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
);
430 static int prealloc_data_structs(struct cache
*cache
, struct prealloc
*p
)
433 p
->mg
= alloc_migration(cache
);
439 p
->cell1
= alloc_prison_cell(cache
);
445 p
->cell2
= alloc_prison_cell(cache
);
453 static void prealloc_free_structs(struct cache
*cache
, struct prealloc
*p
)
456 free_prison_cell(cache
, p
->cell2
);
459 free_prison_cell(cache
, p
->cell1
);
462 free_migration(p
->mg
);
465 static struct dm_cache_migration
*prealloc_get_migration(struct prealloc
*p
)
467 struct dm_cache_migration
*mg
= p
->mg
;
476 * You must have a cell within the prealloc struct to return. If not this
477 * function will BUG() rather than returning NULL.
479 static struct dm_bio_prison_cell
*prealloc_get_cell(struct prealloc
*p
)
481 struct dm_bio_prison_cell
*r
= NULL
;
487 } else if (p
->cell2
) {
497 * You can't have more than two cells in a prealloc struct. BUG() will be
498 * called if you try and overfill.
500 static void prealloc_put_cell(struct prealloc
*p
, struct dm_bio_prison_cell
*cell
)
512 /*----------------------------------------------------------------*/
514 static void build_key(dm_oblock_t begin
, dm_oblock_t end
, struct dm_cell_key
*key
)
518 key
->block_begin
= from_oblock(begin
);
519 key
->block_end
= from_oblock(end
);
523 * The caller hands in a preallocated cell, and a free function for it.
524 * The cell will be freed if there's an error, or if it wasn't used because
525 * a cell with that key already exists.
527 typedef void (*cell_free_fn
)(void *context
, struct dm_bio_prison_cell
*cell
);
529 static int bio_detain_range(struct cache
*cache
, dm_oblock_t oblock_begin
, dm_oblock_t oblock_end
,
530 struct bio
*bio
, struct dm_bio_prison_cell
*cell_prealloc
,
531 cell_free_fn free_fn
, void *free_context
,
532 struct dm_bio_prison_cell
**cell_result
)
535 struct dm_cell_key key
;
537 build_key(oblock_begin
, oblock_end
, &key
);
538 r
= dm_bio_detain(cache
->prison
, &key
, bio
, cell_prealloc
, cell_result
);
540 free_fn(free_context
, cell_prealloc
);
545 static int bio_detain(struct cache
*cache
, dm_oblock_t oblock
,
546 struct bio
*bio
, struct dm_bio_prison_cell
*cell_prealloc
,
547 cell_free_fn free_fn
, void *free_context
,
548 struct dm_bio_prison_cell
**cell_result
)
550 dm_oblock_t end
= to_oblock(from_oblock(oblock
) + 1ULL);
551 return bio_detain_range(cache
, oblock
, end
, bio
,
552 cell_prealloc
, free_fn
, free_context
, cell_result
);
555 static int get_cell(struct cache
*cache
,
557 struct prealloc
*structs
,
558 struct dm_bio_prison_cell
**cell_result
)
561 struct dm_cell_key key
;
562 struct dm_bio_prison_cell
*cell_prealloc
;
564 cell_prealloc
= prealloc_get_cell(structs
);
566 build_key(oblock
, to_oblock(from_oblock(oblock
) + 1ULL), &key
);
567 r
= dm_get_cell(cache
->prison
, &key
, cell_prealloc
, cell_result
);
569 prealloc_put_cell(structs
, cell_prealloc
);
574 /*----------------------------------------------------------------*/
576 static bool is_dirty(struct cache
*cache
, dm_cblock_t b
)
578 return test_bit(from_cblock(b
), cache
->dirty_bitset
);
581 static void set_dirty(struct cache
*cache
, dm_oblock_t oblock
, dm_cblock_t cblock
)
583 if (!test_and_set_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
584 atomic_inc(&cache
->nr_dirty
);
585 policy_set_dirty(cache
->policy
, oblock
);
589 static void clear_dirty(struct cache
*cache
, dm_oblock_t oblock
, dm_cblock_t cblock
)
591 if (test_and_clear_bit(from_cblock(cblock
), cache
->dirty_bitset
)) {
592 policy_clear_dirty(cache
->policy
, oblock
);
593 if (atomic_dec_return(&cache
->nr_dirty
) == 0)
594 dm_table_event(cache
->ti
->table
);
598 /*----------------------------------------------------------------*/
600 static bool block_size_is_power_of_two(struct cache
*cache
)
602 return cache
->sectors_per_block_shift
>= 0;
605 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
606 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
609 static dm_block_t
block_div(dm_block_t b
, uint32_t n
)
616 static dm_block_t
oblocks_per_dblock(struct cache
*cache
)
618 dm_block_t oblocks
= cache
->discard_block_size
;
620 if (block_size_is_power_of_two(cache
))
621 oblocks
>>= cache
->sectors_per_block_shift
;
623 oblocks
= block_div(oblocks
, cache
->sectors_per_block
);
628 static dm_dblock_t
oblock_to_dblock(struct cache
*cache
, dm_oblock_t oblock
)
630 return to_dblock(block_div(from_oblock(oblock
),
631 oblocks_per_dblock(cache
)));
634 static dm_oblock_t
dblock_to_oblock(struct cache
*cache
, dm_dblock_t dblock
)
636 return to_oblock(from_dblock(dblock
) * oblocks_per_dblock(cache
));
639 static void set_discard(struct cache
*cache
, dm_dblock_t b
)
643 BUG_ON(from_dblock(b
) >= from_dblock(cache
->discard_nr_blocks
));
644 atomic_inc(&cache
->stats
.discard_count
);
646 spin_lock_irqsave(&cache
->lock
, flags
);
647 set_bit(from_dblock(b
), cache
->discard_bitset
);
648 spin_unlock_irqrestore(&cache
->lock
, flags
);
651 static void clear_discard(struct cache
*cache
, dm_dblock_t b
)
655 spin_lock_irqsave(&cache
->lock
, flags
);
656 clear_bit(from_dblock(b
), cache
->discard_bitset
);
657 spin_unlock_irqrestore(&cache
->lock
, flags
);
660 static bool is_discarded(struct cache
*cache
, dm_dblock_t b
)
665 spin_lock_irqsave(&cache
->lock
, flags
);
666 r
= test_bit(from_dblock(b
), cache
->discard_bitset
);
667 spin_unlock_irqrestore(&cache
->lock
, flags
);
672 static bool is_discarded_oblock(struct cache
*cache
, dm_oblock_t b
)
677 spin_lock_irqsave(&cache
->lock
, flags
);
678 r
= test_bit(from_dblock(oblock_to_dblock(cache
, b
)),
679 cache
->discard_bitset
);
680 spin_unlock_irqrestore(&cache
->lock
, flags
);
685 /*----------------------------------------------------------------*/
687 static void load_stats(struct cache
*cache
)
689 struct dm_cache_statistics stats
;
691 dm_cache_metadata_get_stats(cache
->cmd
, &stats
);
692 atomic_set(&cache
->stats
.read_hit
, stats
.read_hits
);
693 atomic_set(&cache
->stats
.read_miss
, stats
.read_misses
);
694 atomic_set(&cache
->stats
.write_hit
, stats
.write_hits
);
695 atomic_set(&cache
->stats
.write_miss
, stats
.write_misses
);
698 static void save_stats(struct cache
*cache
)
700 struct dm_cache_statistics stats
;
702 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
705 stats
.read_hits
= atomic_read(&cache
->stats
.read_hit
);
706 stats
.read_misses
= atomic_read(&cache
->stats
.read_miss
);
707 stats
.write_hits
= atomic_read(&cache
->stats
.write_hit
);
708 stats
.write_misses
= atomic_read(&cache
->stats
.write_miss
);
710 dm_cache_metadata_set_stats(cache
->cmd
, &stats
);
713 /*----------------------------------------------------------------
715 *--------------------------------------------------------------*/
718 * If using writeback, leave out struct per_bio_data's writethrough fields.
720 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
721 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
723 static bool writethrough_mode(struct cache_features
*f
)
725 return f
->io_mode
== CM_IO_WRITETHROUGH
;
728 static bool writeback_mode(struct cache_features
*f
)
730 return f
->io_mode
== CM_IO_WRITEBACK
;
733 static bool passthrough_mode(struct cache_features
*f
)
735 return f
->io_mode
== CM_IO_PASSTHROUGH
;
738 static size_t get_per_bio_data_size(struct cache
*cache
)
740 return writethrough_mode(&cache
->features
) ? PB_DATA_SIZE_WT
: PB_DATA_SIZE_WB
;
743 static struct per_bio_data
*get_per_bio_data(struct bio
*bio
, size_t data_size
)
745 struct per_bio_data
*pb
= dm_per_bio_data(bio
, data_size
);
750 static struct per_bio_data
*init_per_bio_data(struct bio
*bio
, size_t data_size
)
752 struct per_bio_data
*pb
= get_per_bio_data(bio
, data_size
);
755 pb
->req_nr
= dm_bio_get_target_bio_nr(bio
);
756 pb
->all_io_entry
= NULL
;
762 /*----------------------------------------------------------------
764 *--------------------------------------------------------------*/
765 static void remap_to_origin(struct cache
*cache
, struct bio
*bio
)
767 bio
->bi_bdev
= cache
->origin_dev
->bdev
;
770 static void remap_to_cache(struct cache
*cache
, struct bio
*bio
,
773 sector_t bi_sector
= bio
->bi_iter
.bi_sector
;
774 sector_t block
= from_cblock(cblock
);
776 bio
->bi_bdev
= cache
->cache_dev
->bdev
;
777 if (!block_size_is_power_of_two(cache
))
778 bio
->bi_iter
.bi_sector
=
779 (block
* cache
->sectors_per_block
) +
780 sector_div(bi_sector
, cache
->sectors_per_block
);
782 bio
->bi_iter
.bi_sector
=
783 (block
<< cache
->sectors_per_block_shift
) |
784 (bi_sector
& (cache
->sectors_per_block
- 1));
787 static void check_if_tick_bio_needed(struct cache
*cache
, struct bio
*bio
)
790 size_t pb_data_size
= get_per_bio_data_size(cache
);
791 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
793 spin_lock_irqsave(&cache
->lock
, flags
);
794 if (cache
->need_tick_bio
&&
795 !(bio
->bi_rw
& (REQ_FUA
| REQ_FLUSH
| REQ_DISCARD
))) {
797 cache
->need_tick_bio
= false;
799 spin_unlock_irqrestore(&cache
->lock
, flags
);
802 static void remap_to_origin_clear_discard(struct cache
*cache
, struct bio
*bio
,
805 check_if_tick_bio_needed(cache
, bio
);
806 remap_to_origin(cache
, bio
);
807 if (bio_data_dir(bio
) == WRITE
)
808 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
811 static void remap_to_cache_dirty(struct cache
*cache
, struct bio
*bio
,
812 dm_oblock_t oblock
, dm_cblock_t cblock
)
814 check_if_tick_bio_needed(cache
, bio
);
815 remap_to_cache(cache
, bio
, cblock
);
816 if (bio_data_dir(bio
) == WRITE
) {
817 set_dirty(cache
, oblock
, cblock
);
818 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
822 static dm_oblock_t
get_bio_block(struct cache
*cache
, struct bio
*bio
)
824 sector_t block_nr
= bio
->bi_iter
.bi_sector
;
826 if (!block_size_is_power_of_two(cache
))
827 (void) sector_div(block_nr
, cache
->sectors_per_block
);
829 block_nr
>>= cache
->sectors_per_block_shift
;
831 return to_oblock(block_nr
);
834 static int bio_triggers_commit(struct cache
*cache
, struct bio
*bio
)
836 return bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
);
840 * You must increment the deferred set whilst the prison cell is held. To
841 * encourage this, we ask for 'cell' to be passed in.
843 static void inc_ds(struct cache
*cache
, struct bio
*bio
,
844 struct dm_bio_prison_cell
*cell
)
846 size_t pb_data_size
= get_per_bio_data_size(cache
);
847 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
850 BUG_ON(pb
->all_io_entry
);
852 pb
->all_io_entry
= dm_deferred_entry_inc(cache
->all_io_ds
);
855 static bool accountable_bio(struct cache
*cache
, struct bio
*bio
)
857 return ((bio
->bi_bdev
== cache
->origin_dev
->bdev
) &&
858 !(bio
->bi_rw
& REQ_DISCARD
));
861 static void accounted_begin(struct cache
*cache
, struct bio
*bio
)
863 size_t pb_data_size
= get_per_bio_data_size(cache
);
864 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
866 if (accountable_bio(cache
, bio
)) {
867 pb
->len
= bio_sectors(bio
);
868 iot_io_begin(&cache
->origin_tracker
, pb
->len
);
872 static void accounted_complete(struct cache
*cache
, struct bio
*bio
)
874 size_t pb_data_size
= get_per_bio_data_size(cache
);
875 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
877 iot_io_end(&cache
->origin_tracker
, pb
->len
);
880 static void accounted_request(struct cache
*cache
, struct bio
*bio
)
882 accounted_begin(cache
, bio
);
883 generic_make_request(bio
);
886 static void issue(struct cache
*cache
, struct bio
*bio
)
890 if (!bio_triggers_commit(cache
, bio
)) {
891 accounted_request(cache
, bio
);
896 * Batch together any bios that trigger commits and then issue a
897 * single commit for them in do_worker().
899 spin_lock_irqsave(&cache
->lock
, flags
);
900 cache
->commit_requested
= true;
901 bio_list_add(&cache
->deferred_flush_bios
, bio
);
902 spin_unlock_irqrestore(&cache
->lock
, flags
);
905 static void inc_and_issue(struct cache
*cache
, struct bio
*bio
, struct dm_bio_prison_cell
*cell
)
907 inc_ds(cache
, bio
, cell
);
911 static void defer_writethrough_bio(struct cache
*cache
, struct bio
*bio
)
915 spin_lock_irqsave(&cache
->lock
, flags
);
916 bio_list_add(&cache
->deferred_writethrough_bios
, bio
);
917 spin_unlock_irqrestore(&cache
->lock
, flags
);
922 static void writethrough_endio(struct bio
*bio
, int err
)
924 struct per_bio_data
*pb
= get_per_bio_data(bio
, PB_DATA_SIZE_WT
);
926 dm_unhook_bio(&pb
->hook_info
, bio
);
933 dm_bio_restore(&pb
->bio_details
, bio
);
934 remap_to_cache(pb
->cache
, bio
, pb
->cblock
);
937 * We can't issue this bio directly, since we're in interrupt
938 * context. So it gets put on a bio list for processing by the
941 defer_writethrough_bio(pb
->cache
, bio
);
945 * When running in writethrough mode we need to send writes to clean blocks
946 * to both the cache and origin devices. In future we'd like to clone the
947 * bio and send them in parallel, but for now we're doing them in
948 * series as this is easier.
950 static void remap_to_origin_then_cache(struct cache
*cache
, struct bio
*bio
,
951 dm_oblock_t oblock
, dm_cblock_t cblock
)
953 struct per_bio_data
*pb
= get_per_bio_data(bio
, PB_DATA_SIZE_WT
);
957 dm_hook_bio(&pb
->hook_info
, bio
, writethrough_endio
, NULL
);
958 dm_bio_record(&pb
->bio_details
, bio
);
960 remap_to_origin_clear_discard(pb
->cache
, bio
, oblock
);
963 /*----------------------------------------------------------------
965 *--------------------------------------------------------------*/
966 static enum cache_metadata_mode
get_cache_mode(struct cache
*cache
)
968 return cache
->features
.mode
;
971 static const char *cache_device_name(struct cache
*cache
)
973 return dm_device_name(dm_table_get_md(cache
->ti
->table
));
976 static void notify_mode_switch(struct cache
*cache
, enum cache_metadata_mode mode
)
978 const char *descs
[] = {
984 dm_table_event(cache
->ti
->table
);
985 DMINFO("%s: switching cache to %s mode",
986 cache_device_name(cache
), descs
[(int)mode
]);
989 static void set_cache_mode(struct cache
*cache
, enum cache_metadata_mode new_mode
)
991 bool needs_check
= dm_cache_metadata_needs_check(cache
->cmd
);
992 enum cache_metadata_mode old_mode
= get_cache_mode(cache
);
994 if (new_mode
== CM_WRITE
&& needs_check
) {
995 DMERR("%s: unable to switch cache to write mode until repaired.",
996 cache_device_name(cache
));
997 if (old_mode
!= new_mode
)
1000 new_mode
= CM_READ_ONLY
;
1003 /* Never move out of fail mode */
1004 if (old_mode
== CM_FAIL
)
1010 dm_cache_metadata_set_read_only(cache
->cmd
);
1014 dm_cache_metadata_set_read_write(cache
->cmd
);
1018 cache
->features
.mode
= new_mode
;
1020 if (new_mode
!= old_mode
)
1021 notify_mode_switch(cache
, new_mode
);
1024 static void abort_transaction(struct cache
*cache
)
1026 const char *dev_name
= cache_device_name(cache
);
1028 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1031 if (dm_cache_metadata_set_needs_check(cache
->cmd
)) {
1032 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name
);
1033 set_cache_mode(cache
, CM_FAIL
);
1036 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name
);
1037 if (dm_cache_metadata_abort(cache
->cmd
)) {
1038 DMERR("%s: failed to abort metadata transaction", dev_name
);
1039 set_cache_mode(cache
, CM_FAIL
);
1043 static void metadata_operation_failed(struct cache
*cache
, const char *op
, int r
)
1045 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1046 cache_device_name(cache
), op
, r
);
1047 abort_transaction(cache
);
1048 set_cache_mode(cache
, CM_READ_ONLY
);
1051 /*----------------------------------------------------------------
1052 * Migration processing
1054 * Migration covers moving data from the origin device to the cache, or
1056 *--------------------------------------------------------------*/
1057 static void inc_io_migrations(struct cache
*cache
)
1059 atomic_inc(&cache
->nr_io_migrations
);
1062 static void dec_io_migrations(struct cache
*cache
)
1064 atomic_dec(&cache
->nr_io_migrations
);
1067 static void __cell_release(struct cache
*cache
, struct dm_bio_prison_cell
*cell
,
1068 bool holder
, struct bio_list
*bios
)
1070 (holder
? dm_cell_release
: dm_cell_release_no_holder
)
1071 (cache
->prison
, cell
, bios
);
1072 free_prison_cell(cache
, cell
);
1075 static bool discard_or_flush(struct bio
*bio
)
1077 return bio
->bi_rw
& (REQ_FLUSH
| REQ_FUA
| REQ_DISCARD
);
1080 static void __cell_defer(struct cache
*cache
, struct dm_bio_prison_cell
*cell
)
1082 if (discard_or_flush(cell
->holder
))
1084 * We have to handle these bios
1087 __cell_release(cache
, cell
, true, &cache
->deferred_bios
);
1090 list_add_tail(&cell
->user_list
, &cache
->deferred_cells
);
1093 static void cell_defer(struct cache
*cache
, struct dm_bio_prison_cell
*cell
, bool holder
)
1095 unsigned long flags
;
1097 if (!holder
&& dm_cell_promote_or_release(cache
->prison
, cell
)) {
1099 * There was no prisoner to promote to holder, the
1100 * cell has been released.
1102 free_prison_cell(cache
, cell
);
1106 spin_lock_irqsave(&cache
->lock
, flags
);
1107 __cell_defer(cache
, cell
);
1108 spin_unlock_irqrestore(&cache
->lock
, flags
);
1113 static void cell_error_with_code(struct cache
*cache
, struct dm_bio_prison_cell
*cell
, int err
)
1115 dm_cell_error(cache
->prison
, cell
, err
);
1116 dm_bio_prison_free_cell(cache
->prison
, cell
);
1119 static void cell_requeue(struct cache
*cache
, struct dm_bio_prison_cell
*cell
)
1121 cell_error_with_code(cache
, cell
, DM_ENDIO_REQUEUE
);
1124 static void free_io_migration(struct dm_cache_migration
*mg
)
1126 dec_io_migrations(mg
->cache
);
1130 static void migration_failure(struct dm_cache_migration
*mg
)
1132 struct cache
*cache
= mg
->cache
;
1133 const char *dev_name
= cache_device_name(cache
);
1135 if (mg
->writeback
) {
1136 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name
);
1137 set_dirty(cache
, mg
->old_oblock
, mg
->cblock
);
1138 cell_defer(cache
, mg
->old_ocell
, false);
1140 } else if (mg
->demote
) {
1141 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name
);
1142 policy_force_mapping(cache
->policy
, mg
->new_oblock
, mg
->old_oblock
);
1144 cell_defer(cache
, mg
->old_ocell
, mg
->promote
? false : true);
1146 cell_defer(cache
, mg
->new_ocell
, true);
1148 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name
);
1149 policy_remove_mapping(cache
->policy
, mg
->new_oblock
);
1150 cell_defer(cache
, mg
->new_ocell
, true);
1153 free_io_migration(mg
);
1156 static void migration_success_pre_commit(struct dm_cache_migration
*mg
)
1159 unsigned long flags
;
1160 struct cache
*cache
= mg
->cache
;
1162 if (mg
->writeback
) {
1163 clear_dirty(cache
, mg
->old_oblock
, mg
->cblock
);
1164 cell_defer(cache
, mg
->old_ocell
, false);
1165 free_io_migration(mg
);
1168 } else if (mg
->demote
) {
1169 r
= dm_cache_remove_mapping(cache
->cmd
, mg
->cblock
);
1171 DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1172 cache_device_name(cache
));
1173 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
1174 policy_force_mapping(cache
->policy
, mg
->new_oblock
,
1177 cell_defer(cache
, mg
->new_ocell
, true);
1178 free_io_migration(mg
);
1182 r
= dm_cache_insert_mapping(cache
->cmd
, mg
->cblock
, mg
->new_oblock
);
1184 DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1185 cache_device_name(cache
));
1186 metadata_operation_failed(cache
, "dm_cache_insert_mapping", r
);
1187 policy_remove_mapping(cache
->policy
, mg
->new_oblock
);
1188 free_io_migration(mg
);
1193 spin_lock_irqsave(&cache
->lock
, flags
);
1194 list_add_tail(&mg
->list
, &cache
->need_commit_migrations
);
1195 cache
->commit_requested
= true;
1196 spin_unlock_irqrestore(&cache
->lock
, flags
);
1199 static void migration_success_post_commit(struct dm_cache_migration
*mg
)
1201 unsigned long flags
;
1202 struct cache
*cache
= mg
->cache
;
1204 if (mg
->writeback
) {
1205 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1206 cache_device_name(cache
));
1209 } else if (mg
->demote
) {
1210 cell_defer(cache
, mg
->old_ocell
, mg
->promote
? false : true);
1215 spin_lock_irqsave(&cache
->lock
, flags
);
1216 list_add_tail(&mg
->list
, &cache
->quiesced_migrations
);
1217 spin_unlock_irqrestore(&cache
->lock
, flags
);
1221 policy_remove_mapping(cache
->policy
, mg
->old_oblock
);
1222 free_io_migration(mg
);
1226 if (mg
->requeue_holder
) {
1227 clear_dirty(cache
, mg
->new_oblock
, mg
->cblock
);
1228 cell_defer(cache
, mg
->new_ocell
, true);
1231 * The block was promoted via an overwrite, so it's dirty.
1233 set_dirty(cache
, mg
->new_oblock
, mg
->cblock
);
1234 bio_endio(mg
->new_ocell
->holder
, 0);
1235 cell_defer(cache
, mg
->new_ocell
, false);
1237 free_io_migration(mg
);
1241 static void copy_complete(int read_err
, unsigned long write_err
, void *context
)
1243 unsigned long flags
;
1244 struct dm_cache_migration
*mg
= (struct dm_cache_migration
*) context
;
1245 struct cache
*cache
= mg
->cache
;
1247 if (read_err
|| write_err
)
1250 spin_lock_irqsave(&cache
->lock
, flags
);
1251 list_add_tail(&mg
->list
, &cache
->completed_migrations
);
1252 spin_unlock_irqrestore(&cache
->lock
, flags
);
1257 static void issue_copy(struct dm_cache_migration
*mg
)
1260 struct dm_io_region o_region
, c_region
;
1261 struct cache
*cache
= mg
->cache
;
1262 sector_t cblock
= from_cblock(mg
->cblock
);
1264 o_region
.bdev
= cache
->origin_dev
->bdev
;
1265 o_region
.count
= cache
->sectors_per_block
;
1267 c_region
.bdev
= cache
->cache_dev
->bdev
;
1268 c_region
.sector
= cblock
* cache
->sectors_per_block
;
1269 c_region
.count
= cache
->sectors_per_block
;
1271 if (mg
->writeback
|| mg
->demote
) {
1273 o_region
.sector
= from_oblock(mg
->old_oblock
) * cache
->sectors_per_block
;
1274 r
= dm_kcopyd_copy(cache
->copier
, &c_region
, 1, &o_region
, 0, copy_complete
, mg
);
1277 o_region
.sector
= from_oblock(mg
->new_oblock
) * cache
->sectors_per_block
;
1278 r
= dm_kcopyd_copy(cache
->copier
, &o_region
, 1, &c_region
, 0, copy_complete
, mg
);
1282 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache
));
1283 migration_failure(mg
);
1287 static void overwrite_endio(struct bio
*bio
, int err
)
1289 struct dm_cache_migration
*mg
= bio
->bi_private
;
1290 struct cache
*cache
= mg
->cache
;
1291 size_t pb_data_size
= get_per_bio_data_size(cache
);
1292 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1293 unsigned long flags
;
1295 dm_unhook_bio(&pb
->hook_info
, bio
);
1300 mg
->requeue_holder
= false;
1302 spin_lock_irqsave(&cache
->lock
, flags
);
1303 list_add_tail(&mg
->list
, &cache
->completed_migrations
);
1304 spin_unlock_irqrestore(&cache
->lock
, flags
);
1309 static void issue_overwrite(struct dm_cache_migration
*mg
, struct bio
*bio
)
1311 size_t pb_data_size
= get_per_bio_data_size(mg
->cache
);
1312 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1314 dm_hook_bio(&pb
->hook_info
, bio
, overwrite_endio
, mg
);
1315 remap_to_cache_dirty(mg
->cache
, bio
, mg
->new_oblock
, mg
->cblock
);
1318 * No need to inc_ds() here, since the cell will be held for the
1319 * duration of the io.
1321 accounted_request(mg
->cache
, bio
);
1324 static bool bio_writes_complete_block(struct cache
*cache
, struct bio
*bio
)
1326 return (bio_data_dir(bio
) == WRITE
) &&
1327 (bio
->bi_iter
.bi_size
== (cache
->sectors_per_block
<< SECTOR_SHIFT
));
1330 static void avoid_copy(struct dm_cache_migration
*mg
)
1332 atomic_inc(&mg
->cache
->stats
.copies_avoided
);
1333 migration_success_pre_commit(mg
);
1336 static void calc_discard_block_range(struct cache
*cache
, struct bio
*bio
,
1337 dm_dblock_t
*b
, dm_dblock_t
*e
)
1339 sector_t sb
= bio
->bi_iter
.bi_sector
;
1340 sector_t se
= bio_end_sector(bio
);
1342 *b
= to_dblock(dm_sector_div_up(sb
, cache
->discard_block_size
));
1344 if (se
- sb
< cache
->discard_block_size
)
1347 *e
= to_dblock(block_div(se
, cache
->discard_block_size
));
1350 static void issue_discard(struct dm_cache_migration
*mg
)
1353 struct bio
*bio
= mg
->new_ocell
->holder
;
1355 calc_discard_block_range(mg
->cache
, bio
, &b
, &e
);
1357 set_discard(mg
->cache
, b
);
1358 b
= to_dblock(from_dblock(b
) + 1);
1362 cell_defer(mg
->cache
, mg
->new_ocell
, false);
1366 static void issue_copy_or_discard(struct dm_cache_migration
*mg
)
1369 struct cache
*cache
= mg
->cache
;
1376 if (mg
->writeback
|| mg
->demote
)
1377 avoid
= !is_dirty(cache
, mg
->cblock
) ||
1378 is_discarded_oblock(cache
, mg
->old_oblock
);
1380 struct bio
*bio
= mg
->new_ocell
->holder
;
1382 avoid
= is_discarded_oblock(cache
, mg
->new_oblock
);
1384 if (writeback_mode(&cache
->features
) &&
1385 !avoid
&& bio_writes_complete_block(cache
, bio
)) {
1386 issue_overwrite(mg
, bio
);
1391 avoid
? avoid_copy(mg
) : issue_copy(mg
);
1394 static void complete_migration(struct dm_cache_migration
*mg
)
1397 migration_failure(mg
);
1399 migration_success_pre_commit(mg
);
1402 static void process_migrations(struct cache
*cache
, struct list_head
*head
,
1403 void (*fn
)(struct dm_cache_migration
*))
1405 unsigned long flags
;
1406 struct list_head list
;
1407 struct dm_cache_migration
*mg
, *tmp
;
1409 INIT_LIST_HEAD(&list
);
1410 spin_lock_irqsave(&cache
->lock
, flags
);
1411 list_splice_init(head
, &list
);
1412 spin_unlock_irqrestore(&cache
->lock
, flags
);
1414 list_for_each_entry_safe(mg
, tmp
, &list
, list
)
1418 static void __queue_quiesced_migration(struct dm_cache_migration
*mg
)
1420 list_add_tail(&mg
->list
, &mg
->cache
->quiesced_migrations
);
1423 static void queue_quiesced_migration(struct dm_cache_migration
*mg
)
1425 unsigned long flags
;
1426 struct cache
*cache
= mg
->cache
;
1428 spin_lock_irqsave(&cache
->lock
, flags
);
1429 __queue_quiesced_migration(mg
);
1430 spin_unlock_irqrestore(&cache
->lock
, flags
);
1435 static void queue_quiesced_migrations(struct cache
*cache
, struct list_head
*work
)
1437 unsigned long flags
;
1438 struct dm_cache_migration
*mg
, *tmp
;
1440 spin_lock_irqsave(&cache
->lock
, flags
);
1441 list_for_each_entry_safe(mg
, tmp
, work
, list
)
1442 __queue_quiesced_migration(mg
);
1443 spin_unlock_irqrestore(&cache
->lock
, flags
);
1448 static void check_for_quiesced_migrations(struct cache
*cache
,
1449 struct per_bio_data
*pb
)
1451 struct list_head work
;
1453 if (!pb
->all_io_entry
)
1456 INIT_LIST_HEAD(&work
);
1457 dm_deferred_entry_dec(pb
->all_io_entry
, &work
);
1459 if (!list_empty(&work
))
1460 queue_quiesced_migrations(cache
, &work
);
1463 static void quiesce_migration(struct dm_cache_migration
*mg
)
1465 if (!dm_deferred_set_add_work(mg
->cache
->all_io_ds
, &mg
->list
))
1466 queue_quiesced_migration(mg
);
1469 static void promote(struct cache
*cache
, struct prealloc
*structs
,
1470 dm_oblock_t oblock
, dm_cblock_t cblock
,
1471 struct dm_bio_prison_cell
*cell
)
1473 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1476 mg
->discard
= false;
1477 mg
->writeback
= false;
1480 mg
->requeue_holder
= true;
1481 mg
->invalidate
= false;
1483 mg
->new_oblock
= oblock
;
1484 mg
->cblock
= cblock
;
1485 mg
->old_ocell
= NULL
;
1486 mg
->new_ocell
= cell
;
1487 mg
->start_jiffies
= jiffies
;
1489 inc_io_migrations(cache
);
1490 quiesce_migration(mg
);
1493 static void writeback(struct cache
*cache
, struct prealloc
*structs
,
1494 dm_oblock_t oblock
, dm_cblock_t cblock
,
1495 struct dm_bio_prison_cell
*cell
)
1497 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1500 mg
->discard
= false;
1501 mg
->writeback
= true;
1503 mg
->promote
= false;
1504 mg
->requeue_holder
= true;
1505 mg
->invalidate
= false;
1507 mg
->old_oblock
= oblock
;
1508 mg
->cblock
= cblock
;
1509 mg
->old_ocell
= cell
;
1510 mg
->new_ocell
= NULL
;
1511 mg
->start_jiffies
= jiffies
;
1513 inc_io_migrations(cache
);
1514 quiesce_migration(mg
);
1517 static void demote_then_promote(struct cache
*cache
, struct prealloc
*structs
,
1518 dm_oblock_t old_oblock
, dm_oblock_t new_oblock
,
1520 struct dm_bio_prison_cell
*old_ocell
,
1521 struct dm_bio_prison_cell
*new_ocell
)
1523 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1526 mg
->discard
= false;
1527 mg
->writeback
= false;
1530 mg
->requeue_holder
= true;
1531 mg
->invalidate
= false;
1533 mg
->old_oblock
= old_oblock
;
1534 mg
->new_oblock
= new_oblock
;
1535 mg
->cblock
= cblock
;
1536 mg
->old_ocell
= old_ocell
;
1537 mg
->new_ocell
= new_ocell
;
1538 mg
->start_jiffies
= jiffies
;
1540 inc_io_migrations(cache
);
1541 quiesce_migration(mg
);
1545 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1546 * block are thrown away.
1548 static void invalidate(struct cache
*cache
, struct prealloc
*structs
,
1549 dm_oblock_t oblock
, dm_cblock_t cblock
,
1550 struct dm_bio_prison_cell
*cell
)
1552 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1555 mg
->discard
= false;
1556 mg
->writeback
= false;
1558 mg
->promote
= false;
1559 mg
->requeue_holder
= true;
1560 mg
->invalidate
= true;
1562 mg
->old_oblock
= oblock
;
1563 mg
->cblock
= cblock
;
1564 mg
->old_ocell
= cell
;
1565 mg
->new_ocell
= NULL
;
1566 mg
->start_jiffies
= jiffies
;
1568 inc_io_migrations(cache
);
1569 quiesce_migration(mg
);
1572 static void discard(struct cache
*cache
, struct prealloc
*structs
,
1573 struct dm_bio_prison_cell
*cell
)
1575 struct dm_cache_migration
*mg
= prealloc_get_migration(structs
);
1579 mg
->writeback
= false;
1581 mg
->promote
= false;
1582 mg
->requeue_holder
= false;
1583 mg
->invalidate
= false;
1585 mg
->old_ocell
= NULL
;
1586 mg
->new_ocell
= cell
;
1587 mg
->start_jiffies
= jiffies
;
1589 quiesce_migration(mg
);
1592 /*----------------------------------------------------------------
1594 *--------------------------------------------------------------*/
1595 static void defer_bio(struct cache
*cache
, struct bio
*bio
)
1597 unsigned long flags
;
1599 spin_lock_irqsave(&cache
->lock
, flags
);
1600 bio_list_add(&cache
->deferred_bios
, bio
);
1601 spin_unlock_irqrestore(&cache
->lock
, flags
);
1606 static void process_flush_bio(struct cache
*cache
, struct bio
*bio
)
1608 size_t pb_data_size
= get_per_bio_data_size(cache
);
1609 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
1611 BUG_ON(bio
->bi_iter
.bi_size
);
1613 remap_to_origin(cache
, bio
);
1615 remap_to_cache(cache
, bio
, 0);
1618 * REQ_FLUSH is not directed at any particular block so we don't
1619 * need to inc_ds(). REQ_FUA's are split into a write + REQ_FLUSH
1625 static void process_discard_bio(struct cache
*cache
, struct prealloc
*structs
,
1630 struct dm_bio_prison_cell
*cell_prealloc
, *new_ocell
;
1632 calc_discard_block_range(cache
, bio
, &b
, &e
);
1638 cell_prealloc
= prealloc_get_cell(structs
);
1639 r
= bio_detain_range(cache
, dblock_to_oblock(cache
, b
), dblock_to_oblock(cache
, e
), bio
, cell_prealloc
,
1640 (cell_free_fn
) prealloc_put_cell
,
1641 structs
, &new_ocell
);
1645 discard(cache
, structs
, new_ocell
);
1648 static bool spare_migration_bandwidth(struct cache
*cache
)
1650 sector_t current_volume
= (atomic_read(&cache
->nr_io_migrations
) + 1) *
1651 cache
->sectors_per_block
;
1652 return current_volume
< cache
->migration_threshold
;
1655 static void inc_hit_counter(struct cache
*cache
, struct bio
*bio
)
1657 atomic_inc(bio_data_dir(bio
) == READ
?
1658 &cache
->stats
.read_hit
: &cache
->stats
.write_hit
);
1661 static void inc_miss_counter(struct cache
*cache
, struct bio
*bio
)
1663 atomic_inc(bio_data_dir(bio
) == READ
?
1664 &cache
->stats
.read_miss
: &cache
->stats
.write_miss
);
1667 /*----------------------------------------------------------------*/
1670 struct cache
*cache
;
1671 struct bio_list bios_for_issue
;
1672 struct bio_list unhandled_bios
;
1676 static void inc_fn(void *context
, struct dm_bio_prison_cell
*cell
)
1679 struct inc_detail
*detail
= context
;
1680 struct cache
*cache
= detail
->cache
;
1682 inc_ds(cache
, cell
->holder
, cell
);
1683 if (bio_data_dir(cell
->holder
) == WRITE
)
1684 detail
->any_writes
= true;
1686 while ((bio
= bio_list_pop(&cell
->bios
))) {
1687 if (discard_or_flush(bio
)) {
1688 bio_list_add(&detail
->unhandled_bios
, bio
);
1692 if (bio_data_dir(bio
) == WRITE
)
1693 detail
->any_writes
= true;
1695 bio_list_add(&detail
->bios_for_issue
, bio
);
1696 inc_ds(cache
, bio
, cell
);
1700 // FIXME: refactor these two
1701 static void remap_cell_to_origin_clear_discard(struct cache
*cache
,
1702 struct dm_bio_prison_cell
*cell
,
1703 dm_oblock_t oblock
, bool issue_holder
)
1706 unsigned long flags
;
1707 struct inc_detail detail
;
1709 detail
.cache
= cache
;
1710 bio_list_init(&detail
.bios_for_issue
);
1711 bio_list_init(&detail
.unhandled_bios
);
1712 detail
.any_writes
= false;
1714 spin_lock_irqsave(&cache
->lock
, flags
);
1715 dm_cell_visit_release(cache
->prison
, inc_fn
, &detail
, cell
);
1716 bio_list_merge(&cache
->deferred_bios
, &detail
.unhandled_bios
);
1717 spin_unlock_irqrestore(&cache
->lock
, flags
);
1719 remap_to_origin(cache
, cell
->holder
);
1721 issue(cache
, cell
->holder
);
1723 accounted_begin(cache
, cell
->holder
);
1725 if (detail
.any_writes
)
1726 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
1728 while ((bio
= bio_list_pop(&detail
.bios_for_issue
))) {
1729 remap_to_origin(cache
, bio
);
1734 static void remap_cell_to_cache_dirty(struct cache
*cache
, struct dm_bio_prison_cell
*cell
,
1735 dm_oblock_t oblock
, dm_cblock_t cblock
, bool issue_holder
)
1738 unsigned long flags
;
1739 struct inc_detail detail
;
1741 detail
.cache
= cache
;
1742 bio_list_init(&detail
.bios_for_issue
);
1743 bio_list_init(&detail
.unhandled_bios
);
1744 detail
.any_writes
= false;
1746 spin_lock_irqsave(&cache
->lock
, flags
);
1747 dm_cell_visit_release(cache
->prison
, inc_fn
, &detail
, cell
);
1748 bio_list_merge(&cache
->deferred_bios
, &detail
.unhandled_bios
);
1749 spin_unlock_irqrestore(&cache
->lock
, flags
);
1751 remap_to_cache(cache
, cell
->holder
, cblock
);
1753 issue(cache
, cell
->holder
);
1755 accounted_begin(cache
, cell
->holder
);
1757 if (detail
.any_writes
) {
1758 set_dirty(cache
, oblock
, cblock
);
1759 clear_discard(cache
, oblock_to_dblock(cache
, oblock
));
1762 while ((bio
= bio_list_pop(&detail
.bios_for_issue
))) {
1763 remap_to_cache(cache
, bio
, cblock
);
1768 /*----------------------------------------------------------------*/
1770 struct old_oblock_lock
{
1771 struct policy_locker locker
;
1772 struct cache
*cache
;
1773 struct prealloc
*structs
;
1774 struct dm_bio_prison_cell
*cell
;
1777 static int null_locker(struct policy_locker
*locker
, dm_oblock_t b
)
1779 /* This should never be called */
1784 static int cell_locker(struct policy_locker
*locker
, dm_oblock_t b
)
1786 struct old_oblock_lock
*l
= container_of(locker
, struct old_oblock_lock
, locker
);
1787 struct dm_bio_prison_cell
*cell_prealloc
= prealloc_get_cell(l
->structs
);
1789 return bio_detain(l
->cache
, b
, NULL
, cell_prealloc
,
1790 (cell_free_fn
) prealloc_put_cell
,
1791 l
->structs
, &l
->cell
);
1794 static void process_cell(struct cache
*cache
, struct prealloc
*structs
,
1795 struct dm_bio_prison_cell
*new_ocell
)
1798 bool release_cell
= true;
1799 struct bio
*bio
= new_ocell
->holder
;
1800 dm_oblock_t block
= get_bio_block(cache
, bio
);
1801 struct policy_result lookup_result
;
1802 bool passthrough
= passthrough_mode(&cache
->features
);
1803 bool fast_promotion
, can_migrate
;
1804 struct old_oblock_lock ool
;
1806 fast_promotion
= is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
);
1807 can_migrate
= !passthrough
&& (fast_promotion
|| spare_migration_bandwidth(cache
));
1809 ool
.locker
.fn
= cell_locker
;
1811 ool
.structs
= structs
;
1813 r
= policy_map(cache
->policy
, block
, true, can_migrate
, fast_promotion
,
1814 bio
, &ool
.locker
, &lookup_result
);
1816 if (r
== -EWOULDBLOCK
)
1817 /* migration has been denied */
1818 lookup_result
.op
= POLICY_MISS
;
1820 switch (lookup_result
.op
) {
1823 inc_miss_counter(cache
, bio
);
1826 * Passthrough always maps to the origin,
1827 * invalidating any cache blocks that are written
1831 if (bio_data_dir(bio
) == WRITE
) {
1832 atomic_inc(&cache
->stats
.demotion
);
1833 invalidate(cache
, structs
, block
, lookup_result
.cblock
, new_ocell
);
1834 release_cell
= false;
1837 /* FIXME: factor out issue_origin() */
1838 remap_to_origin_clear_discard(cache
, bio
, block
);
1839 inc_and_issue(cache
, bio
, new_ocell
);
1842 inc_hit_counter(cache
, bio
);
1844 if (bio_data_dir(bio
) == WRITE
&&
1845 writethrough_mode(&cache
->features
) &&
1846 !is_dirty(cache
, lookup_result
.cblock
)) {
1847 remap_to_origin_then_cache(cache
, bio
, block
, lookup_result
.cblock
);
1848 inc_and_issue(cache
, bio
, new_ocell
);
1851 remap_cell_to_cache_dirty(cache
, new_ocell
, block
, lookup_result
.cblock
, true);
1852 release_cell
= false;
1859 inc_miss_counter(cache
, bio
);
1860 remap_cell_to_origin_clear_discard(cache
, new_ocell
, block
, true);
1861 release_cell
= false;
1865 atomic_inc(&cache
->stats
.promotion
);
1866 promote(cache
, structs
, block
, lookup_result
.cblock
, new_ocell
);
1867 release_cell
= false;
1870 case POLICY_REPLACE
:
1871 atomic_inc(&cache
->stats
.demotion
);
1872 atomic_inc(&cache
->stats
.promotion
);
1873 demote_then_promote(cache
, structs
, lookup_result
.old_oblock
,
1874 block
, lookup_result
.cblock
,
1875 ool
.cell
, new_ocell
);
1876 release_cell
= false;
1880 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1881 cache_device_name(cache
), __func__
,
1882 (unsigned) lookup_result
.op
);
1887 cell_defer(cache
, new_ocell
, false);
1890 static void process_bio(struct cache
*cache
, struct prealloc
*structs
,
1894 dm_oblock_t block
= get_bio_block(cache
, bio
);
1895 struct dm_bio_prison_cell
*cell_prealloc
, *new_ocell
;
1898 * Check to see if that block is currently migrating.
1900 cell_prealloc
= prealloc_get_cell(structs
);
1901 r
= bio_detain(cache
, block
, bio
, cell_prealloc
,
1902 (cell_free_fn
) prealloc_put_cell
,
1903 structs
, &new_ocell
);
1907 process_cell(cache
, structs
, new_ocell
);
1910 static int need_commit_due_to_time(struct cache
*cache
)
1912 return jiffies
< cache
->last_commit_jiffies
||
1913 jiffies
> cache
->last_commit_jiffies
+ COMMIT_PERIOD
;
1917 * A non-zero return indicates read_only or fail_io mode.
1919 static int commit(struct cache
*cache
, bool clean_shutdown
)
1923 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
1926 atomic_inc(&cache
->stats
.commit_count
);
1927 r
= dm_cache_commit(cache
->cmd
, clean_shutdown
);
1929 metadata_operation_failed(cache
, "dm_cache_commit", r
);
1934 static int commit_if_needed(struct cache
*cache
)
1938 if ((cache
->commit_requested
|| need_commit_due_to_time(cache
)) &&
1939 dm_cache_changed_this_transaction(cache
->cmd
)) {
1940 r
= commit(cache
, false);
1941 cache
->commit_requested
= false;
1942 cache
->last_commit_jiffies
= jiffies
;
1948 static void process_deferred_bios(struct cache
*cache
)
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
);
1986 prealloc_free_structs(cache
, &structs
);
1989 static void process_deferred_cells(struct cache
*cache
)
1991 unsigned long flags
;
1992 struct dm_bio_prison_cell
*cell
, *tmp
;
1993 struct list_head cells
;
1994 struct prealloc structs
;
1996 memset(&structs
, 0, sizeof(structs
));
1998 INIT_LIST_HEAD(&cells
);
2000 spin_lock_irqsave(&cache
->lock
, flags
);
2001 list_splice_init(&cache
->deferred_cells
, &cells
);
2002 spin_unlock_irqrestore(&cache
->lock
, flags
);
2004 list_for_each_entry_safe(cell
, tmp
, &cells
, user_list
) {
2006 * If we've got no free migration structs, and processing
2007 * this bio might require one, we pause until there are some
2008 * prepared mappings to process.
2010 if (prealloc_data_structs(cache
, &structs
)) {
2011 spin_lock_irqsave(&cache
->lock
, flags
);
2012 list_splice(&cells
, &cache
->deferred_cells
);
2013 spin_unlock_irqrestore(&cache
->lock
, flags
);
2017 process_cell(cache
, &structs
, cell
);
2020 prealloc_free_structs(cache
, &structs
);
2023 static void process_deferred_flush_bios(struct cache
*cache
, bool submit_bios
)
2025 unsigned long flags
;
2026 struct bio_list bios
;
2029 bio_list_init(&bios
);
2031 spin_lock_irqsave(&cache
->lock
, flags
);
2032 bio_list_merge(&bios
, &cache
->deferred_flush_bios
);
2033 bio_list_init(&cache
->deferred_flush_bios
);
2034 spin_unlock_irqrestore(&cache
->lock
, flags
);
2037 * These bios have already been through inc_ds()
2039 while ((bio
= bio_list_pop(&bios
)))
2040 submit_bios
? accounted_request(cache
, bio
) : bio_io_error(bio
);
2043 static void process_deferred_writethrough_bios(struct cache
*cache
)
2045 unsigned long flags
;
2046 struct bio_list bios
;
2049 bio_list_init(&bios
);
2051 spin_lock_irqsave(&cache
->lock
, flags
);
2052 bio_list_merge(&bios
, &cache
->deferred_writethrough_bios
);
2053 bio_list_init(&cache
->deferred_writethrough_bios
);
2054 spin_unlock_irqrestore(&cache
->lock
, flags
);
2057 * These bios have already been through inc_ds()
2059 while ((bio
= bio_list_pop(&bios
)))
2060 accounted_request(cache
, bio
);
2063 static void writeback_some_dirty_blocks(struct cache
*cache
)
2068 struct prealloc structs
;
2069 struct dm_bio_prison_cell
*old_ocell
;
2070 bool busy
= !iot_idle_for(&cache
->origin_tracker
, HZ
);
2072 memset(&structs
, 0, sizeof(structs
));
2074 while (spare_migration_bandwidth(cache
)) {
2075 if (prealloc_data_structs(cache
, &structs
))
2078 r
= policy_writeback_work(cache
->policy
, &oblock
, &cblock
, busy
);
2082 r
= get_cell(cache
, oblock
, &structs
, &old_ocell
);
2084 policy_set_dirty(cache
->policy
, oblock
);
2088 writeback(cache
, &structs
, oblock
, cblock
, old_ocell
);
2091 prealloc_free_structs(cache
, &structs
);
2094 /*----------------------------------------------------------------
2096 * Dropping something from the cache *without* writing back.
2097 *--------------------------------------------------------------*/
2099 static void process_invalidation_request(struct cache
*cache
, struct invalidation_request
*req
)
2102 uint64_t begin
= from_cblock(req
->cblocks
->begin
);
2103 uint64_t end
= from_cblock(req
->cblocks
->end
);
2105 while (begin
!= end
) {
2106 r
= policy_remove_cblock(cache
->policy
, to_cblock(begin
));
2108 r
= dm_cache_remove_mapping(cache
->cmd
, to_cblock(begin
));
2110 metadata_operation_failed(cache
, "dm_cache_remove_mapping", r
);
2114 } else if (r
== -ENODATA
) {
2115 /* harmless, already unmapped */
2119 DMERR("%s: policy_remove_cblock failed", cache_device_name(cache
));
2126 cache
->commit_requested
= true;
2129 atomic_set(&req
->complete
, 1);
2131 wake_up(&req
->result_wait
);
2134 static void process_invalidation_requests(struct cache
*cache
)
2136 struct list_head list
;
2137 struct invalidation_request
*req
, *tmp
;
2139 INIT_LIST_HEAD(&list
);
2140 spin_lock(&cache
->invalidation_lock
);
2141 list_splice_init(&cache
->invalidation_requests
, &list
);
2142 spin_unlock(&cache
->invalidation_lock
);
2144 list_for_each_entry_safe (req
, tmp
, &list
, list
)
2145 process_invalidation_request(cache
, req
);
2148 /*----------------------------------------------------------------
2150 *--------------------------------------------------------------*/
2151 static bool is_quiescing(struct cache
*cache
)
2153 return atomic_read(&cache
->quiescing
);
2156 static void ack_quiescing(struct cache
*cache
)
2158 if (is_quiescing(cache
)) {
2159 atomic_inc(&cache
->quiescing_ack
);
2160 wake_up(&cache
->quiescing_wait
);
2164 static void wait_for_quiescing_ack(struct cache
*cache
)
2166 wait_event(cache
->quiescing_wait
, atomic_read(&cache
->quiescing_ack
));
2169 static void start_quiescing(struct cache
*cache
)
2171 atomic_inc(&cache
->quiescing
);
2172 wait_for_quiescing_ack(cache
);
2175 static void stop_quiescing(struct cache
*cache
)
2177 atomic_set(&cache
->quiescing
, 0);
2178 atomic_set(&cache
->quiescing_ack
, 0);
2181 static void wait_for_migrations(struct cache
*cache
)
2183 wait_event(cache
->migration_wait
, !atomic_read(&cache
->nr_allocated_migrations
));
2186 static void stop_worker(struct cache
*cache
)
2188 cancel_delayed_work(&cache
->waker
);
2189 flush_workqueue(cache
->wq
);
2192 static void requeue_deferred_cells(struct cache
*cache
)
2194 unsigned long flags
;
2195 struct list_head cells
;
2196 struct dm_bio_prison_cell
*cell
, *tmp
;
2198 INIT_LIST_HEAD(&cells
);
2199 spin_lock_irqsave(&cache
->lock
, flags
);
2200 list_splice_init(&cache
->deferred_cells
, &cells
);
2201 spin_unlock_irqrestore(&cache
->lock
, flags
);
2203 list_for_each_entry_safe(cell
, tmp
, &cells
, user_list
)
2204 cell_requeue(cache
, cell
);
2207 static void requeue_deferred_bios(struct cache
*cache
)
2210 struct bio_list bios
;
2212 bio_list_init(&bios
);
2213 bio_list_merge(&bios
, &cache
->deferred_bios
);
2214 bio_list_init(&cache
->deferred_bios
);
2216 while ((bio
= bio_list_pop(&bios
)))
2217 bio_endio(bio
, DM_ENDIO_REQUEUE
);
2220 static int more_work(struct cache
*cache
)
2222 if (is_quiescing(cache
))
2223 return !list_empty(&cache
->quiesced_migrations
) ||
2224 !list_empty(&cache
->completed_migrations
) ||
2225 !list_empty(&cache
->need_commit_migrations
);
2227 return !bio_list_empty(&cache
->deferred_bios
) ||
2228 !list_empty(&cache
->deferred_cells
) ||
2229 !bio_list_empty(&cache
->deferred_flush_bios
) ||
2230 !bio_list_empty(&cache
->deferred_writethrough_bios
) ||
2231 !list_empty(&cache
->quiesced_migrations
) ||
2232 !list_empty(&cache
->completed_migrations
) ||
2233 !list_empty(&cache
->need_commit_migrations
) ||
2237 static void do_worker(struct work_struct
*ws
)
2239 struct cache
*cache
= container_of(ws
, struct cache
, worker
);
2242 if (!is_quiescing(cache
)) {
2243 writeback_some_dirty_blocks(cache
);
2244 process_deferred_writethrough_bios(cache
);
2245 process_deferred_bios(cache
);
2246 process_deferred_cells(cache
);
2247 process_invalidation_requests(cache
);
2250 process_migrations(cache
, &cache
->quiesced_migrations
, issue_copy_or_discard
);
2251 process_migrations(cache
, &cache
->completed_migrations
, complete_migration
);
2253 if (commit_if_needed(cache
)) {
2254 process_deferred_flush_bios(cache
, false);
2255 process_migrations(cache
, &cache
->need_commit_migrations
, migration_failure
);
2257 process_deferred_flush_bios(cache
, true);
2258 process_migrations(cache
, &cache
->need_commit_migrations
,
2259 migration_success_post_commit
);
2262 ack_quiescing(cache
);
2264 } while (more_work(cache
));
2268 * We want to commit periodically so that not too much
2269 * unwritten metadata builds up.
2271 static void do_waker(struct work_struct
*ws
)
2273 struct cache
*cache
= container_of(to_delayed_work(ws
), struct cache
, waker
);
2274 policy_tick(cache
->policy
, true);
2276 queue_delayed_work(cache
->wq
, &cache
->waker
, COMMIT_PERIOD
);
2279 /*----------------------------------------------------------------*/
2281 static int is_congested(struct dm_dev
*dev
, int bdi_bits
)
2283 struct request_queue
*q
= bdev_get_queue(dev
->bdev
);
2284 return bdi_congested(&q
->backing_dev_info
, bdi_bits
);
2287 static int cache_is_congested(struct dm_target_callbacks
*cb
, int bdi_bits
)
2289 struct cache
*cache
= container_of(cb
, struct cache
, callbacks
);
2291 return is_congested(cache
->origin_dev
, bdi_bits
) ||
2292 is_congested(cache
->cache_dev
, bdi_bits
);
2295 /*----------------------------------------------------------------
2297 *--------------------------------------------------------------*/
2300 * This function gets called on the error paths of the constructor, so we
2301 * have to cope with a partially initialised struct.
2303 static void destroy(struct cache
*cache
)
2307 if (cache
->migration_pool
)
2308 mempool_destroy(cache
->migration_pool
);
2310 if (cache
->all_io_ds
)
2311 dm_deferred_set_destroy(cache
->all_io_ds
);
2314 dm_bio_prison_destroy(cache
->prison
);
2317 destroy_workqueue(cache
->wq
);
2319 if (cache
->dirty_bitset
)
2320 free_bitset(cache
->dirty_bitset
);
2322 if (cache
->discard_bitset
)
2323 free_bitset(cache
->discard_bitset
);
2326 dm_kcopyd_client_destroy(cache
->copier
);
2329 dm_cache_metadata_close(cache
->cmd
);
2331 if (cache
->metadata_dev
)
2332 dm_put_device(cache
->ti
, cache
->metadata_dev
);
2334 if (cache
->origin_dev
)
2335 dm_put_device(cache
->ti
, cache
->origin_dev
);
2337 if (cache
->cache_dev
)
2338 dm_put_device(cache
->ti
, cache
->cache_dev
);
2341 dm_cache_policy_destroy(cache
->policy
);
2343 for (i
= 0; i
< cache
->nr_ctr_args
; i
++)
2344 kfree(cache
->ctr_args
[i
]);
2345 kfree(cache
->ctr_args
);
2350 static void cache_dtr(struct dm_target
*ti
)
2352 struct cache
*cache
= ti
->private;
2357 static sector_t
get_dev_size(struct dm_dev
*dev
)
2359 return i_size_read(dev
->bdev
->bd_inode
) >> SECTOR_SHIFT
;
2362 /*----------------------------------------------------------------*/
2365 * Construct a cache device mapping.
2367 * cache <metadata dev> <cache dev> <origin dev> <block size>
2368 * <#feature args> [<feature arg>]*
2369 * <policy> <#policy args> [<policy arg>]*
2371 * metadata dev : fast device holding the persistent metadata
2372 * cache dev : fast device holding cached data blocks
2373 * origin dev : slow device holding original data blocks
2374 * block size : cache unit size in sectors
2376 * #feature args : number of feature arguments passed
2377 * feature args : writethrough. (The default is writeback.)
2379 * policy : the replacement policy to use
2380 * #policy args : an even number of policy arguments corresponding
2381 * to key/value pairs passed to the policy
2382 * policy args : key/value pairs passed to the policy
2383 * E.g. 'sequential_threshold 1024'
2384 * See cache-policies.txt for details.
2386 * Optional feature arguments are:
2387 * writethrough : write through caching that prohibits cache block
2388 * content from being different from origin block content.
2389 * Without this argument, the default behaviour is to write
2390 * back cache block contents later for performance reasons,
2391 * so they may differ from the corresponding origin blocks.
2394 struct dm_target
*ti
;
2396 struct dm_dev
*metadata_dev
;
2398 struct dm_dev
*cache_dev
;
2399 sector_t cache_sectors
;
2401 struct dm_dev
*origin_dev
;
2402 sector_t origin_sectors
;
2404 uint32_t block_size
;
2406 const char *policy_name
;
2408 const char **policy_argv
;
2410 struct cache_features features
;
2413 static void destroy_cache_args(struct cache_args
*ca
)
2415 if (ca
->metadata_dev
)
2416 dm_put_device(ca
->ti
, ca
->metadata_dev
);
2419 dm_put_device(ca
->ti
, ca
->cache_dev
);
2422 dm_put_device(ca
->ti
, ca
->origin_dev
);
2427 static bool at_least_one_arg(struct dm_arg_set
*as
, char **error
)
2430 *error
= "Insufficient args";
2437 static int parse_metadata_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2441 sector_t metadata_dev_size
;
2442 char b
[BDEVNAME_SIZE
];
2444 if (!at_least_one_arg(as
, error
))
2447 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2450 *error
= "Error opening metadata device";
2454 metadata_dev_size
= get_dev_size(ca
->metadata_dev
);
2455 if (metadata_dev_size
> DM_CACHE_METADATA_MAX_SECTORS_WARNING
)
2456 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2457 bdevname(ca
->metadata_dev
->bdev
, b
), THIN_METADATA_MAX_SECTORS
);
2462 static int parse_cache_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2467 if (!at_least_one_arg(as
, error
))
2470 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2473 *error
= "Error opening cache device";
2476 ca
->cache_sectors
= get_dev_size(ca
->cache_dev
);
2481 static int parse_origin_dev(struct cache_args
*ca
, struct dm_arg_set
*as
,
2486 if (!at_least_one_arg(as
, error
))
2489 r
= dm_get_device(ca
->ti
, dm_shift_arg(as
), FMODE_READ
| FMODE_WRITE
,
2492 *error
= "Error opening origin device";
2496 ca
->origin_sectors
= get_dev_size(ca
->origin_dev
);
2497 if (ca
->ti
->len
> ca
->origin_sectors
) {
2498 *error
= "Device size larger than cached device";
2505 static int parse_block_size(struct cache_args
*ca
, struct dm_arg_set
*as
,
2508 unsigned long block_size
;
2510 if (!at_least_one_arg(as
, error
))
2513 if (kstrtoul(dm_shift_arg(as
), 10, &block_size
) || !block_size
||
2514 block_size
< DATA_DEV_BLOCK_SIZE_MIN_SECTORS
||
2515 block_size
> DATA_DEV_BLOCK_SIZE_MAX_SECTORS
||
2516 block_size
& (DATA_DEV_BLOCK_SIZE_MIN_SECTORS
- 1)) {
2517 *error
= "Invalid data block size";
2521 if (block_size
> ca
->cache_sectors
) {
2522 *error
= "Data block size is larger than the cache device";
2526 ca
->block_size
= block_size
;
2531 static void init_features(struct cache_features
*cf
)
2533 cf
->mode
= CM_WRITE
;
2534 cf
->io_mode
= CM_IO_WRITEBACK
;
2537 static int parse_features(struct cache_args
*ca
, struct dm_arg_set
*as
,
2540 static struct dm_arg _args
[] = {
2541 {0, 1, "Invalid number of cache feature arguments"},
2547 struct cache_features
*cf
= &ca
->features
;
2551 r
= dm_read_arg_group(_args
, as
, &argc
, error
);
2556 arg
= dm_shift_arg(as
);
2558 if (!strcasecmp(arg
, "writeback"))
2559 cf
->io_mode
= CM_IO_WRITEBACK
;
2561 else if (!strcasecmp(arg
, "writethrough"))
2562 cf
->io_mode
= CM_IO_WRITETHROUGH
;
2564 else if (!strcasecmp(arg
, "passthrough"))
2565 cf
->io_mode
= CM_IO_PASSTHROUGH
;
2568 *error
= "Unrecognised cache feature requested";
2576 static int parse_policy(struct cache_args
*ca
, struct dm_arg_set
*as
,
2579 static struct dm_arg _args
[] = {
2580 {0, 1024, "Invalid number of policy arguments"},
2585 if (!at_least_one_arg(as
, error
))
2588 ca
->policy_name
= dm_shift_arg(as
);
2590 r
= dm_read_arg_group(_args
, as
, &ca
->policy_argc
, error
);
2594 ca
->policy_argv
= (const char **)as
->argv
;
2595 dm_consume_args(as
, ca
->policy_argc
);
2600 static int parse_cache_args(struct cache_args
*ca
, int argc
, char **argv
,
2604 struct dm_arg_set as
;
2609 r
= parse_metadata_dev(ca
, &as
, error
);
2613 r
= parse_cache_dev(ca
, &as
, error
);
2617 r
= parse_origin_dev(ca
, &as
, error
);
2621 r
= parse_block_size(ca
, &as
, error
);
2625 r
= parse_features(ca
, &as
, error
);
2629 r
= parse_policy(ca
, &as
, error
);
2636 /*----------------------------------------------------------------*/
2638 static struct kmem_cache
*migration_cache
;
2640 #define NOT_CORE_OPTION 1
2642 static int process_config_option(struct cache
*cache
, const char *key
, const char *value
)
2646 if (!strcasecmp(key
, "migration_threshold")) {
2647 if (kstrtoul(value
, 10, &tmp
))
2650 cache
->migration_threshold
= tmp
;
2654 return NOT_CORE_OPTION
;
2657 static int set_config_value(struct cache
*cache
, const char *key
, const char *value
)
2659 int r
= process_config_option(cache
, key
, value
);
2661 if (r
== NOT_CORE_OPTION
)
2662 r
= policy_set_config_value(cache
->policy
, key
, value
);
2665 DMWARN("bad config value for %s: %s", key
, value
);
2670 static int set_config_values(struct cache
*cache
, int argc
, const char **argv
)
2675 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2680 r
= set_config_value(cache
, argv
[0], argv
[1]);
2691 static int create_cache_policy(struct cache
*cache
, struct cache_args
*ca
,
2694 struct dm_cache_policy
*p
= dm_cache_policy_create(ca
->policy_name
,
2696 cache
->origin_sectors
,
2697 cache
->sectors_per_block
);
2699 *error
= "Error creating cache's policy";
2708 * We want the discard block size to be at least the size of the cache
2709 * block size and have no more than 2^14 discard blocks across the origin.
2711 #define MAX_DISCARD_BLOCKS (1 << 14)
2713 static bool too_many_discard_blocks(sector_t discard_block_size
,
2714 sector_t origin_size
)
2716 (void) sector_div(origin_size
, discard_block_size
);
2718 return origin_size
> MAX_DISCARD_BLOCKS
;
2721 static sector_t
calculate_discard_block_size(sector_t cache_block_size
,
2722 sector_t origin_size
)
2724 sector_t discard_block_size
= cache_block_size
;
2727 while (too_many_discard_blocks(discard_block_size
, origin_size
))
2728 discard_block_size
*= 2;
2730 return discard_block_size
;
2733 static void set_cache_size(struct cache
*cache
, dm_cblock_t size
)
2735 dm_block_t nr_blocks
= from_cblock(size
);
2737 if (nr_blocks
> (1 << 20) && cache
->cache_size
!= size
)
2738 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2739 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2740 "Please consider increasing the cache block size to reduce the overall cache block count.",
2741 (unsigned long long) nr_blocks
);
2743 cache
->cache_size
= size
;
2746 #define DEFAULT_MIGRATION_THRESHOLD 2048
2748 static int cache_create(struct cache_args
*ca
, struct cache
**result
)
2751 char **error
= &ca
->ti
->error
;
2752 struct cache
*cache
;
2753 struct dm_target
*ti
= ca
->ti
;
2754 dm_block_t origin_blocks
;
2755 struct dm_cache_metadata
*cmd
;
2756 bool may_format
= ca
->features
.mode
== CM_WRITE
;
2758 cache
= kzalloc(sizeof(*cache
), GFP_KERNEL
);
2763 ti
->private = cache
;
2764 ti
->num_flush_bios
= 2;
2765 ti
->flush_supported
= true;
2767 ti
->num_discard_bios
= 1;
2768 ti
->discards_supported
= true;
2769 ti
->discard_zeroes_data_unsupported
= true;
2770 ti
->split_discard_bios
= false;
2772 cache
->features
= ca
->features
;
2773 ti
->per_bio_data_size
= get_per_bio_data_size(cache
);
2775 cache
->callbacks
.congested_fn
= cache_is_congested
;
2776 dm_table_add_target_callbacks(ti
->table
, &cache
->callbacks
);
2778 cache
->metadata_dev
= ca
->metadata_dev
;
2779 cache
->origin_dev
= ca
->origin_dev
;
2780 cache
->cache_dev
= ca
->cache_dev
;
2782 ca
->metadata_dev
= ca
->origin_dev
= ca
->cache_dev
= NULL
;
2784 /* FIXME: factor out this whole section */
2785 origin_blocks
= cache
->origin_sectors
= ca
->origin_sectors
;
2786 origin_blocks
= block_div(origin_blocks
, ca
->block_size
);
2787 cache
->origin_blocks
= to_oblock(origin_blocks
);
2789 cache
->sectors_per_block
= ca
->block_size
;
2790 if (dm_set_target_max_io_len(ti
, cache
->sectors_per_block
)) {
2795 if (ca
->block_size
& (ca
->block_size
- 1)) {
2796 dm_block_t cache_size
= ca
->cache_sectors
;
2798 cache
->sectors_per_block_shift
= -1;
2799 cache_size
= block_div(cache_size
, ca
->block_size
);
2800 set_cache_size(cache
, to_cblock(cache_size
));
2802 cache
->sectors_per_block_shift
= __ffs(ca
->block_size
);
2803 set_cache_size(cache
, to_cblock(ca
->cache_sectors
>> cache
->sectors_per_block_shift
));
2806 r
= create_cache_policy(cache
, ca
, error
);
2810 cache
->policy_nr_args
= ca
->policy_argc
;
2811 cache
->migration_threshold
= DEFAULT_MIGRATION_THRESHOLD
;
2813 r
= set_config_values(cache
, ca
->policy_argc
, ca
->policy_argv
);
2815 *error
= "Error setting cache policy's config values";
2819 cmd
= dm_cache_metadata_open(cache
->metadata_dev
->bdev
,
2820 ca
->block_size
, may_format
,
2821 dm_cache_policy_get_hint_size(cache
->policy
));
2823 *error
= "Error creating metadata object";
2828 set_cache_mode(cache
, CM_WRITE
);
2829 if (get_cache_mode(cache
) != CM_WRITE
) {
2830 *error
= "Unable to get write access to metadata, please check/repair metadata.";
2835 if (passthrough_mode(&cache
->features
)) {
2838 r
= dm_cache_metadata_all_clean(cache
->cmd
, &all_clean
);
2840 *error
= "dm_cache_metadata_all_clean() failed";
2845 *error
= "Cannot enter passthrough mode unless all blocks are clean";
2851 spin_lock_init(&cache
->lock
);
2852 INIT_LIST_HEAD(&cache
->deferred_cells
);
2853 bio_list_init(&cache
->deferred_bios
);
2854 bio_list_init(&cache
->deferred_flush_bios
);
2855 bio_list_init(&cache
->deferred_writethrough_bios
);
2856 INIT_LIST_HEAD(&cache
->quiesced_migrations
);
2857 INIT_LIST_HEAD(&cache
->completed_migrations
);
2858 INIT_LIST_HEAD(&cache
->need_commit_migrations
);
2859 atomic_set(&cache
->nr_allocated_migrations
, 0);
2860 atomic_set(&cache
->nr_io_migrations
, 0);
2861 init_waitqueue_head(&cache
->migration_wait
);
2863 init_waitqueue_head(&cache
->quiescing_wait
);
2864 atomic_set(&cache
->quiescing
, 0);
2865 atomic_set(&cache
->quiescing_ack
, 0);
2868 atomic_set(&cache
->nr_dirty
, 0);
2869 cache
->dirty_bitset
= alloc_bitset(from_cblock(cache
->cache_size
));
2870 if (!cache
->dirty_bitset
) {
2871 *error
= "could not allocate dirty bitset";
2874 clear_bitset(cache
->dirty_bitset
, from_cblock(cache
->cache_size
));
2876 cache
->discard_block_size
=
2877 calculate_discard_block_size(cache
->sectors_per_block
,
2878 cache
->origin_sectors
);
2879 cache
->discard_nr_blocks
= to_dblock(dm_sector_div_up(cache
->origin_sectors
,
2880 cache
->discard_block_size
));
2881 cache
->discard_bitset
= alloc_bitset(from_dblock(cache
->discard_nr_blocks
));
2882 if (!cache
->discard_bitset
) {
2883 *error
= "could not allocate discard bitset";
2886 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
2888 cache
->copier
= dm_kcopyd_client_create(&dm_kcopyd_throttle
);
2889 if (IS_ERR(cache
->copier
)) {
2890 *error
= "could not create kcopyd client";
2891 r
= PTR_ERR(cache
->copier
);
2895 cache
->wq
= alloc_ordered_workqueue("dm-" DM_MSG_PREFIX
, WQ_MEM_RECLAIM
);
2897 *error
= "could not create workqueue for metadata object";
2900 INIT_WORK(&cache
->worker
, do_worker
);
2901 INIT_DELAYED_WORK(&cache
->waker
, do_waker
);
2902 cache
->last_commit_jiffies
= jiffies
;
2904 cache
->prison
= dm_bio_prison_create();
2905 if (!cache
->prison
) {
2906 *error
= "could not create bio prison";
2910 cache
->all_io_ds
= dm_deferred_set_create();
2911 if (!cache
->all_io_ds
) {
2912 *error
= "could not create all_io deferred set";
2916 cache
->migration_pool
= mempool_create_slab_pool(MIGRATION_POOL_SIZE
,
2918 if (!cache
->migration_pool
) {
2919 *error
= "Error creating cache's migration mempool";
2923 cache
->need_tick_bio
= true;
2924 cache
->sized
= false;
2925 cache
->invalidate
= false;
2926 cache
->commit_requested
= false;
2927 cache
->loaded_mappings
= false;
2928 cache
->loaded_discards
= false;
2932 atomic_set(&cache
->stats
.demotion
, 0);
2933 atomic_set(&cache
->stats
.promotion
, 0);
2934 atomic_set(&cache
->stats
.copies_avoided
, 0);
2935 atomic_set(&cache
->stats
.cache_cell_clash
, 0);
2936 atomic_set(&cache
->stats
.commit_count
, 0);
2937 atomic_set(&cache
->stats
.discard_count
, 0);
2939 spin_lock_init(&cache
->invalidation_lock
);
2940 INIT_LIST_HEAD(&cache
->invalidation_requests
);
2942 iot_init(&cache
->origin_tracker
);
2952 static int copy_ctr_args(struct cache
*cache
, int argc
, const char **argv
)
2957 copy
= kcalloc(argc
, sizeof(*copy
), GFP_KERNEL
);
2960 for (i
= 0; i
< argc
; i
++) {
2961 copy
[i
] = kstrdup(argv
[i
], GFP_KERNEL
);
2970 cache
->nr_ctr_args
= argc
;
2971 cache
->ctr_args
= copy
;
2976 static int cache_ctr(struct dm_target
*ti
, unsigned argc
, char **argv
)
2979 struct cache_args
*ca
;
2980 struct cache
*cache
= NULL
;
2982 ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
2984 ti
->error
= "Error allocating memory for cache";
2989 r
= parse_cache_args(ca
, argc
, argv
, &ti
->error
);
2993 r
= cache_create(ca
, &cache
);
2997 r
= copy_ctr_args(cache
, argc
- 3, (const char **)argv
+ 3);
3003 ti
->private = cache
;
3006 destroy_cache_args(ca
);
3010 /*----------------------------------------------------------------*/
3012 static int cache_map(struct dm_target
*ti
, struct bio
*bio
)
3014 struct cache
*cache
= ti
->private;
3017 struct dm_bio_prison_cell
*cell
= NULL
;
3018 dm_oblock_t block
= get_bio_block(cache
, bio
);
3019 size_t pb_data_size
= get_per_bio_data_size(cache
);
3020 bool can_migrate
= false;
3021 bool fast_promotion
;
3022 struct policy_result lookup_result
;
3023 struct per_bio_data
*pb
= init_per_bio_data(bio
, pb_data_size
);
3024 struct old_oblock_lock ool
;
3026 ool
.locker
.fn
= null_locker
;
3028 if (unlikely(from_oblock(block
) >= from_oblock(cache
->origin_blocks
))) {
3030 * This can only occur if the io goes to a partial block at
3031 * the end of the origin device. We don't cache these.
3032 * Just remap to the origin and carry on.
3034 remap_to_origin(cache
, bio
);
3035 accounted_begin(cache
, bio
);
3036 return DM_MAPIO_REMAPPED
;
3039 if (discard_or_flush(bio
)) {
3040 defer_bio(cache
, bio
);
3041 return DM_MAPIO_SUBMITTED
;
3045 * Check to see if that block is currently migrating.
3047 cell
= alloc_prison_cell(cache
);
3049 defer_bio(cache
, bio
);
3050 return DM_MAPIO_SUBMITTED
;
3053 r
= bio_detain(cache
, block
, bio
, cell
,
3054 (cell_free_fn
) free_prison_cell
,
3058 defer_bio(cache
, bio
);
3060 return DM_MAPIO_SUBMITTED
;
3063 fast_promotion
= is_discarded_oblock(cache
, block
) || bio_writes_complete_block(cache
, bio
);
3065 r
= policy_map(cache
->policy
, block
, false, can_migrate
, fast_promotion
,
3066 bio
, &ool
.locker
, &lookup_result
);
3067 if (r
== -EWOULDBLOCK
) {
3068 cell_defer(cache
, cell
, true);
3069 return DM_MAPIO_SUBMITTED
;
3072 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3073 cache_device_name(cache
), r
);
3074 cell_defer(cache
, cell
, false);
3076 return DM_MAPIO_SUBMITTED
;
3079 r
= DM_MAPIO_REMAPPED
;
3080 switch (lookup_result
.op
) {
3082 if (passthrough_mode(&cache
->features
)) {
3083 if (bio_data_dir(bio
) == WRITE
) {
3085 * We need to invalidate this block, so
3086 * defer for the worker thread.
3088 cell_defer(cache
, cell
, true);
3089 r
= DM_MAPIO_SUBMITTED
;
3092 inc_miss_counter(cache
, bio
);
3093 remap_to_origin_clear_discard(cache
, bio
, block
);
3094 accounted_begin(cache
, bio
);
3095 inc_ds(cache
, bio
, cell
);
3096 // FIXME: we want to remap hits or misses straight
3097 // away rather than passing over to the worker.
3098 cell_defer(cache
, cell
, false);
3102 inc_hit_counter(cache
, bio
);
3103 if (bio_data_dir(bio
) == WRITE
&& writethrough_mode(&cache
->features
) &&
3104 !is_dirty(cache
, lookup_result
.cblock
)) {
3105 remap_to_origin_then_cache(cache
, bio
, block
, lookup_result
.cblock
);
3106 accounted_begin(cache
, bio
);
3107 inc_ds(cache
, bio
, cell
);
3108 cell_defer(cache
, cell
, false);
3111 remap_cell_to_cache_dirty(cache
, cell
, block
, lookup_result
.cblock
, false);
3116 inc_miss_counter(cache
, bio
);
3117 if (pb
->req_nr
!= 0) {
3119 * This is a duplicate writethrough io that is no
3120 * longer needed because the block has been demoted.
3123 // FIXME: remap everything as a miss
3124 cell_defer(cache
, cell
, false);
3125 r
= DM_MAPIO_SUBMITTED
;
3128 remap_cell_to_origin_clear_discard(cache
, cell
, block
, false);
3132 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3133 cache_device_name(cache
), __func__
,
3134 (unsigned) lookup_result
.op
);
3135 cell_defer(cache
, cell
, false);
3137 r
= DM_MAPIO_SUBMITTED
;
3143 static int cache_end_io(struct dm_target
*ti
, struct bio
*bio
, int error
)
3145 struct cache
*cache
= ti
->private;
3146 unsigned long flags
;
3147 size_t pb_data_size
= get_per_bio_data_size(cache
);
3148 struct per_bio_data
*pb
= get_per_bio_data(bio
, pb_data_size
);
3151 policy_tick(cache
->policy
, false);
3153 spin_lock_irqsave(&cache
->lock
, flags
);
3154 cache
->need_tick_bio
= true;
3155 spin_unlock_irqrestore(&cache
->lock
, flags
);
3158 check_for_quiesced_migrations(cache
, pb
);
3159 accounted_complete(cache
, bio
);
3164 static int write_dirty_bitset(struct cache
*cache
)
3168 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3171 for (i
= 0; i
< from_cblock(cache
->cache_size
); i
++) {
3172 r
= dm_cache_set_dirty(cache
->cmd
, to_cblock(i
),
3173 is_dirty(cache
, to_cblock(i
)));
3175 metadata_operation_failed(cache
, "dm_cache_set_dirty", r
);
3183 static int write_discard_bitset(struct cache
*cache
)
3187 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3190 r
= dm_cache_discard_bitset_resize(cache
->cmd
, cache
->discard_block_size
,
3191 cache
->discard_nr_blocks
);
3193 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache
));
3194 metadata_operation_failed(cache
, "dm_cache_discard_bitset_resize", r
);
3198 for (i
= 0; i
< from_dblock(cache
->discard_nr_blocks
); i
++) {
3199 r
= dm_cache_set_discard(cache
->cmd
, to_dblock(i
),
3200 is_discarded(cache
, to_dblock(i
)));
3202 metadata_operation_failed(cache
, "dm_cache_set_discard", r
);
3210 static int write_hints(struct cache
*cache
)
3214 if (get_cache_mode(cache
) >= CM_READ_ONLY
)
3217 r
= dm_cache_write_hints(cache
->cmd
, cache
->policy
);
3219 metadata_operation_failed(cache
, "dm_cache_write_hints", r
);
3227 * returns true on success
3229 static bool sync_metadata(struct cache
*cache
)
3233 r1
= write_dirty_bitset(cache
);
3235 DMERR("%s: could not write dirty bitset", cache_device_name(cache
));
3237 r2
= write_discard_bitset(cache
);
3239 DMERR("%s: could not write discard bitset", cache_device_name(cache
));
3243 r3
= write_hints(cache
);
3245 DMERR("%s: could not write hints", cache_device_name(cache
));
3248 * If writing the above metadata failed, we still commit, but don't
3249 * set the clean shutdown flag. This will effectively force every
3250 * dirty bit to be set on reload.
3252 r4
= commit(cache
, !r1
&& !r2
&& !r3
);
3254 DMERR("%s: could not write cache metadata", cache_device_name(cache
));
3256 return !r1
&& !r2
&& !r3
&& !r4
;
3259 static void cache_postsuspend(struct dm_target
*ti
)
3261 struct cache
*cache
= ti
->private;
3263 start_quiescing(cache
);
3264 wait_for_migrations(cache
);
3266 requeue_deferred_bios(cache
);
3267 requeue_deferred_cells(cache
);
3268 stop_quiescing(cache
);
3270 if (get_cache_mode(cache
) == CM_WRITE
)
3271 (void) sync_metadata(cache
);
3274 static int load_mapping(void *context
, dm_oblock_t oblock
, dm_cblock_t cblock
,
3275 bool dirty
, uint32_t hint
, bool hint_valid
)
3278 struct cache
*cache
= context
;
3280 r
= policy_load_mapping(cache
->policy
, oblock
, cblock
, hint
, hint_valid
);
3285 set_dirty(cache
, oblock
, cblock
);
3287 clear_dirty(cache
, oblock
, cblock
);
3293 * The discard block size in the on disk metadata is not
3294 * neccessarily the same as we're currently using. So we have to
3295 * be careful to only set the discarded attribute if we know it
3296 * covers a complete block of the new size.
3298 struct discard_load_info
{
3299 struct cache
*cache
;
3302 * These blocks are sized using the on disk dblock size, rather
3303 * than the current one.
3305 dm_block_t block_size
;
3306 dm_block_t discard_begin
, discard_end
;
3309 static void discard_load_info_init(struct cache
*cache
,
3310 struct discard_load_info
*li
)
3313 li
->discard_begin
= li
->discard_end
= 0;
3316 static void set_discard_range(struct discard_load_info
*li
)
3320 if (li
->discard_begin
== li
->discard_end
)
3324 * Convert to sectors.
3326 b
= li
->discard_begin
* li
->block_size
;
3327 e
= li
->discard_end
* li
->block_size
;
3330 * Then convert back to the current dblock size.
3332 b
= dm_sector_div_up(b
, li
->cache
->discard_block_size
);
3333 sector_div(e
, li
->cache
->discard_block_size
);
3336 * The origin may have shrunk, so we need to check we're still in
3339 if (e
> from_dblock(li
->cache
->discard_nr_blocks
))
3340 e
= from_dblock(li
->cache
->discard_nr_blocks
);
3343 set_discard(li
->cache
, to_dblock(b
));
3346 static int load_discard(void *context
, sector_t discard_block_size
,
3347 dm_dblock_t dblock
, bool discard
)
3349 struct discard_load_info
*li
= context
;
3351 li
->block_size
= discard_block_size
;
3354 if (from_dblock(dblock
) == li
->discard_end
)
3356 * We're already in a discard range, just extend it.
3358 li
->discard_end
= li
->discard_end
+ 1ULL;
3362 * Emit the old range and start a new one.
3364 set_discard_range(li
);
3365 li
->discard_begin
= from_dblock(dblock
);
3366 li
->discard_end
= li
->discard_begin
+ 1ULL;
3369 set_discard_range(li
);
3370 li
->discard_begin
= li
->discard_end
= 0;
3376 static dm_cblock_t
get_cache_dev_size(struct cache
*cache
)
3378 sector_t size
= get_dev_size(cache
->cache_dev
);
3379 (void) sector_div(size
, cache
->sectors_per_block
);
3380 return to_cblock(size
);
3383 static bool can_resize(struct cache
*cache
, dm_cblock_t new_size
)
3385 if (from_cblock(new_size
) > from_cblock(cache
->cache_size
))
3389 * We can't drop a dirty block when shrinking the cache.
3391 while (from_cblock(new_size
) < from_cblock(cache
->cache_size
)) {
3392 new_size
= to_cblock(from_cblock(new_size
) + 1);
3393 if (is_dirty(cache
, new_size
)) {
3394 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3395 cache_device_name(cache
),
3396 (unsigned long long) from_cblock(new_size
));
3404 static int resize_cache_dev(struct cache
*cache
, dm_cblock_t new_size
)
3408 r
= dm_cache_resize(cache
->cmd
, new_size
);
3410 DMERR("%s: could not resize cache metadata", cache_device_name(cache
));
3411 metadata_operation_failed(cache
, "dm_cache_resize", r
);
3415 set_cache_size(cache
, new_size
);
3420 static int cache_preresume(struct dm_target
*ti
)
3423 struct cache
*cache
= ti
->private;
3424 dm_cblock_t csize
= get_cache_dev_size(cache
);
3427 * Check to see if the cache has resized.
3429 if (!cache
->sized
) {
3430 r
= resize_cache_dev(cache
, csize
);
3434 cache
->sized
= true;
3436 } else if (csize
!= cache
->cache_size
) {
3437 if (!can_resize(cache
, csize
))
3440 r
= resize_cache_dev(cache
, csize
);
3445 if (!cache
->loaded_mappings
) {
3446 r
= dm_cache_load_mappings(cache
->cmd
, cache
->policy
,
3447 load_mapping
, cache
);
3449 DMERR("%s: could not load cache mappings", cache_device_name(cache
));
3450 metadata_operation_failed(cache
, "dm_cache_load_mappings", r
);
3454 cache
->loaded_mappings
= true;
3457 if (!cache
->loaded_discards
) {
3458 struct discard_load_info li
;
3461 * The discard bitset could have been resized, or the
3462 * discard block size changed. To be safe we start by
3463 * setting every dblock to not discarded.
3465 clear_bitset(cache
->discard_bitset
, from_dblock(cache
->discard_nr_blocks
));
3467 discard_load_info_init(cache
, &li
);
3468 r
= dm_cache_load_discards(cache
->cmd
, load_discard
, &li
);
3470 DMERR("%s: could not load origin discards", cache_device_name(cache
));
3471 metadata_operation_failed(cache
, "dm_cache_load_discards", r
);
3474 set_discard_range(&li
);
3476 cache
->loaded_discards
= true;
3482 static void cache_resume(struct dm_target
*ti
)
3484 struct cache
*cache
= ti
->private;
3486 cache
->need_tick_bio
= true;
3487 do_waker(&cache
->waker
.work
);
3493 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3494 * <cache block size> <#used cache blocks>/<#total cache blocks>
3495 * <#read hits> <#read misses> <#write hits> <#write misses>
3496 * <#demotions> <#promotions> <#dirty>
3497 * <#features> <features>*
3498 * <#core args> <core args>
3499 * <policy name> <#policy args> <policy args>* <cache metadata mode>
3501 static void cache_status(struct dm_target
*ti
, status_type_t type
,
3502 unsigned status_flags
, char *result
, unsigned maxlen
)
3507 dm_block_t nr_free_blocks_metadata
= 0;
3508 dm_block_t nr_blocks_metadata
= 0;
3509 char buf
[BDEVNAME_SIZE
];
3510 struct cache
*cache
= ti
->private;
3511 dm_cblock_t residency
;
3514 case STATUSTYPE_INFO
:
3515 if (get_cache_mode(cache
) == CM_FAIL
) {
3520 /* Commit to ensure statistics aren't out-of-date */
3521 if (!(status_flags
& DM_STATUS_NOFLUSH_FLAG
) && !dm_suspended(ti
))
3522 (void) commit(cache
, false);
3524 r
= dm_cache_get_free_metadata_block_count(cache
->cmd
, &nr_free_blocks_metadata
);
3526 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3527 cache_device_name(cache
), r
);
3531 r
= dm_cache_get_metadata_dev_size(cache
->cmd
, &nr_blocks_metadata
);
3533 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3534 cache_device_name(cache
), r
);
3538 residency
= policy_residency(cache
->policy
);
3540 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3541 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE
,
3542 (unsigned long long)(nr_blocks_metadata
- nr_free_blocks_metadata
),
3543 (unsigned long long)nr_blocks_metadata
,
3544 cache
->sectors_per_block
,
3545 (unsigned long long) from_cblock(residency
),
3546 (unsigned long long) from_cblock(cache
->cache_size
),
3547 (unsigned) atomic_read(&cache
->stats
.read_hit
),
3548 (unsigned) atomic_read(&cache
->stats
.read_miss
),
3549 (unsigned) atomic_read(&cache
->stats
.write_hit
),
3550 (unsigned) atomic_read(&cache
->stats
.write_miss
),
3551 (unsigned) atomic_read(&cache
->stats
.demotion
),
3552 (unsigned) atomic_read(&cache
->stats
.promotion
),
3553 (unsigned long) atomic_read(&cache
->nr_dirty
));
3555 if (writethrough_mode(&cache
->features
))
3556 DMEMIT("1 writethrough ");
3558 else if (passthrough_mode(&cache
->features
))
3559 DMEMIT("1 passthrough ");
3561 else if (writeback_mode(&cache
->features
))
3562 DMEMIT("1 writeback ");
3565 DMERR("%s: internal error: unknown io mode: %d",
3566 cache_device_name(cache
), (int) cache
->features
.io_mode
);
3570 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache
->migration_threshold
);
3572 DMEMIT("%s ", dm_cache_policy_get_name(cache
->policy
));
3574 r
= policy_emit_config_values(cache
->policy
, result
, maxlen
, &sz
);
3576 DMERR("%s: policy_emit_config_values returned %d",
3577 cache_device_name(cache
), r
);
3580 if (get_cache_mode(cache
) == CM_READ_ONLY
)
3587 case STATUSTYPE_TABLE
:
3588 format_dev_t(buf
, cache
->metadata_dev
->bdev
->bd_dev
);
3590 format_dev_t(buf
, cache
->cache_dev
->bdev
->bd_dev
);
3592 format_dev_t(buf
, cache
->origin_dev
->bdev
->bd_dev
);
3595 for (i
= 0; i
< cache
->nr_ctr_args
- 1; i
++)
3596 DMEMIT(" %s", cache
->ctr_args
[i
]);
3597 if (cache
->nr_ctr_args
)
3598 DMEMIT(" %s", cache
->ctr_args
[cache
->nr_ctr_args
- 1]);
3608 * A cache block range can take two forms:
3610 * i) A single cblock, eg. '3456'
3611 * ii) A begin and end cblock with dots between, eg. 123-234
3613 static int parse_cblock_range(struct cache
*cache
, const char *str
,
3614 struct cblock_range
*result
)
3621 * Try and parse form (ii) first.
3623 r
= sscanf(str
, "%llu-%llu%c", &b
, &e
, &dummy
);
3628 result
->begin
= to_cblock(b
);
3629 result
->end
= to_cblock(e
);
3634 * That didn't work, try form (i).
3636 r
= sscanf(str
, "%llu%c", &b
, &dummy
);
3641 result
->begin
= to_cblock(b
);
3642 result
->end
= to_cblock(from_cblock(result
->begin
) + 1u);
3646 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache
), str
);
3650 static int validate_cblock_range(struct cache
*cache
, struct cblock_range
*range
)
3652 uint64_t b
= from_cblock(range
->begin
);
3653 uint64_t e
= from_cblock(range
->end
);
3654 uint64_t n
= from_cblock(cache
->cache_size
);
3657 DMERR("%s: begin cblock out of range: %llu >= %llu",
3658 cache_device_name(cache
), b
, n
);
3663 DMERR("%s: end cblock out of range: %llu > %llu",
3664 cache_device_name(cache
), e
, n
);
3669 DMERR("%s: invalid cblock range: %llu >= %llu",
3670 cache_device_name(cache
), b
, e
);
3677 static int request_invalidation(struct cache
*cache
, struct cblock_range
*range
)
3679 struct invalidation_request req
;
3681 INIT_LIST_HEAD(&req
.list
);
3682 req
.cblocks
= range
;
3683 atomic_set(&req
.complete
, 0);
3685 init_waitqueue_head(&req
.result_wait
);
3687 spin_lock(&cache
->invalidation_lock
);
3688 list_add(&req
.list
, &cache
->invalidation_requests
);
3689 spin_unlock(&cache
->invalidation_lock
);
3692 wait_event(req
.result_wait
, atomic_read(&req
.complete
));
3696 static int process_invalidate_cblocks_message(struct cache
*cache
, unsigned count
,
3697 const char **cblock_ranges
)
3701 struct cblock_range range
;
3703 if (!passthrough_mode(&cache
->features
)) {
3704 DMERR("%s: cache has to be in passthrough mode for invalidation",
3705 cache_device_name(cache
));
3709 for (i
= 0; i
< count
; i
++) {
3710 r
= parse_cblock_range(cache
, cblock_ranges
[i
], &range
);
3714 r
= validate_cblock_range(cache
, &range
);
3719 * Pass begin and end origin blocks to the worker and wake it.
3721 r
= request_invalidation(cache
, &range
);
3733 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3735 * The key migration_threshold is supported by the cache target core.
3737 static int cache_message(struct dm_target
*ti
, unsigned argc
, char **argv
)
3739 struct cache
*cache
= ti
->private;
3744 if (get_cache_mode(cache
) >= CM_READ_ONLY
) {
3745 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3746 cache_device_name(cache
));
3750 if (!strcasecmp(argv
[0], "invalidate_cblocks"))
3751 return process_invalidate_cblocks_message(cache
, argc
- 1, (const char **) argv
+ 1);
3756 return set_config_value(cache
, argv
[0], argv
[1]);
3759 static int cache_iterate_devices(struct dm_target
*ti
,
3760 iterate_devices_callout_fn fn
, void *data
)
3763 struct cache
*cache
= ti
->private;
3765 r
= fn(ti
, cache
->cache_dev
, 0, get_dev_size(cache
->cache_dev
), data
);
3767 r
= fn(ti
, cache
->origin_dev
, 0, ti
->len
, data
);
3773 * We assume I/O is going to the origin (which is the volume
3774 * more likely to have restrictions e.g. by being striped).
3775 * (Looking up the exact location of the data would be expensive
3776 * and could always be out of date by the time the bio is submitted.)
3778 static int cache_bvec_merge(struct dm_target
*ti
,
3779 struct bvec_merge_data
*bvm
,
3780 struct bio_vec
*biovec
, int max_size
)
3782 struct cache
*cache
= ti
->private;
3783 struct request_queue
*q
= bdev_get_queue(cache
->origin_dev
->bdev
);
3785 if (!q
->merge_bvec_fn
)
3788 bvm
->bi_bdev
= cache
->origin_dev
->bdev
;
3789 return min(max_size
, q
->merge_bvec_fn(q
, bvm
, biovec
));
3792 static void set_discard_limits(struct cache
*cache
, struct queue_limits
*limits
)
3795 * FIXME: these limits may be incompatible with the cache device
3797 limits
->max_discard_sectors
= min_t(sector_t
, cache
->discard_block_size
* 1024,
3798 cache
->origin_sectors
);
3799 limits
->discard_granularity
= cache
->discard_block_size
<< SECTOR_SHIFT
;
3802 static void cache_io_hints(struct dm_target
*ti
, struct queue_limits
*limits
)
3804 struct cache
*cache
= ti
->private;
3805 uint64_t io_opt_sectors
= limits
->io_opt
>> SECTOR_SHIFT
;
3808 * If the system-determined stacked limits are compatible with the
3809 * cache's blocksize (io_opt is a factor) do not override them.
3811 if (io_opt_sectors
< cache
->sectors_per_block
||
3812 do_div(io_opt_sectors
, cache
->sectors_per_block
)) {
3813 blk_limits_io_min(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3814 blk_limits_io_opt(limits
, cache
->sectors_per_block
<< SECTOR_SHIFT
);
3816 set_discard_limits(cache
, limits
);
3819 /*----------------------------------------------------------------*/
3821 static struct target_type cache_target
= {
3823 .version
= {1, 7, 0},
3824 .module
= THIS_MODULE
,
3828 .end_io
= cache_end_io
,
3829 .postsuspend
= cache_postsuspend
,
3830 .preresume
= cache_preresume
,
3831 .resume
= cache_resume
,
3832 .status
= cache_status
,
3833 .message
= cache_message
,
3834 .iterate_devices
= cache_iterate_devices
,
3835 .merge
= cache_bvec_merge
,
3836 .io_hints
= cache_io_hints
,
3839 static int __init
dm_cache_init(void)
3843 r
= dm_register_target(&cache_target
);
3845 DMERR("cache target registration failed: %d", r
);
3849 migration_cache
= KMEM_CACHE(dm_cache_migration
, 0);
3850 if (!migration_cache
) {
3851 dm_unregister_target(&cache_target
);
3858 static void __exit
dm_cache_exit(void)
3860 dm_unregister_target(&cache_target
);
3861 kmem_cache_destroy(migration_cache
);
3864 module_init(dm_cache_init
);
3865 module_exit(dm_cache_exit
);
3867 MODULE_DESCRIPTION(DM_NAME
" cache target");
3868 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3869 MODULE_LICENSE("GPL");