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
;
455 mempool_t
*migration_pool
;
457 struct dm_cache_policy
*policy
;
458 unsigned policy_nr_args
;
460 bool need_tick_bio
:1;
463 bool commit_requested
:1;
464 bool loaded_mappings
:1;
465 bool loaded_discards
:1;
468 * Cache features such as write-through.
470 struct cache_features features
;
472 struct cache_stats stats
;
475 * Invalidation fields.
477 spinlock_t invalidation_lock
;
478 struct list_head invalidation_requests
;
480 struct io_tracker tracker
;
482 struct work_struct commit_ws
;
483 struct batcher committer
;
485 struct rw_semaphore background_work_lock
;
488 struct per_bio_data
{
491 struct dm_bio_prison_cell_v2
*cell
;
492 struct dm_hook_info hook_info
;
496 * writethrough fields. These MUST remain at the end of this
497 * structure and the 'cache' member must be the first as it
498 * is used to determine the offset of the writethrough fields.
502 struct dm_bio_details bio_details
;
505 struct dm_cache_migration
{
506 struct continuation k
;
509 struct policy_work
*op
;
510 struct bio
*overwrite_bio
;
511 struct dm_bio_prison_cell_v2
*cell
;
513 dm_cblock_t invalidate_cblock
;
514 dm_oblock_t invalidate_oblock
;
517 /*----------------------------------------------------------------*/
519 static bool writethrough_mode(struct cache
*cache
)
521 return cache
->features
.io_mode
== CM_IO_WRITETHROUGH
;
524 static bool writeback_mode(struct cache
*cache
)
526 return cache
->features
.io_mode
== CM_IO_WRITEBACK
;
529 static inline bool passthrough_mode(struct cache
*cache
)
531 return unlikely(cache
->features
.io_mode
== CM_IO_PASSTHROUGH
);
534 /*----------------------------------------------------------------*/
536 static void wake_deferred_bio_worker(struct cache
*cache
)
538 queue_work(cache
->wq
, &cache
->deferred_bio_worker
);
541 static void wake_deferred_writethrough_worker(struct cache
*cache
)
543 queue_work(cache
->wq
, &cache
->deferred_writethrough_worker
);
546 static void wake_migration_worker(struct cache
*cache
)
548 if (passthrough_mode(cache
))
551 queue_work(cache
->wq
, &cache
->migration_worker
);
554 /*----------------------------------------------------------------*/
556 static struct dm_bio_prison_cell_v2
*alloc_prison_cell(struct cache
*cache
)
558 return dm_bio_prison_alloc_cell_v2(cache
->prison
, GFP_NOWAIT
);
561 static void free_prison_cell(struct cache
*cache
, struct dm_bio_prison_cell_v2
*cell
)
563 dm_bio_prison_free_cell_v2(cache
->prison
, cell
);
566 static struct dm_cache_migration
*alloc_migration(struct cache
*cache
)
568 struct dm_cache_migration
*mg
;
570 mg
= mempool_alloc(cache
->migration_pool
, GFP_NOWAIT
);
573 atomic_inc(&mg
->cache
->nr_allocated_migrations
);
579 static void free_migration(struct dm_cache_migration
*mg
)
581 struct cache
*cache
= mg
->cache
;
583 if (atomic_dec_and_test(&cache
->nr_allocated_migrations
))
584 wake_up(&cache
->migration_wait
);
586 mempool_free(mg
, cache
->migration_pool
);
589 /*----------------------------------------------------------------*/
591 static inline dm_oblock_t
oblock_succ(dm_oblock_t b
)
593 return to_oblock(from_oblock(b
) + 1ull);
596 static void build_key(dm_oblock_t begin
, dm_oblock_t end
, struct dm_cell_key_v2
*key
)
600 key
->block_begin
= from_oblock(begin
);
601 key
->block_end
= from_oblock(end
);
605 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
606 * level 1 which prevents *both* READs and WRITEs.
608 #define WRITE_LOCK_LEVEL 0
609 #define READ_WRITE_LOCK_LEVEL 1
611 static unsigned lock_level(struct bio
*bio
)
613 return bio_data_dir(bio
) == WRITE
?
615 READ_WRITE_LOCK_LEVEL
;
618 /*----------------------------------------------------------------
620 *--------------------------------------------------------------*/
623 * If using writeback, leave out struct per_bio_data's writethrough fields.
625 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
626 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
628 static size_t get_per_bio_data_size(struct cache
*cache
)
630 return writethrough_mode(cache
) ? PB_DATA_SIZE_WT
: PB_DATA_SIZE_WB
;
633 static struct per_bio_data
*get_per_bio_data(struct bio
*bio
, size_t data_size
)
635 struct per_bio_data
*pb
= dm_per_bio_data(bio
, data_size
);
640 static struct per_bio_data
*init_per_bio_data(struct bio
*bio
, size_t data_size
)
642 struct per_bio_data
*pb
= get_per_bio_data(bio
, data_size
);
645 pb
->req_nr
= dm_bio_get_target_bio_nr(bio
);
652 /*----------------------------------------------------------------*/
654 static void defer_bio(struct cache
*cache
, struct bio
*bio
)
658 spin_lock_irqsave(&cache
->lock
, flags
);
659 bio_list_add(&cache
->deferred_bios
, bio
);
660 spin_unlock_irqrestore(&cache
->lock
, flags
);
662 wake_deferred_bio_worker(cache
);
665 static void defer_bios(struct cache
*cache
, struct bio_list
*bios
)
669 spin_lock_irqsave(&cache
->lock
, flags
);
670 bio_list_merge(&cache
->deferred_bios
, bios
);
672 spin_unlock_irqrestore(&cache
->lock
, flags
);
674 wake_deferred_bio_worker(cache
);
677 /*----------------------------------------------------------------*/
679 static bool bio_detain_shared(struct cache
*cache
, dm_oblock_t oblock
, struct bio
*bio
)
683 struct per_bio_data
*pb
;
684 struct dm_cell_key_v2 key
;
685 dm_oblock_t end
= to_oblock(from_oblock(oblock
) + 1ULL);
686 struct dm_bio_prison_cell_v2
*cell_prealloc
, *cell
;
688 cell_prealloc
= alloc_prison_cell(cache
); /* FIXME: allow wait if calling from worker */
689 if (!cell_prealloc
) {
690 defer_bio(cache
, bio
);
694 build_key(oblock
, end
, &key
);
695 r
= dm_cell_get_v2(cache
->prison
, &key
, lock_level(bio
), bio
, cell_prealloc
, &cell
);
698 * Failed to get the lock.
700 free_prison_cell(cache
, cell_prealloc
);
704 if (cell
!= cell_prealloc
)
705 free_prison_cell(cache
, cell_prealloc
);
707 pb_size
= get_per_bio_data_size(cache
);
708 pb
= get_per_bio_data(bio
, pb_size
);
714 /*----------------------------------------------------------------*/
716 static bool is_dirty(struct cache
*cache
, dm_cblock_t b
)
718 return test_bit(from_cblock(b
), cache
->dirty_bitset
);
721 static void set_dirty(struct cache
*cache
, dm_cblock_t cblock
)
723 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
724 atomic_inc(&cache
->nr_dirty
);
725 policy_set_dirty(cache
->policy
, cblock
);
730 * These two are called when setting after migrations to force the policy
731 * and dirty bitset to be in sync.
733 static void force_set_dirty(struct cache
*cache
, dm_cblock_t cblock
)
735 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
))
736 atomic_inc(&cache
->nr_dirty
);
737 policy_set_dirty(cache
->policy
, cblock
);
740 static void force_clear_dirty(struct cache
*cache
, dm_cblock_t cblock
)
742 if (test_and_clear_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
743 if (atomic_dec_return(&cache
->nr_dirty
) == 0)
744 dm_table_event(cache
->ti
->table
);
747 policy_clear_dirty(cache
->policy
, cblock
);
750 /*----------------------------------------------------------------*/
752 static bool block_size_is_power_of_two(struct cache
*cache
)
754 return cache
->sectors_per_block_shift
>= 0;
757 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
758 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
761 static dm_block_t
block_div(dm_block_t b
, uint32_t n
)
768 static dm_block_t
oblocks_per_dblock(struct cache
*cache
)
770 dm_block_t oblocks
= cache
->discard_block_size
;
772 if (block_size_is_power_of_two(cache
))
773 oblocks
>>= cache
->sectors_per_block_shift
;
775 oblocks
= block_div(oblocks
, cache
->sectors_per_block
);
780 static dm_dblock_t
oblock_to_dblock(struct cache
*cache
, dm_oblock_t oblock
)
782 return to_dblock(block_div(from_oblock(oblock
),
783 oblocks_per_dblock(cache
)));
786 static void set_discard(struct cache
*cache
, dm_dblock_t b
)
790 BUG_ON(from_dblock(b
) >= from_dblock(cache
->discard_nr_blocks
));
791 atomic_inc(&cache
->stats
.discard_count
);
793 spin_lock_irqsave(&cache
->lock
, flags
);
794 set_bit(from_dblock(b
), cache
->discard_bitset
);
795 spin_unlock_irqrestore(&cache
->lock
, flags
);
798 static void clear_discard(struct cache
*cache
, dm_dblock_t b
)
802 spin_lock_irqsave(&cache
->lock
, flags
);
803 clear_bit(from_dblock(b
), cache
->discard_bitset
);
804 spin_unlock_irqrestore(&cache
->lock
, flags
);
807 static bool is_discarded(struct cache
*cache
, dm_dblock_t b
)
812 spin_lock_irqsave(&cache
->lock
, flags
);
813 r
= test_bit(from_dblock(b
), cache
->discard_bitset
);
814 spin_unlock_irqrestore(&cache
->lock
, flags
);
819 static bool is_discarded_oblock(struct cache
*cache
, dm_oblock_t b
)
824 spin_lock_irqsave(&cache
->lock
, flags
);
825 r
= test_bit(from_dblock(oblock_to_dblock(cache
, b
)),
826 cache
->discard_bitset
);
827 spin_unlock_irqrestore(&cache
->lock
, flags
);
832 /*----------------------------------------------------------------
834 *--------------------------------------------------------------*/
835 static void remap_to_origin(struct cache
*cache
, struct bio
*bio
)
837 bio_set_dev(bio
, cache
->origin_dev
->bdev
);
840 static void remap_to_cache(struct cache
*cache
, struct bio
*bio
,
843 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
844 sector_t block
= from_cblock(cblock
);
846 bio_set_dev(bio
, cache
->cache_dev
->bdev
);
847 if (!block_size_is_power_of_two(cache
))
848 bio
->bi_iter
.bi_sector
=
849 (block
* cache
->sectors_per_block
) +
850 sector_div(bi_sector
, cache
->sectors_per_block
);
852 bio
->bi_iter
.bi_sector
=
853 (block
<< cache
->sectors_per_block_shift
) |
854 (bi_sector
& (cache
->sectors_per_block
- 1));
857 static void check_if_tick_bio_needed(struct cache
*cache
, struct bio
*bio
)
860 size_t pb_data_size
= get_per_bio_data_size(cache
);
861 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
863 spin_lock_irqsave(&cache
->lock
, flags
);
864 if (cache
->need_tick_bio
&& !op_is_flush(bio
->bi_opf
) &&
865 bio_op(bio
) != REQ_OP_DISCARD
) {
867 cache
->need_tick_bio
= false;
869 spin_unlock_irqrestore(&cache
->lock
, flags
);
872 static void __remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
873 dm_oblock_t oblock
, bool bio_has_pbd
)
876 check_if_tick_bio_needed(cache
, bio
);
877 remap_to_origin(cache
, bio
);
878 if (bio_data_dir(bio
) == WRITE
)
879 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
882 static void remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
885 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
886 __remap_to_origin_clear_discard(cache
, bio
, oblock
, true);
889 static void remap_to_cache_dirty(struct cache
*cache
, struct bio
*bio
,
890 dm_oblock_t oblock
, dm_cblock_t cblock
)
892 check_if_tick_bio_needed(cache
, bio
);
893 remap_to_cache(cache
, bio
, cblock
);
894 if (bio_data_dir(bio
) == WRITE
) {
895 set_dirty(cache
, cblock
);
896 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
900 static dm_oblock_t
get_bio_block(struct cache
*cache
, struct bio
*bio
)
902 sector_t block_nr
= bio
->bi_iter
.bi_sector
;
904 if (!block_size_is_power_of_two(cache
))
905 (void) sector_div(block_nr
, cache
->sectors_per_block
);
907 block_nr
>>= cache
->sectors_per_block_shift
;
909 return to_oblock(block_nr
);
912 static bool accountable_bio(struct cache
*cache
, struct bio
*bio
)
914 return bio_op(bio
) != REQ_OP_DISCARD
;
917 static void accounted_begin(struct cache
*cache
, struct bio
*bio
)
919 size_t pb_data_size
= get_per_bio_data_size(cache
);
920 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
922 if (accountable_bio(cache
, bio
)) {
923 pb
->len
= bio_sectors(bio
);
924 iot_io_begin(&cache
->tracker
, pb
->len
);
928 static void accounted_complete(struct cache
*cache
, struct bio
*bio
)
930 size_t pb_data_size
= get_per_bio_data_size(cache
);
931 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
933 iot_io_end(&cache
->tracker
, pb
->len
);
936 static void accounted_request(struct cache
*cache
, struct bio
*bio
)
938 accounted_begin(cache
, bio
);
939 generic_make_request(bio
);
942 static void issue_op(struct bio
*bio
, void *context
)
944 struct cache
*cache
= context
;
945 accounted_request(cache
, bio
);
948 static void defer_writethrough_bio(struct cache
*cache
, struct bio
*bio
)
952 spin_lock_irqsave(&cache
->lock
, flags
);
953 bio_list_add(&cache
->deferred_writethrough_bios
, bio
);
954 spin_unlock_irqrestore(&cache
->lock
, flags
);
956 wake_deferred_writethrough_worker(cache
);
959 static void writethrough_endio(struct bio
*bio
)
961 struct per_bio_data
*pb
= get_per_bio_data(bio
, PB_DATA_SIZE_WT
);
963 dm_unhook_bio(&pb
->hook_info
, bio
);
965 if (bio
->bi_status
) {
970 dm_bio_restore(&pb
->bio_details
, bio
);
971 remap_to_cache(pb
->cache
, bio
, pb
->cblock
);
974 * We can't issue this bio directly, since we're in interrupt
975 * context. So it gets put on a bio list for processing by the
978 defer_writethrough_bio(pb
->cache
, bio
);
982 * When running in writethrough mode we need to send writes to clean blocks
983 * to both the cache and origin devices. Clone the bio and send them in parallel.
985 static void remap_to_origin_and_cache(struct cache
*cache
, struct bio
*bio
,
986 dm_oblock_t oblock
, dm_cblock_t cblock
)
988 struct bio
*origin_bio
= bio_clone_fast(bio
, GFP_NOIO
, cache
->bs
);
992 bio_chain(origin_bio
, bio
);
994 * Passing false to __remap_to_origin_clear_discard() skips
995 * all code that might use per_bio_data (since clone doesn't have it)
997 __remap_to_origin_clear_discard(cache
, origin_bio
, oblock
, false);
998 submit_bio(origin_bio
);
1000 remap_to_cache(cache
, bio
, cblock
);
1003 /*----------------------------------------------------------------
1005 *--------------------------------------------------------------*/
1006 static enum cache_metadata_mode
get_cache_mode(struct cache
*cache
)
1008 return cache
->features
.mode
;
1011 static const char *cache_device_name(struct cache
*cache
)
1013 return dm_device_name(dm_table_get_md(cache
->ti
->table
));
1016 static void notify_mode_switch(struct cache
*cache
, enum cache_metadata_mode mode
)
1018 const char *descs
[] = {
1024 dm_table_event(cache
->ti
->table
);
1025 DMINFO("%s: switching cache to %s mode",
1026 cache_device_name(cache
), descs
[(int)mode
]);
1029 static void set_cache_mode(struct cache
*cache
, enum cache_metadata_mode new_mode
)
1032 enum cache_metadata_mode old_mode
= get_cache_mode(cache
);
1034 if (dm_cache_metadata_needs_check(cache
->cmd
, &needs_check
)) {
1035 DMERR("%s: unable to read needs_check flag, setting failure mode.",
1036 cache_device_name(cache
));
1040 if (new_mode
== CM_WRITE
&& needs_check
) {
1041 DMERR("%s: unable to switch cache to write mode until repaired.",
1042 cache_device_name(cache
));
1043 if (old_mode
!= new_mode
)
1044 new_mode
= old_mode
;
1046 new_mode
= CM_READ_ONLY
;
1049 /* Never move out of fail mode */
1050 if (old_mode
== CM_FAIL
)
1056 dm_cache_metadata_set_read_only(cache
->cmd
);
1060 dm_cache_metadata_set_read_write(cache
->cmd
);
1064 cache
->features
.mode
= new_mode
;
1066 if (new_mode
!= old_mode
)
1067 notify_mode_switch(cache
, new_mode
);
1070 static void abort_transaction(struct cache
*cache
)
1072 const char *dev_name
= cache_device_name(cache
);
1074 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1077 if (dm_cache_metadata_set_needs_check(cache
->cmd
)) {
1078 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name
);
1079 set_cache_mode(cache
, CM_FAIL
);
1082 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name
);
1083 if (dm_cache_metadata_abort(cache
->cmd
)) {
1084 DMERR("%s: failed to abort metadata transaction", dev_name
);
1085 set_cache_mode(cache
, CM_FAIL
);
1089 static void metadata_operation_failed(struct cache
*cache
, const char *op
, int r
)
1091 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1092 cache_device_name(cache
), op
, r
);
1093 abort_transaction(cache
);
1094 set_cache_mode(cache
, CM_READ_ONLY
);
1097 /*----------------------------------------------------------------*/
1099 static void load_stats(struct cache
*cache
)
1101 struct dm_cache_statistics stats
;
1103 dm_cache_metadata_get_stats(cache
->cmd
, &stats
);
1104 atomic_set(&cache
->stats
.read_hit
, stats
.read_hits
);
1105 atomic_set(&cache
->stats
.read_miss
, stats
.read_misses
);
1106 atomic_set(&cache
->stats
.write_hit
, stats
.write_hits
);
1107 atomic_set(&cache
->stats
.write_miss
, stats
.write_misses
);
1110 static void save_stats(struct cache
*cache
)
1112 struct dm_cache_statistics stats
;
1114 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1117 stats
.read_hits
= atomic_read(&cache
->stats
.read_hit
);
1118 stats
.read_misses
= atomic_read(&cache
->stats
.read_miss
);
1119 stats
.write_hits
= atomic_read(&cache
->stats
.write_hit
);
1120 stats
.write_misses
= atomic_read(&cache
->stats
.write_miss
);
1122 dm_cache_metadata_set_stats(cache
->cmd
, &stats
);
1125 static void update_stats(struct cache_stats
*stats
, enum policy_operation op
)
1128 case POLICY_PROMOTE
:
1129 atomic_inc(&stats
->promotion
);
1133 atomic_inc(&stats
->demotion
);
1136 case POLICY_WRITEBACK
:
1137 atomic_inc(&stats
->writeback
);
1142 /*----------------------------------------------------------------
1143 * Migration processing
1145 * Migration covers moving data from the origin device to the cache, or
1147 *--------------------------------------------------------------*/
1149 static void inc_io_migrations(struct cache
*cache
)
1151 atomic_inc(&cache
->nr_io_migrations
);
1154 static void dec_io_migrations(struct cache
*cache
)
1156 atomic_dec(&cache
->nr_io_migrations
);
1159 static bool discard_or_flush(struct bio
*bio
)
1161 return bio_op(bio
) == REQ_OP_DISCARD
|| op_is_flush(bio
->bi_opf
);
1164 static void calc_discard_block_range(struct cache
*cache
, struct bio
*bio
,
1165 dm_dblock_t
*b
, dm_dblock_t
*e
)
1167 sector_t sb
= bio
->bi_iter
.bi_sector
;
1168 sector_t se
= bio_end_sector(bio
);
1170 *b
= to_dblock(dm_sector_div_up(sb
, cache
->discard_block_size
));
1172 if (se
- sb
< cache
->discard_block_size
)
1175 *e
= to_dblock(block_div(se
, cache
->discard_block_size
));
1178 /*----------------------------------------------------------------*/
1180 static void prevent_background_work(struct cache
*cache
)
1183 down_write(&cache
->background_work_lock
);
1187 static void allow_background_work(struct cache
*cache
)
1190 up_write(&cache
->background_work_lock
);
1194 static bool background_work_begin(struct cache
*cache
)
1199 r
= down_read_trylock(&cache
->background_work_lock
);
1205 static void background_work_end(struct cache
*cache
)
1208 up_read(&cache
->background_work_lock
);
1212 /*----------------------------------------------------------------*/
1214 static bool bio_writes_complete_block(struct cache
*cache
, struct bio
*bio
)
1216 return (bio_data_dir(bio
) == WRITE
) &&
1217 (bio
->bi_iter
.bi_size
== (cache
->sectors_per_block
<< SECTOR_SHIFT
));
1220 static bool optimisable_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
)
1222 return writeback_mode(cache
) &&
1223 (is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
));
1226 static void quiesce(struct dm_cache_migration
*mg
,
1227 void (*continuation
)(struct work_struct
*))
1229 init_continuation(&mg
->k
, continuation
);
1230 dm_cell_quiesce_v2(mg
->cache
->prison
, mg
->cell
, &mg
->k
.ws
);
1233 static struct dm_cache_migration
*ws_to_mg(struct work_struct
*ws
)
1235 struct continuation
*k
= container_of(ws
, struct continuation
, ws
);
1236 return container_of(k
, struct dm_cache_migration
, k
);
1239 static void copy_complete(int read_err
, unsigned long write_err
, void *context
)
1241 struct dm_cache_migration
*mg
= container_of(context
, struct dm_cache_migration
, k
);
1243 if (read_err
|| write_err
)
1244 mg
->k
.input
= BLK_STS_IOERR
;
1246 queue_continuation(mg
->cache
->wq
, &mg
->k
);
1249 static int copy(struct dm_cache_migration
*mg
, bool promote
)
1252 struct dm_io_region o_region
, c_region
;
1253 struct cache
*cache
= mg
->cache
;
1255 o_region
.bdev
= cache
->origin_dev
->bdev
;
1256 o_region
.sector
= from_oblock(mg
->op
->oblock
) * cache
->sectors_per_block
;
1257 o_region
.count
= cache
->sectors_per_block
;
1259 c_region
.bdev
= cache
->cache_dev
->bdev
;
1260 c_region
.sector
= from_cblock(mg
->op
->cblock
) * cache
->sectors_per_block
;
1261 c_region
.count
= cache
->sectors_per_block
;
1264 r
= dm_kcopyd_copy(cache
->copier
, &o_region
, 1, &c_region
, 0, copy_complete
, &mg
->k
);
1266 r
= dm_kcopyd_copy(cache
->copier
, &c_region
, 1, &o_region
, 0, copy_complete
, &mg
->k
);
1271 static void bio_drop_shared_lock(struct cache
*cache
, struct bio
*bio
)
1273 size_t pb_data_size
= get_per_bio_data_size(cache
);
1274 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1276 if (pb
->cell
&& dm_cell_put_v2(cache
->prison
, pb
->cell
))
1277 free_prison_cell(cache
, pb
->cell
);
1281 static void overwrite_endio(struct bio
*bio
)
1283 struct dm_cache_migration
*mg
= bio
->bi_private
;
1284 struct cache
*cache
= mg
->cache
;
1285 size_t pb_data_size
= get_per_bio_data_size(cache
);
1286 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1288 dm_unhook_bio(&pb
->hook_info
, bio
);
1291 mg
->k
.input
= bio
->bi_status
;
1293 queue_continuation(mg
->cache
->wq
, &mg
->k
);
1296 static void overwrite(struct dm_cache_migration
*mg
,
1297 void (*continuation
)(struct work_struct
*))
1299 struct bio
*bio
= mg
->overwrite_bio
;
1300 size_t pb_data_size
= get_per_bio_data_size(mg
->cache
);
1301 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1303 dm_hook_bio(&pb
->hook_info
, bio
, overwrite_endio
, mg
);
1306 * The overwrite bio is part of the copy operation, as such it does
1307 * not set/clear discard or dirty flags.
1309 if (mg
->op
->op
== POLICY_PROMOTE
)
1310 remap_to_cache(mg
->cache
, bio
, mg
->op
->cblock
);
1312 remap_to_origin(mg
->cache
, bio
);
1314 init_continuation(&mg
->k
, continuation
);
1315 accounted_request(mg
->cache
, bio
);
1321 * 1) exclusive lock preventing WRITEs
1323 * 3) copy or issue overwrite bio
1324 * 4) upgrade to exclusive lock preventing READs and WRITEs
1326 * 6) update metadata and commit
1329 static void mg_complete(struct dm_cache_migration
*mg
, bool success
)
1331 struct bio_list bios
;
1332 struct cache
*cache
= mg
->cache
;
1333 struct policy_work
*op
= mg
->op
;
1334 dm_cblock_t cblock
= op
->cblock
;
1337 update_stats(&cache
->stats
, op
->op
);
1340 case POLICY_PROMOTE
:
1341 clear_discard(cache
, oblock_to_dblock(cache
, op
->oblock
));
1342 policy_complete_background_work(cache
->policy
, op
, success
);
1344 if (mg
->overwrite_bio
) {
1346 force_set_dirty(cache
, cblock
);
1347 else if (mg
->k
.input
)
1348 mg
->overwrite_bio
->bi_status
= mg
->k
.input
;
1350 mg
->overwrite_bio
->bi_status
= BLK_STS_IOERR
;
1351 bio_endio(mg
->overwrite_bio
);
1354 force_clear_dirty(cache
, cblock
);
1355 dec_io_migrations(cache
);
1361 * We clear dirty here to update the nr_dirty counter.
1364 force_clear_dirty(cache
, cblock
);
1365 policy_complete_background_work(cache
->policy
, op
, success
);
1366 dec_io_migrations(cache
);
1369 case POLICY_WRITEBACK
:
1371 force_clear_dirty(cache
, cblock
);
1372 policy_complete_background_work(cache
->policy
, op
, success
);
1373 dec_io_migrations(cache
);
1377 bio_list_init(&bios
);
1379 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1380 free_prison_cell(cache
, mg
->cell
);
1384 defer_bios(cache
, &bios
);
1385 wake_migration_worker(cache
);
1387 background_work_end(cache
);
1390 static void mg_success(struct work_struct
*ws
)
1392 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1393 mg_complete(mg
, mg
->k
.input
== 0);
1396 static void mg_update_metadata(struct work_struct
*ws
)
1399 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1400 struct cache
*cache
= mg
->cache
;
1401 struct policy_work
*op
= mg
->op
;
1404 case POLICY_PROMOTE
:
1405 r
= dm_cache_insert_mapping(cache
->cmd
, op
->cblock
, op
->oblock
);
1407 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1408 cache_device_name(cache
));
1409 metadata_operation_failed(cache
, "dm_cache_insert_mapping", r
);
1411 mg_complete(mg
, false);
1414 mg_complete(mg
, true);
1418 r
= dm_cache_remove_mapping(cache
->cmd
, op
->cblock
);
1420 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1421 cache_device_name(cache
));
1422 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1424 mg_complete(mg
, false);
1429 * It would be nice if we only had to commit when a REQ_FLUSH
1430 * comes through. But there's one scenario that we have to
1433 * - vblock x in a cache block
1435 * - cache block gets reallocated and over written
1438 * When we recover, because there was no commit the cache will
1439 * rollback to having the data for vblock x in the cache block.
1440 * But the cache block has since been overwritten, so it'll end
1441 * up pointing to data that was never in 'x' during the history
1444 * To avoid this issue we require a commit as part of the
1445 * demotion operation.
1447 init_continuation(&mg
->k
, mg_success
);
1448 continue_after_commit(&cache
->committer
, &mg
->k
);
1449 schedule_commit(&cache
->committer
);
1452 case POLICY_WRITEBACK
:
1453 mg_complete(mg
, true);
1458 static void mg_update_metadata_after_copy(struct work_struct
*ws
)
1460 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1463 * Did the copy succeed?
1466 mg_complete(mg
, false);
1468 mg_update_metadata(ws
);
1471 static void mg_upgrade_lock(struct work_struct
*ws
)
1474 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1477 * Did the copy succeed?
1480 mg_complete(mg
, false);
1484 * Now we want the lock to prevent both reads and writes.
1486 r
= dm_cell_lock_promote_v2(mg
->cache
->prison
, mg
->cell
,
1487 READ_WRITE_LOCK_LEVEL
);
1489 mg_complete(mg
, false);
1492 quiesce(mg
, mg_update_metadata
);
1495 mg_update_metadata(ws
);
1499 static void mg_full_copy(struct work_struct
*ws
)
1501 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1502 struct cache
*cache
= mg
->cache
;
1503 struct policy_work
*op
= mg
->op
;
1504 bool is_policy_promote
= (op
->op
== POLICY_PROMOTE
);
1506 if ((!is_policy_promote
&& !is_dirty(cache
, op
->cblock
)) ||
1507 is_discarded_oblock(cache
, op
->oblock
)) {
1508 mg_upgrade_lock(ws
);
1512 init_continuation(&mg
->k
, mg_upgrade_lock
);
1514 if (copy(mg
, is_policy_promote
)) {
1515 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache
));
1516 mg
->k
.input
= BLK_STS_IOERR
;
1517 mg_complete(mg
, false);
1521 static void mg_copy(struct work_struct
*ws
)
1523 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1525 if (mg
->overwrite_bio
) {
1527 * No exclusive lock was held when we last checked if the bio
1528 * was optimisable. So we have to check again in case things
1529 * have changed (eg, the block may no longer be discarded).
1531 if (!optimisable_bio(mg
->cache
, mg
->overwrite_bio
, mg
->op
->oblock
)) {
1533 * Fallback to a real full copy after doing some tidying up.
1535 bool rb
= bio_detain_shared(mg
->cache
, mg
->op
->oblock
, mg
->overwrite_bio
);
1536 BUG_ON(rb
); /* An exclussive lock must _not_ be held for this block */
1537 mg
->overwrite_bio
= NULL
;
1538 inc_io_migrations(mg
->cache
);
1544 * It's safe to do this here, even though it's new data
1545 * because all IO has been locked out of the block.
1547 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1548 * so _not_ using mg_upgrade_lock() as continutation.
1550 overwrite(mg
, mg_update_metadata_after_copy
);
1556 static int mg_lock_writes(struct dm_cache_migration
*mg
)
1559 struct dm_cell_key_v2 key
;
1560 struct cache
*cache
= mg
->cache
;
1561 struct dm_bio_prison_cell_v2
*prealloc
;
1563 prealloc
= alloc_prison_cell(cache
);
1565 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache
));
1566 mg_complete(mg
, false);
1571 * Prevent writes to the block, but allow reads to continue.
1572 * Unless we're using an overwrite bio, in which case we lock
1575 build_key(mg
->op
->oblock
, oblock_succ(mg
->op
->oblock
), &key
);
1576 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1577 mg
->overwrite_bio
? READ_WRITE_LOCK_LEVEL
: WRITE_LOCK_LEVEL
,
1578 prealloc
, &mg
->cell
);
1580 free_prison_cell(cache
, prealloc
);
1581 mg_complete(mg
, false);
1585 if (mg
->cell
!= prealloc
)
1586 free_prison_cell(cache
, prealloc
);
1591 quiesce(mg
, mg_copy
);
1596 static int mg_start(struct cache
*cache
, struct policy_work
*op
, struct bio
*bio
)
1598 struct dm_cache_migration
*mg
;
1600 if (!background_work_begin(cache
)) {
1601 policy_complete_background_work(cache
->policy
, op
, false);
1605 mg
= alloc_migration(cache
);
1607 policy_complete_background_work(cache
->policy
, op
, false);
1608 background_work_end(cache
);
1612 memset(mg
, 0, sizeof(*mg
));
1616 mg
->overwrite_bio
= bio
;
1619 inc_io_migrations(cache
);
1621 return mg_lock_writes(mg
);
1624 /*----------------------------------------------------------------
1625 * invalidation processing
1626 *--------------------------------------------------------------*/
1628 static void invalidate_complete(struct dm_cache_migration
*mg
, bool success
)
1630 struct bio_list bios
;
1631 struct cache
*cache
= mg
->cache
;
1633 bio_list_init(&bios
);
1634 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1635 free_prison_cell(cache
, mg
->cell
);
1637 if (!success
&& mg
->overwrite_bio
)
1638 bio_io_error(mg
->overwrite_bio
);
1641 defer_bios(cache
, &bios
);
1643 background_work_end(cache
);
1646 static void invalidate_completed(struct work_struct
*ws
)
1648 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1649 invalidate_complete(mg
, !mg
->k
.input
);
1652 static int invalidate_cblock(struct cache
*cache
, dm_cblock_t cblock
)
1654 int r
= policy_invalidate_mapping(cache
->policy
, cblock
);
1656 r
= dm_cache_remove_mapping(cache
->cmd
, cblock
);
1658 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1659 cache_device_name(cache
));
1660 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1663 } else if (r
== -ENODATA
) {
1665 * Harmless, already unmapped.
1670 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache
));
1675 static void invalidate_remove(struct work_struct
*ws
)
1678 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1679 struct cache
*cache
= mg
->cache
;
1681 r
= invalidate_cblock(cache
, mg
->invalidate_cblock
);
1683 invalidate_complete(mg
, false);
1687 init_continuation(&mg
->k
, invalidate_completed
);
1688 continue_after_commit(&cache
->committer
, &mg
->k
);
1689 remap_to_origin_clear_discard(cache
, mg
->overwrite_bio
, mg
->invalidate_oblock
);
1690 mg
->overwrite_bio
= NULL
;
1691 schedule_commit(&cache
->committer
);
1694 static int invalidate_lock(struct dm_cache_migration
*mg
)
1697 struct dm_cell_key_v2 key
;
1698 struct cache
*cache
= mg
->cache
;
1699 struct dm_bio_prison_cell_v2
*prealloc
;
1701 prealloc
= alloc_prison_cell(cache
);
1703 invalidate_complete(mg
, false);
1707 build_key(mg
->invalidate_oblock
, oblock_succ(mg
->invalidate_oblock
), &key
);
1708 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1709 READ_WRITE_LOCK_LEVEL
, prealloc
, &mg
->cell
);
1711 free_prison_cell(cache
, prealloc
);
1712 invalidate_complete(mg
, false);
1716 if (mg
->cell
!= prealloc
)
1717 free_prison_cell(cache
, prealloc
);
1720 quiesce(mg
, invalidate_remove
);
1724 * We can't call invalidate_remove() directly here because we
1725 * might still be in request context.
1727 init_continuation(&mg
->k
, invalidate_remove
);
1728 queue_work(cache
->wq
, &mg
->k
.ws
);
1734 static int invalidate_start(struct cache
*cache
, dm_cblock_t cblock
,
1735 dm_oblock_t oblock
, struct bio
*bio
)
1737 struct dm_cache_migration
*mg
;
1739 if (!background_work_begin(cache
))
1742 mg
= alloc_migration(cache
);
1744 background_work_end(cache
);
1748 memset(mg
, 0, sizeof(*mg
));
1751 mg
->overwrite_bio
= bio
;
1752 mg
->invalidate_cblock
= cblock
;
1753 mg
->invalidate_oblock
= oblock
;
1755 return invalidate_lock(mg
);
1758 /*----------------------------------------------------------------
1760 *--------------------------------------------------------------*/
1767 static enum busy
spare_migration_bandwidth(struct cache
*cache
)
1769 bool idle
= iot_idle_for(&cache
->tracker
, HZ
);
1770 sector_t current_volume
= (atomic_read(&cache
->nr_io_migrations
) + 1) *
1771 cache
->sectors_per_block
;
1773 if (idle
&& current_volume
<= cache
->migration_threshold
)
1779 static void inc_hit_counter(struct cache
*cache
, struct bio
*bio
)
1781 atomic_inc(bio_data_dir(bio
) == READ
?
1782 &cache
->stats
.read_hit
: &cache
->stats
.write_hit
);
1785 static void inc_miss_counter(struct cache
*cache
, struct bio
*bio
)
1787 atomic_inc(bio_data_dir(bio
) == READ
?
1788 &cache
->stats
.read_miss
: &cache
->stats
.write_miss
);
1791 /*----------------------------------------------------------------*/
1793 static int map_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
,
1794 bool *commit_needed
)
1797 bool rb
, background_queued
;
1799 size_t pb_data_size
= get_per_bio_data_size(cache
);
1800 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1802 *commit_needed
= false;
1804 rb
= bio_detain_shared(cache
, block
, bio
);
1807 * An exclusive lock is held for this block, so we have to
1808 * wait. We set the commit_needed flag so the current
1809 * transaction will be committed asap, allowing this lock
1812 *commit_needed
= true;
1813 return DM_MAPIO_SUBMITTED
;
1816 data_dir
= bio_data_dir(bio
);
1818 if (optimisable_bio(cache
, bio
, block
)) {
1819 struct policy_work
*op
= NULL
;
1821 r
= policy_lookup_with_work(cache
->policy
, block
, &cblock
, data_dir
, true, &op
);
1822 if (unlikely(r
&& r
!= -ENOENT
)) {
1823 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1824 cache_device_name(cache
), r
);
1826 return DM_MAPIO_SUBMITTED
;
1829 if (r
== -ENOENT
&& op
) {
1830 bio_drop_shared_lock(cache
, bio
);
1831 BUG_ON(op
->op
!= POLICY_PROMOTE
);
1832 mg_start(cache
, op
, bio
);
1833 return DM_MAPIO_SUBMITTED
;
1836 r
= policy_lookup(cache
->policy
, block
, &cblock
, data_dir
, false, &background_queued
);
1837 if (unlikely(r
&& r
!= -ENOENT
)) {
1838 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1839 cache_device_name(cache
), r
);
1841 return DM_MAPIO_SUBMITTED
;
1844 if (background_queued
)
1845 wake_migration_worker(cache
);
1852 inc_miss_counter(cache
, bio
);
1853 if (pb
->req_nr
== 0) {
1854 accounted_begin(cache
, bio
);
1855 remap_to_origin_clear_discard(cache
, bio
, block
);
1859 * This is a duplicate writethrough io that is no
1860 * longer needed because the block has been demoted.
1863 return DM_MAPIO_SUBMITTED
;
1869 inc_hit_counter(cache
, bio
);
1872 * Passthrough always maps to the origin, invalidating any
1873 * cache blocks that are written to.
1875 if (passthrough_mode(cache
)) {
1876 if (bio_data_dir(bio
) == WRITE
) {
1877 bio_drop_shared_lock(cache
, bio
);
1878 atomic_inc(&cache
->stats
.demotion
);
1879 invalidate_start(cache
, cblock
, block
, bio
);
1881 remap_to_origin_clear_discard(cache
, bio
, block
);
1884 if (bio_data_dir(bio
) == WRITE
&& writethrough_mode(cache
) &&
1885 !is_dirty(cache
, cblock
)) {
1886 remap_to_origin_and_cache(cache
, bio
, block
, cblock
);
1887 accounted_begin(cache
, bio
);
1889 remap_to_cache_dirty(cache
, bio
, block
, cblock
);
1894 * dm core turns FUA requests into a separate payload and FLUSH req.
1896 if (bio
->bi_opf
& REQ_FUA
) {
1898 * issue_after_commit will call accounted_begin a second time. So
1899 * we call accounted_complete() to avoid double accounting.
1901 accounted_complete(cache
, bio
);
1902 issue_after_commit(&cache
->committer
, bio
);
1903 *commit_needed
= true;
1904 return DM_MAPIO_SUBMITTED
;
1907 return DM_MAPIO_REMAPPED
;
1910 static bool process_bio(struct cache
*cache
, struct bio
*bio
)
1914 if (map_bio(cache
, bio
, get_bio_block(cache
, bio
), &commit_needed
) == DM_MAPIO_REMAPPED
)
1915 generic_make_request(bio
);
1917 return commit_needed
;
1921 * A non-zero return indicates read_only or fail_io mode.
1923 static int commit(struct cache
*cache
, bool clean_shutdown
)
1927 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1930 atomic_inc(&cache
->stats
.commit_count
);
1931 r
= dm_cache_commit(cache
->cmd
, clean_shutdown
);
1933 metadata_operation_failed(cache
, "dm_cache_commit", r
);
1939 * Used by the batcher.
1941 static blk_status_t
commit_op(void *context
)
1943 struct cache
*cache
= context
;
1945 if (dm_cache_changed_this_transaction(cache
->cmd
))
1946 return errno_to_blk_status(commit(cache
, false));
1951 /*----------------------------------------------------------------*/
1953 static bool process_flush_bio(struct cache
*cache
, struct bio
*bio
)
1955 size_t pb_data_size
= get_per_bio_data_size(cache
);
1956 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1959 remap_to_origin(cache
, bio
);
1961 remap_to_cache(cache
, bio
, 0);
1963 issue_after_commit(&cache
->committer
, bio
);
1967 static bool process_discard_bio(struct cache
*cache
, struct bio
*bio
)
1971 // FIXME: do we need to lock the region? Or can we just assume the
1972 // user wont be so foolish as to issue discard concurrently with
1974 calc_discard_block_range(cache
, bio
, &b
, &e
);
1976 set_discard(cache
, b
);
1977 b
= to_dblock(from_dblock(b
) + 1);
1985 static void process_deferred_bios(struct work_struct
*ws
)
1987 struct cache
*cache
= container_of(ws
, struct cache
, deferred_bio_worker
);
1989 unsigned long flags
;
1990 bool commit_needed
= false;
1991 struct bio_list bios
;
1994 bio_list_init(&bios
);
1996 spin_lock_irqsave(&cache
->lock
, flags
);
1997 bio_list_merge(&bios
, &cache
->deferred_bios
);
1998 bio_list_init(&cache
->deferred_bios
);
1999 spin_unlock_irqrestore(&cache
->lock
, flags
);
2001 while ((bio
= bio_list_pop(&bios
))) {
2002 if (bio
->bi_opf
& REQ_PREFLUSH
)
2003 commit_needed
= process_flush_bio(cache
, bio
) || commit_needed
;
2005 else if (bio_op(bio
) == REQ_OP_DISCARD
)
2006 commit_needed
= process_discard_bio(cache
, bio
) || commit_needed
;
2009 commit_needed
= process_bio(cache
, bio
) || commit_needed
;
2013 schedule_commit(&cache
->committer
);
2016 static void process_deferred_writethrough_bios(struct work_struct
*ws
)
2018 struct cache
*cache
= container_of(ws
, struct cache
, deferred_writethrough_worker
);
2020 unsigned long flags
;
2021 struct bio_list bios
;
2024 bio_list_init(&bios
);
2026 spin_lock_irqsave(&cache
->lock
, flags
);
2027 bio_list_merge(&bios
, &cache
->deferred_writethrough_bios
);
2028 bio_list_init(&cache
->deferred_writethrough_bios
);
2029 spin_unlock_irqrestore(&cache
->lock
, flags
);
2032 * These bios have already been through accounted_begin()
2034 while ((bio
= bio_list_pop(&bios
)))
2035 generic_make_request(bio
);
2038 /*----------------------------------------------------------------
2040 *--------------------------------------------------------------*/
2042 static void requeue_deferred_bios(struct cache
*cache
)
2045 struct bio_list bios
;
2047 bio_list_init(&bios
);
2048 bio_list_merge(&bios
, &cache
->deferred_bios
);
2049 bio_list_init(&cache
->deferred_bios
);
2051 while ((bio
= bio_list_pop(&bios
))) {
2052 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
2058 * We want to commit periodically so that not too much
2059 * unwritten metadata builds up.
2061 static void do_waker(struct work_struct
*ws
)
2063 struct cache
*cache
= container_of(to_delayed_work(ws
), struct cache
, waker
);
2065 policy_tick(cache
->policy
, true);
2066 wake_migration_worker(cache
);
2067 schedule_commit(&cache
->committer
);
2068 queue_delayed_work(cache
->wq
, &cache
->waker
, COMMIT_PERIOD
);
2071 static void check_migrations(struct work_struct
*ws
)
2074 struct policy_work
*op
;
2075 struct cache
*cache
= container_of(ws
, struct cache
, migration_worker
);
2079 b
= spare_migration_bandwidth(cache
);
2081 r
= policy_get_background_work(cache
->policy
, b
== IDLE
, &op
);
2086 DMERR_LIMIT("%s: policy_background_work failed",
2087 cache_device_name(cache
));
2091 r
= mg_start(cache
, op
, NULL
);
2097 /*----------------------------------------------------------------
2099 *--------------------------------------------------------------*/
2102 * This function gets called on the error paths of the constructor, so we
2103 * have to cope with a partially initialised struct.
2105 static void destroy(struct cache
*cache
)
2109 mempool_destroy(cache
->migration_pool
);
2112 dm_bio_prison_destroy_v2(cache
->prison
);
2115 destroy_workqueue(cache
->wq
);
2117 if (cache
->dirty_bitset
)
2118 free_bitset(cache
->dirty_bitset
);
2120 if (cache
->discard_bitset
)
2121 free_bitset(cache
->discard_bitset
);
2124 dm_kcopyd_client_destroy(cache
->copier
);
2127 dm_cache_metadata_close(cache
->cmd
);
2129 if (cache
->metadata_dev
)
2130 dm_put_device(cache
->ti
, cache
->metadata_dev
);
2132 if (cache
->origin_dev
)
2133 dm_put_device(cache
->ti
, cache
->origin_dev
);
2135 if (cache
->cache_dev
)
2136 dm_put_device(cache
->ti
, cache
->cache_dev
);
2139 dm_cache_policy_destroy(cache
->policy
);
2141 for (i
= 0; i
< cache
->nr_ctr_args
; i
++)
2142 kfree(cache
->ctr_args
[i
]);
2143 kfree(cache
->ctr_args
);
2146 bioset_free(cache
->bs
);
2151 static void cache_dtr(struct dm_target
*ti
)
2153 struct cache
*cache
= ti
->private;
2158 static sector_t
get_dev_size(struct dm_dev
*dev
)
2160 return i_size_read(dev
->bdev
->bd_inode
) >> SECTOR_SHIFT
;
2163 /*----------------------------------------------------------------*/
2166 * Construct a cache device mapping.
2168 * cache <metadata dev> <cache dev> <origin dev> <block size>
2169 * <#feature args> [<feature arg>]*
2170 * <policy> <#policy args> [<policy arg>]*
2172 * metadata dev : fast device holding the persistent metadata
2173 * cache dev : fast device holding cached data blocks
2174 * origin dev : slow device holding original data blocks
2175 * block size : cache unit size in sectors
2177 * #feature args : number of feature arguments passed
2178 * feature args : writethrough. (The default is writeback.)
2180 * policy : the replacement policy to use
2181 * #policy args : an even number of policy arguments corresponding
2182 * to key/value pairs passed to the policy
2183 * policy args : key/value pairs passed to the policy
2184 * E.g. 'sequential_threshold 1024'
2185 * See cache-policies.txt for details.
2187 * Optional feature arguments are:
2188 * writethrough : write through caching that prohibits cache block
2189 * content from being different from origin block content.
2190 * Without this argument, the default behaviour is to write
2191 * back cache block contents later for performance reasons,
2192 * so they may differ from the corresponding origin blocks.
2195 struct dm_target
*ti
;
2197 struct dm_dev
*metadata_dev
;
2199 struct dm_dev
*cache_dev
;
2200 sector_t cache_sectors
;
2202 struct dm_dev
*origin_dev
;
2203 sector_t origin_sectors
;
2205 uint32_t block_size
;
2207 const char *policy_name
;
2209 const char **policy_argv
;
2211 struct cache_features features
;
2214 static void destroy_cache_args(struct cache_args
*ca
)
2216 if (ca
->metadata_dev
)
2217 dm_put_device(ca
->ti
, ca
->metadata_dev
);
2220 dm_put_device(ca
->ti
, ca
->cache_dev
);
2223 dm_put_device(ca
->ti
, ca
->origin_dev
);
2228 static bool at_least_one_arg(struct dm_arg_set
*as
, char **error
)
2231 *error
= "Insufficient args";
2238 static int parse_metadata_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2242 sector_t metadata_dev_size
;
2243 char b
[BDEVNAME_SIZE
];
2245 if (!at_least_one_arg(as
, error
))
2248 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2251 *error
= "Error opening metadata device";
2255 metadata_dev_size
= get_dev_size(ca
->metadata_dev
);
2256 if (metadata_dev_size
> DM_CACHE_METADATA_MAX_SECTORS_WARNING
)
2257 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2258 bdevname(ca
->metadata_dev
->bdev
, b
), THIN_METADATA_MAX_SECTORS
);
2263 static int parse_cache_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2268 if (!at_least_one_arg(as
, error
))
2271 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2274 *error
= "Error opening cache device";
2277 ca
->cache_sectors
= get_dev_size(ca
->cache_dev
);
2282 static int parse_origin_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2287 if (!at_least_one_arg(as
, error
))
2290 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2293 *error
= "Error opening origin device";
2297 ca
->origin_sectors
= get_dev_size(ca
->origin_dev
);
2298 if (ca
->ti
->len
> ca
->origin_sectors
) {
2299 *error
= "Device size larger than cached device";
2306 static int parse_block_size(struct cache_args
*ca
, struct dm_arg_set
*as
,
2309 unsigned long block_size
;
2311 if (!at_least_one_arg(as
, error
))
2314 if (kstrtoul(dm_shift_arg(as
), 10, &block_size
) || !block_size
||
2315 block_size
< DATA_DEV_BLOCK_SIZE_MIN_SECTORS
||
2316 block_size
> DATA_DEV_BLOCK_SIZE_MAX_SECTORS
||
2317 block_size
& (DATA_DEV_BLOCK_SIZE_MIN_SECTORS
- 1)) {
2318 *error
= "Invalid data block size";
2322 if (block_size
> ca
->cache_sectors
) {
2323 *error
= "Data block size is larger than the cache device";
2327 ca
->block_size
= block_size
;
2332 static void init_features(struct cache_features
*cf
)
2334 cf
->mode
= CM_WRITE
;
2335 cf
->io_mode
= CM_IO_WRITEBACK
;
2336 cf
->metadata_version
= 1;
2339 static int parse_features(struct cache_args
*ca
, struct dm_arg_set
*as
,
2342 static const struct dm_arg _args
[] = {
2343 {0, 2, "Invalid number of cache feature arguments"},
2349 struct cache_features
*cf
= &ca
->features
;
2353 r
= dm_read_arg_group(_args
, as
, &argc
, error
);
2358 arg
= dm_shift_arg(as
);
2360 if (!strcasecmp(arg
, "writeback"))
2361 cf
->io_mode
= CM_IO_WRITEBACK
;
2363 else if (!strcasecmp(arg
, "writethrough"))
2364 cf
->io_mode
= CM_IO_WRITETHROUGH
;
2366 else if (!strcasecmp(arg
, "passthrough"))
2367 cf
->io_mode
= CM_IO_PASSTHROUGH
;
2369 else if (!strcasecmp(arg
, "metadata2"))
2370 cf
->metadata_version
= 2;
2373 *error
= "Unrecognised cache feature requested";
2381 static int parse_policy(struct cache_args
*ca
, struct dm_arg_set
*as
,
2384 static const struct dm_arg _args
[] = {
2385 {0, 1024, "Invalid number of policy arguments"},
2390 if (!at_least_one_arg(as
, error
))
2393 ca
->policy_name
= dm_shift_arg(as
);
2395 r
= dm_read_arg_group(_args
, as
, &ca
->policy_argc
, error
);
2399 ca
->policy_argv
= (const char **)as
->argv
;
2400 dm_consume_args(as
, ca
->policy_argc
);
2405 static int parse_cache_args(struct cache_args
*ca
, int argc
, char **argv
,
2409 struct dm_arg_set as
;
2414 r
= parse_metadata_dev(ca
, &as
, error
);
2418 r
= parse_cache_dev(ca
, &as
, error
);
2422 r
= parse_origin_dev(ca
, &as
, error
);
2426 r
= parse_block_size(ca
, &as
, error
);
2430 r
= parse_features(ca
, &as
, error
);
2434 r
= parse_policy(ca
, &as
, error
);
2441 /*----------------------------------------------------------------*/
2443 static struct kmem_cache
*migration_cache
;
2445 #define NOT_CORE_OPTION 1
2447 static int process_config_option(struct cache
*cache
, const char *key
, const char *value
)
2451 if (!strcasecmp(key
, "migration_threshold")) {
2452 if (kstrtoul(value
, 10, &tmp
))
2455 cache
->migration_threshold
= tmp
;
2459 return NOT_CORE_OPTION
;
2462 static int set_config_value(struct cache
*cache
, const char *key
, const char *value
)
2464 int r
= process_config_option(cache
, key
, value
);
2466 if (r
== NOT_CORE_OPTION
)
2467 r
= policy_set_config_value(cache
->policy
, key
, value
);
2470 DMWARN("bad config value for %s: %s", key
, value
);
2475 static int set_config_values(struct cache
*cache
, int argc
, const char **argv
)
2480 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2485 r
= set_config_value(cache
, argv
[0], argv
[1]);
2496 static int create_cache_policy(struct cache
*cache
, struct cache_args
*ca
,
2499 struct dm_cache_policy
*p
= dm_cache_policy_create(ca
->policy_name
,
2501 cache
->origin_sectors
,
2502 cache
->sectors_per_block
);
2504 *error
= "Error creating cache's policy";
2508 BUG_ON(!cache
->policy
);
2514 * We want the discard block size to be at least the size of the cache
2515 * block size and have no more than 2^14 discard blocks across the origin.
2517 #define MAX_DISCARD_BLOCKS (1 << 14)
2519 static bool too_many_discard_blocks(sector_t discard_block_size
,
2520 sector_t origin_size
)
2522 (void) sector_div(origin_size
, discard_block_size
);
2524 return origin_size
> MAX_DISCARD_BLOCKS
;
2527 static sector_t
calculate_discard_block_size(sector_t cache_block_size
,
2528 sector_t origin_size
)
2530 sector_t discard_block_size
= cache_block_size
;
2533 while (too_many_discard_blocks(discard_block_size
, origin_size
))
2534 discard_block_size
*= 2;
2536 return discard_block_size
;
2539 static void set_cache_size(struct cache
*cache
, dm_cblock_t size
)
2541 dm_block_t nr_blocks
= from_cblock(size
);
2543 if (nr_blocks
> (1 << 20) && cache
->cache_size
!= size
)
2544 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2545 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2546 "Please consider increasing the cache block size to reduce the overall cache block count.",
2547 (unsigned long long) nr_blocks
);
2549 cache
->cache_size
= size
;
2552 static int is_congested(struct dm_dev
*dev
, int bdi_bits
)
2554 struct request_queue
*q
= bdev_get_queue(dev
->bdev
);
2555 return bdi_congested(q
->backing_dev_info
, bdi_bits
);
2558 static int cache_is_congested(struct dm_target_callbacks
*cb
, int bdi_bits
)
2560 struct cache
*cache
= container_of(cb
, struct cache
, callbacks
);
2562 return is_congested(cache
->origin_dev
, bdi_bits
) ||
2563 is_congested(cache
->cache_dev
, bdi_bits
);
2566 #define DEFAULT_MIGRATION_THRESHOLD 2048
2568 static int cache_create(struct cache_args
*ca
, struct cache
**result
)
2571 char **error
= &ca
->ti
->error
;
2572 struct cache
*cache
;
2573 struct dm_target
*ti
= ca
->ti
;
2574 dm_block_t origin_blocks
;
2575 struct dm_cache_metadata
*cmd
;
2576 bool may_format
= ca
->features
.mode
== CM_WRITE
;
2578 cache
= kzalloc(sizeof(*cache
), GFP_KERNEL
);
2583 ti
->private = cache
;
2584 ti
->num_flush_bios
= 2;
2585 ti
->flush_supported
= true;
2587 ti
->num_discard_bios
= 1;
2588 ti
->discards_supported
= true;
2589 ti
->split_discard_bios
= false;
2591 cache
->features
= ca
->features
;
2592 ti
->per_io_data_size
= get_per_bio_data_size(cache
);
2594 if (writethrough_mode(cache
)) {
2595 /* Create bioset for writethrough bios issued to origin */
2596 cache
->bs
= bioset_create(BIO_POOL_SIZE
, 0, 0);
2601 cache
->callbacks
.congested_fn
= cache_is_congested
;
2602 dm_table_add_target_callbacks(ti
->table
, &cache
->callbacks
);
2604 cache
->metadata_dev
= ca
->metadata_dev
;
2605 cache
->origin_dev
= ca
->origin_dev
;
2606 cache
->cache_dev
= ca
->cache_dev
;
2608 ca
->metadata_dev
= ca
->origin_dev
= ca
->cache_dev
= NULL
;
2610 origin_blocks
= cache
->origin_sectors
= ca
->origin_sectors
;
2611 origin_blocks
= block_div(origin_blocks
, ca
->block_size
);
2612 cache
->origin_blocks
= to_oblock(origin_blocks
);
2614 cache
->sectors_per_block
= ca
->block_size
;
2615 if (dm_set_target_max_io_len(ti
, cache
->sectors_per_block
)) {
2620 if (ca
->block_size
& (ca
->block_size
- 1)) {
2621 dm_block_t cache_size
= ca
->cache_sectors
;
2623 cache
->sectors_per_block_shift
= -1;
2624 cache_size
= block_div(cache_size
, ca
->block_size
);
2625 set_cache_size(cache
, to_cblock(cache_size
));
2627 cache
->sectors_per_block_shift
= __ffs(ca
->block_size
);
2628 set_cache_size(cache
, to_cblock(ca
->cache_sectors
>> cache
->sectors_per_block_shift
));
2631 r
= create_cache_policy(cache
, ca
, error
);
2635 cache
->policy_nr_args
= ca
->policy_argc
;
2636 cache
->migration_threshold
= DEFAULT_MIGRATION_THRESHOLD
;
2638 r
= set_config_values(cache
, ca
->policy_argc
, ca
->policy_argv
);
2640 *error
= "Error setting cache policy's config values";
2644 cmd
= dm_cache_metadata_open(cache
->metadata_dev
->bdev
,
2645 ca
->block_size
, may_format
,
2646 dm_cache_policy_get_hint_size(cache
->policy
),
2647 ca
->features
.metadata_version
);
2649 *error
= "Error creating metadata object";
2654 set_cache_mode(cache
, CM_WRITE
);
2655 if (get_cache_mode(cache
) != CM_WRITE
) {
2656 *error
= "Unable to get write access to metadata, please check/repair metadata.";
2661 if (passthrough_mode(cache
)) {
2664 r
= dm_cache_metadata_all_clean(cache
->cmd
, &all_clean
);
2666 *error
= "dm_cache_metadata_all_clean() failed";
2671 *error
= "Cannot enter passthrough mode unless all blocks are clean";
2676 policy_allow_migrations(cache
->policy
, false);
2679 spin_lock_init(&cache
->lock
);
2680 INIT_LIST_HEAD(&cache
->deferred_cells
);
2681 bio_list_init(&cache
->deferred_bios
);
2682 bio_list_init(&cache
->deferred_writethrough_bios
);
2683 atomic_set(&cache
->nr_allocated_migrations
, 0);
2684 atomic_set(&cache
->nr_io_migrations
, 0);
2685 init_waitqueue_head(&cache
->migration_wait
);
2688 atomic_set(&cache
->nr_dirty
, 0);
2689 cache
->dirty_bitset
= alloc_bitset(from_cblock(cache
->cache_size
));
2690 if (!cache
->dirty_bitset
) {
2691 *error
= "could not allocate dirty bitset";
2694 clear_bitset(cache
->dirty_bitset
, from_cblock(cache
->cache_size
));
2696 cache
->discard_block_size
=
2697 calculate_discard_block_size(cache
->sectors_per_block
,
2698 cache
->origin_sectors
);
2699 cache
->discard_nr_blocks
= to_dblock(dm_sector_div_up(cache
->origin_sectors
,
2700 cache
->discard_block_size
));
2701 cache
->discard_bitset
= alloc_bitset(from_dblock(cache
->discard_nr_blocks
));
2702 if (!cache
->discard_bitset
) {
2703 *error
= "could not allocate discard bitset";
2706 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
2708 cache
->copier
= dm_kcopyd_client_create(&dm_kcopyd_throttle
);
2709 if (IS_ERR(cache
->copier
)) {
2710 *error
= "could not create kcopyd client";
2711 r
= PTR_ERR(cache
->copier
);
2715 cache
->wq
= alloc_workqueue("dm-" DM_MSG_PREFIX
, WQ_MEM_RECLAIM
, 0);
2717 *error
= "could not create workqueue for metadata object";
2720 INIT_WORK(&cache
->deferred_bio_worker
, process_deferred_bios
);
2721 INIT_WORK(&cache
->deferred_writethrough_worker
,
2722 process_deferred_writethrough_bios
);
2723 INIT_WORK(&cache
->migration_worker
, check_migrations
);
2724 INIT_DELAYED_WORK(&cache
->waker
, do_waker
);
2726 cache
->prison
= dm_bio_prison_create_v2(cache
->wq
);
2727 if (!cache
->prison
) {
2728 *error
= "could not create bio prison";
2732 cache
->migration_pool
= mempool_create_slab_pool(MIGRATION_POOL_SIZE
,
2734 if (!cache
->migration_pool
) {
2735 *error
= "Error creating cache's migration mempool";
2739 cache
->need_tick_bio
= true;
2740 cache
->sized
= false;
2741 cache
->invalidate
= false;
2742 cache
->commit_requested
= false;
2743 cache
->loaded_mappings
= false;
2744 cache
->loaded_discards
= false;
2748 atomic_set(&cache
->stats
.demotion
, 0);
2749 atomic_set(&cache
->stats
.promotion
, 0);
2750 atomic_set(&cache
->stats
.copies_avoided
, 0);
2751 atomic_set(&cache
->stats
.cache_cell_clash
, 0);
2752 atomic_set(&cache
->stats
.commit_count
, 0);
2753 atomic_set(&cache
->stats
.discard_count
, 0);
2755 spin_lock_init(&cache
->invalidation_lock
);
2756 INIT_LIST_HEAD(&cache
->invalidation_requests
);
2758 batcher_init(&cache
->committer
, commit_op
, cache
,
2759 issue_op
, cache
, cache
->wq
);
2760 iot_init(&cache
->tracker
);
2762 init_rwsem(&cache
->background_work_lock
);
2763 prevent_background_work(cache
);
2772 static int copy_ctr_args(struct cache
*cache
, int argc
, const char **argv
)
2777 copy
= kcalloc(argc
, sizeof(*copy
), GFP_KERNEL
);
2780 for (i
= 0; i
< argc
; i
++) {
2781 copy
[i
] = kstrdup(argv
[i
], GFP_KERNEL
);
2790 cache
->nr_ctr_args
= argc
;
2791 cache
->ctr_args
= copy
;
2796 static int cache_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
2799 struct cache_args
*ca
;
2800 struct cache
*cache
= NULL
;
2802 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
2804 ti
->error
= "Error allocating memory for cache";
2809 r
= parse_cache_args(ca
, argc
, argv
, &ti
->error
);
2813 r
= cache_create(ca
, &cache
);
2817 r
= copy_ctr_args(cache
, argc
- 3, (const char **)argv
+ 3);
2823 ti
->private = cache
;
2825 destroy_cache_args(ca
);
2829 /*----------------------------------------------------------------*/
2831 static int cache_map(struct dm_target
*ti
, struct bio
*bio
)
2833 struct cache
*cache
= ti
->private;
2837 dm_oblock_t block
= get_bio_block(cache
, bio
);
2838 size_t pb_data_size
= get_per_bio_data_size(cache
);
2840 init_per_bio_data(bio
, pb_data_size
);
2841 if (unlikely(from_oblock(block
) >= from_oblock(cache
->origin_blocks
))) {
2843 * This can only occur if the io goes to a partial block at
2844 * the end of the origin device. We don't cache these.
2845 * Just remap to the origin and carry on.
2847 remap_to_origin(cache
, bio
);
2848 accounted_begin(cache
, bio
);
2849 return DM_MAPIO_REMAPPED
;
2852 if (discard_or_flush(bio
)) {
2853 defer_bio(cache
, bio
);
2854 return DM_MAPIO_SUBMITTED
;
2857 r
= map_bio(cache
, bio
, block
, &commit_needed
);
2859 schedule_commit(&cache
->committer
);
2864 static int cache_end_io(struct dm_target
*ti
, struct bio
*bio
,
2865 blk_status_t
*error
)
2867 struct cache
*cache
= ti
->private;
2868 unsigned long flags
;
2869 size_t pb_data_size
= get_per_bio_data_size(cache
);
2870 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
2873 policy_tick(cache
->policy
, false);
2875 spin_lock_irqsave(&cache
->lock
, flags
);
2876 cache
->need_tick_bio
= true;
2877 spin_unlock_irqrestore(&cache
->lock
, flags
);
2880 bio_drop_shared_lock(cache
, bio
);
2881 accounted_complete(cache
, bio
);
2883 return DM_ENDIO_DONE
;
2886 static int write_dirty_bitset(struct cache
*cache
)
2890 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2893 r
= dm_cache_set_dirty_bits(cache
->cmd
, from_cblock(cache
->cache_size
), cache
->dirty_bitset
);
2895 metadata_operation_failed(cache
, "dm_cache_set_dirty_bits", r
);
2900 static int write_discard_bitset(struct cache
*cache
)
2904 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2907 r
= dm_cache_discard_bitset_resize(cache
->cmd
, cache
->discard_block_size
,
2908 cache
->discard_nr_blocks
);
2910 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache
));
2911 metadata_operation_failed(cache
, "dm_cache_discard_bitset_resize", r
);
2915 for (i
= 0; i
< from_dblock(cache
->discard_nr_blocks
); i
++) {
2916 r
= dm_cache_set_discard(cache
->cmd
, to_dblock(i
),
2917 is_discarded(cache
, to_dblock(i
)));
2919 metadata_operation_failed(cache
, "dm_cache_set_discard", r
);
2927 static int write_hints(struct cache
*cache
)
2931 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2934 r
= dm_cache_write_hints(cache
->cmd
, cache
->policy
);
2936 metadata_operation_failed(cache
, "dm_cache_write_hints", r
);
2944 * returns true on success
2946 static bool sync_metadata(struct cache
*cache
)
2950 r1
= write_dirty_bitset(cache
);
2952 DMERR("%s: could not write dirty bitset", cache_device_name(cache
));
2954 r2
= write_discard_bitset(cache
);
2956 DMERR("%s: could not write discard bitset", cache_device_name(cache
));
2960 r3
= write_hints(cache
);
2962 DMERR("%s: could not write hints", cache_device_name(cache
));
2965 * If writing the above metadata failed, we still commit, but don't
2966 * set the clean shutdown flag. This will effectively force every
2967 * dirty bit to be set on reload.
2969 r4
= commit(cache
, !r1
&& !r2
&& !r3
);
2971 DMERR("%s: could not write cache metadata", cache_device_name(cache
));
2973 return !r1
&& !r2
&& !r3
&& !r4
;
2976 static void cache_postsuspend(struct dm_target
*ti
)
2978 struct cache
*cache
= ti
->private;
2980 prevent_background_work(cache
);
2981 BUG_ON(atomic_read(&cache
->nr_io_migrations
));
2983 cancel_delayed_work(&cache
->waker
);
2984 flush_workqueue(cache
->wq
);
2985 WARN_ON(cache
->tracker
.in_flight
);
2988 * If it's a flush suspend there won't be any deferred bios, so this
2991 requeue_deferred_bios(cache
);
2993 if (get_cache_mode(cache
) == CM_WRITE
)
2994 (void) sync_metadata(cache
);
2997 static int load_mapping(void *context
, dm_oblock_t oblock
, dm_cblock_t cblock
,
2998 bool dirty
, uint32_t hint
, bool hint_valid
)
3001 struct cache
*cache
= context
;
3004 set_bit(from_cblock(cblock
), cache
->dirty_bitset
);
3005 atomic_inc(&cache
->nr_dirty
);
3007 clear_bit(from_cblock(cblock
), cache
->dirty_bitset
);
3009 r
= policy_load_mapping(cache
->policy
, oblock
, cblock
, dirty
, hint
, hint_valid
);
3017 * The discard block size in the on disk metadata is not
3018 * neccessarily the same as we're currently using. So we have to
3019 * be careful to only set the discarded attribute if we know it
3020 * covers a complete block of the new size.
3022 struct discard_load_info
{
3023 struct cache
*cache
;
3026 * These blocks are sized using the on disk dblock size, rather
3027 * than the current one.
3029 dm_block_t block_size
;
3030 dm_block_t discard_begin
, discard_end
;
3033 static void discard_load_info_init(struct cache
*cache
,
3034 struct discard_load_info
*li
)
3037 li
->discard_begin
= li
->discard_end
= 0;
3040 static void set_discard_range(struct discard_load_info
*li
)
3044 if (li
->discard_begin
== li
->discard_end
)
3048 * Convert to sectors.
3050 b
= li
->discard_begin
* li
->block_size
;
3051 e
= li
->discard_end
* li
->block_size
;
3054 * Then convert back to the current dblock size.
3056 b
= dm_sector_div_up(b
, li
->cache
->discard_block_size
);
3057 sector_div(e
, li
->cache
->discard_block_size
);
3060 * The origin may have shrunk, so we need to check we're still in
3063 if (e
> from_dblock(li
->cache
->discard_nr_blocks
))
3064 e
= from_dblock(li
->cache
->discard_nr_blocks
);
3067 set_discard(li
->cache
, to_dblock(b
));
3070 static int load_discard(void *context
, sector_t discard_block_size
,
3071 dm_dblock_t dblock
, bool discard
)
3073 struct discard_load_info
*li
= context
;
3075 li
->block_size
= discard_block_size
;
3078 if (from_dblock(dblock
) == li
->discard_end
)
3080 * We're already in a discard range, just extend it.
3082 li
->discard_end
= li
->discard_end
+ 1ULL;
3086 * Emit the old range and start a new one.
3088 set_discard_range(li
);
3089 li
->discard_begin
= from_dblock(dblock
);
3090 li
->discard_end
= li
->discard_begin
+ 1ULL;
3093 set_discard_range(li
);
3094 li
->discard_begin
= li
->discard_end
= 0;
3100 static dm_cblock_t
get_cache_dev_size(struct cache
*cache
)
3102 sector_t size
= get_dev_size(cache
->cache_dev
);
3103 (void) sector_div(size
, cache
->sectors_per_block
);
3104 return to_cblock(size
);
3107 static bool can_resize(struct cache
*cache
, dm_cblock_t new_size
)
3109 if (from_cblock(new_size
) > from_cblock(cache
->cache_size
))
3113 * We can't drop a dirty block when shrinking the cache.
3115 while (from_cblock(new_size
) < from_cblock(cache
->cache_size
)) {
3116 new_size
= to_cblock(from_cblock(new_size
) + 1);
3117 if (is_dirty(cache
, new_size
)) {
3118 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3119 cache_device_name(cache
),
3120 (unsigned long long) from_cblock(new_size
));
3128 static int resize_cache_dev(struct cache
*cache
, dm_cblock_t new_size
)
3132 r
= dm_cache_resize(cache
->cmd
, new_size
);
3134 DMERR("%s: could not resize cache metadata", cache_device_name(cache
));
3135 metadata_operation_failed(cache
, "dm_cache_resize", r
);
3139 set_cache_size(cache
, new_size
);
3144 static int cache_preresume(struct dm_target
*ti
)
3147 struct cache
*cache
= ti
->private;
3148 dm_cblock_t csize
= get_cache_dev_size(cache
);
3151 * Check to see if the cache has resized.
3153 if (!cache
->sized
) {
3154 r
= resize_cache_dev(cache
, csize
);
3158 cache
->sized
= true;
3160 } else if (csize
!= cache
->cache_size
) {
3161 if (!can_resize(cache
, csize
))
3164 r
= resize_cache_dev(cache
, csize
);
3169 if (!cache
->loaded_mappings
) {
3170 r
= dm_cache_load_mappings(cache
->cmd
, cache
->policy
,
3171 load_mapping
, cache
);
3173 DMERR("%s: could not load cache mappings", cache_device_name(cache
));
3174 metadata_operation_failed(cache
, "dm_cache_load_mappings", r
);
3178 cache
->loaded_mappings
= true;
3181 if (!cache
->loaded_discards
) {
3182 struct discard_load_info li
;
3185 * The discard bitset could have been resized, or the
3186 * discard block size changed. To be safe we start by
3187 * setting every dblock to not discarded.
3189 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
3191 discard_load_info_init(cache
, &li
);
3192 r
= dm_cache_load_discards(cache
->cmd
, load_discard
, &li
);
3194 DMERR("%s: could not load origin discards", cache_device_name(cache
));
3195 metadata_operation_failed(cache
, "dm_cache_load_discards", r
);
3198 set_discard_range(&li
);
3200 cache
->loaded_discards
= true;
3206 static void cache_resume(struct dm_target
*ti
)
3208 struct cache
*cache
= ti
->private;
3210 cache
->need_tick_bio
= true;
3211 allow_background_work(cache
);
3212 do_waker(&cache
->waker
.work
);
3218 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3219 * <cache block size> <#used cache blocks>/<#total cache blocks>
3220 * <#read hits> <#read misses> <#write hits> <#write misses>
3221 * <#demotions> <#promotions> <#dirty>
3222 * <#features> <features>*
3223 * <#core args> <core args>
3224 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3226 static void cache_status(struct dm_target
*ti
, status_type_t type
,
3227 unsigned status_flags
, char *result
, unsigned maxlen
)
3232 dm_block_t nr_free_blocks_metadata
= 0;
3233 dm_block_t nr_blocks_metadata
= 0;
3234 char buf
[BDEVNAME_SIZE
];
3235 struct cache
*cache
= ti
->private;
3236 dm_cblock_t residency
;
3240 case STATUSTYPE_INFO
:
3241 if (get_cache_mode(cache
) == CM_FAIL
) {
3246 /* Commit to ensure statistics aren't out-of-date */
3247 if (!(status_flags
& DM_STATUS_NOFLUSH_FLAG
) && !dm_suspended(ti
))
3248 (void) commit(cache
, false);
3250 r
= dm_cache_get_free_metadata_block_count(cache
->cmd
, &nr_free_blocks_metadata
);
3252 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3253 cache_device_name(cache
), r
);
3257 r
= dm_cache_get_metadata_dev_size(cache
->cmd
, &nr_blocks_metadata
);
3259 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3260 cache_device_name(cache
), r
);
3264 residency
= policy_residency(cache
->policy
);
3266 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3267 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE
,
3268 (unsigned long long)(nr_blocks_metadata
- nr_free_blocks_metadata
),
3269 (unsigned long long)nr_blocks_metadata
,
3270 (unsigned long long)cache
->sectors_per_block
,
3271 (unsigned long long) from_cblock(residency
),
3272 (unsigned long long) from_cblock(cache
->cache_size
),
3273 (unsigned) atomic_read(&cache
->stats
.read_hit
),
3274 (unsigned) atomic_read(&cache
->stats
.read_miss
),
3275 (unsigned) atomic_read(&cache
->stats
.write_hit
),
3276 (unsigned) atomic_read(&cache
->stats
.write_miss
),
3277 (unsigned) atomic_read(&cache
->stats
.demotion
),
3278 (unsigned) atomic_read(&cache
->stats
.promotion
),
3279 (unsigned long) atomic_read(&cache
->nr_dirty
));
3281 if (cache
->features
.metadata_version
== 2)
3282 DMEMIT("2 metadata2 ");
3286 if (writethrough_mode(cache
))
3287 DMEMIT("writethrough ");
3289 else if (passthrough_mode(cache
))
3290 DMEMIT("passthrough ");
3292 else if (writeback_mode(cache
))
3293 DMEMIT("writeback ");
3296 DMERR("%s: internal error: unknown io mode: %d",
3297 cache_device_name(cache
), (int) cache
->features
.io_mode
);
3301 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache
->migration_threshold
);
3303 DMEMIT("%s ", dm_cache_policy_get_name(cache
->policy
));
3305 r
= policy_emit_config_values(cache
->policy
, result
, maxlen
, &sz
);
3307 DMERR("%s: policy_emit_config_values returned %d",
3308 cache_device_name(cache
), r
);
3311 if (get_cache_mode(cache
) == CM_READ_ONLY
)
3316 r
= dm_cache_metadata_needs_check(cache
->cmd
, &needs_check
);
3318 if (r
|| needs_check
)
3319 DMEMIT("needs_check ");
3325 case STATUSTYPE_TABLE
:
3326 format_dev_t(buf
, cache
->metadata_dev
->bdev
->bd_dev
);
3328 format_dev_t(buf
, cache
->cache_dev
->bdev
->bd_dev
);
3330 format_dev_t(buf
, cache
->origin_dev
->bdev
->bd_dev
);
3333 for (i
= 0; i
< cache
->nr_ctr_args
- 1; i
++)
3334 DMEMIT(" %s", cache
->ctr_args
[i
]);
3335 if (cache
->nr_ctr_args
)
3336 DMEMIT(" %s", cache
->ctr_args
[cache
->nr_ctr_args
- 1]);
3346 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3347 * the one-past-the-end value.
3349 struct cblock_range
{
3355 * A cache block range can take two forms:
3357 * i) A single cblock, eg. '3456'
3358 * ii) A begin and end cblock with a dash between, eg. 123-234
3360 static int parse_cblock_range(struct cache
*cache
, const char *str
,
3361 struct cblock_range
*result
)
3368 * Try and parse form (ii) first.
3370 r
= sscanf(str
, "%llu-%llu%c", &b
, &e
, &dummy
);
3375 result
->begin
= to_cblock(b
);
3376 result
->end
= to_cblock(e
);
3381 * That didn't work, try form (i).
3383 r
= sscanf(str
, "%llu%c", &b
, &dummy
);
3388 result
->begin
= to_cblock(b
);
3389 result
->end
= to_cblock(from_cblock(result
->begin
) + 1u);
3393 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache
), str
);
3397 static int validate_cblock_range(struct cache
*cache
, struct cblock_range
*range
)
3399 uint64_t b
= from_cblock(range
->begin
);
3400 uint64_t e
= from_cblock(range
->end
);
3401 uint64_t n
= from_cblock(cache
->cache_size
);
3404 DMERR("%s: begin cblock out of range: %llu >= %llu",
3405 cache_device_name(cache
), b
, n
);
3410 DMERR("%s: end cblock out of range: %llu > %llu",
3411 cache_device_name(cache
), e
, n
);
3416 DMERR("%s: invalid cblock range: %llu >= %llu",
3417 cache_device_name(cache
), b
, e
);
3424 static inline dm_cblock_t
cblock_succ(dm_cblock_t b
)
3426 return to_cblock(from_cblock(b
) + 1);
3429 static int request_invalidation(struct cache
*cache
, struct cblock_range
*range
)
3434 * We don't need to do any locking here because we know we're in
3435 * passthrough mode. There's is potential for a race between an
3436 * invalidation triggered by an io and an invalidation message. This
3437 * is harmless, we must not worry if the policy call fails.
3439 while (range
->begin
!= range
->end
) {
3440 r
= invalidate_cblock(cache
, range
->begin
);
3444 range
->begin
= cblock_succ(range
->begin
);
3447 cache
->commit_requested
= true;
3451 static int process_invalidate_cblocks_message(struct cache
*cache
, unsigned count
,
3452 const char **cblock_ranges
)
3456 struct cblock_range range
;
3458 if (!passthrough_mode(cache
)) {
3459 DMERR("%s: cache has to be in passthrough mode for invalidation",
3460 cache_device_name(cache
));
3464 for (i
= 0; i
< count
; i
++) {
3465 r
= parse_cblock_range(cache
, cblock_ranges
[i
], &range
);
3469 r
= validate_cblock_range(cache
, &range
);
3474 * Pass begin and end origin blocks to the worker and wake it.
3476 r
= request_invalidation(cache
, &range
);
3488 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3490 * The key migration_threshold is supported by the cache target core.
3492 static int cache_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
3494 struct cache
*cache
= ti
->private;
3499 if (get_cache_mode(cache
) >= CM_READ_ONLY
) {
3500 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3501 cache_device_name(cache
));
3505 if (!strcasecmp(argv
[0], "invalidate_cblocks"))
3506 return process_invalidate_cblocks_message(cache
, argc
- 1, (const char **) argv
+ 1);
3511 return set_config_value(cache
, argv
[0], argv
[1]);
3514 static int cache_iterate_devices(struct dm_target
*ti
,
3515 iterate_devices_callout_fn fn
, void *data
)
3518 struct cache
*cache
= ti
->private;
3520 r
= fn(ti
, cache
->cache_dev
, 0, get_dev_size(cache
->cache_dev
), data
);
3522 r
= fn(ti
, cache
->origin_dev
, 0, ti
->len
, data
);
3527 static void set_discard_limits(struct cache
*cache
, struct queue_limits
*limits
)
3530 * FIXME: these limits may be incompatible with the cache device
3532 limits
->max_discard_sectors
= min_t(sector_t
, cache
->discard_block_size
* 1024,
3533 cache
->origin_sectors
);
3534 limits
->discard_granularity
= cache
->discard_block_size
<< SECTOR_SHIFT
;
3537 static void cache_io_hints(struct dm_target
*ti
, struct queue_limits
*limits
)
3539 struct cache
*cache
= ti
->private;
3540 uint64_t io_opt_sectors
= limits
->io_opt
>> SECTOR_SHIFT
;
3543 * If the system-determined stacked limits are compatible with the
3544 * cache's blocksize (io_opt is a factor) do not override them.
3546 if (io_opt_sectors
< cache
->sectors_per_block
||
3547 do_div(io_opt_sectors
, cache
->sectors_per_block
)) {
3548 blk_limits_io_min(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3549 blk_limits_io_opt(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3551 set_discard_limits(cache
, limits
);
3554 /*----------------------------------------------------------------*/
3556 static struct target_type cache_target
= {
3558 .version
= {2, 0, 0},
3559 .module
= THIS_MODULE
,
3563 .end_io
= cache_end_io
,
3564 .postsuspend
= cache_postsuspend
,
3565 .preresume
= cache_preresume
,
3566 .resume
= cache_resume
,
3567 .status
= cache_status
,
3568 .message
= cache_message
,
3569 .iterate_devices
= cache_iterate_devices
,
3570 .io_hints
= cache_io_hints
,
3573 static int __init
dm_cache_init(void)
3577 r
= dm_register_target(&cache_target
);
3579 DMERR("cache target registration failed: %d", r
);
3583 migration_cache
= KMEM_CACHE(dm_cache_migration
, 0);
3584 if (!migration_cache
) {
3585 dm_unregister_target(&cache_target
);
3592 static void __exit
dm_cache_exit(void)
3594 dm_unregister_target(&cache_target
);
3595 kmem_cache_destroy(migration_cache
);
3598 module_init(dm_cache_init
);
3599 module_exit(dm_cache_exit
);
3601 MODULE_DESCRIPTION(DM_NAME
" cache target");
3602 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3603 MODULE_LICENSE("GPL");