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 int (*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);
177 struct list_head work_items;
178 struct work_struct *ws, *tmp;
179 struct continuation *k;
181 struct bio_list bios;
183 INIT_LIST_HEAD(&work_items);
184 bio_list_init(&bios);
187 * We have to grab these before the commit_op to avoid a race
190 spin_lock_irqsave(&b->lock, flags);
191 list_splice_init(&b->work_items, &work_items);
192 bio_list_merge(&bios, &b->bios);
193 bio_list_init(&b->bios);
194 b->commit_scheduled = false;
195 spin_unlock_irqrestore(&b->lock, flags);
197 r = b->commit_op(b->commit_context);
199 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
200 k = container_of(ws, struct continuation, ws);
202 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
203 queue_work(b->wq, ws);
206 while ((bio = bio_list_pop(&bios))) {
211 b->issue_op(bio, b->issue_context);
215 static void batcher_init(struct batcher *b,
216 int (*commit_op)(void *),
217 void *commit_context,
218 void (*issue_op)(struct bio *bio, void *),
220 struct workqueue_struct *wq)
222 b->commit_op = commit_op;
223 b->commit_context = commit_context;
224 b->issue_op = issue_op;
225 b->issue_context = issue_context;
228 spin_lock_init(&b->lock);
229 INIT_LIST_HEAD(&b->work_items);
230 bio_list_init(&b->bios);
231 INIT_WORK(&b->commit_work, __commit);
232 b->commit_scheduled = false;
235 static void async_commit(struct batcher *b)
237 queue_work(b->wq, &b->commit_work);
240 static void continue_after_commit(struct batcher *b, struct continuation *k)
243 bool commit_scheduled;
245 spin_lock_irqsave(&b->lock, flags);
246 commit_scheduled = b->commit_scheduled;
247 list_add_tail(&k->ws.entry, &b->work_items);
248 spin_unlock_irqrestore(&b->lock, flags);
250 if (commit_scheduled)
255 * Bios are errored if commit failed.
257 static void issue_after_commit(struct batcher *b, struct bio *bio)
260 bool commit_scheduled;
262 spin_lock_irqsave(&b->lock, flags);
263 commit_scheduled = b->commit_scheduled;
264 bio_list_add(&b->bios, bio);
265 spin_unlock_irqrestore(&b->lock, flags);
267 if (commit_scheduled)
272 * Call this if some urgent work is waiting for the commit to complete.
274 static void schedule_commit(struct batcher *b)
279 spin_lock_irqsave(&b->lock, flags);
280 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
281 b->commit_scheduled = true;
282 spin_unlock_irqrestore(&b->lock, flags);
289 * There are a couple of places where we let a bio run, but want to do some
290 * work before calling its endio function. We do this by temporarily
291 * changing the endio fn.
293 struct dm_hook_info {
294 bio_end_io_t *bi_end_io;
297 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
298 bio_end_io_t *bi_end_io, void *bi_private)
300 h->bi_end_io = bio->bi_end_io;
302 bio->bi_end_io = bi_end_io;
303 bio->bi_private = bi_private;
306 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
308 bio->bi_end_io = h->bi_end_io;
311 /*----------------------------------------------------------------*/
313 #define MIGRATION_POOL_SIZE 128
314 #define COMMIT_PERIOD HZ
315 #define MIGRATION_COUNT_WINDOW 10
318 * The block size of the device holding cache data must be
319 * between 32KB and 1GB.
321 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
322 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
324 enum cache_metadata_mode {
325 CM_WRITE, /* metadata may be changed */
326 CM_READ_ONLY, /* metadata may not be changed */
332 * Data is written to cached blocks only. These blocks are marked
333 * dirty. If you lose the cache device you will lose data.
334 * Potential performance increase for both reads and writes.
339 * Data is written to both cache and origin. Blocks are never
340 * dirty. Potential performance benfit for reads only.
345 * A degraded mode useful for various cache coherency situations
346 * (eg, rolling back snapshots). Reads and writes always go to the
347 * origin. If a write goes to a cached oblock, then the cache
348 * block is invalidated.
353 struct cache_features {
354 enum cache_metadata_mode mode;
355 enum cache_io_mode io_mode;
356 unsigned metadata_version;
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;
375 struct dm_target_callbacks callbacks;
377 struct dm_cache_metadata *cmd;
380 * Metadata is written to this device.
382 struct dm_dev *metadata_dev;
385 * The slower of the two data devices. Typically a spindle.
387 struct dm_dev *origin_dev;
390 * The faster of the two data devices. Typically an SSD.
392 struct dm_dev *cache_dev;
395 * Size of the origin device in _complete_ blocks and native sectors.
397 dm_oblock_t origin_blocks;
398 sector_t origin_sectors;
401 * Size of the cache device in blocks.
403 dm_cblock_t cache_size;
406 * Fields for converting from sectors to blocks.
408 sector_t sectors_per_block;
409 int sectors_per_block_shift;
412 struct list_head deferred_cells;
413 struct bio_list deferred_bios;
414 struct bio_list deferred_writethrough_bios;
415 sector_t migration_threshold;
416 wait_queue_head_t migration_wait;
417 atomic_t nr_allocated_migrations;
420 * The number of in flight migrations that are performing
421 * background io. eg, promotion, writeback.
423 atomic_t nr_io_migrations;
425 struct rw_semaphore quiesce_lock;
428 * cache_size entries, dirty if set
431 unsigned long *dirty_bitset;
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 workqueue_struct *wq;
449 struct work_struct deferred_bio_worker;
450 struct work_struct deferred_writethrough_worker;
451 struct work_struct migration_worker;
452 struct delayed_work waker;
453 struct dm_bio_prison_v2 *prison;
455 mempool_t *migration_pool;
457 struct dm_cache_policy *policy;
458 unsigned policy_nr_args;
460 bool need_tick_bio:1;
463 bool commit_requested:1;
464 bool loaded_mappings:1;
465 bool loaded_discards:1;
468 * Cache features such as write-through.
470 struct cache_features features;
472 struct cache_stats stats;
475 * Invalidation fields.
477 spinlock_t invalidation_lock;
478 struct list_head invalidation_requests;
480 struct io_tracker tracker;
482 struct work_struct commit_ws;
483 struct batcher committer;
485 struct rw_semaphore background_work_lock;
488 struct per_bio_data {
491 struct dm_bio_prison_cell_v2 *cell;
492 struct dm_hook_info hook_info;
496 * writethrough fields. These MUST remain at the end of this
497 * structure and the 'cache' member must be the first as it
498 * is used to determine the offset of the writethrough fields.
502 struct dm_bio_details bio_details;
505 struct dm_cache_migration {
506 struct continuation k;
509 struct policy_work *op;
510 struct bio *overwrite_bio;
511 struct dm_bio_prison_cell_v2 *cell;
513 dm_cblock_t invalidate_cblock;
514 dm_oblock_t invalidate_oblock;
517 /*----------------------------------------------------------------*/
519 static bool writethrough_mode(struct cache_features *f)
521 return f->io_mode == CM_IO_WRITETHROUGH;
524 static bool writeback_mode(struct cache_features *f)
526 return f->io_mode == CM_IO_WRITEBACK;
529 static inline bool passthrough_mode(struct cache_features *f)
531 return unlikely(f->io_mode == CM_IO_PASSTHROUGH);
534 /*----------------------------------------------------------------*/
536 static void wake_deferred_bio_worker(struct cache *cache)
538 queue_work(cache->wq, &cache->deferred_bio_worker);
541 static void wake_deferred_writethrough_worker(struct cache *cache)
543 queue_work(cache->wq, &cache->deferred_writethrough_worker);
546 static void wake_migration_worker(struct cache *cache)
548 if (passthrough_mode(&cache->features))
551 queue_work(cache->wq, &cache->migration_worker);
554 /*----------------------------------------------------------------*/
556 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
558 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
561 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
563 dm_bio_prison_free_cell_v2(cache->prison, cell);
566 static struct dm_cache_migration *alloc_migration(struct cache *cache)
568 struct dm_cache_migration *mg;
570 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
573 atomic_inc(&mg->cache->nr_allocated_migrations);
579 static void free_migration(struct dm_cache_migration *mg)
581 struct cache *cache = mg->cache;
583 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
584 wake_up(&cache->migration_wait);
586 mempool_free(mg, cache->migration_pool);
589 /*----------------------------------------------------------------*/
591 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
593 return to_oblock(from_oblock(b) + 1ull);
596 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
600 key->block_begin = from_oblock(begin);
601 key->block_end = from_oblock(end);
605 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
606 * level 1 which prevents *both* READs and WRITEs.
608 #define WRITE_LOCK_LEVEL 0
609 #define READ_WRITE_LOCK_LEVEL 1
611 static unsigned lock_level(struct bio *bio)
613 return bio_data_dir(bio) == WRITE ?
615 READ_WRITE_LOCK_LEVEL;
618 /*----------------------------------------------------------------
620 *--------------------------------------------------------------*/
623 * If using writeback, leave out struct per_bio_data's writethrough fields.
625 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
626 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
628 static size_t get_per_bio_data_size(struct cache *cache)
630 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
633 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
635 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
640 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
642 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
645 pb->req_nr = dm_bio_get_target_bio_nr(bio);
652 /*----------------------------------------------------------------*/
654 static void defer_bio(struct cache *cache, struct bio *bio)
658 spin_lock_irqsave(&cache->lock, flags);
659 bio_list_add(&cache->deferred_bios, bio);
660 spin_unlock_irqrestore(&cache->lock, flags);
662 wake_deferred_bio_worker(cache);
665 static void defer_bios(struct cache *cache, struct bio_list *bios)
669 spin_lock_irqsave(&cache->lock, flags);
670 bio_list_merge(&cache->deferred_bios, bios);
672 spin_unlock_irqrestore(&cache->lock, flags);
674 wake_deferred_bio_worker(cache);
677 /*----------------------------------------------------------------*/
679 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
683 struct per_bio_data *pb;
684 struct dm_cell_key_v2 key;
685 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
686 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
688 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
689 if (!cell_prealloc) {
690 defer_bio(cache, bio);
694 build_key(oblock, end, &key);
695 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
698 * Failed to get the lock.
700 free_prison_cell(cache, cell_prealloc);
704 if (cell != cell_prealloc)
705 free_prison_cell(cache, cell_prealloc);
707 pb_size = get_per_bio_data_size(cache);
708 pb = get_per_bio_data(bio, pb_size);
714 /*----------------------------------------------------------------*/
716 static bool is_dirty(struct cache *cache, dm_cblock_t b)
718 return test_bit(from_cblock(b), cache->dirty_bitset);
721 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
723 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
724 atomic_inc(&cache->nr_dirty);
725 policy_set_dirty(cache->policy, cblock);
730 * These two are called when setting after migrations to force the policy
731 * and dirty bitset to be in sync.
733 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
735 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
736 atomic_inc(&cache->nr_dirty);
737 policy_set_dirty(cache->policy, cblock);
740 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
742 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
743 if (atomic_dec_return(&cache->nr_dirty) == 0)
744 dm_table_event(cache->ti->table);
747 policy_clear_dirty(cache->policy, cblock);
750 /*----------------------------------------------------------------*/
752 static bool block_size_is_power_of_two(struct cache *cache)
754 return cache->sectors_per_block_shift >= 0;
757 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
758 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
761 static dm_block_t block_div(dm_block_t b, uint32_t n)
768 static dm_block_t oblocks_per_dblock(struct cache *cache)
770 dm_block_t oblocks = cache->discard_block_size;
772 if (block_size_is_power_of_two(cache))
773 oblocks >>= cache->sectors_per_block_shift;
775 oblocks = block_div(oblocks, cache->sectors_per_block);
780 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
782 return to_dblock(block_div(from_oblock(oblock),
783 oblocks_per_dblock(cache)));
786 static void set_discard(struct cache *cache, dm_dblock_t b)
790 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
791 atomic_inc(&cache->stats.discard_count);
793 spin_lock_irqsave(&cache->lock, flags);
794 set_bit(from_dblock(b), cache->discard_bitset);
795 spin_unlock_irqrestore(&cache->lock, flags);
798 static void clear_discard(struct cache *cache, dm_dblock_t b)
802 spin_lock_irqsave(&cache->lock, flags);
803 clear_bit(from_dblock(b), cache->discard_bitset);
804 spin_unlock_irqrestore(&cache->lock, flags);
807 static bool is_discarded(struct cache *cache, dm_dblock_t b)
812 spin_lock_irqsave(&cache->lock, flags);
813 r = test_bit(from_dblock(b), cache->discard_bitset);
814 spin_unlock_irqrestore(&cache->lock, flags);
819 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
824 spin_lock_irqsave(&cache->lock, flags);
825 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
826 cache->discard_bitset);
827 spin_unlock_irqrestore(&cache->lock, flags);
832 /*----------------------------------------------------------------
834 *--------------------------------------------------------------*/
835 static void remap_to_origin(struct cache *cache, struct bio *bio)
837 bio->bi_bdev = cache->origin_dev->bdev;
840 static void remap_to_cache(struct cache *cache, struct bio *bio,
843 sector_t bi_sector = bio->bi_iter.bi_sector;
844 sector_t block = from_cblock(cblock);
846 bio->bi_bdev = cache->cache_dev->bdev;
847 if (!block_size_is_power_of_two(cache))
848 bio->bi_iter.bi_sector =
849 (block * cache->sectors_per_block) +
850 sector_div(bi_sector, cache->sectors_per_block);
852 bio->bi_iter.bi_sector =
853 (block << cache->sectors_per_block_shift) |
854 (bi_sector & (cache->sectors_per_block - 1));
857 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
860 size_t pb_data_size = get_per_bio_data_size(cache);
861 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
863 spin_lock_irqsave(&cache->lock, flags);
864 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
865 bio_op(bio) != REQ_OP_DISCARD) {
867 cache->need_tick_bio = false;
869 spin_unlock_irqrestore(&cache->lock, flags);
872 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
875 // FIXME: this is called way too much.
876 check_if_tick_bio_needed(cache, bio);
877 remap_to_origin(cache, bio);
878 if (bio_data_dir(bio) == WRITE)
879 clear_discard(cache, oblock_to_dblock(cache, oblock));
882 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
883 dm_oblock_t oblock, dm_cblock_t cblock)
885 check_if_tick_bio_needed(cache, bio);
886 remap_to_cache(cache, bio, cblock);
887 if (bio_data_dir(bio) == WRITE) {
888 set_dirty(cache, cblock);
889 clear_discard(cache, oblock_to_dblock(cache, oblock));
893 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
895 sector_t block_nr = bio->bi_iter.bi_sector;
897 if (!block_size_is_power_of_two(cache))
898 (void) sector_div(block_nr, cache->sectors_per_block);
900 block_nr >>= cache->sectors_per_block_shift;
902 return to_oblock(block_nr);
905 static bool accountable_bio(struct cache *cache, struct bio *bio)
907 return bio_op(bio) != REQ_OP_DISCARD;
910 static void accounted_begin(struct cache *cache, struct bio *bio)
912 size_t pb_data_size = get_per_bio_data_size(cache);
913 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
915 if (accountable_bio(cache, bio)) {
916 pb->len = bio_sectors(bio);
917 iot_io_begin(&cache->tracker, pb->len);
921 static void accounted_complete(struct cache *cache, struct bio *bio)
923 size_t pb_data_size = get_per_bio_data_size(cache);
924 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
926 iot_io_end(&cache->tracker, pb->len);
929 static void accounted_request(struct cache *cache, struct bio *bio)
931 accounted_begin(cache, bio);
932 generic_make_request(bio);
935 static void issue_op(struct bio *bio, void *context)
937 struct cache *cache = context;
938 accounted_request(cache, bio);
941 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
945 spin_lock_irqsave(&cache->lock, flags);
946 bio_list_add(&cache->deferred_writethrough_bios, bio);
947 spin_unlock_irqrestore(&cache->lock, flags);
949 wake_deferred_writethrough_worker(cache);
952 static void writethrough_endio(struct bio *bio)
954 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
956 dm_unhook_bio(&pb->hook_info, bio);
963 dm_bio_restore(&pb->bio_details, bio);
964 remap_to_cache(pb->cache, bio, pb->cblock);
967 * We can't issue this bio directly, since we're in interrupt
968 * context. So it gets put on a bio list for processing by the
971 defer_writethrough_bio(pb->cache, bio);
975 * FIXME: send in parallel, huge latency as is.
976 * When running in writethrough mode we need to send writes to clean blocks
977 * to both the cache and origin devices. In future we'd like to clone the
978 * bio and send them in parallel, but for now we're doing them in
979 * series as this is easier.
981 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
982 dm_oblock_t oblock, dm_cblock_t cblock)
984 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
988 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
989 dm_bio_record(&pb->bio_details, bio);
991 remap_to_origin_clear_discard(pb->cache, bio, oblock);
994 /*----------------------------------------------------------------
996 *--------------------------------------------------------------*/
997 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
999 return cache->features.mode;
1002 static const char *cache_device_name(struct cache *cache)
1004 return dm_device_name(dm_table_get_md(cache->ti->table));
1007 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
1009 const char *descs[] = {
1015 dm_table_event(cache->ti->table);
1016 DMINFO("%s: switching cache to %s mode",
1017 cache_device_name(cache), descs[(int)mode]);
1020 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
1023 enum cache_metadata_mode old_mode = get_cache_mode(cache);
1025 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
1026 DMERR("%s: unable to read needs_check flag, setting failure mode.",
1027 cache_device_name(cache));
1031 if (new_mode == CM_WRITE && needs_check) {
1032 DMERR("%s: unable to switch cache to write mode until repaired.",
1033 cache_device_name(cache));
1034 if (old_mode != new_mode)
1035 new_mode = old_mode;
1037 new_mode = CM_READ_ONLY;
1040 /* Never move out of fail mode */
1041 if (old_mode == CM_FAIL)
1047 dm_cache_metadata_set_read_only(cache->cmd);
1051 dm_cache_metadata_set_read_write(cache->cmd);
1055 cache->features.mode = new_mode;
1057 if (new_mode != old_mode)
1058 notify_mode_switch(cache, new_mode);
1061 static void abort_transaction(struct cache *cache)
1063 const char *dev_name = cache_device_name(cache);
1065 if (get_cache_mode(cache) >= CM_READ_ONLY)
1068 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1069 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1070 set_cache_mode(cache, CM_FAIL);
1073 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1074 if (dm_cache_metadata_abort(cache->cmd)) {
1075 DMERR("%s: failed to abort metadata transaction", dev_name);
1076 set_cache_mode(cache, CM_FAIL);
1080 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1082 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1083 cache_device_name(cache), op, r);
1084 abort_transaction(cache);
1085 set_cache_mode(cache, CM_READ_ONLY);
1088 /*----------------------------------------------------------------*/
1090 static void load_stats(struct cache *cache)
1092 struct dm_cache_statistics stats;
1094 dm_cache_metadata_get_stats(cache->cmd, &stats);
1095 atomic_set(&cache->stats.read_hit, stats.read_hits);
1096 atomic_set(&cache->stats.read_miss, stats.read_misses);
1097 atomic_set(&cache->stats.write_hit, stats.write_hits);
1098 atomic_set(&cache->stats.write_miss, stats.write_misses);
1101 static void save_stats(struct cache *cache)
1103 struct dm_cache_statistics stats;
1105 if (get_cache_mode(cache) >= CM_READ_ONLY)
1108 stats.read_hits = atomic_read(&cache->stats.read_hit);
1109 stats.read_misses = atomic_read(&cache->stats.read_miss);
1110 stats.write_hits = atomic_read(&cache->stats.write_hit);
1111 stats.write_misses = atomic_read(&cache->stats.write_miss);
1113 dm_cache_metadata_set_stats(cache->cmd, &stats);
1116 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1119 case POLICY_PROMOTE:
1120 atomic_inc(&stats->promotion);
1124 atomic_inc(&stats->demotion);
1127 case POLICY_WRITEBACK:
1128 atomic_inc(&stats->writeback);
1133 /*----------------------------------------------------------------
1134 * Migration processing
1136 * Migration covers moving data from the origin device to the cache, or
1138 *--------------------------------------------------------------*/
1140 static void inc_io_migrations(struct cache *cache)
1142 atomic_inc(&cache->nr_io_migrations);
1145 static void dec_io_migrations(struct cache *cache)
1147 atomic_dec(&cache->nr_io_migrations);
1150 static bool discard_or_flush(struct bio *bio)
1152 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1155 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1156 dm_dblock_t *b, dm_dblock_t *e)
1158 sector_t sb = bio->bi_iter.bi_sector;
1159 sector_t se = bio_end_sector(bio);
1161 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1163 if (se - sb < cache->discard_block_size)
1166 *e = to_dblock(block_div(se, cache->discard_block_size));
1169 /*----------------------------------------------------------------*/
1171 static void prevent_background_work(struct cache *cache)
1174 down_write(&cache->background_work_lock);
1178 static void allow_background_work(struct cache *cache)
1181 up_write(&cache->background_work_lock);
1185 static bool background_work_begin(struct cache *cache)
1190 r = down_read_trylock(&cache->background_work_lock);
1196 static void background_work_end(struct cache *cache)
1199 up_read(&cache->background_work_lock);
1203 /*----------------------------------------------------------------*/
1205 static void quiesce(struct dm_cache_migration *mg,
1206 void (*continuation)(struct work_struct *))
1208 init_continuation(&mg->k, continuation);
1209 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1212 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1214 struct continuation *k = container_of(ws, struct continuation, ws);
1215 return container_of(k, struct dm_cache_migration, k);
1218 static void copy_complete(int read_err, unsigned long write_err, void *context)
1220 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1222 if (read_err || write_err)
1225 queue_continuation(mg->cache->wq, &mg->k);
1228 static int copy(struct dm_cache_migration *mg, bool promote)
1231 struct dm_io_region o_region, c_region;
1232 struct cache *cache = mg->cache;
1234 o_region.bdev = cache->origin_dev->bdev;
1235 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1236 o_region.count = cache->sectors_per_block;
1238 c_region.bdev = cache->cache_dev->bdev;
1239 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1240 c_region.count = cache->sectors_per_block;
1243 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1245 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1250 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1252 size_t pb_data_size = get_per_bio_data_size(cache);
1253 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1255 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1256 free_prison_cell(cache, pb->cell);
1260 static void overwrite_endio(struct bio *bio)
1262 struct dm_cache_migration *mg = bio->bi_private;
1263 struct cache *cache = mg->cache;
1264 size_t pb_data_size = get_per_bio_data_size(cache);
1265 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1267 dm_unhook_bio(&pb->hook_info, bio);
1270 mg->k.input = bio->bi_error;
1272 queue_continuation(mg->cache->wq, &mg->k);
1275 static void overwrite(struct dm_cache_migration *mg,
1276 void (*continuation)(struct work_struct *))
1278 struct bio *bio = mg->overwrite_bio;
1279 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1280 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1282 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1285 * The overwrite bio is part of the copy operation, as such it does
1286 * not set/clear discard or dirty flags.
1288 if (mg->op->op == POLICY_PROMOTE)
1289 remap_to_cache(mg->cache, bio, mg->op->cblock);
1291 remap_to_origin(mg->cache, bio);
1293 init_continuation(&mg->k, continuation);
1294 accounted_request(mg->cache, bio);
1300 * 1) exclusive lock preventing WRITEs
1302 * 3) copy or issue overwrite bio
1303 * 4) upgrade to exclusive lock preventing READs and WRITEs
1305 * 6) update metadata and commit
1308 static void mg_complete(struct dm_cache_migration *mg, bool success)
1310 struct bio_list bios;
1311 struct cache *cache = mg->cache;
1312 struct policy_work *op = mg->op;
1313 dm_cblock_t cblock = op->cblock;
1316 update_stats(&cache->stats, op->op);
1319 case POLICY_PROMOTE:
1320 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1321 policy_complete_background_work(cache->policy, op, success);
1323 if (mg->overwrite_bio) {
1325 force_set_dirty(cache, cblock);
1327 mg->overwrite_bio->bi_error = (mg->k.input ? : -EIO);
1328 bio_endio(mg->overwrite_bio);
1331 force_clear_dirty(cache, cblock);
1332 dec_io_migrations(cache);
1338 * We clear dirty here to update the nr_dirty counter.
1341 force_clear_dirty(cache, cblock);
1342 policy_complete_background_work(cache->policy, op, success);
1343 dec_io_migrations(cache);
1346 case POLICY_WRITEBACK:
1348 force_clear_dirty(cache, cblock);
1349 policy_complete_background_work(cache->policy, op, success);
1350 dec_io_migrations(cache);
1354 bio_list_init(&bios);
1356 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1357 free_prison_cell(cache, mg->cell);
1361 defer_bios(cache, &bios);
1362 wake_migration_worker(cache);
1364 background_work_end(cache);
1367 static void mg_success(struct work_struct *ws)
1369 struct dm_cache_migration *mg = ws_to_mg(ws);
1370 mg_complete(mg, mg->k.input == 0);
1373 static void mg_update_metadata(struct work_struct *ws)
1376 struct dm_cache_migration *mg = ws_to_mg(ws);
1377 struct cache *cache = mg->cache;
1378 struct policy_work *op = mg->op;
1381 case POLICY_PROMOTE:
1382 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1384 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1385 cache_device_name(cache));
1386 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1388 mg_complete(mg, false);
1391 mg_complete(mg, true);
1395 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1397 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1398 cache_device_name(cache));
1399 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1401 mg_complete(mg, false);
1406 * It would be nice if we only had to commit when a REQ_FLUSH
1407 * comes through. But there's one scenario that we have to
1410 * - vblock x in a cache block
1412 * - cache block gets reallocated and over written
1415 * When we recover, because there was no commit the cache will
1416 * rollback to having the data for vblock x in the cache block.
1417 * But the cache block has since been overwritten, so it'll end
1418 * up pointing to data that was never in 'x' during the history
1421 * To avoid this issue we require a commit as part of the
1422 * demotion operation.
1424 init_continuation(&mg->k, mg_success);
1425 continue_after_commit(&cache->committer, &mg->k);
1426 schedule_commit(&cache->committer);
1429 case POLICY_WRITEBACK:
1430 mg_complete(mg, true);
1435 static void mg_update_metadata_after_copy(struct work_struct *ws)
1437 struct dm_cache_migration *mg = ws_to_mg(ws);
1440 * Did the copy succeed?
1443 mg_complete(mg, false);
1445 mg_update_metadata(ws);
1448 static void mg_upgrade_lock(struct work_struct *ws)
1451 struct dm_cache_migration *mg = ws_to_mg(ws);
1454 * Did the copy succeed?
1457 mg_complete(mg, false);
1461 * Now we want the lock to prevent both reads and writes.
1463 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1464 READ_WRITE_LOCK_LEVEL);
1466 mg_complete(mg, false);
1469 quiesce(mg, mg_update_metadata);
1472 mg_update_metadata(ws);
1476 static void mg_copy(struct work_struct *ws)
1479 struct dm_cache_migration *mg = ws_to_mg(ws);
1481 if (mg->overwrite_bio) {
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);
1492 struct cache *cache = mg->cache;
1493 struct policy_work *op = mg->op;
1494 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1496 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1497 is_discarded_oblock(cache, op->oblock)) {
1498 mg_upgrade_lock(ws);
1502 init_continuation(&mg->k, mg_upgrade_lock);
1504 r = copy(mg, is_policy_promote);
1506 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1508 mg_complete(mg, false);
1513 static int mg_lock_writes(struct dm_cache_migration *mg)
1516 struct dm_cell_key_v2 key;
1517 struct cache *cache = mg->cache;
1518 struct dm_bio_prison_cell_v2 *prealloc;
1520 prealloc = alloc_prison_cell(cache);
1522 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1523 mg_complete(mg, false);
1528 * Prevent writes to the block, but allow reads to continue.
1529 * Unless we're using an overwrite bio, in which case we lock
1532 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1533 r = dm_cell_lock_v2(cache->prison, &key,
1534 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1535 prealloc, &mg->cell);
1537 free_prison_cell(cache, prealloc);
1538 mg_complete(mg, false);
1542 if (mg->cell != prealloc)
1543 free_prison_cell(cache, prealloc);
1548 quiesce(mg, mg_copy);
1553 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1555 struct dm_cache_migration *mg;
1557 if (!background_work_begin(cache)) {
1558 policy_complete_background_work(cache->policy, op, false);
1562 mg = alloc_migration(cache);
1564 policy_complete_background_work(cache->policy, op, false);
1565 background_work_end(cache);
1569 memset(mg, 0, sizeof(*mg));
1573 mg->overwrite_bio = bio;
1576 inc_io_migrations(cache);
1578 return mg_lock_writes(mg);
1581 /*----------------------------------------------------------------
1582 * invalidation processing
1583 *--------------------------------------------------------------*/
1585 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1587 struct bio_list bios;
1588 struct cache *cache = mg->cache;
1590 bio_list_init(&bios);
1591 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1592 free_prison_cell(cache, mg->cell);
1594 if (!success && mg->overwrite_bio)
1595 bio_io_error(mg->overwrite_bio);
1598 defer_bios(cache, &bios);
1600 background_work_end(cache);
1603 static void invalidate_completed(struct work_struct *ws)
1605 struct dm_cache_migration *mg = ws_to_mg(ws);
1606 invalidate_complete(mg, !mg->k.input);
1609 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1611 int r = policy_invalidate_mapping(cache->policy, cblock);
1613 r = dm_cache_remove_mapping(cache->cmd, cblock);
1615 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1616 cache_device_name(cache));
1617 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1620 } else if (r == -ENODATA) {
1622 * Harmless, already unmapped.
1627 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1632 static void invalidate_remove(struct work_struct *ws)
1635 struct dm_cache_migration *mg = ws_to_mg(ws);
1636 struct cache *cache = mg->cache;
1638 r = invalidate_cblock(cache, mg->invalidate_cblock);
1640 invalidate_complete(mg, false);
1644 init_continuation(&mg->k, invalidate_completed);
1645 continue_after_commit(&cache->committer, &mg->k);
1646 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1647 mg->overwrite_bio = NULL;
1648 schedule_commit(&cache->committer);
1651 static int invalidate_lock(struct dm_cache_migration *mg)
1654 struct dm_cell_key_v2 key;
1655 struct cache *cache = mg->cache;
1656 struct dm_bio_prison_cell_v2 *prealloc;
1658 prealloc = alloc_prison_cell(cache);
1660 invalidate_complete(mg, false);
1664 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1665 r = dm_cell_lock_v2(cache->prison, &key,
1666 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1668 free_prison_cell(cache, prealloc);
1669 invalidate_complete(mg, false);
1673 if (mg->cell != prealloc)
1674 free_prison_cell(cache, prealloc);
1677 quiesce(mg, invalidate_remove);
1681 * We can't call invalidate_remove() directly here because we
1682 * might still be in request context.
1684 init_continuation(&mg->k, invalidate_remove);
1685 queue_work(cache->wq, &mg->k.ws);
1691 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1692 dm_oblock_t oblock, struct bio *bio)
1694 struct dm_cache_migration *mg;
1696 if (!background_work_begin(cache))
1699 mg = alloc_migration(cache);
1701 background_work_end(cache);
1705 memset(mg, 0, sizeof(*mg));
1708 mg->overwrite_bio = bio;
1709 mg->invalidate_cblock = cblock;
1710 mg->invalidate_oblock = oblock;
1712 return invalidate_lock(mg);
1715 /*----------------------------------------------------------------
1717 *--------------------------------------------------------------*/
1725 static enum busy spare_migration_bandwidth(struct cache *cache)
1727 bool idle = iot_idle_for(&cache->tracker, HZ);
1728 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1729 cache->sectors_per_block;
1731 if (current_volume <= cache->migration_threshold)
1732 return idle ? IDLE : MODERATE;
1734 return idle ? MODERATE : BUSY;
1737 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1739 atomic_inc(bio_data_dir(bio) == READ ?
1740 &cache->stats.read_hit : &cache->stats.write_hit);
1743 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1745 atomic_inc(bio_data_dir(bio) == READ ?
1746 &cache->stats.read_miss : &cache->stats.write_miss);
1749 /*----------------------------------------------------------------*/
1751 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1753 return (bio_data_dir(bio) == WRITE) &&
1754 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1757 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1759 return writeback_mode(&cache->features) &&
1760 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1763 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1764 bool *commit_needed)
1767 bool rb, background_queued;
1769 size_t pb_data_size = get_per_bio_data_size(cache);
1770 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1772 *commit_needed = false;
1774 rb = bio_detain_shared(cache, block, bio);
1777 * An exclusive lock is held for this block, so we have to
1778 * wait. We set the commit_needed flag so the current
1779 * transaction will be committed asap, allowing this lock
1782 *commit_needed = true;
1783 return DM_MAPIO_SUBMITTED;
1786 data_dir = bio_data_dir(bio);
1788 if (optimisable_bio(cache, bio, block)) {
1789 struct policy_work *op = NULL;
1791 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1792 if (unlikely(r && r != -ENOENT)) {
1793 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1794 cache_device_name(cache), r);
1796 return DM_MAPIO_SUBMITTED;
1799 if (r == -ENOENT && op) {
1800 bio_drop_shared_lock(cache, bio);
1801 BUG_ON(op->op != POLICY_PROMOTE);
1802 mg_start(cache, op, bio);
1803 return DM_MAPIO_SUBMITTED;
1806 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1807 if (unlikely(r && r != -ENOENT)) {
1808 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1809 cache_device_name(cache), r);
1811 return DM_MAPIO_SUBMITTED;
1814 if (background_queued)
1815 wake_migration_worker(cache);
1822 inc_miss_counter(cache, bio);
1823 if (pb->req_nr == 0) {
1824 accounted_begin(cache, bio);
1825 remap_to_origin_clear_discard(cache, bio, block);
1829 * This is a duplicate writethrough io that is no
1830 * longer needed because the block has been demoted.
1833 return DM_MAPIO_SUBMITTED;
1839 inc_hit_counter(cache, bio);
1842 * Passthrough always maps to the origin, invalidating any
1843 * cache blocks that are written to.
1845 if (passthrough_mode(&cache->features)) {
1846 if (bio_data_dir(bio) == WRITE) {
1847 bio_drop_shared_lock(cache, bio);
1848 atomic_inc(&cache->stats.demotion);
1849 invalidate_start(cache, cblock, block, bio);
1851 remap_to_origin_clear_discard(cache, bio, block);
1854 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
1855 !is_dirty(cache, cblock)) {
1856 remap_to_origin_then_cache(cache, bio, block, cblock);
1857 accounted_begin(cache, bio);
1859 remap_to_cache_dirty(cache, bio, block, cblock);
1864 * dm core turns FUA requests into a separate payload and FLUSH req.
1866 if (bio->bi_opf & REQ_FUA) {
1868 * issue_after_commit will call accounted_begin a second time. So
1869 * we call accounted_complete() to avoid double accounting.
1871 accounted_complete(cache, bio);
1872 issue_after_commit(&cache->committer, bio);
1873 *commit_needed = true;
1874 return DM_MAPIO_SUBMITTED;
1877 return DM_MAPIO_REMAPPED;
1880 static bool process_bio(struct cache *cache, struct bio *bio)
1884 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1885 generic_make_request(bio);
1887 return commit_needed;
1891 * A non-zero return indicates read_only or fail_io mode.
1893 static int commit(struct cache *cache, bool clean_shutdown)
1897 if (get_cache_mode(cache) >= CM_READ_ONLY)
1900 atomic_inc(&cache->stats.commit_count);
1901 r = dm_cache_commit(cache->cmd, clean_shutdown);
1903 metadata_operation_failed(cache, "dm_cache_commit", r);
1909 * Used by the batcher.
1911 static int commit_op(void *context)
1913 struct cache *cache = context;
1915 if (dm_cache_changed_this_transaction(cache->cmd))
1916 return commit(cache, false);
1921 /*----------------------------------------------------------------*/
1923 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1925 size_t pb_data_size = get_per_bio_data_size(cache);
1926 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1929 remap_to_origin(cache, bio);
1931 remap_to_cache(cache, bio, 0);
1933 issue_after_commit(&cache->committer, bio);
1937 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1941 // FIXME: do we need to lock the region? Or can we just assume the
1942 // user wont be so foolish as to issue discard concurrently with
1944 calc_discard_block_range(cache, bio, &b, &e);
1946 set_discard(cache, b);
1947 b = to_dblock(from_dblock(b) + 1);
1955 static void process_deferred_bios(struct work_struct *ws)
1957 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1959 unsigned long flags;
1960 bool commit_needed = false;
1961 struct bio_list bios;
1964 bio_list_init(&bios);
1966 spin_lock_irqsave(&cache->lock, flags);
1967 bio_list_merge(&bios, &cache->deferred_bios);
1968 bio_list_init(&cache->deferred_bios);
1969 spin_unlock_irqrestore(&cache->lock, flags);
1971 while ((bio = bio_list_pop(&bios))) {
1972 if (bio->bi_opf & REQ_PREFLUSH)
1973 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1975 else if (bio_op(bio) == REQ_OP_DISCARD)
1976 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1979 commit_needed = process_bio(cache, bio) || commit_needed;
1983 schedule_commit(&cache->committer);
1986 static void process_deferred_writethrough_bios(struct work_struct *ws)
1988 struct cache *cache = container_of(ws, struct cache, deferred_writethrough_worker);
1990 unsigned long flags;
1991 struct bio_list bios;
1994 bio_list_init(&bios);
1996 spin_lock_irqsave(&cache->lock, flags);
1997 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1998 bio_list_init(&cache->deferred_writethrough_bios);
1999 spin_unlock_irqrestore(&cache->lock, flags);
2002 * These bios have already been through accounted_begin()
2004 while ((bio = bio_list_pop(&bios)))
2005 generic_make_request(bio);
2008 /*----------------------------------------------------------------
2010 *--------------------------------------------------------------*/
2012 static void requeue_deferred_bios(struct cache *cache)
2015 struct bio_list bios;
2017 bio_list_init(&bios);
2018 bio_list_merge(&bios, &cache->deferred_bios);
2019 bio_list_init(&cache->deferred_bios);
2021 while ((bio = bio_list_pop(&bios))) {
2022 bio->bi_error = DM_ENDIO_REQUEUE;
2028 * We want to commit periodically so that not too much
2029 * unwritten metadata builds up.
2031 static void do_waker(struct work_struct *ws)
2033 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2035 policy_tick(cache->policy, true);
2036 wake_migration_worker(cache);
2037 schedule_commit(&cache->committer);
2038 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2041 static void check_migrations(struct work_struct *ws)
2044 struct policy_work *op;
2045 struct cache *cache = container_of(ws, struct cache, migration_worker);
2049 b = spare_migration_bandwidth(cache);
2053 r = policy_get_background_work(cache->policy, b == IDLE, &op);
2058 DMERR_LIMIT("%s: policy_background_work failed",
2059 cache_device_name(cache));
2063 r = mg_start(cache, op, NULL);
2069 /*----------------------------------------------------------------
2071 *--------------------------------------------------------------*/
2074 * This function gets called on the error paths of the constructor, so we
2075 * have to cope with a partially initialised struct.
2077 static void destroy(struct cache *cache)
2081 mempool_destroy(cache->migration_pool);
2084 dm_bio_prison_destroy_v2(cache->prison);
2087 destroy_workqueue(cache->wq);
2089 if (cache->dirty_bitset)
2090 free_bitset(cache->dirty_bitset);
2092 if (cache->discard_bitset)
2093 free_bitset(cache->discard_bitset);
2096 dm_kcopyd_client_destroy(cache->copier);
2099 dm_cache_metadata_close(cache->cmd);
2101 if (cache->metadata_dev)
2102 dm_put_device(cache->ti, cache->metadata_dev);
2104 if (cache->origin_dev)
2105 dm_put_device(cache->ti, cache->origin_dev);
2107 if (cache->cache_dev)
2108 dm_put_device(cache->ti, cache->cache_dev);
2111 dm_cache_policy_destroy(cache->policy);
2113 for (i = 0; i < cache->nr_ctr_args ; i++)
2114 kfree(cache->ctr_args[i]);
2115 kfree(cache->ctr_args);
2120 static void cache_dtr(struct dm_target *ti)
2122 struct cache *cache = ti->private;
2127 static sector_t get_dev_size(struct dm_dev *dev)
2129 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2132 /*----------------------------------------------------------------*/
2135 * Construct a cache device mapping.
2137 * cache <metadata dev> <cache dev> <origin dev> <block size>
2138 * <#feature args> [<feature arg>]*
2139 * <policy> <#policy args> [<policy arg>]*
2141 * metadata dev : fast device holding the persistent metadata
2142 * cache dev : fast device holding cached data blocks
2143 * origin dev : slow device holding original data blocks
2144 * block size : cache unit size in sectors
2146 * #feature args : number of feature arguments passed
2147 * feature args : writethrough. (The default is writeback.)
2149 * policy : the replacement policy to use
2150 * #policy args : an even number of policy arguments corresponding
2151 * to key/value pairs passed to the policy
2152 * policy args : key/value pairs passed to the policy
2153 * E.g. 'sequential_threshold 1024'
2154 * See cache-policies.txt for details.
2156 * Optional feature arguments are:
2157 * writethrough : write through caching that prohibits cache block
2158 * content from being different from origin block content.
2159 * Without this argument, the default behaviour is to write
2160 * back cache block contents later for performance reasons,
2161 * so they may differ from the corresponding origin blocks.
2164 struct dm_target *ti;
2166 struct dm_dev *metadata_dev;
2168 struct dm_dev *cache_dev;
2169 sector_t cache_sectors;
2171 struct dm_dev *origin_dev;
2172 sector_t origin_sectors;
2174 uint32_t block_size;
2176 const char *policy_name;
2178 const char **policy_argv;
2180 struct cache_features features;
2183 static void destroy_cache_args(struct cache_args *ca)
2185 if (ca->metadata_dev)
2186 dm_put_device(ca->ti, ca->metadata_dev);
2189 dm_put_device(ca->ti, ca->cache_dev);
2192 dm_put_device(ca->ti, ca->origin_dev);
2197 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2200 *error = "Insufficient args";
2207 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2211 sector_t metadata_dev_size;
2212 char b[BDEVNAME_SIZE];
2214 if (!at_least_one_arg(as, error))
2217 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2220 *error = "Error opening metadata device";
2224 metadata_dev_size = get_dev_size(ca->metadata_dev);
2225 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2226 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2227 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2232 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2237 if (!at_least_one_arg(as, error))
2240 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2243 *error = "Error opening cache device";
2246 ca->cache_sectors = get_dev_size(ca->cache_dev);
2251 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2256 if (!at_least_one_arg(as, error))
2259 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2262 *error = "Error opening origin device";
2266 ca->origin_sectors = get_dev_size(ca->origin_dev);
2267 if (ca->ti->len > ca->origin_sectors) {
2268 *error = "Device size larger than cached device";
2275 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2278 unsigned long block_size;
2280 if (!at_least_one_arg(as, error))
2283 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2284 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2285 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2286 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2287 *error = "Invalid data block size";
2291 if (block_size > ca->cache_sectors) {
2292 *error = "Data block size is larger than the cache device";
2296 ca->block_size = block_size;
2301 static void init_features(struct cache_features *cf)
2303 cf->mode = CM_WRITE;
2304 cf->io_mode = CM_IO_WRITEBACK;
2305 cf->metadata_version = 1;
2308 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2311 static struct dm_arg _args[] = {
2312 {0, 2, "Invalid number of cache feature arguments"},
2318 struct cache_features *cf = &ca->features;
2322 r = dm_read_arg_group(_args, as, &argc, error);
2327 arg = dm_shift_arg(as);
2329 if (!strcasecmp(arg, "writeback"))
2330 cf->io_mode = CM_IO_WRITEBACK;
2332 else if (!strcasecmp(arg, "writethrough"))
2333 cf->io_mode = CM_IO_WRITETHROUGH;
2335 else if (!strcasecmp(arg, "passthrough"))
2336 cf->io_mode = CM_IO_PASSTHROUGH;
2338 else if (!strcasecmp(arg, "metadata2"))
2339 cf->metadata_version = 2;
2342 *error = "Unrecognised cache feature requested";
2350 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2353 static struct dm_arg _args[] = {
2354 {0, 1024, "Invalid number of policy arguments"},
2359 if (!at_least_one_arg(as, error))
2362 ca->policy_name = dm_shift_arg(as);
2364 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2368 ca->policy_argv = (const char **)as->argv;
2369 dm_consume_args(as, ca->policy_argc);
2374 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2378 struct dm_arg_set as;
2383 r = parse_metadata_dev(ca, &as, error);
2387 r = parse_cache_dev(ca, &as, error);
2391 r = parse_origin_dev(ca, &as, error);
2395 r = parse_block_size(ca, &as, error);
2399 r = parse_features(ca, &as, error);
2403 r = parse_policy(ca, &as, error);
2410 /*----------------------------------------------------------------*/
2412 static struct kmem_cache *migration_cache;
2414 #define NOT_CORE_OPTION 1
2416 static int process_config_option(struct cache *cache, const char *key, const char *value)
2420 if (!strcasecmp(key, "migration_threshold")) {
2421 if (kstrtoul(value, 10, &tmp))
2424 cache->migration_threshold = tmp;
2428 return NOT_CORE_OPTION;
2431 static int set_config_value(struct cache *cache, const char *key, const char *value)
2433 int r = process_config_option(cache, key, value);
2435 if (r == NOT_CORE_OPTION)
2436 r = policy_set_config_value(cache->policy, key, value);
2439 DMWARN("bad config value for %s: %s", key, value);
2444 static int set_config_values(struct cache *cache, int argc, const char **argv)
2449 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2454 r = set_config_value(cache, argv[0], argv[1]);
2465 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2468 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2470 cache->origin_sectors,
2471 cache->sectors_per_block);
2473 *error = "Error creating cache's policy";
2477 BUG_ON(!cache->policy);
2483 * We want the discard block size to be at least the size of the cache
2484 * block size and have no more than 2^14 discard blocks across the origin.
2486 #define MAX_DISCARD_BLOCKS (1 << 14)
2488 static bool too_many_discard_blocks(sector_t discard_block_size,
2489 sector_t origin_size)
2491 (void) sector_div(origin_size, discard_block_size);
2493 return origin_size > MAX_DISCARD_BLOCKS;
2496 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2497 sector_t origin_size)
2499 sector_t discard_block_size = cache_block_size;
2502 while (too_many_discard_blocks(discard_block_size, origin_size))
2503 discard_block_size *= 2;
2505 return discard_block_size;
2508 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2510 dm_block_t nr_blocks = from_cblock(size);
2512 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2513 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2514 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2515 "Please consider increasing the cache block size to reduce the overall cache block count.",
2516 (unsigned long long) nr_blocks);
2518 cache->cache_size = size;
2521 static int is_congested(struct dm_dev *dev, int bdi_bits)
2523 struct request_queue *q = bdev_get_queue(dev->bdev);
2524 return bdi_congested(q->backing_dev_info, bdi_bits);
2527 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2529 struct cache *cache = container_of(cb, struct cache, callbacks);
2531 return is_congested(cache->origin_dev, bdi_bits) ||
2532 is_congested(cache->cache_dev, bdi_bits);
2535 #define DEFAULT_MIGRATION_THRESHOLD 2048
2537 static int cache_create(struct cache_args *ca, struct cache **result)
2540 char **error = &ca->ti->error;
2541 struct cache *cache;
2542 struct dm_target *ti = ca->ti;
2543 dm_block_t origin_blocks;
2544 struct dm_cache_metadata *cmd;
2545 bool may_format = ca->features.mode == CM_WRITE;
2547 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2552 ti->private = cache;
2553 ti->num_flush_bios = 2;
2554 ti->flush_supported = true;
2556 ti->num_discard_bios = 1;
2557 ti->discards_supported = true;
2558 ti->split_discard_bios = false;
2560 cache->features = ca->features;
2561 ti->per_io_data_size = get_per_bio_data_size(cache);
2563 cache->callbacks.congested_fn = cache_is_congested;
2564 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2566 cache->metadata_dev = ca->metadata_dev;
2567 cache->origin_dev = ca->origin_dev;
2568 cache->cache_dev = ca->cache_dev;
2570 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2572 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2573 origin_blocks = block_div(origin_blocks, ca->block_size);
2574 cache->origin_blocks = to_oblock(origin_blocks);
2576 cache->sectors_per_block = ca->block_size;
2577 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2582 if (ca->block_size & (ca->block_size - 1)) {
2583 dm_block_t cache_size = ca->cache_sectors;
2585 cache->sectors_per_block_shift = -1;
2586 cache_size = block_div(cache_size, ca->block_size);
2587 set_cache_size(cache, to_cblock(cache_size));
2589 cache->sectors_per_block_shift = __ffs(ca->block_size);
2590 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2593 r = create_cache_policy(cache, ca, error);
2597 cache->policy_nr_args = ca->policy_argc;
2598 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2600 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2602 *error = "Error setting cache policy's config values";
2606 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2607 ca->block_size, may_format,
2608 dm_cache_policy_get_hint_size(cache->policy),
2609 ca->features.metadata_version);
2611 *error = "Error creating metadata object";
2616 set_cache_mode(cache, CM_WRITE);
2617 if (get_cache_mode(cache) != CM_WRITE) {
2618 *error = "Unable to get write access to metadata, please check/repair metadata.";
2623 if (passthrough_mode(&cache->features)) {
2626 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2628 *error = "dm_cache_metadata_all_clean() failed";
2633 *error = "Cannot enter passthrough mode unless all blocks are clean";
2638 policy_allow_migrations(cache->policy, false);
2641 spin_lock_init(&cache->lock);
2642 INIT_LIST_HEAD(&cache->deferred_cells);
2643 bio_list_init(&cache->deferred_bios);
2644 bio_list_init(&cache->deferred_writethrough_bios);
2645 atomic_set(&cache->nr_allocated_migrations, 0);
2646 atomic_set(&cache->nr_io_migrations, 0);
2647 init_waitqueue_head(&cache->migration_wait);
2650 atomic_set(&cache->nr_dirty, 0);
2651 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2652 if (!cache->dirty_bitset) {
2653 *error = "could not allocate dirty bitset";
2656 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2658 cache->discard_block_size =
2659 calculate_discard_block_size(cache->sectors_per_block,
2660 cache->origin_sectors);
2661 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2662 cache->discard_block_size));
2663 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2664 if (!cache->discard_bitset) {
2665 *error = "could not allocate discard bitset";
2668 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2670 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2671 if (IS_ERR(cache->copier)) {
2672 *error = "could not create kcopyd client";
2673 r = PTR_ERR(cache->copier);
2677 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2679 *error = "could not create workqueue for metadata object";
2682 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2683 INIT_WORK(&cache->deferred_writethrough_worker,
2684 process_deferred_writethrough_bios);
2685 INIT_WORK(&cache->migration_worker, check_migrations);
2686 INIT_DELAYED_WORK(&cache->waker, do_waker);
2688 cache->prison = dm_bio_prison_create_v2(cache->wq);
2689 if (!cache->prison) {
2690 *error = "could not create bio prison";
2694 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2696 if (!cache->migration_pool) {
2697 *error = "Error creating cache's migration mempool";
2701 cache->need_tick_bio = true;
2702 cache->sized = false;
2703 cache->invalidate = false;
2704 cache->commit_requested = false;
2705 cache->loaded_mappings = false;
2706 cache->loaded_discards = false;
2710 atomic_set(&cache->stats.demotion, 0);
2711 atomic_set(&cache->stats.promotion, 0);
2712 atomic_set(&cache->stats.copies_avoided, 0);
2713 atomic_set(&cache->stats.cache_cell_clash, 0);
2714 atomic_set(&cache->stats.commit_count, 0);
2715 atomic_set(&cache->stats.discard_count, 0);
2717 spin_lock_init(&cache->invalidation_lock);
2718 INIT_LIST_HEAD(&cache->invalidation_requests);
2720 batcher_init(&cache->committer, commit_op, cache,
2721 issue_op, cache, cache->wq);
2722 iot_init(&cache->tracker);
2724 init_rwsem(&cache->background_work_lock);
2725 prevent_background_work(cache);
2734 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2739 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2742 for (i = 0; i < argc; i++) {
2743 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2752 cache->nr_ctr_args = argc;
2753 cache->ctr_args = copy;
2758 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2761 struct cache_args *ca;
2762 struct cache *cache = NULL;
2764 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2766 ti->error = "Error allocating memory for cache";
2771 r = parse_cache_args(ca, argc, argv, &ti->error);
2775 r = cache_create(ca, &cache);
2779 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2785 ti->private = cache;
2787 destroy_cache_args(ca);
2791 /*----------------------------------------------------------------*/
2793 static int cache_map(struct dm_target *ti, struct bio *bio)
2795 struct cache *cache = ti->private;
2799 dm_oblock_t block = get_bio_block(cache, bio);
2800 size_t pb_data_size = get_per_bio_data_size(cache);
2802 init_per_bio_data(bio, pb_data_size);
2803 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2805 * This can only occur if the io goes to a partial block at
2806 * the end of the origin device. We don't cache these.
2807 * Just remap to the origin and carry on.
2809 remap_to_origin(cache, bio);
2810 accounted_begin(cache, bio);
2811 return DM_MAPIO_REMAPPED;
2814 if (discard_or_flush(bio)) {
2815 defer_bio(cache, bio);
2816 return DM_MAPIO_SUBMITTED;
2819 r = map_bio(cache, bio, block, &commit_needed);
2821 schedule_commit(&cache->committer);
2826 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2828 struct cache *cache = ti->private;
2829 unsigned long flags;
2830 size_t pb_data_size = get_per_bio_data_size(cache);
2831 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2834 policy_tick(cache->policy, false);
2836 spin_lock_irqsave(&cache->lock, flags);
2837 cache->need_tick_bio = true;
2838 spin_unlock_irqrestore(&cache->lock, flags);
2841 bio_drop_shared_lock(cache, bio);
2842 accounted_complete(cache, bio);
2847 static int write_dirty_bitset(struct cache *cache)
2851 if (get_cache_mode(cache) >= CM_READ_ONLY)
2854 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2856 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2861 static int write_discard_bitset(struct cache *cache)
2865 if (get_cache_mode(cache) >= CM_READ_ONLY)
2868 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2869 cache->discard_nr_blocks);
2871 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2872 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2876 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2877 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2878 is_discarded(cache, to_dblock(i)));
2880 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2888 static int write_hints(struct cache *cache)
2892 if (get_cache_mode(cache) >= CM_READ_ONLY)
2895 r = dm_cache_write_hints(cache->cmd, cache->policy);
2897 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2905 * returns true on success
2907 static bool sync_metadata(struct cache *cache)
2911 r1 = write_dirty_bitset(cache);
2913 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2915 r2 = write_discard_bitset(cache);
2917 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2921 r3 = write_hints(cache);
2923 DMERR("%s: could not write hints", cache_device_name(cache));
2926 * If writing the above metadata failed, we still commit, but don't
2927 * set the clean shutdown flag. This will effectively force every
2928 * dirty bit to be set on reload.
2930 r4 = commit(cache, !r1 && !r2 && !r3);
2932 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2934 return !r1 && !r2 && !r3 && !r4;
2937 static void cache_postsuspend(struct dm_target *ti)
2939 struct cache *cache = ti->private;
2941 prevent_background_work(cache);
2942 BUG_ON(atomic_read(&cache->nr_io_migrations));
2944 cancel_delayed_work(&cache->waker);
2945 flush_workqueue(cache->wq);
2946 WARN_ON(cache->tracker.in_flight);
2949 * If it's a flush suspend there won't be any deferred bios, so this
2952 requeue_deferred_bios(cache);
2954 if (get_cache_mode(cache) == CM_WRITE)
2955 (void) sync_metadata(cache);
2958 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2959 bool dirty, uint32_t hint, bool hint_valid)
2962 struct cache *cache = context;
2965 set_bit(from_cblock(cblock), cache->dirty_bitset);
2966 atomic_inc(&cache->nr_dirty);
2968 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2970 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2978 * The discard block size in the on disk metadata is not
2979 * neccessarily the same as we're currently using. So we have to
2980 * be careful to only set the discarded attribute if we know it
2981 * covers a complete block of the new size.
2983 struct discard_load_info {
2984 struct cache *cache;
2987 * These blocks are sized using the on disk dblock size, rather
2988 * than the current one.
2990 dm_block_t block_size;
2991 dm_block_t discard_begin, discard_end;
2994 static void discard_load_info_init(struct cache *cache,
2995 struct discard_load_info *li)
2998 li->discard_begin = li->discard_end = 0;
3001 static void set_discard_range(struct discard_load_info *li)
3005 if (li->discard_begin == li->discard_end)
3009 * Convert to sectors.
3011 b = li->discard_begin * li->block_size;
3012 e = li->discard_end * li->block_size;
3015 * Then convert back to the current dblock size.
3017 b = dm_sector_div_up(b, li->cache->discard_block_size);
3018 sector_div(e, li->cache->discard_block_size);
3021 * The origin may have shrunk, so we need to check we're still in
3024 if (e > from_dblock(li->cache->discard_nr_blocks))
3025 e = from_dblock(li->cache->discard_nr_blocks);
3028 set_discard(li->cache, to_dblock(b));
3031 static int load_discard(void *context, sector_t discard_block_size,
3032 dm_dblock_t dblock, bool discard)
3034 struct discard_load_info *li = context;
3036 li->block_size = discard_block_size;
3039 if (from_dblock(dblock) == li->discard_end)
3041 * We're already in a discard range, just extend it.
3043 li->discard_end = li->discard_end + 1ULL;
3047 * Emit the old range and start a new one.
3049 set_discard_range(li);
3050 li->discard_begin = from_dblock(dblock);
3051 li->discard_end = li->discard_begin + 1ULL;
3054 set_discard_range(li);
3055 li->discard_begin = li->discard_end = 0;
3061 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3063 sector_t size = get_dev_size(cache->cache_dev);
3064 (void) sector_div(size, cache->sectors_per_block);
3065 return to_cblock(size);
3068 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3070 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3074 * We can't drop a dirty block when shrinking the cache.
3076 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3077 new_size = to_cblock(from_cblock(new_size) + 1);
3078 if (is_dirty(cache, new_size)) {
3079 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3080 cache_device_name(cache),
3081 (unsigned long long) from_cblock(new_size));
3089 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3093 r = dm_cache_resize(cache->cmd, new_size);
3095 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3096 metadata_operation_failed(cache, "dm_cache_resize", r);
3100 set_cache_size(cache, new_size);
3105 static int cache_preresume(struct dm_target *ti)
3108 struct cache *cache = ti->private;
3109 dm_cblock_t csize = get_cache_dev_size(cache);
3112 * Check to see if the cache has resized.
3114 if (!cache->sized) {
3115 r = resize_cache_dev(cache, csize);
3119 cache->sized = true;
3121 } else if (csize != cache->cache_size) {
3122 if (!can_resize(cache, csize))
3125 r = resize_cache_dev(cache, csize);
3130 if (!cache->loaded_mappings) {
3131 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3132 load_mapping, cache);
3134 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3135 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3139 cache->loaded_mappings = true;
3142 if (!cache->loaded_discards) {
3143 struct discard_load_info li;
3146 * The discard bitset could have been resized, or the
3147 * discard block size changed. To be safe we start by
3148 * setting every dblock to not discarded.
3150 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3152 discard_load_info_init(cache, &li);
3153 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3155 DMERR("%s: could not load origin discards", cache_device_name(cache));
3156 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3159 set_discard_range(&li);
3161 cache->loaded_discards = true;
3167 static void cache_resume(struct dm_target *ti)
3169 struct cache *cache = ti->private;
3171 cache->need_tick_bio = true;
3172 allow_background_work(cache);
3173 do_waker(&cache->waker.work);
3179 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3180 * <cache block size> <#used cache blocks>/<#total cache blocks>
3181 * <#read hits> <#read misses> <#write hits> <#write misses>
3182 * <#demotions> <#promotions> <#dirty>
3183 * <#features> <features>*
3184 * <#core args> <core args>
3185 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3187 static void cache_status(struct dm_target *ti, status_type_t type,
3188 unsigned status_flags, char *result, unsigned maxlen)
3193 dm_block_t nr_free_blocks_metadata = 0;
3194 dm_block_t nr_blocks_metadata = 0;
3195 char buf[BDEVNAME_SIZE];
3196 struct cache *cache = ti->private;
3197 dm_cblock_t residency;
3201 case STATUSTYPE_INFO:
3202 if (get_cache_mode(cache) == CM_FAIL) {
3207 /* Commit to ensure statistics aren't out-of-date */
3208 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3209 (void) commit(cache, false);
3211 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3213 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3214 cache_device_name(cache), r);
3218 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3220 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3221 cache_device_name(cache), r);
3225 residency = policy_residency(cache->policy);
3227 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3228 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3229 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3230 (unsigned long long)nr_blocks_metadata,
3231 (unsigned long long)cache->sectors_per_block,
3232 (unsigned long long) from_cblock(residency),
3233 (unsigned long long) from_cblock(cache->cache_size),
3234 (unsigned) atomic_read(&cache->stats.read_hit),
3235 (unsigned) atomic_read(&cache->stats.read_miss),
3236 (unsigned) atomic_read(&cache->stats.write_hit),
3237 (unsigned) atomic_read(&cache->stats.write_miss),
3238 (unsigned) atomic_read(&cache->stats.demotion),
3239 (unsigned) atomic_read(&cache->stats.promotion),
3240 (unsigned long) atomic_read(&cache->nr_dirty));
3242 if (cache->features.metadata_version == 2)
3243 DMEMIT("2 metadata2 ");
3247 if (writethrough_mode(&cache->features))
3248 DMEMIT("writethrough ");
3250 else if (passthrough_mode(&cache->features))
3251 DMEMIT("passthrough ");
3253 else if (writeback_mode(&cache->features))
3254 DMEMIT("writeback ");
3257 DMERR("%s: internal error: unknown io mode: %d",
3258 cache_device_name(cache), (int) cache->features.io_mode);
3262 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3264 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3266 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3268 DMERR("%s: policy_emit_config_values returned %d",
3269 cache_device_name(cache), r);
3272 if (get_cache_mode(cache) == CM_READ_ONLY)
3277 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3279 if (r || needs_check)
3280 DMEMIT("needs_check ");
3286 case STATUSTYPE_TABLE:
3287 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3289 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3291 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3294 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3295 DMEMIT(" %s", cache->ctr_args[i]);
3296 if (cache->nr_ctr_args)
3297 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3307 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3308 * the one-past-the-end value.
3310 struct cblock_range {
3316 * A cache block range can take two forms:
3318 * i) A single cblock, eg. '3456'
3319 * ii) A begin and end cblock with a dash between, eg. 123-234
3321 static int parse_cblock_range(struct cache *cache, const char *str,
3322 struct cblock_range *result)
3329 * Try and parse form (ii) first.
3331 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3336 result->begin = to_cblock(b);
3337 result->end = to_cblock(e);
3342 * That didn't work, try form (i).
3344 r = sscanf(str, "%llu%c", &b, &dummy);
3349 result->begin = to_cblock(b);
3350 result->end = to_cblock(from_cblock(result->begin) + 1u);
3354 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3358 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3360 uint64_t b = from_cblock(range->begin);
3361 uint64_t e = from_cblock(range->end);
3362 uint64_t n = from_cblock(cache->cache_size);
3365 DMERR("%s: begin cblock out of range: %llu >= %llu",
3366 cache_device_name(cache), b, n);
3371 DMERR("%s: end cblock out of range: %llu > %llu",
3372 cache_device_name(cache), e, n);
3377 DMERR("%s: invalid cblock range: %llu >= %llu",
3378 cache_device_name(cache), b, e);
3385 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3387 return to_cblock(from_cblock(b) + 1);
3390 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3395 * We don't need to do any locking here because we know we're in
3396 * passthrough mode. There's is potential for a race between an
3397 * invalidation triggered by an io and an invalidation message. This
3398 * is harmless, we must not worry if the policy call fails.
3400 while (range->begin != range->end) {
3401 r = invalidate_cblock(cache, range->begin);
3405 range->begin = cblock_succ(range->begin);
3408 cache->commit_requested = true;
3412 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3413 const char **cblock_ranges)
3417 struct cblock_range range;
3419 if (!passthrough_mode(&cache->features)) {
3420 DMERR("%s: cache has to be in passthrough mode for invalidation",
3421 cache_device_name(cache));
3425 for (i = 0; i < count; i++) {
3426 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3430 r = validate_cblock_range(cache, &range);
3435 * Pass begin and end origin blocks to the worker and wake it.
3437 r = request_invalidation(cache, &range);
3449 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3451 * The key migration_threshold is supported by the cache target core.
3453 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3455 struct cache *cache = ti->private;
3460 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3461 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3462 cache_device_name(cache));
3466 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3467 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3472 return set_config_value(cache, argv[0], argv[1]);
3475 static int cache_iterate_devices(struct dm_target *ti,
3476 iterate_devices_callout_fn fn, void *data)
3479 struct cache *cache = ti->private;
3481 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3483 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3488 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3491 * FIXME: these limits may be incompatible with the cache device
3493 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3494 cache->origin_sectors);
3495 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
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);
3512 set_discard_limits(cache, limits);
3515 /*----------------------------------------------------------------*/
3517 static struct target_type cache_target = {
3519 .version = {2, 0, 0},
3520 .module = THIS_MODULE,
3524 .end_io = cache_end_io,
3525 .postsuspend = cache_postsuspend,
3526 .preresume = cache_preresume,
3527 .resume = cache_resume,
3528 .status = cache_status,
3529 .message = cache_message,
3530 .iterate_devices = cache_iterate_devices,
3531 .io_hints = cache_io_hints,
3534 static int __init dm_cache_init(void)
3538 r = dm_register_target(&cache_target);
3540 DMERR("cache target registration failed: %d", r);
3544 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3545 if (!migration_cache) {
3546 dm_unregister_target(&cache_target);
3553 static void __exit dm_cache_exit(void)
3555 dm_unregister_target(&cache_target);
3556 kmem_cache_destroy(migration_cache);
3559 module_init(dm_cache_init);
3560 module_exit(dm_cache_exit);
3562 MODULE_DESCRIPTION(DM_NAME " cache target");
3563 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3564 MODULE_LICENSE("GPL");