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
)
78 spin_lock_irq(&iot
->lock
);
79 r
= __iot_idle_for(iot
, jifs
);
80 spin_unlock_irq(&iot
->lock
);
85 static void iot_io_begin(struct io_tracker
*iot
, sector_t len
)
87 spin_lock_irq(&iot
->lock
);
88 iot
->in_flight
+= len
;
89 spin_unlock_irq(&iot
->lock
);
92 static void __iot_io_end(struct io_tracker
*iot
, sector_t len
)
97 iot
->in_flight
-= len
;
99 iot
->idle_time
= jiffies
;
102 static void iot_io_end(struct io_tracker
*iot
, sector_t len
)
106 spin_lock_irqsave(&iot
->lock
, flags
);
107 __iot_io_end(iot
, len
);
108 spin_unlock_irqrestore(&iot
->lock
, flags
);
111 /*----------------------------------------------------------------*/
114 * Represents a chunk of future work. 'input' allows continuations to pass
115 * values between themselves, typically error values.
117 struct continuation
{
118 struct work_struct ws
;
122 static inline void init_continuation(struct continuation
*k
,
123 void (*fn
)(struct work_struct
*))
125 INIT_WORK(&k
->ws
, fn
);
129 static inline void queue_continuation(struct workqueue_struct
*wq
,
130 struct continuation
*k
)
132 queue_work(wq
, &k
->ws
);
135 /*----------------------------------------------------------------*/
138 * The batcher collects together pieces of work that need a particular
139 * operation to occur before they can proceed (typically a commit).
143 * The operation that everyone is waiting for.
145 blk_status_t (*commit_op
)(void *context
);
146 void *commit_context
;
149 * This is how bios should be issued once the commit op is complete
150 * (accounted_request).
152 void (*issue_op
)(struct bio
*bio
, void *context
);
156 * Queued work gets put on here after commit.
158 struct workqueue_struct
*wq
;
161 struct list_head work_items
;
162 struct bio_list bios
;
163 struct work_struct commit_work
;
165 bool commit_scheduled
;
168 static void __commit(struct work_struct
*_ws
)
170 struct batcher
*b
= container_of(_ws
, struct batcher
, commit_work
);
172 struct list_head work_items
;
173 struct work_struct
*ws
, *tmp
;
174 struct continuation
*k
;
176 struct bio_list bios
;
178 INIT_LIST_HEAD(&work_items
);
179 bio_list_init(&bios
);
182 * We have to grab these before the commit_op to avoid a race
185 spin_lock_irq(&b
->lock
);
186 list_splice_init(&b
->work_items
, &work_items
);
187 bio_list_merge(&bios
, &b
->bios
);
188 bio_list_init(&b
->bios
);
189 b
->commit_scheduled
= false;
190 spin_unlock_irq(&b
->lock
);
192 r
= b
->commit_op(b
->commit_context
);
194 list_for_each_entry_safe(ws
, tmp
, &work_items
, entry
) {
195 k
= container_of(ws
, struct continuation
, ws
);
197 INIT_LIST_HEAD(&ws
->entry
); /* to avoid a WARN_ON */
198 queue_work(b
->wq
, ws
);
201 while ((bio
= bio_list_pop(&bios
))) {
206 b
->issue_op(bio
, b
->issue_context
);
210 static void batcher_init(struct batcher
*b
,
211 blk_status_t (*commit_op
)(void *),
212 void *commit_context
,
213 void (*issue_op
)(struct bio
*bio
, void *),
215 struct workqueue_struct
*wq
)
217 b
->commit_op
= commit_op
;
218 b
->commit_context
= commit_context
;
219 b
->issue_op
= issue_op
;
220 b
->issue_context
= issue_context
;
223 spin_lock_init(&b
->lock
);
224 INIT_LIST_HEAD(&b
->work_items
);
225 bio_list_init(&b
->bios
);
226 INIT_WORK(&b
->commit_work
, __commit
);
227 b
->commit_scheduled
= false;
230 static void async_commit(struct batcher
*b
)
232 queue_work(b
->wq
, &b
->commit_work
);
235 static void continue_after_commit(struct batcher
*b
, struct continuation
*k
)
237 bool commit_scheduled
;
239 spin_lock_irq(&b
->lock
);
240 commit_scheduled
= b
->commit_scheduled
;
241 list_add_tail(&k
->ws
.entry
, &b
->work_items
);
242 spin_unlock_irq(&b
->lock
);
244 if (commit_scheduled
)
249 * Bios are errored if commit failed.
251 static void issue_after_commit(struct batcher
*b
, struct bio
*bio
)
253 bool commit_scheduled
;
255 spin_lock_irq(&b
->lock
);
256 commit_scheduled
= b
->commit_scheduled
;
257 bio_list_add(&b
->bios
, bio
);
258 spin_unlock_irq(&b
->lock
);
260 if (commit_scheduled
)
265 * Call this if some urgent work is waiting for the commit to complete.
267 static void schedule_commit(struct batcher
*b
)
271 spin_lock_irq(&b
->lock
);
272 immediate
= !list_empty(&b
->work_items
) || !bio_list_empty(&b
->bios
);
273 b
->commit_scheduled
= true;
274 spin_unlock_irq(&b
->lock
);
281 * There are a couple of places where we let a bio run, but want to do some
282 * work before calling its endio function. We do this by temporarily
283 * changing the endio fn.
285 struct dm_hook_info
{
286 bio_end_io_t
*bi_end_io
;
289 static void dm_hook_bio(struct dm_hook_info
*h
, struct bio
*bio
,
290 bio_end_io_t
*bi_end_io
, void *bi_private
)
292 h
->bi_end_io
= bio
->bi_end_io
;
294 bio
->bi_end_io
= bi_end_io
;
295 bio
->bi_private
= bi_private
;
298 static void dm_unhook_bio(struct dm_hook_info
*h
, struct bio
*bio
)
300 bio
->bi_end_io
= h
->bi_end_io
;
303 /*----------------------------------------------------------------*/
305 #define MIGRATION_POOL_SIZE 128
306 #define COMMIT_PERIOD HZ
307 #define MIGRATION_COUNT_WINDOW 10
310 * The block size of the device holding cache data must be
311 * between 32KB and 1GB.
313 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
314 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
316 enum cache_metadata_mode
{
317 CM_WRITE
, /* metadata may be changed */
318 CM_READ_ONLY
, /* metadata may not be changed */
324 * Data is written to cached blocks only. These blocks are marked
325 * dirty. If you lose the cache device you will lose data.
326 * Potential performance increase for both reads and writes.
331 * Data is written to both cache and origin. Blocks are never
332 * dirty. Potential performance benfit for reads only.
337 * A degraded mode useful for various cache coherency situations
338 * (eg, rolling back snapshots). Reads and writes always go to the
339 * origin. If a write goes to a cached oblock, then the cache
340 * block is invalidated.
345 struct cache_features
{
346 enum cache_metadata_mode mode
;
347 enum cache_io_mode io_mode
;
348 unsigned metadata_version
;
349 bool discard_passdown
:1;
360 atomic_t copies_avoided
;
361 atomic_t cache_cell_clash
;
362 atomic_t commit_count
;
363 atomic_t discard_count
;
367 struct dm_target
*ti
;
371 * Fields for converting from sectors to blocks.
373 int sectors_per_block_shift
;
374 sector_t sectors_per_block
;
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 * Invalidation fields.
407 spinlock_t invalidation_lock
;
408 struct list_head invalidation_requests
;
410 sector_t migration_threshold
;
411 wait_queue_head_t migration_wait
;
412 atomic_t nr_allocated_migrations
;
415 * The number of in flight migrations that are performing
416 * background io. eg, promotion, writeback.
418 atomic_t nr_io_migrations
;
420 struct bio_list deferred_bios
;
422 struct rw_semaphore quiesce_lock
;
425 * origin_blocks entries, discarded if set.
427 dm_dblock_t discard_nr_blocks
;
428 unsigned long *discard_bitset
;
429 uint32_t discard_block_size
; /* a power of 2 times sectors per block */
432 * Rather than reconstructing the table line for the status we just
433 * save it and regurgitate.
435 unsigned nr_ctr_args
;
436 const char **ctr_args
;
438 struct dm_kcopyd_client
*copier
;
439 struct work_struct deferred_bio_worker
;
440 struct work_struct migration_worker
;
441 struct workqueue_struct
*wq
;
442 struct delayed_work waker
;
443 struct dm_bio_prison_v2
*prison
;
446 * cache_size entries, dirty if set
448 unsigned long *dirty_bitset
;
451 unsigned policy_nr_args
;
452 struct dm_cache_policy
*policy
;
455 * Cache features such as write-through.
457 struct cache_features features
;
459 struct cache_stats stats
;
461 bool need_tick_bio
:1;
464 bool commit_requested
:1;
465 bool loaded_mappings
:1;
466 bool loaded_discards
:1;
468 struct rw_semaphore background_work_lock
;
470 struct batcher committer
;
471 struct work_struct commit_ws
;
473 struct io_tracker tracker
;
475 mempool_t migration_pool
;
480 struct per_bio_data
{
483 struct dm_bio_prison_cell_v2
*cell
;
484 struct dm_hook_info hook_info
;
488 struct dm_cache_migration
{
489 struct continuation k
;
492 struct policy_work
*op
;
493 struct bio
*overwrite_bio
;
494 struct dm_bio_prison_cell_v2
*cell
;
496 dm_cblock_t invalidate_cblock
;
497 dm_oblock_t invalidate_oblock
;
500 /*----------------------------------------------------------------*/
502 static bool writethrough_mode(struct cache
*cache
)
504 return cache
->features
.io_mode
== CM_IO_WRITETHROUGH
;
507 static bool writeback_mode(struct cache
*cache
)
509 return cache
->features
.io_mode
== CM_IO_WRITEBACK
;
512 static inline bool passthrough_mode(struct cache
*cache
)
514 return unlikely(cache
->features
.io_mode
== CM_IO_PASSTHROUGH
);
517 /*----------------------------------------------------------------*/
519 static void wake_deferred_bio_worker(struct cache
*cache
)
521 queue_work(cache
->wq
, &cache
->deferred_bio_worker
);
524 static void wake_migration_worker(struct cache
*cache
)
526 if (passthrough_mode(cache
))
529 queue_work(cache
->wq
, &cache
->migration_worker
);
532 /*----------------------------------------------------------------*/
534 static struct dm_bio_prison_cell_v2
*alloc_prison_cell(struct cache
*cache
)
536 return dm_bio_prison_alloc_cell_v2(cache
->prison
, GFP_NOIO
);
539 static void free_prison_cell(struct cache
*cache
, struct dm_bio_prison_cell_v2
*cell
)
541 dm_bio_prison_free_cell_v2(cache
->prison
, cell
);
544 static struct dm_cache_migration
*alloc_migration(struct cache
*cache
)
546 struct dm_cache_migration
*mg
;
548 mg
= mempool_alloc(&cache
->migration_pool
, GFP_NOIO
);
550 memset(mg
, 0, sizeof(*mg
));
553 atomic_inc(&cache
->nr_allocated_migrations
);
558 static void free_migration(struct dm_cache_migration
*mg
)
560 struct cache
*cache
= mg
->cache
;
562 if (atomic_dec_and_test(&cache
->nr_allocated_migrations
))
563 wake_up(&cache
->migration_wait
);
565 mempool_free(mg
, &cache
->migration_pool
);
568 /*----------------------------------------------------------------*/
570 static inline dm_oblock_t
oblock_succ(dm_oblock_t b
)
572 return to_oblock(from_oblock(b
) + 1ull);
575 static void build_key(dm_oblock_t begin
, dm_oblock_t end
, struct dm_cell_key_v2
*key
)
579 key
->block_begin
= from_oblock(begin
);
580 key
->block_end
= from_oblock(end
);
584 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
585 * level 1 which prevents *both* READs and WRITEs.
587 #define WRITE_LOCK_LEVEL 0
588 #define READ_WRITE_LOCK_LEVEL 1
590 static unsigned lock_level(struct bio
*bio
)
592 return bio_data_dir(bio
) == WRITE
?
594 READ_WRITE_LOCK_LEVEL
;
597 /*----------------------------------------------------------------
599 *--------------------------------------------------------------*/
601 static struct per_bio_data
*get_per_bio_data(struct bio
*bio
)
603 struct per_bio_data
*pb
= dm_per_bio_data(bio
, sizeof(struct per_bio_data
));
608 static struct per_bio_data
*init_per_bio_data(struct bio
*bio
)
610 struct per_bio_data
*pb
= get_per_bio_data(bio
);
613 pb
->req_nr
= dm_bio_get_target_bio_nr(bio
);
620 /*----------------------------------------------------------------*/
622 static void defer_bio(struct cache
*cache
, struct bio
*bio
)
624 spin_lock_irq(&cache
->lock
);
625 bio_list_add(&cache
->deferred_bios
, bio
);
626 spin_unlock_irq(&cache
->lock
);
628 wake_deferred_bio_worker(cache
);
631 static void defer_bios(struct cache
*cache
, struct bio_list
*bios
)
633 spin_lock_irq(&cache
->lock
);
634 bio_list_merge(&cache
->deferred_bios
, bios
);
636 spin_unlock_irq(&cache
->lock
);
638 wake_deferred_bio_worker(cache
);
641 /*----------------------------------------------------------------*/
643 static bool bio_detain_shared(struct cache
*cache
, dm_oblock_t oblock
, struct bio
*bio
)
646 struct per_bio_data
*pb
;
647 struct dm_cell_key_v2 key
;
648 dm_oblock_t end
= to_oblock(from_oblock(oblock
) + 1ULL);
649 struct dm_bio_prison_cell_v2
*cell_prealloc
, *cell
;
651 cell_prealloc
= alloc_prison_cell(cache
); /* FIXME: allow wait if calling from worker */
653 build_key(oblock
, end
, &key
);
654 r
= dm_cell_get_v2(cache
->prison
, &key
, lock_level(bio
), bio
, cell_prealloc
, &cell
);
657 * Failed to get the lock.
659 free_prison_cell(cache
, cell_prealloc
);
663 if (cell
!= cell_prealloc
)
664 free_prison_cell(cache
, cell_prealloc
);
666 pb
= get_per_bio_data(bio
);
672 /*----------------------------------------------------------------*/
674 static bool is_dirty(struct cache
*cache
, dm_cblock_t b
)
676 return test_bit(from_cblock(b
), cache
->dirty_bitset
);
679 static void set_dirty(struct cache
*cache
, dm_cblock_t cblock
)
681 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
682 atomic_inc(&cache
->nr_dirty
);
683 policy_set_dirty(cache
->policy
, cblock
);
688 * These two are called when setting after migrations to force the policy
689 * and dirty bitset to be in sync.
691 static void force_set_dirty(struct cache
*cache
, dm_cblock_t cblock
)
693 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
))
694 atomic_inc(&cache
->nr_dirty
);
695 policy_set_dirty(cache
->policy
, cblock
);
698 static void force_clear_dirty(struct cache
*cache
, dm_cblock_t cblock
)
700 if (test_and_clear_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
701 if (atomic_dec_return(&cache
->nr_dirty
) == 0)
702 dm_table_event(cache
->ti
->table
);
705 policy_clear_dirty(cache
->policy
, cblock
);
708 /*----------------------------------------------------------------*/
710 static bool block_size_is_power_of_two(struct cache
*cache
)
712 return cache
->sectors_per_block_shift
>= 0;
715 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
716 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
719 static dm_block_t
block_div(dm_block_t b
, uint32_t n
)
726 static dm_block_t
oblocks_per_dblock(struct cache
*cache
)
728 dm_block_t oblocks
= cache
->discard_block_size
;
730 if (block_size_is_power_of_two(cache
))
731 oblocks
>>= cache
->sectors_per_block_shift
;
733 oblocks
= block_div(oblocks
, cache
->sectors_per_block
);
738 static dm_dblock_t
oblock_to_dblock(struct cache
*cache
, dm_oblock_t oblock
)
740 return to_dblock(block_div(from_oblock(oblock
),
741 oblocks_per_dblock(cache
)));
744 static void set_discard(struct cache
*cache
, dm_dblock_t b
)
746 BUG_ON(from_dblock(b
) >= from_dblock(cache
->discard_nr_blocks
));
747 atomic_inc(&cache
->stats
.discard_count
);
749 spin_lock_irq(&cache
->lock
);
750 set_bit(from_dblock(b
), cache
->discard_bitset
);
751 spin_unlock_irq(&cache
->lock
);
754 static void clear_discard(struct cache
*cache
, dm_dblock_t b
)
756 spin_lock_irq(&cache
->lock
);
757 clear_bit(from_dblock(b
), cache
->discard_bitset
);
758 spin_unlock_irq(&cache
->lock
);
761 static bool is_discarded(struct cache
*cache
, dm_dblock_t b
)
764 spin_lock_irq(&cache
->lock
);
765 r
= test_bit(from_dblock(b
), cache
->discard_bitset
);
766 spin_unlock_irq(&cache
->lock
);
771 static bool is_discarded_oblock(struct cache
*cache
, dm_oblock_t b
)
774 spin_lock_irq(&cache
->lock
);
775 r
= test_bit(from_dblock(oblock_to_dblock(cache
, b
)),
776 cache
->discard_bitset
);
777 spin_unlock_irq(&cache
->lock
);
782 /*----------------------------------------------------------------
784 *--------------------------------------------------------------*/
785 static void remap_to_origin(struct cache
*cache
, struct bio
*bio
)
787 bio_set_dev(bio
, cache
->origin_dev
->bdev
);
790 static void remap_to_cache(struct cache
*cache
, struct bio
*bio
,
793 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
794 sector_t block
= from_cblock(cblock
);
796 bio_set_dev(bio
, cache
->cache_dev
->bdev
);
797 if (!block_size_is_power_of_two(cache
))
798 bio
->bi_iter
.bi_sector
=
799 (block
* cache
->sectors_per_block
) +
800 sector_div(bi_sector
, cache
->sectors_per_block
);
802 bio
->bi_iter
.bi_sector
=
803 (block
<< cache
->sectors_per_block_shift
) |
804 (bi_sector
& (cache
->sectors_per_block
- 1));
807 static void check_if_tick_bio_needed(struct cache
*cache
, struct bio
*bio
)
809 struct per_bio_data
*pb
;
811 spin_lock_irq(&cache
->lock
);
812 if (cache
->need_tick_bio
&& !op_is_flush(bio
->bi_opf
) &&
813 bio_op(bio
) != REQ_OP_DISCARD
) {
814 pb
= get_per_bio_data(bio
);
816 cache
->need_tick_bio
= false;
818 spin_unlock_irq(&cache
->lock
);
821 static void __remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
822 dm_oblock_t oblock
, bool bio_has_pbd
)
825 check_if_tick_bio_needed(cache
, bio
);
826 remap_to_origin(cache
, bio
);
827 if (bio_data_dir(bio
) == WRITE
)
828 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
831 static void remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
834 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
835 __remap_to_origin_clear_discard(cache
, bio
, oblock
, true);
838 static void remap_to_cache_dirty(struct cache
*cache
, struct bio
*bio
,
839 dm_oblock_t oblock
, dm_cblock_t cblock
)
841 check_if_tick_bio_needed(cache
, bio
);
842 remap_to_cache(cache
, bio
, cblock
);
843 if (bio_data_dir(bio
) == WRITE
) {
844 set_dirty(cache
, cblock
);
845 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
849 static dm_oblock_t
get_bio_block(struct cache
*cache
, struct bio
*bio
)
851 sector_t block_nr
= bio
->bi_iter
.bi_sector
;
853 if (!block_size_is_power_of_two(cache
))
854 (void) sector_div(block_nr
, cache
->sectors_per_block
);
856 block_nr
>>= cache
->sectors_per_block_shift
;
858 return to_oblock(block_nr
);
861 static bool accountable_bio(struct cache
*cache
, struct bio
*bio
)
863 return bio_op(bio
) != REQ_OP_DISCARD
;
866 static void accounted_begin(struct cache
*cache
, struct bio
*bio
)
868 struct per_bio_data
*pb
;
870 if (accountable_bio(cache
, bio
)) {
871 pb
= get_per_bio_data(bio
);
872 pb
->len
= bio_sectors(bio
);
873 iot_io_begin(&cache
->tracker
, pb
->len
);
877 static void accounted_complete(struct cache
*cache
, struct bio
*bio
)
879 struct per_bio_data
*pb
= get_per_bio_data(bio
);
881 iot_io_end(&cache
->tracker
, pb
->len
);
884 static void accounted_request(struct cache
*cache
, struct bio
*bio
)
886 accounted_begin(cache
, bio
);
887 submit_bio_noacct(bio
);
890 static void issue_op(struct bio
*bio
, void *context
)
892 struct cache
*cache
= context
;
893 accounted_request(cache
, bio
);
897 * When running in writethrough mode we need to send writes to clean blocks
898 * to both the cache and origin devices. Clone the bio and send them in parallel.
900 static void remap_to_origin_and_cache(struct cache
*cache
, struct bio
*bio
,
901 dm_oblock_t oblock
, dm_cblock_t cblock
)
903 struct bio
*origin_bio
= bio_clone_fast(bio
, GFP_NOIO
, &cache
->bs
);
907 bio_chain(origin_bio
, bio
);
909 * Passing false to __remap_to_origin_clear_discard() skips
910 * all code that might use per_bio_data (since clone doesn't have it)
912 __remap_to_origin_clear_discard(cache
, origin_bio
, oblock
, false);
913 submit_bio(origin_bio
);
915 remap_to_cache(cache
, bio
, cblock
);
918 /*----------------------------------------------------------------
920 *--------------------------------------------------------------*/
921 static enum cache_metadata_mode
get_cache_mode(struct cache
*cache
)
923 return cache
->features
.mode
;
926 static const char *cache_device_name(struct cache
*cache
)
928 return dm_device_name(dm_table_get_md(cache
->ti
->table
));
931 static void notify_mode_switch(struct cache
*cache
, enum cache_metadata_mode mode
)
933 const char *descs
[] = {
939 dm_table_event(cache
->ti
->table
);
940 DMINFO("%s: switching cache to %s mode",
941 cache_device_name(cache
), descs
[(int)mode
]);
944 static void set_cache_mode(struct cache
*cache
, enum cache_metadata_mode new_mode
)
947 enum cache_metadata_mode old_mode
= get_cache_mode(cache
);
949 if (dm_cache_metadata_needs_check(cache
->cmd
, &needs_check
)) {
950 DMERR("%s: unable to read needs_check flag, setting failure mode.",
951 cache_device_name(cache
));
955 if (new_mode
== CM_WRITE
&& needs_check
) {
956 DMERR("%s: unable to switch cache to write mode until repaired.",
957 cache_device_name(cache
));
958 if (old_mode
!= new_mode
)
961 new_mode
= CM_READ_ONLY
;
964 /* Never move out of fail mode */
965 if (old_mode
== CM_FAIL
)
971 dm_cache_metadata_set_read_only(cache
->cmd
);
975 dm_cache_metadata_set_read_write(cache
->cmd
);
979 cache
->features
.mode
= new_mode
;
981 if (new_mode
!= old_mode
)
982 notify_mode_switch(cache
, new_mode
);
985 static void abort_transaction(struct cache
*cache
)
987 const char *dev_name
= cache_device_name(cache
);
989 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
992 if (dm_cache_metadata_set_needs_check(cache
->cmd
)) {
993 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name
);
994 set_cache_mode(cache
, CM_FAIL
);
997 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name
);
998 if (dm_cache_metadata_abort(cache
->cmd
)) {
999 DMERR("%s: failed to abort metadata transaction", dev_name
);
1000 set_cache_mode(cache
, CM_FAIL
);
1004 static void metadata_operation_failed(struct cache
*cache
, const char *op
, int r
)
1006 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1007 cache_device_name(cache
), op
, r
);
1008 abort_transaction(cache
);
1009 set_cache_mode(cache
, CM_READ_ONLY
);
1012 /*----------------------------------------------------------------*/
1014 static void load_stats(struct cache
*cache
)
1016 struct dm_cache_statistics stats
;
1018 dm_cache_metadata_get_stats(cache
->cmd
, &stats
);
1019 atomic_set(&cache
->stats
.read_hit
, stats
.read_hits
);
1020 atomic_set(&cache
->stats
.read_miss
, stats
.read_misses
);
1021 atomic_set(&cache
->stats
.write_hit
, stats
.write_hits
);
1022 atomic_set(&cache
->stats
.write_miss
, stats
.write_misses
);
1025 static void save_stats(struct cache
*cache
)
1027 struct dm_cache_statistics stats
;
1029 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1032 stats
.read_hits
= atomic_read(&cache
->stats
.read_hit
);
1033 stats
.read_misses
= atomic_read(&cache
->stats
.read_miss
);
1034 stats
.write_hits
= atomic_read(&cache
->stats
.write_hit
);
1035 stats
.write_misses
= atomic_read(&cache
->stats
.write_miss
);
1037 dm_cache_metadata_set_stats(cache
->cmd
, &stats
);
1040 static void update_stats(struct cache_stats
*stats
, enum policy_operation op
)
1043 case POLICY_PROMOTE
:
1044 atomic_inc(&stats
->promotion
);
1048 atomic_inc(&stats
->demotion
);
1051 case POLICY_WRITEBACK
:
1052 atomic_inc(&stats
->writeback
);
1057 /*----------------------------------------------------------------
1058 * Migration processing
1060 * Migration covers moving data from the origin device to the cache, or
1062 *--------------------------------------------------------------*/
1064 static void inc_io_migrations(struct cache
*cache
)
1066 atomic_inc(&cache
->nr_io_migrations
);
1069 static void dec_io_migrations(struct cache
*cache
)
1071 atomic_dec(&cache
->nr_io_migrations
);
1074 static bool discard_or_flush(struct bio
*bio
)
1076 return bio_op(bio
) == REQ_OP_DISCARD
|| op_is_flush(bio
->bi_opf
);
1079 static void calc_discard_block_range(struct cache
*cache
, struct bio
*bio
,
1080 dm_dblock_t
*b
, dm_dblock_t
*e
)
1082 sector_t sb
= bio
->bi_iter
.bi_sector
;
1083 sector_t se
= bio_end_sector(bio
);
1085 *b
= to_dblock(dm_sector_div_up(sb
, cache
->discard_block_size
));
1087 if (se
- sb
< cache
->discard_block_size
)
1090 *e
= to_dblock(block_div(se
, cache
->discard_block_size
));
1093 /*----------------------------------------------------------------*/
1095 static void prevent_background_work(struct cache
*cache
)
1098 down_write(&cache
->background_work_lock
);
1102 static void allow_background_work(struct cache
*cache
)
1105 up_write(&cache
->background_work_lock
);
1109 static bool background_work_begin(struct cache
*cache
)
1114 r
= down_read_trylock(&cache
->background_work_lock
);
1120 static void background_work_end(struct cache
*cache
)
1123 up_read(&cache
->background_work_lock
);
1127 /*----------------------------------------------------------------*/
1129 static bool bio_writes_complete_block(struct cache
*cache
, struct bio
*bio
)
1131 return (bio_data_dir(bio
) == WRITE
) &&
1132 (bio
->bi_iter
.bi_size
== (cache
->sectors_per_block
<< SECTOR_SHIFT
));
1135 static bool optimisable_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
)
1137 return writeback_mode(cache
) &&
1138 (is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
));
1141 static void quiesce(struct dm_cache_migration
*mg
,
1142 void (*continuation
)(struct work_struct
*))
1144 init_continuation(&mg
->k
, continuation
);
1145 dm_cell_quiesce_v2(mg
->cache
->prison
, mg
->cell
, &mg
->k
.ws
);
1148 static struct dm_cache_migration
*ws_to_mg(struct work_struct
*ws
)
1150 struct continuation
*k
= container_of(ws
, struct continuation
, ws
);
1151 return container_of(k
, struct dm_cache_migration
, k
);
1154 static void copy_complete(int read_err
, unsigned long write_err
, void *context
)
1156 struct dm_cache_migration
*mg
= container_of(context
, struct dm_cache_migration
, k
);
1158 if (read_err
|| write_err
)
1159 mg
->k
.input
= BLK_STS_IOERR
;
1161 queue_continuation(mg
->cache
->wq
, &mg
->k
);
1164 static void copy(struct dm_cache_migration
*mg
, bool promote
)
1166 struct dm_io_region o_region
, c_region
;
1167 struct cache
*cache
= mg
->cache
;
1169 o_region
.bdev
= cache
->origin_dev
->bdev
;
1170 o_region
.sector
= from_oblock(mg
->op
->oblock
) * cache
->sectors_per_block
;
1171 o_region
.count
= cache
->sectors_per_block
;
1173 c_region
.bdev
= cache
->cache_dev
->bdev
;
1174 c_region
.sector
= from_cblock(mg
->op
->cblock
) * cache
->sectors_per_block
;
1175 c_region
.count
= cache
->sectors_per_block
;
1178 dm_kcopyd_copy(cache
->copier
, &o_region
, 1, &c_region
, 0, copy_complete
, &mg
->k
);
1180 dm_kcopyd_copy(cache
->copier
, &c_region
, 1, &o_region
, 0, copy_complete
, &mg
->k
);
1183 static void bio_drop_shared_lock(struct cache
*cache
, struct bio
*bio
)
1185 struct per_bio_data
*pb
= get_per_bio_data(bio
);
1187 if (pb
->cell
&& dm_cell_put_v2(cache
->prison
, pb
->cell
))
1188 free_prison_cell(cache
, pb
->cell
);
1192 static void overwrite_endio(struct bio
*bio
)
1194 struct dm_cache_migration
*mg
= bio
->bi_private
;
1195 struct cache
*cache
= mg
->cache
;
1196 struct per_bio_data
*pb
= get_per_bio_data(bio
);
1198 dm_unhook_bio(&pb
->hook_info
, bio
);
1201 mg
->k
.input
= bio
->bi_status
;
1203 queue_continuation(cache
->wq
, &mg
->k
);
1206 static void overwrite(struct dm_cache_migration
*mg
,
1207 void (*continuation
)(struct work_struct
*))
1209 struct bio
*bio
= mg
->overwrite_bio
;
1210 struct per_bio_data
*pb
= get_per_bio_data(bio
);
1212 dm_hook_bio(&pb
->hook_info
, bio
, overwrite_endio
, mg
);
1215 * The overwrite bio is part of the copy operation, as such it does
1216 * not set/clear discard or dirty flags.
1218 if (mg
->op
->op
== POLICY_PROMOTE
)
1219 remap_to_cache(mg
->cache
, bio
, mg
->op
->cblock
);
1221 remap_to_origin(mg
->cache
, bio
);
1223 init_continuation(&mg
->k
, continuation
);
1224 accounted_request(mg
->cache
, bio
);
1230 * 1) exclusive lock preventing WRITEs
1232 * 3) copy or issue overwrite bio
1233 * 4) upgrade to exclusive lock preventing READs and WRITEs
1235 * 6) update metadata and commit
1238 static void mg_complete(struct dm_cache_migration
*mg
, bool success
)
1240 struct bio_list bios
;
1241 struct cache
*cache
= mg
->cache
;
1242 struct policy_work
*op
= mg
->op
;
1243 dm_cblock_t cblock
= op
->cblock
;
1246 update_stats(&cache
->stats
, op
->op
);
1249 case POLICY_PROMOTE
:
1250 clear_discard(cache
, oblock_to_dblock(cache
, op
->oblock
));
1251 policy_complete_background_work(cache
->policy
, op
, success
);
1253 if (mg
->overwrite_bio
) {
1255 force_set_dirty(cache
, cblock
);
1256 else if (mg
->k
.input
)
1257 mg
->overwrite_bio
->bi_status
= mg
->k
.input
;
1259 mg
->overwrite_bio
->bi_status
= BLK_STS_IOERR
;
1260 bio_endio(mg
->overwrite_bio
);
1263 force_clear_dirty(cache
, cblock
);
1264 dec_io_migrations(cache
);
1270 * We clear dirty here to update the nr_dirty counter.
1273 force_clear_dirty(cache
, cblock
);
1274 policy_complete_background_work(cache
->policy
, op
, success
);
1275 dec_io_migrations(cache
);
1278 case POLICY_WRITEBACK
:
1280 force_clear_dirty(cache
, cblock
);
1281 policy_complete_background_work(cache
->policy
, op
, success
);
1282 dec_io_migrations(cache
);
1286 bio_list_init(&bios
);
1288 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1289 free_prison_cell(cache
, mg
->cell
);
1293 defer_bios(cache
, &bios
);
1294 wake_migration_worker(cache
);
1296 background_work_end(cache
);
1299 static void mg_success(struct work_struct
*ws
)
1301 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1302 mg_complete(mg
, mg
->k
.input
== 0);
1305 static void mg_update_metadata(struct work_struct
*ws
)
1308 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1309 struct cache
*cache
= mg
->cache
;
1310 struct policy_work
*op
= mg
->op
;
1313 case POLICY_PROMOTE
:
1314 r
= dm_cache_insert_mapping(cache
->cmd
, op
->cblock
, op
->oblock
);
1316 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1317 cache_device_name(cache
));
1318 metadata_operation_failed(cache
, "dm_cache_insert_mapping", r
);
1320 mg_complete(mg
, false);
1323 mg_complete(mg
, true);
1327 r
= dm_cache_remove_mapping(cache
->cmd
, op
->cblock
);
1329 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1330 cache_device_name(cache
));
1331 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1333 mg_complete(mg
, false);
1338 * It would be nice if we only had to commit when a REQ_FLUSH
1339 * comes through. But there's one scenario that we have to
1342 * - vblock x in a cache block
1344 * - cache block gets reallocated and over written
1347 * When we recover, because there was no commit the cache will
1348 * rollback to having the data for vblock x in the cache block.
1349 * But the cache block has since been overwritten, so it'll end
1350 * up pointing to data that was never in 'x' during the history
1353 * To avoid this issue we require a commit as part of the
1354 * demotion operation.
1356 init_continuation(&mg
->k
, mg_success
);
1357 continue_after_commit(&cache
->committer
, &mg
->k
);
1358 schedule_commit(&cache
->committer
);
1361 case POLICY_WRITEBACK
:
1362 mg_complete(mg
, true);
1367 static void mg_update_metadata_after_copy(struct work_struct
*ws
)
1369 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1372 * Did the copy succeed?
1375 mg_complete(mg
, false);
1377 mg_update_metadata(ws
);
1380 static void mg_upgrade_lock(struct work_struct
*ws
)
1383 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1386 * Did the copy succeed?
1389 mg_complete(mg
, false);
1393 * Now we want the lock to prevent both reads and writes.
1395 r
= dm_cell_lock_promote_v2(mg
->cache
->prison
, mg
->cell
,
1396 READ_WRITE_LOCK_LEVEL
);
1398 mg_complete(mg
, false);
1401 quiesce(mg
, mg_update_metadata
);
1404 mg_update_metadata(ws
);
1408 static void mg_full_copy(struct work_struct
*ws
)
1410 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1411 struct cache
*cache
= mg
->cache
;
1412 struct policy_work
*op
= mg
->op
;
1413 bool is_policy_promote
= (op
->op
== POLICY_PROMOTE
);
1415 if ((!is_policy_promote
&& !is_dirty(cache
, op
->cblock
)) ||
1416 is_discarded_oblock(cache
, op
->oblock
)) {
1417 mg_upgrade_lock(ws
);
1421 init_continuation(&mg
->k
, mg_upgrade_lock
);
1422 copy(mg
, is_policy_promote
);
1425 static void mg_copy(struct work_struct
*ws
)
1427 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1429 if (mg
->overwrite_bio
) {
1431 * No exclusive lock was held when we last checked if the bio
1432 * was optimisable. So we have to check again in case things
1433 * have changed (eg, the block may no longer be discarded).
1435 if (!optimisable_bio(mg
->cache
, mg
->overwrite_bio
, mg
->op
->oblock
)) {
1437 * Fallback to a real full copy after doing some tidying up.
1439 bool rb
= bio_detain_shared(mg
->cache
, mg
->op
->oblock
, mg
->overwrite_bio
);
1440 BUG_ON(rb
); /* An exclussive lock must _not_ be held for this block */
1441 mg
->overwrite_bio
= NULL
;
1442 inc_io_migrations(mg
->cache
);
1448 * It's safe to do this here, even though it's new data
1449 * because all IO has been locked out of the block.
1451 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1452 * so _not_ using mg_upgrade_lock() as continutation.
1454 overwrite(mg
, mg_update_metadata_after_copy
);
1460 static int mg_lock_writes(struct dm_cache_migration
*mg
)
1463 struct dm_cell_key_v2 key
;
1464 struct cache
*cache
= mg
->cache
;
1465 struct dm_bio_prison_cell_v2
*prealloc
;
1467 prealloc
= alloc_prison_cell(cache
);
1470 * Prevent writes to the block, but allow reads to continue.
1471 * Unless we're using an overwrite bio, in which case we lock
1474 build_key(mg
->op
->oblock
, oblock_succ(mg
->op
->oblock
), &key
);
1475 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1476 mg
->overwrite_bio
? READ_WRITE_LOCK_LEVEL
: WRITE_LOCK_LEVEL
,
1477 prealloc
, &mg
->cell
);
1479 free_prison_cell(cache
, prealloc
);
1480 mg_complete(mg
, false);
1484 if (mg
->cell
!= prealloc
)
1485 free_prison_cell(cache
, prealloc
);
1490 quiesce(mg
, mg_copy
);
1495 static int mg_start(struct cache
*cache
, struct policy_work
*op
, struct bio
*bio
)
1497 struct dm_cache_migration
*mg
;
1499 if (!background_work_begin(cache
)) {
1500 policy_complete_background_work(cache
->policy
, op
, false);
1504 mg
= alloc_migration(cache
);
1507 mg
->overwrite_bio
= bio
;
1510 inc_io_migrations(cache
);
1512 return mg_lock_writes(mg
);
1515 /*----------------------------------------------------------------
1516 * invalidation processing
1517 *--------------------------------------------------------------*/
1519 static void invalidate_complete(struct dm_cache_migration
*mg
, bool success
)
1521 struct bio_list bios
;
1522 struct cache
*cache
= mg
->cache
;
1524 bio_list_init(&bios
);
1525 if (dm_cell_unlock_v2(cache
->prison
, mg
->cell
, &bios
))
1526 free_prison_cell(cache
, mg
->cell
);
1528 if (!success
&& mg
->overwrite_bio
)
1529 bio_io_error(mg
->overwrite_bio
);
1532 defer_bios(cache
, &bios
);
1534 background_work_end(cache
);
1537 static void invalidate_completed(struct work_struct
*ws
)
1539 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1540 invalidate_complete(mg
, !mg
->k
.input
);
1543 static int invalidate_cblock(struct cache
*cache
, dm_cblock_t cblock
)
1545 int r
= policy_invalidate_mapping(cache
->policy
, cblock
);
1547 r
= dm_cache_remove_mapping(cache
->cmd
, cblock
);
1549 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1550 cache_device_name(cache
));
1551 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1554 } else if (r
== -ENODATA
) {
1556 * Harmless, already unmapped.
1561 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache
));
1566 static void invalidate_remove(struct work_struct
*ws
)
1569 struct dm_cache_migration
*mg
= ws_to_mg(ws
);
1570 struct cache
*cache
= mg
->cache
;
1572 r
= invalidate_cblock(cache
, mg
->invalidate_cblock
);
1574 invalidate_complete(mg
, false);
1578 init_continuation(&mg
->k
, invalidate_completed
);
1579 continue_after_commit(&cache
->committer
, &mg
->k
);
1580 remap_to_origin_clear_discard(cache
, mg
->overwrite_bio
, mg
->invalidate_oblock
);
1581 mg
->overwrite_bio
= NULL
;
1582 schedule_commit(&cache
->committer
);
1585 static int invalidate_lock(struct dm_cache_migration
*mg
)
1588 struct dm_cell_key_v2 key
;
1589 struct cache
*cache
= mg
->cache
;
1590 struct dm_bio_prison_cell_v2
*prealloc
;
1592 prealloc
= alloc_prison_cell(cache
);
1594 build_key(mg
->invalidate_oblock
, oblock_succ(mg
->invalidate_oblock
), &key
);
1595 r
= dm_cell_lock_v2(cache
->prison
, &key
,
1596 READ_WRITE_LOCK_LEVEL
, prealloc
, &mg
->cell
);
1598 free_prison_cell(cache
, prealloc
);
1599 invalidate_complete(mg
, false);
1603 if (mg
->cell
!= prealloc
)
1604 free_prison_cell(cache
, prealloc
);
1607 quiesce(mg
, invalidate_remove
);
1611 * We can't call invalidate_remove() directly here because we
1612 * might still be in request context.
1614 init_continuation(&mg
->k
, invalidate_remove
);
1615 queue_work(cache
->wq
, &mg
->k
.ws
);
1621 static int invalidate_start(struct cache
*cache
, dm_cblock_t cblock
,
1622 dm_oblock_t oblock
, struct bio
*bio
)
1624 struct dm_cache_migration
*mg
;
1626 if (!background_work_begin(cache
))
1629 mg
= alloc_migration(cache
);
1631 mg
->overwrite_bio
= bio
;
1632 mg
->invalidate_cblock
= cblock
;
1633 mg
->invalidate_oblock
= oblock
;
1635 return invalidate_lock(mg
);
1638 /*----------------------------------------------------------------
1640 *--------------------------------------------------------------*/
1647 static enum busy
spare_migration_bandwidth(struct cache
*cache
)
1649 bool idle
= iot_idle_for(&cache
->tracker
, HZ
);
1650 sector_t current_volume
= (atomic_read(&cache
->nr_io_migrations
) + 1) *
1651 cache
->sectors_per_block
;
1653 if (idle
&& current_volume
<= cache
->migration_threshold
)
1659 static void inc_hit_counter(struct cache
*cache
, struct bio
*bio
)
1661 atomic_inc(bio_data_dir(bio
) == READ
?
1662 &cache
->stats
.read_hit
: &cache
->stats
.write_hit
);
1665 static void inc_miss_counter(struct cache
*cache
, struct bio
*bio
)
1667 atomic_inc(bio_data_dir(bio
) == READ
?
1668 &cache
->stats
.read_miss
: &cache
->stats
.write_miss
);
1671 /*----------------------------------------------------------------*/
1673 static int map_bio(struct cache
*cache
, struct bio
*bio
, dm_oblock_t block
,
1674 bool *commit_needed
)
1677 bool rb
, background_queued
;
1680 *commit_needed
= false;
1682 rb
= bio_detain_shared(cache
, block
, bio
);
1685 * An exclusive lock is held for this block, so we have to
1686 * wait. We set the commit_needed flag so the current
1687 * transaction will be committed asap, allowing this lock
1690 *commit_needed
= true;
1691 return DM_MAPIO_SUBMITTED
;
1694 data_dir
= bio_data_dir(bio
);
1696 if (optimisable_bio(cache
, bio
, block
)) {
1697 struct policy_work
*op
= NULL
;
1699 r
= policy_lookup_with_work(cache
->policy
, block
, &cblock
, data_dir
, true, &op
);
1700 if (unlikely(r
&& r
!= -ENOENT
)) {
1701 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1702 cache_device_name(cache
), r
);
1704 return DM_MAPIO_SUBMITTED
;
1707 if (r
== -ENOENT
&& op
) {
1708 bio_drop_shared_lock(cache
, bio
);
1709 BUG_ON(op
->op
!= POLICY_PROMOTE
);
1710 mg_start(cache
, op
, bio
);
1711 return DM_MAPIO_SUBMITTED
;
1714 r
= policy_lookup(cache
->policy
, block
, &cblock
, data_dir
, false, &background_queued
);
1715 if (unlikely(r
&& r
!= -ENOENT
)) {
1716 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1717 cache_device_name(cache
), r
);
1719 return DM_MAPIO_SUBMITTED
;
1722 if (background_queued
)
1723 wake_migration_worker(cache
);
1727 struct per_bio_data
*pb
= get_per_bio_data(bio
);
1732 inc_miss_counter(cache
, bio
);
1733 if (pb
->req_nr
== 0) {
1734 accounted_begin(cache
, bio
);
1735 remap_to_origin_clear_discard(cache
, bio
, block
);
1738 * This is a duplicate writethrough io that is no
1739 * longer needed because the block has been demoted.
1742 return DM_MAPIO_SUBMITTED
;
1748 inc_hit_counter(cache
, bio
);
1751 * Passthrough always maps to the origin, invalidating any
1752 * cache blocks that are written to.
1754 if (passthrough_mode(cache
)) {
1755 if (bio_data_dir(bio
) == WRITE
) {
1756 bio_drop_shared_lock(cache
, bio
);
1757 atomic_inc(&cache
->stats
.demotion
);
1758 invalidate_start(cache
, cblock
, block
, bio
);
1760 remap_to_origin_clear_discard(cache
, bio
, block
);
1762 if (bio_data_dir(bio
) == WRITE
&& writethrough_mode(cache
) &&
1763 !is_dirty(cache
, cblock
)) {
1764 remap_to_origin_and_cache(cache
, bio
, block
, cblock
);
1765 accounted_begin(cache
, bio
);
1767 remap_to_cache_dirty(cache
, bio
, block
, cblock
);
1772 * dm core turns FUA requests into a separate payload and FLUSH req.
1774 if (bio
->bi_opf
& REQ_FUA
) {
1776 * issue_after_commit will call accounted_begin a second time. So
1777 * we call accounted_complete() to avoid double accounting.
1779 accounted_complete(cache
, bio
);
1780 issue_after_commit(&cache
->committer
, bio
);
1781 *commit_needed
= true;
1782 return DM_MAPIO_SUBMITTED
;
1785 return DM_MAPIO_REMAPPED
;
1788 static bool process_bio(struct cache
*cache
, struct bio
*bio
)
1792 if (map_bio(cache
, bio
, get_bio_block(cache
, bio
), &commit_needed
) == DM_MAPIO_REMAPPED
)
1793 submit_bio_noacct(bio
);
1795 return commit_needed
;
1799 * A non-zero return indicates read_only or fail_io mode.
1801 static int commit(struct cache
*cache
, bool clean_shutdown
)
1805 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1808 atomic_inc(&cache
->stats
.commit_count
);
1809 r
= dm_cache_commit(cache
->cmd
, clean_shutdown
);
1811 metadata_operation_failed(cache
, "dm_cache_commit", r
);
1817 * Used by the batcher.
1819 static blk_status_t
commit_op(void *context
)
1821 struct cache
*cache
= context
;
1823 if (dm_cache_changed_this_transaction(cache
->cmd
))
1824 return errno_to_blk_status(commit(cache
, false));
1829 /*----------------------------------------------------------------*/
1831 static bool process_flush_bio(struct cache
*cache
, struct bio
*bio
)
1833 struct per_bio_data
*pb
= get_per_bio_data(bio
);
1836 remap_to_origin(cache
, bio
);
1838 remap_to_cache(cache
, bio
, 0);
1840 issue_after_commit(&cache
->committer
, bio
);
1844 static bool process_discard_bio(struct cache
*cache
, struct bio
*bio
)
1848 // FIXME: do we need to lock the region? Or can we just assume the
1849 // user wont be so foolish as to issue discard concurrently with
1851 calc_discard_block_range(cache
, bio
, &b
, &e
);
1853 set_discard(cache
, b
);
1854 b
= to_dblock(from_dblock(b
) + 1);
1857 if (cache
->features
.discard_passdown
) {
1858 remap_to_origin(cache
, bio
);
1859 submit_bio_noacct(bio
);
1866 static void process_deferred_bios(struct work_struct
*ws
)
1868 struct cache
*cache
= container_of(ws
, struct cache
, deferred_bio_worker
);
1870 bool commit_needed
= false;
1871 struct bio_list bios
;
1874 bio_list_init(&bios
);
1876 spin_lock_irq(&cache
->lock
);
1877 bio_list_merge(&bios
, &cache
->deferred_bios
);
1878 bio_list_init(&cache
->deferred_bios
);
1879 spin_unlock_irq(&cache
->lock
);
1881 while ((bio
= bio_list_pop(&bios
))) {
1882 if (bio
->bi_opf
& REQ_PREFLUSH
)
1883 commit_needed
= process_flush_bio(cache
, bio
) || commit_needed
;
1885 else if (bio_op(bio
) == REQ_OP_DISCARD
)
1886 commit_needed
= process_discard_bio(cache
, bio
) || commit_needed
;
1889 commit_needed
= process_bio(cache
, bio
) || commit_needed
;
1893 schedule_commit(&cache
->committer
);
1896 /*----------------------------------------------------------------
1898 *--------------------------------------------------------------*/
1900 static void requeue_deferred_bios(struct cache
*cache
)
1903 struct bio_list bios
;
1905 bio_list_init(&bios
);
1906 bio_list_merge(&bios
, &cache
->deferred_bios
);
1907 bio_list_init(&cache
->deferred_bios
);
1909 while ((bio
= bio_list_pop(&bios
))) {
1910 bio
->bi_status
= BLK_STS_DM_REQUEUE
;
1916 * We want to commit periodically so that not too much
1917 * unwritten metadata builds up.
1919 static void do_waker(struct work_struct
*ws
)
1921 struct cache
*cache
= container_of(to_delayed_work(ws
), struct cache
, waker
);
1923 policy_tick(cache
->policy
, true);
1924 wake_migration_worker(cache
);
1925 schedule_commit(&cache
->committer
);
1926 queue_delayed_work(cache
->wq
, &cache
->waker
, COMMIT_PERIOD
);
1929 static void check_migrations(struct work_struct
*ws
)
1932 struct policy_work
*op
;
1933 struct cache
*cache
= container_of(ws
, struct cache
, migration_worker
);
1937 b
= spare_migration_bandwidth(cache
);
1939 r
= policy_get_background_work(cache
->policy
, b
== IDLE
, &op
);
1944 DMERR_LIMIT("%s: policy_background_work failed",
1945 cache_device_name(cache
));
1949 r
= mg_start(cache
, op
, NULL
);
1955 /*----------------------------------------------------------------
1957 *--------------------------------------------------------------*/
1960 * This function gets called on the error paths of the constructor, so we
1961 * have to cope with a partially initialised struct.
1963 static void destroy(struct cache
*cache
)
1967 mempool_exit(&cache
->migration_pool
);
1970 dm_bio_prison_destroy_v2(cache
->prison
);
1973 destroy_workqueue(cache
->wq
);
1975 if (cache
->dirty_bitset
)
1976 free_bitset(cache
->dirty_bitset
);
1978 if (cache
->discard_bitset
)
1979 free_bitset(cache
->discard_bitset
);
1982 dm_kcopyd_client_destroy(cache
->copier
);
1985 dm_cache_metadata_close(cache
->cmd
);
1987 if (cache
->metadata_dev
)
1988 dm_put_device(cache
->ti
, cache
->metadata_dev
);
1990 if (cache
->origin_dev
)
1991 dm_put_device(cache
->ti
, cache
->origin_dev
);
1993 if (cache
->cache_dev
)
1994 dm_put_device(cache
->ti
, cache
->cache_dev
);
1997 dm_cache_policy_destroy(cache
->policy
);
1999 for (i
= 0; i
< cache
->nr_ctr_args
; i
++)
2000 kfree(cache
->ctr_args
[i
]);
2001 kfree(cache
->ctr_args
);
2003 bioset_exit(&cache
->bs
);
2008 static void cache_dtr(struct dm_target
*ti
)
2010 struct cache
*cache
= ti
->private;
2015 static sector_t
get_dev_size(struct dm_dev
*dev
)
2017 return i_size_read(dev
->bdev
->bd_inode
) >> SECTOR_SHIFT
;
2020 /*----------------------------------------------------------------*/
2023 * Construct a cache device mapping.
2025 * cache <metadata dev> <cache dev> <origin dev> <block size>
2026 * <#feature args> [<feature arg>]*
2027 * <policy> <#policy args> [<policy arg>]*
2029 * metadata dev : fast device holding the persistent metadata
2030 * cache dev : fast device holding cached data blocks
2031 * origin dev : slow device holding original data blocks
2032 * block size : cache unit size in sectors
2034 * #feature args : number of feature arguments passed
2035 * feature args : writethrough. (The default is writeback.)
2037 * policy : the replacement policy to use
2038 * #policy args : an even number of policy arguments corresponding
2039 * to key/value pairs passed to the policy
2040 * policy args : key/value pairs passed to the policy
2041 * E.g. 'sequential_threshold 1024'
2042 * See cache-policies.txt for details.
2044 * Optional feature arguments are:
2045 * writethrough : write through caching that prohibits cache block
2046 * content from being different from origin block content.
2047 * Without this argument, the default behaviour is to write
2048 * back cache block contents later for performance reasons,
2049 * so they may differ from the corresponding origin blocks.
2052 struct dm_target
*ti
;
2054 struct dm_dev
*metadata_dev
;
2056 struct dm_dev
*cache_dev
;
2057 sector_t cache_sectors
;
2059 struct dm_dev
*origin_dev
;
2060 sector_t origin_sectors
;
2062 uint32_t block_size
;
2064 const char *policy_name
;
2066 const char **policy_argv
;
2068 struct cache_features features
;
2071 static void destroy_cache_args(struct cache_args
*ca
)
2073 if (ca
->metadata_dev
)
2074 dm_put_device(ca
->ti
, ca
->metadata_dev
);
2077 dm_put_device(ca
->ti
, ca
->cache_dev
);
2080 dm_put_device(ca
->ti
, ca
->origin_dev
);
2085 static bool at_least_one_arg(struct dm_arg_set
*as
, char **error
)
2088 *error
= "Insufficient args";
2095 static int parse_metadata_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2099 sector_t metadata_dev_size
;
2100 char b
[BDEVNAME_SIZE
];
2102 if (!at_least_one_arg(as
, error
))
2105 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2108 *error
= "Error opening metadata device";
2112 metadata_dev_size
= get_dev_size(ca
->metadata_dev
);
2113 if (metadata_dev_size
> DM_CACHE_METADATA_MAX_SECTORS_WARNING
)
2114 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2115 bdevname(ca
->metadata_dev
->bdev
, b
), THIN_METADATA_MAX_SECTORS
);
2120 static int parse_cache_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2125 if (!at_least_one_arg(as
, error
))
2128 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2131 *error
= "Error opening cache device";
2134 ca
->cache_sectors
= get_dev_size(ca
->cache_dev
);
2139 static int parse_origin_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2144 if (!at_least_one_arg(as
, error
))
2147 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2150 *error
= "Error opening origin device";
2154 ca
->origin_sectors
= get_dev_size(ca
->origin_dev
);
2155 if (ca
->ti
->len
> ca
->origin_sectors
) {
2156 *error
= "Device size larger than cached device";
2163 static int parse_block_size(struct cache_args
*ca
, struct dm_arg_set
*as
,
2166 unsigned long block_size
;
2168 if (!at_least_one_arg(as
, error
))
2171 if (kstrtoul(dm_shift_arg(as
), 10, &block_size
) || !block_size
||
2172 block_size
< DATA_DEV_BLOCK_SIZE_MIN_SECTORS
||
2173 block_size
> DATA_DEV_BLOCK_SIZE_MAX_SECTORS
||
2174 block_size
& (DATA_DEV_BLOCK_SIZE_MIN_SECTORS
- 1)) {
2175 *error
= "Invalid data block size";
2179 if (block_size
> ca
->cache_sectors
) {
2180 *error
= "Data block size is larger than the cache device";
2184 ca
->block_size
= block_size
;
2189 static void init_features(struct cache_features
*cf
)
2191 cf
->mode
= CM_WRITE
;
2192 cf
->io_mode
= CM_IO_WRITEBACK
;
2193 cf
->metadata_version
= 1;
2194 cf
->discard_passdown
= true;
2197 static int parse_features(struct cache_args
*ca
, struct dm_arg_set
*as
,
2200 static const struct dm_arg _args
[] = {
2201 {0, 3, "Invalid number of cache feature arguments"},
2204 int r
, mode_ctr
= 0;
2207 struct cache_features
*cf
= &ca
->features
;
2211 r
= dm_read_arg_group(_args
, as
, &argc
, error
);
2216 arg
= dm_shift_arg(as
);
2218 if (!strcasecmp(arg
, "writeback")) {
2219 cf
->io_mode
= CM_IO_WRITEBACK
;
2223 else if (!strcasecmp(arg
, "writethrough")) {
2224 cf
->io_mode
= CM_IO_WRITETHROUGH
;
2228 else if (!strcasecmp(arg
, "passthrough")) {
2229 cf
->io_mode
= CM_IO_PASSTHROUGH
;
2233 else if (!strcasecmp(arg
, "metadata2"))
2234 cf
->metadata_version
= 2;
2236 else if (!strcasecmp(arg
, "no_discard_passdown"))
2237 cf
->discard_passdown
= false;
2240 *error
= "Unrecognised cache feature requested";
2246 *error
= "Duplicate cache io_mode features requested";
2253 static int parse_policy(struct cache_args
*ca
, struct dm_arg_set
*as
,
2256 static const struct dm_arg _args
[] = {
2257 {0, 1024, "Invalid number of policy arguments"},
2262 if (!at_least_one_arg(as
, error
))
2265 ca
->policy_name
= dm_shift_arg(as
);
2267 r
= dm_read_arg_group(_args
, as
, &ca
->policy_argc
, error
);
2271 ca
->policy_argv
= (const char **)as
->argv
;
2272 dm_consume_args(as
, ca
->policy_argc
);
2277 static int parse_cache_args(struct cache_args
*ca
, int argc
, char **argv
,
2281 struct dm_arg_set as
;
2286 r
= parse_metadata_dev(ca
, &as
, error
);
2290 r
= parse_cache_dev(ca
, &as
, error
);
2294 r
= parse_origin_dev(ca
, &as
, error
);
2298 r
= parse_block_size(ca
, &as
, error
);
2302 r
= parse_features(ca
, &as
, error
);
2306 r
= parse_policy(ca
, &as
, error
);
2313 /*----------------------------------------------------------------*/
2315 static struct kmem_cache
*migration_cache
;
2317 #define NOT_CORE_OPTION 1
2319 static int process_config_option(struct cache
*cache
, const char *key
, const char *value
)
2323 if (!strcasecmp(key
, "migration_threshold")) {
2324 if (kstrtoul(value
, 10, &tmp
))
2327 cache
->migration_threshold
= tmp
;
2331 return NOT_CORE_OPTION
;
2334 static int set_config_value(struct cache
*cache
, const char *key
, const char *value
)
2336 int r
= process_config_option(cache
, key
, value
);
2338 if (r
== NOT_CORE_OPTION
)
2339 r
= policy_set_config_value(cache
->policy
, key
, value
);
2342 DMWARN("bad config value for %s: %s", key
, value
);
2347 static int set_config_values(struct cache
*cache
, int argc
, const char **argv
)
2352 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2357 r
= set_config_value(cache
, argv
[0], argv
[1]);
2368 static int create_cache_policy(struct cache
*cache
, struct cache_args
*ca
,
2371 struct dm_cache_policy
*p
= dm_cache_policy_create(ca
->policy_name
,
2373 cache
->origin_sectors
,
2374 cache
->sectors_per_block
);
2376 *error
= "Error creating cache's policy";
2380 BUG_ON(!cache
->policy
);
2386 * We want the discard block size to be at least the size of the cache
2387 * block size and have no more than 2^14 discard blocks across the origin.
2389 #define MAX_DISCARD_BLOCKS (1 << 14)
2391 static bool too_many_discard_blocks(sector_t discard_block_size
,
2392 sector_t origin_size
)
2394 (void) sector_div(origin_size
, discard_block_size
);
2396 return origin_size
> MAX_DISCARD_BLOCKS
;
2399 static sector_t
calculate_discard_block_size(sector_t cache_block_size
,
2400 sector_t origin_size
)
2402 sector_t discard_block_size
= cache_block_size
;
2405 while (too_many_discard_blocks(discard_block_size
, origin_size
))
2406 discard_block_size
*= 2;
2408 return discard_block_size
;
2411 static void set_cache_size(struct cache
*cache
, dm_cblock_t size
)
2413 dm_block_t nr_blocks
= from_cblock(size
);
2415 if (nr_blocks
> (1 << 20) && cache
->cache_size
!= size
)
2416 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2417 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2418 "Please consider increasing the cache block size to reduce the overall cache block count.",
2419 (unsigned long long) nr_blocks
);
2421 cache
->cache_size
= size
;
2424 #define DEFAULT_MIGRATION_THRESHOLD 2048
2426 static int cache_create(struct cache_args
*ca
, struct cache
**result
)
2429 char **error
= &ca
->ti
->error
;
2430 struct cache
*cache
;
2431 struct dm_target
*ti
= ca
->ti
;
2432 dm_block_t origin_blocks
;
2433 struct dm_cache_metadata
*cmd
;
2434 bool may_format
= ca
->features
.mode
== CM_WRITE
;
2436 cache
= kzalloc(sizeof(*cache
), GFP_KERNEL
);
2441 ti
->private = cache
;
2442 ti
->num_flush_bios
= 2;
2443 ti
->flush_supported
= true;
2445 ti
->num_discard_bios
= 1;
2446 ti
->discards_supported
= true;
2448 ti
->per_io_data_size
= sizeof(struct per_bio_data
);
2450 cache
->features
= ca
->features
;
2451 if (writethrough_mode(cache
)) {
2452 /* Create bioset for writethrough bios issued to origin */
2453 r
= bioset_init(&cache
->bs
, BIO_POOL_SIZE
, 0, 0);
2458 cache
->metadata_dev
= ca
->metadata_dev
;
2459 cache
->origin_dev
= ca
->origin_dev
;
2460 cache
->cache_dev
= ca
->cache_dev
;
2462 ca
->metadata_dev
= ca
->origin_dev
= ca
->cache_dev
= NULL
;
2464 origin_blocks
= cache
->origin_sectors
= ca
->origin_sectors
;
2465 origin_blocks
= block_div(origin_blocks
, ca
->block_size
);
2466 cache
->origin_blocks
= to_oblock(origin_blocks
);
2468 cache
->sectors_per_block
= ca
->block_size
;
2469 if (dm_set_target_max_io_len(ti
, cache
->sectors_per_block
)) {
2474 if (ca
->block_size
& (ca
->block_size
- 1)) {
2475 dm_block_t cache_size
= ca
->cache_sectors
;
2477 cache
->sectors_per_block_shift
= -1;
2478 cache_size
= block_div(cache_size
, ca
->block_size
);
2479 set_cache_size(cache
, to_cblock(cache_size
));
2481 cache
->sectors_per_block_shift
= __ffs(ca
->block_size
);
2482 set_cache_size(cache
, to_cblock(ca
->cache_sectors
>> cache
->sectors_per_block_shift
));
2485 r
= create_cache_policy(cache
, ca
, error
);
2489 cache
->policy_nr_args
= ca
->policy_argc
;
2490 cache
->migration_threshold
= DEFAULT_MIGRATION_THRESHOLD
;
2492 r
= set_config_values(cache
, ca
->policy_argc
, ca
->policy_argv
);
2494 *error
= "Error setting cache policy's config values";
2498 cmd
= dm_cache_metadata_open(cache
->metadata_dev
->bdev
,
2499 ca
->block_size
, may_format
,
2500 dm_cache_policy_get_hint_size(cache
->policy
),
2501 ca
->features
.metadata_version
);
2503 *error
= "Error creating metadata object";
2508 set_cache_mode(cache
, CM_WRITE
);
2509 if (get_cache_mode(cache
) != CM_WRITE
) {
2510 *error
= "Unable to get write access to metadata, please check/repair metadata.";
2515 if (passthrough_mode(cache
)) {
2518 r
= dm_cache_metadata_all_clean(cache
->cmd
, &all_clean
);
2520 *error
= "dm_cache_metadata_all_clean() failed";
2525 *error
= "Cannot enter passthrough mode unless all blocks are clean";
2530 policy_allow_migrations(cache
->policy
, false);
2533 spin_lock_init(&cache
->lock
);
2534 bio_list_init(&cache
->deferred_bios
);
2535 atomic_set(&cache
->nr_allocated_migrations
, 0);
2536 atomic_set(&cache
->nr_io_migrations
, 0);
2537 init_waitqueue_head(&cache
->migration_wait
);
2540 atomic_set(&cache
->nr_dirty
, 0);
2541 cache
->dirty_bitset
= alloc_bitset(from_cblock(cache
->cache_size
));
2542 if (!cache
->dirty_bitset
) {
2543 *error
= "could not allocate dirty bitset";
2546 clear_bitset(cache
->dirty_bitset
, from_cblock(cache
->cache_size
));
2548 cache
->discard_block_size
=
2549 calculate_discard_block_size(cache
->sectors_per_block
,
2550 cache
->origin_sectors
);
2551 cache
->discard_nr_blocks
= to_dblock(dm_sector_div_up(cache
->origin_sectors
,
2552 cache
->discard_block_size
));
2553 cache
->discard_bitset
= alloc_bitset(from_dblock(cache
->discard_nr_blocks
));
2554 if (!cache
->discard_bitset
) {
2555 *error
= "could not allocate discard bitset";
2558 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
2560 cache
->copier
= dm_kcopyd_client_create(&dm_kcopyd_throttle
);
2561 if (IS_ERR(cache
->copier
)) {
2562 *error
= "could not create kcopyd client";
2563 r
= PTR_ERR(cache
->copier
);
2567 cache
->wq
= alloc_workqueue("dm-" DM_MSG_PREFIX
, WQ_MEM_RECLAIM
, 0);
2569 *error
= "could not create workqueue for metadata object";
2572 INIT_WORK(&cache
->deferred_bio_worker
, process_deferred_bios
);
2573 INIT_WORK(&cache
->migration_worker
, check_migrations
);
2574 INIT_DELAYED_WORK(&cache
->waker
, do_waker
);
2576 cache
->prison
= dm_bio_prison_create_v2(cache
->wq
);
2577 if (!cache
->prison
) {
2578 *error
= "could not create bio prison";
2582 r
= mempool_init_slab_pool(&cache
->migration_pool
, MIGRATION_POOL_SIZE
,
2585 *error
= "Error creating cache's migration mempool";
2589 cache
->need_tick_bio
= true;
2590 cache
->sized
= false;
2591 cache
->invalidate
= false;
2592 cache
->commit_requested
= false;
2593 cache
->loaded_mappings
= false;
2594 cache
->loaded_discards
= false;
2598 atomic_set(&cache
->stats
.demotion
, 0);
2599 atomic_set(&cache
->stats
.promotion
, 0);
2600 atomic_set(&cache
->stats
.copies_avoided
, 0);
2601 atomic_set(&cache
->stats
.cache_cell_clash
, 0);
2602 atomic_set(&cache
->stats
.commit_count
, 0);
2603 atomic_set(&cache
->stats
.discard_count
, 0);
2605 spin_lock_init(&cache
->invalidation_lock
);
2606 INIT_LIST_HEAD(&cache
->invalidation_requests
);
2608 batcher_init(&cache
->committer
, commit_op
, cache
,
2609 issue_op
, cache
, cache
->wq
);
2610 iot_init(&cache
->tracker
);
2612 init_rwsem(&cache
->background_work_lock
);
2613 prevent_background_work(cache
);
2622 static int copy_ctr_args(struct cache
*cache
, int argc
, const char **argv
)
2627 copy
= kcalloc(argc
, sizeof(*copy
), GFP_KERNEL
);
2630 for (i
= 0; i
< argc
; i
++) {
2631 copy
[i
] = kstrdup(argv
[i
], GFP_KERNEL
);
2640 cache
->nr_ctr_args
= argc
;
2641 cache
->ctr_args
= copy
;
2646 static int cache_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
2649 struct cache_args
*ca
;
2650 struct cache
*cache
= NULL
;
2652 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
2654 ti
->error
= "Error allocating memory for cache";
2659 r
= parse_cache_args(ca
, argc
, argv
, &ti
->error
);
2663 r
= cache_create(ca
, &cache
);
2667 r
= copy_ctr_args(cache
, argc
- 3, (const char **)argv
+ 3);
2673 ti
->private = cache
;
2675 destroy_cache_args(ca
);
2679 /*----------------------------------------------------------------*/
2681 static int cache_map(struct dm_target
*ti
, struct bio
*bio
)
2683 struct cache
*cache
= ti
->private;
2687 dm_oblock_t block
= get_bio_block(cache
, bio
);
2689 init_per_bio_data(bio
);
2690 if (unlikely(from_oblock(block
) >= from_oblock(cache
->origin_blocks
))) {
2692 * This can only occur if the io goes to a partial block at
2693 * the end of the origin device. We don't cache these.
2694 * Just remap to the origin and carry on.
2696 remap_to_origin(cache
, bio
);
2697 accounted_begin(cache
, bio
);
2698 return DM_MAPIO_REMAPPED
;
2701 if (discard_or_flush(bio
)) {
2702 defer_bio(cache
, bio
);
2703 return DM_MAPIO_SUBMITTED
;
2706 r
= map_bio(cache
, bio
, block
, &commit_needed
);
2708 schedule_commit(&cache
->committer
);
2713 static int cache_end_io(struct dm_target
*ti
, struct bio
*bio
, blk_status_t
*error
)
2715 struct cache
*cache
= ti
->private;
2716 unsigned long flags
;
2717 struct per_bio_data
*pb
= get_per_bio_data(bio
);
2720 policy_tick(cache
->policy
, false);
2722 spin_lock_irqsave(&cache
->lock
, flags
);
2723 cache
->need_tick_bio
= true;
2724 spin_unlock_irqrestore(&cache
->lock
, flags
);
2727 bio_drop_shared_lock(cache
, bio
);
2728 accounted_complete(cache
, bio
);
2730 return DM_ENDIO_DONE
;
2733 static int write_dirty_bitset(struct cache
*cache
)
2737 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2740 r
= dm_cache_set_dirty_bits(cache
->cmd
, from_cblock(cache
->cache_size
), cache
->dirty_bitset
);
2742 metadata_operation_failed(cache
, "dm_cache_set_dirty_bits", r
);
2747 static int write_discard_bitset(struct cache
*cache
)
2751 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2754 r
= dm_cache_discard_bitset_resize(cache
->cmd
, cache
->discard_block_size
,
2755 cache
->discard_nr_blocks
);
2757 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache
));
2758 metadata_operation_failed(cache
, "dm_cache_discard_bitset_resize", r
);
2762 for (i
= 0; i
< from_dblock(cache
->discard_nr_blocks
); i
++) {
2763 r
= dm_cache_set_discard(cache
->cmd
, to_dblock(i
),
2764 is_discarded(cache
, to_dblock(i
)));
2766 metadata_operation_failed(cache
, "dm_cache_set_discard", r
);
2774 static int write_hints(struct cache
*cache
)
2778 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
2781 r
= dm_cache_write_hints(cache
->cmd
, cache
->policy
);
2783 metadata_operation_failed(cache
, "dm_cache_write_hints", r
);
2791 * returns true on success
2793 static bool sync_metadata(struct cache
*cache
)
2797 r1
= write_dirty_bitset(cache
);
2799 DMERR("%s: could not write dirty bitset", cache_device_name(cache
));
2801 r2
= write_discard_bitset(cache
);
2803 DMERR("%s: could not write discard bitset", cache_device_name(cache
));
2807 r3
= write_hints(cache
);
2809 DMERR("%s: could not write hints", cache_device_name(cache
));
2812 * If writing the above metadata failed, we still commit, but don't
2813 * set the clean shutdown flag. This will effectively force every
2814 * dirty bit to be set on reload.
2816 r4
= commit(cache
, !r1
&& !r2
&& !r3
);
2818 DMERR("%s: could not write cache metadata", cache_device_name(cache
));
2820 return !r1
&& !r2
&& !r3
&& !r4
;
2823 static void cache_postsuspend(struct dm_target
*ti
)
2825 struct cache
*cache
= ti
->private;
2827 prevent_background_work(cache
);
2828 BUG_ON(atomic_read(&cache
->nr_io_migrations
));
2830 cancel_delayed_work_sync(&cache
->waker
);
2831 drain_workqueue(cache
->wq
);
2832 WARN_ON(cache
->tracker
.in_flight
);
2835 * If it's a flush suspend there won't be any deferred bios, so this
2838 requeue_deferred_bios(cache
);
2840 if (get_cache_mode(cache
) == CM_WRITE
)
2841 (void) sync_metadata(cache
);
2844 static int load_mapping(void *context
, dm_oblock_t oblock
, dm_cblock_t cblock
,
2845 bool dirty
, uint32_t hint
, bool hint_valid
)
2848 struct cache
*cache
= context
;
2851 set_bit(from_cblock(cblock
), cache
->dirty_bitset
);
2852 atomic_inc(&cache
->nr_dirty
);
2854 clear_bit(from_cblock(cblock
), cache
->dirty_bitset
);
2856 r
= policy_load_mapping(cache
->policy
, oblock
, cblock
, dirty
, hint
, hint_valid
);
2864 * The discard block size in the on disk metadata is not
2865 * neccessarily the same as we're currently using. So we have to
2866 * be careful to only set the discarded attribute if we know it
2867 * covers a complete block of the new size.
2869 struct discard_load_info
{
2870 struct cache
*cache
;
2873 * These blocks are sized using the on disk dblock size, rather
2874 * than the current one.
2876 dm_block_t block_size
;
2877 dm_block_t discard_begin
, discard_end
;
2880 static void discard_load_info_init(struct cache
*cache
,
2881 struct discard_load_info
*li
)
2884 li
->discard_begin
= li
->discard_end
= 0;
2887 static void set_discard_range(struct discard_load_info
*li
)
2891 if (li
->discard_begin
== li
->discard_end
)
2895 * Convert to sectors.
2897 b
= li
->discard_begin
* li
->block_size
;
2898 e
= li
->discard_end
* li
->block_size
;
2901 * Then convert back to the current dblock size.
2903 b
= dm_sector_div_up(b
, li
->cache
->discard_block_size
);
2904 sector_div(e
, li
->cache
->discard_block_size
);
2907 * The origin may have shrunk, so we need to check we're still in
2910 if (e
> from_dblock(li
->cache
->discard_nr_blocks
))
2911 e
= from_dblock(li
->cache
->discard_nr_blocks
);
2914 set_discard(li
->cache
, to_dblock(b
));
2917 static int load_discard(void *context
, sector_t discard_block_size
,
2918 dm_dblock_t dblock
, bool discard
)
2920 struct discard_load_info
*li
= context
;
2922 li
->block_size
= discard_block_size
;
2925 if (from_dblock(dblock
) == li
->discard_end
)
2927 * We're already in a discard range, just extend it.
2929 li
->discard_end
= li
->discard_end
+ 1ULL;
2933 * Emit the old range and start a new one.
2935 set_discard_range(li
);
2936 li
->discard_begin
= from_dblock(dblock
);
2937 li
->discard_end
= li
->discard_begin
+ 1ULL;
2940 set_discard_range(li
);
2941 li
->discard_begin
= li
->discard_end
= 0;
2947 static dm_cblock_t
get_cache_dev_size(struct cache
*cache
)
2949 sector_t size
= get_dev_size(cache
->cache_dev
);
2950 (void) sector_div(size
, cache
->sectors_per_block
);
2951 return to_cblock(size
);
2954 static bool can_resize(struct cache
*cache
, dm_cblock_t new_size
)
2956 if (from_cblock(new_size
) > from_cblock(cache
->cache_size
)) {
2958 DMERR("%s: unable to extend cache due to missing cache table reload",
2959 cache_device_name(cache
));
2965 * We can't drop a dirty block when shrinking the cache.
2967 while (from_cblock(new_size
) < from_cblock(cache
->cache_size
)) {
2968 new_size
= to_cblock(from_cblock(new_size
) + 1);
2969 if (is_dirty(cache
, new_size
)) {
2970 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2971 cache_device_name(cache
),
2972 (unsigned long long) from_cblock(new_size
));
2980 static int resize_cache_dev(struct cache
*cache
, dm_cblock_t new_size
)
2984 r
= dm_cache_resize(cache
->cmd
, new_size
);
2986 DMERR("%s: could not resize cache metadata", cache_device_name(cache
));
2987 metadata_operation_failed(cache
, "dm_cache_resize", r
);
2991 set_cache_size(cache
, new_size
);
2996 static int cache_preresume(struct dm_target
*ti
)
2999 struct cache
*cache
= ti
->private;
3000 dm_cblock_t csize
= get_cache_dev_size(cache
);
3003 * Check to see if the cache has resized.
3005 if (!cache
->sized
) {
3006 r
= resize_cache_dev(cache
, csize
);
3010 cache
->sized
= true;
3012 } else if (csize
!= cache
->cache_size
) {
3013 if (!can_resize(cache
, csize
))
3016 r
= resize_cache_dev(cache
, csize
);
3021 if (!cache
->loaded_mappings
) {
3022 r
= dm_cache_load_mappings(cache
->cmd
, cache
->policy
,
3023 load_mapping
, cache
);
3025 DMERR("%s: could not load cache mappings", cache_device_name(cache
));
3026 metadata_operation_failed(cache
, "dm_cache_load_mappings", r
);
3030 cache
->loaded_mappings
= true;
3033 if (!cache
->loaded_discards
) {
3034 struct discard_load_info li
;
3037 * The discard bitset could have been resized, or the
3038 * discard block size changed. To be safe we start by
3039 * setting every dblock to not discarded.
3041 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
3043 discard_load_info_init(cache
, &li
);
3044 r
= dm_cache_load_discards(cache
->cmd
, load_discard
, &li
);
3046 DMERR("%s: could not load origin discards", cache_device_name(cache
));
3047 metadata_operation_failed(cache
, "dm_cache_load_discards", r
);
3050 set_discard_range(&li
);
3052 cache
->loaded_discards
= true;
3058 static void cache_resume(struct dm_target
*ti
)
3060 struct cache
*cache
= ti
->private;
3062 cache
->need_tick_bio
= true;
3063 allow_background_work(cache
);
3064 do_waker(&cache
->waker
.work
);
3067 static void emit_flags(struct cache
*cache
, char *result
,
3068 unsigned maxlen
, ssize_t
*sz_ptr
)
3070 ssize_t sz
= *sz_ptr
;
3071 struct cache_features
*cf
= &cache
->features
;
3072 unsigned count
= (cf
->metadata_version
== 2) + !cf
->discard_passdown
+ 1;
3074 DMEMIT("%u ", count
);
3076 if (cf
->metadata_version
== 2)
3077 DMEMIT("metadata2 ");
3079 if (writethrough_mode(cache
))
3080 DMEMIT("writethrough ");
3082 else if (passthrough_mode(cache
))
3083 DMEMIT("passthrough ");
3085 else if (writeback_mode(cache
))
3086 DMEMIT("writeback ");
3090 DMERR("%s: internal error: unknown io mode: %d",
3091 cache_device_name(cache
), (int) cf
->io_mode
);
3094 if (!cf
->discard_passdown
)
3095 DMEMIT("no_discard_passdown ");
3103 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3104 * <cache block size> <#used cache blocks>/<#total cache blocks>
3105 * <#read hits> <#read misses> <#write hits> <#write misses>
3106 * <#demotions> <#promotions> <#dirty>
3107 * <#features> <features>*
3108 * <#core args> <core args>
3109 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3111 static void cache_status(struct dm_target
*ti
, status_type_t type
,
3112 unsigned status_flags
, char *result
, unsigned maxlen
)
3117 dm_block_t nr_free_blocks_metadata
= 0;
3118 dm_block_t nr_blocks_metadata
= 0;
3119 char buf
[BDEVNAME_SIZE
];
3120 struct cache
*cache
= ti
->private;
3121 dm_cblock_t residency
;
3125 case STATUSTYPE_INFO
:
3126 if (get_cache_mode(cache
) == CM_FAIL
) {
3131 /* Commit to ensure statistics aren't out-of-date */
3132 if (!(status_flags
& DM_STATUS_NOFLUSH_FLAG
) && !dm_suspended(ti
))
3133 (void) commit(cache
, false);
3135 r
= dm_cache_get_free_metadata_block_count(cache
->cmd
, &nr_free_blocks_metadata
);
3137 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3138 cache_device_name(cache
), r
);
3142 r
= dm_cache_get_metadata_dev_size(cache
->cmd
, &nr_blocks_metadata
);
3144 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3145 cache_device_name(cache
), r
);
3149 residency
= policy_residency(cache
->policy
);
3151 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3152 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE
,
3153 (unsigned long long)(nr_blocks_metadata
- nr_free_blocks_metadata
),
3154 (unsigned long long)nr_blocks_metadata
,
3155 (unsigned long long)cache
->sectors_per_block
,
3156 (unsigned long long) from_cblock(residency
),
3157 (unsigned long long) from_cblock(cache
->cache_size
),
3158 (unsigned) atomic_read(&cache
->stats
.read_hit
),
3159 (unsigned) atomic_read(&cache
->stats
.read_miss
),
3160 (unsigned) atomic_read(&cache
->stats
.write_hit
),
3161 (unsigned) atomic_read(&cache
->stats
.write_miss
),
3162 (unsigned) atomic_read(&cache
->stats
.demotion
),
3163 (unsigned) atomic_read(&cache
->stats
.promotion
),
3164 (unsigned long) atomic_read(&cache
->nr_dirty
));
3166 emit_flags(cache
, result
, maxlen
, &sz
);
3168 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache
->migration_threshold
);
3170 DMEMIT("%s ", dm_cache_policy_get_name(cache
->policy
));
3172 r
= policy_emit_config_values(cache
->policy
, result
, maxlen
, &sz
);
3174 DMERR("%s: policy_emit_config_values returned %d",
3175 cache_device_name(cache
), r
);
3178 if (get_cache_mode(cache
) == CM_READ_ONLY
)
3183 r
= dm_cache_metadata_needs_check(cache
->cmd
, &needs_check
);
3185 if (r
|| needs_check
)
3186 DMEMIT("needs_check ");
3192 case STATUSTYPE_TABLE
:
3193 format_dev_t(buf
, cache
->metadata_dev
->bdev
->bd_dev
);
3195 format_dev_t(buf
, cache
->cache_dev
->bdev
->bd_dev
);
3197 format_dev_t(buf
, cache
->origin_dev
->bdev
->bd_dev
);
3200 for (i
= 0; i
< cache
->nr_ctr_args
- 1; i
++)
3201 DMEMIT(" %s", cache
->ctr_args
[i
]);
3202 if (cache
->nr_ctr_args
)
3203 DMEMIT(" %s", cache
->ctr_args
[cache
->nr_ctr_args
- 1]);
3213 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3214 * the one-past-the-end value.
3216 struct cblock_range
{
3222 * A cache block range can take two forms:
3224 * i) A single cblock, eg. '3456'
3225 * ii) A begin and end cblock with a dash between, eg. 123-234
3227 static int parse_cblock_range(struct cache
*cache
, const char *str
,
3228 struct cblock_range
*result
)
3235 * Try and parse form (ii) first.
3237 r
= sscanf(str
, "%llu-%llu%c", &b
, &e
, &dummy
);
3242 result
->begin
= to_cblock(b
);
3243 result
->end
= to_cblock(e
);
3248 * That didn't work, try form (i).
3250 r
= sscanf(str
, "%llu%c", &b
, &dummy
);
3255 result
->begin
= to_cblock(b
);
3256 result
->end
= to_cblock(from_cblock(result
->begin
) + 1u);
3260 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache
), str
);
3264 static int validate_cblock_range(struct cache
*cache
, struct cblock_range
*range
)
3266 uint64_t b
= from_cblock(range
->begin
);
3267 uint64_t e
= from_cblock(range
->end
);
3268 uint64_t n
= from_cblock(cache
->cache_size
);
3271 DMERR("%s: begin cblock out of range: %llu >= %llu",
3272 cache_device_name(cache
), b
, n
);
3277 DMERR("%s: end cblock out of range: %llu > %llu",
3278 cache_device_name(cache
), e
, n
);
3283 DMERR("%s: invalid cblock range: %llu >= %llu",
3284 cache_device_name(cache
), b
, e
);
3291 static inline dm_cblock_t
cblock_succ(dm_cblock_t b
)
3293 return to_cblock(from_cblock(b
) + 1);
3296 static int request_invalidation(struct cache
*cache
, struct cblock_range
*range
)
3301 * We don't need to do any locking here because we know we're in
3302 * passthrough mode. There's is potential for a race between an
3303 * invalidation triggered by an io and an invalidation message. This
3304 * is harmless, we must not worry if the policy call fails.
3306 while (range
->begin
!= range
->end
) {
3307 r
= invalidate_cblock(cache
, range
->begin
);
3311 range
->begin
= cblock_succ(range
->begin
);
3314 cache
->commit_requested
= true;
3318 static int process_invalidate_cblocks_message(struct cache
*cache
, unsigned count
,
3319 const char **cblock_ranges
)
3323 struct cblock_range range
;
3325 if (!passthrough_mode(cache
)) {
3326 DMERR("%s: cache has to be in passthrough mode for invalidation",
3327 cache_device_name(cache
));
3331 for (i
= 0; i
< count
; i
++) {
3332 r
= parse_cblock_range(cache
, cblock_ranges
[i
], &range
);
3336 r
= validate_cblock_range(cache
, &range
);
3341 * Pass begin and end origin blocks to the worker and wake it.
3343 r
= request_invalidation(cache
, &range
);
3355 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3357 * The key migration_threshold is supported by the cache target core.
3359 static int cache_message(struct dm_target
*ti
, unsigned argc
, char **argv
,
3360 char *result
, unsigned maxlen
)
3362 struct cache
*cache
= ti
->private;
3367 if (get_cache_mode(cache
) >= CM_READ_ONLY
) {
3368 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3369 cache_device_name(cache
));
3373 if (!strcasecmp(argv
[0], "invalidate_cblocks"))
3374 return process_invalidate_cblocks_message(cache
, argc
- 1, (const char **) argv
+ 1);
3379 return set_config_value(cache
, argv
[0], argv
[1]);
3382 static int cache_iterate_devices(struct dm_target
*ti
,
3383 iterate_devices_callout_fn fn
, void *data
)
3386 struct cache
*cache
= ti
->private;
3388 r
= fn(ti
, cache
->cache_dev
, 0, get_dev_size(cache
->cache_dev
), data
);
3390 r
= fn(ti
, cache
->origin_dev
, 0, ti
->len
, data
);
3395 static bool origin_dev_supports_discard(struct block_device
*origin_bdev
)
3397 struct request_queue
*q
= bdev_get_queue(origin_bdev
);
3399 return q
&& blk_queue_discard(q
);
3403 * If discard_passdown was enabled verify that the origin device
3404 * supports discards. Disable discard_passdown if not.
3406 static void disable_passdown_if_not_supported(struct cache
*cache
)
3408 struct block_device
*origin_bdev
= cache
->origin_dev
->bdev
;
3409 struct queue_limits
*origin_limits
= &bdev_get_queue(origin_bdev
)->limits
;
3410 const char *reason
= NULL
;
3411 char buf
[BDEVNAME_SIZE
];
3413 if (!cache
->features
.discard_passdown
)
3416 if (!origin_dev_supports_discard(origin_bdev
))
3417 reason
= "discard unsupported";
3419 else if (origin_limits
->max_discard_sectors
< cache
->sectors_per_block
)
3420 reason
= "max discard sectors smaller than a block";
3423 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3424 bdevname(origin_bdev
, buf
), reason
);
3425 cache
->features
.discard_passdown
= false;
3429 static void set_discard_limits(struct cache
*cache
, struct queue_limits
*limits
)
3431 struct block_device
*origin_bdev
= cache
->origin_dev
->bdev
;
3432 struct queue_limits
*origin_limits
= &bdev_get_queue(origin_bdev
)->limits
;
3434 if (!cache
->features
.discard_passdown
) {
3435 /* No passdown is done so setting own virtual limits */
3436 limits
->max_discard_sectors
= min_t(sector_t
, cache
->discard_block_size
* 1024,
3437 cache
->origin_sectors
);
3438 limits
->discard_granularity
= cache
->discard_block_size
<< SECTOR_SHIFT
;
3443 * cache_iterate_devices() is stacking both origin and fast device limits
3444 * but discards aren't passed to fast device, so inherit origin's limits.
3446 limits
->max_discard_sectors
= origin_limits
->max_discard_sectors
;
3447 limits
->max_hw_discard_sectors
= origin_limits
->max_hw_discard_sectors
;
3448 limits
->discard_granularity
= origin_limits
->discard_granularity
;
3449 limits
->discard_alignment
= origin_limits
->discard_alignment
;
3450 limits
->discard_misaligned
= origin_limits
->discard_misaligned
;
3453 static void cache_io_hints(struct dm_target
*ti
, struct queue_limits
*limits
)
3455 struct cache
*cache
= ti
->private;
3456 uint64_t io_opt_sectors
= limits
->io_opt
>> SECTOR_SHIFT
;
3459 * If the system-determined stacked limits are compatible with the
3460 * cache's blocksize (io_opt is a factor) do not override them.
3462 if (io_opt_sectors
< cache
->sectors_per_block
||
3463 do_div(io_opt_sectors
, cache
->sectors_per_block
)) {
3464 blk_limits_io_min(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3465 blk_limits_io_opt(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3468 disable_passdown_if_not_supported(cache
);
3469 set_discard_limits(cache
, limits
);
3472 /*----------------------------------------------------------------*/
3474 static struct target_type cache_target
= {
3476 .version
= {2, 2, 0},
3477 .module
= THIS_MODULE
,
3481 .end_io
= cache_end_io
,
3482 .postsuspend
= cache_postsuspend
,
3483 .preresume
= cache_preresume
,
3484 .resume
= cache_resume
,
3485 .status
= cache_status
,
3486 .message
= cache_message
,
3487 .iterate_devices
= cache_iterate_devices
,
3488 .io_hints
= cache_io_hints
,
3491 static int __init
dm_cache_init(void)
3495 migration_cache
= KMEM_CACHE(dm_cache_migration
, 0);
3496 if (!migration_cache
)
3499 r
= dm_register_target(&cache_target
);
3501 DMERR("cache target registration failed: %d", r
);
3502 kmem_cache_destroy(migration_cache
);
3509 static void __exit
dm_cache_exit(void)
3511 dm_unregister_target(&cache_target
);
3512 kmem_cache_destroy(migration_cache
);
3515 module_init(dm_cache_init
);
3516 module_exit(dm_cache_exit
);
3518 MODULE_DESCRIPTION(DM_NAME
" cache target");
3519 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3520 MODULE_LICENSE("GPL");