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"
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/rwsem.h>
19 #include <linux/slab.h>
20 #include <linux/vmalloc.h>
22 #define DM_MSG_PREFIX "cache"
24 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
25 "A percentage of time allocated for copying to and/or from cache");
27 /*----------------------------------------------------------------*/
32 * oblock: index of an origin block
33 * cblock: index of a cache block
34 * promotion: movement of a block from origin to cache
35 * demotion: movement of a block from cache to origin
36 * migration: movement of a block between the origin and cache device,
40 /*----------------------------------------------------------------*/
46 * Sectors of in-flight IO.
51 * The time, in jiffies, when this device became idle (if it is
54 unsigned long idle_time;
55 unsigned long last_update_time;
58 static void iot_init(struct io_tracker *iot)
60 spin_lock_init(&iot->lock);
63 iot->last_update_time = jiffies;
66 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
71 return time_after(jiffies, iot->idle_time + jifs);
74 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
79 spin_lock_irqsave(&iot->lock, flags);
80 r = __iot_idle_for(iot, jifs);
81 spin_unlock_irqrestore(&iot->lock, flags);
86 static void iot_io_begin(struct io_tracker *iot, sector_t len)
90 spin_lock_irqsave(&iot->lock, flags);
91 iot->in_flight += len;
92 spin_unlock_irqrestore(&iot->lock, flags);
95 static void __iot_io_end(struct io_tracker *iot, sector_t len)
100 iot->in_flight -= len;
102 iot->idle_time = jiffies;
105 static void iot_io_end(struct io_tracker *iot, sector_t len)
109 spin_lock_irqsave(&iot->lock, flags);
110 __iot_io_end(iot, len);
111 spin_unlock_irqrestore(&iot->lock, flags);
114 /*----------------------------------------------------------------*/
117 * Represents a chunk of future work. 'input' allows continuations to pass
118 * values between themselves, typically error values.
120 struct continuation {
121 struct work_struct ws;
125 static inline void init_continuation(struct continuation *k,
126 void (*fn)(struct work_struct *))
128 INIT_WORK(&k->ws, fn);
132 static inline void queue_continuation(struct workqueue_struct *wq,
133 struct continuation *k)
135 queue_work(wq, &k->ws);
138 /*----------------------------------------------------------------*/
141 * The batcher collects together pieces of work that need a particular
142 * operation to occur before they can proceed (typically a commit).
146 * The operation that everyone is waiting for.
148 blk_status_t (*commit_op)(void *context);
149 void *commit_context;
152 * This is how bios should be issued once the commit op is complete
153 * (accounted_request).
155 void (*issue_op)(struct bio *bio, void *context);
159 * Queued work gets put on here after commit.
161 struct workqueue_struct *wq;
164 struct list_head work_items;
165 struct bio_list bios;
166 struct work_struct commit_work;
168 bool commit_scheduled;
171 static void __commit(struct work_struct *_ws)
173 struct batcher *b = container_of(_ws, struct batcher, commit_work);
176 struct list_head work_items;
177 struct work_struct *ws, *tmp;
178 struct continuation *k;
180 struct bio_list bios;
182 INIT_LIST_HEAD(&work_items);
183 bio_list_init(&bios);
186 * We have to grab these before the commit_op to avoid a race
189 spin_lock_irqsave(&b->lock, flags);
190 list_splice_init(&b->work_items, &work_items);
191 bio_list_merge(&bios, &b->bios);
192 bio_list_init(&b->bios);
193 b->commit_scheduled = false;
194 spin_unlock_irqrestore(&b->lock, flags);
196 r = b->commit_op(b->commit_context);
198 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
199 k = container_of(ws, struct continuation, ws);
201 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
202 queue_work(b->wq, ws);
205 while ((bio = bio_list_pop(&bios))) {
210 b->issue_op(bio, b->issue_context);
214 static void batcher_init(struct batcher *b,
215 blk_status_t (*commit_op)(void *),
216 void *commit_context,
217 void (*issue_op)(struct bio *bio, void *),
219 struct workqueue_struct *wq)
221 b->commit_op = commit_op;
222 b->commit_context = commit_context;
223 b->issue_op = issue_op;
224 b->issue_context = issue_context;
227 spin_lock_init(&b->lock);
228 INIT_LIST_HEAD(&b->work_items);
229 bio_list_init(&b->bios);
230 INIT_WORK(&b->commit_work, __commit);
231 b->commit_scheduled = false;
234 static void async_commit(struct batcher *b)
236 queue_work(b->wq, &b->commit_work);
239 static void continue_after_commit(struct batcher *b, struct continuation *k)
242 bool commit_scheduled;
244 spin_lock_irqsave(&b->lock, flags);
245 commit_scheduled = b->commit_scheduled;
246 list_add_tail(&k->ws.entry, &b->work_items);
247 spin_unlock_irqrestore(&b->lock, flags);
249 if (commit_scheduled)
254 * Bios are errored if commit failed.
256 static void issue_after_commit(struct batcher *b, struct bio *bio)
259 bool commit_scheduled;
261 spin_lock_irqsave(&b->lock, flags);
262 commit_scheduled = b->commit_scheduled;
263 bio_list_add(&b->bios, bio);
264 spin_unlock_irqrestore(&b->lock, flags);
266 if (commit_scheduled)
271 * Call this if some urgent work is waiting for the commit to complete.
273 static void schedule_commit(struct batcher *b)
278 spin_lock_irqsave(&b->lock, flags);
279 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
280 b->commit_scheduled = true;
281 spin_unlock_irqrestore(&b->lock, flags);
288 * There are a couple of places where we let a bio run, but want to do some
289 * work before calling its endio function. We do this by temporarily
290 * changing the endio fn.
292 struct dm_hook_info {
293 bio_end_io_t *bi_end_io;
296 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
297 bio_end_io_t *bi_end_io, void *bi_private)
299 h->bi_end_io = bio->bi_end_io;
301 bio->bi_end_io = bi_end_io;
302 bio->bi_private = bi_private;
305 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
307 bio->bi_end_io = h->bi_end_io;
310 /*----------------------------------------------------------------*/
312 #define MIGRATION_POOL_SIZE 128
313 #define COMMIT_PERIOD HZ
314 #define MIGRATION_COUNT_WINDOW 10
317 * The block size of the device holding cache data must be
318 * between 32KB and 1GB.
320 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
321 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
323 enum cache_metadata_mode {
324 CM_WRITE, /* metadata may be changed */
325 CM_READ_ONLY, /* metadata may not be changed */
331 * Data is written to cached blocks only. These blocks are marked
332 * dirty. If you lose the cache device you will lose data.
333 * Potential performance increase for both reads and writes.
338 * Data is written to both cache and origin. Blocks are never
339 * dirty. Potential performance benfit for reads only.
344 * A degraded mode useful for various cache coherency situations
345 * (eg, rolling back snapshots). Reads and writes always go to the
346 * origin. If a write goes to a cached oblock, then the cache
347 * block is invalidated.
352 struct cache_features {
353 enum cache_metadata_mode mode;
354 enum cache_io_mode io_mode;
355 unsigned metadata_version;
356 bool discard_passdown:1;
367 atomic_t copies_avoided;
368 atomic_t cache_cell_clash;
369 atomic_t commit_count;
370 atomic_t discard_count;
374 struct dm_target *ti;
378 * Fields for converting from sectors to blocks.
380 int sectors_per_block_shift;
381 sector_t sectors_per_block;
383 struct dm_cache_metadata *cmd;
386 * Metadata is written to this device.
388 struct dm_dev *metadata_dev;
391 * The slower of the two data devices. Typically a spindle.
393 struct dm_dev *origin_dev;
396 * The faster of the two data devices. Typically an SSD.
398 struct dm_dev *cache_dev;
401 * Size of the origin device in _complete_ blocks and native sectors.
403 dm_oblock_t origin_blocks;
404 sector_t origin_sectors;
407 * Size of the cache device in blocks.
409 dm_cblock_t cache_size;
412 * Invalidation fields.
414 spinlock_t invalidation_lock;
415 struct list_head invalidation_requests;
417 sector_t migration_threshold;
418 wait_queue_head_t migration_wait;
419 atomic_t nr_allocated_migrations;
422 * The number of in flight migrations that are performing
423 * background io. eg, promotion, writeback.
425 atomic_t nr_io_migrations;
427 struct bio_list deferred_bios;
429 struct rw_semaphore quiesce_lock;
431 struct dm_target_callbacks callbacks;
434 * origin_blocks entries, discarded if set.
436 dm_dblock_t discard_nr_blocks;
437 unsigned long *discard_bitset;
438 uint32_t discard_block_size; /* a power of 2 times sectors per block */
441 * Rather than reconstructing the table line for the status we just
442 * save it and regurgitate.
444 unsigned nr_ctr_args;
445 const char **ctr_args;
447 struct dm_kcopyd_client *copier;
448 struct work_struct deferred_bio_worker;
449 struct work_struct migration_worker;
450 struct workqueue_struct *wq;
451 struct delayed_work waker;
452 struct dm_bio_prison_v2 *prison;
455 * cache_size entries, dirty if set
457 unsigned long *dirty_bitset;
460 unsigned policy_nr_args;
461 struct dm_cache_policy *policy;
464 * Cache features such as write-through.
466 struct cache_features features;
468 struct cache_stats stats;
470 bool need_tick_bio:1;
473 bool commit_requested:1;
474 bool loaded_mappings:1;
475 bool loaded_discards:1;
477 struct rw_semaphore background_work_lock;
479 struct batcher committer;
480 struct work_struct commit_ws;
482 struct io_tracker tracker;
484 mempool_t migration_pool;
489 struct per_bio_data {
492 struct dm_bio_prison_cell_v2 *cell;
493 struct dm_hook_info hook_info;
497 struct dm_cache_migration {
498 struct continuation k;
501 struct policy_work *op;
502 struct bio *overwrite_bio;
503 struct dm_bio_prison_cell_v2 *cell;
505 dm_cblock_t invalidate_cblock;
506 dm_oblock_t invalidate_oblock;
509 /*----------------------------------------------------------------*/
511 static bool writethrough_mode(struct cache *cache)
513 return cache->features.io_mode == CM_IO_WRITETHROUGH;
516 static bool writeback_mode(struct cache *cache)
518 return cache->features.io_mode == CM_IO_WRITEBACK;
521 static inline bool passthrough_mode(struct cache *cache)
523 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
526 /*----------------------------------------------------------------*/
528 static void wake_deferred_bio_worker(struct cache *cache)
530 queue_work(cache->wq, &cache->deferred_bio_worker);
533 static void wake_migration_worker(struct cache *cache)
535 if (passthrough_mode(cache))
538 queue_work(cache->wq, &cache->migration_worker);
541 /*----------------------------------------------------------------*/
543 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
545 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
548 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
550 dm_bio_prison_free_cell_v2(cache->prison, cell);
553 static struct dm_cache_migration *alloc_migration(struct cache *cache)
555 struct dm_cache_migration *mg;
557 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
559 memset(mg, 0, sizeof(*mg));
562 atomic_inc(&cache->nr_allocated_migrations);
567 static void free_migration(struct dm_cache_migration *mg)
569 struct cache *cache = mg->cache;
571 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
572 wake_up(&cache->migration_wait);
574 mempool_free(mg, &cache->migration_pool);
577 /*----------------------------------------------------------------*/
579 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
581 return to_oblock(from_oblock(b) + 1ull);
584 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
588 key->block_begin = from_oblock(begin);
589 key->block_end = from_oblock(end);
593 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
594 * level 1 which prevents *both* READs and WRITEs.
596 #define WRITE_LOCK_LEVEL 0
597 #define READ_WRITE_LOCK_LEVEL 1
599 static unsigned lock_level(struct bio *bio)
601 return bio_data_dir(bio) == WRITE ?
603 READ_WRITE_LOCK_LEVEL;
606 /*----------------------------------------------------------------
608 *--------------------------------------------------------------*/
610 static struct per_bio_data *get_per_bio_data(struct bio *bio)
612 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
617 static struct per_bio_data *init_per_bio_data(struct bio *bio)
619 struct per_bio_data *pb = get_per_bio_data(bio);
622 pb->req_nr = dm_bio_get_target_bio_nr(bio);
629 /*----------------------------------------------------------------*/
631 static void defer_bio(struct cache *cache, struct bio *bio)
635 spin_lock_irqsave(&cache->lock, flags);
636 bio_list_add(&cache->deferred_bios, bio);
637 spin_unlock_irqrestore(&cache->lock, flags);
639 wake_deferred_bio_worker(cache);
642 static void defer_bios(struct cache *cache, struct bio_list *bios)
646 spin_lock_irqsave(&cache->lock, flags);
647 bio_list_merge(&cache->deferred_bios, bios);
649 spin_unlock_irqrestore(&cache->lock, flags);
651 wake_deferred_bio_worker(cache);
654 /*----------------------------------------------------------------*/
656 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
659 struct per_bio_data *pb;
660 struct dm_cell_key_v2 key;
661 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
662 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
664 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
666 build_key(oblock, end, &key);
667 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
670 * Failed to get the lock.
672 free_prison_cell(cache, cell_prealloc);
676 if (cell != cell_prealloc)
677 free_prison_cell(cache, cell_prealloc);
679 pb = get_per_bio_data(bio);
685 /*----------------------------------------------------------------*/
687 static bool is_dirty(struct cache *cache, dm_cblock_t b)
689 return test_bit(from_cblock(b), cache->dirty_bitset);
692 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
694 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
695 atomic_inc(&cache->nr_dirty);
696 policy_set_dirty(cache->policy, cblock);
701 * These two are called when setting after migrations to force the policy
702 * and dirty bitset to be in sync.
704 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
706 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
707 atomic_inc(&cache->nr_dirty);
708 policy_set_dirty(cache->policy, cblock);
711 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
713 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
714 if (atomic_dec_return(&cache->nr_dirty) == 0)
715 dm_table_event(cache->ti->table);
718 policy_clear_dirty(cache->policy, cblock);
721 /*----------------------------------------------------------------*/
723 static bool block_size_is_power_of_two(struct cache *cache)
725 return cache->sectors_per_block_shift >= 0;
728 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
729 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
732 static dm_block_t block_div(dm_block_t b, uint32_t n)
739 static dm_block_t oblocks_per_dblock(struct cache *cache)
741 dm_block_t oblocks = cache->discard_block_size;
743 if (block_size_is_power_of_two(cache))
744 oblocks >>= cache->sectors_per_block_shift;
746 oblocks = block_div(oblocks, cache->sectors_per_block);
751 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
753 return to_dblock(block_div(from_oblock(oblock),
754 oblocks_per_dblock(cache)));
757 static void set_discard(struct cache *cache, dm_dblock_t b)
761 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
762 atomic_inc(&cache->stats.discard_count);
764 spin_lock_irqsave(&cache->lock, flags);
765 set_bit(from_dblock(b), cache->discard_bitset);
766 spin_unlock_irqrestore(&cache->lock, flags);
769 static void clear_discard(struct cache *cache, dm_dblock_t b)
773 spin_lock_irqsave(&cache->lock, flags);
774 clear_bit(from_dblock(b), cache->discard_bitset);
775 spin_unlock_irqrestore(&cache->lock, flags);
778 static bool is_discarded(struct cache *cache, dm_dblock_t b)
783 spin_lock_irqsave(&cache->lock, flags);
784 r = test_bit(from_dblock(b), cache->discard_bitset);
785 spin_unlock_irqrestore(&cache->lock, flags);
790 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
795 spin_lock_irqsave(&cache->lock, flags);
796 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
797 cache->discard_bitset);
798 spin_unlock_irqrestore(&cache->lock, flags);
803 /*----------------------------------------------------------------
805 *--------------------------------------------------------------*/
806 static void remap_to_origin(struct cache *cache, struct bio *bio)
808 bio_set_dev(bio, cache->origin_dev->bdev);
811 static void remap_to_cache(struct cache *cache, struct bio *bio,
814 sector_t bi_sector = bio->bi_iter.bi_sector;
815 sector_t block = from_cblock(cblock);
817 bio_set_dev(bio, cache->cache_dev->bdev);
818 if (!block_size_is_power_of_two(cache))
819 bio->bi_iter.bi_sector =
820 (block * cache->sectors_per_block) +
821 sector_div(bi_sector, cache->sectors_per_block);
823 bio->bi_iter.bi_sector =
824 (block << cache->sectors_per_block_shift) |
825 (bi_sector & (cache->sectors_per_block - 1));
828 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
831 struct per_bio_data *pb;
833 spin_lock_irqsave(&cache->lock, flags);
834 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
835 bio_op(bio) != REQ_OP_DISCARD) {
836 pb = get_per_bio_data(bio);
838 cache->need_tick_bio = false;
840 spin_unlock_irqrestore(&cache->lock, flags);
843 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
844 dm_oblock_t oblock, bool bio_has_pbd)
847 check_if_tick_bio_needed(cache, bio);
848 remap_to_origin(cache, bio);
849 if (bio_data_dir(bio) == WRITE)
850 clear_discard(cache, oblock_to_dblock(cache, oblock));
853 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
856 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
857 __remap_to_origin_clear_discard(cache, bio, oblock, true);
860 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
861 dm_oblock_t oblock, dm_cblock_t cblock)
863 check_if_tick_bio_needed(cache, bio);
864 remap_to_cache(cache, bio, cblock);
865 if (bio_data_dir(bio) == WRITE) {
866 set_dirty(cache, cblock);
867 clear_discard(cache, oblock_to_dblock(cache, oblock));
871 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
873 sector_t block_nr = bio->bi_iter.bi_sector;
875 if (!block_size_is_power_of_two(cache))
876 (void) sector_div(block_nr, cache->sectors_per_block);
878 block_nr >>= cache->sectors_per_block_shift;
880 return to_oblock(block_nr);
883 static bool accountable_bio(struct cache *cache, struct bio *bio)
885 return bio_op(bio) != REQ_OP_DISCARD;
888 static void accounted_begin(struct cache *cache, struct bio *bio)
890 struct per_bio_data *pb;
892 if (accountable_bio(cache, bio)) {
893 pb = get_per_bio_data(bio);
894 pb->len = bio_sectors(bio);
895 iot_io_begin(&cache->tracker, pb->len);
899 static void accounted_complete(struct cache *cache, struct bio *bio)
901 struct per_bio_data *pb = get_per_bio_data(bio);
903 iot_io_end(&cache->tracker, pb->len);
906 static void accounted_request(struct cache *cache, struct bio *bio)
908 accounted_begin(cache, bio);
909 generic_make_request(bio);
912 static void issue_op(struct bio *bio, void *context)
914 struct cache *cache = context;
915 accounted_request(cache, bio);
919 * When running in writethrough mode we need to send writes to clean blocks
920 * to both the cache and origin devices. Clone the bio and send them in parallel.
922 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
923 dm_oblock_t oblock, dm_cblock_t cblock)
925 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
929 bio_chain(origin_bio, bio);
931 * Passing false to __remap_to_origin_clear_discard() skips
932 * all code that might use per_bio_data (since clone doesn't have it)
934 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
935 submit_bio(origin_bio);
937 remap_to_cache(cache, bio, cblock);
940 /*----------------------------------------------------------------
942 *--------------------------------------------------------------*/
943 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
945 return cache->features.mode;
948 static const char *cache_device_name(struct cache *cache)
950 return dm_device_name(dm_table_get_md(cache->ti->table));
953 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
955 const char *descs[] = {
961 dm_table_event(cache->ti->table);
962 DMINFO("%s: switching cache to %s mode",
963 cache_device_name(cache), descs[(int)mode]);
966 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
969 enum cache_metadata_mode old_mode = get_cache_mode(cache);
971 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
972 DMERR("%s: unable to read needs_check flag, setting failure mode.",
973 cache_device_name(cache));
977 if (new_mode == CM_WRITE && needs_check) {
978 DMERR("%s: unable to switch cache to write mode until repaired.",
979 cache_device_name(cache));
980 if (old_mode != new_mode)
983 new_mode = CM_READ_ONLY;
986 /* Never move out of fail mode */
987 if (old_mode == CM_FAIL)
993 dm_cache_metadata_set_read_only(cache->cmd);
997 dm_cache_metadata_set_read_write(cache->cmd);
1001 cache->features.mode = new_mode;
1003 if (new_mode != old_mode)
1004 notify_mode_switch(cache, new_mode);
1007 static void abort_transaction(struct cache *cache)
1009 const char *dev_name = cache_device_name(cache);
1011 if (get_cache_mode(cache) >= CM_READ_ONLY)
1014 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1015 if (dm_cache_metadata_abort(cache->cmd)) {
1016 DMERR("%s: failed to abort metadata transaction", dev_name);
1017 set_cache_mode(cache, CM_FAIL);
1020 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1021 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1022 set_cache_mode(cache, CM_FAIL);
1026 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1028 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1029 cache_device_name(cache), op, r);
1030 abort_transaction(cache);
1031 set_cache_mode(cache, CM_READ_ONLY);
1034 /*----------------------------------------------------------------*/
1036 static void load_stats(struct cache *cache)
1038 struct dm_cache_statistics stats;
1040 dm_cache_metadata_get_stats(cache->cmd, &stats);
1041 atomic_set(&cache->stats.read_hit, stats.read_hits);
1042 atomic_set(&cache->stats.read_miss, stats.read_misses);
1043 atomic_set(&cache->stats.write_hit, stats.write_hits);
1044 atomic_set(&cache->stats.write_miss, stats.write_misses);
1047 static void save_stats(struct cache *cache)
1049 struct dm_cache_statistics stats;
1051 if (get_cache_mode(cache) >= CM_READ_ONLY)
1054 stats.read_hits = atomic_read(&cache->stats.read_hit);
1055 stats.read_misses = atomic_read(&cache->stats.read_miss);
1056 stats.write_hits = atomic_read(&cache->stats.write_hit);
1057 stats.write_misses = atomic_read(&cache->stats.write_miss);
1059 dm_cache_metadata_set_stats(cache->cmd, &stats);
1062 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1065 case POLICY_PROMOTE:
1066 atomic_inc(&stats->promotion);
1070 atomic_inc(&stats->demotion);
1073 case POLICY_WRITEBACK:
1074 atomic_inc(&stats->writeback);
1079 /*----------------------------------------------------------------
1080 * Migration processing
1082 * Migration covers moving data from the origin device to the cache, or
1084 *--------------------------------------------------------------*/
1086 static void inc_io_migrations(struct cache *cache)
1088 atomic_inc(&cache->nr_io_migrations);
1091 static void dec_io_migrations(struct cache *cache)
1093 atomic_dec(&cache->nr_io_migrations);
1096 static bool discard_or_flush(struct bio *bio)
1098 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1101 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1102 dm_dblock_t *b, dm_dblock_t *e)
1104 sector_t sb = bio->bi_iter.bi_sector;
1105 sector_t se = bio_end_sector(bio);
1107 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1109 if (se - sb < cache->discard_block_size)
1112 *e = to_dblock(block_div(se, cache->discard_block_size));
1115 /*----------------------------------------------------------------*/
1117 static void prevent_background_work(struct cache *cache)
1120 down_write(&cache->background_work_lock);
1124 static void allow_background_work(struct cache *cache)
1127 up_write(&cache->background_work_lock);
1131 static bool background_work_begin(struct cache *cache)
1136 r = down_read_trylock(&cache->background_work_lock);
1142 static void background_work_end(struct cache *cache)
1145 up_read(&cache->background_work_lock);
1149 /*----------------------------------------------------------------*/
1151 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1153 return (bio_data_dir(bio) == WRITE) &&
1154 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1157 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1159 return writeback_mode(cache) &&
1160 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1163 static void quiesce(struct dm_cache_migration *mg,
1164 void (*continuation)(struct work_struct *))
1166 init_continuation(&mg->k, continuation);
1167 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1170 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1172 struct continuation *k = container_of(ws, struct continuation, ws);
1173 return container_of(k, struct dm_cache_migration, k);
1176 static void copy_complete(int read_err, unsigned long write_err, void *context)
1178 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1180 if (read_err || write_err)
1181 mg->k.input = BLK_STS_IOERR;
1183 queue_continuation(mg->cache->wq, &mg->k);
1186 static void copy(struct dm_cache_migration *mg, bool promote)
1188 struct dm_io_region o_region, c_region;
1189 struct cache *cache = mg->cache;
1191 o_region.bdev = cache->origin_dev->bdev;
1192 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1193 o_region.count = cache->sectors_per_block;
1195 c_region.bdev = cache->cache_dev->bdev;
1196 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1197 c_region.count = cache->sectors_per_block;
1200 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1202 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1205 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1207 struct per_bio_data *pb = get_per_bio_data(bio);
1209 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1210 free_prison_cell(cache, pb->cell);
1214 static void overwrite_endio(struct bio *bio)
1216 struct dm_cache_migration *mg = bio->bi_private;
1217 struct cache *cache = mg->cache;
1218 struct per_bio_data *pb = get_per_bio_data(bio);
1220 dm_unhook_bio(&pb->hook_info, bio);
1223 mg->k.input = bio->bi_status;
1225 queue_continuation(cache->wq, &mg->k);
1228 static void overwrite(struct dm_cache_migration *mg,
1229 void (*continuation)(struct work_struct *))
1231 struct bio *bio = mg->overwrite_bio;
1232 struct per_bio_data *pb = get_per_bio_data(bio);
1234 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1237 * The overwrite bio is part of the copy operation, as such it does
1238 * not set/clear discard or dirty flags.
1240 if (mg->op->op == POLICY_PROMOTE)
1241 remap_to_cache(mg->cache, bio, mg->op->cblock);
1243 remap_to_origin(mg->cache, bio);
1245 init_continuation(&mg->k, continuation);
1246 accounted_request(mg->cache, bio);
1252 * 1) exclusive lock preventing WRITEs
1254 * 3) copy or issue overwrite bio
1255 * 4) upgrade to exclusive lock preventing READs and WRITEs
1257 * 6) update metadata and commit
1260 static void mg_complete(struct dm_cache_migration *mg, bool success)
1262 struct bio_list bios;
1263 struct cache *cache = mg->cache;
1264 struct policy_work *op = mg->op;
1265 dm_cblock_t cblock = op->cblock;
1268 update_stats(&cache->stats, op->op);
1271 case POLICY_PROMOTE:
1272 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1273 policy_complete_background_work(cache->policy, op, success);
1275 if (mg->overwrite_bio) {
1277 force_set_dirty(cache, cblock);
1278 else if (mg->k.input)
1279 mg->overwrite_bio->bi_status = mg->k.input;
1281 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1282 bio_endio(mg->overwrite_bio);
1285 force_clear_dirty(cache, cblock);
1286 dec_io_migrations(cache);
1292 * We clear dirty here to update the nr_dirty counter.
1295 force_clear_dirty(cache, cblock);
1296 policy_complete_background_work(cache->policy, op, success);
1297 dec_io_migrations(cache);
1300 case POLICY_WRITEBACK:
1302 force_clear_dirty(cache, cblock);
1303 policy_complete_background_work(cache->policy, op, success);
1304 dec_io_migrations(cache);
1308 bio_list_init(&bios);
1310 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1311 free_prison_cell(cache, mg->cell);
1315 defer_bios(cache, &bios);
1316 wake_migration_worker(cache);
1318 background_work_end(cache);
1321 static void mg_success(struct work_struct *ws)
1323 struct dm_cache_migration *mg = ws_to_mg(ws);
1324 mg_complete(mg, mg->k.input == 0);
1327 static void mg_update_metadata(struct work_struct *ws)
1330 struct dm_cache_migration *mg = ws_to_mg(ws);
1331 struct cache *cache = mg->cache;
1332 struct policy_work *op = mg->op;
1335 case POLICY_PROMOTE:
1336 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1338 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1339 cache_device_name(cache));
1340 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1342 mg_complete(mg, false);
1345 mg_complete(mg, true);
1349 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1351 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1352 cache_device_name(cache));
1353 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1355 mg_complete(mg, false);
1360 * It would be nice if we only had to commit when a REQ_FLUSH
1361 * comes through. But there's one scenario that we have to
1364 * - vblock x in a cache block
1366 * - cache block gets reallocated and over written
1369 * When we recover, because there was no commit the cache will
1370 * rollback to having the data for vblock x in the cache block.
1371 * But the cache block has since been overwritten, so it'll end
1372 * up pointing to data that was never in 'x' during the history
1375 * To avoid this issue we require a commit as part of the
1376 * demotion operation.
1378 init_continuation(&mg->k, mg_success);
1379 continue_after_commit(&cache->committer, &mg->k);
1380 schedule_commit(&cache->committer);
1383 case POLICY_WRITEBACK:
1384 mg_complete(mg, true);
1389 static void mg_update_metadata_after_copy(struct work_struct *ws)
1391 struct dm_cache_migration *mg = ws_to_mg(ws);
1394 * Did the copy succeed?
1397 mg_complete(mg, false);
1399 mg_update_metadata(ws);
1402 static void mg_upgrade_lock(struct work_struct *ws)
1405 struct dm_cache_migration *mg = ws_to_mg(ws);
1408 * Did the copy succeed?
1411 mg_complete(mg, false);
1415 * Now we want the lock to prevent both reads and writes.
1417 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1418 READ_WRITE_LOCK_LEVEL);
1420 mg_complete(mg, false);
1423 quiesce(mg, mg_update_metadata);
1426 mg_update_metadata(ws);
1430 static void mg_full_copy(struct work_struct *ws)
1432 struct dm_cache_migration *mg = ws_to_mg(ws);
1433 struct cache *cache = mg->cache;
1434 struct policy_work *op = mg->op;
1435 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1437 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1438 is_discarded_oblock(cache, op->oblock)) {
1439 mg_upgrade_lock(ws);
1443 init_continuation(&mg->k, mg_upgrade_lock);
1444 copy(mg, is_policy_promote);
1447 static void mg_copy(struct work_struct *ws)
1449 struct dm_cache_migration *mg = ws_to_mg(ws);
1451 if (mg->overwrite_bio) {
1453 * No exclusive lock was held when we last checked if the bio
1454 * was optimisable. So we have to check again in case things
1455 * have changed (eg, the block may no longer be discarded).
1457 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1459 * Fallback to a real full copy after doing some tidying up.
1461 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1462 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1463 mg->overwrite_bio = NULL;
1464 inc_io_migrations(mg->cache);
1470 * It's safe to do this here, even though it's new data
1471 * because all IO has been locked out of the block.
1473 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1474 * so _not_ using mg_upgrade_lock() as continutation.
1476 overwrite(mg, mg_update_metadata_after_copy);
1482 static int mg_lock_writes(struct dm_cache_migration *mg)
1485 struct dm_cell_key_v2 key;
1486 struct cache *cache = mg->cache;
1487 struct dm_bio_prison_cell_v2 *prealloc;
1489 prealloc = alloc_prison_cell(cache);
1492 * Prevent writes to the block, but allow reads to continue.
1493 * Unless we're using an overwrite bio, in which case we lock
1496 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1497 r = dm_cell_lock_v2(cache->prison, &key,
1498 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1499 prealloc, &mg->cell);
1501 free_prison_cell(cache, prealloc);
1502 mg_complete(mg, false);
1506 if (mg->cell != prealloc)
1507 free_prison_cell(cache, prealloc);
1512 quiesce(mg, mg_copy);
1517 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1519 struct dm_cache_migration *mg;
1521 if (!background_work_begin(cache)) {
1522 policy_complete_background_work(cache->policy, op, false);
1526 mg = alloc_migration(cache);
1529 mg->overwrite_bio = bio;
1532 inc_io_migrations(cache);
1534 return mg_lock_writes(mg);
1537 /*----------------------------------------------------------------
1538 * invalidation processing
1539 *--------------------------------------------------------------*/
1541 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1543 struct bio_list bios;
1544 struct cache *cache = mg->cache;
1546 bio_list_init(&bios);
1547 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1548 free_prison_cell(cache, mg->cell);
1550 if (!success && mg->overwrite_bio)
1551 bio_io_error(mg->overwrite_bio);
1554 defer_bios(cache, &bios);
1556 background_work_end(cache);
1559 static void invalidate_completed(struct work_struct *ws)
1561 struct dm_cache_migration *mg = ws_to_mg(ws);
1562 invalidate_complete(mg, !mg->k.input);
1565 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1567 int r = policy_invalidate_mapping(cache->policy, cblock);
1569 r = dm_cache_remove_mapping(cache->cmd, cblock);
1571 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1572 cache_device_name(cache));
1573 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1576 } else if (r == -ENODATA) {
1578 * Harmless, already unmapped.
1583 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1588 static void invalidate_remove(struct work_struct *ws)
1591 struct dm_cache_migration *mg = ws_to_mg(ws);
1592 struct cache *cache = mg->cache;
1594 r = invalidate_cblock(cache, mg->invalidate_cblock);
1596 invalidate_complete(mg, false);
1600 init_continuation(&mg->k, invalidate_completed);
1601 continue_after_commit(&cache->committer, &mg->k);
1602 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1603 mg->overwrite_bio = NULL;
1604 schedule_commit(&cache->committer);
1607 static int invalidate_lock(struct dm_cache_migration *mg)
1610 struct dm_cell_key_v2 key;
1611 struct cache *cache = mg->cache;
1612 struct dm_bio_prison_cell_v2 *prealloc;
1614 prealloc = alloc_prison_cell(cache);
1616 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1617 r = dm_cell_lock_v2(cache->prison, &key,
1618 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1620 free_prison_cell(cache, prealloc);
1621 invalidate_complete(mg, false);
1625 if (mg->cell != prealloc)
1626 free_prison_cell(cache, prealloc);
1629 quiesce(mg, invalidate_remove);
1633 * We can't call invalidate_remove() directly here because we
1634 * might still be in request context.
1636 init_continuation(&mg->k, invalidate_remove);
1637 queue_work(cache->wq, &mg->k.ws);
1643 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1644 dm_oblock_t oblock, struct bio *bio)
1646 struct dm_cache_migration *mg;
1648 if (!background_work_begin(cache))
1651 mg = alloc_migration(cache);
1653 mg->overwrite_bio = bio;
1654 mg->invalidate_cblock = cblock;
1655 mg->invalidate_oblock = oblock;
1657 return invalidate_lock(mg);
1660 /*----------------------------------------------------------------
1662 *--------------------------------------------------------------*/
1669 static enum busy spare_migration_bandwidth(struct cache *cache)
1671 bool idle = iot_idle_for(&cache->tracker, HZ);
1672 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1673 cache->sectors_per_block;
1675 if (idle && current_volume <= cache->migration_threshold)
1681 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1683 atomic_inc(bio_data_dir(bio) == READ ?
1684 &cache->stats.read_hit : &cache->stats.write_hit);
1687 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1689 atomic_inc(bio_data_dir(bio) == READ ?
1690 &cache->stats.read_miss : &cache->stats.write_miss);
1693 /*----------------------------------------------------------------*/
1695 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1696 bool *commit_needed)
1699 bool rb, background_queued;
1702 *commit_needed = false;
1704 rb = bio_detain_shared(cache, block, bio);
1707 * An exclusive lock is held for this block, so we have to
1708 * wait. We set the commit_needed flag so the current
1709 * transaction will be committed asap, allowing this lock
1712 *commit_needed = true;
1713 return DM_MAPIO_SUBMITTED;
1716 data_dir = bio_data_dir(bio);
1718 if (optimisable_bio(cache, bio, block)) {
1719 struct policy_work *op = NULL;
1721 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1722 if (unlikely(r && r != -ENOENT)) {
1723 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1724 cache_device_name(cache), r);
1726 return DM_MAPIO_SUBMITTED;
1729 if (r == -ENOENT && op) {
1730 bio_drop_shared_lock(cache, bio);
1731 BUG_ON(op->op != POLICY_PROMOTE);
1732 mg_start(cache, op, bio);
1733 return DM_MAPIO_SUBMITTED;
1736 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1737 if (unlikely(r && r != -ENOENT)) {
1738 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1739 cache_device_name(cache), r);
1741 return DM_MAPIO_SUBMITTED;
1744 if (background_queued)
1745 wake_migration_worker(cache);
1749 struct per_bio_data *pb = get_per_bio_data(bio);
1754 inc_miss_counter(cache, bio);
1755 if (pb->req_nr == 0) {
1756 accounted_begin(cache, bio);
1757 remap_to_origin_clear_discard(cache, bio, block);
1760 * This is a duplicate writethrough io that is no
1761 * longer needed because the block has been demoted.
1764 return DM_MAPIO_SUBMITTED;
1770 inc_hit_counter(cache, bio);
1773 * Passthrough always maps to the origin, invalidating any
1774 * cache blocks that are written to.
1776 if (passthrough_mode(cache)) {
1777 if (bio_data_dir(bio) == WRITE) {
1778 bio_drop_shared_lock(cache, bio);
1779 atomic_inc(&cache->stats.demotion);
1780 invalidate_start(cache, cblock, block, bio);
1782 remap_to_origin_clear_discard(cache, bio, block);
1784 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1785 !is_dirty(cache, cblock)) {
1786 remap_to_origin_and_cache(cache, bio, block, cblock);
1787 accounted_begin(cache, bio);
1789 remap_to_cache_dirty(cache, bio, block, cblock);
1794 * dm core turns FUA requests into a separate payload and FLUSH req.
1796 if (bio->bi_opf & REQ_FUA) {
1798 * issue_after_commit will call accounted_begin a second time. So
1799 * we call accounted_complete() to avoid double accounting.
1801 accounted_complete(cache, bio);
1802 issue_after_commit(&cache->committer, bio);
1803 *commit_needed = true;
1804 return DM_MAPIO_SUBMITTED;
1807 return DM_MAPIO_REMAPPED;
1810 static bool process_bio(struct cache *cache, struct bio *bio)
1814 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1815 generic_make_request(bio);
1817 return commit_needed;
1821 * A non-zero return indicates read_only or fail_io mode.
1823 static int commit(struct cache *cache, bool clean_shutdown)
1827 if (get_cache_mode(cache) >= CM_READ_ONLY)
1830 atomic_inc(&cache->stats.commit_count);
1831 r = dm_cache_commit(cache->cmd, clean_shutdown);
1833 metadata_operation_failed(cache, "dm_cache_commit", r);
1839 * Used by the batcher.
1841 static blk_status_t commit_op(void *context)
1843 struct cache *cache = context;
1845 if (dm_cache_changed_this_transaction(cache->cmd))
1846 return errno_to_blk_status(commit(cache, false));
1851 /*----------------------------------------------------------------*/
1853 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1855 struct per_bio_data *pb = get_per_bio_data(bio);
1858 remap_to_origin(cache, bio);
1860 remap_to_cache(cache, bio, 0);
1862 issue_after_commit(&cache->committer, bio);
1866 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1870 // FIXME: do we need to lock the region? Or can we just assume the
1871 // user wont be so foolish as to issue discard concurrently with
1873 calc_discard_block_range(cache, bio, &b, &e);
1875 set_discard(cache, b);
1876 b = to_dblock(from_dblock(b) + 1);
1879 if (cache->features.discard_passdown) {
1880 remap_to_origin(cache, bio);
1881 generic_make_request(bio);
1888 static void process_deferred_bios(struct work_struct *ws)
1890 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1892 unsigned long flags;
1893 bool commit_needed = false;
1894 struct bio_list bios;
1897 bio_list_init(&bios);
1899 spin_lock_irqsave(&cache->lock, flags);
1900 bio_list_merge(&bios, &cache->deferred_bios);
1901 bio_list_init(&cache->deferred_bios);
1902 spin_unlock_irqrestore(&cache->lock, flags);
1904 while ((bio = bio_list_pop(&bios))) {
1905 if (bio->bi_opf & REQ_PREFLUSH)
1906 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1908 else if (bio_op(bio) == REQ_OP_DISCARD)
1909 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1912 commit_needed = process_bio(cache, bio) || commit_needed;
1917 schedule_commit(&cache->committer);
1920 /*----------------------------------------------------------------
1922 *--------------------------------------------------------------*/
1924 static void requeue_deferred_bios(struct cache *cache)
1927 struct bio_list bios;
1929 bio_list_init(&bios);
1930 bio_list_merge(&bios, &cache->deferred_bios);
1931 bio_list_init(&cache->deferred_bios);
1933 while ((bio = bio_list_pop(&bios))) {
1934 bio->bi_status = BLK_STS_DM_REQUEUE;
1941 * We want to commit periodically so that not too much
1942 * unwritten metadata builds up.
1944 static void do_waker(struct work_struct *ws)
1946 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1948 policy_tick(cache->policy, true);
1949 wake_migration_worker(cache);
1950 schedule_commit(&cache->committer);
1951 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1954 static void check_migrations(struct work_struct *ws)
1957 struct policy_work *op;
1958 struct cache *cache = container_of(ws, struct cache, migration_worker);
1962 b = spare_migration_bandwidth(cache);
1964 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1969 DMERR_LIMIT("%s: policy_background_work failed",
1970 cache_device_name(cache));
1974 r = mg_start(cache, op, NULL);
1982 /*----------------------------------------------------------------
1984 *--------------------------------------------------------------*/
1987 * This function gets called on the error paths of the constructor, so we
1988 * have to cope with a partially initialised struct.
1990 static void destroy(struct cache *cache)
1994 mempool_exit(&cache->migration_pool);
1997 dm_bio_prison_destroy_v2(cache->prison);
1999 cancel_delayed_work_sync(&cache->waker);
2001 destroy_workqueue(cache->wq);
2003 if (cache->dirty_bitset)
2004 free_bitset(cache->dirty_bitset);
2006 if (cache->discard_bitset)
2007 free_bitset(cache->discard_bitset);
2010 dm_kcopyd_client_destroy(cache->copier);
2013 dm_cache_metadata_close(cache->cmd);
2015 if (cache->metadata_dev)
2016 dm_put_device(cache->ti, cache->metadata_dev);
2018 if (cache->origin_dev)
2019 dm_put_device(cache->ti, cache->origin_dev);
2021 if (cache->cache_dev)
2022 dm_put_device(cache->ti, cache->cache_dev);
2025 dm_cache_policy_destroy(cache->policy);
2027 for (i = 0; i < cache->nr_ctr_args ; i++)
2028 kfree(cache->ctr_args[i]);
2029 kfree(cache->ctr_args);
2031 bioset_exit(&cache->bs);
2036 static void cache_dtr(struct dm_target *ti)
2038 struct cache *cache = ti->private;
2043 static sector_t get_dev_size(struct dm_dev *dev)
2045 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2048 /*----------------------------------------------------------------*/
2051 * Construct a cache device mapping.
2053 * cache <metadata dev> <cache dev> <origin dev> <block size>
2054 * <#feature args> [<feature arg>]*
2055 * <policy> <#policy args> [<policy arg>]*
2057 * metadata dev : fast device holding the persistent metadata
2058 * cache dev : fast device holding cached data blocks
2059 * origin dev : slow device holding original data blocks
2060 * block size : cache unit size in sectors
2062 * #feature args : number of feature arguments passed
2063 * feature args : writethrough. (The default is writeback.)
2065 * policy : the replacement policy to use
2066 * #policy args : an even number of policy arguments corresponding
2067 * to key/value pairs passed to the policy
2068 * policy args : key/value pairs passed to the policy
2069 * E.g. 'sequential_threshold 1024'
2070 * See cache-policies.txt for details.
2072 * Optional feature arguments are:
2073 * writethrough : write through caching that prohibits cache block
2074 * content from being different from origin block content.
2075 * Without this argument, the default behaviour is to write
2076 * back cache block contents later for performance reasons,
2077 * so they may differ from the corresponding origin blocks.
2080 struct dm_target *ti;
2082 struct dm_dev *metadata_dev;
2084 struct dm_dev *cache_dev;
2085 sector_t cache_sectors;
2087 struct dm_dev *origin_dev;
2088 sector_t origin_sectors;
2090 uint32_t block_size;
2092 const char *policy_name;
2094 const char **policy_argv;
2096 struct cache_features features;
2099 static void destroy_cache_args(struct cache_args *ca)
2101 if (ca->metadata_dev)
2102 dm_put_device(ca->ti, ca->metadata_dev);
2105 dm_put_device(ca->ti, ca->cache_dev);
2108 dm_put_device(ca->ti, ca->origin_dev);
2113 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2116 *error = "Insufficient args";
2123 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2127 sector_t metadata_dev_size;
2128 char b[BDEVNAME_SIZE];
2130 if (!at_least_one_arg(as, error))
2133 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2136 *error = "Error opening metadata device";
2140 metadata_dev_size = get_dev_size(ca->metadata_dev);
2141 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2142 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2143 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2148 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2153 if (!at_least_one_arg(as, error))
2156 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2159 *error = "Error opening cache device";
2162 ca->cache_sectors = get_dev_size(ca->cache_dev);
2167 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2172 if (!at_least_one_arg(as, error))
2175 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2178 *error = "Error opening origin device";
2182 ca->origin_sectors = get_dev_size(ca->origin_dev);
2183 if (ca->ti->len > ca->origin_sectors) {
2184 *error = "Device size larger than cached device";
2191 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2194 unsigned long block_size;
2196 if (!at_least_one_arg(as, error))
2199 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2200 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2201 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2202 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2203 *error = "Invalid data block size";
2207 if (block_size > ca->cache_sectors) {
2208 *error = "Data block size is larger than the cache device";
2212 ca->block_size = block_size;
2217 static void init_features(struct cache_features *cf)
2219 cf->mode = CM_WRITE;
2220 cf->io_mode = CM_IO_WRITEBACK;
2221 cf->metadata_version = 1;
2222 cf->discard_passdown = true;
2225 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2228 static const struct dm_arg _args[] = {
2229 {0, 3, "Invalid number of cache feature arguments"},
2232 int r, mode_ctr = 0;
2235 struct cache_features *cf = &ca->features;
2239 r = dm_read_arg_group(_args, as, &argc, error);
2244 arg = dm_shift_arg(as);
2246 if (!strcasecmp(arg, "writeback")) {
2247 cf->io_mode = CM_IO_WRITEBACK;
2251 else if (!strcasecmp(arg, "writethrough")) {
2252 cf->io_mode = CM_IO_WRITETHROUGH;
2256 else if (!strcasecmp(arg, "passthrough")) {
2257 cf->io_mode = CM_IO_PASSTHROUGH;
2261 else if (!strcasecmp(arg, "metadata2"))
2262 cf->metadata_version = 2;
2264 else if (!strcasecmp(arg, "no_discard_passdown"))
2265 cf->discard_passdown = false;
2268 *error = "Unrecognised cache feature requested";
2274 *error = "Duplicate cache io_mode features requested";
2281 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2284 static const struct dm_arg _args[] = {
2285 {0, 1024, "Invalid number of policy arguments"},
2290 if (!at_least_one_arg(as, error))
2293 ca->policy_name = dm_shift_arg(as);
2295 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2299 ca->policy_argv = (const char **)as->argv;
2300 dm_consume_args(as, ca->policy_argc);
2305 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2309 struct dm_arg_set as;
2314 r = parse_metadata_dev(ca, &as, error);
2318 r = parse_cache_dev(ca, &as, error);
2322 r = parse_origin_dev(ca, &as, error);
2326 r = parse_block_size(ca, &as, error);
2330 r = parse_features(ca, &as, error);
2334 r = parse_policy(ca, &as, error);
2341 /*----------------------------------------------------------------*/
2343 static struct kmem_cache *migration_cache;
2345 #define NOT_CORE_OPTION 1
2347 static int process_config_option(struct cache *cache, const char *key, const char *value)
2351 if (!strcasecmp(key, "migration_threshold")) {
2352 if (kstrtoul(value, 10, &tmp))
2355 cache->migration_threshold = tmp;
2359 return NOT_CORE_OPTION;
2362 static int set_config_value(struct cache *cache, const char *key, const char *value)
2364 int r = process_config_option(cache, key, value);
2366 if (r == NOT_CORE_OPTION)
2367 r = policy_set_config_value(cache->policy, key, value);
2370 DMWARN("bad config value for %s: %s", key, value);
2375 static int set_config_values(struct cache *cache, int argc, const char **argv)
2380 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2385 r = set_config_value(cache, argv[0], argv[1]);
2396 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2399 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2401 cache->origin_sectors,
2402 cache->sectors_per_block);
2404 *error = "Error creating cache's policy";
2408 BUG_ON(!cache->policy);
2414 * We want the discard block size to be at least the size of the cache
2415 * block size and have no more than 2^14 discard blocks across the origin.
2417 #define MAX_DISCARD_BLOCKS (1 << 14)
2419 static bool too_many_discard_blocks(sector_t discard_block_size,
2420 sector_t origin_size)
2422 (void) sector_div(origin_size, discard_block_size);
2424 return origin_size > MAX_DISCARD_BLOCKS;
2427 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2428 sector_t origin_size)
2430 sector_t discard_block_size = cache_block_size;
2433 while (too_many_discard_blocks(discard_block_size, origin_size))
2434 discard_block_size *= 2;
2436 return discard_block_size;
2439 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2441 dm_block_t nr_blocks = from_cblock(size);
2443 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2444 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2445 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2446 "Please consider increasing the cache block size to reduce the overall cache block count.",
2447 (unsigned long long) nr_blocks);
2449 cache->cache_size = size;
2452 static int is_congested(struct dm_dev *dev, int bdi_bits)
2454 struct request_queue *q = bdev_get_queue(dev->bdev);
2455 return bdi_congested(q->backing_dev_info, bdi_bits);
2458 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2460 struct cache *cache = container_of(cb, struct cache, callbacks);
2462 return is_congested(cache->origin_dev, bdi_bits) ||
2463 is_congested(cache->cache_dev, bdi_bits);
2466 #define DEFAULT_MIGRATION_THRESHOLD 2048
2468 static int cache_create(struct cache_args *ca, struct cache **result)
2471 char **error = &ca->ti->error;
2472 struct cache *cache;
2473 struct dm_target *ti = ca->ti;
2474 dm_block_t origin_blocks;
2475 struct dm_cache_metadata *cmd;
2476 bool may_format = ca->features.mode == CM_WRITE;
2478 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2483 ti->private = cache;
2484 ti->num_flush_bios = 2;
2485 ti->flush_supported = true;
2487 ti->num_discard_bios = 1;
2488 ti->discards_supported = true;
2490 ti->per_io_data_size = sizeof(struct per_bio_data);
2492 cache->features = ca->features;
2493 if (writethrough_mode(cache)) {
2494 /* Create bioset for writethrough bios issued to origin */
2495 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2500 cache->callbacks.congested_fn = cache_is_congested;
2501 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2503 cache->metadata_dev = ca->metadata_dev;
2504 cache->origin_dev = ca->origin_dev;
2505 cache->cache_dev = ca->cache_dev;
2507 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2509 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2510 origin_blocks = block_div(origin_blocks, ca->block_size);
2511 cache->origin_blocks = to_oblock(origin_blocks);
2513 cache->sectors_per_block = ca->block_size;
2514 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2519 if (ca->block_size & (ca->block_size - 1)) {
2520 dm_block_t cache_size = ca->cache_sectors;
2522 cache->sectors_per_block_shift = -1;
2523 cache_size = block_div(cache_size, ca->block_size);
2524 set_cache_size(cache, to_cblock(cache_size));
2526 cache->sectors_per_block_shift = __ffs(ca->block_size);
2527 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2530 r = create_cache_policy(cache, ca, error);
2534 cache->policy_nr_args = ca->policy_argc;
2535 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2537 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2539 *error = "Error setting cache policy's config values";
2543 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2544 ca->block_size, may_format,
2545 dm_cache_policy_get_hint_size(cache->policy),
2546 ca->features.metadata_version);
2548 *error = "Error creating metadata object";
2553 set_cache_mode(cache, CM_WRITE);
2554 if (get_cache_mode(cache) != CM_WRITE) {
2555 *error = "Unable to get write access to metadata, please check/repair metadata.";
2560 if (passthrough_mode(cache)) {
2563 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2565 *error = "dm_cache_metadata_all_clean() failed";
2570 *error = "Cannot enter passthrough mode unless all blocks are clean";
2575 policy_allow_migrations(cache->policy, false);
2578 spin_lock_init(&cache->lock);
2579 bio_list_init(&cache->deferred_bios);
2580 atomic_set(&cache->nr_allocated_migrations, 0);
2581 atomic_set(&cache->nr_io_migrations, 0);
2582 init_waitqueue_head(&cache->migration_wait);
2585 atomic_set(&cache->nr_dirty, 0);
2586 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2587 if (!cache->dirty_bitset) {
2588 *error = "could not allocate dirty bitset";
2591 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2593 cache->discard_block_size =
2594 calculate_discard_block_size(cache->sectors_per_block,
2595 cache->origin_sectors);
2596 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2597 cache->discard_block_size));
2598 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2599 if (!cache->discard_bitset) {
2600 *error = "could not allocate discard bitset";
2603 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2605 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2606 if (IS_ERR(cache->copier)) {
2607 *error = "could not create kcopyd client";
2608 r = PTR_ERR(cache->copier);
2612 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2614 *error = "could not create workqueue for metadata object";
2617 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2618 INIT_WORK(&cache->migration_worker, check_migrations);
2619 INIT_DELAYED_WORK(&cache->waker, do_waker);
2621 cache->prison = dm_bio_prison_create_v2(cache->wq);
2622 if (!cache->prison) {
2623 *error = "could not create bio prison";
2627 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2630 *error = "Error creating cache's migration mempool";
2634 cache->need_tick_bio = true;
2635 cache->sized = false;
2636 cache->invalidate = false;
2637 cache->commit_requested = false;
2638 cache->loaded_mappings = false;
2639 cache->loaded_discards = false;
2643 atomic_set(&cache->stats.demotion, 0);
2644 atomic_set(&cache->stats.promotion, 0);
2645 atomic_set(&cache->stats.copies_avoided, 0);
2646 atomic_set(&cache->stats.cache_cell_clash, 0);
2647 atomic_set(&cache->stats.commit_count, 0);
2648 atomic_set(&cache->stats.discard_count, 0);
2650 spin_lock_init(&cache->invalidation_lock);
2651 INIT_LIST_HEAD(&cache->invalidation_requests);
2653 batcher_init(&cache->committer, commit_op, cache,
2654 issue_op, cache, cache->wq);
2655 iot_init(&cache->tracker);
2657 init_rwsem(&cache->background_work_lock);
2658 prevent_background_work(cache);
2667 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2672 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2675 for (i = 0; i < argc; i++) {
2676 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2685 cache->nr_ctr_args = argc;
2686 cache->ctr_args = copy;
2691 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2694 struct cache_args *ca;
2695 struct cache *cache = NULL;
2697 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2699 ti->error = "Error allocating memory for cache";
2704 r = parse_cache_args(ca, argc, argv, &ti->error);
2708 r = cache_create(ca, &cache);
2712 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2718 ti->private = cache;
2720 destroy_cache_args(ca);
2724 /*----------------------------------------------------------------*/
2726 static int cache_map(struct dm_target *ti, struct bio *bio)
2728 struct cache *cache = ti->private;
2732 dm_oblock_t block = get_bio_block(cache, bio);
2734 init_per_bio_data(bio);
2735 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2737 * This can only occur if the io goes to a partial block at
2738 * the end of the origin device. We don't cache these.
2739 * Just remap to the origin and carry on.
2741 remap_to_origin(cache, bio);
2742 accounted_begin(cache, bio);
2743 return DM_MAPIO_REMAPPED;
2746 if (discard_or_flush(bio)) {
2747 defer_bio(cache, bio);
2748 return DM_MAPIO_SUBMITTED;
2751 r = map_bio(cache, bio, block, &commit_needed);
2753 schedule_commit(&cache->committer);
2758 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2760 struct cache *cache = ti->private;
2761 unsigned long flags;
2762 struct per_bio_data *pb = get_per_bio_data(bio);
2765 policy_tick(cache->policy, false);
2767 spin_lock_irqsave(&cache->lock, flags);
2768 cache->need_tick_bio = true;
2769 spin_unlock_irqrestore(&cache->lock, flags);
2772 bio_drop_shared_lock(cache, bio);
2773 accounted_complete(cache, bio);
2775 return DM_ENDIO_DONE;
2778 static int write_dirty_bitset(struct cache *cache)
2782 if (get_cache_mode(cache) >= CM_READ_ONLY)
2785 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2787 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2792 static int write_discard_bitset(struct cache *cache)
2796 if (get_cache_mode(cache) >= CM_READ_ONLY)
2799 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2800 cache->discard_nr_blocks);
2802 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2803 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2807 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2808 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2809 is_discarded(cache, to_dblock(i)));
2811 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2819 static int write_hints(struct cache *cache)
2823 if (get_cache_mode(cache) >= CM_READ_ONLY)
2826 r = dm_cache_write_hints(cache->cmd, cache->policy);
2828 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2836 * returns true on success
2838 static bool sync_metadata(struct cache *cache)
2842 r1 = write_dirty_bitset(cache);
2844 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2846 r2 = write_discard_bitset(cache);
2848 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2852 r3 = write_hints(cache);
2854 DMERR("%s: could not write hints", cache_device_name(cache));
2857 * If writing the above metadata failed, we still commit, but don't
2858 * set the clean shutdown flag. This will effectively force every
2859 * dirty bit to be set on reload.
2861 r4 = commit(cache, !r1 && !r2 && !r3);
2863 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2865 return !r1 && !r2 && !r3 && !r4;
2868 static void cache_postsuspend(struct dm_target *ti)
2870 struct cache *cache = ti->private;
2872 prevent_background_work(cache);
2873 BUG_ON(atomic_read(&cache->nr_io_migrations));
2875 cancel_delayed_work_sync(&cache->waker);
2876 drain_workqueue(cache->wq);
2877 WARN_ON(cache->tracker.in_flight);
2880 * If it's a flush suspend there won't be any deferred bios, so this
2883 requeue_deferred_bios(cache);
2885 if (get_cache_mode(cache) == CM_WRITE)
2886 (void) sync_metadata(cache);
2889 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2890 bool dirty, uint32_t hint, bool hint_valid)
2893 struct cache *cache = context;
2896 set_bit(from_cblock(cblock), cache->dirty_bitset);
2897 atomic_inc(&cache->nr_dirty);
2899 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2901 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2909 * The discard block size in the on disk metadata is not
2910 * neccessarily the same as we're currently using. So we have to
2911 * be careful to only set the discarded attribute if we know it
2912 * covers a complete block of the new size.
2914 struct discard_load_info {
2915 struct cache *cache;
2918 * These blocks are sized using the on disk dblock size, rather
2919 * than the current one.
2921 dm_block_t block_size;
2922 dm_block_t discard_begin, discard_end;
2925 static void discard_load_info_init(struct cache *cache,
2926 struct discard_load_info *li)
2929 li->discard_begin = li->discard_end = 0;
2932 static void set_discard_range(struct discard_load_info *li)
2936 if (li->discard_begin == li->discard_end)
2940 * Convert to sectors.
2942 b = li->discard_begin * li->block_size;
2943 e = li->discard_end * li->block_size;
2946 * Then convert back to the current dblock size.
2948 b = dm_sector_div_up(b, li->cache->discard_block_size);
2949 sector_div(e, li->cache->discard_block_size);
2952 * The origin may have shrunk, so we need to check we're still in
2955 if (e > from_dblock(li->cache->discard_nr_blocks))
2956 e = from_dblock(li->cache->discard_nr_blocks);
2959 set_discard(li->cache, to_dblock(b));
2962 static int load_discard(void *context, sector_t discard_block_size,
2963 dm_dblock_t dblock, bool discard)
2965 struct discard_load_info *li = context;
2967 li->block_size = discard_block_size;
2970 if (from_dblock(dblock) == li->discard_end)
2972 * We're already in a discard range, just extend it.
2974 li->discard_end = li->discard_end + 1ULL;
2978 * Emit the old range and start a new one.
2980 set_discard_range(li);
2981 li->discard_begin = from_dblock(dblock);
2982 li->discard_end = li->discard_begin + 1ULL;
2985 set_discard_range(li);
2986 li->discard_begin = li->discard_end = 0;
2992 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2994 sector_t size = get_dev_size(cache->cache_dev);
2995 (void) sector_div(size, cache->sectors_per_block);
2996 return to_cblock(size);
2999 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3001 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
3003 DMERR("%s: unable to extend cache due to missing cache table reload",
3004 cache_device_name(cache));
3010 * We can't drop a dirty block when shrinking the cache.
3012 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3013 new_size = to_cblock(from_cblock(new_size) + 1);
3014 if (is_dirty(cache, new_size)) {
3015 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3016 cache_device_name(cache),
3017 (unsigned long long) from_cblock(new_size));
3025 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3029 r = dm_cache_resize(cache->cmd, new_size);
3031 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3032 metadata_operation_failed(cache, "dm_cache_resize", r);
3036 set_cache_size(cache, new_size);
3041 static int cache_preresume(struct dm_target *ti)
3044 struct cache *cache = ti->private;
3045 dm_cblock_t csize = get_cache_dev_size(cache);
3048 * Check to see if the cache has resized.
3050 if (!cache->sized) {
3051 r = resize_cache_dev(cache, csize);
3055 cache->sized = true;
3057 } else if (csize != cache->cache_size) {
3058 if (!can_resize(cache, csize))
3061 r = resize_cache_dev(cache, csize);
3066 if (!cache->loaded_mappings) {
3067 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3068 load_mapping, cache);
3070 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3071 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3075 cache->loaded_mappings = true;
3078 if (!cache->loaded_discards) {
3079 struct discard_load_info li;
3082 * The discard bitset could have been resized, or the
3083 * discard block size changed. To be safe we start by
3084 * setting every dblock to not discarded.
3086 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3088 discard_load_info_init(cache, &li);
3089 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3091 DMERR("%s: could not load origin discards", cache_device_name(cache));
3092 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3095 set_discard_range(&li);
3097 cache->loaded_discards = true;
3103 static void cache_resume(struct dm_target *ti)
3105 struct cache *cache = ti->private;
3107 cache->need_tick_bio = true;
3108 allow_background_work(cache);
3109 do_waker(&cache->waker.work);
3112 static void emit_flags(struct cache *cache, char *result,
3113 unsigned maxlen, ssize_t *sz_ptr)
3115 ssize_t sz = *sz_ptr;
3116 struct cache_features *cf = &cache->features;
3117 unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3119 DMEMIT("%u ", count);
3121 if (cf->metadata_version == 2)
3122 DMEMIT("metadata2 ");
3124 if (writethrough_mode(cache))
3125 DMEMIT("writethrough ");
3127 else if (passthrough_mode(cache))
3128 DMEMIT("passthrough ");
3130 else if (writeback_mode(cache))
3131 DMEMIT("writeback ");
3135 DMERR("%s: internal error: unknown io mode: %d",
3136 cache_device_name(cache), (int) cf->io_mode);
3139 if (!cf->discard_passdown)
3140 DMEMIT("no_discard_passdown ");
3148 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3149 * <cache block size> <#used cache blocks>/<#total cache blocks>
3150 * <#read hits> <#read misses> <#write hits> <#write misses>
3151 * <#demotions> <#promotions> <#dirty>
3152 * <#features> <features>*
3153 * <#core args> <core args>
3154 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3156 static void cache_status(struct dm_target *ti, status_type_t type,
3157 unsigned status_flags, char *result, unsigned maxlen)
3162 dm_block_t nr_free_blocks_metadata = 0;
3163 dm_block_t nr_blocks_metadata = 0;
3164 char buf[BDEVNAME_SIZE];
3165 struct cache *cache = ti->private;
3166 dm_cblock_t residency;
3170 case STATUSTYPE_INFO:
3171 if (get_cache_mode(cache) == CM_FAIL) {
3176 /* Commit to ensure statistics aren't out-of-date */
3177 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3178 (void) commit(cache, false);
3180 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3182 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3183 cache_device_name(cache), r);
3187 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3189 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3190 cache_device_name(cache), r);
3194 residency = policy_residency(cache->policy);
3196 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3197 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3198 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3199 (unsigned long long)nr_blocks_metadata,
3200 (unsigned long long)cache->sectors_per_block,
3201 (unsigned long long) from_cblock(residency),
3202 (unsigned long long) from_cblock(cache->cache_size),
3203 (unsigned) atomic_read(&cache->stats.read_hit),
3204 (unsigned) atomic_read(&cache->stats.read_miss),
3205 (unsigned) atomic_read(&cache->stats.write_hit),
3206 (unsigned) atomic_read(&cache->stats.write_miss),
3207 (unsigned) atomic_read(&cache->stats.demotion),
3208 (unsigned) atomic_read(&cache->stats.promotion),
3209 (unsigned long) atomic_read(&cache->nr_dirty));
3211 emit_flags(cache, result, maxlen, &sz);
3213 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3215 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3217 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3219 DMERR("%s: policy_emit_config_values returned %d",
3220 cache_device_name(cache), r);
3223 if (get_cache_mode(cache) == CM_READ_ONLY)
3228 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3230 if (r || needs_check)
3231 DMEMIT("needs_check ");
3237 case STATUSTYPE_TABLE:
3238 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3240 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3242 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3245 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3246 DMEMIT(" %s", cache->ctr_args[i]);
3247 if (cache->nr_ctr_args)
3248 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3258 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3259 * the one-past-the-end value.
3261 struct cblock_range {
3267 * A cache block range can take two forms:
3269 * i) A single cblock, eg. '3456'
3270 * ii) A begin and end cblock with a dash between, eg. 123-234
3272 static int parse_cblock_range(struct cache *cache, const char *str,
3273 struct cblock_range *result)
3280 * Try and parse form (ii) first.
3282 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3287 result->begin = to_cblock(b);
3288 result->end = to_cblock(e);
3293 * That didn't work, try form (i).
3295 r = sscanf(str, "%llu%c", &b, &dummy);
3300 result->begin = to_cblock(b);
3301 result->end = to_cblock(from_cblock(result->begin) + 1u);
3305 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3309 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3311 uint64_t b = from_cblock(range->begin);
3312 uint64_t e = from_cblock(range->end);
3313 uint64_t n = from_cblock(cache->cache_size);
3316 DMERR("%s: begin cblock out of range: %llu >= %llu",
3317 cache_device_name(cache), b, n);
3322 DMERR("%s: end cblock out of range: %llu > %llu",
3323 cache_device_name(cache), e, n);
3328 DMERR("%s: invalid cblock range: %llu >= %llu",
3329 cache_device_name(cache), b, e);
3336 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3338 return to_cblock(from_cblock(b) + 1);
3341 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3346 * We don't need to do any locking here because we know we're in
3347 * passthrough mode. There's is potential for a race between an
3348 * invalidation triggered by an io and an invalidation message. This
3349 * is harmless, we must not worry if the policy call fails.
3351 while (range->begin != range->end) {
3352 r = invalidate_cblock(cache, range->begin);
3356 range->begin = cblock_succ(range->begin);
3359 cache->commit_requested = true;
3363 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3364 const char **cblock_ranges)
3368 struct cblock_range range;
3370 if (!passthrough_mode(cache)) {
3371 DMERR("%s: cache has to be in passthrough mode for invalidation",
3372 cache_device_name(cache));
3376 for (i = 0; i < count; i++) {
3377 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3381 r = validate_cblock_range(cache, &range);
3386 * Pass begin and end origin blocks to the worker and wake it.
3388 r = request_invalidation(cache, &range);
3400 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3402 * The key migration_threshold is supported by the cache target core.
3404 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3405 char *result, unsigned maxlen)
3407 struct cache *cache = ti->private;
3412 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3413 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3414 cache_device_name(cache));
3418 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3419 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3424 return set_config_value(cache, argv[0], argv[1]);
3427 static int cache_iterate_devices(struct dm_target *ti,
3428 iterate_devices_callout_fn fn, void *data)
3431 struct cache *cache = ti->private;
3433 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3435 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3440 static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3442 struct request_queue *q = bdev_get_queue(origin_bdev);
3444 return q && blk_queue_discard(q);
3448 * If discard_passdown was enabled verify that the origin device
3449 * supports discards. Disable discard_passdown if not.
3451 static void disable_passdown_if_not_supported(struct cache *cache)
3453 struct block_device *origin_bdev = cache->origin_dev->bdev;
3454 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3455 const char *reason = NULL;
3456 char buf[BDEVNAME_SIZE];
3458 if (!cache->features.discard_passdown)
3461 if (!origin_dev_supports_discard(origin_bdev))
3462 reason = "discard unsupported";
3464 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3465 reason = "max discard sectors smaller than a block";
3468 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3469 bdevname(origin_bdev, buf), reason);
3470 cache->features.discard_passdown = false;
3474 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3476 struct block_device *origin_bdev = cache->origin_dev->bdev;
3477 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3479 if (!cache->features.discard_passdown) {
3480 /* No passdown is done so setting own virtual limits */
3481 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3482 cache->origin_sectors);
3483 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3488 * cache_iterate_devices() is stacking both origin and fast device limits
3489 * but discards aren't passed to fast device, so inherit origin's limits.
3491 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3492 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3493 limits->discard_granularity = origin_limits->discard_granularity;
3494 limits->discard_alignment = origin_limits->discard_alignment;
3495 limits->discard_misaligned = origin_limits->discard_misaligned;
3498 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3500 struct cache *cache = ti->private;
3501 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3504 * If the system-determined stacked limits are compatible with the
3505 * cache's blocksize (io_opt is a factor) do not override them.
3507 if (io_opt_sectors < cache->sectors_per_block ||
3508 do_div(io_opt_sectors, cache->sectors_per_block)) {
3509 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3510 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3513 disable_passdown_if_not_supported(cache);
3514 set_discard_limits(cache, limits);
3517 /*----------------------------------------------------------------*/
3519 static struct target_type cache_target = {
3521 .version = {2, 1, 0},
3522 .module = THIS_MODULE,
3526 .end_io = cache_end_io,
3527 .postsuspend = cache_postsuspend,
3528 .preresume = cache_preresume,
3529 .resume = cache_resume,
3530 .status = cache_status,
3531 .message = cache_message,
3532 .iterate_devices = cache_iterate_devices,
3533 .io_hints = cache_io_hints,
3536 static int __init dm_cache_init(void)
3540 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3541 if (!migration_cache)
3544 r = dm_register_target(&cache_target);
3546 DMERR("cache target registration failed: %d", r);
3547 kmem_cache_destroy(migration_cache);
3554 static void __exit dm_cache_exit(void)
3556 dm_unregister_target(&cache_target);
3557 kmem_cache_destroy(migration_cache);
3560 module_init(dm_cache_init);
3561 module_exit(dm_cache_exit);
3563 MODULE_DESCRIPTION(DM_NAME " cache target");
3564 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3565 MODULE_LICENSE("GPL");