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;
366 atomic_t copies_avoided;
367 atomic_t cache_cell_clash;
368 atomic_t commit_count;
369 atomic_t discard_count;
373 struct dm_target *ti;
377 * Fields for converting from sectors to blocks.
379 int sectors_per_block_shift;
380 sector_t sectors_per_block;
382 struct dm_cache_metadata *cmd;
385 * Metadata is written to this device.
387 struct dm_dev *metadata_dev;
390 * The slower of the two data devices. Typically a spindle.
392 struct dm_dev *origin_dev;
395 * The faster of the two data devices. Typically an SSD.
397 struct dm_dev *cache_dev;
400 * Size of the origin device in _complete_ blocks and native sectors.
402 dm_oblock_t origin_blocks;
403 sector_t origin_sectors;
406 * Size of the cache device in blocks.
408 dm_cblock_t cache_size;
411 * Invalidation fields.
413 spinlock_t invalidation_lock;
414 struct list_head invalidation_requests;
416 sector_t migration_threshold;
417 wait_queue_head_t migration_wait;
418 atomic_t nr_allocated_migrations;
421 * The number of in flight migrations that are performing
422 * background io. eg, promotion, writeback.
424 atomic_t nr_io_migrations;
426 struct bio_list deferred_bios;
428 struct rw_semaphore quiesce_lock;
430 struct dm_target_callbacks callbacks;
433 * origin_blocks entries, discarded if set.
435 dm_dblock_t discard_nr_blocks;
436 unsigned long *discard_bitset;
437 uint32_t discard_block_size; /* a power of 2 times sectors per block */
440 * Rather than reconstructing the table line for the status we just
441 * save it and regurgitate.
443 unsigned nr_ctr_args;
444 const char **ctr_args;
446 struct dm_kcopyd_client *copier;
447 struct work_struct deferred_bio_worker;
448 struct work_struct migration_worker;
449 struct workqueue_struct *wq;
450 struct delayed_work waker;
451 struct dm_bio_prison_v2 *prison;
454 * cache_size entries, dirty if set
456 unsigned long *dirty_bitset;
459 unsigned policy_nr_args;
460 struct dm_cache_policy *policy;
463 * Cache features such as write-through.
465 struct cache_features features;
467 struct cache_stats stats;
469 bool need_tick_bio:1;
472 bool commit_requested:1;
473 bool loaded_mappings:1;
474 bool loaded_discards:1;
476 struct rw_semaphore background_work_lock;
478 struct batcher committer;
479 struct work_struct commit_ws;
481 struct io_tracker tracker;
483 mempool_t migration_pool;
488 struct per_bio_data {
491 struct dm_bio_prison_cell_v2 *cell;
492 struct dm_hook_info hook_info;
496 struct dm_cache_migration {
497 struct continuation k;
500 struct policy_work *op;
501 struct bio *overwrite_bio;
502 struct dm_bio_prison_cell_v2 *cell;
504 dm_cblock_t invalidate_cblock;
505 dm_oblock_t invalidate_oblock;
508 /*----------------------------------------------------------------*/
510 static bool writethrough_mode(struct cache *cache)
512 return cache->features.io_mode == CM_IO_WRITETHROUGH;
515 static bool writeback_mode(struct cache *cache)
517 return cache->features.io_mode == CM_IO_WRITEBACK;
520 static inline bool passthrough_mode(struct cache *cache)
522 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
525 /*----------------------------------------------------------------*/
527 static void wake_deferred_bio_worker(struct cache *cache)
529 queue_work(cache->wq, &cache->deferred_bio_worker);
532 static void wake_migration_worker(struct cache *cache)
534 if (passthrough_mode(cache))
537 queue_work(cache->wq, &cache->migration_worker);
540 /*----------------------------------------------------------------*/
542 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
544 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
547 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
549 dm_bio_prison_free_cell_v2(cache->prison, cell);
552 static struct dm_cache_migration *alloc_migration(struct cache *cache)
554 struct dm_cache_migration *mg;
556 mg = mempool_alloc(&cache->migration_pool, GFP_NOWAIT);
560 memset(mg, 0, sizeof(*mg));
563 atomic_inc(&cache->nr_allocated_migrations);
568 static void free_migration(struct dm_cache_migration *mg)
570 struct cache *cache = mg->cache;
572 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
573 wake_up(&cache->migration_wait);
575 mempool_free(mg, &cache->migration_pool);
578 /*----------------------------------------------------------------*/
580 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
582 return to_oblock(from_oblock(b) + 1ull);
585 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
589 key->block_begin = from_oblock(begin);
590 key->block_end = from_oblock(end);
594 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
595 * level 1 which prevents *both* READs and WRITEs.
597 #define WRITE_LOCK_LEVEL 0
598 #define READ_WRITE_LOCK_LEVEL 1
600 static unsigned lock_level(struct bio *bio)
602 return bio_data_dir(bio) == WRITE ?
604 READ_WRITE_LOCK_LEVEL;
607 /*----------------------------------------------------------------
609 *--------------------------------------------------------------*/
611 static struct per_bio_data *get_per_bio_data(struct bio *bio)
613 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
618 static struct per_bio_data *init_per_bio_data(struct bio *bio)
620 struct per_bio_data *pb = get_per_bio_data(bio);
623 pb->req_nr = dm_bio_get_target_bio_nr(bio);
630 /*----------------------------------------------------------------*/
632 static void defer_bio(struct cache *cache, struct bio *bio)
636 spin_lock_irqsave(&cache->lock, flags);
637 bio_list_add(&cache->deferred_bios, bio);
638 spin_unlock_irqrestore(&cache->lock, flags);
640 wake_deferred_bio_worker(cache);
643 static void defer_bios(struct cache *cache, struct bio_list *bios)
647 spin_lock_irqsave(&cache->lock, flags);
648 bio_list_merge(&cache->deferred_bios, bios);
650 spin_unlock_irqrestore(&cache->lock, flags);
652 wake_deferred_bio_worker(cache);
655 /*----------------------------------------------------------------*/
657 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
660 struct per_bio_data *pb;
661 struct dm_cell_key_v2 key;
662 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
663 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
665 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
666 if (!cell_prealloc) {
667 defer_bio(cache, bio);
671 build_key(oblock, end, &key);
672 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
675 * Failed to get the lock.
677 free_prison_cell(cache, cell_prealloc);
681 if (cell != cell_prealloc)
682 free_prison_cell(cache, cell_prealloc);
684 pb = get_per_bio_data(bio);
690 /*----------------------------------------------------------------*/
692 static bool is_dirty(struct cache *cache, dm_cblock_t b)
694 return test_bit(from_cblock(b), cache->dirty_bitset);
697 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
699 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
700 atomic_inc(&cache->nr_dirty);
701 policy_set_dirty(cache->policy, cblock);
706 * These two are called when setting after migrations to force the policy
707 * and dirty bitset to be in sync.
709 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
711 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
712 atomic_inc(&cache->nr_dirty);
713 policy_set_dirty(cache->policy, cblock);
716 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
718 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
719 if (atomic_dec_return(&cache->nr_dirty) == 0)
720 dm_table_event(cache->ti->table);
723 policy_clear_dirty(cache->policy, cblock);
726 /*----------------------------------------------------------------*/
728 static bool block_size_is_power_of_two(struct cache *cache)
730 return cache->sectors_per_block_shift >= 0;
733 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
734 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
737 static dm_block_t block_div(dm_block_t b, uint32_t n)
744 static dm_block_t oblocks_per_dblock(struct cache *cache)
746 dm_block_t oblocks = cache->discard_block_size;
748 if (block_size_is_power_of_two(cache))
749 oblocks >>= cache->sectors_per_block_shift;
751 oblocks = block_div(oblocks, cache->sectors_per_block);
756 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
758 return to_dblock(block_div(from_oblock(oblock),
759 oblocks_per_dblock(cache)));
762 static void set_discard(struct cache *cache, dm_dblock_t b)
766 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
767 atomic_inc(&cache->stats.discard_count);
769 spin_lock_irqsave(&cache->lock, flags);
770 set_bit(from_dblock(b), cache->discard_bitset);
771 spin_unlock_irqrestore(&cache->lock, flags);
774 static void clear_discard(struct cache *cache, dm_dblock_t b)
778 spin_lock_irqsave(&cache->lock, flags);
779 clear_bit(from_dblock(b), cache->discard_bitset);
780 spin_unlock_irqrestore(&cache->lock, flags);
783 static bool is_discarded(struct cache *cache, dm_dblock_t b)
788 spin_lock_irqsave(&cache->lock, flags);
789 r = test_bit(from_dblock(b), cache->discard_bitset);
790 spin_unlock_irqrestore(&cache->lock, flags);
795 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
800 spin_lock_irqsave(&cache->lock, flags);
801 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
802 cache->discard_bitset);
803 spin_unlock_irqrestore(&cache->lock, flags);
808 /*----------------------------------------------------------------
810 *--------------------------------------------------------------*/
811 static void remap_to_origin(struct cache *cache, struct bio *bio)
813 bio_set_dev(bio, cache->origin_dev->bdev);
816 static void remap_to_cache(struct cache *cache, struct bio *bio,
819 sector_t bi_sector = bio->bi_iter.bi_sector;
820 sector_t block = from_cblock(cblock);
822 bio_set_dev(bio, cache->cache_dev->bdev);
823 if (!block_size_is_power_of_two(cache))
824 bio->bi_iter.bi_sector =
825 (block * cache->sectors_per_block) +
826 sector_div(bi_sector, cache->sectors_per_block);
828 bio->bi_iter.bi_sector =
829 (block << cache->sectors_per_block_shift) |
830 (bi_sector & (cache->sectors_per_block - 1));
833 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
836 struct per_bio_data *pb;
838 spin_lock_irqsave(&cache->lock, flags);
839 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
840 bio_op(bio) != REQ_OP_DISCARD) {
841 pb = get_per_bio_data(bio);
843 cache->need_tick_bio = false;
845 spin_unlock_irqrestore(&cache->lock, flags);
848 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
849 dm_oblock_t oblock, bool bio_has_pbd)
852 check_if_tick_bio_needed(cache, bio);
853 remap_to_origin(cache, bio);
854 if (bio_data_dir(bio) == WRITE)
855 clear_discard(cache, oblock_to_dblock(cache, oblock));
858 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
861 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
862 __remap_to_origin_clear_discard(cache, bio, oblock, true);
865 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
866 dm_oblock_t oblock, dm_cblock_t cblock)
868 check_if_tick_bio_needed(cache, bio);
869 remap_to_cache(cache, bio, cblock);
870 if (bio_data_dir(bio) == WRITE) {
871 set_dirty(cache, cblock);
872 clear_discard(cache, oblock_to_dblock(cache, oblock));
876 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
878 sector_t block_nr = bio->bi_iter.bi_sector;
880 if (!block_size_is_power_of_two(cache))
881 (void) sector_div(block_nr, cache->sectors_per_block);
883 block_nr >>= cache->sectors_per_block_shift;
885 return to_oblock(block_nr);
888 static bool accountable_bio(struct cache *cache, struct bio *bio)
890 return bio_op(bio) != REQ_OP_DISCARD;
893 static void accounted_begin(struct cache *cache, struct bio *bio)
895 struct per_bio_data *pb;
897 if (accountable_bio(cache, bio)) {
898 pb = get_per_bio_data(bio);
899 pb->len = bio_sectors(bio);
900 iot_io_begin(&cache->tracker, pb->len);
904 static void accounted_complete(struct cache *cache, struct bio *bio)
906 struct per_bio_data *pb = get_per_bio_data(bio);
908 iot_io_end(&cache->tracker, pb->len);
911 static void accounted_request(struct cache *cache, struct bio *bio)
913 accounted_begin(cache, bio);
914 generic_make_request(bio);
917 static void issue_op(struct bio *bio, void *context)
919 struct cache *cache = context;
920 accounted_request(cache, bio);
924 * When running in writethrough mode we need to send writes to clean blocks
925 * to both the cache and origin devices. Clone the bio and send them in parallel.
927 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
928 dm_oblock_t oblock, dm_cblock_t cblock)
930 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
934 bio_chain(origin_bio, bio);
936 * Passing false to __remap_to_origin_clear_discard() skips
937 * all code that might use per_bio_data (since clone doesn't have it)
939 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
940 submit_bio(origin_bio);
942 remap_to_cache(cache, bio, cblock);
945 /*----------------------------------------------------------------
947 *--------------------------------------------------------------*/
948 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
950 return cache->features.mode;
953 static const char *cache_device_name(struct cache *cache)
955 return dm_device_name(dm_table_get_md(cache->ti->table));
958 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
960 const char *descs[] = {
966 dm_table_event(cache->ti->table);
967 DMINFO("%s: switching cache to %s mode",
968 cache_device_name(cache), descs[(int)mode]);
971 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
974 enum cache_metadata_mode old_mode = get_cache_mode(cache);
976 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
977 DMERR("%s: unable to read needs_check flag, setting failure mode.",
978 cache_device_name(cache));
982 if (new_mode == CM_WRITE && needs_check) {
983 DMERR("%s: unable to switch cache to write mode until repaired.",
984 cache_device_name(cache));
985 if (old_mode != new_mode)
988 new_mode = CM_READ_ONLY;
991 /* Never move out of fail mode */
992 if (old_mode == CM_FAIL)
998 dm_cache_metadata_set_read_only(cache->cmd);
1002 dm_cache_metadata_set_read_write(cache->cmd);
1006 cache->features.mode = new_mode;
1008 if (new_mode != old_mode)
1009 notify_mode_switch(cache, new_mode);
1012 static void abort_transaction(struct cache *cache)
1014 const char *dev_name = cache_device_name(cache);
1016 if (get_cache_mode(cache) >= CM_READ_ONLY)
1019 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1020 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1021 set_cache_mode(cache, CM_FAIL);
1024 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1025 if (dm_cache_metadata_abort(cache->cmd)) {
1026 DMERR("%s: failed to abort metadata transaction", dev_name);
1027 set_cache_mode(cache, CM_FAIL);
1031 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1033 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1034 cache_device_name(cache), op, r);
1035 abort_transaction(cache);
1036 set_cache_mode(cache, CM_READ_ONLY);
1039 /*----------------------------------------------------------------*/
1041 static void load_stats(struct cache *cache)
1043 struct dm_cache_statistics stats;
1045 dm_cache_metadata_get_stats(cache->cmd, &stats);
1046 atomic_set(&cache->stats.read_hit, stats.read_hits);
1047 atomic_set(&cache->stats.read_miss, stats.read_misses);
1048 atomic_set(&cache->stats.write_hit, stats.write_hits);
1049 atomic_set(&cache->stats.write_miss, stats.write_misses);
1052 static void save_stats(struct cache *cache)
1054 struct dm_cache_statistics stats;
1056 if (get_cache_mode(cache) >= CM_READ_ONLY)
1059 stats.read_hits = atomic_read(&cache->stats.read_hit);
1060 stats.read_misses = atomic_read(&cache->stats.read_miss);
1061 stats.write_hits = atomic_read(&cache->stats.write_hit);
1062 stats.write_misses = atomic_read(&cache->stats.write_miss);
1064 dm_cache_metadata_set_stats(cache->cmd, &stats);
1067 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1070 case POLICY_PROMOTE:
1071 atomic_inc(&stats->promotion);
1075 atomic_inc(&stats->demotion);
1078 case POLICY_WRITEBACK:
1079 atomic_inc(&stats->writeback);
1084 /*----------------------------------------------------------------
1085 * Migration processing
1087 * Migration covers moving data from the origin device to the cache, or
1089 *--------------------------------------------------------------*/
1091 static void inc_io_migrations(struct cache *cache)
1093 atomic_inc(&cache->nr_io_migrations);
1096 static void dec_io_migrations(struct cache *cache)
1098 atomic_dec(&cache->nr_io_migrations);
1101 static bool discard_or_flush(struct bio *bio)
1103 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1106 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1107 dm_dblock_t *b, dm_dblock_t *e)
1109 sector_t sb = bio->bi_iter.bi_sector;
1110 sector_t se = bio_end_sector(bio);
1112 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1114 if (se - sb < cache->discard_block_size)
1117 *e = to_dblock(block_div(se, cache->discard_block_size));
1120 /*----------------------------------------------------------------*/
1122 static void prevent_background_work(struct cache *cache)
1125 down_write(&cache->background_work_lock);
1129 static void allow_background_work(struct cache *cache)
1132 up_write(&cache->background_work_lock);
1136 static bool background_work_begin(struct cache *cache)
1141 r = down_read_trylock(&cache->background_work_lock);
1147 static void background_work_end(struct cache *cache)
1150 up_read(&cache->background_work_lock);
1154 /*----------------------------------------------------------------*/
1156 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1158 return (bio_data_dir(bio) == WRITE) &&
1159 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1162 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1164 return writeback_mode(cache) &&
1165 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1168 static void quiesce(struct dm_cache_migration *mg,
1169 void (*continuation)(struct work_struct *))
1171 init_continuation(&mg->k, continuation);
1172 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1175 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1177 struct continuation *k = container_of(ws, struct continuation, ws);
1178 return container_of(k, struct dm_cache_migration, k);
1181 static void copy_complete(int read_err, unsigned long write_err, void *context)
1183 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1185 if (read_err || write_err)
1186 mg->k.input = BLK_STS_IOERR;
1188 queue_continuation(mg->cache->wq, &mg->k);
1191 static int copy(struct dm_cache_migration *mg, bool promote)
1194 struct dm_io_region o_region, c_region;
1195 struct cache *cache = mg->cache;
1197 o_region.bdev = cache->origin_dev->bdev;
1198 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1199 o_region.count = cache->sectors_per_block;
1201 c_region.bdev = cache->cache_dev->bdev;
1202 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1203 c_region.count = cache->sectors_per_block;
1206 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1208 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1213 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1215 struct per_bio_data *pb = get_per_bio_data(bio);
1217 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1218 free_prison_cell(cache, pb->cell);
1222 static void overwrite_endio(struct bio *bio)
1224 struct dm_cache_migration *mg = bio->bi_private;
1225 struct cache *cache = mg->cache;
1226 struct per_bio_data *pb = get_per_bio_data(bio);
1228 dm_unhook_bio(&pb->hook_info, bio);
1231 mg->k.input = bio->bi_status;
1233 queue_continuation(cache->wq, &mg->k);
1236 static void overwrite(struct dm_cache_migration *mg,
1237 void (*continuation)(struct work_struct *))
1239 struct bio *bio = mg->overwrite_bio;
1240 struct per_bio_data *pb = get_per_bio_data(bio);
1242 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1245 * The overwrite bio is part of the copy operation, as such it does
1246 * not set/clear discard or dirty flags.
1248 if (mg->op->op == POLICY_PROMOTE)
1249 remap_to_cache(mg->cache, bio, mg->op->cblock);
1251 remap_to_origin(mg->cache, bio);
1253 init_continuation(&mg->k, continuation);
1254 accounted_request(mg->cache, bio);
1260 * 1) exclusive lock preventing WRITEs
1262 * 3) copy or issue overwrite bio
1263 * 4) upgrade to exclusive lock preventing READs and WRITEs
1265 * 6) update metadata and commit
1268 static void mg_complete(struct dm_cache_migration *mg, bool success)
1270 struct bio_list bios;
1271 struct cache *cache = mg->cache;
1272 struct policy_work *op = mg->op;
1273 dm_cblock_t cblock = op->cblock;
1276 update_stats(&cache->stats, op->op);
1279 case POLICY_PROMOTE:
1280 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1281 policy_complete_background_work(cache->policy, op, success);
1283 if (mg->overwrite_bio) {
1285 force_set_dirty(cache, cblock);
1286 else if (mg->k.input)
1287 mg->overwrite_bio->bi_status = mg->k.input;
1289 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1290 bio_endio(mg->overwrite_bio);
1293 force_clear_dirty(cache, cblock);
1294 dec_io_migrations(cache);
1300 * We clear dirty here to update the nr_dirty counter.
1303 force_clear_dirty(cache, cblock);
1304 policy_complete_background_work(cache->policy, op, success);
1305 dec_io_migrations(cache);
1308 case POLICY_WRITEBACK:
1310 force_clear_dirty(cache, cblock);
1311 policy_complete_background_work(cache->policy, op, success);
1312 dec_io_migrations(cache);
1316 bio_list_init(&bios);
1318 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1319 free_prison_cell(cache, mg->cell);
1323 defer_bios(cache, &bios);
1324 wake_migration_worker(cache);
1326 background_work_end(cache);
1329 static void mg_success(struct work_struct *ws)
1331 struct dm_cache_migration *mg = ws_to_mg(ws);
1332 mg_complete(mg, mg->k.input == 0);
1335 static void mg_update_metadata(struct work_struct *ws)
1338 struct dm_cache_migration *mg = ws_to_mg(ws);
1339 struct cache *cache = mg->cache;
1340 struct policy_work *op = mg->op;
1343 case POLICY_PROMOTE:
1344 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1346 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1347 cache_device_name(cache));
1348 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1350 mg_complete(mg, false);
1353 mg_complete(mg, true);
1357 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1359 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1360 cache_device_name(cache));
1361 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1363 mg_complete(mg, false);
1368 * It would be nice if we only had to commit when a REQ_FLUSH
1369 * comes through. But there's one scenario that we have to
1372 * - vblock x in a cache block
1374 * - cache block gets reallocated and over written
1377 * When we recover, because there was no commit the cache will
1378 * rollback to having the data for vblock x in the cache block.
1379 * But the cache block has since been overwritten, so it'll end
1380 * up pointing to data that was never in 'x' during the history
1383 * To avoid this issue we require a commit as part of the
1384 * demotion operation.
1386 init_continuation(&mg->k, mg_success);
1387 continue_after_commit(&cache->committer, &mg->k);
1388 schedule_commit(&cache->committer);
1391 case POLICY_WRITEBACK:
1392 mg_complete(mg, true);
1397 static void mg_update_metadata_after_copy(struct work_struct *ws)
1399 struct dm_cache_migration *mg = ws_to_mg(ws);
1402 * Did the copy succeed?
1405 mg_complete(mg, false);
1407 mg_update_metadata(ws);
1410 static void mg_upgrade_lock(struct work_struct *ws)
1413 struct dm_cache_migration *mg = ws_to_mg(ws);
1416 * Did the copy succeed?
1419 mg_complete(mg, false);
1423 * Now we want the lock to prevent both reads and writes.
1425 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1426 READ_WRITE_LOCK_LEVEL);
1428 mg_complete(mg, false);
1431 quiesce(mg, mg_update_metadata);
1434 mg_update_metadata(ws);
1438 static void mg_full_copy(struct work_struct *ws)
1440 struct dm_cache_migration *mg = ws_to_mg(ws);
1441 struct cache *cache = mg->cache;
1442 struct policy_work *op = mg->op;
1443 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1445 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1446 is_discarded_oblock(cache, op->oblock)) {
1447 mg_upgrade_lock(ws);
1451 init_continuation(&mg->k, mg_upgrade_lock);
1453 if (copy(mg, is_policy_promote)) {
1454 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1455 mg->k.input = BLK_STS_IOERR;
1456 mg_complete(mg, false);
1460 static void mg_copy(struct work_struct *ws)
1462 struct dm_cache_migration *mg = ws_to_mg(ws);
1464 if (mg->overwrite_bio) {
1466 * No exclusive lock was held when we last checked if the bio
1467 * was optimisable. So we have to check again in case things
1468 * have changed (eg, the block may no longer be discarded).
1470 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1472 * Fallback to a real full copy after doing some tidying up.
1474 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1475 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1476 mg->overwrite_bio = NULL;
1477 inc_io_migrations(mg->cache);
1483 * It's safe to do this here, even though it's new data
1484 * because all IO has been locked out of the block.
1486 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1487 * so _not_ using mg_upgrade_lock() as continutation.
1489 overwrite(mg, mg_update_metadata_after_copy);
1495 static int mg_lock_writes(struct dm_cache_migration *mg)
1498 struct dm_cell_key_v2 key;
1499 struct cache *cache = mg->cache;
1500 struct dm_bio_prison_cell_v2 *prealloc;
1502 prealloc = alloc_prison_cell(cache);
1504 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1505 mg_complete(mg, false);
1510 * Prevent writes to the block, but allow reads to continue.
1511 * Unless we're using an overwrite bio, in which case we lock
1514 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1515 r = dm_cell_lock_v2(cache->prison, &key,
1516 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1517 prealloc, &mg->cell);
1519 free_prison_cell(cache, prealloc);
1520 mg_complete(mg, false);
1524 if (mg->cell != prealloc)
1525 free_prison_cell(cache, prealloc);
1530 quiesce(mg, mg_copy);
1535 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1537 struct dm_cache_migration *mg;
1539 if (!background_work_begin(cache)) {
1540 policy_complete_background_work(cache->policy, op, false);
1544 mg = alloc_migration(cache);
1546 policy_complete_background_work(cache->policy, op, false);
1547 background_work_end(cache);
1552 mg->overwrite_bio = bio;
1555 inc_io_migrations(cache);
1557 return mg_lock_writes(mg);
1560 /*----------------------------------------------------------------
1561 * invalidation processing
1562 *--------------------------------------------------------------*/
1564 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1566 struct bio_list bios;
1567 struct cache *cache = mg->cache;
1569 bio_list_init(&bios);
1570 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1571 free_prison_cell(cache, mg->cell);
1573 if (!success && mg->overwrite_bio)
1574 bio_io_error(mg->overwrite_bio);
1577 defer_bios(cache, &bios);
1579 background_work_end(cache);
1582 static void invalidate_completed(struct work_struct *ws)
1584 struct dm_cache_migration *mg = ws_to_mg(ws);
1585 invalidate_complete(mg, !mg->k.input);
1588 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1590 int r = policy_invalidate_mapping(cache->policy, cblock);
1592 r = dm_cache_remove_mapping(cache->cmd, cblock);
1594 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1595 cache_device_name(cache));
1596 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1599 } else if (r == -ENODATA) {
1601 * Harmless, already unmapped.
1606 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1611 static void invalidate_remove(struct work_struct *ws)
1614 struct dm_cache_migration *mg = ws_to_mg(ws);
1615 struct cache *cache = mg->cache;
1617 r = invalidate_cblock(cache, mg->invalidate_cblock);
1619 invalidate_complete(mg, false);
1623 init_continuation(&mg->k, invalidate_completed);
1624 continue_after_commit(&cache->committer, &mg->k);
1625 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1626 mg->overwrite_bio = NULL;
1627 schedule_commit(&cache->committer);
1630 static int invalidate_lock(struct dm_cache_migration *mg)
1633 struct dm_cell_key_v2 key;
1634 struct cache *cache = mg->cache;
1635 struct dm_bio_prison_cell_v2 *prealloc;
1637 prealloc = alloc_prison_cell(cache);
1639 invalidate_complete(mg, false);
1643 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1644 r = dm_cell_lock_v2(cache->prison, &key,
1645 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1647 free_prison_cell(cache, prealloc);
1648 invalidate_complete(mg, false);
1652 if (mg->cell != prealloc)
1653 free_prison_cell(cache, prealloc);
1656 quiesce(mg, invalidate_remove);
1660 * We can't call invalidate_remove() directly here because we
1661 * might still be in request context.
1663 init_continuation(&mg->k, invalidate_remove);
1664 queue_work(cache->wq, &mg->k.ws);
1670 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1671 dm_oblock_t oblock, struct bio *bio)
1673 struct dm_cache_migration *mg;
1675 if (!background_work_begin(cache))
1678 mg = alloc_migration(cache);
1680 background_work_end(cache);
1684 mg->overwrite_bio = bio;
1685 mg->invalidate_cblock = cblock;
1686 mg->invalidate_oblock = oblock;
1688 return invalidate_lock(mg);
1691 /*----------------------------------------------------------------
1693 *--------------------------------------------------------------*/
1700 static enum busy spare_migration_bandwidth(struct cache *cache)
1702 bool idle = iot_idle_for(&cache->tracker, HZ);
1703 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1704 cache->sectors_per_block;
1706 if (idle && current_volume <= cache->migration_threshold)
1712 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1714 atomic_inc(bio_data_dir(bio) == READ ?
1715 &cache->stats.read_hit : &cache->stats.write_hit);
1718 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1720 atomic_inc(bio_data_dir(bio) == READ ?
1721 &cache->stats.read_miss : &cache->stats.write_miss);
1724 /*----------------------------------------------------------------*/
1726 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1727 bool *commit_needed)
1730 bool rb, background_queued;
1733 *commit_needed = false;
1735 rb = bio_detain_shared(cache, block, bio);
1738 * An exclusive lock is held for this block, so we have to
1739 * wait. We set the commit_needed flag so the current
1740 * transaction will be committed asap, allowing this lock
1743 *commit_needed = true;
1744 return DM_MAPIO_SUBMITTED;
1747 data_dir = bio_data_dir(bio);
1749 if (optimisable_bio(cache, bio, block)) {
1750 struct policy_work *op = NULL;
1752 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1753 if (unlikely(r && r != -ENOENT)) {
1754 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1755 cache_device_name(cache), r);
1757 return DM_MAPIO_SUBMITTED;
1760 if (r == -ENOENT && op) {
1761 bio_drop_shared_lock(cache, bio);
1762 BUG_ON(op->op != POLICY_PROMOTE);
1763 mg_start(cache, op, bio);
1764 return DM_MAPIO_SUBMITTED;
1767 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1768 if (unlikely(r && r != -ENOENT)) {
1769 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1770 cache_device_name(cache), r);
1772 return DM_MAPIO_SUBMITTED;
1775 if (background_queued)
1776 wake_migration_worker(cache);
1780 struct per_bio_data *pb = get_per_bio_data(bio);
1785 inc_miss_counter(cache, bio);
1786 if (pb->req_nr == 0) {
1787 accounted_begin(cache, bio);
1788 remap_to_origin_clear_discard(cache, bio, block);
1791 * This is a duplicate writethrough io that is no
1792 * longer needed because the block has been demoted.
1795 return DM_MAPIO_SUBMITTED;
1801 inc_hit_counter(cache, bio);
1804 * Passthrough always maps to the origin, invalidating any
1805 * cache blocks that are written to.
1807 if (passthrough_mode(cache)) {
1808 if (bio_data_dir(bio) == WRITE) {
1809 bio_drop_shared_lock(cache, bio);
1810 atomic_inc(&cache->stats.demotion);
1811 invalidate_start(cache, cblock, block, bio);
1813 remap_to_origin_clear_discard(cache, bio, block);
1815 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1816 !is_dirty(cache, cblock)) {
1817 remap_to_origin_and_cache(cache, bio, block, cblock);
1818 accounted_begin(cache, bio);
1820 remap_to_cache_dirty(cache, bio, block, cblock);
1825 * dm core turns FUA requests into a separate payload and FLUSH req.
1827 if (bio->bi_opf & REQ_FUA) {
1829 * issue_after_commit will call accounted_begin a second time. So
1830 * we call accounted_complete() to avoid double accounting.
1832 accounted_complete(cache, bio);
1833 issue_after_commit(&cache->committer, bio);
1834 *commit_needed = true;
1835 return DM_MAPIO_SUBMITTED;
1838 return DM_MAPIO_REMAPPED;
1841 static bool process_bio(struct cache *cache, struct bio *bio)
1845 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1846 generic_make_request(bio);
1848 return commit_needed;
1852 * A non-zero return indicates read_only or fail_io mode.
1854 static int commit(struct cache *cache, bool clean_shutdown)
1858 if (get_cache_mode(cache) >= CM_READ_ONLY)
1861 atomic_inc(&cache->stats.commit_count);
1862 r = dm_cache_commit(cache->cmd, clean_shutdown);
1864 metadata_operation_failed(cache, "dm_cache_commit", r);
1870 * Used by the batcher.
1872 static blk_status_t commit_op(void *context)
1874 struct cache *cache = context;
1876 if (dm_cache_changed_this_transaction(cache->cmd))
1877 return errno_to_blk_status(commit(cache, false));
1882 /*----------------------------------------------------------------*/
1884 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1886 struct per_bio_data *pb = get_per_bio_data(bio);
1889 remap_to_origin(cache, bio);
1891 remap_to_cache(cache, bio, 0);
1893 issue_after_commit(&cache->committer, bio);
1897 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1901 // FIXME: do we need to lock the region? Or can we just assume the
1902 // user wont be so foolish as to issue discard concurrently with
1904 calc_discard_block_range(cache, bio, &b, &e);
1906 set_discard(cache, b);
1907 b = to_dblock(from_dblock(b) + 1);
1915 static void process_deferred_bios(struct work_struct *ws)
1917 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1919 unsigned long flags;
1920 bool commit_needed = false;
1921 struct bio_list bios;
1924 bio_list_init(&bios);
1926 spin_lock_irqsave(&cache->lock, flags);
1927 bio_list_merge(&bios, &cache->deferred_bios);
1928 bio_list_init(&cache->deferred_bios);
1929 spin_unlock_irqrestore(&cache->lock, flags);
1931 while ((bio = bio_list_pop(&bios))) {
1932 if (bio->bi_opf & REQ_PREFLUSH)
1933 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1935 else if (bio_op(bio) == REQ_OP_DISCARD)
1936 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1939 commit_needed = process_bio(cache, bio) || commit_needed;
1943 schedule_commit(&cache->committer);
1946 /*----------------------------------------------------------------
1948 *--------------------------------------------------------------*/
1950 static void requeue_deferred_bios(struct cache *cache)
1953 struct bio_list bios;
1955 bio_list_init(&bios);
1956 bio_list_merge(&bios, &cache->deferred_bios);
1957 bio_list_init(&cache->deferred_bios);
1959 while ((bio = bio_list_pop(&bios))) {
1960 bio->bi_status = BLK_STS_DM_REQUEUE;
1966 * We want to commit periodically so that not too much
1967 * unwritten metadata builds up.
1969 static void do_waker(struct work_struct *ws)
1971 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1973 policy_tick(cache->policy, true);
1974 wake_migration_worker(cache);
1975 schedule_commit(&cache->committer);
1976 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1979 static void check_migrations(struct work_struct *ws)
1982 struct policy_work *op;
1983 struct cache *cache = container_of(ws, struct cache, migration_worker);
1987 b = spare_migration_bandwidth(cache);
1989 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1994 DMERR_LIMIT("%s: policy_background_work failed",
1995 cache_device_name(cache));
1999 r = mg_start(cache, op, NULL);
2005 /*----------------------------------------------------------------
2007 *--------------------------------------------------------------*/
2010 * This function gets called on the error paths of the constructor, so we
2011 * have to cope with a partially initialised struct.
2013 static void destroy(struct cache *cache)
2017 mempool_exit(&cache->migration_pool);
2020 dm_bio_prison_destroy_v2(cache->prison);
2023 destroy_workqueue(cache->wq);
2025 if (cache->dirty_bitset)
2026 free_bitset(cache->dirty_bitset);
2028 if (cache->discard_bitset)
2029 free_bitset(cache->discard_bitset);
2032 dm_kcopyd_client_destroy(cache->copier);
2035 dm_cache_metadata_close(cache->cmd);
2037 if (cache->metadata_dev)
2038 dm_put_device(cache->ti, cache->metadata_dev);
2040 if (cache->origin_dev)
2041 dm_put_device(cache->ti, cache->origin_dev);
2043 if (cache->cache_dev)
2044 dm_put_device(cache->ti, cache->cache_dev);
2047 dm_cache_policy_destroy(cache->policy);
2049 for (i = 0; i < cache->nr_ctr_args ; i++)
2050 kfree(cache->ctr_args[i]);
2051 kfree(cache->ctr_args);
2053 bioset_exit(&cache->bs);
2058 static void cache_dtr(struct dm_target *ti)
2060 struct cache *cache = ti->private;
2065 static sector_t get_dev_size(struct dm_dev *dev)
2067 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2070 /*----------------------------------------------------------------*/
2073 * Construct a cache device mapping.
2075 * cache <metadata dev> <cache dev> <origin dev> <block size>
2076 * <#feature args> [<feature arg>]*
2077 * <policy> <#policy args> [<policy arg>]*
2079 * metadata dev : fast device holding the persistent metadata
2080 * cache dev : fast device holding cached data blocks
2081 * origin dev : slow device holding original data blocks
2082 * block size : cache unit size in sectors
2084 * #feature args : number of feature arguments passed
2085 * feature args : writethrough. (The default is writeback.)
2087 * policy : the replacement policy to use
2088 * #policy args : an even number of policy arguments corresponding
2089 * to key/value pairs passed to the policy
2090 * policy args : key/value pairs passed to the policy
2091 * E.g. 'sequential_threshold 1024'
2092 * See cache-policies.txt for details.
2094 * Optional feature arguments are:
2095 * writethrough : write through caching that prohibits cache block
2096 * content from being different from origin block content.
2097 * Without this argument, the default behaviour is to write
2098 * back cache block contents later for performance reasons,
2099 * so they may differ from the corresponding origin blocks.
2102 struct dm_target *ti;
2104 struct dm_dev *metadata_dev;
2106 struct dm_dev *cache_dev;
2107 sector_t cache_sectors;
2109 struct dm_dev *origin_dev;
2110 sector_t origin_sectors;
2112 uint32_t block_size;
2114 const char *policy_name;
2116 const char **policy_argv;
2118 struct cache_features features;
2121 static void destroy_cache_args(struct cache_args *ca)
2123 if (ca->metadata_dev)
2124 dm_put_device(ca->ti, ca->metadata_dev);
2127 dm_put_device(ca->ti, ca->cache_dev);
2130 dm_put_device(ca->ti, ca->origin_dev);
2135 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2138 *error = "Insufficient args";
2145 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2149 sector_t metadata_dev_size;
2150 char b[BDEVNAME_SIZE];
2152 if (!at_least_one_arg(as, error))
2155 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2158 *error = "Error opening metadata device";
2162 metadata_dev_size = get_dev_size(ca->metadata_dev);
2163 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2164 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2165 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2170 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2175 if (!at_least_one_arg(as, error))
2178 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2181 *error = "Error opening cache device";
2184 ca->cache_sectors = get_dev_size(ca->cache_dev);
2189 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2194 if (!at_least_one_arg(as, error))
2197 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2200 *error = "Error opening origin device";
2204 ca->origin_sectors = get_dev_size(ca->origin_dev);
2205 if (ca->ti->len > ca->origin_sectors) {
2206 *error = "Device size larger than cached device";
2213 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2216 unsigned long block_size;
2218 if (!at_least_one_arg(as, error))
2221 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2222 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2223 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2224 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2225 *error = "Invalid data block size";
2229 if (block_size > ca->cache_sectors) {
2230 *error = "Data block size is larger than the cache device";
2234 ca->block_size = block_size;
2239 static void init_features(struct cache_features *cf)
2241 cf->mode = CM_WRITE;
2242 cf->io_mode = CM_IO_WRITEBACK;
2243 cf->metadata_version = 1;
2246 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2249 static const struct dm_arg _args[] = {
2250 {0, 2, "Invalid number of cache feature arguments"},
2256 struct cache_features *cf = &ca->features;
2260 r = dm_read_arg_group(_args, as, &argc, error);
2265 arg = dm_shift_arg(as);
2267 if (!strcasecmp(arg, "writeback"))
2268 cf->io_mode = CM_IO_WRITEBACK;
2270 else if (!strcasecmp(arg, "writethrough"))
2271 cf->io_mode = CM_IO_WRITETHROUGH;
2273 else if (!strcasecmp(arg, "passthrough"))
2274 cf->io_mode = CM_IO_PASSTHROUGH;
2276 else if (!strcasecmp(arg, "metadata2"))
2277 cf->metadata_version = 2;
2280 *error = "Unrecognised cache feature requested";
2288 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2291 static const struct dm_arg _args[] = {
2292 {0, 1024, "Invalid number of policy arguments"},
2297 if (!at_least_one_arg(as, error))
2300 ca->policy_name = dm_shift_arg(as);
2302 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2306 ca->policy_argv = (const char **)as->argv;
2307 dm_consume_args(as, ca->policy_argc);
2312 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2316 struct dm_arg_set as;
2321 r = parse_metadata_dev(ca, &as, error);
2325 r = parse_cache_dev(ca, &as, error);
2329 r = parse_origin_dev(ca, &as, error);
2333 r = parse_block_size(ca, &as, error);
2337 r = parse_features(ca, &as, error);
2341 r = parse_policy(ca, &as, error);
2348 /*----------------------------------------------------------------*/
2350 static struct kmem_cache *migration_cache;
2352 #define NOT_CORE_OPTION 1
2354 static int process_config_option(struct cache *cache, const char *key, const char *value)
2358 if (!strcasecmp(key, "migration_threshold")) {
2359 if (kstrtoul(value, 10, &tmp))
2362 cache->migration_threshold = tmp;
2366 return NOT_CORE_OPTION;
2369 static int set_config_value(struct cache *cache, const char *key, const char *value)
2371 int r = process_config_option(cache, key, value);
2373 if (r == NOT_CORE_OPTION)
2374 r = policy_set_config_value(cache->policy, key, value);
2377 DMWARN("bad config value for %s: %s", key, value);
2382 static int set_config_values(struct cache *cache, int argc, const char **argv)
2387 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2392 r = set_config_value(cache, argv[0], argv[1]);
2403 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2406 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2408 cache->origin_sectors,
2409 cache->sectors_per_block);
2411 *error = "Error creating cache's policy";
2415 BUG_ON(!cache->policy);
2421 * We want the discard block size to be at least the size of the cache
2422 * block size and have no more than 2^14 discard blocks across the origin.
2424 #define MAX_DISCARD_BLOCKS (1 << 14)
2426 static bool too_many_discard_blocks(sector_t discard_block_size,
2427 sector_t origin_size)
2429 (void) sector_div(origin_size, discard_block_size);
2431 return origin_size > MAX_DISCARD_BLOCKS;
2434 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2435 sector_t origin_size)
2437 sector_t discard_block_size = cache_block_size;
2440 while (too_many_discard_blocks(discard_block_size, origin_size))
2441 discard_block_size *= 2;
2443 return discard_block_size;
2446 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2448 dm_block_t nr_blocks = from_cblock(size);
2450 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2451 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2452 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2453 "Please consider increasing the cache block size to reduce the overall cache block count.",
2454 (unsigned long long) nr_blocks);
2456 cache->cache_size = size;
2459 static int is_congested(struct dm_dev *dev, int bdi_bits)
2461 struct request_queue *q = bdev_get_queue(dev->bdev);
2462 return bdi_congested(q->backing_dev_info, bdi_bits);
2465 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2467 struct cache *cache = container_of(cb, struct cache, callbacks);
2469 return is_congested(cache->origin_dev, bdi_bits) ||
2470 is_congested(cache->cache_dev, bdi_bits);
2473 #define DEFAULT_MIGRATION_THRESHOLD 2048
2475 static int cache_create(struct cache_args *ca, struct cache **result)
2478 char **error = &ca->ti->error;
2479 struct cache *cache;
2480 struct dm_target *ti = ca->ti;
2481 dm_block_t origin_blocks;
2482 struct dm_cache_metadata *cmd;
2483 bool may_format = ca->features.mode == CM_WRITE;
2485 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2490 ti->private = cache;
2491 ti->num_flush_bios = 2;
2492 ti->flush_supported = true;
2494 ti->num_discard_bios = 1;
2495 ti->discards_supported = true;
2496 ti->split_discard_bios = false;
2498 ti->per_io_data_size = sizeof(struct per_bio_data);
2500 cache->features = ca->features;
2501 if (writethrough_mode(cache)) {
2502 /* Create bioset for writethrough bios issued to origin */
2503 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2508 cache->callbacks.congested_fn = cache_is_congested;
2509 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2511 cache->metadata_dev = ca->metadata_dev;
2512 cache->origin_dev = ca->origin_dev;
2513 cache->cache_dev = ca->cache_dev;
2515 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2517 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2518 origin_blocks = block_div(origin_blocks, ca->block_size);
2519 cache->origin_blocks = to_oblock(origin_blocks);
2521 cache->sectors_per_block = ca->block_size;
2522 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2527 if (ca->block_size & (ca->block_size - 1)) {
2528 dm_block_t cache_size = ca->cache_sectors;
2530 cache->sectors_per_block_shift = -1;
2531 cache_size = block_div(cache_size, ca->block_size);
2532 set_cache_size(cache, to_cblock(cache_size));
2534 cache->sectors_per_block_shift = __ffs(ca->block_size);
2535 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2538 r = create_cache_policy(cache, ca, error);
2542 cache->policy_nr_args = ca->policy_argc;
2543 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2545 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2547 *error = "Error setting cache policy's config values";
2551 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2552 ca->block_size, may_format,
2553 dm_cache_policy_get_hint_size(cache->policy),
2554 ca->features.metadata_version);
2556 *error = "Error creating metadata object";
2561 set_cache_mode(cache, CM_WRITE);
2562 if (get_cache_mode(cache) != CM_WRITE) {
2563 *error = "Unable to get write access to metadata, please check/repair metadata.";
2568 if (passthrough_mode(cache)) {
2571 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2573 *error = "dm_cache_metadata_all_clean() failed";
2578 *error = "Cannot enter passthrough mode unless all blocks are clean";
2583 policy_allow_migrations(cache->policy, false);
2586 spin_lock_init(&cache->lock);
2587 bio_list_init(&cache->deferred_bios);
2588 atomic_set(&cache->nr_allocated_migrations, 0);
2589 atomic_set(&cache->nr_io_migrations, 0);
2590 init_waitqueue_head(&cache->migration_wait);
2593 atomic_set(&cache->nr_dirty, 0);
2594 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2595 if (!cache->dirty_bitset) {
2596 *error = "could not allocate dirty bitset";
2599 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2601 cache->discard_block_size =
2602 calculate_discard_block_size(cache->sectors_per_block,
2603 cache->origin_sectors);
2604 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2605 cache->discard_block_size));
2606 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2607 if (!cache->discard_bitset) {
2608 *error = "could not allocate discard bitset";
2611 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2613 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2614 if (IS_ERR(cache->copier)) {
2615 *error = "could not create kcopyd client";
2616 r = PTR_ERR(cache->copier);
2620 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2622 *error = "could not create workqueue for metadata object";
2625 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2626 INIT_WORK(&cache->migration_worker, check_migrations);
2627 INIT_DELAYED_WORK(&cache->waker, do_waker);
2629 cache->prison = dm_bio_prison_create_v2(cache->wq);
2630 if (!cache->prison) {
2631 *error = "could not create bio prison";
2635 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2638 *error = "Error creating cache's migration mempool";
2642 cache->need_tick_bio = true;
2643 cache->sized = false;
2644 cache->invalidate = false;
2645 cache->commit_requested = false;
2646 cache->loaded_mappings = false;
2647 cache->loaded_discards = false;
2651 atomic_set(&cache->stats.demotion, 0);
2652 atomic_set(&cache->stats.promotion, 0);
2653 atomic_set(&cache->stats.copies_avoided, 0);
2654 atomic_set(&cache->stats.cache_cell_clash, 0);
2655 atomic_set(&cache->stats.commit_count, 0);
2656 atomic_set(&cache->stats.discard_count, 0);
2658 spin_lock_init(&cache->invalidation_lock);
2659 INIT_LIST_HEAD(&cache->invalidation_requests);
2661 batcher_init(&cache->committer, commit_op, cache,
2662 issue_op, cache, cache->wq);
2663 iot_init(&cache->tracker);
2665 init_rwsem(&cache->background_work_lock);
2666 prevent_background_work(cache);
2675 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2680 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2683 for (i = 0; i < argc; i++) {
2684 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2693 cache->nr_ctr_args = argc;
2694 cache->ctr_args = copy;
2699 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2702 struct cache_args *ca;
2703 struct cache *cache = NULL;
2705 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2707 ti->error = "Error allocating memory for cache";
2712 r = parse_cache_args(ca, argc, argv, &ti->error);
2716 r = cache_create(ca, &cache);
2720 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2726 ti->private = cache;
2728 destroy_cache_args(ca);
2732 /*----------------------------------------------------------------*/
2734 static int cache_map(struct dm_target *ti, struct bio *bio)
2736 struct cache *cache = ti->private;
2740 dm_oblock_t block = get_bio_block(cache, bio);
2742 init_per_bio_data(bio);
2743 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2745 * This can only occur if the io goes to a partial block at
2746 * the end of the origin device. We don't cache these.
2747 * Just remap to the origin and carry on.
2749 remap_to_origin(cache, bio);
2750 accounted_begin(cache, bio);
2751 return DM_MAPIO_REMAPPED;
2754 if (discard_or_flush(bio)) {
2755 defer_bio(cache, bio);
2756 return DM_MAPIO_SUBMITTED;
2759 r = map_bio(cache, bio, block, &commit_needed);
2761 schedule_commit(&cache->committer);
2766 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2768 struct cache *cache = ti->private;
2769 unsigned long flags;
2770 struct per_bio_data *pb = get_per_bio_data(bio);
2773 policy_tick(cache->policy, false);
2775 spin_lock_irqsave(&cache->lock, flags);
2776 cache->need_tick_bio = true;
2777 spin_unlock_irqrestore(&cache->lock, flags);
2780 bio_drop_shared_lock(cache, bio);
2781 accounted_complete(cache, bio);
2783 return DM_ENDIO_DONE;
2786 static int write_dirty_bitset(struct cache *cache)
2790 if (get_cache_mode(cache) >= CM_READ_ONLY)
2793 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2795 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2800 static int write_discard_bitset(struct cache *cache)
2804 if (get_cache_mode(cache) >= CM_READ_ONLY)
2807 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2808 cache->discard_nr_blocks);
2810 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2811 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2815 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2816 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2817 is_discarded(cache, to_dblock(i)));
2819 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2827 static int write_hints(struct cache *cache)
2831 if (get_cache_mode(cache) >= CM_READ_ONLY)
2834 r = dm_cache_write_hints(cache->cmd, cache->policy);
2836 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2844 * returns true on success
2846 static bool sync_metadata(struct cache *cache)
2850 r1 = write_dirty_bitset(cache);
2852 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2854 r2 = write_discard_bitset(cache);
2856 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2860 r3 = write_hints(cache);
2862 DMERR("%s: could not write hints", cache_device_name(cache));
2865 * If writing the above metadata failed, we still commit, but don't
2866 * set the clean shutdown flag. This will effectively force every
2867 * dirty bit to be set on reload.
2869 r4 = commit(cache, !r1 && !r2 && !r3);
2871 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2873 return !r1 && !r2 && !r3 && !r4;
2876 static void cache_postsuspend(struct dm_target *ti)
2878 struct cache *cache = ti->private;
2880 prevent_background_work(cache);
2881 BUG_ON(atomic_read(&cache->nr_io_migrations));
2883 cancel_delayed_work(&cache->waker);
2884 flush_workqueue(cache->wq);
2885 WARN_ON(cache->tracker.in_flight);
2888 * If it's a flush suspend there won't be any deferred bios, so this
2891 requeue_deferred_bios(cache);
2893 if (get_cache_mode(cache) == CM_WRITE)
2894 (void) sync_metadata(cache);
2897 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2898 bool dirty, uint32_t hint, bool hint_valid)
2901 struct cache *cache = context;
2904 set_bit(from_cblock(cblock), cache->dirty_bitset);
2905 atomic_inc(&cache->nr_dirty);
2907 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2909 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2917 * The discard block size in the on disk metadata is not
2918 * neccessarily the same as we're currently using. So we have to
2919 * be careful to only set the discarded attribute if we know it
2920 * covers a complete block of the new size.
2922 struct discard_load_info {
2923 struct cache *cache;
2926 * These blocks are sized using the on disk dblock size, rather
2927 * than the current one.
2929 dm_block_t block_size;
2930 dm_block_t discard_begin, discard_end;
2933 static void discard_load_info_init(struct cache *cache,
2934 struct discard_load_info *li)
2937 li->discard_begin = li->discard_end = 0;
2940 static void set_discard_range(struct discard_load_info *li)
2944 if (li->discard_begin == li->discard_end)
2948 * Convert to sectors.
2950 b = li->discard_begin * li->block_size;
2951 e = li->discard_end * li->block_size;
2954 * Then convert back to the current dblock size.
2956 b = dm_sector_div_up(b, li->cache->discard_block_size);
2957 sector_div(e, li->cache->discard_block_size);
2960 * The origin may have shrunk, so we need to check we're still in
2963 if (e > from_dblock(li->cache->discard_nr_blocks))
2964 e = from_dblock(li->cache->discard_nr_blocks);
2967 set_discard(li->cache, to_dblock(b));
2970 static int load_discard(void *context, sector_t discard_block_size,
2971 dm_dblock_t dblock, bool discard)
2973 struct discard_load_info *li = context;
2975 li->block_size = discard_block_size;
2978 if (from_dblock(dblock) == li->discard_end)
2980 * We're already in a discard range, just extend it.
2982 li->discard_end = li->discard_end + 1ULL;
2986 * Emit the old range and start a new one.
2988 set_discard_range(li);
2989 li->discard_begin = from_dblock(dblock);
2990 li->discard_end = li->discard_begin + 1ULL;
2993 set_discard_range(li);
2994 li->discard_begin = li->discard_end = 0;
3000 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3002 sector_t size = get_dev_size(cache->cache_dev);
3003 (void) sector_div(size, cache->sectors_per_block);
3004 return to_cblock(size);
3007 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3009 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3013 * We can't drop a dirty block when shrinking the cache.
3015 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3016 new_size = to_cblock(from_cblock(new_size) + 1);
3017 if (is_dirty(cache, new_size)) {
3018 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3019 cache_device_name(cache),
3020 (unsigned long long) from_cblock(new_size));
3028 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3032 r = dm_cache_resize(cache->cmd, new_size);
3034 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3035 metadata_operation_failed(cache, "dm_cache_resize", r);
3039 set_cache_size(cache, new_size);
3044 static int cache_preresume(struct dm_target *ti)
3047 struct cache *cache = ti->private;
3048 dm_cblock_t csize = get_cache_dev_size(cache);
3051 * Check to see if the cache has resized.
3053 if (!cache->sized) {
3054 r = resize_cache_dev(cache, csize);
3058 cache->sized = true;
3060 } else if (csize != cache->cache_size) {
3061 if (!can_resize(cache, csize))
3064 r = resize_cache_dev(cache, csize);
3069 if (!cache->loaded_mappings) {
3070 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3071 load_mapping, cache);
3073 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3074 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3078 cache->loaded_mappings = true;
3081 if (!cache->loaded_discards) {
3082 struct discard_load_info li;
3085 * The discard bitset could have been resized, or the
3086 * discard block size changed. To be safe we start by
3087 * setting every dblock to not discarded.
3089 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3091 discard_load_info_init(cache, &li);
3092 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3094 DMERR("%s: could not load origin discards", cache_device_name(cache));
3095 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3098 set_discard_range(&li);
3100 cache->loaded_discards = true;
3106 static void cache_resume(struct dm_target *ti)
3108 struct cache *cache = ti->private;
3110 cache->need_tick_bio = true;
3111 allow_background_work(cache);
3112 do_waker(&cache->waker.work);
3118 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3119 * <cache block size> <#used cache blocks>/<#total cache blocks>
3120 * <#read hits> <#read misses> <#write hits> <#write misses>
3121 * <#demotions> <#promotions> <#dirty>
3122 * <#features> <features>*
3123 * <#core args> <core args>
3124 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3126 static void cache_status(struct dm_target *ti, status_type_t type,
3127 unsigned status_flags, char *result, unsigned maxlen)
3132 dm_block_t nr_free_blocks_metadata = 0;
3133 dm_block_t nr_blocks_metadata = 0;
3134 char buf[BDEVNAME_SIZE];
3135 struct cache *cache = ti->private;
3136 dm_cblock_t residency;
3140 case STATUSTYPE_INFO:
3141 if (get_cache_mode(cache) == CM_FAIL) {
3146 /* Commit to ensure statistics aren't out-of-date */
3147 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3148 (void) commit(cache, false);
3150 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3152 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3153 cache_device_name(cache), r);
3157 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3159 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3160 cache_device_name(cache), r);
3164 residency = policy_residency(cache->policy);
3166 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3167 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3168 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3169 (unsigned long long)nr_blocks_metadata,
3170 (unsigned long long)cache->sectors_per_block,
3171 (unsigned long long) from_cblock(residency),
3172 (unsigned long long) from_cblock(cache->cache_size),
3173 (unsigned) atomic_read(&cache->stats.read_hit),
3174 (unsigned) atomic_read(&cache->stats.read_miss),
3175 (unsigned) atomic_read(&cache->stats.write_hit),
3176 (unsigned) atomic_read(&cache->stats.write_miss),
3177 (unsigned) atomic_read(&cache->stats.demotion),
3178 (unsigned) atomic_read(&cache->stats.promotion),
3179 (unsigned long) atomic_read(&cache->nr_dirty));
3181 if (cache->features.metadata_version == 2)
3182 DMEMIT("2 metadata2 ");
3186 if (writethrough_mode(cache))
3187 DMEMIT("writethrough ");
3189 else if (passthrough_mode(cache))
3190 DMEMIT("passthrough ");
3192 else if (writeback_mode(cache))
3193 DMEMIT("writeback ");
3196 DMERR("%s: internal error: unknown io mode: %d",
3197 cache_device_name(cache), (int) cache->features.io_mode);
3201 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3203 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3205 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3207 DMERR("%s: policy_emit_config_values returned %d",
3208 cache_device_name(cache), r);
3211 if (get_cache_mode(cache) == CM_READ_ONLY)
3216 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3218 if (r || needs_check)
3219 DMEMIT("needs_check ");
3225 case STATUSTYPE_TABLE:
3226 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3228 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3230 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3233 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3234 DMEMIT(" %s", cache->ctr_args[i]);
3235 if (cache->nr_ctr_args)
3236 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3246 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3247 * the one-past-the-end value.
3249 struct cblock_range {
3255 * A cache block range can take two forms:
3257 * i) A single cblock, eg. '3456'
3258 * ii) A begin and end cblock with a dash between, eg. 123-234
3260 static int parse_cblock_range(struct cache *cache, const char *str,
3261 struct cblock_range *result)
3268 * Try and parse form (ii) first.
3270 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3275 result->begin = to_cblock(b);
3276 result->end = to_cblock(e);
3281 * That didn't work, try form (i).
3283 r = sscanf(str, "%llu%c", &b, &dummy);
3288 result->begin = to_cblock(b);
3289 result->end = to_cblock(from_cblock(result->begin) + 1u);
3293 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3297 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3299 uint64_t b = from_cblock(range->begin);
3300 uint64_t e = from_cblock(range->end);
3301 uint64_t n = from_cblock(cache->cache_size);
3304 DMERR("%s: begin cblock out of range: %llu >= %llu",
3305 cache_device_name(cache), b, n);
3310 DMERR("%s: end cblock out of range: %llu > %llu",
3311 cache_device_name(cache), e, n);
3316 DMERR("%s: invalid cblock range: %llu >= %llu",
3317 cache_device_name(cache), b, e);
3324 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3326 return to_cblock(from_cblock(b) + 1);
3329 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3334 * We don't need to do any locking here because we know we're in
3335 * passthrough mode. There's is potential for a race between an
3336 * invalidation triggered by an io and an invalidation message. This
3337 * is harmless, we must not worry if the policy call fails.
3339 while (range->begin != range->end) {
3340 r = invalidate_cblock(cache, range->begin);
3344 range->begin = cblock_succ(range->begin);
3347 cache->commit_requested = true;
3351 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3352 const char **cblock_ranges)
3356 struct cblock_range range;
3358 if (!passthrough_mode(cache)) {
3359 DMERR("%s: cache has to be in passthrough mode for invalidation",
3360 cache_device_name(cache));
3364 for (i = 0; i < count; i++) {
3365 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3369 r = validate_cblock_range(cache, &range);
3374 * Pass begin and end origin blocks to the worker and wake it.
3376 r = request_invalidation(cache, &range);
3388 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3390 * The key migration_threshold is supported by the cache target core.
3392 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3393 char *result, unsigned maxlen)
3395 struct cache *cache = ti->private;
3400 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3401 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3402 cache_device_name(cache));
3406 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3407 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3412 return set_config_value(cache, argv[0], argv[1]);
3415 static int cache_iterate_devices(struct dm_target *ti,
3416 iterate_devices_callout_fn fn, void *data)
3419 struct cache *cache = ti->private;
3421 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3423 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3428 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3431 * FIXME: these limits may be incompatible with the cache device
3433 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3434 cache->origin_sectors);
3435 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3438 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3440 struct cache *cache = ti->private;
3441 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3444 * If the system-determined stacked limits are compatible with the
3445 * cache's blocksize (io_opt is a factor) do not override them.
3447 if (io_opt_sectors < cache->sectors_per_block ||
3448 do_div(io_opt_sectors, cache->sectors_per_block)) {
3449 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3450 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3452 set_discard_limits(cache, limits);
3455 /*----------------------------------------------------------------*/
3457 static struct target_type cache_target = {
3459 .version = {2, 0, 0},
3460 .module = THIS_MODULE,
3464 .end_io = cache_end_io,
3465 .postsuspend = cache_postsuspend,
3466 .preresume = cache_preresume,
3467 .resume = cache_resume,
3468 .status = cache_status,
3469 .message = cache_message,
3470 .iterate_devices = cache_iterate_devices,
3471 .io_hints = cache_io_hints,
3474 static int __init dm_cache_init(void)
3478 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3479 if (!migration_cache) {
3480 dm_unregister_target(&cache_target);
3484 r = dm_register_target(&cache_target);
3486 DMERR("cache target registration failed: %d", r);
3493 static void __exit dm_cache_exit(void)
3495 dm_unregister_target(&cache_target);
3496 kmem_cache_destroy(migration_cache);
3499 module_init(dm_cache_init);
3500 module_exit(dm_cache_exit);
3502 MODULE_DESCRIPTION(DM_NAME " cache target");
3503 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3504 MODULE_LICENSE("GPL");