2 * Copyright (C) 2012 Red Hat. All rights reserved.
4 * This file is released under the GPL.
8 #include "dm-bio-prison-v2.h"
9 #include "dm-bio-record.h"
10 #include "dm-cache-metadata.h"
12 #include <linux/dm-io.h>
13 #include <linux/dm-kcopyd.h>
14 #include <linux/jiffies.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/rwsem.h>
19 #include <linux/slab.h>
20 #include <linux/vmalloc.h>
22 #define DM_MSG_PREFIX "cache"
24 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
25 "A percentage of time allocated for copying to and/or from cache");
27 /*----------------------------------------------------------------*/
32 * oblock: index of an origin block
33 * cblock: index of a cache block
34 * promotion: movement of a block from origin to cache
35 * demotion: movement of a block from cache to origin
36 * migration: movement of a block between the origin and cache device,
40 /*----------------------------------------------------------------*/
46 * Sectors of in-flight IO.
51 * The time, in jiffies, when this device became idle (if it is
54 unsigned long idle_time;
55 unsigned long last_update_time;
58 static void iot_init(struct io_tracker *iot)
60 spin_lock_init(&iot->lock);
63 iot->last_update_time = jiffies;
66 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
71 return time_after(jiffies, iot->idle_time + jifs);
74 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
79 spin_lock_irqsave(&iot->lock, flags);
80 r = __iot_idle_for(iot, jifs);
81 spin_unlock_irqrestore(&iot->lock, flags);
86 static void iot_io_begin(struct io_tracker *iot, sector_t len)
90 spin_lock_irqsave(&iot->lock, flags);
91 iot->in_flight += len;
92 spin_unlock_irqrestore(&iot->lock, flags);
95 static void __iot_io_end(struct io_tracker *iot, sector_t len)
100 iot->in_flight -= len;
102 iot->idle_time = jiffies;
105 static void iot_io_end(struct io_tracker *iot, sector_t len)
109 spin_lock_irqsave(&iot->lock, flags);
110 __iot_io_end(iot, len);
111 spin_unlock_irqrestore(&iot->lock, flags);
114 /*----------------------------------------------------------------*/
117 * Represents a chunk of future work. 'input' allows continuations to pass
118 * values between themselves, typically error values.
120 struct continuation {
121 struct work_struct ws;
125 static inline void init_continuation(struct continuation *k,
126 void (*fn)(struct work_struct *))
128 INIT_WORK(&k->ws, fn);
132 static inline void queue_continuation(struct workqueue_struct *wq,
133 struct continuation *k)
135 queue_work(wq, &k->ws);
138 /*----------------------------------------------------------------*/
141 * The batcher collects together pieces of work that need a particular
142 * operation to occur before they can proceed (typically a commit).
146 * The operation that everyone is waiting for.
148 blk_status_t (*commit_op)(void *context);
149 void *commit_context;
152 * This is how bios should be issued once the commit op is complete
153 * (accounted_request).
155 void (*issue_op)(struct bio *bio, void *context);
159 * Queued work gets put on here after commit.
161 struct workqueue_struct *wq;
164 struct list_head work_items;
165 struct bio_list bios;
166 struct work_struct commit_work;
168 bool commit_scheduled;
171 static void __commit(struct work_struct *_ws)
173 struct batcher *b = container_of(_ws, struct batcher, commit_work);
176 struct list_head work_items;
177 struct work_struct *ws, *tmp;
178 struct continuation *k;
180 struct bio_list bios;
182 INIT_LIST_HEAD(&work_items);
183 bio_list_init(&bios);
186 * We have to grab these before the commit_op to avoid a race
189 spin_lock_irqsave(&b->lock, flags);
190 list_splice_init(&b->work_items, &work_items);
191 bio_list_merge(&bios, &b->bios);
192 bio_list_init(&b->bios);
193 b->commit_scheduled = false;
194 spin_unlock_irqrestore(&b->lock, flags);
196 r = b->commit_op(b->commit_context);
198 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
199 k = container_of(ws, struct continuation, ws);
201 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
202 queue_work(b->wq, ws);
205 while ((bio = bio_list_pop(&bios))) {
210 b->issue_op(bio, b->issue_context);
214 static void batcher_init(struct batcher *b,
215 blk_status_t (*commit_op)(void *),
216 void *commit_context,
217 void (*issue_op)(struct bio *bio, void *),
219 struct workqueue_struct *wq)
221 b->commit_op = commit_op;
222 b->commit_context = commit_context;
223 b->issue_op = issue_op;
224 b->issue_context = issue_context;
227 spin_lock_init(&b->lock);
228 INIT_LIST_HEAD(&b->work_items);
229 bio_list_init(&b->bios);
230 INIT_WORK(&b->commit_work, __commit);
231 b->commit_scheduled = false;
234 static void async_commit(struct batcher *b)
236 queue_work(b->wq, &b->commit_work);
239 static void continue_after_commit(struct batcher *b, struct continuation *k)
242 bool commit_scheduled;
244 spin_lock_irqsave(&b->lock, flags);
245 commit_scheduled = b->commit_scheduled;
246 list_add_tail(&k->ws.entry, &b->work_items);
247 spin_unlock_irqrestore(&b->lock, flags);
249 if (commit_scheduled)
254 * Bios are errored if commit failed.
256 static void issue_after_commit(struct batcher *b, struct bio *bio)
259 bool commit_scheduled;
261 spin_lock_irqsave(&b->lock, flags);
262 commit_scheduled = b->commit_scheduled;
263 bio_list_add(&b->bios, bio);
264 spin_unlock_irqrestore(&b->lock, flags);
266 if (commit_scheduled)
271 * Call this if some urgent work is waiting for the commit to complete.
273 static void schedule_commit(struct batcher *b)
278 spin_lock_irqsave(&b->lock, flags);
279 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
280 b->commit_scheduled = true;
281 spin_unlock_irqrestore(&b->lock, flags);
288 * There are a couple of places where we let a bio run, but want to do some
289 * work before calling its endio function. We do this by temporarily
290 * changing the endio fn.
292 struct dm_hook_info {
293 bio_end_io_t *bi_end_io;
296 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
297 bio_end_io_t *bi_end_io, void *bi_private)
299 h->bi_end_io = bio->bi_end_io;
301 bio->bi_end_io = bi_end_io;
302 bio->bi_private = bi_private;
305 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
307 bio->bi_end_io = h->bi_end_io;
310 /*----------------------------------------------------------------*/
312 #define MIGRATION_POOL_SIZE 128
313 #define COMMIT_PERIOD HZ
314 #define MIGRATION_COUNT_WINDOW 10
317 * The block size of the device holding cache data must be
318 * between 32KB and 1GB.
320 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
321 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
323 enum cache_metadata_mode {
324 CM_WRITE, /* metadata may be changed */
325 CM_READ_ONLY, /* metadata may not be changed */
331 * Data is written to cached blocks only. These blocks are marked
332 * dirty. If you lose the cache device you will lose data.
333 * Potential performance increase for both reads and writes.
338 * Data is written to both cache and origin. Blocks are never
339 * dirty. Potential performance benfit for reads only.
344 * A degraded mode useful for various cache coherency situations
345 * (eg, rolling back snapshots). Reads and writes always go to the
346 * origin. If a write goes to a cached oblock, then the cache
347 * block is invalidated.
352 struct cache_features {
353 enum cache_metadata_mode mode;
354 enum cache_io_mode io_mode;
355 unsigned metadata_version;
366 atomic_t copies_avoided;
367 atomic_t cache_cell_clash;
368 atomic_t commit_count;
369 atomic_t discard_count;
373 struct dm_target *ti;
374 struct dm_target_callbacks callbacks;
376 struct dm_cache_metadata *cmd;
379 * Metadata is written to this device.
381 struct dm_dev *metadata_dev;
384 * The slower of the two data devices. Typically a spindle.
386 struct dm_dev *origin_dev;
389 * The faster of the two data devices. Typically an SSD.
391 struct dm_dev *cache_dev;
394 * Size of the origin device in _complete_ blocks and native sectors.
396 dm_oblock_t origin_blocks;
397 sector_t origin_sectors;
400 * Size of the cache device in blocks.
402 dm_cblock_t cache_size;
405 * Fields for converting from sectors to blocks.
407 sector_t sectors_per_block;
408 int sectors_per_block_shift;
411 struct list_head deferred_cells;
412 struct bio_list deferred_bios;
413 struct bio_list deferred_writethrough_bios;
414 sector_t migration_threshold;
415 wait_queue_head_t migration_wait;
416 atomic_t nr_allocated_migrations;
419 * The number of in flight migrations that are performing
420 * background io. eg, promotion, writeback.
422 atomic_t nr_io_migrations;
424 struct rw_semaphore quiesce_lock;
427 * cache_size entries, dirty if set
430 unsigned long *dirty_bitset;
433 * origin_blocks entries, discarded if set.
435 dm_dblock_t discard_nr_blocks;
436 unsigned long *discard_bitset;
437 uint32_t discard_block_size; /* a power of 2 times sectors per block */
440 * Rather than reconstructing the table line for the status we just
441 * save it and regurgitate.
443 unsigned nr_ctr_args;
444 const char **ctr_args;
446 struct dm_kcopyd_client *copier;
447 struct workqueue_struct *wq;
448 struct work_struct deferred_bio_worker;
449 struct work_struct deferred_writethrough_worker;
450 struct work_struct migration_worker;
451 struct delayed_work waker;
452 struct dm_bio_prison_v2 *prison;
454 mempool_t *migration_pool;
456 struct dm_cache_policy *policy;
457 unsigned policy_nr_args;
459 bool need_tick_bio:1;
462 bool commit_requested:1;
463 bool loaded_mappings:1;
464 bool loaded_discards:1;
467 * Cache features such as write-through.
469 struct cache_features features;
471 struct cache_stats stats;
474 * Invalidation fields.
476 spinlock_t invalidation_lock;
477 struct list_head invalidation_requests;
479 struct io_tracker tracker;
481 struct work_struct commit_ws;
482 struct batcher committer;
484 struct rw_semaphore background_work_lock;
487 struct per_bio_data {
490 struct dm_bio_prison_cell_v2 *cell;
491 struct dm_hook_info hook_info;
495 * writethrough fields. These MUST remain at the end of this
496 * structure and the 'cache' member must be the first as it
497 * is used to determine the offset of the writethrough fields.
501 struct dm_bio_details bio_details;
504 struct dm_cache_migration {
505 struct continuation k;
508 struct policy_work *op;
509 struct bio *overwrite_bio;
510 struct dm_bio_prison_cell_v2 *cell;
512 dm_cblock_t invalidate_cblock;
513 dm_oblock_t invalidate_oblock;
516 /*----------------------------------------------------------------*/
518 static bool writethrough_mode(struct cache_features *f)
520 return f->io_mode == CM_IO_WRITETHROUGH;
523 static bool writeback_mode(struct cache_features *f)
525 return f->io_mode == CM_IO_WRITEBACK;
528 static inline bool passthrough_mode(struct cache_features *f)
530 return unlikely(f->io_mode == CM_IO_PASSTHROUGH);
533 /*----------------------------------------------------------------*/
535 static void wake_deferred_bio_worker(struct cache *cache)
537 queue_work(cache->wq, &cache->deferred_bio_worker);
540 static void wake_deferred_writethrough_worker(struct cache *cache)
542 queue_work(cache->wq, &cache->deferred_writethrough_worker);
545 static void wake_migration_worker(struct cache *cache)
547 if (passthrough_mode(&cache->features))
550 queue_work(cache->wq, &cache->migration_worker);
553 /*----------------------------------------------------------------*/
555 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
557 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
560 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
562 dm_bio_prison_free_cell_v2(cache->prison, cell);
565 static struct dm_cache_migration *alloc_migration(struct cache *cache)
567 struct dm_cache_migration *mg;
569 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
572 atomic_inc(&mg->cache->nr_allocated_migrations);
578 static void free_migration(struct dm_cache_migration *mg)
580 struct cache *cache = mg->cache;
582 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
583 wake_up(&cache->migration_wait);
585 mempool_free(mg, cache->migration_pool);
588 /*----------------------------------------------------------------*/
590 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
592 return to_oblock(from_oblock(b) + 1ull);
595 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
599 key->block_begin = from_oblock(begin);
600 key->block_end = from_oblock(end);
604 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
605 * level 1 which prevents *both* READs and WRITEs.
607 #define WRITE_LOCK_LEVEL 0
608 #define READ_WRITE_LOCK_LEVEL 1
610 static unsigned lock_level(struct bio *bio)
612 return bio_data_dir(bio) == WRITE ?
614 READ_WRITE_LOCK_LEVEL;
617 /*----------------------------------------------------------------
619 *--------------------------------------------------------------*/
622 * If using writeback, leave out struct per_bio_data's writethrough fields.
624 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
625 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
627 static size_t get_per_bio_data_size(struct cache *cache)
629 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
632 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
634 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
639 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
641 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
644 pb->req_nr = dm_bio_get_target_bio_nr(bio);
651 /*----------------------------------------------------------------*/
653 static void defer_bio(struct cache *cache, struct bio *bio)
657 spin_lock_irqsave(&cache->lock, flags);
658 bio_list_add(&cache->deferred_bios, bio);
659 spin_unlock_irqrestore(&cache->lock, flags);
661 wake_deferred_bio_worker(cache);
664 static void defer_bios(struct cache *cache, struct bio_list *bios)
668 spin_lock_irqsave(&cache->lock, flags);
669 bio_list_merge(&cache->deferred_bios, bios);
671 spin_unlock_irqrestore(&cache->lock, flags);
673 wake_deferred_bio_worker(cache);
676 /*----------------------------------------------------------------*/
678 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
682 struct per_bio_data *pb;
683 struct dm_cell_key_v2 key;
684 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
685 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
687 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
688 if (!cell_prealloc) {
689 defer_bio(cache, bio);
693 build_key(oblock, end, &key);
694 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
697 * Failed to get the lock.
699 free_prison_cell(cache, cell_prealloc);
703 if (cell != cell_prealloc)
704 free_prison_cell(cache, cell_prealloc);
706 pb_size = get_per_bio_data_size(cache);
707 pb = get_per_bio_data(bio, pb_size);
713 /*----------------------------------------------------------------*/
715 static bool is_dirty(struct cache *cache, dm_cblock_t b)
717 return test_bit(from_cblock(b), cache->dirty_bitset);
720 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
722 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
723 atomic_inc(&cache->nr_dirty);
724 policy_set_dirty(cache->policy, cblock);
729 * These two are called when setting after migrations to force the policy
730 * and dirty bitset to be in sync.
732 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
734 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
735 atomic_inc(&cache->nr_dirty);
736 policy_set_dirty(cache->policy, cblock);
739 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
741 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
742 if (atomic_dec_return(&cache->nr_dirty) == 0)
743 dm_table_event(cache->ti->table);
746 policy_clear_dirty(cache->policy, cblock);
749 /*----------------------------------------------------------------*/
751 static bool block_size_is_power_of_two(struct cache *cache)
753 return cache->sectors_per_block_shift >= 0;
756 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
757 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
760 static dm_block_t block_div(dm_block_t b, uint32_t n)
767 static dm_block_t oblocks_per_dblock(struct cache *cache)
769 dm_block_t oblocks = cache->discard_block_size;
771 if (block_size_is_power_of_two(cache))
772 oblocks >>= cache->sectors_per_block_shift;
774 oblocks = block_div(oblocks, cache->sectors_per_block);
779 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
781 return to_dblock(block_div(from_oblock(oblock),
782 oblocks_per_dblock(cache)));
785 static void set_discard(struct cache *cache, dm_dblock_t b)
789 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
790 atomic_inc(&cache->stats.discard_count);
792 spin_lock_irqsave(&cache->lock, flags);
793 set_bit(from_dblock(b), cache->discard_bitset);
794 spin_unlock_irqrestore(&cache->lock, flags);
797 static void clear_discard(struct cache *cache, dm_dblock_t b)
801 spin_lock_irqsave(&cache->lock, flags);
802 clear_bit(from_dblock(b), cache->discard_bitset);
803 spin_unlock_irqrestore(&cache->lock, flags);
806 static bool is_discarded(struct cache *cache, dm_dblock_t b)
811 spin_lock_irqsave(&cache->lock, flags);
812 r = test_bit(from_dblock(b), cache->discard_bitset);
813 spin_unlock_irqrestore(&cache->lock, flags);
818 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
823 spin_lock_irqsave(&cache->lock, flags);
824 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
825 cache->discard_bitset);
826 spin_unlock_irqrestore(&cache->lock, flags);
831 /*----------------------------------------------------------------
833 *--------------------------------------------------------------*/
834 static void remap_to_origin(struct cache *cache, struct bio *bio)
836 bio->bi_bdev = cache->origin_dev->bdev;
839 static void remap_to_cache(struct cache *cache, struct bio *bio,
842 sector_t bi_sector = bio->bi_iter.bi_sector;
843 sector_t block = from_cblock(cblock);
845 bio->bi_bdev = cache->cache_dev->bdev;
846 if (!block_size_is_power_of_two(cache))
847 bio->bi_iter.bi_sector =
848 (block * cache->sectors_per_block) +
849 sector_div(bi_sector, cache->sectors_per_block);
851 bio->bi_iter.bi_sector =
852 (block << cache->sectors_per_block_shift) |
853 (bi_sector & (cache->sectors_per_block - 1));
856 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
859 size_t pb_data_size = get_per_bio_data_size(cache);
860 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
862 spin_lock_irqsave(&cache->lock, flags);
863 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
864 bio_op(bio) != REQ_OP_DISCARD) {
866 cache->need_tick_bio = false;
868 spin_unlock_irqrestore(&cache->lock, flags);
871 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
874 // FIXME: this is called way too much.
875 check_if_tick_bio_needed(cache, bio);
876 remap_to_origin(cache, bio);
877 if (bio_data_dir(bio) == WRITE)
878 clear_discard(cache, oblock_to_dblock(cache, oblock));
881 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
882 dm_oblock_t oblock, dm_cblock_t cblock)
884 check_if_tick_bio_needed(cache, bio);
885 remap_to_cache(cache, bio, cblock);
886 if (bio_data_dir(bio) == WRITE) {
887 set_dirty(cache, cblock);
888 clear_discard(cache, oblock_to_dblock(cache, oblock));
892 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
894 sector_t block_nr = bio->bi_iter.bi_sector;
896 if (!block_size_is_power_of_two(cache))
897 (void) sector_div(block_nr, cache->sectors_per_block);
899 block_nr >>= cache->sectors_per_block_shift;
901 return to_oblock(block_nr);
904 static bool accountable_bio(struct cache *cache, struct bio *bio)
906 return bio_op(bio) != REQ_OP_DISCARD;
909 static void accounted_begin(struct cache *cache, struct bio *bio)
911 size_t pb_data_size = get_per_bio_data_size(cache);
912 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
914 if (accountable_bio(cache, bio)) {
915 pb->len = bio_sectors(bio);
916 iot_io_begin(&cache->tracker, pb->len);
920 static void accounted_complete(struct cache *cache, struct bio *bio)
922 size_t pb_data_size = get_per_bio_data_size(cache);
923 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
925 iot_io_end(&cache->tracker, pb->len);
928 static void accounted_request(struct cache *cache, struct bio *bio)
930 accounted_begin(cache, bio);
931 generic_make_request(bio);
934 static void issue_op(struct bio *bio, void *context)
936 struct cache *cache = context;
937 accounted_request(cache, bio);
940 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
944 spin_lock_irqsave(&cache->lock, flags);
945 bio_list_add(&cache->deferred_writethrough_bios, bio);
946 spin_unlock_irqrestore(&cache->lock, flags);
948 wake_deferred_writethrough_worker(cache);
951 static void writethrough_endio(struct bio *bio)
953 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
955 dm_unhook_bio(&pb->hook_info, bio);
957 if (bio->bi_status) {
962 dm_bio_restore(&pb->bio_details, bio);
963 remap_to_cache(pb->cache, bio, pb->cblock);
966 * We can't issue this bio directly, since we're in interrupt
967 * context. So it gets put on a bio list for processing by the
970 defer_writethrough_bio(pb->cache, bio);
974 * FIXME: send in parallel, huge latency as is.
975 * When running in writethrough mode we need to send writes to clean blocks
976 * to both the cache and origin devices. In future we'd like to clone the
977 * bio and send them in parallel, but for now we're doing them in
978 * series as this is easier.
980 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
981 dm_oblock_t oblock, dm_cblock_t cblock)
983 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
987 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
988 dm_bio_record(&pb->bio_details, bio);
990 remap_to_origin_clear_discard(pb->cache, bio, oblock);
993 /*----------------------------------------------------------------
995 *--------------------------------------------------------------*/
996 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
998 return cache->features.mode;
1001 static const char *cache_device_name(struct cache *cache)
1003 return dm_device_name(dm_table_get_md(cache->ti->table));
1006 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
1008 const char *descs[] = {
1014 dm_table_event(cache->ti->table);
1015 DMINFO("%s: switching cache to %s mode",
1016 cache_device_name(cache), descs[(int)mode]);
1019 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
1022 enum cache_metadata_mode old_mode = get_cache_mode(cache);
1024 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
1025 DMERR("%s: unable to read needs_check flag, setting failure mode.",
1026 cache_device_name(cache));
1030 if (new_mode == CM_WRITE && needs_check) {
1031 DMERR("%s: unable to switch cache to write mode until repaired.",
1032 cache_device_name(cache));
1033 if (old_mode != new_mode)
1034 new_mode = old_mode;
1036 new_mode = CM_READ_ONLY;
1039 /* Never move out of fail mode */
1040 if (old_mode == CM_FAIL)
1046 dm_cache_metadata_set_read_only(cache->cmd);
1050 dm_cache_metadata_set_read_write(cache->cmd);
1054 cache->features.mode = new_mode;
1056 if (new_mode != old_mode)
1057 notify_mode_switch(cache, new_mode);
1060 static void abort_transaction(struct cache *cache)
1062 const char *dev_name = cache_device_name(cache);
1064 if (get_cache_mode(cache) >= CM_READ_ONLY)
1067 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1068 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1069 set_cache_mode(cache, CM_FAIL);
1072 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1073 if (dm_cache_metadata_abort(cache->cmd)) {
1074 DMERR("%s: failed to abort metadata transaction", dev_name);
1075 set_cache_mode(cache, CM_FAIL);
1079 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1081 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1082 cache_device_name(cache), op, r);
1083 abort_transaction(cache);
1084 set_cache_mode(cache, CM_READ_ONLY);
1087 /*----------------------------------------------------------------*/
1089 static void load_stats(struct cache *cache)
1091 struct dm_cache_statistics stats;
1093 dm_cache_metadata_get_stats(cache->cmd, &stats);
1094 atomic_set(&cache->stats.read_hit, stats.read_hits);
1095 atomic_set(&cache->stats.read_miss, stats.read_misses);
1096 atomic_set(&cache->stats.write_hit, stats.write_hits);
1097 atomic_set(&cache->stats.write_miss, stats.write_misses);
1100 static void save_stats(struct cache *cache)
1102 struct dm_cache_statistics stats;
1104 if (get_cache_mode(cache) >= CM_READ_ONLY)
1107 stats.read_hits = atomic_read(&cache->stats.read_hit);
1108 stats.read_misses = atomic_read(&cache->stats.read_miss);
1109 stats.write_hits = atomic_read(&cache->stats.write_hit);
1110 stats.write_misses = atomic_read(&cache->stats.write_miss);
1112 dm_cache_metadata_set_stats(cache->cmd, &stats);
1115 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1118 case POLICY_PROMOTE:
1119 atomic_inc(&stats->promotion);
1123 atomic_inc(&stats->demotion);
1126 case POLICY_WRITEBACK:
1127 atomic_inc(&stats->writeback);
1132 /*----------------------------------------------------------------
1133 * Migration processing
1135 * Migration covers moving data from the origin device to the cache, or
1137 *--------------------------------------------------------------*/
1139 static void inc_io_migrations(struct cache *cache)
1141 atomic_inc(&cache->nr_io_migrations);
1144 static void dec_io_migrations(struct cache *cache)
1146 atomic_dec(&cache->nr_io_migrations);
1149 static bool discard_or_flush(struct bio *bio)
1151 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1154 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1155 dm_dblock_t *b, dm_dblock_t *e)
1157 sector_t sb = bio->bi_iter.bi_sector;
1158 sector_t se = bio_end_sector(bio);
1160 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1162 if (se - sb < cache->discard_block_size)
1165 *e = to_dblock(block_div(se, cache->discard_block_size));
1168 /*----------------------------------------------------------------*/
1170 static void prevent_background_work(struct cache *cache)
1173 down_write(&cache->background_work_lock);
1177 static void allow_background_work(struct cache *cache)
1180 up_write(&cache->background_work_lock);
1184 static bool background_work_begin(struct cache *cache)
1189 r = down_read_trylock(&cache->background_work_lock);
1195 static void background_work_end(struct cache *cache)
1198 up_read(&cache->background_work_lock);
1202 /*----------------------------------------------------------------*/
1204 static void quiesce(struct dm_cache_migration *mg,
1205 void (*continuation)(struct work_struct *))
1207 init_continuation(&mg->k, continuation);
1208 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1211 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1213 struct continuation *k = container_of(ws, struct continuation, ws);
1214 return container_of(k, struct dm_cache_migration, k);
1217 static void copy_complete(int read_err, unsigned long write_err, void *context)
1219 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1221 if (read_err || write_err)
1222 mg->k.input = BLK_STS_IOERR;
1224 queue_continuation(mg->cache->wq, &mg->k);
1227 static int copy(struct dm_cache_migration *mg, bool promote)
1230 struct dm_io_region o_region, c_region;
1231 struct cache *cache = mg->cache;
1233 o_region.bdev = cache->origin_dev->bdev;
1234 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1235 o_region.count = cache->sectors_per_block;
1237 c_region.bdev = cache->cache_dev->bdev;
1238 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1239 c_region.count = cache->sectors_per_block;
1242 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1244 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1249 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1251 size_t pb_data_size = get_per_bio_data_size(cache);
1252 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1254 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1255 free_prison_cell(cache, pb->cell);
1259 static void overwrite_endio(struct bio *bio)
1261 struct dm_cache_migration *mg = bio->bi_private;
1262 struct cache *cache = mg->cache;
1263 size_t pb_data_size = get_per_bio_data_size(cache);
1264 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1266 dm_unhook_bio(&pb->hook_info, bio);
1269 mg->k.input = bio->bi_status;
1271 queue_continuation(mg->cache->wq, &mg->k);
1274 static void overwrite(struct dm_cache_migration *mg,
1275 void (*continuation)(struct work_struct *))
1277 struct bio *bio = mg->overwrite_bio;
1278 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1279 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1281 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1284 * The overwrite bio is part of the copy operation, as such it does
1285 * not set/clear discard or dirty flags.
1287 if (mg->op->op == POLICY_PROMOTE)
1288 remap_to_cache(mg->cache, bio, mg->op->cblock);
1290 remap_to_origin(mg->cache, bio);
1292 init_continuation(&mg->k, continuation);
1293 accounted_request(mg->cache, bio);
1299 * 1) exclusive lock preventing WRITEs
1301 * 3) copy or issue overwrite bio
1302 * 4) upgrade to exclusive lock preventing READs and WRITEs
1304 * 6) update metadata and commit
1307 static void mg_complete(struct dm_cache_migration *mg, bool success)
1309 struct bio_list bios;
1310 struct cache *cache = mg->cache;
1311 struct policy_work *op = mg->op;
1312 dm_cblock_t cblock = op->cblock;
1315 update_stats(&cache->stats, op->op);
1318 case POLICY_PROMOTE:
1319 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1320 policy_complete_background_work(cache->policy, op, success);
1322 if (mg->overwrite_bio) {
1324 force_set_dirty(cache, cblock);
1325 else if (mg->k.input)
1326 mg->overwrite_bio->bi_status = mg->k.input;
1328 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1329 bio_endio(mg->overwrite_bio);
1332 force_clear_dirty(cache, cblock);
1333 dec_io_migrations(cache);
1339 * We clear dirty here to update the nr_dirty counter.
1342 force_clear_dirty(cache, cblock);
1343 policy_complete_background_work(cache->policy, op, success);
1344 dec_io_migrations(cache);
1347 case POLICY_WRITEBACK:
1349 force_clear_dirty(cache, cblock);
1350 policy_complete_background_work(cache->policy, op, success);
1351 dec_io_migrations(cache);
1355 bio_list_init(&bios);
1357 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1358 free_prison_cell(cache, mg->cell);
1362 defer_bios(cache, &bios);
1363 wake_migration_worker(cache);
1365 background_work_end(cache);
1368 static void mg_success(struct work_struct *ws)
1370 struct dm_cache_migration *mg = ws_to_mg(ws);
1371 mg_complete(mg, mg->k.input == 0);
1374 static void mg_update_metadata(struct work_struct *ws)
1377 struct dm_cache_migration *mg = ws_to_mg(ws);
1378 struct cache *cache = mg->cache;
1379 struct policy_work *op = mg->op;
1382 case POLICY_PROMOTE:
1383 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1385 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1386 cache_device_name(cache));
1387 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1389 mg_complete(mg, false);
1392 mg_complete(mg, true);
1396 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1398 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1399 cache_device_name(cache));
1400 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1402 mg_complete(mg, false);
1407 * It would be nice if we only had to commit when a REQ_FLUSH
1408 * comes through. But there's one scenario that we have to
1411 * - vblock x in a cache block
1413 * - cache block gets reallocated and over written
1416 * When we recover, because there was no commit the cache will
1417 * rollback to having the data for vblock x in the cache block.
1418 * But the cache block has since been overwritten, so it'll end
1419 * up pointing to data that was never in 'x' during the history
1422 * To avoid this issue we require a commit as part of the
1423 * demotion operation.
1425 init_continuation(&mg->k, mg_success);
1426 continue_after_commit(&cache->committer, &mg->k);
1427 schedule_commit(&cache->committer);
1430 case POLICY_WRITEBACK:
1431 mg_complete(mg, true);
1436 static void mg_update_metadata_after_copy(struct work_struct *ws)
1438 struct dm_cache_migration *mg = ws_to_mg(ws);
1441 * Did the copy succeed?
1444 mg_complete(mg, false);
1446 mg_update_metadata(ws);
1449 static void mg_upgrade_lock(struct work_struct *ws)
1452 struct dm_cache_migration *mg = ws_to_mg(ws);
1455 * Did the copy succeed?
1458 mg_complete(mg, false);
1462 * Now we want the lock to prevent both reads and writes.
1464 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1465 READ_WRITE_LOCK_LEVEL);
1467 mg_complete(mg, false);
1470 quiesce(mg, mg_update_metadata);
1473 mg_update_metadata(ws);
1477 static void mg_copy(struct work_struct *ws)
1480 struct dm_cache_migration *mg = ws_to_mg(ws);
1482 if (mg->overwrite_bio) {
1484 * It's safe to do this here, even though it's new data
1485 * because all IO has been locked out of the block.
1487 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1488 * so _not_ using mg_upgrade_lock() as continutation.
1490 overwrite(mg, mg_update_metadata_after_copy);
1493 struct cache *cache = mg->cache;
1494 struct policy_work *op = mg->op;
1495 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1497 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1498 is_discarded_oblock(cache, op->oblock)) {
1499 mg_upgrade_lock(ws);
1503 init_continuation(&mg->k, mg_upgrade_lock);
1505 r = copy(mg, is_policy_promote);
1507 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1508 mg->k.input = BLK_STS_IOERR;
1509 mg_complete(mg, false);
1514 static int mg_lock_writes(struct dm_cache_migration *mg)
1517 struct dm_cell_key_v2 key;
1518 struct cache *cache = mg->cache;
1519 struct dm_bio_prison_cell_v2 *prealloc;
1521 prealloc = alloc_prison_cell(cache);
1523 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1524 mg_complete(mg, false);
1529 * Prevent writes to the block, but allow reads to continue.
1530 * Unless we're using an overwrite bio, in which case we lock
1533 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1534 r = dm_cell_lock_v2(cache->prison, &key,
1535 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1536 prealloc, &mg->cell);
1538 free_prison_cell(cache, prealloc);
1539 mg_complete(mg, false);
1543 if (mg->cell != prealloc)
1544 free_prison_cell(cache, prealloc);
1549 quiesce(mg, mg_copy);
1554 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1556 struct dm_cache_migration *mg;
1558 if (!background_work_begin(cache)) {
1559 policy_complete_background_work(cache->policy, op, false);
1563 mg = alloc_migration(cache);
1565 policy_complete_background_work(cache->policy, op, false);
1566 background_work_end(cache);
1570 memset(mg, 0, sizeof(*mg));
1574 mg->overwrite_bio = bio;
1577 inc_io_migrations(cache);
1579 return mg_lock_writes(mg);
1582 /*----------------------------------------------------------------
1583 * invalidation processing
1584 *--------------------------------------------------------------*/
1586 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1588 struct bio_list bios;
1589 struct cache *cache = mg->cache;
1591 bio_list_init(&bios);
1592 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1593 free_prison_cell(cache, mg->cell);
1595 if (!success && mg->overwrite_bio)
1596 bio_io_error(mg->overwrite_bio);
1599 defer_bios(cache, &bios);
1601 background_work_end(cache);
1604 static void invalidate_completed(struct work_struct *ws)
1606 struct dm_cache_migration *mg = ws_to_mg(ws);
1607 invalidate_complete(mg, !mg->k.input);
1610 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1612 int r = policy_invalidate_mapping(cache->policy, cblock);
1614 r = dm_cache_remove_mapping(cache->cmd, cblock);
1616 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1617 cache_device_name(cache));
1618 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1621 } else if (r == -ENODATA) {
1623 * Harmless, already unmapped.
1628 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1633 static void invalidate_remove(struct work_struct *ws)
1636 struct dm_cache_migration *mg = ws_to_mg(ws);
1637 struct cache *cache = mg->cache;
1639 r = invalidate_cblock(cache, mg->invalidate_cblock);
1641 invalidate_complete(mg, false);
1645 init_continuation(&mg->k, invalidate_completed);
1646 continue_after_commit(&cache->committer, &mg->k);
1647 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1648 mg->overwrite_bio = NULL;
1649 schedule_commit(&cache->committer);
1652 static int invalidate_lock(struct dm_cache_migration *mg)
1655 struct dm_cell_key_v2 key;
1656 struct cache *cache = mg->cache;
1657 struct dm_bio_prison_cell_v2 *prealloc;
1659 prealloc = alloc_prison_cell(cache);
1661 invalidate_complete(mg, false);
1665 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1666 r = dm_cell_lock_v2(cache->prison, &key,
1667 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1669 free_prison_cell(cache, prealloc);
1670 invalidate_complete(mg, false);
1674 if (mg->cell != prealloc)
1675 free_prison_cell(cache, prealloc);
1678 quiesce(mg, invalidate_remove);
1682 * We can't call invalidate_remove() directly here because we
1683 * might still be in request context.
1685 init_continuation(&mg->k, invalidate_remove);
1686 queue_work(cache->wq, &mg->k.ws);
1692 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1693 dm_oblock_t oblock, struct bio *bio)
1695 struct dm_cache_migration *mg;
1697 if (!background_work_begin(cache))
1700 mg = alloc_migration(cache);
1702 background_work_end(cache);
1706 memset(mg, 0, sizeof(*mg));
1709 mg->overwrite_bio = bio;
1710 mg->invalidate_cblock = cblock;
1711 mg->invalidate_oblock = oblock;
1713 return invalidate_lock(mg);
1716 /*----------------------------------------------------------------
1718 *--------------------------------------------------------------*/
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 (idle && current_volume <= cache->migration_threshold)
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 blk_status_t commit_op(void *context)
1913 struct cache *cache = context;
1915 if (dm_cache_changed_this_transaction(cache->cmd))
1916 return errno_to_blk_status(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_status = BLK_STS_DM_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);
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->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,
2825 blk_status_t *error)
2827 struct cache *cache = ti->private;
2828 unsigned long flags;
2829 size_t pb_data_size = get_per_bio_data_size(cache);
2830 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2833 policy_tick(cache->policy, false);
2835 spin_lock_irqsave(&cache->lock, flags);
2836 cache->need_tick_bio = true;
2837 spin_unlock_irqrestore(&cache->lock, flags);
2840 bio_drop_shared_lock(cache, bio);
2841 accounted_complete(cache, bio);
2843 return DM_ENDIO_DONE;
2846 static int write_dirty_bitset(struct cache *cache)
2850 if (get_cache_mode(cache) >= CM_READ_ONLY)
2853 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2855 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2860 static int write_discard_bitset(struct cache *cache)
2864 if (get_cache_mode(cache) >= CM_READ_ONLY)
2867 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2868 cache->discard_nr_blocks);
2870 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2871 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2875 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2876 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2877 is_discarded(cache, to_dblock(i)));
2879 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2887 static int write_hints(struct cache *cache)
2891 if (get_cache_mode(cache) >= CM_READ_ONLY)
2894 r = dm_cache_write_hints(cache->cmd, cache->policy);
2896 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2904 * returns true on success
2906 static bool sync_metadata(struct cache *cache)
2910 r1 = write_dirty_bitset(cache);
2912 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2914 r2 = write_discard_bitset(cache);
2916 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2920 r3 = write_hints(cache);
2922 DMERR("%s: could not write hints", cache_device_name(cache));
2925 * If writing the above metadata failed, we still commit, but don't
2926 * set the clean shutdown flag. This will effectively force every
2927 * dirty bit to be set on reload.
2929 r4 = commit(cache, !r1 && !r2 && !r3);
2931 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2933 return !r1 && !r2 && !r3 && !r4;
2936 static void cache_postsuspend(struct dm_target *ti)
2938 struct cache *cache = ti->private;
2940 prevent_background_work(cache);
2941 BUG_ON(atomic_read(&cache->nr_io_migrations));
2943 cancel_delayed_work(&cache->waker);
2944 flush_workqueue(cache->wq);
2945 WARN_ON(cache->tracker.in_flight);
2948 * If it's a flush suspend there won't be any deferred bios, so this
2951 requeue_deferred_bios(cache);
2953 if (get_cache_mode(cache) == CM_WRITE)
2954 (void) sync_metadata(cache);
2957 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2958 bool dirty, uint32_t hint, bool hint_valid)
2961 struct cache *cache = context;
2964 set_bit(from_cblock(cblock), cache->dirty_bitset);
2965 atomic_inc(&cache->nr_dirty);
2967 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2969 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2977 * The discard block size in the on disk metadata is not
2978 * neccessarily the same as we're currently using. So we have to
2979 * be careful to only set the discarded attribute if we know it
2980 * covers a complete block of the new size.
2982 struct discard_load_info {
2983 struct cache *cache;
2986 * These blocks are sized using the on disk dblock size, rather
2987 * than the current one.
2989 dm_block_t block_size;
2990 dm_block_t discard_begin, discard_end;
2993 static void discard_load_info_init(struct cache *cache,
2994 struct discard_load_info *li)
2997 li->discard_begin = li->discard_end = 0;
3000 static void set_discard_range(struct discard_load_info *li)
3004 if (li->discard_begin == li->discard_end)
3008 * Convert to sectors.
3010 b = li->discard_begin * li->block_size;
3011 e = li->discard_end * li->block_size;
3014 * Then convert back to the current dblock size.
3016 b = dm_sector_div_up(b, li->cache->discard_block_size);
3017 sector_div(e, li->cache->discard_block_size);
3020 * The origin may have shrunk, so we need to check we're still in
3023 if (e > from_dblock(li->cache->discard_nr_blocks))
3024 e = from_dblock(li->cache->discard_nr_blocks);
3027 set_discard(li->cache, to_dblock(b));
3030 static int load_discard(void *context, sector_t discard_block_size,
3031 dm_dblock_t dblock, bool discard)
3033 struct discard_load_info *li = context;
3035 li->block_size = discard_block_size;
3038 if (from_dblock(dblock) == li->discard_end)
3040 * We're already in a discard range, just extend it.
3042 li->discard_end = li->discard_end + 1ULL;
3046 * Emit the old range and start a new one.
3048 set_discard_range(li);
3049 li->discard_begin = from_dblock(dblock);
3050 li->discard_end = li->discard_begin + 1ULL;
3053 set_discard_range(li);
3054 li->discard_begin = li->discard_end = 0;
3060 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3062 sector_t size = get_dev_size(cache->cache_dev);
3063 (void) sector_div(size, cache->sectors_per_block);
3064 return to_cblock(size);
3067 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3069 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3073 * We can't drop a dirty block when shrinking the cache.
3075 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3076 new_size = to_cblock(from_cblock(new_size) + 1);
3077 if (is_dirty(cache, new_size)) {
3078 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3079 cache_device_name(cache),
3080 (unsigned long long) from_cblock(new_size));
3088 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3092 r = dm_cache_resize(cache->cmd, new_size);
3094 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3095 metadata_operation_failed(cache, "dm_cache_resize", r);
3099 set_cache_size(cache, new_size);
3104 static int cache_preresume(struct dm_target *ti)
3107 struct cache *cache = ti->private;
3108 dm_cblock_t csize = get_cache_dev_size(cache);
3111 * Check to see if the cache has resized.
3113 if (!cache->sized) {
3114 r = resize_cache_dev(cache, csize);
3118 cache->sized = true;
3120 } else if (csize != cache->cache_size) {
3121 if (!can_resize(cache, csize))
3124 r = resize_cache_dev(cache, csize);
3129 if (!cache->loaded_mappings) {
3130 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3131 load_mapping, cache);
3133 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3134 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3138 cache->loaded_mappings = true;
3141 if (!cache->loaded_discards) {
3142 struct discard_load_info li;
3145 * The discard bitset could have been resized, or the
3146 * discard block size changed. To be safe we start by
3147 * setting every dblock to not discarded.
3149 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3151 discard_load_info_init(cache, &li);
3152 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3154 DMERR("%s: could not load origin discards", cache_device_name(cache));
3155 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3158 set_discard_range(&li);
3160 cache->loaded_discards = true;
3166 static void cache_resume(struct dm_target *ti)
3168 struct cache *cache = ti->private;
3170 cache->need_tick_bio = true;
3171 allow_background_work(cache);
3172 do_waker(&cache->waker.work);
3178 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3179 * <cache block size> <#used cache blocks>/<#total cache blocks>
3180 * <#read hits> <#read misses> <#write hits> <#write misses>
3181 * <#demotions> <#promotions> <#dirty>
3182 * <#features> <features>*
3183 * <#core args> <core args>
3184 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3186 static void cache_status(struct dm_target *ti, status_type_t type,
3187 unsigned status_flags, char *result, unsigned maxlen)
3192 dm_block_t nr_free_blocks_metadata = 0;
3193 dm_block_t nr_blocks_metadata = 0;
3194 char buf[BDEVNAME_SIZE];
3195 struct cache *cache = ti->private;
3196 dm_cblock_t residency;
3200 case STATUSTYPE_INFO:
3201 if (get_cache_mode(cache) == CM_FAIL) {
3206 /* Commit to ensure statistics aren't out-of-date */
3207 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3208 (void) commit(cache, false);
3210 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3212 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3213 cache_device_name(cache), r);
3217 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3219 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3220 cache_device_name(cache), r);
3224 residency = policy_residency(cache->policy);
3226 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3227 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3228 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3229 (unsigned long long)nr_blocks_metadata,
3230 (unsigned long long)cache->sectors_per_block,
3231 (unsigned long long) from_cblock(residency),
3232 (unsigned long long) from_cblock(cache->cache_size),
3233 (unsigned) atomic_read(&cache->stats.read_hit),
3234 (unsigned) atomic_read(&cache->stats.read_miss),
3235 (unsigned) atomic_read(&cache->stats.write_hit),
3236 (unsigned) atomic_read(&cache->stats.write_miss),
3237 (unsigned) atomic_read(&cache->stats.demotion),
3238 (unsigned) atomic_read(&cache->stats.promotion),
3239 (unsigned long) atomic_read(&cache->nr_dirty));
3241 if (cache->features.metadata_version == 2)
3242 DMEMIT("2 metadata2 ");
3246 if (writethrough_mode(&cache->features))
3247 DMEMIT("writethrough ");
3249 else if (passthrough_mode(&cache->features))
3250 DMEMIT("passthrough ");
3252 else if (writeback_mode(&cache->features))
3253 DMEMIT("writeback ");
3256 DMERR("%s: internal error: unknown io mode: %d",
3257 cache_device_name(cache), (int) cache->features.io_mode);
3261 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3263 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3265 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3267 DMERR("%s: policy_emit_config_values returned %d",
3268 cache_device_name(cache), r);
3271 if (get_cache_mode(cache) == CM_READ_ONLY)
3276 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3278 if (r || needs_check)
3279 DMEMIT("needs_check ");
3285 case STATUSTYPE_TABLE:
3286 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3288 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3290 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3293 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3294 DMEMIT(" %s", cache->ctr_args[i]);
3295 if (cache->nr_ctr_args)
3296 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3306 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3307 * the one-past-the-end value.
3309 struct cblock_range {
3315 * A cache block range can take two forms:
3317 * i) A single cblock, eg. '3456'
3318 * ii) A begin and end cblock with a dash between, eg. 123-234
3320 static int parse_cblock_range(struct cache *cache, const char *str,
3321 struct cblock_range *result)
3328 * Try and parse form (ii) first.
3330 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3335 result->begin = to_cblock(b);
3336 result->end = to_cblock(e);
3341 * That didn't work, try form (i).
3343 r = sscanf(str, "%llu%c", &b, &dummy);
3348 result->begin = to_cblock(b);
3349 result->end = to_cblock(from_cblock(result->begin) + 1u);
3353 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3357 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3359 uint64_t b = from_cblock(range->begin);
3360 uint64_t e = from_cblock(range->end);
3361 uint64_t n = from_cblock(cache->cache_size);
3364 DMERR("%s: begin cblock out of range: %llu >= %llu",
3365 cache_device_name(cache), b, n);
3370 DMERR("%s: end cblock out of range: %llu > %llu",
3371 cache_device_name(cache), e, n);
3376 DMERR("%s: invalid cblock range: %llu >= %llu",
3377 cache_device_name(cache), b, e);
3384 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3386 return to_cblock(from_cblock(b) + 1);
3389 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3394 * We don't need to do any locking here because we know we're in
3395 * passthrough mode. There's is potential for a race between an
3396 * invalidation triggered by an io and an invalidation message. This
3397 * is harmless, we must not worry if the policy call fails.
3399 while (range->begin != range->end) {
3400 r = invalidate_cblock(cache, range->begin);
3404 range->begin = cblock_succ(range->begin);
3407 cache->commit_requested = true;
3411 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3412 const char **cblock_ranges)
3416 struct cblock_range range;
3418 if (!passthrough_mode(&cache->features)) {
3419 DMERR("%s: cache has to be in passthrough mode for invalidation",
3420 cache_device_name(cache));
3424 for (i = 0; i < count; i++) {
3425 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3429 r = validate_cblock_range(cache, &range);
3434 * Pass begin and end origin blocks to the worker and wake it.
3436 r = request_invalidation(cache, &range);
3448 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3450 * The key migration_threshold is supported by the cache target core.
3452 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3454 struct cache *cache = ti->private;
3459 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3460 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3461 cache_device_name(cache));
3465 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3466 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3471 return set_config_value(cache, argv[0], argv[1]);
3474 static int cache_iterate_devices(struct dm_target *ti,
3475 iterate_devices_callout_fn fn, void *data)
3478 struct cache *cache = ti->private;
3480 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3482 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3487 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3490 * FIXME: these limits may be incompatible with the cache device
3492 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3493 cache->origin_sectors);
3494 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3497 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3499 struct cache *cache = ti->private;
3500 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3503 * If the system-determined stacked limits are compatible with the
3504 * cache's blocksize (io_opt is a factor) do not override them.
3506 if (io_opt_sectors < cache->sectors_per_block ||
3507 do_div(io_opt_sectors, cache->sectors_per_block)) {
3508 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3509 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3511 set_discard_limits(cache, limits);
3514 /*----------------------------------------------------------------*/
3516 static struct target_type cache_target = {
3518 .version = {2, 0, 0},
3519 .module = THIS_MODULE,
3523 .end_io = cache_end_io,
3524 .postsuspend = cache_postsuspend,
3525 .preresume = cache_preresume,
3526 .resume = cache_resume,
3527 .status = cache_status,
3528 .message = cache_message,
3529 .iterate_devices = cache_iterate_devices,
3530 .io_hints = cache_io_hints,
3533 static int __init dm_cache_init(void)
3537 r = dm_register_target(&cache_target);
3539 DMERR("cache target registration failed: %d", r);
3543 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3544 if (!migration_cache) {
3545 dm_unregister_target(&cache_target);
3552 static void __exit dm_cache_exit(void)
3554 dm_unregister_target(&cache_target);
3555 kmem_cache_destroy(migration_cache);
3558 module_init(dm_cache_init);
3559 module_exit(dm_cache_exit);
3561 MODULE_DESCRIPTION(DM_NAME " cache target");
3562 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3563 MODULE_LICENSE("GPL");