0)
632 dm_table_event(cache->ti->table);
635 policy_clear_dirty(cache->policy, cblock);
638 /*----------------------------------------------------------------*/
640 static bool block_size_is_power_of_two(struct cache *cache)
642 return cache->sectors_per_block_shift >= 0;
645 static dm_block_t block_div(dm_block_t b, uint32_t n)
652 static dm_block_t oblocks_per_dblock(struct cache *cache)
654 dm_block_t oblocks = cache->discard_block_size;
656 if (block_size_is_power_of_two(cache))
657 oblocks >>= cache->sectors_per_block_shift;
659 oblocks = block_div(oblocks, cache->sectors_per_block);
664 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
666 return to_dblock(block_div(from_oblock(oblock),
667 oblocks_per_dblock(cache)));
670 static void set_discard(struct cache *cache, dm_dblock_t b)
672 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
673 atomic_inc(&cache->stats.discard_count);
675 spin_lock_irq(&cache->lock);
676 set_bit(from_dblock(b), cache->discard_bitset);
677 spin_unlock_irq(&cache->lock);
680 static void clear_discard(struct cache *cache, dm_dblock_t b)
682 spin_lock_irq(&cache->lock);
683 clear_bit(from_dblock(b), cache->discard_bitset);
684 spin_unlock_irq(&cache->lock);
687 static bool is_discarded(struct cache *cache, dm_dblock_t b)
690 spin_lock_irq(&cache->lock);
691 r = test_bit(from_dblock(b), cache->discard_bitset);
692 spin_unlock_irq(&cache->lock);
697 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
700 spin_lock_irq(&cache->lock);
701 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
702 cache->discard_bitset);
703 spin_unlock_irq(&cache->lock);
708 /*----------------------------------------------------------------
710 *--------------------------------------------------------------*/
711 static void remap_to_origin(struct cache *cache, struct bio *bio)
713 bio_set_dev(bio, cache->origin_dev->bdev);
716 static void remap_to_cache(struct cache *cache, struct bio *bio,
719 sector_t bi_sector = bio->bi_iter.bi_sector;
720 sector_t block = from_cblock(cblock);
722 bio_set_dev(bio, cache->cache_dev->bdev);
723 if (!block_size_is_power_of_two(cache))
724 bio->bi_iter.bi_sector =
725 (block * cache->sectors_per_block) +
726 sector_div(bi_sector, cache->sectors_per_block);
728 bio->bi_iter.bi_sector =
729 (block << cache->sectors_per_block_shift) |
730 (bi_sector & (cache->sectors_per_block - 1));
733 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
735 struct per_bio_data *pb;
737 spin_lock_irq(&cache->lock);
738 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
739 bio_op(bio) != REQ_OP_DISCARD) {
740 pb = get_per_bio_data(bio);
742 cache->need_tick_bio = false;
744 spin_unlock_irq(&cache->lock);
747 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
748 dm_oblock_t oblock, bool bio_has_pbd)
751 check_if_tick_bio_needed(cache, bio);
752 remap_to_origin(cache, bio);
753 if (bio_data_dir(bio) == WRITE)
754 clear_discard(cache, oblock_to_dblock(cache, oblock));
757 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
760 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
761 __remap_to_origin_clear_discard(cache, bio, oblock, true);
764 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
765 dm_oblock_t oblock, dm_cblock_t cblock)
767 check_if_tick_bio_needed(cache, bio);
768 remap_to_cache(cache, bio, cblock);
769 if (bio_data_dir(bio) == WRITE) {
770 set_dirty(cache, cblock);
771 clear_discard(cache, oblock_to_dblock(cache, oblock));
775 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
777 sector_t block_nr = bio->bi_iter.bi_sector;
779 if (!block_size_is_power_of_two(cache))
780 (void) sector_div(block_nr, cache->sectors_per_block);
782 block_nr >>= cache->sectors_per_block_shift;
784 return to_oblock(block_nr);
787 static bool accountable_bio(struct cache *cache, struct bio *bio)
789 return bio_op(bio) != REQ_OP_DISCARD;
792 static void accounted_begin(struct cache *cache, struct bio *bio)
794 struct per_bio_data *pb;
796 if (accountable_bio(cache, bio)) {
797 pb = get_per_bio_data(bio);
798 pb->len = bio_sectors(bio);
799 dm_iot_io_begin(&cache->tracker, pb->len);
803 static void accounted_complete(struct cache *cache, struct bio *bio)
805 struct per_bio_data *pb = get_per_bio_data(bio);
807 dm_iot_io_end(&cache->tracker, pb->len);
810 static void accounted_request(struct cache *cache, struct bio *bio)
812 accounted_begin(cache, bio);
813 submit_bio_noacct(bio);
816 static void issue_op(struct bio *bio, void *context)
818 struct cache *cache = context;
819 accounted_request(cache, bio);
823 * When running in writethrough mode we need to send writes to clean blocks
824 * to both the cache and origin devices. Clone the bio and send them in parallel.
826 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
827 dm_oblock_t oblock, dm_cblock_t cblock)
829 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
833 bio_chain(origin_bio, bio);
835 * Passing false to __remap_to_origin_clear_discard() skips
836 * all code that might use per_bio_data (since clone doesn't have it)
838 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
839 submit_bio(origin_bio);
841 remap_to_cache(cache, bio, cblock);
844 /*----------------------------------------------------------------
846 *--------------------------------------------------------------*/
847 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
849 return cache->features.mode;
852 static const char *cache_device_name(struct cache *cache)
854 return dm_table_device_name(cache->ti->table);
857 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
859 const char *descs[] = {
865 dm_table_event(cache->ti->table);
866 DMINFO("%s: switching cache to %s mode",
867 cache_device_name(cache), descs[(int)mode]);
870 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
873 enum cache_metadata_mode old_mode = get_cache_mode(cache);
875 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
876 DMERR("%s: unable to read needs_check flag, setting failure mode.",
877 cache_device_name(cache));
881 if (new_mode == CM_WRITE && needs_check) {
882 DMERR("%s: unable to switch cache to write mode until repaired.",
883 cache_device_name(cache));
884 if (old_mode != new_mode)
887 new_mode = CM_READ_ONLY;
890 /* Never move out of fail mode */
891 if (old_mode == CM_FAIL)
897 dm_cache_metadata_set_read_only(cache->cmd);
901 dm_cache_metadata_set_read_write(cache->cmd);
905 cache->features.mode = new_mode;
907 if (new_mode != old_mode)
908 notify_mode_switch(cache, new_mode);
911 static void abort_transaction(struct cache *cache)
913 const char *dev_name = cache_device_name(cache);
915 if (get_cache_mode(cache) >= CM_READ_ONLY)
918 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
919 if (dm_cache_metadata_abort(cache->cmd)) {
920 DMERR("%s: failed to abort metadata transaction", dev_name);
921 set_cache_mode(cache, CM_FAIL);
924 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
925 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
926 set_cache_mode(cache, CM_FAIL);
930 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
932 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
933 cache_device_name(cache), op, r);
934 abort_transaction(cache);
935 set_cache_mode(cache, CM_READ_ONLY);
938 /*----------------------------------------------------------------*/
940 static void load_stats(struct cache *cache)
942 struct dm_cache_statistics stats;
944 dm_cache_metadata_get_stats(cache->cmd, &stats);
945 atomic_set(&cache->stats.read_hit, stats.read_hits);
946 atomic_set(&cache->stats.read_miss, stats.read_misses);
947 atomic_set(&cache->stats.write_hit, stats.write_hits);
948 atomic_set(&cache->stats.write_miss, stats.write_misses);
951 static void save_stats(struct cache *cache)
953 struct dm_cache_statistics stats;
955 if (get_cache_mode(cache) >= CM_READ_ONLY)
958 stats.read_hits = atomic_read(&cache->stats.read_hit);
959 stats.read_misses = atomic_read(&cache->stats.read_miss);
960 stats.write_hits = atomic_read(&cache->stats.write_hit);
961 stats.write_misses = atomic_read(&cache->stats.write_miss);
963 dm_cache_metadata_set_stats(cache->cmd, &stats);
966 static void update_stats(struct cache_stats *stats, enum policy_operation op)
970 atomic_inc(&stats->promotion);
974 atomic_inc(&stats->demotion);
977 case POLICY_WRITEBACK:
978 atomic_inc(&stats->writeback);
983 /*----------------------------------------------------------------
984 * Migration processing
986 * Migration covers moving data from the origin device to the cache, or
988 *--------------------------------------------------------------*/
990 static void inc_io_migrations(struct cache *cache)
992 atomic_inc(&cache->nr_io_migrations);
995 static void dec_io_migrations(struct cache *cache)
997 atomic_dec(&cache->nr_io_migrations);
1000 static bool discard_or_flush(struct bio *bio)
1002 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1005 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1006 dm_dblock_t *b, dm_dblock_t *e)
1008 sector_t sb = bio->bi_iter.bi_sector;
1009 sector_t se = bio_end_sector(bio);
1011 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1013 if (se - sb < cache->discard_block_size)
1016 *e = to_dblock(block_div(se, cache->discard_block_size));
1019 /*----------------------------------------------------------------*/
1021 static void prevent_background_work(struct cache *cache)
1024 down_write(&cache->background_work_lock);
1028 static void allow_background_work(struct cache *cache)
1031 up_write(&cache->background_work_lock);
1035 static bool background_work_begin(struct cache *cache)
1040 r = down_read_trylock(&cache->background_work_lock);
1046 static void background_work_end(struct cache *cache)
1049 up_read(&cache->background_work_lock);
1053 /*----------------------------------------------------------------*/
1055 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1057 return (bio_data_dir(bio) == WRITE) &&
1058 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1061 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1063 return writeback_mode(cache) &&
1064 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1067 static void quiesce(struct dm_cache_migration *mg,
1068 void (*continuation)(struct work_struct *))
1070 init_continuation(&mg->k, continuation);
1071 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1074 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1076 struct continuation *k = container_of(ws, struct continuation, ws);
1077 return container_of(k, struct dm_cache_migration, k);
1080 static void copy_complete(int read_err, unsigned long write_err, void *context)
1082 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1084 if (read_err || write_err)
1085 mg->k.input = BLK_STS_IOERR;
1087 queue_continuation(mg->cache->wq, &mg->k);
1090 static void copy(struct dm_cache_migration *mg, bool promote)
1092 struct dm_io_region o_region, c_region;
1093 struct cache *cache = mg->cache;
1095 o_region.bdev = cache->origin_dev->bdev;
1096 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1097 o_region.count = cache->sectors_per_block;
1099 c_region.bdev = cache->cache_dev->bdev;
1100 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1101 c_region.count = cache->sectors_per_block;
1104 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1106 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1109 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1111 struct per_bio_data *pb = get_per_bio_data(bio);
1113 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1114 free_prison_cell(cache, pb->cell);
1118 static void overwrite_endio(struct bio *bio)
1120 struct dm_cache_migration *mg = bio->bi_private;
1121 struct cache *cache = mg->cache;
1122 struct per_bio_data *pb = get_per_bio_data(bio);
1124 dm_unhook_bio(&pb->hook_info, bio);
1127 mg->k.input = bio->bi_status;
1129 queue_continuation(cache->wq, &mg->k);
1132 static void overwrite(struct dm_cache_migration *mg,
1133 void (*continuation)(struct work_struct *))
1135 struct bio *bio = mg->overwrite_bio;
1136 struct per_bio_data *pb = get_per_bio_data(bio);
1138 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1141 * The overwrite bio is part of the copy operation, as such it does
1142 * not set/clear discard or dirty flags.
1144 if (mg->op->op == POLICY_PROMOTE)
1145 remap_to_cache(mg->cache, bio, mg->op->cblock);
1147 remap_to_origin(mg->cache, bio);
1149 init_continuation(&mg->k, continuation);
1150 accounted_request(mg->cache, bio);
1156 * 1) exclusive lock preventing WRITEs
1158 * 3) copy or issue overwrite bio
1159 * 4) upgrade to exclusive lock preventing READs and WRITEs
1161 * 6) update metadata and commit
1164 static void mg_complete(struct dm_cache_migration *mg, bool success)
1166 struct bio_list bios;
1167 struct cache *cache = mg->cache;
1168 struct policy_work *op = mg->op;
1169 dm_cblock_t cblock = op->cblock;
1172 update_stats(&cache->stats, op->op);
1175 case POLICY_PROMOTE:
1176 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1177 policy_complete_background_work(cache->policy, op, success);
1179 if (mg->overwrite_bio) {
1181 force_set_dirty(cache, cblock);
1182 else if (mg->k.input)
1183 mg->overwrite_bio->bi_status = mg->k.input;
1185 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1186 bio_endio(mg->overwrite_bio);
1189 force_clear_dirty(cache, cblock);
1190 dec_io_migrations(cache);
1196 * We clear dirty here to update the nr_dirty counter.
1199 force_clear_dirty(cache, cblock);
1200 policy_complete_background_work(cache->policy, op, success);
1201 dec_io_migrations(cache);
1204 case POLICY_WRITEBACK:
1206 force_clear_dirty(cache, cblock);
1207 policy_complete_background_work(cache->policy, op, success);
1208 dec_io_migrations(cache);
1212 bio_list_init(&bios);
1214 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1215 free_prison_cell(cache, mg->cell);
1219 defer_bios(cache, &bios);
1220 wake_migration_worker(cache);
1222 background_work_end(cache);
1225 static void mg_success(struct work_struct *ws)
1227 struct dm_cache_migration *mg = ws_to_mg(ws);
1228 mg_complete(mg, mg->k.input == 0);
1231 static void mg_update_metadata(struct work_struct *ws)
1234 struct dm_cache_migration *mg = ws_to_mg(ws);
1235 struct cache *cache = mg->cache;
1236 struct policy_work *op = mg->op;
1239 case POLICY_PROMOTE:
1240 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1242 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1243 cache_device_name(cache));
1244 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1246 mg_complete(mg, false);
1249 mg_complete(mg, true);
1253 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1255 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1256 cache_device_name(cache));
1257 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1259 mg_complete(mg, false);
1264 * It would be nice if we only had to commit when a REQ_FLUSH
1265 * comes through. But there's one scenario that we have to
1268 * - vblock x in a cache block
1270 * - cache block gets reallocated and over written
1273 * When we recover, because there was no commit the cache will
1274 * rollback to having the data for vblock x in the cache block.
1275 * But the cache block has since been overwritten, so it'll end
1276 * up pointing to data that was never in 'x' during the history
1279 * To avoid this issue we require a commit as part of the
1280 * demotion operation.
1282 init_continuation(&mg->k, mg_success);
1283 continue_after_commit(&cache->committer, &mg->k);
1284 schedule_commit(&cache->committer);
1287 case POLICY_WRITEBACK:
1288 mg_complete(mg, true);
1293 static void mg_update_metadata_after_copy(struct work_struct *ws)
1295 struct dm_cache_migration *mg = ws_to_mg(ws);
1298 * Did the copy succeed?
1301 mg_complete(mg, false);
1303 mg_update_metadata(ws);
1306 static void mg_upgrade_lock(struct work_struct *ws)
1309 struct dm_cache_migration *mg = ws_to_mg(ws);
1312 * Did the copy succeed?
1315 mg_complete(mg, false);
1319 * Now we want the lock to prevent both reads and writes.
1321 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1322 READ_WRITE_LOCK_LEVEL);
1324 mg_complete(mg, false);
1327 quiesce(mg, mg_update_metadata);
1330 mg_update_metadata(ws);
1334 static void mg_full_copy(struct work_struct *ws)
1336 struct dm_cache_migration *mg = ws_to_mg(ws);
1337 struct cache *cache = mg->cache;
1338 struct policy_work *op = mg->op;
1339 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1341 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1342 is_discarded_oblock(cache, op->oblock)) {
1343 mg_upgrade_lock(ws);
1347 init_continuation(&mg->k, mg_upgrade_lock);
1348 copy(mg, is_policy_promote);
1351 static void mg_copy(struct work_struct *ws)
1353 struct dm_cache_migration *mg = ws_to_mg(ws);
1355 if (mg->overwrite_bio) {
1357 * No exclusive lock was held when we last checked if the bio
1358 * was optimisable. So we have to check again in case things
1359 * have changed (eg, the block may no longer be discarded).
1361 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1363 * Fallback to a real full copy after doing some tidying up.
1365 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1366 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1367 mg->overwrite_bio = NULL;
1368 inc_io_migrations(mg->cache);
1374 * It's safe to do this here, even though it's new data
1375 * because all IO has been locked out of the block.
1377 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1378 * so _not_ using mg_upgrade_lock() as continutation.
1380 overwrite(mg, mg_update_metadata_after_copy);
1386 static int mg_lock_writes(struct dm_cache_migration *mg)
1389 struct dm_cell_key_v2 key;
1390 struct cache *cache = mg->cache;
1391 struct dm_bio_prison_cell_v2 *prealloc;
1393 prealloc = alloc_prison_cell(cache);
1396 * Prevent writes to the block, but allow reads to continue.
1397 * Unless we're using an overwrite bio, in which case we lock
1400 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1401 r = dm_cell_lock_v2(cache->prison, &key,
1402 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1403 prealloc, &mg->cell);
1405 free_prison_cell(cache, prealloc);
1406 mg_complete(mg, false);
1410 if (mg->cell != prealloc)
1411 free_prison_cell(cache, prealloc);
1416 quiesce(mg, mg_copy);
1421 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1423 struct dm_cache_migration *mg;
1425 if (!background_work_begin(cache)) {
1426 policy_complete_background_work(cache->policy, op, false);
1430 mg = alloc_migration(cache);
1433 mg->overwrite_bio = bio;
1436 inc_io_migrations(cache);
1438 return mg_lock_writes(mg);
1441 /*----------------------------------------------------------------
1442 * invalidation processing
1443 *--------------------------------------------------------------*/
1445 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1447 struct bio_list bios;
1448 struct cache *cache = mg->cache;
1450 bio_list_init(&bios);
1451 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1452 free_prison_cell(cache, mg->cell);
1454 if (!success && mg->overwrite_bio)
1455 bio_io_error(mg->overwrite_bio);
1458 defer_bios(cache, &bios);
1460 background_work_end(cache);
1463 static void invalidate_completed(struct work_struct *ws)
1465 struct dm_cache_migration *mg = ws_to_mg(ws);
1466 invalidate_complete(mg, !mg->k.input);
1469 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1471 int r = policy_invalidate_mapping(cache->policy, cblock);
1473 r = dm_cache_remove_mapping(cache->cmd, cblock);
1475 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1476 cache_device_name(cache));
1477 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1480 } else if (r == -ENODATA) {
1482 * Harmless, already unmapped.
1487 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1492 static void invalidate_remove(struct work_struct *ws)
1495 struct dm_cache_migration *mg = ws_to_mg(ws);
1496 struct cache *cache = mg->cache;
1498 r = invalidate_cblock(cache, mg->invalidate_cblock);
1500 invalidate_complete(mg, false);
1504 init_continuation(&mg->k, invalidate_completed);
1505 continue_after_commit(&cache->committer, &mg->k);
1506 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1507 mg->overwrite_bio = NULL;
1508 schedule_commit(&cache->committer);
1511 static int invalidate_lock(struct dm_cache_migration *mg)
1514 struct dm_cell_key_v2 key;
1515 struct cache *cache = mg->cache;
1516 struct dm_bio_prison_cell_v2 *prealloc;
1518 prealloc = alloc_prison_cell(cache);
1520 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1521 r = dm_cell_lock_v2(cache->prison, &key,
1522 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1524 free_prison_cell(cache, prealloc);
1525 invalidate_complete(mg, false);
1529 if (mg->cell != prealloc)
1530 free_prison_cell(cache, prealloc);
1533 quiesce(mg, invalidate_remove);
1537 * We can't call invalidate_remove() directly here because we
1538 * might still be in request context.
1540 init_continuation(&mg->k, invalidate_remove);
1541 queue_work(cache->wq, &mg->k.ws);
1547 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1548 dm_oblock_t oblock, struct bio *bio)
1550 struct dm_cache_migration *mg;
1552 if (!background_work_begin(cache))
1555 mg = alloc_migration(cache);
1557 mg->overwrite_bio = bio;
1558 mg->invalidate_cblock = cblock;
1559 mg->invalidate_oblock = oblock;
1561 return invalidate_lock(mg);
1564 /*----------------------------------------------------------------
1566 *--------------------------------------------------------------*/
1573 static enum busy spare_migration_bandwidth(struct cache *cache)
1575 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1576 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1577 cache->sectors_per_block;
1579 if (idle && current_volume <= cache->migration_threshold)
1585 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1587 atomic_inc(bio_data_dir(bio) == READ ?
1588 &cache->stats.read_hit : &cache->stats.write_hit);
1591 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1593 atomic_inc(bio_data_dir(bio) == READ ?
1594 &cache->stats.read_miss : &cache->stats.write_miss);
1597 /*----------------------------------------------------------------*/
1599 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1600 bool *commit_needed)
1603 bool rb, background_queued;
1606 *commit_needed = false;
1608 rb = bio_detain_shared(cache, block, bio);
1611 * An exclusive lock is held for this block, so we have to
1612 * wait. We set the commit_needed flag so the current
1613 * transaction will be committed asap, allowing this lock
1616 *commit_needed = true;
1617 return DM_MAPIO_SUBMITTED;
1620 data_dir = bio_data_dir(bio);
1622 if (optimisable_bio(cache, bio, block)) {
1623 struct policy_work *op = NULL;
1625 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1626 if (unlikely(r && r != -ENOENT)) {
1627 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1628 cache_device_name(cache), r);
1630 return DM_MAPIO_SUBMITTED;
1633 if (r == -ENOENT && op) {
1634 bio_drop_shared_lock(cache, bio);
1635 BUG_ON(op->op != POLICY_PROMOTE);
1636 mg_start(cache, op, bio);
1637 return DM_MAPIO_SUBMITTED;
1640 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1641 if (unlikely(r && r != -ENOENT)) {
1642 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1643 cache_device_name(cache), r);
1645 return DM_MAPIO_SUBMITTED;
1648 if (background_queued)
1649 wake_migration_worker(cache);
1653 struct per_bio_data *pb = get_per_bio_data(bio);
1658 inc_miss_counter(cache, bio);
1659 if (pb->req_nr == 0) {
1660 accounted_begin(cache, bio);
1661 remap_to_origin_clear_discard(cache, bio, block);
1664 * This is a duplicate writethrough io that is no
1665 * longer needed because the block has been demoted.
1668 return DM_MAPIO_SUBMITTED;
1674 inc_hit_counter(cache, bio);
1677 * Passthrough always maps to the origin, invalidating any
1678 * cache blocks that are written to.
1680 if (passthrough_mode(cache)) {
1681 if (bio_data_dir(bio) == WRITE) {
1682 bio_drop_shared_lock(cache, bio);
1683 atomic_inc(&cache->stats.demotion);
1684 invalidate_start(cache, cblock, block, bio);
1686 remap_to_origin_clear_discard(cache, bio, block);
1688 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1689 !is_dirty(cache, cblock)) {
1690 remap_to_origin_and_cache(cache, bio, block, cblock);
1691 accounted_begin(cache, bio);
1693 remap_to_cache_dirty(cache, bio, block, cblock);
1698 * dm core turns FUA requests into a separate payload and FLUSH req.
1700 if (bio->bi_opf & REQ_FUA) {
1702 * issue_after_commit will call accounted_begin a second time. So
1703 * we call accounted_complete() to avoid double accounting.
1705 accounted_complete(cache, bio);
1706 issue_after_commit(&cache->committer, bio);
1707 *commit_needed = true;
1708 return DM_MAPIO_SUBMITTED;
1711 return DM_MAPIO_REMAPPED;
1714 static bool process_bio(struct cache *cache, struct bio *bio)
1718 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1719 submit_bio_noacct(bio);
1721 return commit_needed;
1725 * A non-zero return indicates read_only or fail_io mode.
1727 static int commit(struct cache *cache, bool clean_shutdown)
1731 if (get_cache_mode(cache) >= CM_READ_ONLY)
1734 atomic_inc(&cache->stats.commit_count);
1735 r = dm_cache_commit(cache->cmd, clean_shutdown);
1737 metadata_operation_failed(cache, "dm_cache_commit", r);
1743 * Used by the batcher.
1745 static blk_status_t commit_op(void *context)
1747 struct cache *cache = context;
1749 if (dm_cache_changed_this_transaction(cache->cmd))
1750 return errno_to_blk_status(commit(cache, false));
1755 /*----------------------------------------------------------------*/
1757 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1759 struct per_bio_data *pb = get_per_bio_data(bio);
1762 remap_to_origin(cache, bio);
1764 remap_to_cache(cache, bio, 0);
1766 issue_after_commit(&cache->committer, bio);
1770 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1774 // FIXME: do we need to lock the region? Or can we just assume the
1775 // user wont be so foolish as to issue discard concurrently with
1777 calc_discard_block_range(cache, bio, &b, &e);
1779 set_discard(cache, b);
1780 b = to_dblock(from_dblock(b) + 1);
1783 if (cache->features.discard_passdown) {
1784 remap_to_origin(cache, bio);
1785 submit_bio_noacct(bio);
1792 static void process_deferred_bios(struct work_struct *ws)
1794 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1796 bool commit_needed = false;
1797 struct bio_list bios;
1800 bio_list_init(&bios);
1802 spin_lock_irq(&cache->lock);
1803 bio_list_merge(&bios, &cache->deferred_bios);
1804 bio_list_init(&cache->deferred_bios);
1805 spin_unlock_irq(&cache->lock);
1807 while ((bio = bio_list_pop(&bios))) {
1808 if (bio->bi_opf & REQ_PREFLUSH)
1809 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1811 else if (bio_op(bio) == REQ_OP_DISCARD)
1812 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1815 commit_needed = process_bio(cache, bio) || commit_needed;
1819 schedule_commit(&cache->committer);
1822 /*----------------------------------------------------------------
1824 *--------------------------------------------------------------*/
1826 static void requeue_deferred_bios(struct cache *cache)
1829 struct bio_list bios;
1831 bio_list_init(&bios);
1832 bio_list_merge(&bios, &cache->deferred_bios);
1833 bio_list_init(&cache->deferred_bios);
1835 while ((bio = bio_list_pop(&bios))) {
1836 bio->bi_status = BLK_STS_DM_REQUEUE;
1842 * We want to commit periodically so that not too much
1843 * unwritten metadata builds up.
1845 static void do_waker(struct work_struct *ws)
1847 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1849 policy_tick(cache->policy, true);
1850 wake_migration_worker(cache);
1851 schedule_commit(&cache->committer);
1852 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1855 static void check_migrations(struct work_struct *ws)
1858 struct policy_work *op;
1859 struct cache *cache = container_of(ws, struct cache, migration_worker);
1863 b = spare_migration_bandwidth(cache);
1865 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1870 DMERR_LIMIT("%s: policy_background_work failed",
1871 cache_device_name(cache));
1875 r = mg_start(cache, op, NULL);
1881 /*----------------------------------------------------------------
1883 *--------------------------------------------------------------*/
1886 * This function gets called on the error paths of the constructor, so we
1887 * have to cope with a partially initialised struct.
1889 static void destroy(struct cache *cache)
1893 mempool_exit(&cache->migration_pool);
1896 dm_bio_prison_destroy_v2(cache->prison);
1898 cancel_delayed_work_sync(&cache->waker);
1900 destroy_workqueue(cache->wq);
1902 if (cache->dirty_bitset)
1903 free_bitset(cache->dirty_bitset);
1905 if (cache->discard_bitset)
1906 free_bitset(cache->discard_bitset);
1909 dm_kcopyd_client_destroy(cache->copier);
1912 dm_cache_metadata_close(cache->cmd);
1914 if (cache->metadata_dev)
1915 dm_put_device(cache->ti, cache->metadata_dev);
1917 if (cache->origin_dev)
1918 dm_put_device(cache->ti, cache->origin_dev);
1920 if (cache->cache_dev)
1921 dm_put_device(cache->ti, cache->cache_dev);
1924 dm_cache_policy_destroy(cache->policy);
1926 for (i = 0; i < cache->nr_ctr_args ; i++)
1927 kfree(cache->ctr_args[i]);
1928 kfree(cache->ctr_args);
1930 bioset_exit(&cache->bs);
1935 static void cache_dtr(struct dm_target *ti)
1937 struct cache *cache = ti->private;
1942 static sector_t get_dev_size(struct dm_dev *dev)
1944 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1947 /*----------------------------------------------------------------*/
1950 * Construct a cache device mapping.
1952 * cache <metadata dev> <cache dev> <origin dev> <block size>
1953 * <#feature args> [<feature arg>]*
1954 * <policy> <#policy args> [<policy arg>]*
1956 * metadata dev : fast device holding the persistent metadata
1957 * cache dev : fast device holding cached data blocks
1958 * origin dev : slow device holding original data blocks
1959 * block size : cache unit size in sectors
1961 * #feature args : number of feature arguments passed
1962 * feature args : writethrough. (The default is writeback.)
1964 * policy : the replacement policy to use
1965 * #policy args : an even number of policy arguments corresponding
1966 * to key/value pairs passed to the policy
1967 * policy args : key/value pairs passed to the policy
1968 * E.g. 'sequential_threshold 1024'
1969 * See cache-policies.txt for details.
1971 * Optional feature arguments are:
1972 * writethrough : write through caching that prohibits cache block
1973 * content from being different from origin block content.
1974 * Without this argument, the default behaviour is to write
1975 * back cache block contents later for performance reasons,
1976 * so they may differ from the corresponding origin blocks.
1979 struct dm_target *ti;
1981 struct dm_dev *metadata_dev;
1983 struct dm_dev *cache_dev;
1984 sector_t cache_sectors;
1986 struct dm_dev *origin_dev;
1987 sector_t origin_sectors;
1989 uint32_t block_size;
1991 const char *policy_name;
1993 const char **policy_argv;
1995 struct cache_features features;
1998 static void destroy_cache_args(struct cache_args *ca)
2000 if (ca->metadata_dev)
2001 dm_put_device(ca->ti, ca->metadata_dev);
2004 dm_put_device(ca->ti, ca->cache_dev);
2007 dm_put_device(ca->ti, ca->origin_dev);
2012 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2015 *error = "Insufficient args";
2022 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2026 sector_t metadata_dev_size;
2027 char b[BDEVNAME_SIZE];
2029 if (!at_least_one_arg(as, error))
2032 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2035 *error = "Error opening metadata device";
2039 metadata_dev_size = get_dev_size(ca->metadata_dev);
2040 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2041 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2042 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2047 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2052 if (!at_least_one_arg(as, error))
2055 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2058 *error = "Error opening cache device";
2061 ca->cache_sectors = get_dev_size(ca->cache_dev);
2066 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2071 if (!at_least_one_arg(as, error))
2074 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2077 *error = "Error opening origin device";
2081 ca->origin_sectors = get_dev_size(ca->origin_dev);
2082 if (ca->ti->len > ca->origin_sectors) {
2083 *error = "Device size larger than cached device";
2090 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2093 unsigned long block_size;
2095 if (!at_least_one_arg(as, error))
2098 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2099 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2100 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2101 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2102 *error = "Invalid data block size";
2106 if (block_size > ca->cache_sectors) {
2107 *error = "Data block size is larger than the cache device";
2111 ca->block_size = block_size;
2116 static void init_features(struct cache_features *cf)
2118 cf->mode = CM_WRITE;
2119 cf->io_mode = CM_IO_WRITEBACK;
2120 cf->metadata_version = 1;
2121 cf->discard_passdown = true;
2124 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2127 static const struct dm_arg _args[] = {
2128 {0, 3, "Invalid number of cache feature arguments"},
2131 int r, mode_ctr = 0;
2134 struct cache_features *cf = &ca->features;
2138 r = dm_read_arg_group(_args, as, &argc, error);
2143 arg = dm_shift_arg(as);
2145 if (!strcasecmp(arg, "writeback")) {
2146 cf->io_mode = CM_IO_WRITEBACK;
2150 else if (!strcasecmp(arg, "writethrough")) {
2151 cf->io_mode = CM_IO_WRITETHROUGH;
2155 else if (!strcasecmp(arg, "passthrough")) {
2156 cf->io_mode = CM_IO_PASSTHROUGH;
2160 else if (!strcasecmp(arg, "metadata2"))
2161 cf->metadata_version = 2;
2163 else if (!strcasecmp(arg, "no_discard_passdown"))
2164 cf->discard_passdown = false;
2167 *error = "Unrecognised cache feature requested";
2173 *error = "Duplicate cache io_mode features requested";
2180 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2183 static const struct dm_arg _args[] = {
2184 {0, 1024, "Invalid number of policy arguments"},
2189 if (!at_least_one_arg(as, error))
2192 ca->policy_name = dm_shift_arg(as);
2194 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2198 ca->policy_argv = (const char **)as->argv;
2199 dm_consume_args(as, ca->policy_argc);
2204 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2208 struct dm_arg_set as;
2213 r = parse_metadata_dev(ca, &as, error);
2217 r = parse_cache_dev(ca, &as, error);
2221 r = parse_origin_dev(ca, &as, error);
2225 r = parse_block_size(ca, &as, error);
2229 r = parse_features(ca, &as, error);
2233 r = parse_policy(ca, &as, error);
2240 /*----------------------------------------------------------------*/
2242 static struct kmem_cache *migration_cache;
2244 #define NOT_CORE_OPTION 1
2246 static int process_config_option(struct cache *cache, const char *key, const char *value)
2250 if (!strcasecmp(key, "migration_threshold")) {
2251 if (kstrtoul(value, 10, &tmp))
2254 cache->migration_threshold = tmp;
2258 return NOT_CORE_OPTION;
2261 static int set_config_value(struct cache *cache, const char *key, const char *value)
2263 int r = process_config_option(cache, key, value);
2265 if (r == NOT_CORE_OPTION)
2266 r = policy_set_config_value(cache->policy, key, value);
2269 DMWARN("bad config value for %s: %s", key, value);
2274 static int set_config_values(struct cache *cache, int argc, const char **argv)
2279 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2284 r = set_config_value(cache, argv[0], argv[1]);
2295 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2298 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2300 cache->origin_sectors,
2301 cache->sectors_per_block);
2303 *error = "Error creating cache's policy";
2307 BUG_ON(!cache->policy);
2313 * We want the discard block size to be at least the size of the cache
2314 * block size and have no more than 2^14 discard blocks across the origin.
2316 #define MAX_DISCARD_BLOCKS (1 << 14)
2318 static bool too_many_discard_blocks(sector_t discard_block_size,
2319 sector_t origin_size)
2321 (void) sector_div(origin_size, discard_block_size);
2323 return origin_size > MAX_DISCARD_BLOCKS;
2326 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2327 sector_t origin_size)
2329 sector_t discard_block_size = cache_block_size;
2332 while (too_many_discard_blocks(discard_block_size, origin_size))
2333 discard_block_size *= 2;
2335 return discard_block_size;
2338 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2340 dm_block_t nr_blocks = from_cblock(size);
2342 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2343 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2344 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2345 "Please consider increasing the cache block size to reduce the overall cache block count.",
2346 (unsigned long long) nr_blocks);
2348 cache->cache_size = size;
2351 #define DEFAULT_MIGRATION_THRESHOLD 2048
2353 static int cache_create(struct cache_args *ca, struct cache **result)
2356 char **error = &ca->ti->error;
2357 struct cache *cache;
2358 struct dm_target *ti = ca->ti;
2359 dm_block_t origin_blocks;
2360 struct dm_cache_metadata *cmd;
2361 bool may_format = ca->features.mode == CM_WRITE;
2363 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2368 ti->private = cache;
2369 ti->num_flush_bios = 2;
2370 ti->flush_supported = true;
2372 ti->num_discard_bios = 1;
2373 ti->discards_supported = true;
2375 ti->per_io_data_size = sizeof(struct per_bio_data);
2377 cache->features = ca->features;
2378 if (writethrough_mode(cache)) {
2379 /* Create bioset for writethrough bios issued to origin */
2380 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2385 cache->metadata_dev = ca->metadata_dev;
2386 cache->origin_dev = ca->origin_dev;
2387 cache->cache_dev = ca->cache_dev;
2389 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2391 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2392 origin_blocks = block_div(origin_blocks, ca->block_size);
2393 cache->origin_blocks = to_oblock(origin_blocks);
2395 cache->sectors_per_block = ca->block_size;
2396 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2401 if (ca->block_size & (ca->block_size - 1)) {
2402 dm_block_t cache_size = ca->cache_sectors;
2404 cache->sectors_per_block_shift = -1;
2405 cache_size = block_div(cache_size, ca->block_size);
2406 set_cache_size(cache, to_cblock(cache_size));
2408 cache->sectors_per_block_shift = __ffs(ca->block_size);
2409 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2412 r = create_cache_policy(cache, ca, error);
2416 cache->policy_nr_args = ca->policy_argc;
2417 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2419 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2421 *error = "Error setting cache policy's config values";
2425 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2426 ca->block_size, may_format,
2427 dm_cache_policy_get_hint_size(cache->policy),
2428 ca->features.metadata_version);
2430 *error = "Error creating metadata object";
2435 set_cache_mode(cache, CM_WRITE);
2436 if (get_cache_mode(cache) != CM_WRITE) {
2437 *error = "Unable to get write access to metadata, please check/repair metadata.";
2442 if (passthrough_mode(cache)) {
2445 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2447 *error = "dm_cache_metadata_all_clean() failed";
2452 *error = "Cannot enter passthrough mode unless all blocks are clean";
2457 policy_allow_migrations(cache->policy, false);
2460 spin_lock_init(&cache->lock);
2461 bio_list_init(&cache->deferred_bios);
2462 atomic_set(&cache->nr_allocated_migrations, 0);
2463 atomic_set(&cache->nr_io_migrations, 0);
2464 init_waitqueue_head(&cache->migration_wait);
2467 atomic_set(&cache->nr_dirty, 0);
2468 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2469 if (!cache->dirty_bitset) {
2470 *error = "could not allocate dirty bitset";
2473 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2475 cache->discard_block_size =
2476 calculate_discard_block_size(cache->sectors_per_block,
2477 cache->origin_sectors);
2478 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2479 cache->discard_block_size));
2480 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2481 if (!cache->discard_bitset) {
2482 *error = "could not allocate discard bitset";
2485 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2487 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2488 if (IS_ERR(cache->copier)) {
2489 *error = "could not create kcopyd client";
2490 r = PTR_ERR(cache->copier);
2494 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2496 *error = "could not create workqueue for metadata object";
2499 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2500 INIT_WORK(&cache->migration_worker, check_migrations);
2501 INIT_DELAYED_WORK(&cache->waker, do_waker);
2503 cache->prison = dm_bio_prison_create_v2(cache->wq);
2504 if (!cache->prison) {
2505 *error = "could not create bio prison";
2509 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2512 *error = "Error creating cache's migration mempool";
2516 cache->need_tick_bio = true;
2517 cache->sized = false;
2518 cache->invalidate = false;
2519 cache->commit_requested = false;
2520 cache->loaded_mappings = false;
2521 cache->loaded_discards = false;
2525 atomic_set(&cache->stats.demotion, 0);
2526 atomic_set(&cache->stats.promotion, 0);
2527 atomic_set(&cache->stats.copies_avoided, 0);
2528 atomic_set(&cache->stats.cache_cell_clash, 0);
2529 atomic_set(&cache->stats.commit_count, 0);
2530 atomic_set(&cache->stats.discard_count, 0);
2532 spin_lock_init(&cache->invalidation_lock);
2533 INIT_LIST_HEAD(&cache->invalidation_requests);
2535 batcher_init(&cache->committer, commit_op, cache,
2536 issue_op, cache, cache->wq);
2537 dm_iot_init(&cache->tracker);
2539 init_rwsem(&cache->background_work_lock);
2540 prevent_background_work(cache);
2549 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2554 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2557 for (i = 0; i < argc; i++) {
2558 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2567 cache->nr_ctr_args = argc;
2568 cache->ctr_args = copy;
2573 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2576 struct cache_args *ca;
2577 struct cache *cache = NULL;
2579 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2581 ti->error = "Error allocating memory for cache";
2586 r = parse_cache_args(ca, argc, argv, &ti->error);
2590 r = cache_create(ca, &cache);
2594 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2600 ti->private = cache;
2602 destroy_cache_args(ca);
2606 /*----------------------------------------------------------------*/
2608 static int cache_map(struct dm_target *ti, struct bio *bio)
2610 struct cache *cache = ti->private;
2614 dm_oblock_t block = get_bio_block(cache, bio);
2616 init_per_bio_data(bio);
2617 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2619 * This can only occur if the io goes to a partial block at
2620 * the end of the origin device. We don't cache these.
2621 * Just remap to the origin and carry on.
2623 remap_to_origin(cache, bio);
2624 accounted_begin(cache, bio);
2625 return DM_MAPIO_REMAPPED;
2628 if (discard_or_flush(bio)) {
2629 defer_bio(cache, bio);
2630 return DM_MAPIO_SUBMITTED;
2633 r = map_bio(cache, bio, block, &commit_needed);
2635 schedule_commit(&cache->committer);
2640 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2642 struct cache *cache = ti->private;
2643 unsigned long flags;
2644 struct per_bio_data *pb = get_per_bio_data(bio);
2647 policy_tick(cache->policy, false);
2649 spin_lock_irqsave(&cache->lock, flags);
2650 cache->need_tick_bio = true;
2651 spin_unlock_irqrestore(&cache->lock, flags);
2654 bio_drop_shared_lock(cache, bio);
2655 accounted_complete(cache, bio);
2657 return DM_ENDIO_DONE;
2660 static int write_dirty_bitset(struct cache *cache)
2664 if (get_cache_mode(cache) >= CM_READ_ONLY)
2667 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2669 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2674 static int write_discard_bitset(struct cache *cache)
2678 if (get_cache_mode(cache) >= CM_READ_ONLY)
2681 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2682 cache->discard_nr_blocks);
2684 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2685 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2689 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2690 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2691 is_discarded(cache, to_dblock(i)));
2693 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2701 static int write_hints(struct cache *cache)
2705 if (get_cache_mode(cache) >= CM_READ_ONLY)
2708 r = dm_cache_write_hints(cache->cmd, cache->policy);
2710 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2718 * returns true on success
2720 static bool sync_metadata(struct cache *cache)
2724 r1 = write_dirty_bitset(cache);
2726 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2728 r2 = write_discard_bitset(cache);
2730 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2734 r3 = write_hints(cache);
2736 DMERR("%s: could not write hints", cache_device_name(cache));
2739 * If writing the above metadata failed, we still commit, but don't
2740 * set the clean shutdown flag. This will effectively force every
2741 * dirty bit to be set on reload.
2743 r4 = commit(cache, !r1 && !r2 && !r3);
2745 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2747 return !r1 && !r2 && !r3 && !r4;
2750 static void cache_postsuspend(struct dm_target *ti)
2752 struct cache *cache = ti->private;
2754 prevent_background_work(cache);
2755 BUG_ON(atomic_read(&cache->nr_io_migrations));
2757 cancel_delayed_work_sync(&cache->waker);
2758 drain_workqueue(cache->wq);
2759 WARN_ON(cache->tracker.in_flight);
2762 * If it's a flush suspend there won't be any deferred bios, so this
2765 requeue_deferred_bios(cache);
2767 if (get_cache_mode(cache) == CM_WRITE)
2768 (void) sync_metadata(cache);
2771 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2772 bool dirty, uint32_t hint, bool hint_valid)
2774 struct cache *cache = context;
2777 set_bit(from_cblock(cblock), cache->dirty_bitset);
2778 atomic_inc(&cache->nr_dirty);
2780 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2782 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2786 * The discard block size in the on disk metadata is not
2787 * neccessarily the same as we're currently using. So we have to
2788 * be careful to only set the discarded attribute if we know it
2789 * covers a complete block of the new size.
2791 struct discard_load_info {
2792 struct cache *cache;
2795 * These blocks are sized using the on disk dblock size, rather
2796 * than the current one.
2798 dm_block_t block_size;
2799 dm_block_t discard_begin, discard_end;
2802 static void discard_load_info_init(struct cache *cache,
2803 struct discard_load_info *li)
2806 li->discard_begin = li->discard_end = 0;
2809 static void set_discard_range(struct discard_load_info *li)
2813 if (li->discard_begin == li->discard_end)
2817 * Convert to sectors.
2819 b = li->discard_begin * li->block_size;
2820 e = li->discard_end * li->block_size;
2823 * Then convert back to the current dblock size.
2825 b = dm_sector_div_up(b, li->cache->discard_block_size);
2826 sector_div(e, li->cache->discard_block_size);
2829 * The origin may have shrunk, so we need to check we're still in
2832 if (e > from_dblock(li->cache->discard_nr_blocks))
2833 e = from_dblock(li->cache->discard_nr_blocks);
2836 set_discard(li->cache, to_dblock(b));
2839 static int load_discard(void *context, sector_t discard_block_size,
2840 dm_dblock_t dblock, bool discard)
2842 struct discard_load_info *li = context;
2844 li->block_size = discard_block_size;
2847 if (from_dblock(dblock) == li->discard_end)
2849 * We're already in a discard range, just extend it.
2851 li->discard_end = li->discard_end + 1ULL;
2855 * Emit the old range and start a new one.
2857 set_discard_range(li);
2858 li->discard_begin = from_dblock(dblock);
2859 li->discard_end = li->discard_begin + 1ULL;
2862 set_discard_range(li);
2863 li->discard_begin = li->discard_end = 0;
2869 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2871 sector_t size = get_dev_size(cache->cache_dev);
2872 (void) sector_div(size, cache->sectors_per_block);
2873 return to_cblock(size);
2876 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2878 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2880 DMERR("%s: unable to extend cache due to missing cache table reload",
2881 cache_device_name(cache));
2887 * We can't drop a dirty block when shrinking the cache.
2889 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2890 new_size = to_cblock(from_cblock(new_size) + 1);
2891 if (is_dirty(cache, new_size)) {
2892 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2893 cache_device_name(cache),
2894 (unsigned long long) from_cblock(new_size));
2902 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2906 r = dm_cache_resize(cache->cmd, new_size);
2908 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2909 metadata_operation_failed(cache, "dm_cache_resize", r);
2913 set_cache_size(cache, new_size);
2918 static int cache_preresume(struct dm_target *ti)
2921 struct cache *cache = ti->private;
2922 dm_cblock_t csize = get_cache_dev_size(cache);
2925 * Check to see if the cache has resized.
2927 if (!cache->sized) {
2928 r = resize_cache_dev(cache, csize);
2932 cache->sized = true;
2934 } else if (csize != cache->cache_size) {
2935 if (!can_resize(cache, csize))
2938 r = resize_cache_dev(cache, csize);
2943 if (!cache->loaded_mappings) {
2944 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2945 load_mapping, cache);
2947 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2948 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2952 cache->loaded_mappings = true;
2955 if (!cache->loaded_discards) {
2956 struct discard_load_info li;
2959 * The discard bitset could have been resized, or the
2960 * discard block size changed. To be safe we start by
2961 * setting every dblock to not discarded.
2963 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2965 discard_load_info_init(cache, &li);
2966 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2968 DMERR("%s: could not load origin discards", cache_device_name(cache));
2969 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2972 set_discard_range(&li);
2974 cache->loaded_discards = true;
2980 static void cache_resume(struct dm_target *ti)
2982 struct cache *cache = ti->private;
2984 cache->need_tick_bio = true;
2985 allow_background_work(cache);
2986 do_waker(&cache->waker.work);
2989 static void emit_flags(struct cache *cache, char *result,
2990 unsigned maxlen, ssize_t *sz_ptr)
2992 ssize_t sz = *sz_ptr;
2993 struct cache_features *cf = &cache->features;
2994 unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
2996 DMEMIT("%u ", count);
2998 if (cf->metadata_version == 2)
2999 DMEMIT("metadata2 ");
3001 if (writethrough_mode(cache))
3002 DMEMIT("writethrough ");
3004 else if (passthrough_mode(cache))
3005 DMEMIT("passthrough ");
3007 else if (writeback_mode(cache))
3008 DMEMIT("writeback ");
3012 DMERR("%s: internal error: unknown io mode: %d",
3013 cache_device_name(cache), (int) cf->io_mode);
3016 if (!cf->discard_passdown)
3017 DMEMIT("no_discard_passdown ");
3025 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3026 * <cache block size> <#used cache blocks>/<#total cache blocks>
3027 * <#read hits> <#read misses> <#write hits> <#write misses>
3028 * <#demotions> <#promotions> <#dirty>
3029 * <#features> <features>*
3030 * <#core args> <core args>
3031 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3033 static void cache_status(struct dm_target *ti, status_type_t type,
3034 unsigned status_flags, char *result, unsigned maxlen)
3039 dm_block_t nr_free_blocks_metadata = 0;
3040 dm_block_t nr_blocks_metadata = 0;
3041 char buf[BDEVNAME_SIZE];
3042 struct cache *cache = ti->private;
3043 dm_cblock_t residency;
3047 case STATUSTYPE_INFO:
3048 if (get_cache_mode(cache) == CM_FAIL) {
3053 /* Commit to ensure statistics aren't out-of-date */
3054 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3055 (void) commit(cache, false);
3057 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3059 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3060 cache_device_name(cache), r);
3064 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3066 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3067 cache_device_name(cache), r);
3071 residency = policy_residency(cache->policy);
3073 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3074 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3075 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3076 (unsigned long long)nr_blocks_metadata,
3077 (unsigned long long)cache->sectors_per_block,
3078 (unsigned long long) from_cblock(residency),
3079 (unsigned long long) from_cblock(cache->cache_size),
3080 (unsigned) atomic_read(&cache->stats.read_hit),
3081 (unsigned) atomic_read(&cache->stats.read_miss),
3082 (unsigned) atomic_read(&cache->stats.write_hit),
3083 (unsigned) atomic_read(&cache->stats.write_miss),
3084 (unsigned) atomic_read(&cache->stats.demotion),
3085 (unsigned) atomic_read(&cache->stats.promotion),
3086 (unsigned long) atomic_read(&cache->nr_dirty));
3088 emit_flags(cache, result, maxlen, &sz);
3090 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3092 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3094 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3096 DMERR("%s: policy_emit_config_values returned %d",
3097 cache_device_name(cache), r);
3100 if (get_cache_mode(cache) == CM_READ_ONLY)
3105 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3107 if (r || needs_check)
3108 DMEMIT("needs_check ");
3114 case STATUSTYPE_TABLE:
3115 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3117 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3119 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3122 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3123 DMEMIT(" %s", cache->ctr_args[i]);
3124 if (cache->nr_ctr_args)
3125 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3128 case STATUSTYPE_IMA:
3129 DMEMIT_TARGET_NAME_VERSION(ti->type);
3130 if (get_cache_mode(cache) == CM_FAIL)
3131 DMEMIT(",metadata_mode=fail");
3132 else if (get_cache_mode(cache) == CM_READ_ONLY)
3133 DMEMIT(",metadata_mode=ro");
3135 DMEMIT(",metadata_mode=rw");
3137 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3138 DMEMIT(",cache_metadata_device=%s", buf);
3139 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3140 DMEMIT(",cache_device=%s", buf);
3141 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3142 DMEMIT(",cache_origin_device=%s", buf);
3143 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3144 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3145 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3146 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3147 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3159 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3160 * the one-past-the-end value.
3162 struct cblock_range {
3168 * A cache block range can take two forms:
3170 * i) A single cblock, eg. '3456'
3171 * ii) A begin and end cblock with a dash between, eg. 123-234
3173 static int parse_cblock_range(struct cache *cache, const char *str,
3174 struct cblock_range *result)
3181 * Try and parse form (ii) first.
3183 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3188 result->begin = to_cblock(b);
3189 result->end = to_cblock(e);
3194 * That didn't work, try form (i).
3196 r = sscanf(str, "%llu%c", &b, &dummy);
3201 result->begin = to_cblock(b);
3202 result->end = to_cblock(from_cblock(result->begin) + 1u);
3206 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3210 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3212 uint64_t b = from_cblock(range->begin);
3213 uint64_t e = from_cblock(range->end);
3214 uint64_t n = from_cblock(cache->cache_size);
3217 DMERR("%s: begin cblock out of range: %llu >= %llu",
3218 cache_device_name(cache), b, n);
3223 DMERR("%s: end cblock out of range: %llu > %llu",
3224 cache_device_name(cache), e, n);
3229 DMERR("%s: invalid cblock range: %llu >= %llu",
3230 cache_device_name(cache), b, e);
3237 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3239 return to_cblock(from_cblock(b) + 1);
3242 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3247 * We don't need to do any locking here because we know we're in
3248 * passthrough mode. There's is potential for a race between an
3249 * invalidation triggered by an io and an invalidation message. This
3250 * is harmless, we must not worry if the policy call fails.
3252 while (range->begin != range->end) {
3253 r = invalidate_cblock(cache, range->begin);
3257 range->begin = cblock_succ(range->begin);
3260 cache->commit_requested = true;
3264 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3265 const char **cblock_ranges)
3269 struct cblock_range range;
3271 if (!passthrough_mode(cache)) {
3272 DMERR("%s: cache has to be in passthrough mode for invalidation",
3273 cache_device_name(cache));
3277 for (i = 0; i < count; i++) {
3278 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3282 r = validate_cblock_range(cache, &range);
3287 * Pass begin and end origin blocks to the worker and wake it.
3289 r = request_invalidation(cache, &range);
3301 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3303 * The key migration_threshold is supported by the cache target core.
3305 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3306 char *result, unsigned maxlen)
3308 struct cache *cache = ti->private;
3313 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3314 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3315 cache_device_name(cache));
3319 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3320 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3325 return set_config_value(cache, argv[0], argv[1]);
3328 static int cache_iterate_devices(struct dm_target *ti,
3329 iterate_devices_callout_fn fn, void *data)
3332 struct cache *cache = ti->private;
3334 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3336 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3341 static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3343 struct request_queue *q = bdev_get_queue(origin_bdev);
3345 return blk_queue_discard(q);
3349 * If discard_passdown was enabled verify that the origin device
3350 * supports discards. Disable discard_passdown if not.
3352 static void disable_passdown_if_not_supported(struct cache *cache)
3354 struct block_device *origin_bdev = cache->origin_dev->bdev;
3355 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3356 const char *reason = NULL;
3357 char buf[BDEVNAME_SIZE];
3359 if (!cache->features.discard_passdown)
3362 if (!origin_dev_supports_discard(origin_bdev))
3363 reason = "discard unsupported";
3365 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3366 reason = "max discard sectors smaller than a block";
3369 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3370 bdevname(origin_bdev, buf), reason);
3371 cache->features.discard_passdown = false;
3375 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3377 struct block_device *origin_bdev = cache->origin_dev->bdev;
3378 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3380 if (!cache->features.discard_passdown) {
3381 /* No passdown is done so setting own virtual limits */
3382 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3383 cache->origin_sectors);
3384 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3389 * cache_iterate_devices() is stacking both origin and fast device limits
3390 * but discards aren't passed to fast device, so inherit origin's limits.
3392 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3393 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3394 limits->discard_granularity = origin_limits->discard_granularity;
3395 limits->discard_alignment = origin_limits->discard_alignment;
3396 limits->discard_misaligned = origin_limits->discard_misaligned;
3399 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3401 struct cache *cache = ti->private;
3402 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3405 * If the system-determined stacked limits are compatible with the
3406 * cache's blocksize (io_opt is a factor) do not override them.
3408 if (io_opt_sectors < cache->sectors_per_block ||
3409 do_div(io_opt_sectors, cache->sectors_per_block)) {
3410 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3411 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3414 disable_passdown_if_not_supported(cache);
3415 set_discard_limits(cache, limits);
3418 /*----------------------------------------------------------------*/
3420 static struct target_type cache_target = {
3422 .version = {2, 2, 0},
3423 .module = THIS_MODULE,
3427 .end_io = cache_end_io,
3428 .postsuspend = cache_postsuspend,
3429 .preresume = cache_preresume,
3430 .resume = cache_resume,
3431 .status = cache_status,
3432 .message = cache_message,
3433 .iterate_devices = cache_iterate_devices,
3434 .io_hints = cache_io_hints,
3437 static int __init dm_cache_init(void)
3441 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3442 if (!migration_cache)
3445 r = dm_register_target(&cache_target);
3447 DMERR("cache target registration failed: %d", r);
3448 kmem_cache_destroy(migration_cache);
3455 static void __exit dm_cache_exit(void)
3457 dm_unregister_target(&cache_target);
3458 kmem_cache_destroy(migration_cache);
3461 module_init(dm_cache_init);
3462 module_exit(dm_cache_exit);
3464 MODULE_DESCRIPTION(DM_NAME " cache target");
3465 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3466 MODULE_LICENSE("GPL");