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)
97 iot->in_flight -= len;
99 iot->idle_time = jiffies;
102 static void iot_io_end(struct io_tracker *iot, sector_t len)
106 spin_lock_irqsave(&iot->lock, flags);
107 __iot_io_end(iot, len);
108 spin_unlock_irqrestore(&iot->lock, flags);
111 /*----------------------------------------------------------------*/
114 * Represents a chunk of future work. 'input' allows continuations to pass
115 * values between themselves, typically error values.
117 struct continuation {
118 struct work_struct ws;
122 static inline void init_continuation(struct continuation *k,
123 void (*fn)(struct work_struct *))
125 INIT_WORK(&k->ws, fn);
129 static inline void queue_continuation(struct workqueue_struct *wq,
130 struct continuation *k)
132 queue_work(wq, &k->ws);
135 /*----------------------------------------------------------------*/
138 * The batcher collects together pieces of work that need a particular
139 * operation to occur before they can proceed (typically a commit).
143 * The operation that everyone is waiting for.
145 int (*commit_op)(void *context);
146 void *commit_context;
149 * This is how bios should be issued once the commit op is complete
150 * (accounted_request).
152 void (*issue_op)(struct bio *bio, void *context);
156 * Queued work gets put on here after commit.
158 struct workqueue_struct *wq;
161 struct list_head work_items;
162 struct bio_list bios;
163 struct work_struct commit_work;
165 bool commit_scheduled;
168 static void __commit(struct work_struct *_ws)
170 struct batcher *b = container_of(_ws, struct batcher, commit_work);
174 struct list_head work_items;
175 struct work_struct *ws, *tmp;
176 struct continuation *k;
178 struct bio_list bios;
180 INIT_LIST_HEAD(&work_items);
181 bio_list_init(&bios);
184 * We have to grab these before the commit_op to avoid a race
187 spin_lock_irqsave(&b->lock, flags);
188 list_splice_init(&b->work_items, &work_items);
189 bio_list_merge(&bios, &b->bios);
190 bio_list_init(&b->bios);
191 b->commit_scheduled = false;
192 spin_unlock_irqrestore(&b->lock, flags);
194 r = b->commit_op(b->commit_context);
196 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
197 k = container_of(ws, struct continuation, ws);
199 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
200 queue_work(b->wq, ws);
203 while ((bio = bio_list_pop(&bios))) {
208 b->issue_op(bio, b->issue_context);
212 static void batcher_init(struct batcher *b,
213 int (*commit_op)(void *),
214 void *commit_context,
215 void (*issue_op)(struct bio *bio, void *),
217 struct workqueue_struct *wq)
219 b->commit_op = commit_op;
220 b->commit_context = commit_context;
221 b->issue_op = issue_op;
222 b->issue_context = issue_context;
225 spin_lock_init(&b->lock);
226 INIT_LIST_HEAD(&b->work_items);
227 bio_list_init(&b->bios);
228 INIT_WORK(&b->commit_work, __commit);
229 b->commit_scheduled = false;
232 static void async_commit(struct batcher *b)
234 queue_work(b->wq, &b->commit_work);
237 static void continue_after_commit(struct batcher *b, struct continuation *k)
240 bool commit_scheduled;
242 spin_lock_irqsave(&b->lock, flags);
243 commit_scheduled = b->commit_scheduled;
244 list_add_tail(&k->ws.entry, &b->work_items);
245 spin_unlock_irqrestore(&b->lock, flags);
247 if (commit_scheduled)
252 * Bios are errored if commit failed.
254 static void issue_after_commit(struct batcher *b, struct bio *bio)
257 bool commit_scheduled;
259 spin_lock_irqsave(&b->lock, flags);
260 commit_scheduled = b->commit_scheduled;
261 bio_list_add(&b->bios, bio);
262 spin_unlock_irqrestore(&b->lock, flags);
264 if (commit_scheduled)
269 * Call this if some urgent work is waiting for the commit to complete.
271 static void schedule_commit(struct batcher *b)
276 spin_lock_irqsave(&b->lock, flags);
277 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
278 b->commit_scheduled = true;
279 spin_unlock_irqrestore(&b->lock, flags);
286 * There are a couple of places where we let a bio run, but want to do some
287 * work before calling its endio function. We do this by temporarily
288 * changing the endio fn.
290 struct dm_hook_info {
291 bio_end_io_t *bi_end_io;
294 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
295 bio_end_io_t *bi_end_io, void *bi_private)
297 h->bi_end_io = bio->bi_end_io;
299 bio->bi_end_io = bi_end_io;
300 bio->bi_private = bi_private;
303 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
305 bio->bi_end_io = h->bi_end_io;
308 /*----------------------------------------------------------------*/
310 #define MIGRATION_POOL_SIZE 128
311 #define COMMIT_PERIOD HZ
312 #define MIGRATION_COUNT_WINDOW 10
315 * The block size of the device holding cache data must be
316 * between 32KB and 1GB.
318 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
319 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
321 enum cache_metadata_mode {
322 CM_WRITE, /* metadata may be changed */
323 CM_READ_ONLY, /* metadata may not be changed */
329 * Data is written to cached blocks only. These blocks are marked
330 * dirty. If you lose the cache device you will lose data.
331 * Potential performance increase for both reads and writes.
336 * Data is written to both cache and origin. Blocks are never
337 * dirty. Potential performance benfit for reads only.
342 * A degraded mode useful for various cache coherency situations
343 * (eg, rolling back snapshots). Reads and writes always go to the
344 * origin. If a write goes to a cached oblock, then the cache
345 * block is invalidated.
350 struct cache_features {
351 enum cache_metadata_mode mode;
352 enum cache_io_mode io_mode;
353 unsigned metadata_version;
364 atomic_t copies_avoided;
365 atomic_t cache_cell_clash;
366 atomic_t commit_count;
367 atomic_t discard_count;
371 struct dm_target *ti;
372 struct dm_target_callbacks callbacks;
374 struct dm_cache_metadata *cmd;
377 * Metadata is written to this device.
379 struct dm_dev *metadata_dev;
382 * The slower of the two data devices. Typically a spindle.
384 struct dm_dev *origin_dev;
387 * The faster of the two data devices. Typically an SSD.
389 struct dm_dev *cache_dev;
392 * Size of the origin device in _complete_ blocks and native sectors.
394 dm_oblock_t origin_blocks;
395 sector_t origin_sectors;
398 * Size of the cache device in blocks.
400 dm_cblock_t cache_size;
403 * Fields for converting from sectors to blocks.
405 sector_t sectors_per_block;
406 int sectors_per_block_shift;
409 struct list_head deferred_cells;
410 struct bio_list deferred_bios;
411 struct bio_list deferred_writethrough_bios;
412 sector_t migration_threshold;
413 wait_queue_head_t migration_wait;
414 atomic_t nr_allocated_migrations;
417 * The number of in flight migrations that are performing
418 * background io. eg, promotion, writeback.
420 atomic_t nr_io_migrations;
422 struct rw_semaphore quiesce_lock;
425 * cache_size entries, dirty if set
428 unsigned long *dirty_bitset;
431 * origin_blocks entries, discarded if set.
433 dm_dblock_t discard_nr_blocks;
434 unsigned long *discard_bitset;
435 uint32_t discard_block_size; /* a power of 2 times sectors per block */
438 * Rather than reconstructing the table line for the status we just
439 * save it and regurgitate.
441 unsigned nr_ctr_args;
442 const char **ctr_args;
444 struct dm_kcopyd_client *copier;
445 struct workqueue_struct *wq;
446 struct work_struct deferred_bio_worker;
447 struct work_struct deferred_writethrough_worker;
448 struct work_struct migration_worker;
449 struct delayed_work waker;
450 struct dm_bio_prison_v2 *prison;
452 mempool_t *migration_pool;
454 struct dm_cache_policy *policy;
455 unsigned policy_nr_args;
457 bool need_tick_bio:1;
460 bool commit_requested:1;
461 bool loaded_mappings:1;
462 bool loaded_discards:1;
465 * Cache features such as write-through.
467 struct cache_features features;
469 struct cache_stats stats;
472 * Invalidation fields.
474 spinlock_t invalidation_lock;
475 struct list_head invalidation_requests;
477 struct io_tracker origin_tracker;
479 struct work_struct commit_ws;
480 struct batcher committer;
482 struct rw_semaphore background_work_lock;
485 struct per_bio_data {
488 struct dm_bio_prison_cell_v2 *cell;
489 struct dm_hook_info hook_info;
493 * writethrough fields. These MUST remain at the end of this
494 * structure and the 'cache' member must be the first as it
495 * is used to determine the offset of the writethrough fields.
499 struct dm_bio_details bio_details;
502 struct dm_cache_migration {
503 struct continuation k;
506 struct policy_work *op;
507 struct bio *overwrite_bio;
508 struct dm_bio_prison_cell_v2 *cell;
510 dm_cblock_t invalidate_cblock;
511 dm_oblock_t invalidate_oblock;
514 /*----------------------------------------------------------------*/
516 static bool writethrough_mode(struct cache_features *f)
518 return f->io_mode == CM_IO_WRITETHROUGH;
521 static bool writeback_mode(struct cache_features *f)
523 return f->io_mode == CM_IO_WRITEBACK;
526 static inline bool passthrough_mode(struct cache_features *f)
528 return unlikely(f->io_mode == CM_IO_PASSTHROUGH);
531 /*----------------------------------------------------------------*/
533 static void wake_deferred_bio_worker(struct cache *cache)
535 queue_work(cache->wq, &cache->deferred_bio_worker);
538 static void wake_deferred_writethrough_worker(struct cache *cache)
540 queue_work(cache->wq, &cache->deferred_writethrough_worker);
543 static void wake_migration_worker(struct cache *cache)
545 if (passthrough_mode(&cache->features))
548 queue_work(cache->wq, &cache->migration_worker);
551 /*----------------------------------------------------------------*/
553 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
555 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
558 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
560 dm_bio_prison_free_cell_v2(cache->prison, cell);
563 static struct dm_cache_migration *alloc_migration(struct cache *cache)
565 struct dm_cache_migration *mg;
567 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
570 atomic_inc(&mg->cache->nr_allocated_migrations);
576 static void free_migration(struct dm_cache_migration *mg)
578 struct cache *cache = mg->cache;
580 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
581 wake_up(&cache->migration_wait);
583 mempool_free(mg, cache->migration_pool);
586 /*----------------------------------------------------------------*/
588 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
590 return to_oblock(from_oblock(b) + 1ull);
593 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
597 key->block_begin = from_oblock(begin);
598 key->block_end = from_oblock(end);
602 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
603 * level 1 which prevents *both* READs and WRITEs.
605 #define WRITE_LOCK_LEVEL 0
606 #define READ_WRITE_LOCK_LEVEL 1
608 static unsigned lock_level(struct bio *bio)
610 return bio_data_dir(bio) == WRITE ?
612 READ_WRITE_LOCK_LEVEL;
615 /*----------------------------------------------------------------
617 *--------------------------------------------------------------*/
620 * If using writeback, leave out struct per_bio_data's writethrough fields.
622 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
623 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
625 static size_t get_per_bio_data_size(struct cache *cache)
627 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
630 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
632 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
637 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
639 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
642 pb->req_nr = dm_bio_get_target_bio_nr(bio);
649 /*----------------------------------------------------------------*/
651 static void defer_bio(struct cache *cache, struct bio *bio)
655 spin_lock_irqsave(&cache->lock, flags);
656 bio_list_add(&cache->deferred_bios, bio);
657 spin_unlock_irqrestore(&cache->lock, flags);
659 wake_deferred_bio_worker(cache);
662 static void defer_bios(struct cache *cache, struct bio_list *bios)
666 spin_lock_irqsave(&cache->lock, flags);
667 bio_list_merge(&cache->deferred_bios, bios);
669 spin_unlock_irqrestore(&cache->lock, flags);
671 wake_deferred_bio_worker(cache);
674 /*----------------------------------------------------------------*/
676 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
680 struct per_bio_data *pb;
681 struct dm_cell_key_v2 key;
682 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
683 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
685 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
686 if (!cell_prealloc) {
687 defer_bio(cache, bio);
691 build_key(oblock, end, &key);
692 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
695 * Failed to get the lock.
697 free_prison_cell(cache, cell_prealloc);
701 if (cell != cell_prealloc)
702 free_prison_cell(cache, cell_prealloc);
704 pb_size = get_per_bio_data_size(cache);
705 pb = get_per_bio_data(bio, pb_size);
711 /*----------------------------------------------------------------*/
713 static bool is_dirty(struct cache *cache, dm_cblock_t b)
715 return test_bit(from_cblock(b), cache->dirty_bitset);
718 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
720 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
721 atomic_inc(&cache->nr_dirty);
722 policy_set_dirty(cache->policy, cblock);
727 * These two are called when setting after migrations to force the policy
728 * and dirty bitset to be in sync.
730 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
732 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
733 atomic_inc(&cache->nr_dirty);
734 policy_set_dirty(cache->policy, cblock);
737 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
739 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
740 if (atomic_dec_return(&cache->nr_dirty) == 0)
741 dm_table_event(cache->ti->table);
744 policy_clear_dirty(cache->policy, cblock);
747 /*----------------------------------------------------------------*/
749 static bool block_size_is_power_of_two(struct cache *cache)
751 return cache->sectors_per_block_shift >= 0;
754 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
755 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
758 static dm_block_t block_div(dm_block_t b, uint32_t n)
765 static dm_block_t oblocks_per_dblock(struct cache *cache)
767 dm_block_t oblocks = cache->discard_block_size;
769 if (block_size_is_power_of_two(cache))
770 oblocks >>= cache->sectors_per_block_shift;
772 oblocks = block_div(oblocks, cache->sectors_per_block);
777 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
779 return to_dblock(block_div(from_oblock(oblock),
780 oblocks_per_dblock(cache)));
783 static void set_discard(struct cache *cache, dm_dblock_t b)
787 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
788 atomic_inc(&cache->stats.discard_count);
790 spin_lock_irqsave(&cache->lock, flags);
791 set_bit(from_dblock(b), cache->discard_bitset);
792 spin_unlock_irqrestore(&cache->lock, flags);
795 static void clear_discard(struct cache *cache, dm_dblock_t b)
799 spin_lock_irqsave(&cache->lock, flags);
800 clear_bit(from_dblock(b), cache->discard_bitset);
801 spin_unlock_irqrestore(&cache->lock, flags);
804 static bool is_discarded(struct cache *cache, dm_dblock_t b)
809 spin_lock_irqsave(&cache->lock, flags);
810 r = test_bit(from_dblock(b), cache->discard_bitset);
811 spin_unlock_irqrestore(&cache->lock, flags);
816 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
821 spin_lock_irqsave(&cache->lock, flags);
822 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
823 cache->discard_bitset);
824 spin_unlock_irqrestore(&cache->lock, flags);
829 /*----------------------------------------------------------------
831 *--------------------------------------------------------------*/
832 static void remap_to_origin(struct cache *cache, struct bio *bio)
834 bio->bi_bdev = cache->origin_dev->bdev;
837 static void remap_to_cache(struct cache *cache, struct bio *bio,
840 sector_t bi_sector = bio->bi_iter.bi_sector;
841 sector_t block = from_cblock(cblock);
843 bio->bi_bdev = cache->cache_dev->bdev;
844 if (!block_size_is_power_of_two(cache))
845 bio->bi_iter.bi_sector =
846 (block * cache->sectors_per_block) +
847 sector_div(bi_sector, cache->sectors_per_block);
849 bio->bi_iter.bi_sector =
850 (block << cache->sectors_per_block_shift) |
851 (bi_sector & (cache->sectors_per_block - 1));
854 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
857 size_t pb_data_size = get_per_bio_data_size(cache);
858 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
860 spin_lock_irqsave(&cache->lock, flags);
861 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
862 bio_op(bio) != REQ_OP_DISCARD) {
864 cache->need_tick_bio = false;
866 spin_unlock_irqrestore(&cache->lock, flags);
869 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
872 // FIXME: this is called way too much.
873 check_if_tick_bio_needed(cache, bio);
874 remap_to_origin(cache, bio);
875 if (bio_data_dir(bio) == WRITE)
876 clear_discard(cache, oblock_to_dblock(cache, oblock));
879 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
880 dm_oblock_t oblock, dm_cblock_t cblock)
882 check_if_tick_bio_needed(cache, bio);
883 remap_to_cache(cache, bio, cblock);
884 if (bio_data_dir(bio) == WRITE) {
885 set_dirty(cache, cblock);
886 clear_discard(cache, oblock_to_dblock(cache, oblock));
890 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
892 sector_t block_nr = bio->bi_iter.bi_sector;
894 if (!block_size_is_power_of_two(cache))
895 (void) sector_div(block_nr, cache->sectors_per_block);
897 block_nr >>= cache->sectors_per_block_shift;
899 return to_oblock(block_nr);
902 static bool accountable_bio(struct cache *cache, struct bio *bio)
904 return ((bio->bi_bdev == cache->origin_dev->bdev) &&
905 bio_op(bio) != REQ_OP_DISCARD);
908 static void accounted_begin(struct cache *cache, struct bio *bio)
910 size_t pb_data_size = get_per_bio_data_size(cache);
911 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
913 if (accountable_bio(cache, bio)) {
914 pb->len = bio_sectors(bio);
915 iot_io_begin(&cache->origin_tracker, pb->len);
919 static void accounted_complete(struct cache *cache, struct bio *bio)
921 size_t pb_data_size = get_per_bio_data_size(cache);
922 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
924 iot_io_end(&cache->origin_tracker, pb->len);
927 static void accounted_request(struct cache *cache, struct bio *bio)
929 accounted_begin(cache, bio);
930 generic_make_request(bio);
933 static void issue_op(struct bio *bio, void *context)
935 struct cache *cache = context;
936 accounted_request(cache, bio);
939 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
943 spin_lock_irqsave(&cache->lock, flags);
944 bio_list_add(&cache->deferred_writethrough_bios, bio);
945 spin_unlock_irqrestore(&cache->lock, flags);
947 wake_deferred_writethrough_worker(cache);
950 static void writethrough_endio(struct bio *bio)
952 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
954 dm_unhook_bio(&pb->hook_info, bio);
961 dm_bio_restore(&pb->bio_details, bio);
962 remap_to_cache(pb->cache, bio, pb->cblock);
965 * We can't issue this bio directly, since we're in interrupt
966 * context. So it gets put on a bio list for processing by the
969 defer_writethrough_bio(pb->cache, bio);
973 * FIXME: send in parallel, huge latency as is.
974 * When running in writethrough mode we need to send writes to clean blocks
975 * to both the cache and origin devices. In future we'd like to clone the
976 * bio and send them in parallel, but for now we're doing them in
977 * series as this is easier.
979 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
980 dm_oblock_t oblock, dm_cblock_t cblock)
982 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
986 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
987 dm_bio_record(&pb->bio_details, bio);
989 remap_to_origin_clear_discard(pb->cache, bio, oblock);
992 /*----------------------------------------------------------------
994 *--------------------------------------------------------------*/
995 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
997 return cache->features.mode;
1000 static const char *cache_device_name(struct cache *cache)
1002 return dm_device_name(dm_table_get_md(cache->ti->table));
1005 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
1007 const char *descs[] = {
1013 dm_table_event(cache->ti->table);
1014 DMINFO("%s: switching cache to %s mode",
1015 cache_device_name(cache), descs[(int)mode]);
1018 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
1021 enum cache_metadata_mode old_mode = get_cache_mode(cache);
1023 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
1024 DMERR("%s: unable to read needs_check flag, setting failure mode.",
1025 cache_device_name(cache));
1029 if (new_mode == CM_WRITE && needs_check) {
1030 DMERR("%s: unable to switch cache to write mode until repaired.",
1031 cache_device_name(cache));
1032 if (old_mode != new_mode)
1033 new_mode = old_mode;
1035 new_mode = CM_READ_ONLY;
1038 /* Never move out of fail mode */
1039 if (old_mode == CM_FAIL)
1045 dm_cache_metadata_set_read_only(cache->cmd);
1049 dm_cache_metadata_set_read_write(cache->cmd);
1053 cache->features.mode = new_mode;
1055 if (new_mode != old_mode)
1056 notify_mode_switch(cache, new_mode);
1059 static void abort_transaction(struct cache *cache)
1061 const char *dev_name = cache_device_name(cache);
1063 if (get_cache_mode(cache) >= CM_READ_ONLY)
1066 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1067 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1068 set_cache_mode(cache, CM_FAIL);
1071 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1072 if (dm_cache_metadata_abort(cache->cmd)) {
1073 DMERR("%s: failed to abort metadata transaction", dev_name);
1074 set_cache_mode(cache, CM_FAIL);
1078 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1080 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1081 cache_device_name(cache), op, r);
1082 abort_transaction(cache);
1083 set_cache_mode(cache, CM_READ_ONLY);
1086 /*----------------------------------------------------------------*/
1088 static void load_stats(struct cache *cache)
1090 struct dm_cache_statistics stats;
1092 dm_cache_metadata_get_stats(cache->cmd, &stats);
1093 atomic_set(&cache->stats.read_hit, stats.read_hits);
1094 atomic_set(&cache->stats.read_miss, stats.read_misses);
1095 atomic_set(&cache->stats.write_hit, stats.write_hits);
1096 atomic_set(&cache->stats.write_miss, stats.write_misses);
1099 static void save_stats(struct cache *cache)
1101 struct dm_cache_statistics stats;
1103 if (get_cache_mode(cache) >= CM_READ_ONLY)
1106 stats.read_hits = atomic_read(&cache->stats.read_hit);
1107 stats.read_misses = atomic_read(&cache->stats.read_miss);
1108 stats.write_hits = atomic_read(&cache->stats.write_hit);
1109 stats.write_misses = atomic_read(&cache->stats.write_miss);
1111 dm_cache_metadata_set_stats(cache->cmd, &stats);
1114 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1117 case POLICY_PROMOTE:
1118 atomic_inc(&stats->promotion);
1122 atomic_inc(&stats->demotion);
1125 case POLICY_WRITEBACK:
1126 atomic_inc(&stats->writeback);
1131 /*----------------------------------------------------------------
1132 * Migration processing
1134 * Migration covers moving data from the origin device to the cache, or
1136 *--------------------------------------------------------------*/
1138 static void inc_io_migrations(struct cache *cache)
1140 atomic_inc(&cache->nr_io_migrations);
1143 static void dec_io_migrations(struct cache *cache)
1145 atomic_dec(&cache->nr_io_migrations);
1148 static bool discard_or_flush(struct bio *bio)
1150 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1153 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1154 dm_dblock_t *b, dm_dblock_t *e)
1156 sector_t sb = bio->bi_iter.bi_sector;
1157 sector_t se = bio_end_sector(bio);
1159 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1161 if (se - sb < cache->discard_block_size)
1164 *e = to_dblock(block_div(se, cache->discard_block_size));
1167 /*----------------------------------------------------------------*/
1169 static void prevent_background_work(struct cache *cache)
1172 down_write(&cache->background_work_lock);
1176 static void allow_background_work(struct cache *cache)
1179 up_write(&cache->background_work_lock);
1183 static bool background_work_begin(struct cache *cache)
1188 r = down_read_trylock(&cache->background_work_lock);
1194 static void background_work_end(struct cache *cache)
1197 up_read(&cache->background_work_lock);
1201 /*----------------------------------------------------------------*/
1203 static void quiesce(struct dm_cache_migration *mg,
1204 void (*continuation)(struct work_struct *))
1206 init_continuation(&mg->k, continuation);
1207 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1210 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1212 struct continuation *k = container_of(ws, struct continuation, ws);
1213 return container_of(k, struct dm_cache_migration, k);
1216 static void copy_complete(int read_err, unsigned long write_err, void *context)
1218 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1220 if (read_err || write_err)
1223 queue_continuation(mg->cache->wq, &mg->k);
1226 static int copy(struct dm_cache_migration *mg, bool promote)
1229 struct dm_io_region o_region, c_region;
1230 struct cache *cache = mg->cache;
1232 o_region.bdev = cache->origin_dev->bdev;
1233 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1234 o_region.count = cache->sectors_per_block;
1236 c_region.bdev = cache->cache_dev->bdev;
1237 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1238 c_region.count = cache->sectors_per_block;
1241 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1243 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1248 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1250 size_t pb_data_size = get_per_bio_data_size(cache);
1251 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1253 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1254 free_prison_cell(cache, pb->cell);
1258 static void overwrite_endio(struct bio *bio)
1260 struct dm_cache_migration *mg = bio->bi_private;
1261 struct cache *cache = mg->cache;
1262 size_t pb_data_size = get_per_bio_data_size(cache);
1263 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1265 dm_unhook_bio(&pb->hook_info, bio);
1268 mg->k.input = bio->bi_error;
1270 queue_continuation(mg->cache->wq, &mg->k);
1273 static void overwrite(struct dm_cache_migration *mg,
1274 void (*continuation)(struct work_struct *))
1276 struct bio *bio = mg->overwrite_bio;
1277 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1278 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1280 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1283 * The overwrite bio is part of the copy operation, as such it does
1284 * not set/clear discard or dirty flags.
1286 if (mg->op->op == POLICY_PROMOTE)
1287 remap_to_cache(mg->cache, bio, mg->op->cblock);
1289 remap_to_origin(mg->cache, bio);
1291 init_continuation(&mg->k, continuation);
1292 accounted_request(mg->cache, bio);
1298 * 1) exclusive lock preventing WRITEs
1300 * 3) copy or issue overwrite bio
1301 * 4) upgrade to exclusive lock preventing READs and WRITEs
1303 * 6) update metadata and commit
1306 static void mg_complete(struct dm_cache_migration *mg, bool success)
1308 struct bio_list bios;
1309 struct cache *cache = mg->cache;
1310 struct policy_work *op = mg->op;
1311 dm_cblock_t cblock = op->cblock;
1314 update_stats(&cache->stats, op->op);
1317 case POLICY_PROMOTE:
1318 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1319 policy_complete_background_work(cache->policy, op, success);
1321 if (mg->overwrite_bio) {
1323 force_set_dirty(cache, cblock);
1325 mg->overwrite_bio->bi_error = (mg->k.input ? : -EIO);
1326 bio_endio(mg->overwrite_bio);
1329 force_clear_dirty(cache, cblock);
1330 dec_io_migrations(cache);
1336 * We clear dirty here to update the nr_dirty counter.
1339 force_clear_dirty(cache, cblock);
1340 policy_complete_background_work(cache->policy, op, success);
1341 dec_io_migrations(cache);
1344 case POLICY_WRITEBACK:
1346 force_clear_dirty(cache, cblock);
1347 policy_complete_background_work(cache->policy, op, success);
1348 dec_io_migrations(cache);
1352 bio_list_init(&bios);
1354 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1355 free_prison_cell(cache, mg->cell);
1359 defer_bios(cache, &bios);
1360 wake_migration_worker(cache);
1362 background_work_end(cache);
1365 static void mg_success(struct work_struct *ws)
1367 struct dm_cache_migration *mg = ws_to_mg(ws);
1368 mg_complete(mg, mg->k.input == 0);
1371 static void mg_update_metadata(struct work_struct *ws)
1374 struct dm_cache_migration *mg = ws_to_mg(ws);
1375 struct cache *cache = mg->cache;
1376 struct policy_work *op = mg->op;
1379 case POLICY_PROMOTE:
1380 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1382 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1383 cache_device_name(cache));
1384 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1386 mg_complete(mg, false);
1389 mg_complete(mg, true);
1393 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1395 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1396 cache_device_name(cache));
1397 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1399 mg_complete(mg, false);
1404 * It would be nice if we only had to commit when a REQ_FLUSH
1405 * comes through. But there's one scenario that we have to
1408 * - vblock x in a cache block
1410 * - cache block gets reallocated and over written
1413 * When we recover, because there was no commit the cache will
1414 * rollback to having the data for vblock x in the cache block.
1415 * But the cache block has since been overwritten, so it'll end
1416 * up pointing to data that was never in 'x' during the history
1419 * To avoid this issue we require a commit as part of the
1420 * demotion operation.
1422 init_continuation(&mg->k, mg_success);
1423 continue_after_commit(&cache->committer, &mg->k);
1424 schedule_commit(&cache->committer);
1427 case POLICY_WRITEBACK:
1428 mg_complete(mg, true);
1433 static void mg_update_metadata_after_copy(struct work_struct *ws)
1435 struct dm_cache_migration *mg = ws_to_mg(ws);
1438 * Did the copy succeed?
1441 mg_complete(mg, false);
1443 mg_update_metadata(ws);
1446 static void mg_upgrade_lock(struct work_struct *ws)
1449 struct dm_cache_migration *mg = ws_to_mg(ws);
1452 * Did the copy succeed?
1455 mg_complete(mg, false);
1459 * Now we want the lock to prevent both reads and writes.
1461 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1462 READ_WRITE_LOCK_LEVEL);
1464 mg_complete(mg, false);
1467 quiesce(mg, mg_update_metadata);
1470 mg_update_metadata(ws);
1474 static void mg_copy(struct work_struct *ws)
1477 struct dm_cache_migration *mg = ws_to_mg(ws);
1479 if (mg->overwrite_bio) {
1481 * It's safe to do this here, even though it's new data
1482 * because all IO has been locked out of the block.
1484 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1485 * so _not_ using mg_upgrade_lock() as continutation.
1487 overwrite(mg, mg_update_metadata_after_copy);
1490 struct cache *cache = mg->cache;
1491 struct policy_work *op = mg->op;
1492 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1494 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1495 is_discarded_oblock(cache, op->oblock)) {
1496 mg_upgrade_lock(ws);
1500 init_continuation(&mg->k, mg_upgrade_lock);
1502 r = copy(mg, is_policy_promote);
1504 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1506 mg_complete(mg, false);
1511 static int mg_lock_writes(struct dm_cache_migration *mg)
1514 struct dm_cell_key_v2 key;
1515 struct cache *cache = mg->cache;
1516 struct dm_bio_prison_cell_v2 *prealloc;
1518 prealloc = alloc_prison_cell(cache);
1520 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1521 mg_complete(mg, false);
1526 * Prevent writes to the block, but allow reads to continue.
1527 * Unless we're using an overwrite bio, in which case we lock
1530 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1531 r = dm_cell_lock_v2(cache->prison, &key,
1532 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1533 prealloc, &mg->cell);
1535 free_prison_cell(cache, prealloc);
1536 mg_complete(mg, false);
1540 if (mg->cell != prealloc)
1541 free_prison_cell(cache, prealloc);
1546 quiesce(mg, mg_copy);
1551 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1553 struct dm_cache_migration *mg;
1555 if (!background_work_begin(cache)) {
1556 policy_complete_background_work(cache->policy, op, false);
1560 mg = alloc_migration(cache);
1562 policy_complete_background_work(cache->policy, op, false);
1563 background_work_end(cache);
1567 memset(mg, 0, sizeof(*mg));
1571 mg->overwrite_bio = bio;
1574 inc_io_migrations(cache);
1576 return mg_lock_writes(mg);
1579 /*----------------------------------------------------------------
1580 * invalidation processing
1581 *--------------------------------------------------------------*/
1583 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1585 struct bio_list bios;
1586 struct cache *cache = mg->cache;
1588 bio_list_init(&bios);
1589 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1590 free_prison_cell(cache, mg->cell);
1592 if (!success && mg->overwrite_bio)
1593 bio_io_error(mg->overwrite_bio);
1596 defer_bios(cache, &bios);
1598 background_work_end(cache);
1601 static void invalidate_completed(struct work_struct *ws)
1603 struct dm_cache_migration *mg = ws_to_mg(ws);
1604 invalidate_complete(mg, !mg->k.input);
1607 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1609 int r = policy_invalidate_mapping(cache->policy, cblock);
1611 r = dm_cache_remove_mapping(cache->cmd, cblock);
1613 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1614 cache_device_name(cache));
1615 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1618 } else if (r == -ENODATA) {
1620 * Harmless, already unmapped.
1625 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1630 static void invalidate_remove(struct work_struct *ws)
1633 struct dm_cache_migration *mg = ws_to_mg(ws);
1634 struct cache *cache = mg->cache;
1636 r = invalidate_cblock(cache, mg->invalidate_cblock);
1638 invalidate_complete(mg, false);
1642 init_continuation(&mg->k, invalidate_completed);
1643 continue_after_commit(&cache->committer, &mg->k);
1644 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1645 mg->overwrite_bio = NULL;
1646 schedule_commit(&cache->committer);
1649 static int invalidate_lock(struct dm_cache_migration *mg)
1652 struct dm_cell_key_v2 key;
1653 struct cache *cache = mg->cache;
1654 struct dm_bio_prison_cell_v2 *prealloc;
1656 prealloc = alloc_prison_cell(cache);
1658 invalidate_complete(mg, false);
1662 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1663 r = dm_cell_lock_v2(cache->prison, &key,
1664 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1666 free_prison_cell(cache, prealloc);
1667 invalidate_complete(mg, false);
1671 if (mg->cell != prealloc)
1672 free_prison_cell(cache, prealloc);
1675 quiesce(mg, invalidate_remove);
1679 * We can't call invalidate_remove() directly here because we
1680 * might still be in request context.
1682 init_continuation(&mg->k, invalidate_remove);
1683 queue_work(cache->wq, &mg->k.ws);
1689 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1690 dm_oblock_t oblock, struct bio *bio)
1692 struct dm_cache_migration *mg;
1694 if (!background_work_begin(cache))
1697 mg = alloc_migration(cache);
1699 background_work_end(cache);
1703 memset(mg, 0, sizeof(*mg));
1706 mg->overwrite_bio = bio;
1707 mg->invalidate_cblock = cblock;
1708 mg->invalidate_oblock = oblock;
1710 return invalidate_lock(mg);
1713 /*----------------------------------------------------------------
1715 *--------------------------------------------------------------*/
1723 static enum busy spare_migration_bandwidth(struct cache *cache)
1725 bool idle = iot_idle_for(&cache->origin_tracker, HZ);
1726 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1727 cache->sectors_per_block;
1729 if (current_volume <= cache->migration_threshold)
1730 return idle ? IDLE : MODERATE;
1732 return idle ? MODERATE : BUSY;
1735 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1737 atomic_inc(bio_data_dir(bio) == READ ?
1738 &cache->stats.read_hit : &cache->stats.write_hit);
1741 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1743 atomic_inc(bio_data_dir(bio) == READ ?
1744 &cache->stats.read_miss : &cache->stats.write_miss);
1747 /*----------------------------------------------------------------*/
1749 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1751 return (bio_data_dir(bio) == WRITE) &&
1752 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1755 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1757 return writeback_mode(&cache->features) &&
1758 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1761 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1762 bool *commit_needed)
1765 bool rb, background_queued;
1767 size_t pb_data_size = get_per_bio_data_size(cache);
1768 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1770 *commit_needed = false;
1772 rb = bio_detain_shared(cache, block, bio);
1775 * An exclusive lock is held for this block, so we have to
1776 * wait. We set the commit_needed flag so the current
1777 * transaction will be committed asap, allowing this lock
1780 *commit_needed = true;
1781 return DM_MAPIO_SUBMITTED;
1784 data_dir = bio_data_dir(bio);
1786 if (optimisable_bio(cache, bio, block)) {
1787 struct policy_work *op = NULL;
1789 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1790 if (unlikely(r && r != -ENOENT)) {
1791 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1792 cache_device_name(cache), r);
1794 return DM_MAPIO_SUBMITTED;
1797 if (r == -ENOENT && op) {
1798 bio_drop_shared_lock(cache, bio);
1799 BUG_ON(op->op != POLICY_PROMOTE);
1800 mg_start(cache, op, bio);
1801 return DM_MAPIO_SUBMITTED;
1804 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1805 if (unlikely(r && r != -ENOENT)) {
1806 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1807 cache_device_name(cache), r);
1809 return DM_MAPIO_SUBMITTED;
1812 if (background_queued)
1813 wake_migration_worker(cache);
1820 inc_miss_counter(cache, bio);
1821 if (pb->req_nr == 0) {
1822 accounted_begin(cache, bio);
1823 remap_to_origin_clear_discard(cache, bio, block);
1827 * This is a duplicate writethrough io that is no
1828 * longer needed because the block has been demoted.
1831 return DM_MAPIO_SUBMITTED;
1837 inc_hit_counter(cache, bio);
1840 * Passthrough always maps to the origin, invalidating any
1841 * cache blocks that are written to.
1843 if (passthrough_mode(&cache->features)) {
1844 if (bio_data_dir(bio) == WRITE) {
1845 bio_drop_shared_lock(cache, bio);
1846 atomic_inc(&cache->stats.demotion);
1847 invalidate_start(cache, cblock, block, bio);
1849 remap_to_origin_clear_discard(cache, bio, block);
1852 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
1853 !is_dirty(cache, cblock)) {
1854 remap_to_origin_then_cache(cache, bio, block, cblock);
1855 accounted_begin(cache, bio);
1857 remap_to_cache_dirty(cache, bio, block, cblock);
1862 * dm core turns FUA requests into a separate payload and FLUSH req.
1864 if (bio->bi_opf & REQ_FUA) {
1866 * issue_after_commit will call accounted_begin a second time. So
1867 * we call accounted_complete() to avoid double accounting.
1869 accounted_complete(cache, bio);
1870 issue_after_commit(&cache->committer, bio);
1871 *commit_needed = true;
1872 return DM_MAPIO_SUBMITTED;
1875 return DM_MAPIO_REMAPPED;
1878 static bool process_bio(struct cache *cache, struct bio *bio)
1882 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1883 generic_make_request(bio);
1885 return commit_needed;
1889 * A non-zero return indicates read_only or fail_io mode.
1891 static int commit(struct cache *cache, bool clean_shutdown)
1895 if (get_cache_mode(cache) >= CM_READ_ONLY)
1898 atomic_inc(&cache->stats.commit_count);
1899 r = dm_cache_commit(cache->cmd, clean_shutdown);
1901 metadata_operation_failed(cache, "dm_cache_commit", r);
1907 * Used by the batcher.
1909 static int commit_op(void *context)
1911 struct cache *cache = context;
1913 if (dm_cache_changed_this_transaction(cache->cmd))
1914 return commit(cache, false);
1919 /*----------------------------------------------------------------*/
1921 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1923 size_t pb_data_size = get_per_bio_data_size(cache);
1924 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1927 remap_to_origin(cache, bio);
1929 remap_to_cache(cache, bio, 0);
1931 issue_after_commit(&cache->committer, bio);
1935 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1939 // FIXME: do we need to lock the region? Or can we just assume the
1940 // user wont be so foolish as to issue discard concurrently with
1942 calc_discard_block_range(cache, bio, &b, &e);
1944 set_discard(cache, b);
1945 b = to_dblock(from_dblock(b) + 1);
1953 static void process_deferred_bios(struct work_struct *ws)
1955 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1957 unsigned long flags;
1958 bool commit_needed = false;
1959 struct bio_list bios;
1962 bio_list_init(&bios);
1964 spin_lock_irqsave(&cache->lock, flags);
1965 bio_list_merge(&bios, &cache->deferred_bios);
1966 bio_list_init(&cache->deferred_bios);
1967 spin_unlock_irqrestore(&cache->lock, flags);
1969 while ((bio = bio_list_pop(&bios))) {
1970 if (bio->bi_opf & REQ_PREFLUSH)
1971 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1973 else if (bio_op(bio) == REQ_OP_DISCARD)
1974 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1977 commit_needed = process_bio(cache, bio) || commit_needed;
1981 schedule_commit(&cache->committer);
1984 static void process_deferred_writethrough_bios(struct work_struct *ws)
1986 struct cache *cache = container_of(ws, struct cache, deferred_writethrough_worker);
1988 unsigned long flags;
1989 struct bio_list bios;
1992 bio_list_init(&bios);
1994 spin_lock_irqsave(&cache->lock, flags);
1995 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1996 bio_list_init(&cache->deferred_writethrough_bios);
1997 spin_unlock_irqrestore(&cache->lock, flags);
2000 * These bios have already been through accounted_begin()
2002 while ((bio = bio_list_pop(&bios)))
2003 generic_make_request(bio);
2006 /*----------------------------------------------------------------
2008 *--------------------------------------------------------------*/
2010 static void requeue_deferred_bios(struct cache *cache)
2013 struct bio_list bios;
2015 bio_list_init(&bios);
2016 bio_list_merge(&bios, &cache->deferred_bios);
2017 bio_list_init(&cache->deferred_bios);
2019 while ((bio = bio_list_pop(&bios))) {
2020 bio->bi_error = DM_ENDIO_REQUEUE;
2026 * We want to commit periodically so that not too much
2027 * unwritten metadata builds up.
2029 static void do_waker(struct work_struct *ws)
2031 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2033 policy_tick(cache->policy, true);
2034 wake_migration_worker(cache);
2035 schedule_commit(&cache->committer);
2036 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2039 static void check_migrations(struct work_struct *ws)
2042 struct policy_work *op;
2043 struct cache *cache = container_of(ws, struct cache, migration_worker);
2047 b = spare_migration_bandwidth(cache);
2051 r = policy_get_background_work(cache->policy, b == IDLE, &op);
2056 DMERR_LIMIT("%s: policy_background_work failed",
2057 cache_device_name(cache));
2061 r = mg_start(cache, op, NULL);
2067 /*----------------------------------------------------------------
2069 *--------------------------------------------------------------*/
2072 * This function gets called on the error paths of the constructor, so we
2073 * have to cope with a partially initialised struct.
2075 static void destroy(struct cache *cache)
2079 mempool_destroy(cache->migration_pool);
2082 dm_bio_prison_destroy_v2(cache->prison);
2085 destroy_workqueue(cache->wq);
2087 if (cache->dirty_bitset)
2088 free_bitset(cache->dirty_bitset);
2090 if (cache->discard_bitset)
2091 free_bitset(cache->discard_bitset);
2094 dm_kcopyd_client_destroy(cache->copier);
2097 dm_cache_metadata_close(cache->cmd);
2099 if (cache->metadata_dev)
2100 dm_put_device(cache->ti, cache->metadata_dev);
2102 if (cache->origin_dev)
2103 dm_put_device(cache->ti, cache->origin_dev);
2105 if (cache->cache_dev)
2106 dm_put_device(cache->ti, cache->cache_dev);
2109 dm_cache_policy_destroy(cache->policy);
2111 for (i = 0; i < cache->nr_ctr_args ; i++)
2112 kfree(cache->ctr_args[i]);
2113 kfree(cache->ctr_args);
2118 static void cache_dtr(struct dm_target *ti)
2120 struct cache *cache = ti->private;
2125 static sector_t get_dev_size(struct dm_dev *dev)
2127 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2130 /*----------------------------------------------------------------*/
2133 * Construct a cache device mapping.
2135 * cache <metadata dev> <cache dev> <origin dev> <block size>
2136 * <#feature args> [<feature arg>]*
2137 * <policy> <#policy args> [<policy arg>]*
2139 * metadata dev : fast device holding the persistent metadata
2140 * cache dev : fast device holding cached data blocks
2141 * origin dev : slow device holding original data blocks
2142 * block size : cache unit size in sectors
2144 * #feature args : number of feature arguments passed
2145 * feature args : writethrough. (The default is writeback.)
2147 * policy : the replacement policy to use
2148 * #policy args : an even number of policy arguments corresponding
2149 * to key/value pairs passed to the policy
2150 * policy args : key/value pairs passed to the policy
2151 * E.g. 'sequential_threshold 1024'
2152 * See cache-policies.txt for details.
2154 * Optional feature arguments are:
2155 * writethrough : write through caching that prohibits cache block
2156 * content from being different from origin block content.
2157 * Without this argument, the default behaviour is to write
2158 * back cache block contents later for performance reasons,
2159 * so they may differ from the corresponding origin blocks.
2162 struct dm_target *ti;
2164 struct dm_dev *metadata_dev;
2166 struct dm_dev *cache_dev;
2167 sector_t cache_sectors;
2169 struct dm_dev *origin_dev;
2170 sector_t origin_sectors;
2172 uint32_t block_size;
2174 const char *policy_name;
2176 const char **policy_argv;
2178 struct cache_features features;
2181 static void destroy_cache_args(struct cache_args *ca)
2183 if (ca->metadata_dev)
2184 dm_put_device(ca->ti, ca->metadata_dev);
2187 dm_put_device(ca->ti, ca->cache_dev);
2190 dm_put_device(ca->ti, ca->origin_dev);
2195 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2198 *error = "Insufficient args";
2205 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2209 sector_t metadata_dev_size;
2210 char b[BDEVNAME_SIZE];
2212 if (!at_least_one_arg(as, error))
2215 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2218 *error = "Error opening metadata device";
2222 metadata_dev_size = get_dev_size(ca->metadata_dev);
2223 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2224 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2225 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2230 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2235 if (!at_least_one_arg(as, error))
2238 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2241 *error = "Error opening cache device";
2244 ca->cache_sectors = get_dev_size(ca->cache_dev);
2249 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2254 if (!at_least_one_arg(as, error))
2257 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2260 *error = "Error opening origin device";
2264 ca->origin_sectors = get_dev_size(ca->origin_dev);
2265 if (ca->ti->len > ca->origin_sectors) {
2266 *error = "Device size larger than cached device";
2273 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2276 unsigned long block_size;
2278 if (!at_least_one_arg(as, error))
2281 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2282 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2283 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2284 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2285 *error = "Invalid data block size";
2289 if (block_size > ca->cache_sectors) {
2290 *error = "Data block size is larger than the cache device";
2294 ca->block_size = block_size;
2299 static void init_features(struct cache_features *cf)
2301 cf->mode = CM_WRITE;
2302 cf->io_mode = CM_IO_WRITEBACK;
2303 cf->metadata_version = 1;
2306 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2309 static struct dm_arg _args[] = {
2310 {0, 2, "Invalid number of cache feature arguments"},
2316 struct cache_features *cf = &ca->features;
2320 r = dm_read_arg_group(_args, as, &argc, error);
2325 arg = dm_shift_arg(as);
2327 if (!strcasecmp(arg, "writeback"))
2328 cf->io_mode = CM_IO_WRITEBACK;
2330 else if (!strcasecmp(arg, "writethrough"))
2331 cf->io_mode = CM_IO_WRITETHROUGH;
2333 else if (!strcasecmp(arg, "passthrough"))
2334 cf->io_mode = CM_IO_PASSTHROUGH;
2336 else if (!strcasecmp(arg, "metadata2"))
2337 cf->metadata_version = 2;
2340 *error = "Unrecognised cache feature requested";
2348 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2351 static struct dm_arg _args[] = {
2352 {0, 1024, "Invalid number of policy arguments"},
2357 if (!at_least_one_arg(as, error))
2360 ca->policy_name = dm_shift_arg(as);
2362 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2366 ca->policy_argv = (const char **)as->argv;
2367 dm_consume_args(as, ca->policy_argc);
2372 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2376 struct dm_arg_set as;
2381 r = parse_metadata_dev(ca, &as, error);
2385 r = parse_cache_dev(ca, &as, error);
2389 r = parse_origin_dev(ca, &as, error);
2393 r = parse_block_size(ca, &as, error);
2397 r = parse_features(ca, &as, error);
2401 r = parse_policy(ca, &as, error);
2408 /*----------------------------------------------------------------*/
2410 static struct kmem_cache *migration_cache;
2412 #define NOT_CORE_OPTION 1
2414 static int process_config_option(struct cache *cache, const char *key, const char *value)
2418 if (!strcasecmp(key, "migration_threshold")) {
2419 if (kstrtoul(value, 10, &tmp))
2422 cache->migration_threshold = tmp;
2426 return NOT_CORE_OPTION;
2429 static int set_config_value(struct cache *cache, const char *key, const char *value)
2431 int r = process_config_option(cache, key, value);
2433 if (r == NOT_CORE_OPTION)
2434 r = policy_set_config_value(cache->policy, key, value);
2437 DMWARN("bad config value for %s: %s", key, value);
2442 static int set_config_values(struct cache *cache, int argc, const char **argv)
2447 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2452 r = set_config_value(cache, argv[0], argv[1]);
2463 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2466 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2468 cache->origin_sectors,
2469 cache->sectors_per_block);
2471 *error = "Error creating cache's policy";
2475 BUG_ON(!cache->policy);
2481 * We want the discard block size to be at least the size of the cache
2482 * block size and have no more than 2^14 discard blocks across the origin.
2484 #define MAX_DISCARD_BLOCKS (1 << 14)
2486 static bool too_many_discard_blocks(sector_t discard_block_size,
2487 sector_t origin_size)
2489 (void) sector_div(origin_size, discard_block_size);
2491 return origin_size > MAX_DISCARD_BLOCKS;
2494 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2495 sector_t origin_size)
2497 sector_t discard_block_size = cache_block_size;
2500 while (too_many_discard_blocks(discard_block_size, origin_size))
2501 discard_block_size *= 2;
2503 return discard_block_size;
2506 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2508 dm_block_t nr_blocks = from_cblock(size);
2510 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2511 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2512 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2513 "Please consider increasing the cache block size to reduce the overall cache block count.",
2514 (unsigned long long) nr_blocks);
2516 cache->cache_size = size;
2519 static int is_congested(struct dm_dev *dev, int bdi_bits)
2521 struct request_queue *q = bdev_get_queue(dev->bdev);
2522 return bdi_congested(q->backing_dev_info, bdi_bits);
2525 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2527 struct cache *cache = container_of(cb, struct cache, callbacks);
2529 return is_congested(cache->origin_dev, bdi_bits) ||
2530 is_congested(cache->cache_dev, bdi_bits);
2533 #define DEFAULT_MIGRATION_THRESHOLD 2048
2535 static int cache_create(struct cache_args *ca, struct cache **result)
2538 char **error = &ca->ti->error;
2539 struct cache *cache;
2540 struct dm_target *ti = ca->ti;
2541 dm_block_t origin_blocks;
2542 struct dm_cache_metadata *cmd;
2543 bool may_format = ca->features.mode == CM_WRITE;
2545 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2550 ti->private = cache;
2551 ti->num_flush_bios = 2;
2552 ti->flush_supported = true;
2554 ti->num_discard_bios = 1;
2555 ti->discards_supported = true;
2556 ti->split_discard_bios = false;
2558 cache->features = ca->features;
2559 ti->per_io_data_size = get_per_bio_data_size(cache);
2561 cache->callbacks.congested_fn = cache_is_congested;
2562 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2564 cache->metadata_dev = ca->metadata_dev;
2565 cache->origin_dev = ca->origin_dev;
2566 cache->cache_dev = ca->cache_dev;
2568 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2570 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2571 origin_blocks = block_div(origin_blocks, ca->block_size);
2572 cache->origin_blocks = to_oblock(origin_blocks);
2574 cache->sectors_per_block = ca->block_size;
2575 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2580 if (ca->block_size & (ca->block_size - 1)) {
2581 dm_block_t cache_size = ca->cache_sectors;
2583 cache->sectors_per_block_shift = -1;
2584 cache_size = block_div(cache_size, ca->block_size);
2585 set_cache_size(cache, to_cblock(cache_size));
2587 cache->sectors_per_block_shift = __ffs(ca->block_size);
2588 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2591 r = create_cache_policy(cache, ca, error);
2595 cache->policy_nr_args = ca->policy_argc;
2596 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2598 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2600 *error = "Error setting cache policy's config values";
2604 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2605 ca->block_size, may_format,
2606 dm_cache_policy_get_hint_size(cache->policy),
2607 ca->features.metadata_version);
2609 *error = "Error creating metadata object";
2614 set_cache_mode(cache, CM_WRITE);
2615 if (get_cache_mode(cache) != CM_WRITE) {
2616 *error = "Unable to get write access to metadata, please check/repair metadata.";
2621 if (passthrough_mode(&cache->features)) {
2624 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2626 *error = "dm_cache_metadata_all_clean() failed";
2631 *error = "Cannot enter passthrough mode unless all blocks are clean";
2636 policy_allow_migrations(cache->policy, false);
2639 spin_lock_init(&cache->lock);
2640 INIT_LIST_HEAD(&cache->deferred_cells);
2641 bio_list_init(&cache->deferred_bios);
2642 bio_list_init(&cache->deferred_writethrough_bios);
2643 atomic_set(&cache->nr_allocated_migrations, 0);
2644 atomic_set(&cache->nr_io_migrations, 0);
2645 init_waitqueue_head(&cache->migration_wait);
2648 atomic_set(&cache->nr_dirty, 0);
2649 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2650 if (!cache->dirty_bitset) {
2651 *error = "could not allocate dirty bitset";
2654 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2656 cache->discard_block_size =
2657 calculate_discard_block_size(cache->sectors_per_block,
2658 cache->origin_sectors);
2659 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2660 cache->discard_block_size));
2661 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2662 if (!cache->discard_bitset) {
2663 *error = "could not allocate discard bitset";
2666 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2668 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2669 if (IS_ERR(cache->copier)) {
2670 *error = "could not create kcopyd client";
2671 r = PTR_ERR(cache->copier);
2675 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2677 *error = "could not create workqueue for metadata object";
2680 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2681 INIT_WORK(&cache->deferred_writethrough_worker,
2682 process_deferred_writethrough_bios);
2683 INIT_WORK(&cache->migration_worker, check_migrations);
2684 INIT_DELAYED_WORK(&cache->waker, do_waker);
2686 cache->prison = dm_bio_prison_create_v2(cache->wq);
2687 if (!cache->prison) {
2688 *error = "could not create bio prison";
2692 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2694 if (!cache->migration_pool) {
2695 *error = "Error creating cache's migration mempool";
2699 cache->need_tick_bio = true;
2700 cache->sized = false;
2701 cache->invalidate = false;
2702 cache->commit_requested = false;
2703 cache->loaded_mappings = false;
2704 cache->loaded_discards = false;
2708 atomic_set(&cache->stats.demotion, 0);
2709 atomic_set(&cache->stats.promotion, 0);
2710 atomic_set(&cache->stats.copies_avoided, 0);
2711 atomic_set(&cache->stats.cache_cell_clash, 0);
2712 atomic_set(&cache->stats.commit_count, 0);
2713 atomic_set(&cache->stats.discard_count, 0);
2715 spin_lock_init(&cache->invalidation_lock);
2716 INIT_LIST_HEAD(&cache->invalidation_requests);
2718 batcher_init(&cache->committer, commit_op, cache,
2719 issue_op, cache, cache->wq);
2720 iot_init(&cache->origin_tracker);
2722 init_rwsem(&cache->background_work_lock);
2723 prevent_background_work(cache);
2732 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2737 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2740 for (i = 0; i < argc; i++) {
2741 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2750 cache->nr_ctr_args = argc;
2751 cache->ctr_args = copy;
2756 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2759 struct cache_args *ca;
2760 struct cache *cache = NULL;
2762 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2764 ti->error = "Error allocating memory for cache";
2769 r = parse_cache_args(ca, argc, argv, &ti->error);
2773 r = cache_create(ca, &cache);
2777 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2783 ti->private = cache;
2785 destroy_cache_args(ca);
2789 /*----------------------------------------------------------------*/
2791 static int cache_map(struct dm_target *ti, struct bio *bio)
2793 struct cache *cache = ti->private;
2797 dm_oblock_t block = get_bio_block(cache, bio);
2798 size_t pb_data_size = get_per_bio_data_size(cache);
2800 init_per_bio_data(bio, pb_data_size);
2801 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2803 * This can only occur if the io goes to a partial block at
2804 * the end of the origin device. We don't cache these.
2805 * Just remap to the origin and carry on.
2807 remap_to_origin(cache, bio);
2808 accounted_begin(cache, bio);
2809 return DM_MAPIO_REMAPPED;
2812 if (discard_or_flush(bio)) {
2813 defer_bio(cache, bio);
2814 return DM_MAPIO_SUBMITTED;
2817 r = map_bio(cache, bio, block, &commit_needed);
2819 schedule_commit(&cache->committer);
2824 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2826 struct cache *cache = ti->private;
2827 unsigned long flags;
2828 size_t pb_data_size = get_per_bio_data_size(cache);
2829 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2832 policy_tick(cache->policy, false);
2834 spin_lock_irqsave(&cache->lock, flags);
2835 cache->need_tick_bio = true;
2836 spin_unlock_irqrestore(&cache->lock, flags);
2839 bio_drop_shared_lock(cache, bio);
2840 accounted_complete(cache, bio);
2845 static int write_dirty_bitset(struct cache *cache)
2849 if (get_cache_mode(cache) >= CM_READ_ONLY)
2852 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2854 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2859 static int write_discard_bitset(struct cache *cache)
2863 if (get_cache_mode(cache) >= CM_READ_ONLY)
2866 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2867 cache->discard_nr_blocks);
2869 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2870 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2874 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2875 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2876 is_discarded(cache, to_dblock(i)));
2878 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2886 static int write_hints(struct cache *cache)
2890 if (get_cache_mode(cache) >= CM_READ_ONLY)
2893 r = dm_cache_write_hints(cache->cmd, cache->policy);
2895 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2903 * returns true on success
2905 static bool sync_metadata(struct cache *cache)
2909 r1 = write_dirty_bitset(cache);
2911 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2913 r2 = write_discard_bitset(cache);
2915 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2919 r3 = write_hints(cache);
2921 DMERR("%s: could not write hints", cache_device_name(cache));
2924 * If writing the above metadata failed, we still commit, but don't
2925 * set the clean shutdown flag. This will effectively force every
2926 * dirty bit to be set on reload.
2928 r4 = commit(cache, !r1 && !r2 && !r3);
2930 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2932 return !r1 && !r2 && !r3 && !r4;
2935 static void cache_postsuspend(struct dm_target *ti)
2937 struct cache *cache = ti->private;
2939 prevent_background_work(cache);
2940 BUG_ON(atomic_read(&cache->nr_io_migrations));
2942 cancel_delayed_work(&cache->waker);
2943 flush_workqueue(cache->wq);
2944 WARN_ON(cache->origin_tracker.in_flight);
2947 * If it's a flush suspend there won't be any deferred bios, so this
2950 requeue_deferred_bios(cache);
2952 if (get_cache_mode(cache) == CM_WRITE)
2953 (void) sync_metadata(cache);
2956 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2957 bool dirty, uint32_t hint, bool hint_valid)
2960 struct cache *cache = context;
2963 set_bit(from_cblock(cblock), cache->dirty_bitset);
2964 atomic_inc(&cache->nr_dirty);
2966 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2968 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2976 * The discard block size in the on disk metadata is not
2977 * neccessarily the same as we're currently using. So we have to
2978 * be careful to only set the discarded attribute if we know it
2979 * covers a complete block of the new size.
2981 struct discard_load_info {
2982 struct cache *cache;
2985 * These blocks are sized using the on disk dblock size, rather
2986 * than the current one.
2988 dm_block_t block_size;
2989 dm_block_t discard_begin, discard_end;
2992 static void discard_load_info_init(struct cache *cache,
2993 struct discard_load_info *li)
2996 li->discard_begin = li->discard_end = 0;
2999 static void set_discard_range(struct discard_load_info *li)
3003 if (li->discard_begin == li->discard_end)
3007 * Convert to sectors.
3009 b = li->discard_begin * li->block_size;
3010 e = li->discard_end * li->block_size;
3013 * Then convert back to the current dblock size.
3015 b = dm_sector_div_up(b, li->cache->discard_block_size);
3016 sector_div(e, li->cache->discard_block_size);
3019 * The origin may have shrunk, so we need to check we're still in
3022 if (e > from_dblock(li->cache->discard_nr_blocks))
3023 e = from_dblock(li->cache->discard_nr_blocks);
3026 set_discard(li->cache, to_dblock(b));
3029 static int load_discard(void *context, sector_t discard_block_size,
3030 dm_dblock_t dblock, bool discard)
3032 struct discard_load_info *li = context;
3034 li->block_size = discard_block_size;
3037 if (from_dblock(dblock) == li->discard_end)
3039 * We're already in a discard range, just extend it.
3041 li->discard_end = li->discard_end + 1ULL;
3045 * Emit the old range and start a new one.
3047 set_discard_range(li);
3048 li->discard_begin = from_dblock(dblock);
3049 li->discard_end = li->discard_begin + 1ULL;
3052 set_discard_range(li);
3053 li->discard_begin = li->discard_end = 0;
3059 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3061 sector_t size = get_dev_size(cache->cache_dev);
3062 (void) sector_div(size, cache->sectors_per_block);
3063 return to_cblock(size);
3066 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3068 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3072 * We can't drop a dirty block when shrinking the cache.
3074 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3075 new_size = to_cblock(from_cblock(new_size) + 1);
3076 if (is_dirty(cache, new_size)) {
3077 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3078 cache_device_name(cache),
3079 (unsigned long long) from_cblock(new_size));
3087 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3091 r = dm_cache_resize(cache->cmd, new_size);
3093 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3094 metadata_operation_failed(cache, "dm_cache_resize", r);
3098 set_cache_size(cache, new_size);
3103 static int cache_preresume(struct dm_target *ti)
3106 struct cache *cache = ti->private;
3107 dm_cblock_t csize = get_cache_dev_size(cache);
3110 * Check to see if the cache has resized.
3112 if (!cache->sized) {
3113 r = resize_cache_dev(cache, csize);
3117 cache->sized = true;
3119 } else if (csize != cache->cache_size) {
3120 if (!can_resize(cache, csize))
3123 r = resize_cache_dev(cache, csize);
3128 if (!cache->loaded_mappings) {
3129 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3130 load_mapping, cache);
3132 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3133 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3137 cache->loaded_mappings = true;
3140 if (!cache->loaded_discards) {
3141 struct discard_load_info li;
3144 * The discard bitset could have been resized, or the
3145 * discard block size changed. To be safe we start by
3146 * setting every dblock to not discarded.
3148 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3150 discard_load_info_init(cache, &li);
3151 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3153 DMERR("%s: could not load origin discards", cache_device_name(cache));
3154 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3157 set_discard_range(&li);
3159 cache->loaded_discards = true;
3165 static void cache_resume(struct dm_target *ti)
3167 struct cache *cache = ti->private;
3169 cache->need_tick_bio = true;
3170 allow_background_work(cache);
3171 do_waker(&cache->waker.work);
3177 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3178 * <cache block size> <#used cache blocks>/<#total cache blocks>
3179 * <#read hits> <#read misses> <#write hits> <#write misses>
3180 * <#demotions> <#promotions> <#dirty>
3181 * <#features> <features>*
3182 * <#core args> <core args>
3183 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3185 static void cache_status(struct dm_target *ti, status_type_t type,
3186 unsigned status_flags, char *result, unsigned maxlen)
3191 dm_block_t nr_free_blocks_metadata = 0;
3192 dm_block_t nr_blocks_metadata = 0;
3193 char buf[BDEVNAME_SIZE];
3194 struct cache *cache = ti->private;
3195 dm_cblock_t residency;
3199 case STATUSTYPE_INFO:
3200 if (get_cache_mode(cache) == CM_FAIL) {
3205 /* Commit to ensure statistics aren't out-of-date */
3206 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3207 (void) commit(cache, false);
3209 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3211 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3212 cache_device_name(cache), r);
3216 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3218 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3219 cache_device_name(cache), r);
3223 residency = policy_residency(cache->policy);
3225 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3226 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3227 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3228 (unsigned long long)nr_blocks_metadata,
3229 (unsigned long long)cache->sectors_per_block,
3230 (unsigned long long) from_cblock(residency),
3231 (unsigned long long) from_cblock(cache->cache_size),
3232 (unsigned) atomic_read(&cache->stats.read_hit),
3233 (unsigned) atomic_read(&cache->stats.read_miss),
3234 (unsigned) atomic_read(&cache->stats.write_hit),
3235 (unsigned) atomic_read(&cache->stats.write_miss),
3236 (unsigned) atomic_read(&cache->stats.demotion),
3237 (unsigned) atomic_read(&cache->stats.promotion),
3238 (unsigned long) atomic_read(&cache->nr_dirty));
3240 if (cache->features.metadata_version == 2)
3241 DMEMIT("2 metadata2 ");
3245 if (writethrough_mode(&cache->features))
3246 DMEMIT("writethrough ");
3248 else if (passthrough_mode(&cache->features))
3249 DMEMIT("passthrough ");
3251 else if (writeback_mode(&cache->features))
3252 DMEMIT("writeback ");
3255 DMERR("%s: internal error: unknown io mode: %d",
3256 cache_device_name(cache), (int) cache->features.io_mode);
3260 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3262 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3264 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3266 DMERR("%s: policy_emit_config_values returned %d",
3267 cache_device_name(cache), r);
3270 if (get_cache_mode(cache) == CM_READ_ONLY)
3275 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3277 if (r || needs_check)
3278 DMEMIT("needs_check ");
3284 case STATUSTYPE_TABLE:
3285 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3287 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3289 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3292 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3293 DMEMIT(" %s", cache->ctr_args[i]);
3294 if (cache->nr_ctr_args)
3295 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3305 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3306 * the one-past-the-end value.
3308 struct cblock_range {
3314 * A cache block range can take two forms:
3316 * i) A single cblock, eg. '3456'
3317 * ii) A begin and end cblock with a dash between, eg. 123-234
3319 static int parse_cblock_range(struct cache *cache, const char *str,
3320 struct cblock_range *result)
3327 * Try and parse form (ii) first.
3329 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3334 result->begin = to_cblock(b);
3335 result->end = to_cblock(e);
3340 * That didn't work, try form (i).
3342 r = sscanf(str, "%llu%c", &b, &dummy);
3347 result->begin = to_cblock(b);
3348 result->end = to_cblock(from_cblock(result->begin) + 1u);
3352 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3356 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3358 uint64_t b = from_cblock(range->begin);
3359 uint64_t e = from_cblock(range->end);
3360 uint64_t n = from_cblock(cache->cache_size);
3363 DMERR("%s: begin cblock out of range: %llu >= %llu",
3364 cache_device_name(cache), b, n);
3369 DMERR("%s: end cblock out of range: %llu > %llu",
3370 cache_device_name(cache), e, n);
3375 DMERR("%s: invalid cblock range: %llu >= %llu",
3376 cache_device_name(cache), b, e);
3383 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3385 return to_cblock(from_cblock(b) + 1);
3388 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3393 * We don't need to do any locking here because we know we're in
3394 * passthrough mode. There's is potential for a race between an
3395 * invalidation triggered by an io and an invalidation message. This
3396 * is harmless, we must not worry if the policy call fails.
3398 while (range->begin != range->end) {
3399 r = invalidate_cblock(cache, range->begin);
3403 range->begin = cblock_succ(range->begin);
3406 cache->commit_requested = true;
3410 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3411 const char **cblock_ranges)
3415 struct cblock_range range;
3417 if (!passthrough_mode(&cache->features)) {
3418 DMERR("%s: cache has to be in passthrough mode for invalidation",
3419 cache_device_name(cache));
3423 for (i = 0; i < count; i++) {
3424 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3428 r = validate_cblock_range(cache, &range);
3433 * Pass begin and end origin blocks to the worker and wake it.
3435 r = request_invalidation(cache, &range);
3447 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3449 * The key migration_threshold is supported by the cache target core.
3451 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3453 struct cache *cache = ti->private;
3458 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3459 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3460 cache_device_name(cache));
3464 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3465 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3470 return set_config_value(cache, argv[0], argv[1]);
3473 static int cache_iterate_devices(struct dm_target *ti,
3474 iterate_devices_callout_fn fn, void *data)
3477 struct cache *cache = ti->private;
3479 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3481 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3486 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3489 * FIXME: these limits may be incompatible with the cache device
3491 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3492 cache->origin_sectors);
3493 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3496 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3498 struct cache *cache = ti->private;
3499 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3502 * If the system-determined stacked limits are compatible with the
3503 * cache's blocksize (io_opt is a factor) do not override them.
3505 if (io_opt_sectors < cache->sectors_per_block ||
3506 do_div(io_opt_sectors, cache->sectors_per_block)) {
3507 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3508 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3510 set_discard_limits(cache, limits);
3513 /*----------------------------------------------------------------*/
3515 static struct target_type cache_target = {
3517 .version = {2, 0, 0},
3518 .module = THIS_MODULE,
3522 .end_io = cache_end_io,
3523 .postsuspend = cache_postsuspend,
3524 .preresume = cache_preresume,
3525 .resume = cache_resume,
3526 .status = cache_status,
3527 .message = cache_message,
3528 .iterate_devices = cache_iterate_devices,
3529 .io_hints = cache_io_hints,
3532 static int __init dm_cache_init(void)
3536 r = dm_register_target(&cache_target);
3538 DMERR("cache target registration failed: %d", r);
3542 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3543 if (!migration_cache) {
3544 dm_unregister_target(&cache_target);
3551 static void __exit dm_cache_exit(void)
3553 dm_unregister_target(&cache_target);
3554 kmem_cache_destroy(migration_cache);
3557 module_init(dm_cache_init);
3558 module_exit(dm_cache_exit);
3560 MODULE_DESCRIPTION(DM_NAME " cache target");
3561 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3562 MODULE_LICENSE("GPL");