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_set_dev(bio
, 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_set_dev(bio
, 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 bool bio_writes_complete_block(struct cache
*cache
, struct bio
*bio
)
1206 return (bio_data_dir(bio
) == WRITE
) &&
1207 (bio
->bi_iter
.bi_size
== (cache
->sectors_per_block
<< SECTOR_SHIFT
));
1210 static bool optimisable_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
)
1212 return writeback_mode(&cache
->features
) &&
1213 (is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
));
1216 static void quiesce(struct dm_cache_migration
*mg
,
1217 void (*continuation
)(struct work_struct
*))
1219 init_continuation(&mg
->k
, continuation
);
1220 dm_cell_quiesce_v2(mg
->cache
->prison
, mg
->cell
, &mg
->k
.ws
);
1223 static struct dm_cache_migration
*ws_to_mg(struct work_struct
*ws
)
1225 struct continuation
*k
= container_of(ws
, struct continuation
, ws
);
1226 return container_of(k
, struct dm_cache_migration
, k
);
1229 static void copy_complete(int read_err
, unsigned long write_err
, void *context
)
1231 struct dm_cache_migration
*mg
= container_of(context
, struct dm_cache_migration
, k
);
1233 if (read_err
|| write_err
)
1234 mg
->k
.input
= BLK_STS_IOERR
;
1236 queue_continuation(mg
->cache
->wq
, &mg
->k
);
1239 static int copy(struct dm_cache_migration
*mg
, bool promote
)
1242 struct dm_io_region o_region
, c_region
;
1243 struct cache
*cache
= mg
->cache
;
1245 o_region
.bdev
= cache
->origin_dev
->bdev
;
1246 o_region
.sector
= from_oblock(mg
->op
->oblock
) * cache
->sectors_per_block
;
1247 o_region
.count
= cache
->sectors_per_block
;
1249 c_region
.bdev
= cache
->cache_dev
->bdev
;
1250 c_region
.sector
= from_cblock(mg
->op
->cblock
) * cache
->sectors_per_block
;
1251 c_region
.count
= cache
->sectors_per_block
;
1254 r
= dm_kcopyd_copy(cache
->copier
, &o_region
, 1, &c_region
, 0, copy_complete
, &mg
->k
);
1256 r
= dm_kcopyd_copy(cache
->copier
, &c_region
, 1, &o_region
, 0, copy_complete
, &mg
->k
);
1261 static void bio_drop_shared_lock(struct cache
*cache
, struct bio
*bio
)
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 if (pb
->cell
&& dm_cell_put_v2(cache
->prison
, pb
->cell
))
1267 free_prison_cell(cache
, pb
->cell
);
1271 static void overwrite_endio(struct bio
*bio
)
1273 struct dm_cache_migration
*mg
= bio
->bi_private
;
1274 struct cache
*cache
= mg
->cache
;
1275 size_t pb_data_size
= get_per_bio_data_size(cache
);
1276 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1278 dm_unhook_bio(&pb
->hook_info
, bio
);
1281 mg
->k
.input
= bio
->bi_status
;
1283 queue_continuation(mg
->cache
->wq
, &mg
->k
);
1286 static void overwrite(struct dm_cache_migration
*mg
,
1287 void (*continuation
)(struct work_struct
*))
1289 struct bio
*bio
= mg
->overwrite_bio
;
1290 size_t pb_data_size
= get_per_bio_data_size(mg
->cache
);
1291 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1293 dm_hook_bio(&pb
->hook_info
, bio
, overwrite_endio
, mg
);
1296 * The overwrite bio is part of the copy operation, as such it does
1297 * not set/clear discard or dirty flags.
1299 if (mg
->op
->op
== POLICY_PROMOTE
)
1300 remap_to_cache(mg
->cache
, bio
, mg
->op
->cblock
);
1302 remap_to_origin(mg
->cache
, bio
);
1304 init_continuation(&mg
->k
, continuation
);
1305 accounted_request(mg
->cache
, bio
);
1311 * 1) exclusive lock preventing WRITEs
1313 * 3) copy or issue overwrite bio
1314 * 4) upgrade to exclusive lock preventing READs and WRITEs
1316 * 6) update metadata and commit
1319 static void mg_complete(struct dm_cache_migration
*mg
, bool success
)
1321 struct bio_list bios
;
1322 struct cache
*cache
= mg
->cache
;
1323 struct policy_work
*op
= mg
->op
;
1324 dm_cblock_t cblock
= op
->cblock
;
1327 update_stats(&cache
->stats
, op
->op
);
1330 case POLICY_PROMOTE
:
1331 clear_discard(cache
, oblock_to_dblock(cache
, op
->oblock
));
1332 policy_complete_background_work(cache
->policy
, op
, success
);
1334 if (mg
->overwrite_bio
) {
1336 force_set_dirty(cache
, cblock
);
1337 else if (mg
->k
.input
)
1338 mg
->overwrite_bio
->bi_status
= mg
->k
.input
;
1340 mg
->overwrite_bio
->bi_status
= BLK_STS_IOERR
;
1341 bio_endio(mg
->overwrite_bio
);
1344 force_clear_dirty(cache
, cblock
);
1345 dec_io_migrations(cache
);
1351 * We clear dirty here to update the nr_dirty counter.
1354 force_clear_dirty(cache
, cblock
);
1355 policy_complete_background_work(cache
->policy
, op
, success
);
1356 dec_io_migrations(cache
);
1359 case POLICY_WRITEBACK
:
1361 force_clear_dirty(cache
, cblock
);
1362 policy_complete_background_work(cache
->policy
, op
, success
);
1363 dec_io_migrations(cache
);
1367 bio_list_init(&bios
);
1369 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1370 free_prison_cell(cache
, mg
->cell
);
1374 defer_bios(cache
, &bios
);
1375 wake_migration_worker(cache
);
1377 background_work_end(cache
);
1380 static void mg_success(struct work_struct
*ws
)
1382 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1383 mg_complete(mg
, mg
->k
.input
== 0);
1386 static void mg_update_metadata(struct work_struct
*ws
)
1389 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1390 struct cache
*cache
= mg
->cache
;
1391 struct policy_work
*op
= mg
->op
;
1394 case POLICY_PROMOTE
:
1395 r
= dm_cache_insert_mapping(cache
->cmd
, op
->cblock
, op
->oblock
);
1397 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1398 cache_device_name(cache
));
1399 metadata_operation_failed(cache
, "dm_cache_insert_mapping", r
);
1401 mg_complete(mg
, false);
1404 mg_complete(mg
, true);
1408 r
= dm_cache_remove_mapping(cache
->cmd
, op
->cblock
);
1410 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1411 cache_device_name(cache
));
1412 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1414 mg_complete(mg
, false);
1419 * It would be nice if we only had to commit when a REQ_FLUSH
1420 * comes through. But there's one scenario that we have to
1423 * - vblock x in a cache block
1425 * - cache block gets reallocated and over written
1428 * When we recover, because there was no commit the cache will
1429 * rollback to having the data for vblock x in the cache block.
1430 * But the cache block has since been overwritten, so it'll end
1431 * up pointing to data that was never in 'x' during the history
1434 * To avoid this issue we require a commit as part of the
1435 * demotion operation.
1437 init_continuation(&mg
->k
, mg_success
);
1438 continue_after_commit(&cache
->committer
, &mg
->k
);
1439 schedule_commit(&cache
->committer
);
1442 case POLICY_WRITEBACK
:
1443 mg_complete(mg
, true);
1448 static void mg_update_metadata_after_copy(struct work_struct
*ws
)
1450 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1453 * Did the copy succeed?
1456 mg_complete(mg
, false);
1458 mg_update_metadata(ws
);
1461 static void mg_upgrade_lock(struct work_struct
*ws
)
1464 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1467 * Did the copy succeed?
1470 mg_complete(mg
, false);
1474 * Now we want the lock to prevent both reads and writes.
1476 r
= dm_cell_lock_promote_v2(mg
->cache
->prison
, mg
->cell
,
1477 READ_WRITE_LOCK_LEVEL
);
1479 mg_complete(mg
, false);
1482 quiesce(mg
, mg_update_metadata
);
1485 mg_update_metadata(ws
);
1489 static void mg_full_copy(struct work_struct
*ws
)
1491 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1492 struct cache
*cache
= mg
->cache
;
1493 struct policy_work
*op
= mg
->op
;
1494 bool is_policy_promote
= (op
->op
== POLICY_PROMOTE
);
1496 if ((!is_policy_promote
&& !is_dirty(cache
, op
->cblock
)) ||
1497 is_discarded_oblock(cache
, op
->oblock
)) {
1498 mg_upgrade_lock(ws
);
1502 init_continuation(&mg
->k
, mg_upgrade_lock
);
1504 if (copy(mg
, is_policy_promote
)) {
1505 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache
));
1506 mg
->k
.input
= BLK_STS_IOERR
;
1507 mg_complete(mg
, false);
1511 static void mg_copy(struct work_struct
*ws
)
1513 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1515 if (mg
->overwrite_bio
) {
1517 * No exclusive lock was held when we last checked if the bio
1518 * was optimisable. So we have to check again in case things
1519 * have changed (eg, the block may no longer be discarded).
1521 if (!optimisable_bio(mg
->cache
, mg
->overwrite_bio
, mg
->op
->oblock
)) {
1523 * Fallback to a real full copy after doing some tidying up.
1525 bool rb
= bio_detain_shared(mg
->cache
, mg
->op
->oblock
, mg
->overwrite_bio
);
1526 BUG_ON(rb
); /* An exclussive lock must _not_ be held for this block */
1527 mg
->overwrite_bio
= NULL
;
1528 inc_io_migrations(mg
->cache
);
1534 * It's safe to do this here, even though it's new data
1535 * because all IO has been locked out of the block.
1537 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1538 * so _not_ using mg_upgrade_lock() as continutation.
1540 overwrite(mg
, mg_update_metadata_after_copy
);
1546 static int mg_lock_writes(struct dm_cache_migration
*mg
)
1549 struct dm_cell_key_v2 key
;
1550 struct cache
*cache
= mg
->cache
;
1551 struct dm_bio_prison_cell_v2
*prealloc
;
1553 prealloc
= alloc_prison_cell(cache
);
1555 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache
));
1556 mg_complete(mg
, false);
1561 * Prevent writes to the block, but allow reads to continue.
1562 * Unless we're using an overwrite bio, in which case we lock
1565 build_key(mg
->op
->oblock
, oblock_succ(mg
->op
->oblock
), &key
);
1566 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1567 mg
->overwrite_bio
? READ_WRITE_LOCK_LEVEL
: WRITE_LOCK_LEVEL
,
1568 prealloc
, &mg
->cell
);
1570 free_prison_cell(cache
, prealloc
);
1571 mg_complete(mg
, false);
1575 if (mg
->cell
!= prealloc
)
1576 free_prison_cell(cache
, prealloc
);
1581 quiesce(mg
, mg_copy
);
1586 static int mg_start(struct cache
*cache
, struct policy_work
*op
, struct bio
*bio
)
1588 struct dm_cache_migration
*mg
;
1590 if (!background_work_begin(cache
)) {
1591 policy_complete_background_work(cache
->policy
, op
, false);
1595 mg
= alloc_migration(cache
);
1597 policy_complete_background_work(cache
->policy
, op
, false);
1598 background_work_end(cache
);
1602 memset(mg
, 0, sizeof(*mg
));
1606 mg
->overwrite_bio
= bio
;
1609 inc_io_migrations(cache
);
1611 return mg_lock_writes(mg
);
1614 /*----------------------------------------------------------------
1615 * invalidation processing
1616 *--------------------------------------------------------------*/
1618 static void invalidate_complete(struct dm_cache_migration
*mg
, bool success
)
1620 struct bio_list bios
;
1621 struct cache
*cache
= mg
->cache
;
1623 bio_list_init(&bios
);
1624 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1625 free_prison_cell(cache
, mg
->cell
);
1627 if (!success
&& mg
->overwrite_bio
)
1628 bio_io_error(mg
->overwrite_bio
);
1631 defer_bios(cache
, &bios
);
1633 background_work_end(cache
);
1636 static void invalidate_completed(struct work_struct
*ws
)
1638 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1639 invalidate_complete(mg
, !mg
->k
.input
);
1642 static int invalidate_cblock(struct cache
*cache
, dm_cblock_t cblock
)
1644 int r
= policy_invalidate_mapping(cache
->policy
, cblock
);
1646 r
= dm_cache_remove_mapping(cache
->cmd
, cblock
);
1648 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1649 cache_device_name(cache
));
1650 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1653 } else if (r
== -ENODATA
) {
1655 * Harmless, already unmapped.
1660 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache
));
1665 static void invalidate_remove(struct work_struct
*ws
)
1668 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1669 struct cache
*cache
= mg
->cache
;
1671 r
= invalidate_cblock(cache
, mg
->invalidate_cblock
);
1673 invalidate_complete(mg
, false);
1677 init_continuation(&mg
->k
, invalidate_completed
);
1678 continue_after_commit(&cache
->committer
, &mg
->k
);
1679 remap_to_origin_clear_discard(cache
, mg
->overwrite_bio
, mg
->invalidate_oblock
);
1680 mg
->overwrite_bio
= NULL
;
1681 schedule_commit(&cache
->committer
);
1684 static int invalidate_lock(struct dm_cache_migration
*mg
)
1687 struct dm_cell_key_v2 key
;
1688 struct cache
*cache
= mg
->cache
;
1689 struct dm_bio_prison_cell_v2
*prealloc
;
1691 prealloc
= alloc_prison_cell(cache
);
1693 invalidate_complete(mg
, false);
1697 build_key(mg
->invalidate_oblock
, oblock_succ(mg
->invalidate_oblock
), &key
);
1698 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1699 READ_WRITE_LOCK_LEVEL
, prealloc
, &mg
->cell
);
1701 free_prison_cell(cache
, prealloc
);
1702 invalidate_complete(mg
, false);
1706 if (mg
->cell
!= prealloc
)
1707 free_prison_cell(cache
, prealloc
);
1710 quiesce(mg
, invalidate_remove
);
1714 * We can't call invalidate_remove() directly here because we
1715 * might still be in request context.
1717 init_continuation(&mg
->k
, invalidate_remove
);
1718 queue_work(cache
->wq
, &mg
->k
.ws
);
1724 static int invalidate_start(struct cache
*cache
, dm_cblock_t cblock
,
1725 dm_oblock_t oblock
, struct bio
*bio
)
1727 struct dm_cache_migration
*mg
;
1729 if (!background_work_begin(cache
))
1732 mg
= alloc_migration(cache
);
1734 background_work_end(cache
);
1738 memset(mg
, 0, sizeof(*mg
));
1741 mg
->overwrite_bio
= bio
;
1742 mg
->invalidate_cblock
= cblock
;
1743 mg
->invalidate_oblock
= oblock
;
1745 return invalidate_lock(mg
);
1748 /*----------------------------------------------------------------
1750 *--------------------------------------------------------------*/
1757 static enum busy
spare_migration_bandwidth(struct cache
*cache
)
1759 bool idle
= iot_idle_for(&cache
->tracker
, HZ
);
1760 sector_t current_volume
= (atomic_read(&cache
->nr_io_migrations
) + 1) *
1761 cache
->sectors_per_block
;
1763 if (idle
&& current_volume
<= cache
->migration_threshold
)
1769 static void inc_hit_counter(struct cache
*cache
, struct bio
*bio
)
1771 atomic_inc(bio_data_dir(bio
) == READ
?
1772 &cache
->stats
.read_hit
: &cache
->stats
.write_hit
);
1775 static void inc_miss_counter(struct cache
*cache
, struct bio
*bio
)
1777 atomic_inc(bio_data_dir(bio
) == READ
?
1778 &cache
->stats
.read_miss
: &cache
->stats
.write_miss
);
1781 /*----------------------------------------------------------------*/
1783 static int map_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
,
1784 bool *commit_needed
)
1787 bool rb
, background_queued
;
1789 size_t pb_data_size
= get_per_bio_data_size(cache
);
1790 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1792 *commit_needed
= false;
1794 rb
= bio_detain_shared(cache
, block
, bio
);
1797 * An exclusive lock is held for this block, so we have to
1798 * wait. We set the commit_needed flag so the current
1799 * transaction will be committed asap, allowing this lock
1802 *commit_needed
= true;
1803 return DM_MAPIO_SUBMITTED
;
1806 data_dir
= bio_data_dir(bio
);
1808 if (optimisable_bio(cache
, bio
, block
)) {
1809 struct policy_work
*op
= NULL
;
1811 r
= policy_lookup_with_work(cache
->policy
, block
, &cblock
, data_dir
, true, &op
);
1812 if (unlikely(r
&& r
!= -ENOENT
)) {
1813 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1814 cache_device_name(cache
), r
);
1816 return DM_MAPIO_SUBMITTED
;
1819 if (r
== -ENOENT
&& op
) {
1820 bio_drop_shared_lock(cache
, bio
);
1821 BUG_ON(op
->op
!= POLICY_PROMOTE
);
1822 mg_start(cache
, op
, bio
);
1823 return DM_MAPIO_SUBMITTED
;
1826 r
= policy_lookup(cache
->policy
, block
, &cblock
, data_dir
, false, &background_queued
);
1827 if (unlikely(r
&& r
!= -ENOENT
)) {
1828 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1829 cache_device_name(cache
), r
);
1831 return DM_MAPIO_SUBMITTED
;
1834 if (background_queued
)
1835 wake_migration_worker(cache
);
1842 inc_miss_counter(cache
, bio
);
1843 if (pb
->req_nr
== 0) {
1844 accounted_begin(cache
, bio
);
1845 remap_to_origin_clear_discard(cache
, bio
, block
);
1849 * This is a duplicate writethrough io that is no
1850 * longer needed because the block has been demoted.
1853 return DM_MAPIO_SUBMITTED
;
1859 inc_hit_counter(cache
, bio
);
1862 * Passthrough always maps to the origin, invalidating any
1863 * cache blocks that are written to.
1865 if (passthrough_mode(&cache
->features
)) {
1866 if (bio_data_dir(bio
) == WRITE
) {
1867 bio_drop_shared_lock(cache
, bio
);
1868 atomic_inc(&cache
->stats
.demotion
);
1869 invalidate_start(cache
, cblock
, block
, bio
);
1871 remap_to_origin_clear_discard(cache
, bio
, block
);
1874 if (bio_data_dir(bio
) == WRITE
&& writethrough_mode(&cache
->features
) &&
1875 !is_dirty(cache
, cblock
)) {
1876 remap_to_origin_then_cache(cache
, bio
, block
, cblock
);
1877 accounted_begin(cache
, bio
);
1879 remap_to_cache_dirty(cache
, bio
, block
, cblock
);
1884 * dm core turns FUA requests into a separate payload and FLUSH req.
1886 if (bio
->bi_opf
& REQ_FUA
) {
1888 * issue_after_commit will call accounted_begin a second time. So
1889 * we call accounted_complete() to avoid double accounting.
1891 accounted_complete(cache
, bio
);
1892 issue_after_commit(&cache
->committer
, bio
);
1893 *commit_needed
= true;
1894 return DM_MAPIO_SUBMITTED
;
1897 return DM_MAPIO_REMAPPED
;
1900 static bool process_bio(struct cache
*cache
, struct bio
*bio
)
1904 if (map_bio(cache
, bio
, get_bio_block(cache
, bio
), &commit_needed
) == DM_MAPIO_REMAPPED
)
1905 generic_make_request(bio
);
1907 return commit_needed
;
1911 * A non-zero return indicates read_only or fail_io mode.
1913 static int commit(struct cache
*cache
, bool clean_shutdown
)
1917 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1920 atomic_inc(&cache
->stats
.commit_count
);
1921 r
= dm_cache_commit(cache
->cmd
, clean_shutdown
);
1923 metadata_operation_failed(cache
, "dm_cache_commit", r
);
1929 * Used by the batcher.
1931 static blk_status_t
commit_op(void *context
)
1933 struct cache
*cache
= context
;
1935 if (dm_cache_changed_this_transaction(cache
->cmd
))
1936 return errno_to_blk_status(commit(cache
, false));
1941 /*----------------------------------------------------------------*/
1943 static bool process_flush_bio(struct cache
*cache
, struct bio
*bio
)
1945 size_t pb_data_size
= get_per_bio_data_size(cache
);
1946 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1949 remap_to_origin(cache
, bio
);
1951 remap_to_cache(cache
, bio
, 0);
1953 issue_after_commit(&cache
->committer
, bio
);
1957 static bool process_discard_bio(struct cache
*cache
, struct bio
*bio
)
1961 // FIXME: do we need to lock the region? Or can we just assume the
1962 // user wont be so foolish as to issue discard concurrently with
1964 calc_discard_block_range(cache
, bio
, &b
, &e
);
1966 set_discard(cache
, b
);
1967 b
= to_dblock(from_dblock(b
) + 1);
1975 static void process_deferred_bios(struct work_struct
*ws
)
1977 struct cache
*cache
= container_of(ws
, struct cache
, deferred_bio_worker
);
1979 unsigned long flags
;
1980 bool commit_needed
= false;
1981 struct bio_list bios
;
1984 bio_list_init(&bios
);
1986 spin_lock_irqsave(&cache
->lock
, flags
);
1987 bio_list_merge(&bios
, &cache
->deferred_bios
);
1988 bio_list_init(&cache
->deferred_bios
);
1989 spin_unlock_irqrestore(&cache
->lock
, flags
);
1991 while ((bio
= bio_list_pop(&bios
))) {
1992 if (bio
->bi_opf
& REQ_PREFLUSH
)
1993 commit_needed
= process_flush_bio(cache
, bio
) || commit_needed
;
1995 else if (bio_op(bio
) == REQ_OP_DISCARD
)
1996 commit_needed
= process_discard_bio(cache
, bio
) || commit_needed
;
1999 commit_needed
= process_bio(cache
, bio
) || commit_needed
;
2003 schedule_commit(&cache
->committer
);
2006 static void process_deferred_writethrough_bios(struct work_struct
*ws
)
2008 struct cache
*cache
= container_of(ws
, struct cache
, deferred_writethrough_worker
);
2010 unsigned long flags
;
2011 struct bio_list bios
;
2014 bio_list_init(&bios
);
2016 spin_lock_irqsave(&cache
->lock
, flags
);
2017 bio_list_merge(&bios
, &cache
->deferred_writethrough_bios
);
2018 bio_list_init(&cache
->deferred_writethrough_bios
);
2019 spin_unlock_irqrestore(&cache
->lock
, flags
);
2022 * These bios have already been through accounted_begin()
2024 while ((bio
= bio_list_pop(&bios
)))
2025 generic_make_request(bio
);
2028 /*----------------------------------------------------------------
2030 *--------------------------------------------------------------*/
2032 static void requeue_deferred_bios(struct cache
*cache
)
2035 struct bio_list bios
;
2037 bio_list_init(&bios
);
2038 bio_list_merge(&bios
, &cache
->deferred_bios
);
2039 bio_list_init(&cache
->deferred_bios
);
2041 while ((bio
= bio_list_pop(&bios
))) {
2042 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
2048 * We want to commit periodically so that not too much
2049 * unwritten metadata builds up.
2051 static void do_waker(struct work_struct
*ws
)
2053 struct cache
*cache
= container_of(to_delayed_work(ws
), struct cache
, waker
);
2055 policy_tick(cache
->policy
, true);
2056 wake_migration_worker(cache
);
2057 schedule_commit(&cache
->committer
);
2058 queue_delayed_work(cache
->wq
, &cache
->waker
, COMMIT_PERIOD
);
2061 static void check_migrations(struct work_struct
*ws
)
2064 struct policy_work
*op
;
2065 struct cache
*cache
= container_of(ws
, struct cache
, migration_worker
);
2069 b
= spare_migration_bandwidth(cache
);
2071 r
= policy_get_background_work(cache
->policy
, b
== IDLE
, &op
);
2076 DMERR_LIMIT("%s: policy_background_work failed",
2077 cache_device_name(cache
));
2081 r
= mg_start(cache
, op
, NULL
);
2087 /*----------------------------------------------------------------
2089 *--------------------------------------------------------------*/
2092 * This function gets called on the error paths of the constructor, so we
2093 * have to cope with a partially initialised struct.
2095 static void destroy(struct cache
*cache
)
2099 mempool_destroy(cache
->migration_pool
);
2102 dm_bio_prison_destroy_v2(cache
->prison
);
2105 destroy_workqueue(cache
->wq
);
2107 if (cache
->dirty_bitset
)
2108 free_bitset(cache
->dirty_bitset
);
2110 if (cache
->discard_bitset
)
2111 free_bitset(cache
->discard_bitset
);
2114 dm_kcopyd_client_destroy(cache
->copier
);
2117 dm_cache_metadata_close(cache
->cmd
);
2119 if (cache
->metadata_dev
)
2120 dm_put_device(cache
->ti
, cache
->metadata_dev
);
2122 if (cache
->origin_dev
)
2123 dm_put_device(cache
->ti
, cache
->origin_dev
);
2125 if (cache
->cache_dev
)
2126 dm_put_device(cache
->ti
, cache
->cache_dev
);
2129 dm_cache_policy_destroy(cache
->policy
);
2131 for (i
= 0; i
< cache
->nr_ctr_args
; i
++)
2132 kfree(cache
->ctr_args
[i
]);
2133 kfree(cache
->ctr_args
);
2138 static void cache_dtr(struct dm_target
*ti
)
2140 struct cache
*cache
= ti
->private;
2145 static sector_t
get_dev_size(struct dm_dev
*dev
)
2147 return i_size_read(dev
->bdev
->bd_inode
) >> SECTOR_SHIFT
;
2150 /*----------------------------------------------------------------*/
2153 * Construct a cache device mapping.
2155 * cache <metadata dev> <cache dev> <origin dev> <block size>
2156 * <#feature args> [<feature arg>]*
2157 * <policy> <#policy args> [<policy arg>]*
2159 * metadata dev : fast device holding the persistent metadata
2160 * cache dev : fast device holding cached data blocks
2161 * origin dev : slow device holding original data blocks
2162 * block size : cache unit size in sectors
2164 * #feature args : number of feature arguments passed
2165 * feature args : writethrough. (The default is writeback.)
2167 * policy : the replacement policy to use
2168 * #policy args : an even number of policy arguments corresponding
2169 * to key/value pairs passed to the policy
2170 * policy args : key/value pairs passed to the policy
2171 * E.g. 'sequential_threshold 1024'
2172 * See cache-policies.txt for details.
2174 * Optional feature arguments are:
2175 * writethrough : write through caching that prohibits cache block
2176 * content from being different from origin block content.
2177 * Without this argument, the default behaviour is to write
2178 * back cache block contents later for performance reasons,
2179 * so they may differ from the corresponding origin blocks.
2182 struct dm_target
*ti
;
2184 struct dm_dev
*metadata_dev
;
2186 struct dm_dev
*cache_dev
;
2187 sector_t cache_sectors
;
2189 struct dm_dev
*origin_dev
;
2190 sector_t origin_sectors
;
2192 uint32_t block_size
;
2194 const char *policy_name
;
2196 const char **policy_argv
;
2198 struct cache_features features
;
2201 static void destroy_cache_args(struct cache_args
*ca
)
2203 if (ca
->metadata_dev
)
2204 dm_put_device(ca
->ti
, ca
->metadata_dev
);
2207 dm_put_device(ca
->ti
, ca
->cache_dev
);
2210 dm_put_device(ca
->ti
, ca
->origin_dev
);
2215 static bool at_least_one_arg(struct dm_arg_set
*as
, char **error
)
2218 *error
= "Insufficient args";
2225 static int parse_metadata_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2229 sector_t metadata_dev_size
;
2230 char b
[BDEVNAME_SIZE
];
2232 if (!at_least_one_arg(as
, error
))
2235 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2238 *error
= "Error opening metadata device";
2242 metadata_dev_size
= get_dev_size(ca
->metadata_dev
);
2243 if (metadata_dev_size
> DM_CACHE_METADATA_MAX_SECTORS_WARNING
)
2244 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2245 bdevname(ca
->metadata_dev
->bdev
, b
), THIN_METADATA_MAX_SECTORS
);
2250 static int parse_cache_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2255 if (!at_least_one_arg(as
, error
))
2258 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2261 *error
= "Error opening cache device";
2264 ca
->cache_sectors
= get_dev_size(ca
->cache_dev
);
2269 static int parse_origin_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2274 if (!at_least_one_arg(as
, error
))
2277 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2280 *error
= "Error opening origin device";
2284 ca
->origin_sectors
= get_dev_size(ca
->origin_dev
);
2285 if (ca
->ti
->len
> ca
->origin_sectors
) {
2286 *error
= "Device size larger than cached device";
2293 static int parse_block_size(struct cache_args
*ca
, struct dm_arg_set
*as
,
2296 unsigned long block_size
;
2298 if (!at_least_one_arg(as
, error
))
2301 if (kstrtoul(dm_shift_arg(as
), 10, &block_size
) || !block_size
||
2302 block_size
< DATA_DEV_BLOCK_SIZE_MIN_SECTORS
||
2303 block_size
> DATA_DEV_BLOCK_SIZE_MAX_SECTORS
||
2304 block_size
& (DATA_DEV_BLOCK_SIZE_MIN_SECTORS
- 1)) {
2305 *error
= "Invalid data block size";
2309 if (block_size
> ca
->cache_sectors
) {
2310 *error
= "Data block size is larger than the cache device";
2314 ca
->block_size
= block_size
;
2319 static void init_features(struct cache_features
*cf
)
2321 cf
->mode
= CM_WRITE
;
2322 cf
->io_mode
= CM_IO_WRITEBACK
;
2323 cf
->metadata_version
= 1;
2326 static int parse_features(struct cache_args
*ca
, struct dm_arg_set
*as
,
2329 static const struct dm_arg _args
[] = {
2330 {0, 2, "Invalid number of cache feature arguments"},
2333 int r
, mode_ctr
= 0;
2336 struct cache_features
*cf
= &ca
->features
;
2340 r
= dm_read_arg_group(_args
, as
, &argc
, error
);
2345 arg
= dm_shift_arg(as
);
2347 if (!strcasecmp(arg
, "writeback")) {
2348 cf
->io_mode
= CM_IO_WRITEBACK
;
2352 else if (!strcasecmp(arg
, "writethrough")) {
2353 cf
->io_mode
= CM_IO_WRITETHROUGH
;
2357 else if (!strcasecmp(arg
, "passthrough")) {
2358 cf
->io_mode
= CM_IO_PASSTHROUGH
;
2362 else if (!strcasecmp(arg
, "metadata2"))
2363 cf
->metadata_version
= 2;
2366 *error
= "Unrecognised cache feature requested";
2372 *error
= "Duplicate cache io_mode features requested";
2379 static int parse_policy(struct cache_args
*ca
, struct dm_arg_set
*as
,
2382 static const struct dm_arg _args
[] = {
2383 {0, 1024, "Invalid number of policy arguments"},
2388 if (!at_least_one_arg(as
, error
))
2391 ca
->policy_name
= dm_shift_arg(as
);
2393 r
= dm_read_arg_group(_args
, as
, &ca
->policy_argc
, error
);
2397 ca
->policy_argv
= (const char **)as
->argv
;
2398 dm_consume_args(as
, ca
->policy_argc
);
2403 static int parse_cache_args(struct cache_args
*ca
, int argc
, char **argv
,
2407 struct dm_arg_set as
;
2412 r
= parse_metadata_dev(ca
, &as
, error
);
2416 r
= parse_cache_dev(ca
, &as
, error
);
2420 r
= parse_origin_dev(ca
, &as
, error
);
2424 r
= parse_block_size(ca
, &as
, error
);
2428 r
= parse_features(ca
, &as
, error
);
2432 r
= parse_policy(ca
, &as
, error
);
2439 /*----------------------------------------------------------------*/
2441 static struct kmem_cache
*migration_cache
;
2443 #define NOT_CORE_OPTION 1
2445 static int process_config_option(struct cache
*cache
, const char *key
, const char *value
)
2449 if (!strcasecmp(key
, "migration_threshold")) {
2450 if (kstrtoul(value
, 10, &tmp
))
2453 cache
->migration_threshold
= tmp
;
2457 return NOT_CORE_OPTION
;
2460 static int set_config_value(struct cache
*cache
, const char *key
, const char *value
)
2462 int r
= process_config_option(cache
, key
, value
);
2464 if (r
== NOT_CORE_OPTION
)
2465 r
= policy_set_config_value(cache
->policy
, key
, value
);
2468 DMWARN("bad config value for %s: %s", key
, value
);
2473 static int set_config_values(struct cache
*cache
, int argc
, const char **argv
)
2478 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2483 r
= set_config_value(cache
, argv
[0], argv
[1]);
2494 static int create_cache_policy(struct cache
*cache
, struct cache_args
*ca
,
2497 struct dm_cache_policy
*p
= dm_cache_policy_create(ca
->policy_name
,
2499 cache
->origin_sectors
,
2500 cache
->sectors_per_block
);
2502 *error
= "Error creating cache's policy";
2506 BUG_ON(!cache
->policy
);
2512 * We want the discard block size to be at least the size of the cache
2513 * block size and have no more than 2^14 discard blocks across the origin.
2515 #define MAX_DISCARD_BLOCKS (1 << 14)
2517 static bool too_many_discard_blocks(sector_t discard_block_size
,
2518 sector_t origin_size
)
2520 (void) sector_div(origin_size
, discard_block_size
);
2522 return origin_size
> MAX_DISCARD_BLOCKS
;
2525 static sector_t
calculate_discard_block_size(sector_t cache_block_size
,
2526 sector_t origin_size
)
2528 sector_t discard_block_size
= cache_block_size
;
2531 while (too_many_discard_blocks(discard_block_size
, origin_size
))
2532 discard_block_size
*= 2;
2534 return discard_block_size
;
2537 static void set_cache_size(struct cache
*cache
, dm_cblock_t size
)
2539 dm_block_t nr_blocks
= from_cblock(size
);
2541 if (nr_blocks
> (1 << 20) && cache
->cache_size
!= size
)
2542 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2543 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2544 "Please consider increasing the cache block size to reduce the overall cache block count.",
2545 (unsigned long long) nr_blocks
);
2547 cache
->cache_size
= size
;
2550 static int is_congested(struct dm_dev
*dev
, int bdi_bits
)
2552 struct request_queue
*q
= bdev_get_queue(dev
->bdev
);
2553 return bdi_congested(q
->backing_dev_info
, bdi_bits
);
2556 static int cache_is_congested(struct dm_target_callbacks
*cb
, int bdi_bits
)
2558 struct cache
*cache
= container_of(cb
, struct cache
, callbacks
);
2560 return is_congested(cache
->origin_dev
, bdi_bits
) ||
2561 is_congested(cache
->cache_dev
, bdi_bits
);
2564 #define DEFAULT_MIGRATION_THRESHOLD 2048
2566 static int cache_create(struct cache_args
*ca
, struct cache
**result
)
2569 char **error
= &ca
->ti
->error
;
2570 struct cache
*cache
;
2571 struct dm_target
*ti
= ca
->ti
;
2572 dm_block_t origin_blocks
;
2573 struct dm_cache_metadata
*cmd
;
2574 bool may_format
= ca
->features
.mode
== CM_WRITE
;
2576 cache
= kzalloc(sizeof(*cache
), GFP_KERNEL
);
2581 ti
->private = cache
;
2582 ti
->num_flush_bios
= 2;
2583 ti
->flush_supported
= true;
2585 ti
->num_discard_bios
= 1;
2586 ti
->discards_supported
= true;
2587 ti
->split_discard_bios
= false;
2589 cache
->features
= ca
->features
;
2590 ti
->per_io_data_size
= get_per_bio_data_size(cache
);
2592 cache
->callbacks
.congested_fn
= cache_is_congested
;
2593 dm_table_add_target_callbacks(ti
->table
, &cache
->callbacks
);
2595 cache
->metadata_dev
= ca
->metadata_dev
;
2596 cache
->origin_dev
= ca
->origin_dev
;
2597 cache
->cache_dev
= ca
->cache_dev
;
2599 ca
->metadata_dev
= ca
->origin_dev
= ca
->cache_dev
= NULL
;
2601 origin_blocks
= cache
->origin_sectors
= ca
->origin_sectors
;
2602 origin_blocks
= block_div(origin_blocks
, ca
->block_size
);
2603 cache
->origin_blocks
= to_oblock(origin_blocks
);
2605 cache
->sectors_per_block
= ca
->block_size
;
2606 if (dm_set_target_max_io_len(ti
, cache
->sectors_per_block
)) {
2611 if (ca
->block_size
& (ca
->block_size
- 1)) {
2612 dm_block_t cache_size
= ca
->cache_sectors
;
2614 cache
->sectors_per_block_shift
= -1;
2615 cache_size
= block_div(cache_size
, ca
->block_size
);
2616 set_cache_size(cache
, to_cblock(cache_size
));
2618 cache
->sectors_per_block_shift
= __ffs(ca
->block_size
);
2619 set_cache_size(cache
, to_cblock(ca
->cache_sectors
>> cache
->sectors_per_block_shift
));
2622 r
= create_cache_policy(cache
, ca
, error
);
2626 cache
->policy_nr_args
= ca
->policy_argc
;
2627 cache
->migration_threshold
= DEFAULT_MIGRATION_THRESHOLD
;
2629 r
= set_config_values(cache
, ca
->policy_argc
, ca
->policy_argv
);
2631 *error
= "Error setting cache policy's config values";
2635 cmd
= dm_cache_metadata_open(cache
->metadata_dev
->bdev
,
2636 ca
->block_size
, may_format
,
2637 dm_cache_policy_get_hint_size(cache
->policy
),
2638 ca
->features
.metadata_version
);
2640 *error
= "Error creating metadata object";
2645 set_cache_mode(cache
, CM_WRITE
);
2646 if (get_cache_mode(cache
) != CM_WRITE
) {
2647 *error
= "Unable to get write access to metadata, please check/repair metadata.";
2652 if (passthrough_mode(&cache
->features
)) {
2655 r
= dm_cache_metadata_all_clean(cache
->cmd
, &all_clean
);
2657 *error
= "dm_cache_metadata_all_clean() failed";
2662 *error
= "Cannot enter passthrough mode unless all blocks are clean";
2667 policy_allow_migrations(cache
->policy
, false);
2670 spin_lock_init(&cache
->lock
);
2671 INIT_LIST_HEAD(&cache
->deferred_cells
);
2672 bio_list_init(&cache
->deferred_bios
);
2673 bio_list_init(&cache
->deferred_writethrough_bios
);
2674 atomic_set(&cache
->nr_allocated_migrations
, 0);
2675 atomic_set(&cache
->nr_io_migrations
, 0);
2676 init_waitqueue_head(&cache
->migration_wait
);
2679 atomic_set(&cache
->nr_dirty
, 0);
2680 cache
->dirty_bitset
= alloc_bitset(from_cblock(cache
->cache_size
));
2681 if (!cache
->dirty_bitset
) {
2682 *error
= "could not allocate dirty bitset";
2685 clear_bitset(cache
->dirty_bitset
, from_cblock(cache
->cache_size
));
2687 cache
->discard_block_size
=
2688 calculate_discard_block_size(cache
->sectors_per_block
,
2689 cache
->origin_sectors
);
2690 cache
->discard_nr_blocks
= to_dblock(dm_sector_div_up(cache
->origin_sectors
,
2691 cache
->discard_block_size
));
2692 cache
->discard_bitset
= alloc_bitset(from_dblock(cache
->discard_nr_blocks
));
2693 if (!cache
->discard_bitset
) {
2694 *error
= "could not allocate discard bitset";
2697 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
2699 cache
->copier
= dm_kcopyd_client_create(&dm_kcopyd_throttle
);
2700 if (IS_ERR(cache
->copier
)) {
2701 *error
= "could not create kcopyd client";
2702 r
= PTR_ERR(cache
->copier
);
2706 cache
->wq
= alloc_workqueue("dm-" DM_MSG_PREFIX
, WQ_MEM_RECLAIM
, 0);
2708 *error
= "could not create workqueue for metadata object";
2711 INIT_WORK(&cache
->deferred_bio_worker
, process_deferred_bios
);
2712 INIT_WORK(&cache
->deferred_writethrough_worker
,
2713 process_deferred_writethrough_bios
);
2714 INIT_WORK(&cache
->migration_worker
, check_migrations
);
2715 INIT_DELAYED_WORK(&cache
->waker
, do_waker
);
2717 cache
->prison
= dm_bio_prison_create_v2(cache
->wq
);
2718 if (!cache
->prison
) {
2719 *error
= "could not create bio prison";
2723 cache
->migration_pool
= mempool_create_slab_pool(MIGRATION_POOL_SIZE
,
2725 if (!cache
->migration_pool
) {
2726 *error
= "Error creating cache's migration mempool";
2730 cache
->need_tick_bio
= true;
2731 cache
->sized
= false;
2732 cache
->invalidate
= false;
2733 cache
->commit_requested
= false;
2734 cache
->loaded_mappings
= false;
2735 cache
->loaded_discards
= false;
2739 atomic_set(&cache
->stats
.demotion
, 0);
2740 atomic_set(&cache
->stats
.promotion
, 0);
2741 atomic_set(&cache
->stats
.copies_avoided
, 0);
2742 atomic_set(&cache
->stats
.cache_cell_clash
, 0);
2743 atomic_set(&cache
->stats
.commit_count
, 0);
2744 atomic_set(&cache
->stats
.discard_count
, 0);
2746 spin_lock_init(&cache
->invalidation_lock
);
2747 INIT_LIST_HEAD(&cache
->invalidation_requests
);
2749 batcher_init(&cache
->committer
, commit_op
, cache
,
2750 issue_op
, cache
, cache
->wq
);
2751 iot_init(&cache
->tracker
);
2753 init_rwsem(&cache
->background_work_lock
);
2754 prevent_background_work(cache
);
2763 static int copy_ctr_args(struct cache
*cache
, int argc
, const char **argv
)
2768 copy
= kcalloc(argc
, sizeof(*copy
), GFP_KERNEL
);
2771 for (i
= 0; i
< argc
; i
++) {
2772 copy
[i
] = kstrdup(argv
[i
], GFP_KERNEL
);
2781 cache
->nr_ctr_args
= argc
;
2782 cache
->ctr_args
= copy
;
2787 static int cache_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
2790 struct cache_args
*ca
;
2791 struct cache
*cache
= NULL
;
2793 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
2795 ti
->error
= "Error allocating memory for cache";
2800 r
= parse_cache_args(ca
, argc
, argv
, &ti
->error
);
2804 r
= cache_create(ca
, &cache
);
2808 r
= copy_ctr_args(cache
, argc
- 3, (const char **)argv
+ 3);
2814 ti
->private = cache
;
2816 destroy_cache_args(ca
);
2820 /*----------------------------------------------------------------*/
2822 static int cache_map(struct dm_target
*ti
, struct bio
*bio
)
2824 struct cache
*cache
= ti
->private;
2828 dm_oblock_t block
= get_bio_block(cache
, bio
);
2829 size_t pb_data_size
= get_per_bio_data_size(cache
);
2831 init_per_bio_data(bio
, pb_data_size
);
2832 if (unlikely(from_oblock(block
) >= from_oblock(cache
->origin_blocks
))) {
2834 * This can only occur if the io goes to a partial block at
2835 * the end of the origin device. We don't cache these.
2836 * Just remap to the origin and carry on.
2838 remap_to_origin(cache
, bio
);
2839 accounted_begin(cache
, bio
);
2840 return DM_MAPIO_REMAPPED
;
2843 if (discard_or_flush(bio
)) {
2844 defer_bio(cache
, bio
);
2845 return DM_MAPIO_SUBMITTED
;
2848 r
= map_bio(cache
, bio
, block
, &commit_needed
);
2850 schedule_commit(&cache
->committer
);
2855 static int cache_end_io(struct dm_target
*ti
, struct bio
*bio
,
2856 blk_status_t
*error
)
2858 struct cache
*cache
= ti
->private;
2859 unsigned long flags
;
2860 size_t pb_data_size
= get_per_bio_data_size(cache
);
2861 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
2864 policy_tick(cache
->policy
, false);
2866 spin_lock_irqsave(&cache
->lock
, flags
);
2867 cache
->need_tick_bio
= true;
2868 spin_unlock_irqrestore(&cache
->lock
, flags
);
2871 bio_drop_shared_lock(cache
, bio
);
2872 accounted_complete(cache
, bio
);
2874 return DM_ENDIO_DONE
;
2877 static int write_dirty_bitset(struct cache
*cache
)
2881 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2884 r
= dm_cache_set_dirty_bits(cache
->cmd
, from_cblock(cache
->cache_size
), cache
->dirty_bitset
);
2886 metadata_operation_failed(cache
, "dm_cache_set_dirty_bits", r
);
2891 static int write_discard_bitset(struct cache
*cache
)
2895 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2898 r
= dm_cache_discard_bitset_resize(cache
->cmd
, cache
->discard_block_size
,
2899 cache
->discard_nr_blocks
);
2901 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache
));
2902 metadata_operation_failed(cache
, "dm_cache_discard_bitset_resize", r
);
2906 for (i
= 0; i
< from_dblock(cache
->discard_nr_blocks
); i
++) {
2907 r
= dm_cache_set_discard(cache
->cmd
, to_dblock(i
),
2908 is_discarded(cache
, to_dblock(i
)));
2910 metadata_operation_failed(cache
, "dm_cache_set_discard", r
);
2918 static int write_hints(struct cache
*cache
)
2922 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2925 r
= dm_cache_write_hints(cache
->cmd
, cache
->policy
);
2927 metadata_operation_failed(cache
, "dm_cache_write_hints", r
);
2935 * returns true on success
2937 static bool sync_metadata(struct cache
*cache
)
2941 r1
= write_dirty_bitset(cache
);
2943 DMERR("%s: could not write dirty bitset", cache_device_name(cache
));
2945 r2
= write_discard_bitset(cache
);
2947 DMERR("%s: could not write discard bitset", cache_device_name(cache
));
2951 r3
= write_hints(cache
);
2953 DMERR("%s: could not write hints", cache_device_name(cache
));
2956 * If writing the above metadata failed, we still commit, but don't
2957 * set the clean shutdown flag. This will effectively force every
2958 * dirty bit to be set on reload.
2960 r4
= commit(cache
, !r1
&& !r2
&& !r3
);
2962 DMERR("%s: could not write cache metadata", cache_device_name(cache
));
2964 return !r1
&& !r2
&& !r3
&& !r4
;
2967 static void cache_postsuspend(struct dm_target
*ti
)
2969 struct cache
*cache
= ti
->private;
2971 prevent_background_work(cache
);
2972 BUG_ON(atomic_read(&cache
->nr_io_migrations
));
2974 cancel_delayed_work(&cache
->waker
);
2975 flush_workqueue(cache
->wq
);
2976 WARN_ON(cache
->tracker
.in_flight
);
2979 * If it's a flush suspend there won't be any deferred bios, so this
2982 requeue_deferred_bios(cache
);
2984 if (get_cache_mode(cache
) == CM_WRITE
)
2985 (void) sync_metadata(cache
);
2988 static int load_mapping(void *context
, dm_oblock_t oblock
, dm_cblock_t cblock
,
2989 bool dirty
, uint32_t hint
, bool hint_valid
)
2992 struct cache
*cache
= context
;
2995 set_bit(from_cblock(cblock
), cache
->dirty_bitset
);
2996 atomic_inc(&cache
->nr_dirty
);
2998 clear_bit(from_cblock(cblock
), cache
->dirty_bitset
);
3000 r
= policy_load_mapping(cache
->policy
, oblock
, cblock
, dirty
, hint
, hint_valid
);
3008 * The discard block size in the on disk metadata is not
3009 * neccessarily the same as we're currently using. So we have to
3010 * be careful to only set the discarded attribute if we know it
3011 * covers a complete block of the new size.
3013 struct discard_load_info
{
3014 struct cache
*cache
;
3017 * These blocks are sized using the on disk dblock size, rather
3018 * than the current one.
3020 dm_block_t block_size
;
3021 dm_block_t discard_begin
, discard_end
;
3024 static void discard_load_info_init(struct cache
*cache
,
3025 struct discard_load_info
*li
)
3028 li
->discard_begin
= li
->discard_end
= 0;
3031 static void set_discard_range(struct discard_load_info
*li
)
3035 if (li
->discard_begin
== li
->discard_end
)
3039 * Convert to sectors.
3041 b
= li
->discard_begin
* li
->block_size
;
3042 e
= li
->discard_end
* li
->block_size
;
3045 * Then convert back to the current dblock size.
3047 b
= dm_sector_div_up(b
, li
->cache
->discard_block_size
);
3048 sector_div(e
, li
->cache
->discard_block_size
);
3051 * The origin may have shrunk, so we need to check we're still in
3054 if (e
> from_dblock(li
->cache
->discard_nr_blocks
))
3055 e
= from_dblock(li
->cache
->discard_nr_blocks
);
3058 set_discard(li
->cache
, to_dblock(b
));
3061 static int load_discard(void *context
, sector_t discard_block_size
,
3062 dm_dblock_t dblock
, bool discard
)
3064 struct discard_load_info
*li
= context
;
3066 li
->block_size
= discard_block_size
;
3069 if (from_dblock(dblock
) == li
->discard_end
)
3071 * We're already in a discard range, just extend it.
3073 li
->discard_end
= li
->discard_end
+ 1ULL;
3077 * Emit the old range and start a new one.
3079 set_discard_range(li
);
3080 li
->discard_begin
= from_dblock(dblock
);
3081 li
->discard_end
= li
->discard_begin
+ 1ULL;
3084 set_discard_range(li
);
3085 li
->discard_begin
= li
->discard_end
= 0;
3091 static dm_cblock_t
get_cache_dev_size(struct cache
*cache
)
3093 sector_t size
= get_dev_size(cache
->cache_dev
);
3094 (void) sector_div(size
, cache
->sectors_per_block
);
3095 return to_cblock(size
);
3098 static bool can_resize(struct cache
*cache
, dm_cblock_t new_size
)
3100 if (from_cblock(new_size
) > from_cblock(cache
->cache_size
)) {
3102 DMERR("%s: unable to extend cache due to missing cache table reload",
3103 cache_device_name(cache
));
3109 * We can't drop a dirty block when shrinking the cache.
3111 while (from_cblock(new_size
) < from_cblock(cache
->cache_size
)) {
3112 new_size
= to_cblock(from_cblock(new_size
) + 1);
3113 if (is_dirty(cache
, new_size
)) {
3114 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3115 cache_device_name(cache
),
3116 (unsigned long long) from_cblock(new_size
));
3124 static int resize_cache_dev(struct cache
*cache
, dm_cblock_t new_size
)
3128 r
= dm_cache_resize(cache
->cmd
, new_size
);
3130 DMERR("%s: could not resize cache metadata", cache_device_name(cache
));
3131 metadata_operation_failed(cache
, "dm_cache_resize", r
);
3135 set_cache_size(cache
, new_size
);
3140 static int cache_preresume(struct dm_target
*ti
)
3143 struct cache
*cache
= ti
->private;
3144 dm_cblock_t csize
= get_cache_dev_size(cache
);
3147 * Check to see if the cache has resized.
3149 if (!cache
->sized
) {
3150 r
= resize_cache_dev(cache
, csize
);
3154 cache
->sized
= true;
3156 } else if (csize
!= cache
->cache_size
) {
3157 if (!can_resize(cache
, csize
))
3160 r
= resize_cache_dev(cache
, csize
);
3165 if (!cache
->loaded_mappings
) {
3166 r
= dm_cache_load_mappings(cache
->cmd
, cache
->policy
,
3167 load_mapping
, cache
);
3169 DMERR("%s: could not load cache mappings", cache_device_name(cache
));
3170 metadata_operation_failed(cache
, "dm_cache_load_mappings", r
);
3174 cache
->loaded_mappings
= true;
3177 if (!cache
->loaded_discards
) {
3178 struct discard_load_info li
;
3181 * The discard bitset could have been resized, or the
3182 * discard block size changed. To be safe we start by
3183 * setting every dblock to not discarded.
3185 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
3187 discard_load_info_init(cache
, &li
);
3188 r
= dm_cache_load_discards(cache
->cmd
, load_discard
, &li
);
3190 DMERR("%s: could not load origin discards", cache_device_name(cache
));
3191 metadata_operation_failed(cache
, "dm_cache_load_discards", r
);
3194 set_discard_range(&li
);
3196 cache
->loaded_discards
= true;
3202 static void cache_resume(struct dm_target
*ti
)
3204 struct cache
*cache
= ti
->private;
3206 cache
->need_tick_bio
= true;
3207 allow_background_work(cache
);
3208 do_waker(&cache
->waker
.work
);
3214 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3215 * <cache block size> <#used cache blocks>/<#total cache blocks>
3216 * <#read hits> <#read misses> <#write hits> <#write misses>
3217 * <#demotions> <#promotions> <#dirty>
3218 * <#features> <features>*
3219 * <#core args> <core args>
3220 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3222 static void cache_status(struct dm_target
*ti
, status_type_t type
,
3223 unsigned status_flags
, char *result
, unsigned maxlen
)
3228 dm_block_t nr_free_blocks_metadata
= 0;
3229 dm_block_t nr_blocks_metadata
= 0;
3230 char buf
[BDEVNAME_SIZE
];
3231 struct cache
*cache
= ti
->private;
3232 dm_cblock_t residency
;
3236 case STATUSTYPE_INFO
:
3237 if (get_cache_mode(cache
) == CM_FAIL
) {
3242 /* Commit to ensure statistics aren't out-of-date */
3243 if (!(status_flags
& DM_STATUS_NOFLUSH_FLAG
) && !dm_suspended(ti
))
3244 (void) commit(cache
, false);
3246 r
= dm_cache_get_free_metadata_block_count(cache
->cmd
, &nr_free_blocks_metadata
);
3248 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3249 cache_device_name(cache
), r
);
3253 r
= dm_cache_get_metadata_dev_size(cache
->cmd
, &nr_blocks_metadata
);
3255 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3256 cache_device_name(cache
), r
);
3260 residency
= policy_residency(cache
->policy
);
3262 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3263 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE
,
3264 (unsigned long long)(nr_blocks_metadata
- nr_free_blocks_metadata
),
3265 (unsigned long long)nr_blocks_metadata
,
3266 (unsigned long long)cache
->sectors_per_block
,
3267 (unsigned long long) from_cblock(residency
),
3268 (unsigned long long) from_cblock(cache
->cache_size
),
3269 (unsigned) atomic_read(&cache
->stats
.read_hit
),
3270 (unsigned) atomic_read(&cache
->stats
.read_miss
),
3271 (unsigned) atomic_read(&cache
->stats
.write_hit
),
3272 (unsigned) atomic_read(&cache
->stats
.write_miss
),
3273 (unsigned) atomic_read(&cache
->stats
.demotion
),
3274 (unsigned) atomic_read(&cache
->stats
.promotion
),
3275 (unsigned long) atomic_read(&cache
->nr_dirty
));
3277 if (cache
->features
.metadata_version
== 2)
3278 DMEMIT("2 metadata2 ");
3282 if (writethrough_mode(&cache
->features
))
3283 DMEMIT("writethrough ");
3285 else if (passthrough_mode(&cache
->features
))
3286 DMEMIT("passthrough ");
3288 else if (writeback_mode(&cache
->features
))
3289 DMEMIT("writeback ");
3292 DMERR("%s: internal error: unknown io mode: %d",
3293 cache_device_name(cache
), (int) cache
->features
.io_mode
);
3297 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache
->migration_threshold
);
3299 DMEMIT("%s ", dm_cache_policy_get_name(cache
->policy
));
3301 r
= policy_emit_config_values(cache
->policy
, result
, maxlen
, &sz
);
3303 DMERR("%s: policy_emit_config_values returned %d",
3304 cache_device_name(cache
), r
);
3307 if (get_cache_mode(cache
) == CM_READ_ONLY
)
3312 r
= dm_cache_metadata_needs_check(cache
->cmd
, &needs_check
);
3314 if (r
|| needs_check
)
3315 DMEMIT("needs_check ");
3321 case STATUSTYPE_TABLE
:
3322 format_dev_t(buf
, cache
->metadata_dev
->bdev
->bd_dev
);
3324 format_dev_t(buf
, cache
->cache_dev
->bdev
->bd_dev
);
3326 format_dev_t(buf
, cache
->origin_dev
->bdev
->bd_dev
);
3329 for (i
= 0; i
< cache
->nr_ctr_args
- 1; i
++)
3330 DMEMIT(" %s", cache
->ctr_args
[i
]);
3331 if (cache
->nr_ctr_args
)
3332 DMEMIT(" %s", cache
->ctr_args
[cache
->nr_ctr_args
- 1]);
3342 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3343 * the one-past-the-end value.
3345 struct cblock_range
{
3351 * A cache block range can take two forms:
3353 * i) A single cblock, eg. '3456'
3354 * ii) A begin and end cblock with a dash between, eg. 123-234
3356 static int parse_cblock_range(struct cache
*cache
, const char *str
,
3357 struct cblock_range
*result
)
3364 * Try and parse form (ii) first.
3366 r
= sscanf(str
, "%llu-%llu%c", &b
, &e
, &dummy
);
3371 result
->begin
= to_cblock(b
);
3372 result
->end
= to_cblock(e
);
3377 * That didn't work, try form (i).
3379 r
= sscanf(str
, "%llu%c", &b
, &dummy
);
3384 result
->begin
= to_cblock(b
);
3385 result
->end
= to_cblock(from_cblock(result
->begin
) + 1u);
3389 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache
), str
);
3393 static int validate_cblock_range(struct cache
*cache
, struct cblock_range
*range
)
3395 uint64_t b
= from_cblock(range
->begin
);
3396 uint64_t e
= from_cblock(range
->end
);
3397 uint64_t n
= from_cblock(cache
->cache_size
);
3400 DMERR("%s: begin cblock out of range: %llu >= %llu",
3401 cache_device_name(cache
), b
, n
);
3406 DMERR("%s: end cblock out of range: %llu > %llu",
3407 cache_device_name(cache
), e
, n
);
3412 DMERR("%s: invalid cblock range: %llu >= %llu",
3413 cache_device_name(cache
), b
, e
);
3420 static inline dm_cblock_t
cblock_succ(dm_cblock_t b
)
3422 return to_cblock(from_cblock(b
) + 1);
3425 static int request_invalidation(struct cache
*cache
, struct cblock_range
*range
)
3430 * We don't need to do any locking here because we know we're in
3431 * passthrough mode. There's is potential for a race between an
3432 * invalidation triggered by an io and an invalidation message. This
3433 * is harmless, we must not worry if the policy call fails.
3435 while (range
->begin
!= range
->end
) {
3436 r
= invalidate_cblock(cache
, range
->begin
);
3440 range
->begin
= cblock_succ(range
->begin
);
3443 cache
->commit_requested
= true;
3447 static int process_invalidate_cblocks_message(struct cache
*cache
, unsigned count
,
3448 const char **cblock_ranges
)
3452 struct cblock_range range
;
3454 if (!passthrough_mode(&cache
->features
)) {
3455 DMERR("%s: cache has to be in passthrough mode for invalidation",
3456 cache_device_name(cache
));
3460 for (i
= 0; i
< count
; i
++) {
3461 r
= parse_cblock_range(cache
, cblock_ranges
[i
], &range
);
3465 r
= validate_cblock_range(cache
, &range
);
3470 * Pass begin and end origin blocks to the worker and wake it.
3472 r
= request_invalidation(cache
, &range
);
3484 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3486 * The key migration_threshold is supported by the cache target core.
3488 static int cache_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
3490 struct cache
*cache
= ti
->private;
3495 if (get_cache_mode(cache
) >= CM_READ_ONLY
) {
3496 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3497 cache_device_name(cache
));
3501 if (!strcasecmp(argv
[0], "invalidate_cblocks"))
3502 return process_invalidate_cblocks_message(cache
, argc
- 1, (const char **) argv
+ 1);
3507 return set_config_value(cache
, argv
[0], argv
[1]);
3510 static int cache_iterate_devices(struct dm_target
*ti
,
3511 iterate_devices_callout_fn fn
, void *data
)
3514 struct cache
*cache
= ti
->private;
3516 r
= fn(ti
, cache
->cache_dev
, 0, get_dev_size(cache
->cache_dev
), data
);
3518 r
= fn(ti
, cache
->origin_dev
, 0, ti
->len
, data
);
3523 static void set_discard_limits(struct cache
*cache
, struct queue_limits
*limits
)
3526 * FIXME: these limits may be incompatible with the cache device
3528 limits
->max_discard_sectors
= min_t(sector_t
, cache
->discard_block_size
* 1024,
3529 cache
->origin_sectors
);
3530 limits
->discard_granularity
= cache
->discard_block_size
<< SECTOR_SHIFT
;
3533 static void cache_io_hints(struct dm_target
*ti
, struct queue_limits
*limits
)
3535 struct cache
*cache
= ti
->private;
3536 uint64_t io_opt_sectors
= limits
->io_opt
>> SECTOR_SHIFT
;
3539 * If the system-determined stacked limits are compatible with the
3540 * cache's blocksize (io_opt is a factor) do not override them.
3542 if (io_opt_sectors
< cache
->sectors_per_block
||
3543 do_div(io_opt_sectors
, cache
->sectors_per_block
)) {
3544 blk_limits_io_min(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3545 blk_limits_io_opt(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3547 set_discard_limits(cache
, limits
);
3550 /*----------------------------------------------------------------*/
3552 static struct target_type cache_target
= {
3554 .version
= {2, 0, 0},
3555 .module
= THIS_MODULE
,
3559 .end_io
= cache_end_io
,
3560 .postsuspend
= cache_postsuspend
,
3561 .preresume
= cache_preresume
,
3562 .resume
= cache_resume
,
3563 .status
= cache_status
,
3564 .message
= cache_message
,
3565 .iterate_devices
= cache_iterate_devices
,
3566 .io_hints
= cache_io_hints
,
3569 static int __init
dm_cache_init(void)
3573 migration_cache
= KMEM_CACHE(dm_cache_migration
, 0);
3574 if (!migration_cache
)
3577 r
= dm_register_target(&cache_target
);
3579 DMERR("cache target registration failed: %d", r
);
3580 kmem_cache_destroy(migration_cache
);
3587 static void __exit
dm_cache_exit(void)
3589 dm_unregister_target(&cache_target
);
3590 kmem_cache_destroy(migration_cache
);
3593 module_init(dm_cache_init
);
3594 module_exit(dm_cache_exit
);
3596 MODULE_DESCRIPTION(DM_NAME
" cache target");
3597 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3598 MODULE_LICENSE("GPL");