2 * Copyright (C) 2012 Red Hat. All rights reserved.
4 * This file is released under the GPL.
8 #include "dm-bio-prison.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/slab.h>
19 #include <linux/vmalloc.h>
21 #define DM_MSG_PREFIX "cache"
23 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle
,
24 "A percentage of time allocated for copying to and/or from cache");
26 /*----------------------------------------------------------------*/
28 #define IOT_RESOLUTION 4
34 * Sectors of in-flight IO.
39 * The time, in jiffies, when this device became idle (if it is
42 unsigned long idle_time
;
43 unsigned long last_update_time
;
46 static void iot_init(struct io_tracker
*iot
)
48 spin_lock_init(&iot
->lock
);
51 iot
->last_update_time
= jiffies
;
54 static bool __iot_idle_for(struct io_tracker
*iot
, unsigned long jifs
)
59 return time_after(jiffies
, iot
->idle_time
+ jifs
);
62 static bool iot_idle_for(struct io_tracker
*iot
, unsigned long jifs
)
67 spin_lock_irqsave(&iot
->lock
, flags
);
68 r
= __iot_idle_for(iot
, jifs
);
69 spin_unlock_irqrestore(&iot
->lock
, flags
);
74 static void iot_io_begin(struct io_tracker
*iot
, sector_t len
)
78 spin_lock_irqsave(&iot
->lock
, flags
);
79 iot
->in_flight
+= len
;
80 spin_unlock_irqrestore(&iot
->lock
, flags
);
83 static void __iot_io_end(struct io_tracker
*iot
, sector_t len
)
85 iot
->in_flight
-= len
;
87 iot
->idle_time
= jiffies
;
90 static void iot_io_end(struct io_tracker
*iot
, sector_t len
)
94 spin_lock_irqsave(&iot
->lock
, flags
);
95 __iot_io_end(iot
, len
);
96 spin_unlock_irqrestore(&iot
->lock
, flags
);
99 /*----------------------------------------------------------------*/
104 * oblock: index of an origin block
105 * cblock: index of a cache block
106 * promotion: movement of a block from origin to cache
107 * demotion: movement of a block from cache to origin
108 * migration: movement of a block between the origin and cache device,
112 /*----------------------------------------------------------------*/
115 * There are a couple of places where we let a bio run, but want to do some
116 * work before calling its endio function. We do this by temporarily
117 * changing the endio fn.
119 struct dm_hook_info
{
120 bio_end_io_t
*bi_end_io
;
124 static void dm_hook_bio(struct dm_hook_info
*h
, struct bio
*bio
,
125 bio_end_io_t
*bi_end_io
, void *bi_private
)
127 h
->bi_end_io
= bio
->bi_end_io
;
128 h
->bi_private
= bio
->bi_private
;
130 bio
->bi_end_io
= bi_end_io
;
131 bio
->bi_private
= bi_private
;
134 static void dm_unhook_bio(struct dm_hook_info
*h
, struct bio
*bio
)
136 bio
->bi_end_io
= h
->bi_end_io
;
137 bio
->bi_private
= h
->bi_private
;
140 /*----------------------------------------------------------------*/
142 #define MIGRATION_POOL_SIZE 128
143 #define COMMIT_PERIOD HZ
144 #define MIGRATION_COUNT_WINDOW 10
147 * The block size of the device holding cache data must be
148 * between 32KB and 1GB.
150 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
151 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
153 enum cache_metadata_mode
{
154 CM_WRITE
, /* metadata may be changed */
155 CM_READ_ONLY
, /* metadata may not be changed */
161 * Data is written to cached blocks only. These blocks are marked
162 * dirty. If you lose the cache device you will lose data.
163 * Potential performance increase for both reads and writes.
168 * Data is written to both cache and origin. Blocks are never
169 * dirty. Potential performance benfit for reads only.
174 * A degraded mode useful for various cache coherency situations
175 * (eg, rolling back snapshots). Reads and writes always go to the
176 * origin. If a write goes to a cached oblock, then the cache
177 * block is invalidated.
182 struct cache_features
{
183 enum cache_metadata_mode mode
;
184 enum cache_io_mode io_mode
;
194 atomic_t copies_avoided
;
195 atomic_t cache_cell_clash
;
196 atomic_t commit_count
;
197 atomic_t discard_count
;
201 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
202 * the one-past-the-end value.
204 struct cblock_range
{
209 struct invalidation_request
{
210 struct list_head list
;
211 struct cblock_range
*cblocks
;
216 wait_queue_head_t result_wait
;
220 struct dm_target
*ti
;
221 struct dm_target_callbacks callbacks
;
223 struct dm_cache_metadata
*cmd
;
226 * Metadata is written to this device.
228 struct dm_dev
*metadata_dev
;
231 * The slower of the two data devices. Typically a spindle.
233 struct dm_dev
*origin_dev
;
236 * The faster of the two data devices. Typically an SSD.
238 struct dm_dev
*cache_dev
;
241 * Size of the origin device in _complete_ blocks and native sectors.
243 dm_oblock_t origin_blocks
;
244 sector_t origin_sectors
;
247 * Size of the cache device in blocks.
249 dm_cblock_t cache_size
;
252 * Fields for converting from sectors to blocks.
254 uint32_t sectors_per_block
;
255 int sectors_per_block_shift
;
258 struct list_head deferred_cells
;
259 struct bio_list deferred_bios
;
260 struct bio_list deferred_flush_bios
;
261 struct bio_list deferred_writethrough_bios
;
262 struct list_head quiesced_migrations
;
263 struct list_head completed_migrations
;
264 struct list_head need_commit_migrations
;
265 sector_t migration_threshold
;
266 wait_queue_head_t migration_wait
;
267 atomic_t nr_allocated_migrations
;
270 * The number of in flight migrations that are performing
271 * background io. eg, promotion, writeback.
273 atomic_t nr_io_migrations
;
275 wait_queue_head_t quiescing_wait
;
277 atomic_t quiescing_ack
;
280 * cache_size entries, dirty if set
283 unsigned long *dirty_bitset
;
286 * origin_blocks entries, discarded if set.
288 dm_dblock_t discard_nr_blocks
;
289 unsigned long *discard_bitset
;
290 uint32_t discard_block_size
; /* a power of 2 times sectors per block */
293 * Rather than reconstructing the table line for the status we just
294 * save it and regurgitate.
296 unsigned nr_ctr_args
;
297 const char **ctr_args
;
299 struct dm_kcopyd_client
*copier
;
300 struct workqueue_struct
*wq
;
301 struct work_struct worker
;
303 struct delayed_work waker
;
304 unsigned long last_commit_jiffies
;
306 struct dm_bio_prison
*prison
;
307 struct dm_deferred_set
*all_io_ds
;
309 mempool_t
*migration_pool
;
311 struct dm_cache_policy
*policy
;
312 unsigned policy_nr_args
;
314 bool need_tick_bio
:1;
317 bool commit_requested
:1;
318 bool loaded_mappings
:1;
319 bool loaded_discards
:1;
322 * Cache features such as write-through.
324 struct cache_features features
;
326 struct cache_stats stats
;
329 * Invalidation fields.
331 spinlock_t invalidation_lock
;
332 struct list_head invalidation_requests
;
334 struct io_tracker origin_tracker
;
337 struct per_bio_data
{
340 struct dm_deferred_entry
*all_io_entry
;
341 struct dm_hook_info hook_info
;
345 * writethrough fields. These MUST remain at the end of this
346 * structure and the 'cache' member must be the first as it
347 * is used to determine the offset of the writethrough fields.
351 struct dm_bio_details bio_details
;
354 struct dm_cache_migration
{
355 struct list_head list
;
358 unsigned long start_jiffies
;
359 dm_oblock_t old_oblock
;
360 dm_oblock_t new_oblock
;
368 bool requeue_holder
:1;
371 struct dm_bio_prison_cell
*old_ocell
;
372 struct dm_bio_prison_cell
*new_ocell
;
376 * Processing a bio in the worker thread may require these memory
377 * allocations. We prealloc to avoid deadlocks (the same worker thread
378 * frees them back to the mempool).
381 struct dm_cache_migration
*mg
;
382 struct dm_bio_prison_cell
*cell1
;
383 struct dm_bio_prison_cell
*cell2
;
386 static enum cache_metadata_mode
get_cache_mode(struct cache
*cache
);
388 static void wake_worker(struct cache
*cache
)
390 queue_work(cache
->wq
, &cache
->worker
);
393 /*----------------------------------------------------------------*/
395 static struct dm_bio_prison_cell
*alloc_prison_cell(struct cache
*cache
)
397 /* FIXME: change to use a local slab. */
398 return dm_bio_prison_alloc_cell(cache
->prison
, GFP_NOWAIT
);
401 static void free_prison_cell(struct cache
*cache
, struct dm_bio_prison_cell
*cell
)
403 dm_bio_prison_free_cell(cache
->prison
, cell
);
406 static struct dm_cache_migration
*alloc_migration(struct cache
*cache
)
408 struct dm_cache_migration
*mg
;
410 mg
= mempool_alloc(cache
->migration_pool
, GFP_NOWAIT
);
413 atomic_inc(&mg
->cache
->nr_allocated_migrations
);
419 static void free_migration(struct dm_cache_migration
*mg
)
421 struct cache
*cache
= mg
->cache
;
423 if (atomic_dec_and_test(&cache
->nr_allocated_migrations
))
424 wake_up(&cache
->migration_wait
);
426 mempool_free(mg
, cache
->migration_pool
);
429 static int prealloc_data_structs(struct cache
*cache
, struct prealloc
*p
)
432 p
->mg
= alloc_migration(cache
);
438 p
->cell1
= alloc_prison_cell(cache
);
444 p
->cell2
= alloc_prison_cell(cache
);
452 static void prealloc_free_structs(struct cache
*cache
, struct prealloc
*p
)
455 free_prison_cell(cache
, p
->cell2
);
458 free_prison_cell(cache
, p
->cell1
);
461 free_migration(p
->mg
);
464 static struct dm_cache_migration
*prealloc_get_migration(struct prealloc
*p
)
466 struct dm_cache_migration
*mg
= p
->mg
;
475 * You must have a cell within the prealloc struct to return. If not this
476 * function will BUG() rather than returning NULL.
478 static struct dm_bio_prison_cell
*prealloc_get_cell(struct prealloc
*p
)
480 struct dm_bio_prison_cell
*r
= NULL
;
486 } else if (p
->cell2
) {
496 * You can't have more than two cells in a prealloc struct. BUG() will be
497 * called if you try and overfill.
499 static void prealloc_put_cell(struct prealloc
*p
, struct dm_bio_prison_cell
*cell
)
511 /*----------------------------------------------------------------*/
513 static void build_key(dm_oblock_t begin
, dm_oblock_t end
, struct dm_cell_key
*key
)
517 key
->block_begin
= from_oblock(begin
);
518 key
->block_end
= from_oblock(end
);
522 * The caller hands in a preallocated cell, and a free function for it.
523 * The cell will be freed if there's an error, or if it wasn't used because
524 * a cell with that key already exists.
526 typedef void (*cell_free_fn
)(void *context
, struct dm_bio_prison_cell
*cell
);
528 static int bio_detain_range(struct cache
*cache
, dm_oblock_t oblock_begin
, dm_oblock_t oblock_end
,
529 struct bio
*bio
, struct dm_bio_prison_cell
*cell_prealloc
,
530 cell_free_fn free_fn
, void *free_context
,
531 struct dm_bio_prison_cell
**cell_result
)
534 struct dm_cell_key key
;
536 build_key(oblock_begin
, oblock_end
, &key
);
537 r
= dm_bio_detain(cache
->prison
, &key
, bio
, cell_prealloc
, cell_result
);
539 free_fn(free_context
, cell_prealloc
);
544 static int bio_detain(struct cache
*cache
, dm_oblock_t oblock
,
545 struct bio
*bio
, struct dm_bio_prison_cell
*cell_prealloc
,
546 cell_free_fn free_fn
, void *free_context
,
547 struct dm_bio_prison_cell
**cell_result
)
549 dm_oblock_t end
= to_oblock(from_oblock(oblock
) + 1ULL);
550 return bio_detain_range(cache
, oblock
, end
, bio
,
551 cell_prealloc
, free_fn
, free_context
, cell_result
);
554 static int get_cell(struct cache
*cache
,
556 struct prealloc
*structs
,
557 struct dm_bio_prison_cell
**cell_result
)
560 struct dm_cell_key key
;
561 struct dm_bio_prison_cell
*cell_prealloc
;
563 cell_prealloc
= prealloc_get_cell(structs
);
565 build_key(oblock
, to_oblock(from_oblock(oblock
) + 1ULL), &key
);
566 r
= dm_get_cell(cache
->prison
, &key
, cell_prealloc
, cell_result
);
568 prealloc_put_cell(structs
, cell_prealloc
);
573 /*----------------------------------------------------------------*/
575 static bool is_dirty(struct cache
*cache
, dm_cblock_t b
)
577 return test_bit(from_cblock(b
), cache
->dirty_bitset
);
580 static void set_dirty(struct cache
*cache
, dm_oblock_t oblock
, dm_cblock_t cblock
)
582 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
583 atomic_inc(&cache
->nr_dirty
);
584 policy_set_dirty(cache
->policy
, oblock
);
588 static void clear_dirty(struct cache
*cache
, dm_oblock_t oblock
, dm_cblock_t cblock
)
590 if (test_and_clear_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
591 policy_clear_dirty(cache
->policy
, oblock
);
592 if (atomic_dec_return(&cache
->nr_dirty
) == 0)
593 dm_table_event(cache
->ti
->table
);
597 /*----------------------------------------------------------------*/
599 static bool block_size_is_power_of_two(struct cache
*cache
)
601 return cache
->sectors_per_block_shift
>= 0;
604 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
605 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
608 static dm_block_t
block_div(dm_block_t b
, uint32_t n
)
615 static dm_block_t
oblocks_per_dblock(struct cache
*cache
)
617 dm_block_t oblocks
= cache
->discard_block_size
;
619 if (block_size_is_power_of_two(cache
))
620 oblocks
>>= cache
->sectors_per_block_shift
;
622 oblocks
= block_div(oblocks
, cache
->sectors_per_block
);
627 static dm_dblock_t
oblock_to_dblock(struct cache
*cache
, dm_oblock_t oblock
)
629 return to_dblock(block_div(from_oblock(oblock
),
630 oblocks_per_dblock(cache
)));
633 static dm_oblock_t
dblock_to_oblock(struct cache
*cache
, dm_dblock_t dblock
)
635 return to_oblock(from_dblock(dblock
) * oblocks_per_dblock(cache
));
638 static void set_discard(struct cache
*cache
, dm_dblock_t b
)
642 BUG_ON(from_dblock(b
) >= from_dblock(cache
->discard_nr_blocks
));
643 atomic_inc(&cache
->stats
.discard_count
);
645 spin_lock_irqsave(&cache
->lock
, flags
);
646 set_bit(from_dblock(b
), cache
->discard_bitset
);
647 spin_unlock_irqrestore(&cache
->lock
, flags
);
650 static void clear_discard(struct cache
*cache
, dm_dblock_t b
)
654 spin_lock_irqsave(&cache
->lock
, flags
);
655 clear_bit(from_dblock(b
), cache
->discard_bitset
);
656 spin_unlock_irqrestore(&cache
->lock
, flags
);
659 static bool is_discarded(struct cache
*cache
, dm_dblock_t b
)
664 spin_lock_irqsave(&cache
->lock
, flags
);
665 r
= test_bit(from_dblock(b
), cache
->discard_bitset
);
666 spin_unlock_irqrestore(&cache
->lock
, flags
);
671 static bool is_discarded_oblock(struct cache
*cache
, dm_oblock_t b
)
676 spin_lock_irqsave(&cache
->lock
, flags
);
677 r
= test_bit(from_dblock(oblock_to_dblock(cache
, b
)),
678 cache
->discard_bitset
);
679 spin_unlock_irqrestore(&cache
->lock
, flags
);
684 /*----------------------------------------------------------------*/
686 static void load_stats(struct cache
*cache
)
688 struct dm_cache_statistics stats
;
690 dm_cache_metadata_get_stats(cache
->cmd
, &stats
);
691 atomic_set(&cache
->stats
.read_hit
, stats
.read_hits
);
692 atomic_set(&cache
->stats
.read_miss
, stats
.read_misses
);
693 atomic_set(&cache
->stats
.write_hit
, stats
.write_hits
);
694 atomic_set(&cache
->stats
.write_miss
, stats
.write_misses
);
697 static void save_stats(struct cache
*cache
)
699 struct dm_cache_statistics stats
;
701 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
704 stats
.read_hits
= atomic_read(&cache
->stats
.read_hit
);
705 stats
.read_misses
= atomic_read(&cache
->stats
.read_miss
);
706 stats
.write_hits
= atomic_read(&cache
->stats
.write_hit
);
707 stats
.write_misses
= atomic_read(&cache
->stats
.write_miss
);
709 dm_cache_metadata_set_stats(cache
->cmd
, &stats
);
712 /*----------------------------------------------------------------
714 *--------------------------------------------------------------*/
717 * If using writeback, leave out struct per_bio_data's writethrough fields.
719 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
720 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
722 static bool writethrough_mode(struct cache_features
*f
)
724 return f
->io_mode
== CM_IO_WRITETHROUGH
;
727 static bool writeback_mode(struct cache_features
*f
)
729 return f
->io_mode
== CM_IO_WRITEBACK
;
732 static bool passthrough_mode(struct cache_features
*f
)
734 return f
->io_mode
== CM_IO_PASSTHROUGH
;
737 static size_t get_per_bio_data_size(struct cache
*cache
)
739 return writethrough_mode(&cache
->features
) ? PB_DATA_SIZE_WT
: PB_DATA_SIZE_WB
;
742 static struct per_bio_data
*get_per_bio_data(struct bio
*bio
, size_t data_size
)
744 struct per_bio_data
*pb
= dm_per_bio_data(bio
, data_size
);
749 static struct per_bio_data
*init_per_bio_data(struct bio
*bio
, size_t data_size
)
751 struct per_bio_data
*pb
= get_per_bio_data(bio
, data_size
);
754 pb
->req_nr
= dm_bio_get_target_bio_nr(bio
);
755 pb
->all_io_entry
= NULL
;
761 /*----------------------------------------------------------------
763 *--------------------------------------------------------------*/
764 static void remap_to_origin(struct cache
*cache
, struct bio
*bio
)
766 bio
->bi_bdev
= cache
->origin_dev
->bdev
;
769 static void remap_to_cache(struct cache
*cache
, struct bio
*bio
,
772 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
773 sector_t block
= from_cblock(cblock
);
775 bio
->bi_bdev
= cache
->cache_dev
->bdev
;
776 if (!block_size_is_power_of_two(cache
))
777 bio
->bi_iter
.bi_sector
=
778 (block
* cache
->sectors_per_block
) +
779 sector_div(bi_sector
, cache
->sectors_per_block
);
781 bio
->bi_iter
.bi_sector
=
782 (block
<< cache
->sectors_per_block_shift
) |
783 (bi_sector
& (cache
->sectors_per_block
- 1));
786 static void check_if_tick_bio_needed(struct cache
*cache
, struct bio
*bio
)
789 size_t pb_data_size
= get_per_bio_data_size(cache
);
790 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
792 spin_lock_irqsave(&cache
->lock
, flags
);
793 if (cache
->need_tick_bio
&&
794 !(bio
->bi_rw
& (REQ_FUA
| REQ_FLUSH
| REQ_DISCARD
))) {
796 cache
->need_tick_bio
= false;
798 spin_unlock_irqrestore(&cache
->lock
, flags
);
801 static void remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
804 check_if_tick_bio_needed(cache
, bio
);
805 remap_to_origin(cache
, bio
);
806 if (bio_data_dir(bio
) == WRITE
)
807 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
810 static void remap_to_cache_dirty(struct cache
*cache
, struct bio
*bio
,
811 dm_oblock_t oblock
, dm_cblock_t cblock
)
813 check_if_tick_bio_needed(cache
, bio
);
814 remap_to_cache(cache
, bio
, cblock
);
815 if (bio_data_dir(bio
) == WRITE
) {
816 set_dirty(cache
, oblock
, cblock
);
817 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
821 static dm_oblock_t
get_bio_block(struct cache
*cache
, struct bio
*bio
)
823 sector_t block_nr
= bio
->bi_iter
.bi_sector
;
825 if (!block_size_is_power_of_two(cache
))
826 (void) sector_div(block_nr
, cache
->sectors_per_block
);
828 block_nr
>>= cache
->sectors_per_block_shift
;
830 return to_oblock(block_nr
);
833 static int bio_triggers_commit(struct cache
*cache
, struct bio
*bio
)
835 return bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
839 * You must increment the deferred set whilst the prison cell is held. To
840 * encourage this, we ask for 'cell' to be passed in.
842 static void inc_ds(struct cache
*cache
, struct bio
*bio
,
843 struct dm_bio_prison_cell
*cell
)
845 size_t pb_data_size
= get_per_bio_data_size(cache
);
846 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
849 BUG_ON(pb
->all_io_entry
);
851 pb
->all_io_entry
= dm_deferred_entry_inc(cache
->all_io_ds
);
854 static bool accountable_bio(struct cache
*cache
, struct bio
*bio
)
856 return ((bio
->bi_bdev
== cache
->origin_dev
->bdev
) &&
857 !(bio
->bi_rw
& REQ_DISCARD
));
860 static void accounted_begin(struct cache
*cache
, struct bio
*bio
)
862 size_t pb_data_size
= get_per_bio_data_size(cache
);
863 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
865 if (accountable_bio(cache
, bio
)) {
866 pb
->len
= bio_sectors(bio
);
867 iot_io_begin(&cache
->origin_tracker
, pb
->len
);
871 static void accounted_complete(struct cache
*cache
, struct bio
*bio
)
873 size_t pb_data_size
= get_per_bio_data_size(cache
);
874 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
876 iot_io_end(&cache
->origin_tracker
, pb
->len
);
879 static void accounted_request(struct cache
*cache
, struct bio
*bio
)
881 accounted_begin(cache
, bio
);
882 generic_make_request(bio
);
885 static void issue(struct cache
*cache
, struct bio
*bio
)
889 if (!bio_triggers_commit(cache
, bio
)) {
890 accounted_request(cache
, bio
);
895 * Batch together any bios that trigger commits and then issue a
896 * single commit for them in do_worker().
898 spin_lock_irqsave(&cache
->lock
, flags
);
899 cache
->commit_requested
= true;
900 bio_list_add(&cache
->deferred_flush_bios
, bio
);
901 spin_unlock_irqrestore(&cache
->lock
, flags
);
904 static void inc_and_issue(struct cache
*cache
, struct bio
*bio
, struct dm_bio_prison_cell
*cell
)
906 inc_ds(cache
, bio
, cell
);
910 static void defer_writethrough_bio(struct cache
*cache
, struct bio
*bio
)
914 spin_lock_irqsave(&cache
->lock
, flags
);
915 bio_list_add(&cache
->deferred_writethrough_bios
, bio
);
916 spin_unlock_irqrestore(&cache
->lock
, flags
);
921 static void writethrough_endio(struct bio
*bio
, int err
)
923 struct per_bio_data
*pb
= get_per_bio_data(bio
, PB_DATA_SIZE_WT
);
925 dm_unhook_bio(&pb
->hook_info
, bio
);
932 dm_bio_restore(&pb
->bio_details
, bio
);
933 remap_to_cache(pb
->cache
, bio
, pb
->cblock
);
936 * We can't issue this bio directly, since we're in interrupt
937 * context. So it gets put on a bio list for processing by the
940 defer_writethrough_bio(pb
->cache
, bio
);
944 * When running in writethrough mode we need to send writes to clean blocks
945 * to both the cache and origin devices. In future we'd like to clone the
946 * bio and send them in parallel, but for now we're doing them in
947 * series as this is easier.
949 static void remap_to_origin_then_cache(struct cache
*cache
, struct bio
*bio
,
950 dm_oblock_t oblock
, dm_cblock_t cblock
)
952 struct per_bio_data
*pb
= get_per_bio_data(bio
, PB_DATA_SIZE_WT
);
956 dm_hook_bio(&pb
->hook_info
, bio
, writethrough_endio
, NULL
);
957 dm_bio_record(&pb
->bio_details
, bio
);
959 remap_to_origin_clear_discard(pb
->cache
, bio
, oblock
);
962 /*----------------------------------------------------------------
964 *--------------------------------------------------------------*/
965 static enum cache_metadata_mode
get_cache_mode(struct cache
*cache
)
967 return cache
->features
.mode
;
970 static const char *cache_device_name(struct cache
*cache
)
972 return dm_device_name(dm_table_get_md(cache
->ti
->table
));
975 static void notify_mode_switch(struct cache
*cache
, enum cache_metadata_mode mode
)
977 const char *descs
[] = {
983 dm_table_event(cache
->ti
->table
);
984 DMINFO("%s: switching cache to %s mode",
985 cache_device_name(cache
), descs
[(int)mode
]);
988 static void set_cache_mode(struct cache
*cache
, enum cache_metadata_mode new_mode
)
990 bool needs_check
= dm_cache_metadata_needs_check(cache
->cmd
);
991 enum cache_metadata_mode old_mode
= get_cache_mode(cache
);
993 if (new_mode
== CM_WRITE
&& needs_check
) {
994 DMERR("%s: unable to switch cache to write mode until repaired.",
995 cache_device_name(cache
));
996 if (old_mode
!= new_mode
)
999 new_mode
= CM_READ_ONLY
;
1002 /* Never move out of fail mode */
1003 if (old_mode
== CM_FAIL
)
1009 dm_cache_metadata_set_read_only(cache
->cmd
);
1013 dm_cache_metadata_set_read_write(cache
->cmd
);
1017 cache
->features
.mode
= new_mode
;
1019 if (new_mode
!= old_mode
)
1020 notify_mode_switch(cache
, new_mode
);
1023 static void abort_transaction(struct cache
*cache
)
1025 const char *dev_name
= cache_device_name(cache
);
1027 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1030 if (dm_cache_metadata_set_needs_check(cache
->cmd
)) {
1031 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name
);
1032 set_cache_mode(cache
, CM_FAIL
);
1035 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name
);
1036 if (dm_cache_metadata_abort(cache
->cmd
)) {
1037 DMERR("%s: failed to abort metadata transaction", dev_name
);
1038 set_cache_mode(cache
, CM_FAIL
);
1042 static void metadata_operation_failed(struct cache
*cache
, const char *op
, int r
)
1044 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1045 cache_device_name(cache
), op
, r
);
1046 abort_transaction(cache
);
1047 set_cache_mode(cache
, CM_READ_ONLY
);
1050 /*----------------------------------------------------------------
1051 * Migration processing
1053 * Migration covers moving data from the origin device to the cache, or
1055 *--------------------------------------------------------------*/
1056 static void inc_io_migrations(struct cache
*cache
)
1058 atomic_inc(&cache
->nr_io_migrations
);
1061 static void dec_io_migrations(struct cache
*cache
)
1063 atomic_dec(&cache
->nr_io_migrations
);
1066 static void __cell_release(struct cache
*cache
, struct dm_bio_prison_cell
*cell
,
1067 bool holder
, struct bio_list
*bios
)
1069 (holder
? dm_cell_release
: dm_cell_release_no_holder
)
1070 (cache
->prison
, cell
, bios
);
1071 free_prison_cell(cache
, cell
);
1074 static bool discard_or_flush(struct bio
*bio
)
1076 return bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
| REQ_DISCARD
);
1079 static void __cell_defer(struct cache
*cache
, struct dm_bio_prison_cell
*cell
)
1081 if (discard_or_flush(cell
->holder
))
1083 * We have to handle these bios
1086 __cell_release(cache
, cell
, true, &cache
->deferred_bios
);
1089 list_add_tail(&cell
->user_list
, &cache
->deferred_cells
);
1092 static void cell_defer(struct cache
*cache
, struct dm_bio_prison_cell
*cell
, bool holder
)
1094 unsigned long flags
;
1096 if (!holder
&& dm_cell_promote_or_release(cache
->prison
, cell
)) {
1098 * There was no prisoner to promote to holder, the
1099 * cell has been released.
1101 free_prison_cell(cache
, cell
);
1105 spin_lock_irqsave(&cache
->lock
, flags
);
1106 __cell_defer(cache
, cell
);
1107 spin_unlock_irqrestore(&cache
->lock
, flags
);
1112 static void cell_error_with_code(struct cache
*cache
, struct dm_bio_prison_cell
*cell
, int err
)
1114 dm_cell_error(cache
->prison
, cell
, err
);
1115 dm_bio_prison_free_cell(cache
->prison
, cell
);
1118 static void cell_requeue(struct cache
*cache
, struct dm_bio_prison_cell
*cell
)
1120 cell_error_with_code(cache
, cell
, DM_ENDIO_REQUEUE
);
1123 static void free_io_migration(struct dm_cache_migration
*mg
)
1125 dec_io_migrations(mg
->cache
);
1129 static void migration_failure(struct dm_cache_migration
*mg
)
1131 struct cache
*cache
= mg
->cache
;
1132 const char *dev_name
= cache_device_name(cache
);
1134 if (mg
->writeback
) {
1135 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name
);
1136 set_dirty(cache
, mg
->old_oblock
, mg
->cblock
);
1137 cell_defer(cache
, mg
->old_ocell
, false);
1139 } else if (mg
->demote
) {
1140 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name
);
1141 policy_force_mapping(cache
->policy
, mg
->new_oblock
, mg
->old_oblock
);
1143 cell_defer(cache
, mg
->old_ocell
, mg
->promote
? false : true);
1145 cell_defer(cache
, mg
->new_ocell
, true);
1147 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name
);
1148 policy_remove_mapping(cache
->policy
, mg
->new_oblock
);
1149 cell_defer(cache
, mg
->new_ocell
, true);
1152 free_io_migration(mg
);
1155 static void migration_success_pre_commit(struct dm_cache_migration
*mg
)
1158 unsigned long flags
;
1159 struct cache
*cache
= mg
->cache
;
1161 if (mg
->writeback
) {
1162 clear_dirty(cache
, mg
->old_oblock
, mg
->cblock
);
1163 cell_defer(cache
, mg
->old_ocell
, false);
1164 free_io_migration(mg
);
1167 } else if (mg
->demote
) {
1168 r
= dm_cache_remove_mapping(cache
->cmd
, mg
->cblock
);
1170 DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1171 cache_device_name(cache
));
1172 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1173 policy_force_mapping(cache
->policy
, mg
->new_oblock
,
1176 cell_defer(cache
, mg
->new_ocell
, true);
1177 free_io_migration(mg
);
1181 r
= dm_cache_insert_mapping(cache
->cmd
, mg
->cblock
, mg
->new_oblock
);
1183 DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1184 cache_device_name(cache
));
1185 metadata_operation_failed(cache
, "dm_cache_insert_mapping", r
);
1186 policy_remove_mapping(cache
->policy
, mg
->new_oblock
);
1187 free_io_migration(mg
);
1192 spin_lock_irqsave(&cache
->lock
, flags
);
1193 list_add_tail(&mg
->list
, &cache
->need_commit_migrations
);
1194 cache
->commit_requested
= true;
1195 spin_unlock_irqrestore(&cache
->lock
, flags
);
1198 static void migration_success_post_commit(struct dm_cache_migration
*mg
)
1200 unsigned long flags
;
1201 struct cache
*cache
= mg
->cache
;
1203 if (mg
->writeback
) {
1204 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1205 cache_device_name(cache
));
1208 } else if (mg
->demote
) {
1209 cell_defer(cache
, mg
->old_ocell
, mg
->promote
? false : true);
1214 spin_lock_irqsave(&cache
->lock
, flags
);
1215 list_add_tail(&mg
->list
, &cache
->quiesced_migrations
);
1216 spin_unlock_irqrestore(&cache
->lock
, flags
);
1220 policy_remove_mapping(cache
->policy
, mg
->old_oblock
);
1221 free_io_migration(mg
);
1225 if (mg
->requeue_holder
) {
1226 clear_dirty(cache
, mg
->new_oblock
, mg
->cblock
);
1227 cell_defer(cache
, mg
->new_ocell
, true);
1230 * The block was promoted via an overwrite, so it's dirty.
1232 set_dirty(cache
, mg
->new_oblock
, mg
->cblock
);
1233 bio_endio(mg
->new_ocell
->holder
, 0);
1234 cell_defer(cache
, mg
->new_ocell
, false);
1236 free_io_migration(mg
);
1240 static void copy_complete(int read_err
, unsigned long write_err
, void *context
)
1242 unsigned long flags
;
1243 struct dm_cache_migration
*mg
= (struct dm_cache_migration
*) context
;
1244 struct cache
*cache
= mg
->cache
;
1246 if (read_err
|| write_err
)
1249 spin_lock_irqsave(&cache
->lock
, flags
);
1250 list_add_tail(&mg
->list
, &cache
->completed_migrations
);
1251 spin_unlock_irqrestore(&cache
->lock
, flags
);
1256 static void issue_copy(struct dm_cache_migration
*mg
)
1259 struct dm_io_region o_region
, c_region
;
1260 struct cache
*cache
= mg
->cache
;
1261 sector_t cblock
= from_cblock(mg
->cblock
);
1263 o_region
.bdev
= cache
->origin_dev
->bdev
;
1264 o_region
.count
= cache
->sectors_per_block
;
1266 c_region
.bdev
= cache
->cache_dev
->bdev
;
1267 c_region
.sector
= cblock
* cache
->sectors_per_block
;
1268 c_region
.count
= cache
->sectors_per_block
;
1270 if (mg
->writeback
|| mg
->demote
) {
1272 o_region
.sector
= from_oblock(mg
->old_oblock
) * cache
->sectors_per_block
;
1273 r
= dm_kcopyd_copy(cache
->copier
, &c_region
, 1, &o_region
, 0, copy_complete
, mg
);
1276 o_region
.sector
= from_oblock(mg
->new_oblock
) * cache
->sectors_per_block
;
1277 r
= dm_kcopyd_copy(cache
->copier
, &o_region
, 1, &c_region
, 0, copy_complete
, mg
);
1281 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache
));
1282 migration_failure(mg
);
1286 static void overwrite_endio(struct bio
*bio
, int err
)
1288 struct dm_cache_migration
*mg
= bio
->bi_private
;
1289 struct cache
*cache
= mg
->cache
;
1290 size_t pb_data_size
= get_per_bio_data_size(cache
);
1291 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1292 unsigned long flags
;
1294 dm_unhook_bio(&pb
->hook_info
, bio
);
1299 mg
->requeue_holder
= false;
1301 spin_lock_irqsave(&cache
->lock
, flags
);
1302 list_add_tail(&mg
->list
, &cache
->completed_migrations
);
1303 spin_unlock_irqrestore(&cache
->lock
, flags
);
1308 static void issue_overwrite(struct dm_cache_migration
*mg
, struct bio
*bio
)
1310 size_t pb_data_size
= get_per_bio_data_size(mg
->cache
);
1311 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1313 dm_hook_bio(&pb
->hook_info
, bio
, overwrite_endio
, mg
);
1314 remap_to_cache_dirty(mg
->cache
, bio
, mg
->new_oblock
, mg
->cblock
);
1317 * No need to inc_ds() here, since the cell will be held for the
1318 * duration of the io.
1320 accounted_request(mg
->cache
, bio
);
1323 static bool bio_writes_complete_block(struct cache
*cache
, struct bio
*bio
)
1325 return (bio_data_dir(bio
) == WRITE
) &&
1326 (bio
->bi_iter
.bi_size
== (cache
->sectors_per_block
<< SECTOR_SHIFT
));
1329 static void avoid_copy(struct dm_cache_migration
*mg
)
1331 atomic_inc(&mg
->cache
->stats
.copies_avoided
);
1332 migration_success_pre_commit(mg
);
1335 static void calc_discard_block_range(struct cache
*cache
, struct bio
*bio
,
1336 dm_dblock_t
*b
, dm_dblock_t
*e
)
1338 sector_t sb
= bio
->bi_iter
.bi_sector
;
1339 sector_t se
= bio_end_sector(bio
);
1341 *b
= to_dblock(dm_sector_div_up(sb
, cache
->discard_block_size
));
1343 if (se
- sb
< cache
->discard_block_size
)
1346 *e
= to_dblock(block_div(se
, cache
->discard_block_size
));
1349 static void issue_discard(struct dm_cache_migration
*mg
)
1352 struct bio
*bio
= mg
->new_ocell
->holder
;
1354 calc_discard_block_range(mg
->cache
, bio
, &b
, &e
);
1356 set_discard(mg
->cache
, b
);
1357 b
= to_dblock(from_dblock(b
) + 1);
1361 cell_defer(mg
->cache
, mg
->new_ocell
, false);
1365 static void issue_copy_or_discard(struct dm_cache_migration
*mg
)
1368 struct cache
*cache
= mg
->cache
;
1375 if (mg
->writeback
|| mg
->demote
)
1376 avoid
= !is_dirty(cache
, mg
->cblock
) ||
1377 is_discarded_oblock(cache
, mg
->old_oblock
);
1379 struct bio
*bio
= mg
->new_ocell
->holder
;
1381 avoid
= is_discarded_oblock(cache
, mg
->new_oblock
);
1383 if (writeback_mode(&cache
->features
) &&
1384 !avoid
&& bio_writes_complete_block(cache
, bio
)) {
1385 issue_overwrite(mg
, bio
);
1390 avoid
? avoid_copy(mg
) : issue_copy(mg
);
1393 static void complete_migration(struct dm_cache_migration
*mg
)
1396 migration_failure(mg
);
1398 migration_success_pre_commit(mg
);
1401 static void process_migrations(struct cache
*cache
, struct list_head
*head
,
1402 void (*fn
)(struct dm_cache_migration
*))
1404 unsigned long flags
;
1405 struct list_head list
;
1406 struct dm_cache_migration
*mg
, *tmp
;
1408 INIT_LIST_HEAD(&list
);
1409 spin_lock_irqsave(&cache
->lock
, flags
);
1410 list_splice_init(head
, &list
);
1411 spin_unlock_irqrestore(&cache
->lock
, flags
);
1413 list_for_each_entry_safe(mg
, tmp
, &list
, list
)
1417 static void __queue_quiesced_migration(struct dm_cache_migration
*mg
)
1419 list_add_tail(&mg
->list
, &mg
->cache
->quiesced_migrations
);
1422 static void queue_quiesced_migration(struct dm_cache_migration
*mg
)
1424 unsigned long flags
;
1425 struct cache
*cache
= mg
->cache
;
1427 spin_lock_irqsave(&cache
->lock
, flags
);
1428 __queue_quiesced_migration(mg
);
1429 spin_unlock_irqrestore(&cache
->lock
, flags
);
1434 static void queue_quiesced_migrations(struct cache
*cache
, struct list_head
*work
)
1436 unsigned long flags
;
1437 struct dm_cache_migration
*mg
, *tmp
;
1439 spin_lock_irqsave(&cache
->lock
, flags
);
1440 list_for_each_entry_safe(mg
, tmp
, work
, list
)
1441 __queue_quiesced_migration(mg
);
1442 spin_unlock_irqrestore(&cache
->lock
, flags
);
1447 static void check_for_quiesced_migrations(struct cache
*cache
,
1448 struct per_bio_data
*pb
)
1450 struct list_head work
;
1452 if (!pb
->all_io_entry
)
1455 INIT_LIST_HEAD(&work
);
1456 dm_deferred_entry_dec(pb
->all_io_entry
, &work
);
1458 if (!list_empty(&work
))
1459 queue_quiesced_migrations(cache
, &work
);
1462 static void quiesce_migration(struct dm_cache_migration
*mg
)
1464 if (!dm_deferred_set_add_work(mg
->cache
->all_io_ds
, &mg
->list
))
1465 queue_quiesced_migration(mg
);
1468 static void promote(struct cache
*cache
, struct prealloc
*structs
,
1469 dm_oblock_t oblock
, dm_cblock_t cblock
,
1470 struct dm_bio_prison_cell
*cell
)
1472 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1475 mg
->discard
= false;
1476 mg
->writeback
= false;
1479 mg
->requeue_holder
= true;
1480 mg
->invalidate
= false;
1482 mg
->new_oblock
= oblock
;
1483 mg
->cblock
= cblock
;
1484 mg
->old_ocell
= NULL
;
1485 mg
->new_ocell
= cell
;
1486 mg
->start_jiffies
= jiffies
;
1488 inc_io_migrations(cache
);
1489 quiesce_migration(mg
);
1492 static void writeback(struct cache
*cache
, struct prealloc
*structs
,
1493 dm_oblock_t oblock
, dm_cblock_t cblock
,
1494 struct dm_bio_prison_cell
*cell
)
1496 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1499 mg
->discard
= false;
1500 mg
->writeback
= true;
1502 mg
->promote
= false;
1503 mg
->requeue_holder
= true;
1504 mg
->invalidate
= false;
1506 mg
->old_oblock
= oblock
;
1507 mg
->cblock
= cblock
;
1508 mg
->old_ocell
= cell
;
1509 mg
->new_ocell
= NULL
;
1510 mg
->start_jiffies
= jiffies
;
1512 inc_io_migrations(cache
);
1513 quiesce_migration(mg
);
1516 static void demote_then_promote(struct cache
*cache
, struct prealloc
*structs
,
1517 dm_oblock_t old_oblock
, dm_oblock_t new_oblock
,
1519 struct dm_bio_prison_cell
*old_ocell
,
1520 struct dm_bio_prison_cell
*new_ocell
)
1522 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1525 mg
->discard
= false;
1526 mg
->writeback
= false;
1529 mg
->requeue_holder
= true;
1530 mg
->invalidate
= false;
1532 mg
->old_oblock
= old_oblock
;
1533 mg
->new_oblock
= new_oblock
;
1534 mg
->cblock
= cblock
;
1535 mg
->old_ocell
= old_ocell
;
1536 mg
->new_ocell
= new_ocell
;
1537 mg
->start_jiffies
= jiffies
;
1539 inc_io_migrations(cache
);
1540 quiesce_migration(mg
);
1544 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1545 * block are thrown away.
1547 static void invalidate(struct cache
*cache
, struct prealloc
*structs
,
1548 dm_oblock_t oblock
, dm_cblock_t cblock
,
1549 struct dm_bio_prison_cell
*cell
)
1551 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1554 mg
->discard
= false;
1555 mg
->writeback
= false;
1557 mg
->promote
= false;
1558 mg
->requeue_holder
= true;
1559 mg
->invalidate
= true;
1561 mg
->old_oblock
= oblock
;
1562 mg
->cblock
= cblock
;
1563 mg
->old_ocell
= cell
;
1564 mg
->new_ocell
= NULL
;
1565 mg
->start_jiffies
= jiffies
;
1567 inc_io_migrations(cache
);
1568 quiesce_migration(mg
);
1571 static void discard(struct cache
*cache
, struct prealloc
*structs
,
1572 struct dm_bio_prison_cell
*cell
)
1574 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1578 mg
->writeback
= false;
1580 mg
->promote
= false;
1581 mg
->requeue_holder
= false;
1582 mg
->invalidate
= false;
1584 mg
->old_ocell
= NULL
;
1585 mg
->new_ocell
= cell
;
1586 mg
->start_jiffies
= jiffies
;
1588 quiesce_migration(mg
);
1591 /*----------------------------------------------------------------
1593 *--------------------------------------------------------------*/
1594 static void defer_bio(struct cache
*cache
, struct bio
*bio
)
1596 unsigned long flags
;
1598 spin_lock_irqsave(&cache
->lock
, flags
);
1599 bio_list_add(&cache
->deferred_bios
, bio
);
1600 spin_unlock_irqrestore(&cache
->lock
, flags
);
1605 static void process_flush_bio(struct cache
*cache
, struct bio
*bio
)
1607 size_t pb_data_size
= get_per_bio_data_size(cache
);
1608 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1610 BUG_ON(bio
->bi_iter
.bi_size
);
1612 remap_to_origin(cache
, bio
);
1614 remap_to_cache(cache
, bio
, 0);
1617 * REQ_FLUSH is not directed at any particular block so we don't
1618 * need to inc_ds(). REQ_FUA's are split into a write + REQ_FLUSH
1624 static void process_discard_bio(struct cache
*cache
, struct prealloc
*structs
,
1629 struct dm_bio_prison_cell
*cell_prealloc
, *new_ocell
;
1631 calc_discard_block_range(cache
, bio
, &b
, &e
);
1637 cell_prealloc
= prealloc_get_cell(structs
);
1638 r
= bio_detain_range(cache
, dblock_to_oblock(cache
, b
), dblock_to_oblock(cache
, e
), bio
, cell_prealloc
,
1639 (cell_free_fn
) prealloc_put_cell
,
1640 structs
, &new_ocell
);
1644 discard(cache
, structs
, new_ocell
);
1647 static bool spare_migration_bandwidth(struct cache
*cache
)
1649 sector_t current_volume
= (atomic_read(&cache
->nr_io_migrations
) + 1) *
1650 cache
->sectors_per_block
;
1651 return current_volume
< cache
->migration_threshold
;
1654 static void inc_hit_counter(struct cache
*cache
, struct bio
*bio
)
1656 atomic_inc(bio_data_dir(bio
) == READ
?
1657 &cache
->stats
.read_hit
: &cache
->stats
.write_hit
);
1660 static void inc_miss_counter(struct cache
*cache
, struct bio
*bio
)
1662 atomic_inc(bio_data_dir(bio
) == READ
?
1663 &cache
->stats
.read_miss
: &cache
->stats
.write_miss
);
1666 /*----------------------------------------------------------------*/
1669 struct cache
*cache
;
1670 struct bio_list bios_for_issue
;
1671 struct bio_list unhandled_bios
;
1675 static void inc_fn(void *context
, struct dm_bio_prison_cell
*cell
)
1678 struct inc_detail
*detail
= context
;
1679 struct cache
*cache
= detail
->cache
;
1681 inc_ds(cache
, cell
->holder
, cell
);
1682 if (bio_data_dir(cell
->holder
) == WRITE
)
1683 detail
->any_writes
= true;
1685 while ((bio
= bio_list_pop(&cell
->bios
))) {
1686 if (discard_or_flush(bio
)) {
1687 bio_list_add(&detail
->unhandled_bios
, bio
);
1691 if (bio_data_dir(bio
) == WRITE
)
1692 detail
->any_writes
= true;
1694 bio_list_add(&detail
->bios_for_issue
, bio
);
1695 inc_ds(cache
, bio
, cell
);
1699 // FIXME: refactor these two
1700 static void remap_cell_to_origin_clear_discard(struct cache
*cache
,
1701 struct dm_bio_prison_cell
*cell
,
1702 dm_oblock_t oblock
, bool issue_holder
)
1705 unsigned long flags
;
1706 struct inc_detail detail
;
1708 detail
.cache
= cache
;
1709 bio_list_init(&detail
.bios_for_issue
);
1710 bio_list_init(&detail
.unhandled_bios
);
1711 detail
.any_writes
= false;
1713 spin_lock_irqsave(&cache
->lock
, flags
);
1714 dm_cell_visit_release(cache
->prison
, inc_fn
, &detail
, cell
);
1715 bio_list_merge(&cache
->deferred_bios
, &detail
.unhandled_bios
);
1716 spin_unlock_irqrestore(&cache
->lock
, flags
);
1718 remap_to_origin(cache
, cell
->holder
);
1720 issue(cache
, cell
->holder
);
1722 accounted_begin(cache
, cell
->holder
);
1724 if (detail
.any_writes
)
1725 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
1727 while ((bio
= bio_list_pop(&detail
.bios_for_issue
))) {
1728 remap_to_origin(cache
, bio
);
1733 static void remap_cell_to_cache_dirty(struct cache
*cache
, struct dm_bio_prison_cell
*cell
,
1734 dm_oblock_t oblock
, dm_cblock_t cblock
, bool issue_holder
)
1737 unsigned long flags
;
1738 struct inc_detail detail
;
1740 detail
.cache
= cache
;
1741 bio_list_init(&detail
.bios_for_issue
);
1742 bio_list_init(&detail
.unhandled_bios
);
1743 detail
.any_writes
= false;
1745 spin_lock_irqsave(&cache
->lock
, flags
);
1746 dm_cell_visit_release(cache
->prison
, inc_fn
, &detail
, cell
);
1747 bio_list_merge(&cache
->deferred_bios
, &detail
.unhandled_bios
);
1748 spin_unlock_irqrestore(&cache
->lock
, flags
);
1750 remap_to_cache(cache
, cell
->holder
, cblock
);
1752 issue(cache
, cell
->holder
);
1754 accounted_begin(cache
, cell
->holder
);
1756 if (detail
.any_writes
) {
1757 set_dirty(cache
, oblock
, cblock
);
1758 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
1761 while ((bio
= bio_list_pop(&detail
.bios_for_issue
))) {
1762 remap_to_cache(cache
, bio
, cblock
);
1767 /*----------------------------------------------------------------*/
1769 struct old_oblock_lock
{
1770 struct policy_locker locker
;
1771 struct cache
*cache
;
1772 struct prealloc
*structs
;
1773 struct dm_bio_prison_cell
*cell
;
1776 static int null_locker(struct policy_locker
*locker
, dm_oblock_t b
)
1778 /* This should never be called */
1783 static int cell_locker(struct policy_locker
*locker
, dm_oblock_t b
)
1785 struct old_oblock_lock
*l
= container_of(locker
, struct old_oblock_lock
, locker
);
1786 struct dm_bio_prison_cell
*cell_prealloc
= prealloc_get_cell(l
->structs
);
1788 return bio_detain(l
->cache
, b
, NULL
, cell_prealloc
,
1789 (cell_free_fn
) prealloc_put_cell
,
1790 l
->structs
, &l
->cell
);
1793 static void process_cell(struct cache
*cache
, struct prealloc
*structs
,
1794 struct dm_bio_prison_cell
*new_ocell
)
1797 bool release_cell
= true;
1798 struct bio
*bio
= new_ocell
->holder
;
1799 dm_oblock_t block
= get_bio_block(cache
, bio
);
1800 struct policy_result lookup_result
;
1801 bool passthrough
= passthrough_mode(&cache
->features
);
1802 bool fast_promotion
, can_migrate
;
1803 struct old_oblock_lock ool
;
1805 fast_promotion
= is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
);
1806 can_migrate
= !passthrough
&& (fast_promotion
|| spare_migration_bandwidth(cache
));
1808 ool
.locker
.fn
= cell_locker
;
1810 ool
.structs
= structs
;
1812 r
= policy_map(cache
->policy
, block
, true, can_migrate
, fast_promotion
,
1813 bio
, &ool
.locker
, &lookup_result
);
1815 if (r
== -EWOULDBLOCK
)
1816 /* migration has been denied */
1817 lookup_result
.op
= POLICY_MISS
;
1819 switch (lookup_result
.op
) {
1822 inc_miss_counter(cache
, bio
);
1825 * Passthrough always maps to the origin,
1826 * invalidating any cache blocks that are written
1830 if (bio_data_dir(bio
) == WRITE
) {
1831 atomic_inc(&cache
->stats
.demotion
);
1832 invalidate(cache
, structs
, block
, lookup_result
.cblock
, new_ocell
);
1833 release_cell
= false;
1836 /* FIXME: factor out issue_origin() */
1837 remap_to_origin_clear_discard(cache
, bio
, block
);
1838 inc_and_issue(cache
, bio
, new_ocell
);
1841 inc_hit_counter(cache
, bio
);
1843 if (bio_data_dir(bio
) == WRITE
&&
1844 writethrough_mode(&cache
->features
) &&
1845 !is_dirty(cache
, lookup_result
.cblock
)) {
1846 remap_to_origin_then_cache(cache
, bio
, block
, lookup_result
.cblock
);
1847 inc_and_issue(cache
, bio
, new_ocell
);
1850 remap_cell_to_cache_dirty(cache
, new_ocell
, block
, lookup_result
.cblock
, true);
1851 release_cell
= false;
1858 inc_miss_counter(cache
, bio
);
1859 remap_cell_to_origin_clear_discard(cache
, new_ocell
, block
, true);
1860 release_cell
= false;
1864 atomic_inc(&cache
->stats
.promotion
);
1865 promote(cache
, structs
, block
, lookup_result
.cblock
, new_ocell
);
1866 release_cell
= false;
1869 case POLICY_REPLACE
:
1870 atomic_inc(&cache
->stats
.demotion
);
1871 atomic_inc(&cache
->stats
.promotion
);
1872 demote_then_promote(cache
, structs
, lookup_result
.old_oblock
,
1873 block
, lookup_result
.cblock
,
1874 ool
.cell
, new_ocell
);
1875 release_cell
= false;
1879 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1880 cache_device_name(cache
), __func__
,
1881 (unsigned) lookup_result
.op
);
1886 cell_defer(cache
, new_ocell
, false);
1889 static void process_bio(struct cache
*cache
, struct prealloc
*structs
,
1893 dm_oblock_t block
= get_bio_block(cache
, bio
);
1894 struct dm_bio_prison_cell
*cell_prealloc
, *new_ocell
;
1897 * Check to see if that block is currently migrating.
1899 cell_prealloc
= prealloc_get_cell(structs
);
1900 r
= bio_detain(cache
, block
, bio
, cell_prealloc
,
1901 (cell_free_fn
) prealloc_put_cell
,
1902 structs
, &new_ocell
);
1906 process_cell(cache
, structs
, new_ocell
);
1909 static int need_commit_due_to_time(struct cache
*cache
)
1911 return jiffies
< cache
->last_commit_jiffies
||
1912 jiffies
> cache
->last_commit_jiffies
+ COMMIT_PERIOD
;
1916 * A non-zero return indicates read_only or fail_io mode.
1918 static int commit(struct cache
*cache
, bool clean_shutdown
)
1922 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1925 atomic_inc(&cache
->stats
.commit_count
);
1926 r
= dm_cache_commit(cache
->cmd
, clean_shutdown
);
1928 metadata_operation_failed(cache
, "dm_cache_commit", r
);
1933 static int commit_if_needed(struct cache
*cache
)
1937 if ((cache
->commit_requested
|| need_commit_due_to_time(cache
)) &&
1938 dm_cache_changed_this_transaction(cache
->cmd
)) {
1939 r
= commit(cache
, false);
1940 cache
->commit_requested
= false;
1941 cache
->last_commit_jiffies
= jiffies
;
1947 static void process_deferred_bios(struct cache
*cache
)
1949 bool prealloc_used
= false;
1950 unsigned long flags
;
1951 struct bio_list bios
;
1953 struct prealloc structs
;
1955 memset(&structs
, 0, sizeof(structs
));
1956 bio_list_init(&bios
);
1958 spin_lock_irqsave(&cache
->lock
, flags
);
1959 bio_list_merge(&bios
, &cache
->deferred_bios
);
1960 bio_list_init(&cache
->deferred_bios
);
1961 spin_unlock_irqrestore(&cache
->lock
, flags
);
1963 while (!bio_list_empty(&bios
)) {
1965 * If we've got no free migration structs, and processing
1966 * this bio might require one, we pause until there are some
1967 * prepared mappings to process.
1969 if (prealloc_data_structs(cache
, &structs
)) {
1970 spin_lock_irqsave(&cache
->lock
, flags
);
1971 bio_list_merge(&cache
->deferred_bios
, &bios
);
1972 spin_unlock_irqrestore(&cache
->lock
, flags
);
1976 bio
= bio_list_pop(&bios
);
1978 if (bio
->bi_rw
& REQ_FLUSH
)
1979 process_flush_bio(cache
, bio
);
1980 else if (bio
->bi_rw
& REQ_DISCARD
)
1981 process_discard_bio(cache
, &structs
, bio
);
1983 process_bio(cache
, &structs
, bio
);
1984 prealloc_used
= true;
1988 prealloc_free_structs(cache
, &structs
);
1991 static void process_deferred_cells(struct cache
*cache
)
1993 bool prealloc_used
= false;
1994 unsigned long flags
;
1995 struct dm_bio_prison_cell
*cell
, *tmp
;
1996 struct list_head cells
;
1997 struct prealloc structs
;
1999 memset(&structs
, 0, sizeof(structs
));
2001 INIT_LIST_HEAD(&cells
);
2003 spin_lock_irqsave(&cache
->lock
, flags
);
2004 list_splice_init(&cache
->deferred_cells
, &cells
);
2005 spin_unlock_irqrestore(&cache
->lock
, flags
);
2007 list_for_each_entry_safe(cell
, tmp
, &cells
, user_list
) {
2009 * If we've got no free migration structs, and processing
2010 * this bio might require one, we pause until there are some
2011 * prepared mappings to process.
2013 if (prealloc_data_structs(cache
, &structs
)) {
2014 spin_lock_irqsave(&cache
->lock
, flags
);
2015 list_splice(&cells
, &cache
->deferred_cells
);
2016 spin_unlock_irqrestore(&cache
->lock
, flags
);
2020 process_cell(cache
, &structs
, cell
);
2021 prealloc_used
= true;
2025 prealloc_free_structs(cache
, &structs
);
2028 static void process_deferred_flush_bios(struct cache
*cache
, bool submit_bios
)
2030 unsigned long flags
;
2031 struct bio_list bios
;
2034 bio_list_init(&bios
);
2036 spin_lock_irqsave(&cache
->lock
, flags
);
2037 bio_list_merge(&bios
, &cache
->deferred_flush_bios
);
2038 bio_list_init(&cache
->deferred_flush_bios
);
2039 spin_unlock_irqrestore(&cache
->lock
, flags
);
2042 * These bios have already been through inc_ds()
2044 while ((bio
= bio_list_pop(&bios
)))
2045 submit_bios
? accounted_request(cache
, bio
) : bio_io_error(bio
);
2048 static void process_deferred_writethrough_bios(struct cache
*cache
)
2050 unsigned long flags
;
2051 struct bio_list bios
;
2054 bio_list_init(&bios
);
2056 spin_lock_irqsave(&cache
->lock
, flags
);
2057 bio_list_merge(&bios
, &cache
->deferred_writethrough_bios
);
2058 bio_list_init(&cache
->deferred_writethrough_bios
);
2059 spin_unlock_irqrestore(&cache
->lock
, flags
);
2062 * These bios have already been through inc_ds()
2064 while ((bio
= bio_list_pop(&bios
)))
2065 accounted_request(cache
, bio
);
2068 static void writeback_some_dirty_blocks(struct cache
*cache
)
2070 bool prealloc_used
= false;
2073 struct prealloc structs
;
2074 struct dm_bio_prison_cell
*old_ocell
;
2075 bool busy
= !iot_idle_for(&cache
->origin_tracker
, HZ
);
2077 memset(&structs
, 0, sizeof(structs
));
2079 while (spare_migration_bandwidth(cache
)) {
2080 if (policy_writeback_work(cache
->policy
, &oblock
, &cblock
, busy
))
2081 break; /* no work to do */
2083 if (prealloc_data_structs(cache
, &structs
) ||
2084 get_cell(cache
, oblock
, &structs
, &old_ocell
)) {
2085 policy_set_dirty(cache
->policy
, oblock
);
2089 writeback(cache
, &structs
, oblock
, cblock
, old_ocell
);
2090 prealloc_used
= true;
2094 prealloc_free_structs(cache
, &structs
);
2097 /*----------------------------------------------------------------
2099 * Dropping something from the cache *without* writing back.
2100 *--------------------------------------------------------------*/
2102 static void process_invalidation_request(struct cache
*cache
, struct invalidation_request
*req
)
2105 uint64_t begin
= from_cblock(req
->cblocks
->begin
);
2106 uint64_t end
= from_cblock(req
->cblocks
->end
);
2108 while (begin
!= end
) {
2109 r
= policy_remove_cblock(cache
->policy
, to_cblock(begin
));
2111 r
= dm_cache_remove_mapping(cache
->cmd
, to_cblock(begin
));
2113 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
2117 } else if (r
== -ENODATA
) {
2118 /* harmless, already unmapped */
2122 DMERR("%s: policy_remove_cblock failed", cache_device_name(cache
));
2129 cache
->commit_requested
= true;
2132 atomic_set(&req
->complete
, 1);
2134 wake_up(&req
->result_wait
);
2137 static void process_invalidation_requests(struct cache
*cache
)
2139 struct list_head list
;
2140 struct invalidation_request
*req
, *tmp
;
2142 INIT_LIST_HEAD(&list
);
2143 spin_lock(&cache
->invalidation_lock
);
2144 list_splice_init(&cache
->invalidation_requests
, &list
);
2145 spin_unlock(&cache
->invalidation_lock
);
2147 list_for_each_entry_safe (req
, tmp
, &list
, list
)
2148 process_invalidation_request(cache
, req
);
2151 /*----------------------------------------------------------------
2153 *--------------------------------------------------------------*/
2154 static bool is_quiescing(struct cache
*cache
)
2156 return atomic_read(&cache
->quiescing
);
2159 static void ack_quiescing(struct cache
*cache
)
2161 if (is_quiescing(cache
)) {
2162 atomic_inc(&cache
->quiescing_ack
);
2163 wake_up(&cache
->quiescing_wait
);
2167 static void wait_for_quiescing_ack(struct cache
*cache
)
2169 wait_event(cache
->quiescing_wait
, atomic_read(&cache
->quiescing_ack
));
2172 static void start_quiescing(struct cache
*cache
)
2174 atomic_inc(&cache
->quiescing
);
2175 wait_for_quiescing_ack(cache
);
2178 static void stop_quiescing(struct cache
*cache
)
2180 atomic_set(&cache
->quiescing
, 0);
2181 atomic_set(&cache
->quiescing_ack
, 0);
2184 static void wait_for_migrations(struct cache
*cache
)
2186 wait_event(cache
->migration_wait
, !atomic_read(&cache
->nr_allocated_migrations
));
2189 static void stop_worker(struct cache
*cache
)
2191 cancel_delayed_work(&cache
->waker
);
2192 flush_workqueue(cache
->wq
);
2195 static void requeue_deferred_cells(struct cache
*cache
)
2197 unsigned long flags
;
2198 struct list_head cells
;
2199 struct dm_bio_prison_cell
*cell
, *tmp
;
2201 INIT_LIST_HEAD(&cells
);
2202 spin_lock_irqsave(&cache
->lock
, flags
);
2203 list_splice_init(&cache
->deferred_cells
, &cells
);
2204 spin_unlock_irqrestore(&cache
->lock
, flags
);
2206 list_for_each_entry_safe(cell
, tmp
, &cells
, user_list
)
2207 cell_requeue(cache
, cell
);
2210 static void requeue_deferred_bios(struct cache
*cache
)
2213 struct bio_list bios
;
2215 bio_list_init(&bios
);
2216 bio_list_merge(&bios
, &cache
->deferred_bios
);
2217 bio_list_init(&cache
->deferred_bios
);
2219 while ((bio
= bio_list_pop(&bios
)))
2220 bio_endio(bio
, DM_ENDIO_REQUEUE
);
2223 static int more_work(struct cache
*cache
)
2225 if (is_quiescing(cache
))
2226 return !list_empty(&cache
->quiesced_migrations
) ||
2227 !list_empty(&cache
->completed_migrations
) ||
2228 !list_empty(&cache
->need_commit_migrations
);
2230 return !bio_list_empty(&cache
->deferred_bios
) ||
2231 !list_empty(&cache
->deferred_cells
) ||
2232 !bio_list_empty(&cache
->deferred_flush_bios
) ||
2233 !bio_list_empty(&cache
->deferred_writethrough_bios
) ||
2234 !list_empty(&cache
->quiesced_migrations
) ||
2235 !list_empty(&cache
->completed_migrations
) ||
2236 !list_empty(&cache
->need_commit_migrations
) ||
2240 static void do_worker(struct work_struct
*ws
)
2242 struct cache
*cache
= container_of(ws
, struct cache
, worker
);
2245 if (!is_quiescing(cache
)) {
2246 writeback_some_dirty_blocks(cache
);
2247 process_deferred_writethrough_bios(cache
);
2248 process_deferred_bios(cache
);
2249 process_deferred_cells(cache
);
2250 process_invalidation_requests(cache
);
2253 process_migrations(cache
, &cache
->quiesced_migrations
, issue_copy_or_discard
);
2254 process_migrations(cache
, &cache
->completed_migrations
, complete_migration
);
2256 if (commit_if_needed(cache
)) {
2257 process_deferred_flush_bios(cache
, false);
2258 process_migrations(cache
, &cache
->need_commit_migrations
, migration_failure
);
2260 process_deferred_flush_bios(cache
, true);
2261 process_migrations(cache
, &cache
->need_commit_migrations
,
2262 migration_success_post_commit
);
2265 ack_quiescing(cache
);
2267 } while (more_work(cache
));
2271 * We want to commit periodically so that not too much
2272 * unwritten metadata builds up.
2274 static void do_waker(struct work_struct
*ws
)
2276 struct cache
*cache
= container_of(to_delayed_work(ws
), struct cache
, waker
);
2277 policy_tick(cache
->policy
, true);
2279 queue_delayed_work(cache
->wq
, &cache
->waker
, COMMIT_PERIOD
);
2282 /*----------------------------------------------------------------*/
2284 static int is_congested(struct dm_dev
*dev
, int bdi_bits
)
2286 struct request_queue
*q
= bdev_get_queue(dev
->bdev
);
2287 return bdi_congested(&q
->backing_dev_info
, bdi_bits
);
2290 static int cache_is_congested(struct dm_target_callbacks
*cb
, int bdi_bits
)
2292 struct cache
*cache
= container_of(cb
, struct cache
, callbacks
);
2294 return is_congested(cache
->origin_dev
, bdi_bits
) ||
2295 is_congested(cache
->cache_dev
, bdi_bits
);
2298 /*----------------------------------------------------------------
2300 *--------------------------------------------------------------*/
2303 * This function gets called on the error paths of the constructor, so we
2304 * have to cope with a partially initialised struct.
2306 static void destroy(struct cache
*cache
)
2310 if (cache
->migration_pool
)
2311 mempool_destroy(cache
->migration_pool
);
2313 if (cache
->all_io_ds
)
2314 dm_deferred_set_destroy(cache
->all_io_ds
);
2317 dm_bio_prison_destroy(cache
->prison
);
2320 destroy_workqueue(cache
->wq
);
2322 if (cache
->dirty_bitset
)
2323 free_bitset(cache
->dirty_bitset
);
2325 if (cache
->discard_bitset
)
2326 free_bitset(cache
->discard_bitset
);
2329 dm_kcopyd_client_destroy(cache
->copier
);
2332 dm_cache_metadata_close(cache
->cmd
);
2334 if (cache
->metadata_dev
)
2335 dm_put_device(cache
->ti
, cache
->metadata_dev
);
2337 if (cache
->origin_dev
)
2338 dm_put_device(cache
->ti
, cache
->origin_dev
);
2340 if (cache
->cache_dev
)
2341 dm_put_device(cache
->ti
, cache
->cache_dev
);
2344 dm_cache_policy_destroy(cache
->policy
);
2346 for (i
= 0; i
< cache
->nr_ctr_args
; i
++)
2347 kfree(cache
->ctr_args
[i
]);
2348 kfree(cache
->ctr_args
);
2353 static void cache_dtr(struct dm_target
*ti
)
2355 struct cache
*cache
= ti
->private;
2360 static sector_t
get_dev_size(struct dm_dev
*dev
)
2362 return i_size_read(dev
->bdev
->bd_inode
) >> SECTOR_SHIFT
;
2365 /*----------------------------------------------------------------*/
2368 * Construct a cache device mapping.
2370 * cache <metadata dev> <cache dev> <origin dev> <block size>
2371 * <#feature args> [<feature arg>]*
2372 * <policy> <#policy args> [<policy arg>]*
2374 * metadata dev : fast device holding the persistent metadata
2375 * cache dev : fast device holding cached data blocks
2376 * origin dev : slow device holding original data blocks
2377 * block size : cache unit size in sectors
2379 * #feature args : number of feature arguments passed
2380 * feature args : writethrough. (The default is writeback.)
2382 * policy : the replacement policy to use
2383 * #policy args : an even number of policy arguments corresponding
2384 * to key/value pairs passed to the policy
2385 * policy args : key/value pairs passed to the policy
2386 * E.g. 'sequential_threshold 1024'
2387 * See cache-policies.txt for details.
2389 * Optional feature arguments are:
2390 * writethrough : write through caching that prohibits cache block
2391 * content from being different from origin block content.
2392 * Without this argument, the default behaviour is to write
2393 * back cache block contents later for performance reasons,
2394 * so they may differ from the corresponding origin blocks.
2397 struct dm_target
*ti
;
2399 struct dm_dev
*metadata_dev
;
2401 struct dm_dev
*cache_dev
;
2402 sector_t cache_sectors
;
2404 struct dm_dev
*origin_dev
;
2405 sector_t origin_sectors
;
2407 uint32_t block_size
;
2409 const char *policy_name
;
2411 const char **policy_argv
;
2413 struct cache_features features
;
2416 static void destroy_cache_args(struct cache_args
*ca
)
2418 if (ca
->metadata_dev
)
2419 dm_put_device(ca
->ti
, ca
->metadata_dev
);
2422 dm_put_device(ca
->ti
, ca
->cache_dev
);
2425 dm_put_device(ca
->ti
, ca
->origin_dev
);
2430 static bool at_least_one_arg(struct dm_arg_set
*as
, char **error
)
2433 *error
= "Insufficient args";
2440 static int parse_metadata_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2444 sector_t metadata_dev_size
;
2445 char b
[BDEVNAME_SIZE
];
2447 if (!at_least_one_arg(as
, error
))
2450 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2453 *error
= "Error opening metadata device";
2457 metadata_dev_size
= get_dev_size(ca
->metadata_dev
);
2458 if (metadata_dev_size
> DM_CACHE_METADATA_MAX_SECTORS_WARNING
)
2459 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2460 bdevname(ca
->metadata_dev
->bdev
, b
), THIN_METADATA_MAX_SECTORS
);
2465 static int parse_cache_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2470 if (!at_least_one_arg(as
, error
))
2473 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2476 *error
= "Error opening cache device";
2479 ca
->cache_sectors
= get_dev_size(ca
->cache_dev
);
2484 static int parse_origin_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2489 if (!at_least_one_arg(as
, error
))
2492 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2495 *error
= "Error opening origin device";
2499 ca
->origin_sectors
= get_dev_size(ca
->origin_dev
);
2500 if (ca
->ti
->len
> ca
->origin_sectors
) {
2501 *error
= "Device size larger than cached device";
2508 static int parse_block_size(struct cache_args
*ca
, struct dm_arg_set
*as
,
2511 unsigned long block_size
;
2513 if (!at_least_one_arg(as
, error
))
2516 if (kstrtoul(dm_shift_arg(as
), 10, &block_size
) || !block_size
||
2517 block_size
< DATA_DEV_BLOCK_SIZE_MIN_SECTORS
||
2518 block_size
> DATA_DEV_BLOCK_SIZE_MAX_SECTORS
||
2519 block_size
& (DATA_DEV_BLOCK_SIZE_MIN_SECTORS
- 1)) {
2520 *error
= "Invalid data block size";
2524 if (block_size
> ca
->cache_sectors
) {
2525 *error
= "Data block size is larger than the cache device";
2529 ca
->block_size
= block_size
;
2534 static void init_features(struct cache_features
*cf
)
2536 cf
->mode
= CM_WRITE
;
2537 cf
->io_mode
= CM_IO_WRITEBACK
;
2540 static int parse_features(struct cache_args
*ca
, struct dm_arg_set
*as
,
2543 static struct dm_arg _args
[] = {
2544 {0, 1, "Invalid number of cache feature arguments"},
2550 struct cache_features
*cf
= &ca
->features
;
2554 r
= dm_read_arg_group(_args
, as
, &argc
, error
);
2559 arg
= dm_shift_arg(as
);
2561 if (!strcasecmp(arg
, "writeback"))
2562 cf
->io_mode
= CM_IO_WRITEBACK
;
2564 else if (!strcasecmp(arg
, "writethrough"))
2565 cf
->io_mode
= CM_IO_WRITETHROUGH
;
2567 else if (!strcasecmp(arg
, "passthrough"))
2568 cf
->io_mode
= CM_IO_PASSTHROUGH
;
2571 *error
= "Unrecognised cache feature requested";
2579 static int parse_policy(struct cache_args
*ca
, struct dm_arg_set
*as
,
2582 static struct dm_arg _args
[] = {
2583 {0, 1024, "Invalid number of policy arguments"},
2588 if (!at_least_one_arg(as
, error
))
2591 ca
->policy_name
= dm_shift_arg(as
);
2593 r
= dm_read_arg_group(_args
, as
, &ca
->policy_argc
, error
);
2597 ca
->policy_argv
= (const char **)as
->argv
;
2598 dm_consume_args(as
, ca
->policy_argc
);
2603 static int parse_cache_args(struct cache_args
*ca
, int argc
, char **argv
,
2607 struct dm_arg_set as
;
2612 r
= parse_metadata_dev(ca
, &as
, error
);
2616 r
= parse_cache_dev(ca
, &as
, error
);
2620 r
= parse_origin_dev(ca
, &as
, error
);
2624 r
= parse_block_size(ca
, &as
, error
);
2628 r
= parse_features(ca
, &as
, error
);
2632 r
= parse_policy(ca
, &as
, error
);
2639 /*----------------------------------------------------------------*/
2641 static struct kmem_cache
*migration_cache
;
2643 #define NOT_CORE_OPTION 1
2645 static int process_config_option(struct cache
*cache
, const char *key
, const char *value
)
2649 if (!strcasecmp(key
, "migration_threshold")) {
2650 if (kstrtoul(value
, 10, &tmp
))
2653 cache
->migration_threshold
= tmp
;
2657 return NOT_CORE_OPTION
;
2660 static int set_config_value(struct cache
*cache
, const char *key
, const char *value
)
2662 int r
= process_config_option(cache
, key
, value
);
2664 if (r
== NOT_CORE_OPTION
)
2665 r
= policy_set_config_value(cache
->policy
, key
, value
);
2668 DMWARN("bad config value for %s: %s", key
, value
);
2673 static int set_config_values(struct cache
*cache
, int argc
, const char **argv
)
2678 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2683 r
= set_config_value(cache
, argv
[0], argv
[1]);
2694 static int create_cache_policy(struct cache
*cache
, struct cache_args
*ca
,
2697 struct dm_cache_policy
*p
= dm_cache_policy_create(ca
->policy_name
,
2699 cache
->origin_sectors
,
2700 cache
->sectors_per_block
);
2702 *error
= "Error creating cache's policy";
2711 * We want the discard block size to be at least the size of the cache
2712 * block size and have no more than 2^14 discard blocks across the origin.
2714 #define MAX_DISCARD_BLOCKS (1 << 14)
2716 static bool too_many_discard_blocks(sector_t discard_block_size
,
2717 sector_t origin_size
)
2719 (void) sector_div(origin_size
, discard_block_size
);
2721 return origin_size
> MAX_DISCARD_BLOCKS
;
2724 static sector_t
calculate_discard_block_size(sector_t cache_block_size
,
2725 sector_t origin_size
)
2727 sector_t discard_block_size
= cache_block_size
;
2730 while (too_many_discard_blocks(discard_block_size
, origin_size
))
2731 discard_block_size
*= 2;
2733 return discard_block_size
;
2736 static void set_cache_size(struct cache
*cache
, dm_cblock_t size
)
2738 dm_block_t nr_blocks
= from_cblock(size
);
2740 if (nr_blocks
> (1 << 20) && cache
->cache_size
!= size
)
2741 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2742 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2743 "Please consider increasing the cache block size to reduce the overall cache block count.",
2744 (unsigned long long) nr_blocks
);
2746 cache
->cache_size
= size
;
2749 #define DEFAULT_MIGRATION_THRESHOLD 2048
2751 static int cache_create(struct cache_args
*ca
, struct cache
**result
)
2754 char **error
= &ca
->ti
->error
;
2755 struct cache
*cache
;
2756 struct dm_target
*ti
= ca
->ti
;
2757 dm_block_t origin_blocks
;
2758 struct dm_cache_metadata
*cmd
;
2759 bool may_format
= ca
->features
.mode
== CM_WRITE
;
2761 cache
= kzalloc(sizeof(*cache
), GFP_KERNEL
);
2766 ti
->private = cache
;
2767 ti
->num_flush_bios
= 2;
2768 ti
->flush_supported
= true;
2770 ti
->num_discard_bios
= 1;
2771 ti
->discards_supported
= true;
2772 ti
->discard_zeroes_data_unsupported
= true;
2773 ti
->split_discard_bios
= false;
2775 cache
->features
= ca
->features
;
2776 ti
->per_bio_data_size
= get_per_bio_data_size(cache
);
2778 cache
->callbacks
.congested_fn
= cache_is_congested
;
2779 dm_table_add_target_callbacks(ti
->table
, &cache
->callbacks
);
2781 cache
->metadata_dev
= ca
->metadata_dev
;
2782 cache
->origin_dev
= ca
->origin_dev
;
2783 cache
->cache_dev
= ca
->cache_dev
;
2785 ca
->metadata_dev
= ca
->origin_dev
= ca
->cache_dev
= NULL
;
2787 /* FIXME: factor out this whole section */
2788 origin_blocks
= cache
->origin_sectors
= ca
->origin_sectors
;
2789 origin_blocks
= block_div(origin_blocks
, ca
->block_size
);
2790 cache
->origin_blocks
= to_oblock(origin_blocks
);
2792 cache
->sectors_per_block
= ca
->block_size
;
2793 if (dm_set_target_max_io_len(ti
, cache
->sectors_per_block
)) {
2798 if (ca
->block_size
& (ca
->block_size
- 1)) {
2799 dm_block_t cache_size
= ca
->cache_sectors
;
2801 cache
->sectors_per_block_shift
= -1;
2802 cache_size
= block_div(cache_size
, ca
->block_size
);
2803 set_cache_size(cache
, to_cblock(cache_size
));
2805 cache
->sectors_per_block_shift
= __ffs(ca
->block_size
);
2806 set_cache_size(cache
, to_cblock(ca
->cache_sectors
>> cache
->sectors_per_block_shift
));
2809 r
= create_cache_policy(cache
, ca
, error
);
2813 cache
->policy_nr_args
= ca
->policy_argc
;
2814 cache
->migration_threshold
= DEFAULT_MIGRATION_THRESHOLD
;
2816 r
= set_config_values(cache
, ca
->policy_argc
, ca
->policy_argv
);
2818 *error
= "Error setting cache policy's config values";
2822 cmd
= dm_cache_metadata_open(cache
->metadata_dev
->bdev
,
2823 ca
->block_size
, may_format
,
2824 dm_cache_policy_get_hint_size(cache
->policy
));
2826 *error
= "Error creating metadata object";
2831 set_cache_mode(cache
, CM_WRITE
);
2832 if (get_cache_mode(cache
) != CM_WRITE
) {
2833 *error
= "Unable to get write access to metadata, please check/repair metadata.";
2838 if (passthrough_mode(&cache
->features
)) {
2841 r
= dm_cache_metadata_all_clean(cache
->cmd
, &all_clean
);
2843 *error
= "dm_cache_metadata_all_clean() failed";
2848 *error
= "Cannot enter passthrough mode unless all blocks are clean";
2854 spin_lock_init(&cache
->lock
);
2855 INIT_LIST_HEAD(&cache
->deferred_cells
);
2856 bio_list_init(&cache
->deferred_bios
);
2857 bio_list_init(&cache
->deferred_flush_bios
);
2858 bio_list_init(&cache
->deferred_writethrough_bios
);
2859 INIT_LIST_HEAD(&cache
->quiesced_migrations
);
2860 INIT_LIST_HEAD(&cache
->completed_migrations
);
2861 INIT_LIST_HEAD(&cache
->need_commit_migrations
);
2862 atomic_set(&cache
->nr_allocated_migrations
, 0);
2863 atomic_set(&cache
->nr_io_migrations
, 0);
2864 init_waitqueue_head(&cache
->migration_wait
);
2866 init_waitqueue_head(&cache
->quiescing_wait
);
2867 atomic_set(&cache
->quiescing
, 0);
2868 atomic_set(&cache
->quiescing_ack
, 0);
2871 atomic_set(&cache
->nr_dirty
, 0);
2872 cache
->dirty_bitset
= alloc_bitset(from_cblock(cache
->cache_size
));
2873 if (!cache
->dirty_bitset
) {
2874 *error
= "could not allocate dirty bitset";
2877 clear_bitset(cache
->dirty_bitset
, from_cblock(cache
->cache_size
));
2879 cache
->discard_block_size
=
2880 calculate_discard_block_size(cache
->sectors_per_block
,
2881 cache
->origin_sectors
);
2882 cache
->discard_nr_blocks
= to_dblock(dm_sector_div_up(cache
->origin_sectors
,
2883 cache
->discard_block_size
));
2884 cache
->discard_bitset
= alloc_bitset(from_dblock(cache
->discard_nr_blocks
));
2885 if (!cache
->discard_bitset
) {
2886 *error
= "could not allocate discard bitset";
2889 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
2891 cache
->copier
= dm_kcopyd_client_create(&dm_kcopyd_throttle
);
2892 if (IS_ERR(cache
->copier
)) {
2893 *error
= "could not create kcopyd client";
2894 r
= PTR_ERR(cache
->copier
);
2898 cache
->wq
= alloc_ordered_workqueue("dm-" DM_MSG_PREFIX
, WQ_MEM_RECLAIM
);
2900 *error
= "could not create workqueue for metadata object";
2903 INIT_WORK(&cache
->worker
, do_worker
);
2904 INIT_DELAYED_WORK(&cache
->waker
, do_waker
);
2905 cache
->last_commit_jiffies
= jiffies
;
2907 cache
->prison
= dm_bio_prison_create();
2908 if (!cache
->prison
) {
2909 *error
= "could not create bio prison";
2913 cache
->all_io_ds
= dm_deferred_set_create();
2914 if (!cache
->all_io_ds
) {
2915 *error
= "could not create all_io deferred set";
2919 cache
->migration_pool
= mempool_create_slab_pool(MIGRATION_POOL_SIZE
,
2921 if (!cache
->migration_pool
) {
2922 *error
= "Error creating cache's migration mempool";
2926 cache
->need_tick_bio
= true;
2927 cache
->sized
= false;
2928 cache
->invalidate
= false;
2929 cache
->commit_requested
= false;
2930 cache
->loaded_mappings
= false;
2931 cache
->loaded_discards
= false;
2935 atomic_set(&cache
->stats
.demotion
, 0);
2936 atomic_set(&cache
->stats
.promotion
, 0);
2937 atomic_set(&cache
->stats
.copies_avoided
, 0);
2938 atomic_set(&cache
->stats
.cache_cell_clash
, 0);
2939 atomic_set(&cache
->stats
.commit_count
, 0);
2940 atomic_set(&cache
->stats
.discard_count
, 0);
2942 spin_lock_init(&cache
->invalidation_lock
);
2943 INIT_LIST_HEAD(&cache
->invalidation_requests
);
2945 iot_init(&cache
->origin_tracker
);
2955 static int copy_ctr_args(struct cache
*cache
, int argc
, const char **argv
)
2960 copy
= kcalloc(argc
, sizeof(*copy
), GFP_KERNEL
);
2963 for (i
= 0; i
< argc
; i
++) {
2964 copy
[i
] = kstrdup(argv
[i
], GFP_KERNEL
);
2973 cache
->nr_ctr_args
= argc
;
2974 cache
->ctr_args
= copy
;
2979 static int cache_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
2982 struct cache_args
*ca
;
2983 struct cache
*cache
= NULL
;
2985 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
2987 ti
->error
= "Error allocating memory for cache";
2992 r
= parse_cache_args(ca
, argc
, argv
, &ti
->error
);
2996 r
= cache_create(ca
, &cache
);
3000 r
= copy_ctr_args(cache
, argc
- 3, (const char **)argv
+ 3);
3006 ti
->private = cache
;
3009 destroy_cache_args(ca
);
3013 /*----------------------------------------------------------------*/
3015 static int cache_map(struct dm_target
*ti
, struct bio
*bio
)
3017 struct cache
*cache
= ti
->private;
3020 struct dm_bio_prison_cell
*cell
= NULL
;
3021 dm_oblock_t block
= get_bio_block(cache
, bio
);
3022 size_t pb_data_size
= get_per_bio_data_size(cache
);
3023 bool can_migrate
= false;
3024 bool fast_promotion
;
3025 struct policy_result lookup_result
;
3026 struct per_bio_data
*pb
= init_per_bio_data(bio
, pb_data_size
);
3027 struct old_oblock_lock ool
;
3029 ool
.locker
.fn
= null_locker
;
3031 if (unlikely(from_oblock(block
) >= from_oblock(cache
->origin_blocks
))) {
3033 * This can only occur if the io goes to a partial block at
3034 * the end of the origin device. We don't cache these.
3035 * Just remap to the origin and carry on.
3037 remap_to_origin(cache
, bio
);
3038 accounted_begin(cache
, bio
);
3039 return DM_MAPIO_REMAPPED
;
3042 if (discard_or_flush(bio
)) {
3043 defer_bio(cache
, bio
);
3044 return DM_MAPIO_SUBMITTED
;
3048 * Check to see if that block is currently migrating.
3050 cell
= alloc_prison_cell(cache
);
3052 defer_bio(cache
, bio
);
3053 return DM_MAPIO_SUBMITTED
;
3056 r
= bio_detain(cache
, block
, bio
, cell
,
3057 (cell_free_fn
) free_prison_cell
,
3061 defer_bio(cache
, bio
);
3063 return DM_MAPIO_SUBMITTED
;
3066 fast_promotion
= is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
);
3068 r
= policy_map(cache
->policy
, block
, false, can_migrate
, fast_promotion
,
3069 bio
, &ool
.locker
, &lookup_result
);
3070 if (r
== -EWOULDBLOCK
) {
3071 cell_defer(cache
, cell
, true);
3072 return DM_MAPIO_SUBMITTED
;
3075 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3076 cache_device_name(cache
), r
);
3077 cell_defer(cache
, cell
, false);
3079 return DM_MAPIO_SUBMITTED
;
3082 r
= DM_MAPIO_REMAPPED
;
3083 switch (lookup_result
.op
) {
3085 if (passthrough_mode(&cache
->features
)) {
3086 if (bio_data_dir(bio
) == WRITE
) {
3088 * We need to invalidate this block, so
3089 * defer for the worker thread.
3091 cell_defer(cache
, cell
, true);
3092 r
= DM_MAPIO_SUBMITTED
;
3095 inc_miss_counter(cache
, bio
);
3096 remap_to_origin_clear_discard(cache
, bio
, block
);
3097 accounted_begin(cache
, bio
);
3098 inc_ds(cache
, bio
, cell
);
3099 // FIXME: we want to remap hits or misses straight
3100 // away rather than passing over to the worker.
3101 cell_defer(cache
, cell
, false);
3105 inc_hit_counter(cache
, bio
);
3106 if (bio_data_dir(bio
) == WRITE
&& writethrough_mode(&cache
->features
) &&
3107 !is_dirty(cache
, lookup_result
.cblock
)) {
3108 remap_to_origin_then_cache(cache
, bio
, block
, lookup_result
.cblock
);
3109 accounted_begin(cache
, bio
);
3110 inc_ds(cache
, bio
, cell
);
3111 cell_defer(cache
, cell
, false);
3114 remap_cell_to_cache_dirty(cache
, cell
, block
, lookup_result
.cblock
, false);
3119 inc_miss_counter(cache
, bio
);
3120 if (pb
->req_nr
!= 0) {
3122 * This is a duplicate writethrough io that is no
3123 * longer needed because the block has been demoted.
3126 // FIXME: remap everything as a miss
3127 cell_defer(cache
, cell
, false);
3128 r
= DM_MAPIO_SUBMITTED
;
3131 remap_cell_to_origin_clear_discard(cache
, cell
, block
, false);
3135 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3136 cache_device_name(cache
), __func__
,
3137 (unsigned) lookup_result
.op
);
3138 cell_defer(cache
, cell
, false);
3140 r
= DM_MAPIO_SUBMITTED
;
3146 static int cache_end_io(struct dm_target
*ti
, struct bio
*bio
, int error
)
3148 struct cache
*cache
= ti
->private;
3149 unsigned long flags
;
3150 size_t pb_data_size
= get_per_bio_data_size(cache
);
3151 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
3154 policy_tick(cache
->policy
, false);
3156 spin_lock_irqsave(&cache
->lock
, flags
);
3157 cache
->need_tick_bio
= true;
3158 spin_unlock_irqrestore(&cache
->lock
, flags
);
3161 check_for_quiesced_migrations(cache
, pb
);
3162 accounted_complete(cache
, bio
);
3167 static int write_dirty_bitset(struct cache
*cache
)
3171 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3174 for (i
= 0; i
< from_cblock(cache
->cache_size
); i
++) {
3175 r
= dm_cache_set_dirty(cache
->cmd
, to_cblock(i
),
3176 is_dirty(cache
, to_cblock(i
)));
3178 metadata_operation_failed(cache
, "dm_cache_set_dirty", r
);
3186 static int write_discard_bitset(struct cache
*cache
)
3190 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3193 r
= dm_cache_discard_bitset_resize(cache
->cmd
, cache
->discard_block_size
,
3194 cache
->discard_nr_blocks
);
3196 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache
));
3197 metadata_operation_failed(cache
, "dm_cache_discard_bitset_resize", r
);
3201 for (i
= 0; i
< from_dblock(cache
->discard_nr_blocks
); i
++) {
3202 r
= dm_cache_set_discard(cache
->cmd
, to_dblock(i
),
3203 is_discarded(cache
, to_dblock(i
)));
3205 metadata_operation_failed(cache
, "dm_cache_set_discard", r
);
3213 static int write_hints(struct cache
*cache
)
3217 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3220 r
= dm_cache_write_hints(cache
->cmd
, cache
->policy
);
3222 metadata_operation_failed(cache
, "dm_cache_write_hints", r
);
3230 * returns true on success
3232 static bool sync_metadata(struct cache
*cache
)
3236 r1
= write_dirty_bitset(cache
);
3238 DMERR("%s: could not write dirty bitset", cache_device_name(cache
));
3240 r2
= write_discard_bitset(cache
);
3242 DMERR("%s: could not write discard bitset", cache_device_name(cache
));
3246 r3
= write_hints(cache
);
3248 DMERR("%s: could not write hints", cache_device_name(cache
));
3251 * If writing the above metadata failed, we still commit, but don't
3252 * set the clean shutdown flag. This will effectively force every
3253 * dirty bit to be set on reload.
3255 r4
= commit(cache
, !r1
&& !r2
&& !r3
);
3257 DMERR("%s: could not write cache metadata", cache_device_name(cache
));
3259 return !r1
&& !r2
&& !r3
&& !r4
;
3262 static void cache_postsuspend(struct dm_target
*ti
)
3264 struct cache
*cache
= ti
->private;
3266 start_quiescing(cache
);
3267 wait_for_migrations(cache
);
3269 requeue_deferred_bios(cache
);
3270 requeue_deferred_cells(cache
);
3271 stop_quiescing(cache
);
3273 if (get_cache_mode(cache
) == CM_WRITE
)
3274 (void) sync_metadata(cache
);
3277 static int load_mapping(void *context
, dm_oblock_t oblock
, dm_cblock_t cblock
,
3278 bool dirty
, uint32_t hint
, bool hint_valid
)
3281 struct cache
*cache
= context
;
3283 r
= policy_load_mapping(cache
->policy
, oblock
, cblock
, hint
, hint_valid
);
3288 set_dirty(cache
, oblock
, cblock
);
3290 clear_dirty(cache
, oblock
, cblock
);
3296 * The discard block size in the on disk metadata is not
3297 * neccessarily the same as we're currently using. So we have to
3298 * be careful to only set the discarded attribute if we know it
3299 * covers a complete block of the new size.
3301 struct discard_load_info
{
3302 struct cache
*cache
;
3305 * These blocks are sized using the on disk dblock size, rather
3306 * than the current one.
3308 dm_block_t block_size
;
3309 dm_block_t discard_begin
, discard_end
;
3312 static void discard_load_info_init(struct cache
*cache
,
3313 struct discard_load_info
*li
)
3316 li
->discard_begin
= li
->discard_end
= 0;
3319 static void set_discard_range(struct discard_load_info
*li
)
3323 if (li
->discard_begin
== li
->discard_end
)
3327 * Convert to sectors.
3329 b
= li
->discard_begin
* li
->block_size
;
3330 e
= li
->discard_end
* li
->block_size
;
3333 * Then convert back to the current dblock size.
3335 b
= dm_sector_div_up(b
, li
->cache
->discard_block_size
);
3336 sector_div(e
, li
->cache
->discard_block_size
);
3339 * The origin may have shrunk, so we need to check we're still in
3342 if (e
> from_dblock(li
->cache
->discard_nr_blocks
))
3343 e
= from_dblock(li
->cache
->discard_nr_blocks
);
3346 set_discard(li
->cache
, to_dblock(b
));
3349 static int load_discard(void *context
, sector_t discard_block_size
,
3350 dm_dblock_t dblock
, bool discard
)
3352 struct discard_load_info
*li
= context
;
3354 li
->block_size
= discard_block_size
;
3357 if (from_dblock(dblock
) == li
->discard_end
)
3359 * We're already in a discard range, just extend it.
3361 li
->discard_end
= li
->discard_end
+ 1ULL;
3365 * Emit the old range and start a new one.
3367 set_discard_range(li
);
3368 li
->discard_begin
= from_dblock(dblock
);
3369 li
->discard_end
= li
->discard_begin
+ 1ULL;
3372 set_discard_range(li
);
3373 li
->discard_begin
= li
->discard_end
= 0;
3379 static dm_cblock_t
get_cache_dev_size(struct cache
*cache
)
3381 sector_t size
= get_dev_size(cache
->cache_dev
);
3382 (void) sector_div(size
, cache
->sectors_per_block
);
3383 return to_cblock(size
);
3386 static bool can_resize(struct cache
*cache
, dm_cblock_t new_size
)
3388 if (from_cblock(new_size
) > from_cblock(cache
->cache_size
))
3392 * We can't drop a dirty block when shrinking the cache.
3394 while (from_cblock(new_size
) < from_cblock(cache
->cache_size
)) {
3395 new_size
= to_cblock(from_cblock(new_size
) + 1);
3396 if (is_dirty(cache
, new_size
)) {
3397 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3398 cache_device_name(cache
),
3399 (unsigned long long) from_cblock(new_size
));
3407 static int resize_cache_dev(struct cache
*cache
, dm_cblock_t new_size
)
3411 r
= dm_cache_resize(cache
->cmd
, new_size
);
3413 DMERR("%s: could not resize cache metadata", cache_device_name(cache
));
3414 metadata_operation_failed(cache
, "dm_cache_resize", r
);
3418 set_cache_size(cache
, new_size
);
3423 static int cache_preresume(struct dm_target
*ti
)
3426 struct cache
*cache
= ti
->private;
3427 dm_cblock_t csize
= get_cache_dev_size(cache
);
3430 * Check to see if the cache has resized.
3432 if (!cache
->sized
) {
3433 r
= resize_cache_dev(cache
, csize
);
3437 cache
->sized
= true;
3439 } else if (csize
!= cache
->cache_size
) {
3440 if (!can_resize(cache
, csize
))
3443 r
= resize_cache_dev(cache
, csize
);
3448 if (!cache
->loaded_mappings
) {
3449 r
= dm_cache_load_mappings(cache
->cmd
, cache
->policy
,
3450 load_mapping
, cache
);
3452 DMERR("%s: could not load cache mappings", cache_device_name(cache
));
3453 metadata_operation_failed(cache
, "dm_cache_load_mappings", r
);
3457 cache
->loaded_mappings
= true;
3460 if (!cache
->loaded_discards
) {
3461 struct discard_load_info li
;
3464 * The discard bitset could have been resized, or the
3465 * discard block size changed. To be safe we start by
3466 * setting every dblock to not discarded.
3468 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
3470 discard_load_info_init(cache
, &li
);
3471 r
= dm_cache_load_discards(cache
->cmd
, load_discard
, &li
);
3473 DMERR("%s: could not load origin discards", cache_device_name(cache
));
3474 metadata_operation_failed(cache
, "dm_cache_load_discards", r
);
3477 set_discard_range(&li
);
3479 cache
->loaded_discards
= true;
3485 static void cache_resume(struct dm_target
*ti
)
3487 struct cache
*cache
= ti
->private;
3489 cache
->need_tick_bio
= true;
3490 do_waker(&cache
->waker
.work
);
3496 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3497 * <cache block size> <#used cache blocks>/<#total cache blocks>
3498 * <#read hits> <#read misses> <#write hits> <#write misses>
3499 * <#demotions> <#promotions> <#dirty>
3500 * <#features> <features>*
3501 * <#core args> <core args>
3502 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3504 static void cache_status(struct dm_target
*ti
, status_type_t type
,
3505 unsigned status_flags
, char *result
, unsigned maxlen
)
3510 dm_block_t nr_free_blocks_metadata
= 0;
3511 dm_block_t nr_blocks_metadata
= 0;
3512 char buf
[BDEVNAME_SIZE
];
3513 struct cache
*cache
= ti
->private;
3514 dm_cblock_t residency
;
3517 case STATUSTYPE_INFO
:
3518 if (get_cache_mode(cache
) == CM_FAIL
) {
3523 /* Commit to ensure statistics aren't out-of-date */
3524 if (!(status_flags
& DM_STATUS_NOFLUSH_FLAG
) && !dm_suspended(ti
))
3525 (void) commit(cache
, false);
3527 r
= dm_cache_get_free_metadata_block_count(cache
->cmd
, &nr_free_blocks_metadata
);
3529 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3530 cache_device_name(cache
), r
);
3534 r
= dm_cache_get_metadata_dev_size(cache
->cmd
, &nr_blocks_metadata
);
3536 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3537 cache_device_name(cache
), r
);
3541 residency
= policy_residency(cache
->policy
);
3543 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3544 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE
,
3545 (unsigned long long)(nr_blocks_metadata
- nr_free_blocks_metadata
),
3546 (unsigned long long)nr_blocks_metadata
,
3547 cache
->sectors_per_block
,
3548 (unsigned long long) from_cblock(residency
),
3549 (unsigned long long) from_cblock(cache
->cache_size
),
3550 (unsigned) atomic_read(&cache
->stats
.read_hit
),
3551 (unsigned) atomic_read(&cache
->stats
.read_miss
),
3552 (unsigned) atomic_read(&cache
->stats
.write_hit
),
3553 (unsigned) atomic_read(&cache
->stats
.write_miss
),
3554 (unsigned) atomic_read(&cache
->stats
.demotion
),
3555 (unsigned) atomic_read(&cache
->stats
.promotion
),
3556 (unsigned long) atomic_read(&cache
->nr_dirty
));
3558 if (writethrough_mode(&cache
->features
))
3559 DMEMIT("1 writethrough ");
3561 else if (passthrough_mode(&cache
->features
))
3562 DMEMIT("1 passthrough ");
3564 else if (writeback_mode(&cache
->features
))
3565 DMEMIT("1 writeback ");
3568 DMERR("%s: internal error: unknown io mode: %d",
3569 cache_device_name(cache
), (int) cache
->features
.io_mode
);
3573 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache
->migration_threshold
);
3575 DMEMIT("%s ", dm_cache_policy_get_name(cache
->policy
));
3577 r
= policy_emit_config_values(cache
->policy
, result
, maxlen
, &sz
);
3579 DMERR("%s: policy_emit_config_values returned %d",
3580 cache_device_name(cache
), r
);
3583 if (get_cache_mode(cache
) == CM_READ_ONLY
)
3588 if (dm_cache_metadata_needs_check(cache
->cmd
))
3589 DMEMIT("needs_check ");
3595 case STATUSTYPE_TABLE
:
3596 format_dev_t(buf
, cache
->metadata_dev
->bdev
->bd_dev
);
3598 format_dev_t(buf
, cache
->cache_dev
->bdev
->bd_dev
);
3600 format_dev_t(buf
, cache
->origin_dev
->bdev
->bd_dev
);
3603 for (i
= 0; i
< cache
->nr_ctr_args
- 1; i
++)
3604 DMEMIT(" %s", cache
->ctr_args
[i
]);
3605 if (cache
->nr_ctr_args
)
3606 DMEMIT(" %s", cache
->ctr_args
[cache
->nr_ctr_args
- 1]);
3616 * A cache block range can take two forms:
3618 * i) A single cblock, eg. '3456'
3619 * ii) A begin and end cblock with dots between, eg. 123-234
3621 static int parse_cblock_range(struct cache
*cache
, const char *str
,
3622 struct cblock_range
*result
)
3629 * Try and parse form (ii) first.
3631 r
= sscanf(str
, "%llu-%llu%c", &b
, &e
, &dummy
);
3636 result
->begin
= to_cblock(b
);
3637 result
->end
= to_cblock(e
);
3642 * That didn't work, try form (i).
3644 r
= sscanf(str
, "%llu%c", &b
, &dummy
);
3649 result
->begin
= to_cblock(b
);
3650 result
->end
= to_cblock(from_cblock(result
->begin
) + 1u);
3654 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache
), str
);
3658 static int validate_cblock_range(struct cache
*cache
, struct cblock_range
*range
)
3660 uint64_t b
= from_cblock(range
->begin
);
3661 uint64_t e
= from_cblock(range
->end
);
3662 uint64_t n
= from_cblock(cache
->cache_size
);
3665 DMERR("%s: begin cblock out of range: %llu >= %llu",
3666 cache_device_name(cache
), b
, n
);
3671 DMERR("%s: end cblock out of range: %llu > %llu",
3672 cache_device_name(cache
), e
, n
);
3677 DMERR("%s: invalid cblock range: %llu >= %llu",
3678 cache_device_name(cache
), b
, e
);
3685 static int request_invalidation(struct cache
*cache
, struct cblock_range
*range
)
3687 struct invalidation_request req
;
3689 INIT_LIST_HEAD(&req
.list
);
3690 req
.cblocks
= range
;
3691 atomic_set(&req
.complete
, 0);
3693 init_waitqueue_head(&req
.result_wait
);
3695 spin_lock(&cache
->invalidation_lock
);
3696 list_add(&req
.list
, &cache
->invalidation_requests
);
3697 spin_unlock(&cache
->invalidation_lock
);
3700 wait_event(req
.result_wait
, atomic_read(&req
.complete
));
3704 static int process_invalidate_cblocks_message(struct cache
*cache
, unsigned count
,
3705 const char **cblock_ranges
)
3709 struct cblock_range range
;
3711 if (!passthrough_mode(&cache
->features
)) {
3712 DMERR("%s: cache has to be in passthrough mode for invalidation",
3713 cache_device_name(cache
));
3717 for (i
= 0; i
< count
; i
++) {
3718 r
= parse_cblock_range(cache
, cblock_ranges
[i
], &range
);
3722 r
= validate_cblock_range(cache
, &range
);
3727 * Pass begin and end origin blocks to the worker and wake it.
3729 r
= request_invalidation(cache
, &range
);
3741 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3743 * The key migration_threshold is supported by the cache target core.
3745 static int cache_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
3747 struct cache
*cache
= ti
->private;
3752 if (get_cache_mode(cache
) >= CM_READ_ONLY
) {
3753 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3754 cache_device_name(cache
));
3758 if (!strcasecmp(argv
[0], "invalidate_cblocks"))
3759 return process_invalidate_cblocks_message(cache
, argc
- 1, (const char **) argv
+ 1);
3764 return set_config_value(cache
, argv
[0], argv
[1]);
3767 static int cache_iterate_devices(struct dm_target
*ti
,
3768 iterate_devices_callout_fn fn
, void *data
)
3771 struct cache
*cache
= ti
->private;
3773 r
= fn(ti
, cache
->cache_dev
, 0, get_dev_size(cache
->cache_dev
), data
);
3775 r
= fn(ti
, cache
->origin_dev
, 0, ti
->len
, data
);
3781 * We assume I/O is going to the origin (which is the volume
3782 * more likely to have restrictions e.g. by being striped).
3783 * (Looking up the exact location of the data would be expensive
3784 * and could always be out of date by the time the bio is submitted.)
3786 static int cache_bvec_merge(struct dm_target
*ti
,
3787 struct bvec_merge_data
*bvm
,
3788 struct bio_vec
*biovec
, int max_size
)
3790 struct cache
*cache
= ti
->private;
3791 struct request_queue
*q
= bdev_get_queue(cache
->origin_dev
->bdev
);
3793 if (!q
->merge_bvec_fn
)
3796 bvm
->bi_bdev
= cache
->origin_dev
->bdev
;
3797 return min(max_size
, q
->merge_bvec_fn(q
, bvm
, biovec
));
3800 static void set_discard_limits(struct cache
*cache
, struct queue_limits
*limits
)
3803 * FIXME: these limits may be incompatible with the cache device
3805 limits
->max_discard_sectors
= min_t(sector_t
, cache
->discard_block_size
* 1024,
3806 cache
->origin_sectors
);
3807 limits
->discard_granularity
= cache
->discard_block_size
<< SECTOR_SHIFT
;
3810 static void cache_io_hints(struct dm_target
*ti
, struct queue_limits
*limits
)
3812 struct cache
*cache
= ti
->private;
3813 uint64_t io_opt_sectors
= limits
->io_opt
>> SECTOR_SHIFT
;
3816 * If the system-determined stacked limits are compatible with the
3817 * cache's blocksize (io_opt is a factor) do not override them.
3819 if (io_opt_sectors
< cache
->sectors_per_block
||
3820 do_div(io_opt_sectors
, cache
->sectors_per_block
)) {
3821 blk_limits_io_min(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3822 blk_limits_io_opt(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3824 set_discard_limits(cache
, limits
);
3827 /*----------------------------------------------------------------*/
3829 static struct target_type cache_target
= {
3831 .version
= {1, 8, 0},
3832 .module
= THIS_MODULE
,
3836 .end_io
= cache_end_io
,
3837 .postsuspend
= cache_postsuspend
,
3838 .preresume
= cache_preresume
,
3839 .resume
= cache_resume
,
3840 .status
= cache_status
,
3841 .message
= cache_message
,
3842 .iterate_devices
= cache_iterate_devices
,
3843 .merge
= cache_bvec_merge
,
3844 .io_hints
= cache_io_hints
,
3847 static int __init
dm_cache_init(void)
3851 r
= dm_register_target(&cache_target
);
3853 DMERR("cache target registration failed: %d", r
);
3857 migration_cache
= KMEM_CACHE(dm_cache_migration
, 0);
3858 if (!migration_cache
) {
3859 dm_unregister_target(&cache_target
);
3866 static void __exit
dm_cache_exit(void)
3868 dm_unregister_target(&cache_target
);
3869 kmem_cache_destroy(migration_cache
);
3872 module_init(dm_cache_init
);
3873 module_exit(dm_cache_exit
);
3875 MODULE_DESCRIPTION(DM_NAME
" cache target");
3876 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3877 MODULE_LICENSE("GPL");