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 sector_t migration_threshold;
414 wait_queue_head_t migration_wait;
415 atomic_t nr_allocated_migrations;
418 * The number of in flight migrations that are performing
419 * background io. eg, promotion, writeback.
421 atomic_t nr_io_migrations;
423 struct rw_semaphore quiesce_lock;
426 * cache_size entries, dirty if set
429 unsigned long *dirty_bitset;
432 * origin_blocks entries, discarded if set.
434 dm_dblock_t discard_nr_blocks;
435 unsigned long *discard_bitset;
436 uint32_t discard_block_size; /* a power of 2 times sectors per block */
439 * Rather than reconstructing the table line for the status we just
440 * save it and regurgitate.
442 unsigned nr_ctr_args;
443 const char **ctr_args;
445 struct dm_kcopyd_client *copier;
446 struct workqueue_struct *wq;
447 struct work_struct deferred_bio_worker;
448 struct work_struct migration_worker;
449 struct delayed_work waker;
450 struct dm_bio_prison_v2 *prison;
453 mempool_t *migration_pool;
455 struct dm_cache_policy *policy;
456 unsigned policy_nr_args;
458 bool need_tick_bio:1;
461 bool commit_requested:1;
462 bool loaded_mappings:1;
463 bool loaded_discards:1;
466 * Cache features such as write-through.
468 struct cache_features features;
470 struct cache_stats stats;
473 * Invalidation fields.
475 spinlock_t invalidation_lock;
476 struct list_head invalidation_requests;
478 struct io_tracker tracker;
480 struct work_struct commit_ws;
481 struct batcher committer;
483 struct rw_semaphore background_work_lock;
486 struct per_bio_data {
489 struct dm_bio_prison_cell_v2 *cell;
490 struct dm_hook_info hook_info;
494 struct dm_cache_migration {
495 struct continuation k;
498 struct policy_work *op;
499 struct bio *overwrite_bio;
500 struct dm_bio_prison_cell_v2 *cell;
502 dm_cblock_t invalidate_cblock;
503 dm_oblock_t invalidate_oblock;
506 /*----------------------------------------------------------------*/
508 static bool writethrough_mode(struct cache *cache)
510 return cache->features.io_mode == CM_IO_WRITETHROUGH;
513 static bool writeback_mode(struct cache *cache)
515 return cache->features.io_mode == CM_IO_WRITEBACK;
518 static inline bool passthrough_mode(struct cache *cache)
520 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
523 /*----------------------------------------------------------------*/
525 static void wake_deferred_bio_worker(struct cache *cache)
527 queue_work(cache->wq, &cache->deferred_bio_worker);
530 static void wake_migration_worker(struct cache *cache)
532 if (passthrough_mode(cache))
535 queue_work(cache->wq, &cache->migration_worker);
538 /*----------------------------------------------------------------*/
540 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
542 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
545 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
547 dm_bio_prison_free_cell_v2(cache->prison, cell);
550 static struct dm_cache_migration *alloc_migration(struct cache *cache)
552 struct dm_cache_migration *mg;
554 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
557 atomic_inc(&mg->cache->nr_allocated_migrations);
563 static void free_migration(struct dm_cache_migration *mg)
565 struct cache *cache = mg->cache;
567 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
568 wake_up(&cache->migration_wait);
570 mempool_free(mg, cache->migration_pool);
573 /*----------------------------------------------------------------*/
575 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
577 return to_oblock(from_oblock(b) + 1ull);
580 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
584 key->block_begin = from_oblock(begin);
585 key->block_end = from_oblock(end);
589 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
590 * level 1 which prevents *both* READs and WRITEs.
592 #define WRITE_LOCK_LEVEL 0
593 #define READ_WRITE_LOCK_LEVEL 1
595 static unsigned lock_level(struct bio *bio)
597 return bio_data_dir(bio) == WRITE ?
599 READ_WRITE_LOCK_LEVEL;
602 /*----------------------------------------------------------------
604 *--------------------------------------------------------------*/
606 static size_t get_per_bio_data_size(struct cache *cache)
608 return sizeof(struct per_bio_data);
611 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
613 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
618 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
620 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
623 pb->req_nr = dm_bio_get_target_bio_nr(bio);
630 /*----------------------------------------------------------------*/
632 static void defer_bio(struct cache *cache, struct bio *bio)
636 spin_lock_irqsave(&cache->lock, flags);
637 bio_list_add(&cache->deferred_bios, bio);
638 spin_unlock_irqrestore(&cache->lock, flags);
640 wake_deferred_bio_worker(cache);
643 static void defer_bios(struct cache *cache, struct bio_list *bios)
647 spin_lock_irqsave(&cache->lock, flags);
648 bio_list_merge(&cache->deferred_bios, bios);
650 spin_unlock_irqrestore(&cache->lock, flags);
652 wake_deferred_bio_worker(cache);
655 /*----------------------------------------------------------------*/
657 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
661 struct per_bio_data *pb;
662 struct dm_cell_key_v2 key;
663 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
664 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
666 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
667 if (!cell_prealloc) {
668 defer_bio(cache, bio);
672 build_key(oblock, end, &key);
673 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
676 * Failed to get the lock.
678 free_prison_cell(cache, cell_prealloc);
682 if (cell != cell_prealloc)
683 free_prison_cell(cache, cell_prealloc);
685 pb_size = get_per_bio_data_size(cache);
686 pb = get_per_bio_data(bio, pb_size);
692 /*----------------------------------------------------------------*/
694 static bool is_dirty(struct cache *cache, dm_cblock_t b)
696 return test_bit(from_cblock(b), cache->dirty_bitset);
699 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
701 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
702 atomic_inc(&cache->nr_dirty);
703 policy_set_dirty(cache->policy, cblock);
708 * These two are called when setting after migrations to force the policy
709 * and dirty bitset to be in sync.
711 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
713 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
714 atomic_inc(&cache->nr_dirty);
715 policy_set_dirty(cache->policy, cblock);
718 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
720 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
721 if (atomic_dec_return(&cache->nr_dirty) == 0)
722 dm_table_event(cache->ti->table);
725 policy_clear_dirty(cache->policy, cblock);
728 /*----------------------------------------------------------------*/
730 static bool block_size_is_power_of_two(struct cache *cache)
732 return cache->sectors_per_block_shift >= 0;
735 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
736 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
739 static dm_block_t block_div(dm_block_t b, uint32_t n)
746 static dm_block_t oblocks_per_dblock(struct cache *cache)
748 dm_block_t oblocks = cache->discard_block_size;
750 if (block_size_is_power_of_two(cache))
751 oblocks >>= cache->sectors_per_block_shift;
753 oblocks = block_div(oblocks, cache->sectors_per_block);
758 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
760 return to_dblock(block_div(from_oblock(oblock),
761 oblocks_per_dblock(cache)));
764 static void set_discard(struct cache *cache, dm_dblock_t b)
768 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
769 atomic_inc(&cache->stats.discard_count);
771 spin_lock_irqsave(&cache->lock, flags);
772 set_bit(from_dblock(b), cache->discard_bitset);
773 spin_unlock_irqrestore(&cache->lock, flags);
776 static void clear_discard(struct cache *cache, dm_dblock_t b)
780 spin_lock_irqsave(&cache->lock, flags);
781 clear_bit(from_dblock(b), cache->discard_bitset);
782 spin_unlock_irqrestore(&cache->lock, flags);
785 static bool is_discarded(struct cache *cache, dm_dblock_t b)
790 spin_lock_irqsave(&cache->lock, flags);
791 r = test_bit(from_dblock(b), cache->discard_bitset);
792 spin_unlock_irqrestore(&cache->lock, flags);
797 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
802 spin_lock_irqsave(&cache->lock, flags);
803 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
804 cache->discard_bitset);
805 spin_unlock_irqrestore(&cache->lock, flags);
810 /*----------------------------------------------------------------
812 *--------------------------------------------------------------*/
813 static void remap_to_origin(struct cache *cache, struct bio *bio)
815 bio_set_dev(bio, cache->origin_dev->bdev);
818 static void remap_to_cache(struct cache *cache, struct bio *bio,
821 sector_t bi_sector = bio->bi_iter.bi_sector;
822 sector_t block = from_cblock(cblock);
824 bio_set_dev(bio, cache->cache_dev->bdev);
825 if (!block_size_is_power_of_two(cache))
826 bio->bi_iter.bi_sector =
827 (block * cache->sectors_per_block) +
828 sector_div(bi_sector, cache->sectors_per_block);
830 bio->bi_iter.bi_sector =
831 (block << cache->sectors_per_block_shift) |
832 (bi_sector & (cache->sectors_per_block - 1));
835 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
838 size_t pb_data_size = get_per_bio_data_size(cache);
839 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
841 spin_lock_irqsave(&cache->lock, flags);
842 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
843 bio_op(bio) != REQ_OP_DISCARD) {
845 cache->need_tick_bio = false;
847 spin_unlock_irqrestore(&cache->lock, flags);
850 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
851 dm_oblock_t oblock, bool bio_has_pbd)
854 check_if_tick_bio_needed(cache, bio);
855 remap_to_origin(cache, bio);
856 if (bio_data_dir(bio) == WRITE)
857 clear_discard(cache, oblock_to_dblock(cache, oblock));
860 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
863 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
864 __remap_to_origin_clear_discard(cache, bio, oblock, true);
867 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
868 dm_oblock_t oblock, dm_cblock_t cblock)
870 check_if_tick_bio_needed(cache, bio);
871 remap_to_cache(cache, bio, cblock);
872 if (bio_data_dir(bio) == WRITE) {
873 set_dirty(cache, cblock);
874 clear_discard(cache, oblock_to_dblock(cache, oblock));
878 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
880 sector_t block_nr = bio->bi_iter.bi_sector;
882 if (!block_size_is_power_of_two(cache))
883 (void) sector_div(block_nr, cache->sectors_per_block);
885 block_nr >>= cache->sectors_per_block_shift;
887 return to_oblock(block_nr);
890 static bool accountable_bio(struct cache *cache, struct bio *bio)
892 return bio_op(bio) != REQ_OP_DISCARD;
895 static void accounted_begin(struct cache *cache, struct bio *bio)
897 size_t pb_data_size = get_per_bio_data_size(cache);
898 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
900 if (accountable_bio(cache, bio)) {
901 pb->len = bio_sectors(bio);
902 iot_io_begin(&cache->tracker, pb->len);
906 static void accounted_complete(struct cache *cache, struct bio *bio)
908 size_t pb_data_size = get_per_bio_data_size(cache);
909 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
911 iot_io_end(&cache->tracker, pb->len);
914 static void accounted_request(struct cache *cache, struct bio *bio)
916 accounted_begin(cache, bio);
917 generic_make_request(bio);
920 static void issue_op(struct bio *bio, void *context)
922 struct cache *cache = context;
923 accounted_request(cache, bio);
927 * When running in writethrough mode we need to send writes to clean blocks
928 * to both the cache and origin devices. Clone the bio and send them in parallel.
930 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
931 dm_oblock_t oblock, dm_cblock_t cblock)
933 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, cache->bs);
937 bio_chain(origin_bio, bio);
939 * Passing false to __remap_to_origin_clear_discard() skips
940 * all code that might use per_bio_data (since clone doesn't have it)
942 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
943 submit_bio(origin_bio);
945 remap_to_cache(cache, bio, cblock);
948 /*----------------------------------------------------------------
950 *--------------------------------------------------------------*/
951 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
953 return cache->features.mode;
956 static const char *cache_device_name(struct cache *cache)
958 return dm_device_name(dm_table_get_md(cache->ti->table));
961 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
963 const char *descs[] = {
969 dm_table_event(cache->ti->table);
970 DMINFO("%s: switching cache to %s mode",
971 cache_device_name(cache), descs[(int)mode]);
974 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
977 enum cache_metadata_mode old_mode = get_cache_mode(cache);
979 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
980 DMERR("%s: unable to read needs_check flag, setting failure mode.",
981 cache_device_name(cache));
985 if (new_mode == CM_WRITE && needs_check) {
986 DMERR("%s: unable to switch cache to write mode until repaired.",
987 cache_device_name(cache));
988 if (old_mode != new_mode)
991 new_mode = CM_READ_ONLY;
994 /* Never move out of fail mode */
995 if (old_mode == CM_FAIL)
1001 dm_cache_metadata_set_read_only(cache->cmd);
1005 dm_cache_metadata_set_read_write(cache->cmd);
1009 cache->features.mode = new_mode;
1011 if (new_mode != old_mode)
1012 notify_mode_switch(cache, new_mode);
1015 static void abort_transaction(struct cache *cache)
1017 const char *dev_name = cache_device_name(cache);
1019 if (get_cache_mode(cache) >= CM_READ_ONLY)
1022 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1023 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1024 set_cache_mode(cache, CM_FAIL);
1027 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1028 if (dm_cache_metadata_abort(cache->cmd)) {
1029 DMERR("%s: failed to abort metadata transaction", dev_name);
1030 set_cache_mode(cache, CM_FAIL);
1034 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1036 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1037 cache_device_name(cache), op, r);
1038 abort_transaction(cache);
1039 set_cache_mode(cache, CM_READ_ONLY);
1042 /*----------------------------------------------------------------*/
1044 static void load_stats(struct cache *cache)
1046 struct dm_cache_statistics stats;
1048 dm_cache_metadata_get_stats(cache->cmd, &stats);
1049 atomic_set(&cache->stats.read_hit, stats.read_hits);
1050 atomic_set(&cache->stats.read_miss, stats.read_misses);
1051 atomic_set(&cache->stats.write_hit, stats.write_hits);
1052 atomic_set(&cache->stats.write_miss, stats.write_misses);
1055 static void save_stats(struct cache *cache)
1057 struct dm_cache_statistics stats;
1059 if (get_cache_mode(cache) >= CM_READ_ONLY)
1062 stats.read_hits = atomic_read(&cache->stats.read_hit);
1063 stats.read_misses = atomic_read(&cache->stats.read_miss);
1064 stats.write_hits = atomic_read(&cache->stats.write_hit);
1065 stats.write_misses = atomic_read(&cache->stats.write_miss);
1067 dm_cache_metadata_set_stats(cache->cmd, &stats);
1070 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1073 case POLICY_PROMOTE:
1074 atomic_inc(&stats->promotion);
1078 atomic_inc(&stats->demotion);
1081 case POLICY_WRITEBACK:
1082 atomic_inc(&stats->writeback);
1087 /*----------------------------------------------------------------
1088 * Migration processing
1090 * Migration covers moving data from the origin device to the cache, or
1092 *--------------------------------------------------------------*/
1094 static void inc_io_migrations(struct cache *cache)
1096 atomic_inc(&cache->nr_io_migrations);
1099 static void dec_io_migrations(struct cache *cache)
1101 atomic_dec(&cache->nr_io_migrations);
1104 static bool discard_or_flush(struct bio *bio)
1106 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1109 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1110 dm_dblock_t *b, dm_dblock_t *e)
1112 sector_t sb = bio->bi_iter.bi_sector;
1113 sector_t se = bio_end_sector(bio);
1115 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1117 if (se - sb < cache->discard_block_size)
1120 *e = to_dblock(block_div(se, cache->discard_block_size));
1123 /*----------------------------------------------------------------*/
1125 static void prevent_background_work(struct cache *cache)
1128 down_write(&cache->background_work_lock);
1132 static void allow_background_work(struct cache *cache)
1135 up_write(&cache->background_work_lock);
1139 static bool background_work_begin(struct cache *cache)
1144 r = down_read_trylock(&cache->background_work_lock);
1150 static void background_work_end(struct cache *cache)
1153 up_read(&cache->background_work_lock);
1157 /*----------------------------------------------------------------*/
1159 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1161 return (bio_data_dir(bio) == WRITE) &&
1162 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1165 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1167 return writeback_mode(cache) &&
1168 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1171 static void quiesce(struct dm_cache_migration *mg,
1172 void (*continuation)(struct work_struct *))
1174 init_continuation(&mg->k, continuation);
1175 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1178 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1180 struct continuation *k = container_of(ws, struct continuation, ws);
1181 return container_of(k, struct dm_cache_migration, k);
1184 static void copy_complete(int read_err, unsigned long write_err, void *context)
1186 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1188 if (read_err || write_err)
1189 mg->k.input = BLK_STS_IOERR;
1191 queue_continuation(mg->cache->wq, &mg->k);
1194 static int copy(struct dm_cache_migration *mg, bool promote)
1197 struct dm_io_region o_region, c_region;
1198 struct cache *cache = mg->cache;
1200 o_region.bdev = cache->origin_dev->bdev;
1201 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1202 o_region.count = cache->sectors_per_block;
1204 c_region.bdev = cache->cache_dev->bdev;
1205 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1206 c_region.count = cache->sectors_per_block;
1209 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1211 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1216 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1218 size_t pb_data_size = get_per_bio_data_size(cache);
1219 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1221 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1222 free_prison_cell(cache, pb->cell);
1226 static void overwrite_endio(struct bio *bio)
1228 struct dm_cache_migration *mg = bio->bi_private;
1229 struct cache *cache = mg->cache;
1230 size_t pb_data_size = get_per_bio_data_size(cache);
1231 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1233 dm_unhook_bio(&pb->hook_info, bio);
1236 mg->k.input = bio->bi_status;
1238 queue_continuation(mg->cache->wq, &mg->k);
1241 static void overwrite(struct dm_cache_migration *mg,
1242 void (*continuation)(struct work_struct *))
1244 struct bio *bio = mg->overwrite_bio;
1245 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1246 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1248 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1251 * The overwrite bio is part of the copy operation, as such it does
1252 * not set/clear discard or dirty flags.
1254 if (mg->op->op == POLICY_PROMOTE)
1255 remap_to_cache(mg->cache, bio, mg->op->cblock);
1257 remap_to_origin(mg->cache, bio);
1259 init_continuation(&mg->k, continuation);
1260 accounted_request(mg->cache, bio);
1266 * 1) exclusive lock preventing WRITEs
1268 * 3) copy or issue overwrite bio
1269 * 4) upgrade to exclusive lock preventing READs and WRITEs
1271 * 6) update metadata and commit
1274 static void mg_complete(struct dm_cache_migration *mg, bool success)
1276 struct bio_list bios;
1277 struct cache *cache = mg->cache;
1278 struct policy_work *op = mg->op;
1279 dm_cblock_t cblock = op->cblock;
1282 update_stats(&cache->stats, op->op);
1285 case POLICY_PROMOTE:
1286 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1287 policy_complete_background_work(cache->policy, op, success);
1289 if (mg->overwrite_bio) {
1291 force_set_dirty(cache, cblock);
1292 else if (mg->k.input)
1293 mg->overwrite_bio->bi_status = mg->k.input;
1295 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1296 bio_endio(mg->overwrite_bio);
1299 force_clear_dirty(cache, cblock);
1300 dec_io_migrations(cache);
1306 * We clear dirty here to update the nr_dirty counter.
1309 force_clear_dirty(cache, cblock);
1310 policy_complete_background_work(cache->policy, op, success);
1311 dec_io_migrations(cache);
1314 case POLICY_WRITEBACK:
1316 force_clear_dirty(cache, cblock);
1317 policy_complete_background_work(cache->policy, op, success);
1318 dec_io_migrations(cache);
1322 bio_list_init(&bios);
1324 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1325 free_prison_cell(cache, mg->cell);
1329 defer_bios(cache, &bios);
1330 wake_migration_worker(cache);
1332 background_work_end(cache);
1335 static void mg_success(struct work_struct *ws)
1337 struct dm_cache_migration *mg = ws_to_mg(ws);
1338 mg_complete(mg, mg->k.input == 0);
1341 static void mg_update_metadata(struct work_struct *ws)
1344 struct dm_cache_migration *mg = ws_to_mg(ws);
1345 struct cache *cache = mg->cache;
1346 struct policy_work *op = mg->op;
1349 case POLICY_PROMOTE:
1350 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1352 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1353 cache_device_name(cache));
1354 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1356 mg_complete(mg, false);
1359 mg_complete(mg, true);
1363 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1365 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1366 cache_device_name(cache));
1367 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1369 mg_complete(mg, false);
1374 * It would be nice if we only had to commit when a REQ_FLUSH
1375 * comes through. But there's one scenario that we have to
1378 * - vblock x in a cache block
1380 * - cache block gets reallocated and over written
1383 * When we recover, because there was no commit the cache will
1384 * rollback to having the data for vblock x in the cache block.
1385 * But the cache block has since been overwritten, so it'll end
1386 * up pointing to data that was never in 'x' during the history
1389 * To avoid this issue we require a commit as part of the
1390 * demotion operation.
1392 init_continuation(&mg->k, mg_success);
1393 continue_after_commit(&cache->committer, &mg->k);
1394 schedule_commit(&cache->committer);
1397 case POLICY_WRITEBACK:
1398 mg_complete(mg, true);
1403 static void mg_update_metadata_after_copy(struct work_struct *ws)
1405 struct dm_cache_migration *mg = ws_to_mg(ws);
1408 * Did the copy succeed?
1411 mg_complete(mg, false);
1413 mg_update_metadata(ws);
1416 static void mg_upgrade_lock(struct work_struct *ws)
1419 struct dm_cache_migration *mg = ws_to_mg(ws);
1422 * Did the copy succeed?
1425 mg_complete(mg, false);
1429 * Now we want the lock to prevent both reads and writes.
1431 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1432 READ_WRITE_LOCK_LEVEL);
1434 mg_complete(mg, false);
1437 quiesce(mg, mg_update_metadata);
1440 mg_update_metadata(ws);
1444 static void mg_full_copy(struct work_struct *ws)
1446 struct dm_cache_migration *mg = ws_to_mg(ws);
1447 struct cache *cache = mg->cache;
1448 struct policy_work *op = mg->op;
1449 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1451 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1452 is_discarded_oblock(cache, op->oblock)) {
1453 mg_upgrade_lock(ws);
1457 init_continuation(&mg->k, mg_upgrade_lock);
1459 if (copy(mg, is_policy_promote)) {
1460 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1461 mg->k.input = BLK_STS_IOERR;
1462 mg_complete(mg, false);
1466 static void mg_copy(struct work_struct *ws)
1468 struct dm_cache_migration *mg = ws_to_mg(ws);
1470 if (mg->overwrite_bio) {
1472 * No exclusive lock was held when we last checked if the bio
1473 * was optimisable. So we have to check again in case things
1474 * have changed (eg, the block may no longer be discarded).
1476 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1478 * Fallback to a real full copy after doing some tidying up.
1480 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1481 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1482 mg->overwrite_bio = NULL;
1483 inc_io_migrations(mg->cache);
1489 * It's safe to do this here, even though it's new data
1490 * because all IO has been locked out of the block.
1492 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1493 * so _not_ using mg_upgrade_lock() as continutation.
1495 overwrite(mg, mg_update_metadata_after_copy);
1501 static int mg_lock_writes(struct dm_cache_migration *mg)
1504 struct dm_cell_key_v2 key;
1505 struct cache *cache = mg->cache;
1506 struct dm_bio_prison_cell_v2 *prealloc;
1508 prealloc = alloc_prison_cell(cache);
1510 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1511 mg_complete(mg, false);
1516 * Prevent writes to the block, but allow reads to continue.
1517 * Unless we're using an overwrite bio, in which case we lock
1520 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1521 r = dm_cell_lock_v2(cache->prison, &key,
1522 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1523 prealloc, &mg->cell);
1525 free_prison_cell(cache, prealloc);
1526 mg_complete(mg, false);
1530 if (mg->cell != prealloc)
1531 free_prison_cell(cache, prealloc);
1536 quiesce(mg, mg_copy);
1541 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1543 struct dm_cache_migration *mg;
1545 if (!background_work_begin(cache)) {
1546 policy_complete_background_work(cache->policy, op, false);
1550 mg = alloc_migration(cache);
1552 policy_complete_background_work(cache->policy, op, false);
1553 background_work_end(cache);
1557 memset(mg, 0, sizeof(*mg));
1561 mg->overwrite_bio = bio;
1564 inc_io_migrations(cache);
1566 return mg_lock_writes(mg);
1569 /*----------------------------------------------------------------
1570 * invalidation processing
1571 *--------------------------------------------------------------*/
1573 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1575 struct bio_list bios;
1576 struct cache *cache = mg->cache;
1578 bio_list_init(&bios);
1579 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1580 free_prison_cell(cache, mg->cell);
1582 if (!success && mg->overwrite_bio)
1583 bio_io_error(mg->overwrite_bio);
1586 defer_bios(cache, &bios);
1588 background_work_end(cache);
1591 static void invalidate_completed(struct work_struct *ws)
1593 struct dm_cache_migration *mg = ws_to_mg(ws);
1594 invalidate_complete(mg, !mg->k.input);
1597 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1599 int r = policy_invalidate_mapping(cache->policy, cblock);
1601 r = dm_cache_remove_mapping(cache->cmd, cblock);
1603 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1604 cache_device_name(cache));
1605 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1608 } else if (r == -ENODATA) {
1610 * Harmless, already unmapped.
1615 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1620 static void invalidate_remove(struct work_struct *ws)
1623 struct dm_cache_migration *mg = ws_to_mg(ws);
1624 struct cache *cache = mg->cache;
1626 r = invalidate_cblock(cache, mg->invalidate_cblock);
1628 invalidate_complete(mg, false);
1632 init_continuation(&mg->k, invalidate_completed);
1633 continue_after_commit(&cache->committer, &mg->k);
1634 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1635 mg->overwrite_bio = NULL;
1636 schedule_commit(&cache->committer);
1639 static int invalidate_lock(struct dm_cache_migration *mg)
1642 struct dm_cell_key_v2 key;
1643 struct cache *cache = mg->cache;
1644 struct dm_bio_prison_cell_v2 *prealloc;
1646 prealloc = alloc_prison_cell(cache);
1648 invalidate_complete(mg, false);
1652 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1653 r = dm_cell_lock_v2(cache->prison, &key,
1654 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1656 free_prison_cell(cache, prealloc);
1657 invalidate_complete(mg, false);
1661 if (mg->cell != prealloc)
1662 free_prison_cell(cache, prealloc);
1665 quiesce(mg, invalidate_remove);
1669 * We can't call invalidate_remove() directly here because we
1670 * might still be in request context.
1672 init_continuation(&mg->k, invalidate_remove);
1673 queue_work(cache->wq, &mg->k.ws);
1679 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1680 dm_oblock_t oblock, struct bio *bio)
1682 struct dm_cache_migration *mg;
1684 if (!background_work_begin(cache))
1687 mg = alloc_migration(cache);
1689 background_work_end(cache);
1693 memset(mg, 0, sizeof(*mg));
1696 mg->overwrite_bio = bio;
1697 mg->invalidate_cblock = cblock;
1698 mg->invalidate_oblock = oblock;
1700 return invalidate_lock(mg);
1703 /*----------------------------------------------------------------
1705 *--------------------------------------------------------------*/
1712 static enum busy spare_migration_bandwidth(struct cache *cache)
1714 bool idle = iot_idle_for(&cache->tracker, HZ);
1715 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1716 cache->sectors_per_block;
1718 if (idle && current_volume <= cache->migration_threshold)
1724 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1726 atomic_inc(bio_data_dir(bio) == READ ?
1727 &cache->stats.read_hit : &cache->stats.write_hit);
1730 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1732 atomic_inc(bio_data_dir(bio) == READ ?
1733 &cache->stats.read_miss : &cache->stats.write_miss);
1736 /*----------------------------------------------------------------*/
1738 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1739 bool *commit_needed)
1742 bool rb, background_queued;
1744 size_t pb_data_size = get_per_bio_data_size(cache);
1745 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1747 *commit_needed = false;
1749 rb = bio_detain_shared(cache, block, bio);
1752 * An exclusive lock is held for this block, so we have to
1753 * wait. We set the commit_needed flag so the current
1754 * transaction will be committed asap, allowing this lock
1757 *commit_needed = true;
1758 return DM_MAPIO_SUBMITTED;
1761 data_dir = bio_data_dir(bio);
1763 if (optimisable_bio(cache, bio, block)) {
1764 struct policy_work *op = NULL;
1766 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1767 if (unlikely(r && r != -ENOENT)) {
1768 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1769 cache_device_name(cache), r);
1771 return DM_MAPIO_SUBMITTED;
1774 if (r == -ENOENT && op) {
1775 bio_drop_shared_lock(cache, bio);
1776 BUG_ON(op->op != POLICY_PROMOTE);
1777 mg_start(cache, op, bio);
1778 return DM_MAPIO_SUBMITTED;
1781 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1782 if (unlikely(r && r != -ENOENT)) {
1783 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1784 cache_device_name(cache), r);
1786 return DM_MAPIO_SUBMITTED;
1789 if (background_queued)
1790 wake_migration_worker(cache);
1797 inc_miss_counter(cache, bio);
1798 if (pb->req_nr == 0) {
1799 accounted_begin(cache, bio);
1800 remap_to_origin_clear_discard(cache, bio, block);
1804 * This is a duplicate writethrough io that is no
1805 * longer needed because the block has been demoted.
1808 return DM_MAPIO_SUBMITTED;
1814 inc_hit_counter(cache, bio);
1817 * Passthrough always maps to the origin, invalidating any
1818 * cache blocks that are written to.
1820 if (passthrough_mode(cache)) {
1821 if (bio_data_dir(bio) == WRITE) {
1822 bio_drop_shared_lock(cache, bio);
1823 atomic_inc(&cache->stats.demotion);
1824 invalidate_start(cache, cblock, block, bio);
1826 remap_to_origin_clear_discard(cache, bio, block);
1829 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1830 !is_dirty(cache, cblock)) {
1831 remap_to_origin_and_cache(cache, bio, block, cblock);
1832 accounted_begin(cache, bio);
1834 remap_to_cache_dirty(cache, bio, block, cblock);
1839 * dm core turns FUA requests into a separate payload and FLUSH req.
1841 if (bio->bi_opf & REQ_FUA) {
1843 * issue_after_commit will call accounted_begin a second time. So
1844 * we call accounted_complete() to avoid double accounting.
1846 accounted_complete(cache, bio);
1847 issue_after_commit(&cache->committer, bio);
1848 *commit_needed = true;
1849 return DM_MAPIO_SUBMITTED;
1852 return DM_MAPIO_REMAPPED;
1855 static bool process_bio(struct cache *cache, struct bio *bio)
1859 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1860 generic_make_request(bio);
1862 return commit_needed;
1866 * A non-zero return indicates read_only or fail_io mode.
1868 static int commit(struct cache *cache, bool clean_shutdown)
1872 if (get_cache_mode(cache) >= CM_READ_ONLY)
1875 atomic_inc(&cache->stats.commit_count);
1876 r = dm_cache_commit(cache->cmd, clean_shutdown);
1878 metadata_operation_failed(cache, "dm_cache_commit", r);
1884 * Used by the batcher.
1886 static blk_status_t commit_op(void *context)
1888 struct cache *cache = context;
1890 if (dm_cache_changed_this_transaction(cache->cmd))
1891 return errno_to_blk_status(commit(cache, false));
1896 /*----------------------------------------------------------------*/
1898 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1900 size_t pb_data_size = get_per_bio_data_size(cache);
1901 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1904 remap_to_origin(cache, bio);
1906 remap_to_cache(cache, bio, 0);
1908 issue_after_commit(&cache->committer, bio);
1912 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1916 // FIXME: do we need to lock the region? Or can we just assume the
1917 // user wont be so foolish as to issue discard concurrently with
1919 calc_discard_block_range(cache, bio, &b, &e);
1921 set_discard(cache, b);
1922 b = to_dblock(from_dblock(b) + 1);
1930 static void process_deferred_bios(struct work_struct *ws)
1932 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1934 unsigned long flags;
1935 bool commit_needed = false;
1936 struct bio_list bios;
1939 bio_list_init(&bios);
1941 spin_lock_irqsave(&cache->lock, flags);
1942 bio_list_merge(&bios, &cache->deferred_bios);
1943 bio_list_init(&cache->deferred_bios);
1944 spin_unlock_irqrestore(&cache->lock, flags);
1946 while ((bio = bio_list_pop(&bios))) {
1947 if (bio->bi_opf & REQ_PREFLUSH)
1948 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1950 else if (bio_op(bio) == REQ_OP_DISCARD)
1951 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1954 commit_needed = process_bio(cache, bio) || commit_needed;
1958 schedule_commit(&cache->committer);
1961 /*----------------------------------------------------------------
1963 *--------------------------------------------------------------*/
1965 static void requeue_deferred_bios(struct cache *cache)
1968 struct bio_list bios;
1970 bio_list_init(&bios);
1971 bio_list_merge(&bios, &cache->deferred_bios);
1972 bio_list_init(&cache->deferred_bios);
1974 while ((bio = bio_list_pop(&bios))) {
1975 bio->bi_status = BLK_STS_DM_REQUEUE;
1981 * We want to commit periodically so that not too much
1982 * unwritten metadata builds up.
1984 static void do_waker(struct work_struct *ws)
1986 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1988 policy_tick(cache->policy, true);
1989 wake_migration_worker(cache);
1990 schedule_commit(&cache->committer);
1991 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1994 static void check_migrations(struct work_struct *ws)
1997 struct policy_work *op;
1998 struct cache *cache = container_of(ws, struct cache, migration_worker);
2002 b = spare_migration_bandwidth(cache);
2004 r = policy_get_background_work(cache->policy, b == IDLE, &op);
2009 DMERR_LIMIT("%s: policy_background_work failed",
2010 cache_device_name(cache));
2014 r = mg_start(cache, op, NULL);
2020 /*----------------------------------------------------------------
2022 *--------------------------------------------------------------*/
2025 * This function gets called on the error paths of the constructor, so we
2026 * have to cope with a partially initialised struct.
2028 static void destroy(struct cache *cache)
2032 mempool_destroy(cache->migration_pool);
2035 dm_bio_prison_destroy_v2(cache->prison);
2038 destroy_workqueue(cache->wq);
2040 if (cache->dirty_bitset)
2041 free_bitset(cache->dirty_bitset);
2043 if (cache->discard_bitset)
2044 free_bitset(cache->discard_bitset);
2047 dm_kcopyd_client_destroy(cache->copier);
2050 dm_cache_metadata_close(cache->cmd);
2052 if (cache->metadata_dev)
2053 dm_put_device(cache->ti, cache->metadata_dev);
2055 if (cache->origin_dev)
2056 dm_put_device(cache->ti, cache->origin_dev);
2058 if (cache->cache_dev)
2059 dm_put_device(cache->ti, cache->cache_dev);
2062 dm_cache_policy_destroy(cache->policy);
2064 for (i = 0; i < cache->nr_ctr_args ; i++)
2065 kfree(cache->ctr_args[i]);
2066 kfree(cache->ctr_args);
2069 bioset_free(cache->bs);
2074 static void cache_dtr(struct dm_target *ti)
2076 struct cache *cache = ti->private;
2081 static sector_t get_dev_size(struct dm_dev *dev)
2083 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2086 /*----------------------------------------------------------------*/
2089 * Construct a cache device mapping.
2091 * cache <metadata dev> <cache dev> <origin dev> <block size>
2092 * <#feature args> [<feature arg>]*
2093 * <policy> <#policy args> [<policy arg>]*
2095 * metadata dev : fast device holding the persistent metadata
2096 * cache dev : fast device holding cached data blocks
2097 * origin dev : slow device holding original data blocks
2098 * block size : cache unit size in sectors
2100 * #feature args : number of feature arguments passed
2101 * feature args : writethrough. (The default is writeback.)
2103 * policy : the replacement policy to use
2104 * #policy args : an even number of policy arguments corresponding
2105 * to key/value pairs passed to the policy
2106 * policy args : key/value pairs passed to the policy
2107 * E.g. 'sequential_threshold 1024'
2108 * See cache-policies.txt for details.
2110 * Optional feature arguments are:
2111 * writethrough : write through caching that prohibits cache block
2112 * content from being different from origin block content.
2113 * Without this argument, the default behaviour is to write
2114 * back cache block contents later for performance reasons,
2115 * so they may differ from the corresponding origin blocks.
2118 struct dm_target *ti;
2120 struct dm_dev *metadata_dev;
2122 struct dm_dev *cache_dev;
2123 sector_t cache_sectors;
2125 struct dm_dev *origin_dev;
2126 sector_t origin_sectors;
2128 uint32_t block_size;
2130 const char *policy_name;
2132 const char **policy_argv;
2134 struct cache_features features;
2137 static void destroy_cache_args(struct cache_args *ca)
2139 if (ca->metadata_dev)
2140 dm_put_device(ca->ti, ca->metadata_dev);
2143 dm_put_device(ca->ti, ca->cache_dev);
2146 dm_put_device(ca->ti, ca->origin_dev);
2151 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2154 *error = "Insufficient args";
2161 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2165 sector_t metadata_dev_size;
2166 char b[BDEVNAME_SIZE];
2168 if (!at_least_one_arg(as, error))
2171 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2174 *error = "Error opening metadata device";
2178 metadata_dev_size = get_dev_size(ca->metadata_dev);
2179 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2180 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2181 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2186 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2191 if (!at_least_one_arg(as, error))
2194 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2197 *error = "Error opening cache device";
2200 ca->cache_sectors = get_dev_size(ca->cache_dev);
2205 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2210 if (!at_least_one_arg(as, error))
2213 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2216 *error = "Error opening origin device";
2220 ca->origin_sectors = get_dev_size(ca->origin_dev);
2221 if (ca->ti->len > ca->origin_sectors) {
2222 *error = "Device size larger than cached device";
2229 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2232 unsigned long block_size;
2234 if (!at_least_one_arg(as, error))
2237 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2238 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2239 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2240 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2241 *error = "Invalid data block size";
2245 if (block_size > ca->cache_sectors) {
2246 *error = "Data block size is larger than the cache device";
2250 ca->block_size = block_size;
2255 static void init_features(struct cache_features *cf)
2257 cf->mode = CM_WRITE;
2258 cf->io_mode = CM_IO_WRITEBACK;
2259 cf->metadata_version = 1;
2262 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2265 static const struct dm_arg _args[] = {
2266 {0, 2, "Invalid number of cache feature arguments"},
2269 int r, mode_ctr = 0;
2272 struct cache_features *cf = &ca->features;
2276 r = dm_read_arg_group(_args, as, &argc, error);
2281 arg = dm_shift_arg(as);
2283 if (!strcasecmp(arg, "writeback")) {
2284 cf->io_mode = CM_IO_WRITEBACK;
2288 else if (!strcasecmp(arg, "writethrough")) {
2289 cf->io_mode = CM_IO_WRITETHROUGH;
2293 else if (!strcasecmp(arg, "passthrough")) {
2294 cf->io_mode = CM_IO_PASSTHROUGH;
2298 else if (!strcasecmp(arg, "metadata2"))
2299 cf->metadata_version = 2;
2302 *error = "Unrecognised cache feature requested";
2308 *error = "Duplicate cache io_mode features requested";
2315 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2318 static const struct dm_arg _args[] = {
2319 {0, 1024, "Invalid number of policy arguments"},
2324 if (!at_least_one_arg(as, error))
2327 ca->policy_name = dm_shift_arg(as);
2329 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2333 ca->policy_argv = (const char **)as->argv;
2334 dm_consume_args(as, ca->policy_argc);
2339 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2343 struct dm_arg_set as;
2348 r = parse_metadata_dev(ca, &as, error);
2352 r = parse_cache_dev(ca, &as, error);
2356 r = parse_origin_dev(ca, &as, error);
2360 r = parse_block_size(ca, &as, error);
2364 r = parse_features(ca, &as, error);
2368 r = parse_policy(ca, &as, error);
2375 /*----------------------------------------------------------------*/
2377 static struct kmem_cache *migration_cache;
2379 #define NOT_CORE_OPTION 1
2381 static int process_config_option(struct cache *cache, const char *key, const char *value)
2385 if (!strcasecmp(key, "migration_threshold")) {
2386 if (kstrtoul(value, 10, &tmp))
2389 cache->migration_threshold = tmp;
2393 return NOT_CORE_OPTION;
2396 static int set_config_value(struct cache *cache, const char *key, const char *value)
2398 int r = process_config_option(cache, key, value);
2400 if (r == NOT_CORE_OPTION)
2401 r = policy_set_config_value(cache->policy, key, value);
2404 DMWARN("bad config value for %s: %s", key, value);
2409 static int set_config_values(struct cache *cache, int argc, const char **argv)
2414 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2419 r = set_config_value(cache, argv[0], argv[1]);
2430 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2433 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2435 cache->origin_sectors,
2436 cache->sectors_per_block);
2438 *error = "Error creating cache's policy";
2442 BUG_ON(!cache->policy);
2448 * We want the discard block size to be at least the size of the cache
2449 * block size and have no more than 2^14 discard blocks across the origin.
2451 #define MAX_DISCARD_BLOCKS (1 << 14)
2453 static bool too_many_discard_blocks(sector_t discard_block_size,
2454 sector_t origin_size)
2456 (void) sector_div(origin_size, discard_block_size);
2458 return origin_size > MAX_DISCARD_BLOCKS;
2461 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2462 sector_t origin_size)
2464 sector_t discard_block_size = cache_block_size;
2467 while (too_many_discard_blocks(discard_block_size, origin_size))
2468 discard_block_size *= 2;
2470 return discard_block_size;
2473 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2475 dm_block_t nr_blocks = from_cblock(size);
2477 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2478 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2479 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2480 "Please consider increasing the cache block size to reduce the overall cache block count.",
2481 (unsigned long long) nr_blocks);
2483 cache->cache_size = size;
2486 static int is_congested(struct dm_dev *dev, int bdi_bits)
2488 struct request_queue *q = bdev_get_queue(dev->bdev);
2489 return bdi_congested(q->backing_dev_info, bdi_bits);
2492 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2494 struct cache *cache = container_of(cb, struct cache, callbacks);
2496 return is_congested(cache->origin_dev, bdi_bits) ||
2497 is_congested(cache->cache_dev, bdi_bits);
2500 #define DEFAULT_MIGRATION_THRESHOLD 2048
2502 static int cache_create(struct cache_args *ca, struct cache **result)
2505 char **error = &ca->ti->error;
2506 struct cache *cache;
2507 struct dm_target *ti = ca->ti;
2508 dm_block_t origin_blocks;
2509 struct dm_cache_metadata *cmd;
2510 bool may_format = ca->features.mode == CM_WRITE;
2512 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2517 ti->private = cache;
2518 ti->num_flush_bios = 2;
2519 ti->flush_supported = true;
2521 ti->num_discard_bios = 1;
2522 ti->discards_supported = true;
2523 ti->split_discard_bios = false;
2525 cache->features = ca->features;
2526 ti->per_io_data_size = get_per_bio_data_size(cache);
2528 if (writethrough_mode(cache)) {
2529 /* Create bioset for writethrough bios issued to origin */
2530 cache->bs = bioset_create(BIO_POOL_SIZE, 0, 0);
2535 cache->callbacks.congested_fn = cache_is_congested;
2536 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2538 cache->metadata_dev = ca->metadata_dev;
2539 cache->origin_dev = ca->origin_dev;
2540 cache->cache_dev = ca->cache_dev;
2542 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2544 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2545 origin_blocks = block_div(origin_blocks, ca->block_size);
2546 cache->origin_blocks = to_oblock(origin_blocks);
2548 cache->sectors_per_block = ca->block_size;
2549 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2554 if (ca->block_size & (ca->block_size - 1)) {
2555 dm_block_t cache_size = ca->cache_sectors;
2557 cache->sectors_per_block_shift = -1;
2558 cache_size = block_div(cache_size, ca->block_size);
2559 set_cache_size(cache, to_cblock(cache_size));
2561 cache->sectors_per_block_shift = __ffs(ca->block_size);
2562 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2565 r = create_cache_policy(cache, ca, error);
2569 cache->policy_nr_args = ca->policy_argc;
2570 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2572 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2574 *error = "Error setting cache policy's config values";
2578 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2579 ca->block_size, may_format,
2580 dm_cache_policy_get_hint_size(cache->policy),
2581 ca->features.metadata_version);
2583 *error = "Error creating metadata object";
2588 set_cache_mode(cache, CM_WRITE);
2589 if (get_cache_mode(cache) != CM_WRITE) {
2590 *error = "Unable to get write access to metadata, please check/repair metadata.";
2595 if (passthrough_mode(cache)) {
2598 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2600 *error = "dm_cache_metadata_all_clean() failed";
2605 *error = "Cannot enter passthrough mode unless all blocks are clean";
2610 policy_allow_migrations(cache->policy, false);
2613 spin_lock_init(&cache->lock);
2614 INIT_LIST_HEAD(&cache->deferred_cells);
2615 bio_list_init(&cache->deferred_bios);
2616 atomic_set(&cache->nr_allocated_migrations, 0);
2617 atomic_set(&cache->nr_io_migrations, 0);
2618 init_waitqueue_head(&cache->migration_wait);
2621 atomic_set(&cache->nr_dirty, 0);
2622 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2623 if (!cache->dirty_bitset) {
2624 *error = "could not allocate dirty bitset";
2627 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2629 cache->discard_block_size =
2630 calculate_discard_block_size(cache->sectors_per_block,
2631 cache->origin_sectors);
2632 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2633 cache->discard_block_size));
2634 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2635 if (!cache->discard_bitset) {
2636 *error = "could not allocate discard bitset";
2639 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2641 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2642 if (IS_ERR(cache->copier)) {
2643 *error = "could not create kcopyd client";
2644 r = PTR_ERR(cache->copier);
2648 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2650 *error = "could not create workqueue for metadata object";
2653 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2654 INIT_WORK(&cache->migration_worker, check_migrations);
2655 INIT_DELAYED_WORK(&cache->waker, do_waker);
2657 cache->prison = dm_bio_prison_create_v2(cache->wq);
2658 if (!cache->prison) {
2659 *error = "could not create bio prison";
2663 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2665 if (!cache->migration_pool) {
2666 *error = "Error creating cache's migration mempool";
2670 cache->need_tick_bio = true;
2671 cache->sized = false;
2672 cache->invalidate = false;
2673 cache->commit_requested = false;
2674 cache->loaded_mappings = false;
2675 cache->loaded_discards = false;
2679 atomic_set(&cache->stats.demotion, 0);
2680 atomic_set(&cache->stats.promotion, 0);
2681 atomic_set(&cache->stats.copies_avoided, 0);
2682 atomic_set(&cache->stats.cache_cell_clash, 0);
2683 atomic_set(&cache->stats.commit_count, 0);
2684 atomic_set(&cache->stats.discard_count, 0);
2686 spin_lock_init(&cache->invalidation_lock);
2687 INIT_LIST_HEAD(&cache->invalidation_requests);
2689 batcher_init(&cache->committer, commit_op, cache,
2690 issue_op, cache, cache->wq);
2691 iot_init(&cache->tracker);
2693 init_rwsem(&cache->background_work_lock);
2694 prevent_background_work(cache);
2703 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2708 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2711 for (i = 0; i < argc; i++) {
2712 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2721 cache->nr_ctr_args = argc;
2722 cache->ctr_args = copy;
2727 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2730 struct cache_args *ca;
2731 struct cache *cache = NULL;
2733 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2735 ti->error = "Error allocating memory for cache";
2740 r = parse_cache_args(ca, argc, argv, &ti->error);
2744 r = cache_create(ca, &cache);
2748 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2754 ti->private = cache;
2756 destroy_cache_args(ca);
2760 /*----------------------------------------------------------------*/
2762 static int cache_map(struct dm_target *ti, struct bio *bio)
2764 struct cache *cache = ti->private;
2768 dm_oblock_t block = get_bio_block(cache, bio);
2769 size_t pb_data_size = get_per_bio_data_size(cache);
2771 init_per_bio_data(bio, pb_data_size);
2772 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2774 * This can only occur if the io goes to a partial block at
2775 * the end of the origin device. We don't cache these.
2776 * Just remap to the origin and carry on.
2778 remap_to_origin(cache, bio);
2779 accounted_begin(cache, bio);
2780 return DM_MAPIO_REMAPPED;
2783 if (discard_or_flush(bio)) {
2784 defer_bio(cache, bio);
2785 return DM_MAPIO_SUBMITTED;
2788 r = map_bio(cache, bio, block, &commit_needed);
2790 schedule_commit(&cache->committer);
2795 static int cache_end_io(struct dm_target *ti, struct bio *bio,
2796 blk_status_t *error)
2798 struct cache *cache = ti->private;
2799 unsigned long flags;
2800 size_t pb_data_size = get_per_bio_data_size(cache);
2801 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2804 policy_tick(cache->policy, false);
2806 spin_lock_irqsave(&cache->lock, flags);
2807 cache->need_tick_bio = true;
2808 spin_unlock_irqrestore(&cache->lock, flags);
2811 bio_drop_shared_lock(cache, bio);
2812 accounted_complete(cache, bio);
2814 return DM_ENDIO_DONE;
2817 static int write_dirty_bitset(struct cache *cache)
2821 if (get_cache_mode(cache) >= CM_READ_ONLY)
2824 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2826 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2831 static int write_discard_bitset(struct cache *cache)
2835 if (get_cache_mode(cache) >= CM_READ_ONLY)
2838 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2839 cache->discard_nr_blocks);
2841 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2842 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2846 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2847 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2848 is_discarded(cache, to_dblock(i)));
2850 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2858 static int write_hints(struct cache *cache)
2862 if (get_cache_mode(cache) >= CM_READ_ONLY)
2865 r = dm_cache_write_hints(cache->cmd, cache->policy);
2867 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2875 * returns true on success
2877 static bool sync_metadata(struct cache *cache)
2881 r1 = write_dirty_bitset(cache);
2883 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2885 r2 = write_discard_bitset(cache);
2887 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2891 r3 = write_hints(cache);
2893 DMERR("%s: could not write hints", cache_device_name(cache));
2896 * If writing the above metadata failed, we still commit, but don't
2897 * set the clean shutdown flag. This will effectively force every
2898 * dirty bit to be set on reload.
2900 r4 = commit(cache, !r1 && !r2 && !r3);
2902 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2904 return !r1 && !r2 && !r3 && !r4;
2907 static void cache_postsuspend(struct dm_target *ti)
2909 struct cache *cache = ti->private;
2911 prevent_background_work(cache);
2912 BUG_ON(atomic_read(&cache->nr_io_migrations));
2914 cancel_delayed_work_sync(&cache->waker);
2915 drain_workqueue(cache->wq);
2916 WARN_ON(cache->tracker.in_flight);
2919 * If it's a flush suspend there won't be any deferred bios, so this
2922 requeue_deferred_bios(cache);
2924 if (get_cache_mode(cache) == CM_WRITE)
2925 (void) sync_metadata(cache);
2928 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2929 bool dirty, uint32_t hint, bool hint_valid)
2932 struct cache *cache = context;
2935 set_bit(from_cblock(cblock), cache->dirty_bitset);
2936 atomic_inc(&cache->nr_dirty);
2938 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2940 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2948 * The discard block size in the on disk metadata is not
2949 * neccessarily the same as we're currently using. So we have to
2950 * be careful to only set the discarded attribute if we know it
2951 * covers a complete block of the new size.
2953 struct discard_load_info {
2954 struct cache *cache;
2957 * These blocks are sized using the on disk dblock size, rather
2958 * than the current one.
2960 dm_block_t block_size;
2961 dm_block_t discard_begin, discard_end;
2964 static void discard_load_info_init(struct cache *cache,
2965 struct discard_load_info *li)
2968 li->discard_begin = li->discard_end = 0;
2971 static void set_discard_range(struct discard_load_info *li)
2975 if (li->discard_begin == li->discard_end)
2979 * Convert to sectors.
2981 b = li->discard_begin * li->block_size;
2982 e = li->discard_end * li->block_size;
2985 * Then convert back to the current dblock size.
2987 b = dm_sector_div_up(b, li->cache->discard_block_size);
2988 sector_div(e, li->cache->discard_block_size);
2991 * The origin may have shrunk, so we need to check we're still in
2994 if (e > from_dblock(li->cache->discard_nr_blocks))
2995 e = from_dblock(li->cache->discard_nr_blocks);
2998 set_discard(li->cache, to_dblock(b));
3001 static int load_discard(void *context, sector_t discard_block_size,
3002 dm_dblock_t dblock, bool discard)
3004 struct discard_load_info *li = context;
3006 li->block_size = discard_block_size;
3009 if (from_dblock(dblock) == li->discard_end)
3011 * We're already in a discard range, just extend it.
3013 li->discard_end = li->discard_end + 1ULL;
3017 * Emit the old range and start a new one.
3019 set_discard_range(li);
3020 li->discard_begin = from_dblock(dblock);
3021 li->discard_end = li->discard_begin + 1ULL;
3024 set_discard_range(li);
3025 li->discard_begin = li->discard_end = 0;
3031 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3033 sector_t size = get_dev_size(cache->cache_dev);
3034 (void) sector_div(size, cache->sectors_per_block);
3035 return to_cblock(size);
3038 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3040 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
3042 DMERR("%s: unable to extend cache due to missing cache table reload",
3043 cache_device_name(cache));
3049 * We can't drop a dirty block when shrinking the cache.
3051 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3052 new_size = to_cblock(from_cblock(new_size) + 1);
3053 if (is_dirty(cache, new_size)) {
3054 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3055 cache_device_name(cache),
3056 (unsigned long long) from_cblock(new_size));
3064 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3068 r = dm_cache_resize(cache->cmd, new_size);
3070 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3071 metadata_operation_failed(cache, "dm_cache_resize", r);
3075 set_cache_size(cache, new_size);
3080 static int cache_preresume(struct dm_target *ti)
3083 struct cache *cache = ti->private;
3084 dm_cblock_t csize = get_cache_dev_size(cache);
3087 * Check to see if the cache has resized.
3089 if (!cache->sized) {
3090 r = resize_cache_dev(cache, csize);
3094 cache->sized = true;
3096 } else if (csize != cache->cache_size) {
3097 if (!can_resize(cache, csize))
3100 r = resize_cache_dev(cache, csize);
3105 if (!cache->loaded_mappings) {
3106 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3107 load_mapping, cache);
3109 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3110 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3114 cache->loaded_mappings = true;
3117 if (!cache->loaded_discards) {
3118 struct discard_load_info li;
3121 * The discard bitset could have been resized, or the
3122 * discard block size changed. To be safe we start by
3123 * setting every dblock to not discarded.
3125 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3127 discard_load_info_init(cache, &li);
3128 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3130 DMERR("%s: could not load origin discards", cache_device_name(cache));
3131 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3134 set_discard_range(&li);
3136 cache->loaded_discards = true;
3142 static void cache_resume(struct dm_target *ti)
3144 struct cache *cache = ti->private;
3146 cache->need_tick_bio = true;
3147 allow_background_work(cache);
3148 do_waker(&cache->waker.work);
3154 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3155 * <cache block size> <#used cache blocks>/<#total cache blocks>
3156 * <#read hits> <#read misses> <#write hits> <#write misses>
3157 * <#demotions> <#promotions> <#dirty>
3158 * <#features> <features>*
3159 * <#core args> <core args>
3160 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3162 static void cache_status(struct dm_target *ti, status_type_t type,
3163 unsigned status_flags, char *result, unsigned maxlen)
3168 dm_block_t nr_free_blocks_metadata = 0;
3169 dm_block_t nr_blocks_metadata = 0;
3170 char buf[BDEVNAME_SIZE];
3171 struct cache *cache = ti->private;
3172 dm_cblock_t residency;
3176 case STATUSTYPE_INFO:
3177 if (get_cache_mode(cache) == CM_FAIL) {
3182 /* Commit to ensure statistics aren't out-of-date */
3183 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3184 (void) commit(cache, false);
3186 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3188 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3189 cache_device_name(cache), r);
3193 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3195 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3196 cache_device_name(cache), r);
3200 residency = policy_residency(cache->policy);
3202 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3203 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3204 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3205 (unsigned long long)nr_blocks_metadata,
3206 (unsigned long long)cache->sectors_per_block,
3207 (unsigned long long) from_cblock(residency),
3208 (unsigned long long) from_cblock(cache->cache_size),
3209 (unsigned) atomic_read(&cache->stats.read_hit),
3210 (unsigned) atomic_read(&cache->stats.read_miss),
3211 (unsigned) atomic_read(&cache->stats.write_hit),
3212 (unsigned) atomic_read(&cache->stats.write_miss),
3213 (unsigned) atomic_read(&cache->stats.demotion),
3214 (unsigned) atomic_read(&cache->stats.promotion),
3215 (unsigned long) atomic_read(&cache->nr_dirty));
3217 if (cache->features.metadata_version == 2)
3218 DMEMIT("2 metadata2 ");
3222 if (writethrough_mode(cache))
3223 DMEMIT("writethrough ");
3225 else if (passthrough_mode(cache))
3226 DMEMIT("passthrough ");
3228 else if (writeback_mode(cache))
3229 DMEMIT("writeback ");
3232 DMERR("%s: internal error: unknown io mode: %d",
3233 cache_device_name(cache), (int) cache->features.io_mode);
3237 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3239 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3241 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3243 DMERR("%s: policy_emit_config_values returned %d",
3244 cache_device_name(cache), r);
3247 if (get_cache_mode(cache) == CM_READ_ONLY)
3252 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3254 if (r || needs_check)
3255 DMEMIT("needs_check ");
3261 case STATUSTYPE_TABLE:
3262 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3264 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3266 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3269 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3270 DMEMIT(" %s", cache->ctr_args[i]);
3271 if (cache->nr_ctr_args)
3272 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3282 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3283 * the one-past-the-end value.
3285 struct cblock_range {
3291 * A cache block range can take two forms:
3293 * i) A single cblock, eg. '3456'
3294 * ii) A begin and end cblock with a dash between, eg. 123-234
3296 static int parse_cblock_range(struct cache *cache, const char *str,
3297 struct cblock_range *result)
3304 * Try and parse form (ii) first.
3306 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3311 result->begin = to_cblock(b);
3312 result->end = to_cblock(e);
3317 * That didn't work, try form (i).
3319 r = sscanf(str, "%llu%c", &b, &dummy);
3324 result->begin = to_cblock(b);
3325 result->end = to_cblock(from_cblock(result->begin) + 1u);
3329 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3333 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3335 uint64_t b = from_cblock(range->begin);
3336 uint64_t e = from_cblock(range->end);
3337 uint64_t n = from_cblock(cache->cache_size);
3340 DMERR("%s: begin cblock out of range: %llu >= %llu",
3341 cache_device_name(cache), b, n);
3346 DMERR("%s: end cblock out of range: %llu > %llu",
3347 cache_device_name(cache), e, n);
3352 DMERR("%s: invalid cblock range: %llu >= %llu",
3353 cache_device_name(cache), b, e);
3360 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3362 return to_cblock(from_cblock(b) + 1);
3365 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3370 * We don't need to do any locking here because we know we're in
3371 * passthrough mode. There's is potential for a race between an
3372 * invalidation triggered by an io and an invalidation message. This
3373 * is harmless, we must not worry if the policy call fails.
3375 while (range->begin != range->end) {
3376 r = invalidate_cblock(cache, range->begin);
3380 range->begin = cblock_succ(range->begin);
3383 cache->commit_requested = true;
3387 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3388 const char **cblock_ranges)
3392 struct cblock_range range;
3394 if (!passthrough_mode(cache)) {
3395 DMERR("%s: cache has to be in passthrough mode for invalidation",
3396 cache_device_name(cache));
3400 for (i = 0; i < count; i++) {
3401 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3405 r = validate_cblock_range(cache, &range);
3410 * Pass begin and end origin blocks to the worker and wake it.
3412 r = request_invalidation(cache, &range);
3424 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3426 * The key migration_threshold is supported by the cache target core.
3428 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3430 struct cache *cache = ti->private;
3435 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3436 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3437 cache_device_name(cache));
3441 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3442 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3447 return set_config_value(cache, argv[0], argv[1]);
3450 static int cache_iterate_devices(struct dm_target *ti,
3451 iterate_devices_callout_fn fn, void *data)
3454 struct cache *cache = ti->private;
3456 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3458 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3463 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3466 * FIXME: these limits may be incompatible with the cache device
3468 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3469 cache->origin_sectors);
3470 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3473 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3475 struct cache *cache = ti->private;
3476 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3479 * If the system-determined stacked limits are compatible with the
3480 * cache's blocksize (io_opt is a factor) do not override them.
3482 if (io_opt_sectors < cache->sectors_per_block ||
3483 do_div(io_opt_sectors, cache->sectors_per_block)) {
3484 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3485 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3487 set_discard_limits(cache, limits);
3490 /*----------------------------------------------------------------*/
3492 static struct target_type cache_target = {
3494 .version = {2, 0, 0},
3495 .module = THIS_MODULE,
3499 .end_io = cache_end_io,
3500 .postsuspend = cache_postsuspend,
3501 .preresume = cache_preresume,
3502 .resume = cache_resume,
3503 .status = cache_status,
3504 .message = cache_message,
3505 .iterate_devices = cache_iterate_devices,
3506 .io_hints = cache_io_hints,
3509 static int __init dm_cache_init(void)
3513 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3514 if (!migration_cache)
3517 r = dm_register_target(&cache_target);
3519 DMERR("cache target registration failed: %d", r);
3520 kmem_cache_destroy(migration_cache);
3527 static void __exit dm_cache_exit(void)
3529 dm_unregister_target(&cache_target);
3530 kmem_cache_destroy(migration_cache);
3533 module_init(dm_cache_init);
3534 module_exit(dm_cache_exit);
3536 MODULE_DESCRIPTION(DM_NAME " cache target");
3537 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3538 MODULE_LICENSE("GPL");