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/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/slab.h>
18 #include <linux/vmalloc.h>
20 #define DM_MSG_PREFIX "cache"
22 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
23 "A percentage of time allocated for copying to and/or from cache");
25 /*----------------------------------------------------------------*/
30 * oblock: index of an origin block
31 * cblock: index of a cache block
32 * promotion: movement of a block from origin to cache
33 * demotion: movement of a block from cache to origin
34 * migration: movement of a block between the origin and cache device,
38 /*----------------------------------------------------------------*/
40 static size_t bitset_size_in_bytes(unsigned nr_entries)
42 return sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
45 static unsigned long *alloc_bitset(unsigned nr_entries)
47 size_t s = bitset_size_in_bytes(nr_entries);
51 static void clear_bitset(void *bitset, unsigned nr_entries)
53 size_t s = bitset_size_in_bytes(nr_entries);
57 static void free_bitset(unsigned long *bits)
62 /*----------------------------------------------------------------*/
65 * There are a couple of places where we let a bio run, but want to do some
66 * work before calling its endio function. We do this by temporarily
67 * changing the endio fn.
70 bio_end_io_t *bi_end_io;
74 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
75 bio_end_io_t *bi_end_io, void *bi_private)
77 h->bi_end_io = bio->bi_end_io;
78 h->bi_private = bio->bi_private;
80 bio->bi_end_io = bi_end_io;
81 bio->bi_private = bi_private;
84 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
86 bio->bi_end_io = h->bi_end_io;
87 bio->bi_private = h->bi_private;
90 * Must bump bi_remaining to allow bio to complete with
93 atomic_inc(&bio->bi_remaining);
96 /*----------------------------------------------------------------*/
98 #define PRISON_CELLS 1024
99 #define MIGRATION_POOL_SIZE 128
100 #define COMMIT_PERIOD HZ
101 #define MIGRATION_COUNT_WINDOW 10
104 * The block size of the device holding cache data must be
105 * between 32KB and 1GB.
107 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
108 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
111 * FIXME: the cache is read/write for the time being.
113 enum cache_metadata_mode {
114 CM_WRITE, /* metadata may be changed */
115 CM_READ_ONLY, /* metadata may not be changed */
120 * Data is written to cached blocks only. These blocks are marked
121 * dirty. If you lose the cache device you will lose data.
122 * Potential performance increase for both reads and writes.
127 * Data is written to both cache and origin. Blocks are never
128 * dirty. Potential performance benfit for reads only.
133 * A degraded mode useful for various cache coherency situations
134 * (eg, rolling back snapshots). Reads and writes always go to the
135 * origin. If a write goes to a cached oblock, then the cache
136 * block is invalidated.
141 struct cache_features {
142 enum cache_metadata_mode mode;
143 enum cache_io_mode io_mode;
153 atomic_t copies_avoided;
154 atomic_t cache_cell_clash;
155 atomic_t commit_count;
156 atomic_t discard_count;
160 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
161 * the one-past-the-end value.
163 struct cblock_range {
168 struct invalidation_request {
169 struct list_head list;
170 struct cblock_range *cblocks;
175 wait_queue_head_t result_wait;
179 struct dm_target *ti;
180 struct dm_target_callbacks callbacks;
182 struct dm_cache_metadata *cmd;
185 * Metadata is written to this device.
187 struct dm_dev *metadata_dev;
190 * The slower of the two data devices. Typically a spindle.
192 struct dm_dev *origin_dev;
195 * The faster of the two data devices. Typically an SSD.
197 struct dm_dev *cache_dev;
200 * Size of the origin device in _complete_ blocks and native sectors.
202 dm_oblock_t origin_blocks;
203 sector_t origin_sectors;
206 * Size of the cache device in blocks.
208 dm_cblock_t cache_size;
211 * Fields for converting from sectors to blocks.
213 uint32_t sectors_per_block;
214 int sectors_per_block_shift;
217 struct bio_list deferred_bios;
218 struct bio_list deferred_flush_bios;
219 struct bio_list deferred_writethrough_bios;
220 struct list_head quiesced_migrations;
221 struct list_head completed_migrations;
222 struct list_head need_commit_migrations;
223 sector_t migration_threshold;
224 wait_queue_head_t migration_wait;
225 atomic_t nr_migrations;
227 wait_queue_head_t quiescing_wait;
229 atomic_t quiescing_ack;
232 * cache_size entries, dirty if set
235 unsigned long *dirty_bitset;
238 * origin_blocks entries, discarded if set.
240 dm_oblock_t discard_nr_blocks;
241 unsigned long *discard_bitset;
244 * Rather than reconstructing the table line for the status we just
245 * save it and regurgitate.
247 unsigned nr_ctr_args;
248 const char **ctr_args;
250 struct dm_kcopyd_client *copier;
251 struct workqueue_struct *wq;
252 struct work_struct worker;
254 struct delayed_work waker;
255 unsigned long last_commit_jiffies;
257 struct dm_bio_prison *prison;
258 struct dm_deferred_set *all_io_ds;
260 mempool_t *migration_pool;
261 struct dm_cache_migration *next_migration;
263 struct dm_cache_policy *policy;
264 unsigned policy_nr_args;
266 bool need_tick_bio:1;
269 bool commit_requested:1;
270 bool loaded_mappings:1;
271 bool loaded_discards:1;
274 * Cache features such as write-through.
276 struct cache_features features;
278 struct cache_stats stats;
281 * Invalidation fields.
283 spinlock_t invalidation_lock;
284 struct list_head invalidation_requests;
287 struct per_bio_data {
290 struct dm_deferred_entry *all_io_entry;
291 struct dm_hook_info hook_info;
294 * writethrough fields. These MUST remain at the end of this
295 * structure and the 'cache' member must be the first as it
296 * is used to determine the offset of the writethrough fields.
300 struct dm_bio_details bio_details;
303 struct dm_cache_migration {
304 struct list_head list;
307 unsigned long start_jiffies;
308 dm_oblock_t old_oblock;
309 dm_oblock_t new_oblock;
316 bool requeue_holder:1;
319 struct dm_bio_prison_cell *old_ocell;
320 struct dm_bio_prison_cell *new_ocell;
324 * Processing a bio in the worker thread may require these memory
325 * allocations. We prealloc to avoid deadlocks (the same worker thread
326 * frees them back to the mempool).
329 struct dm_cache_migration *mg;
330 struct dm_bio_prison_cell *cell1;
331 struct dm_bio_prison_cell *cell2;
334 static void wake_worker(struct cache *cache)
336 queue_work(cache->wq, &cache->worker);
339 /*----------------------------------------------------------------*/
341 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
343 /* FIXME: change to use a local slab. */
344 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
347 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
349 dm_bio_prison_free_cell(cache->prison, cell);
352 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
355 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
361 p->cell1 = alloc_prison_cell(cache);
367 p->cell2 = alloc_prison_cell(cache);
375 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
378 free_prison_cell(cache, p->cell2);
381 free_prison_cell(cache, p->cell1);
384 mempool_free(p->mg, cache->migration_pool);
387 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
389 struct dm_cache_migration *mg = p->mg;
398 * You must have a cell within the prealloc struct to return. If not this
399 * function will BUG() rather than returning NULL.
401 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
403 struct dm_bio_prison_cell *r = NULL;
409 } else if (p->cell2) {
419 * You can't have more than two cells in a prealloc struct. BUG() will be
420 * called if you try and overfill.
422 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
434 /*----------------------------------------------------------------*/
436 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
440 key->block = from_oblock(oblock);
444 * The caller hands in a preallocated cell, and a free function for it.
445 * The cell will be freed if there's an error, or if it wasn't used because
446 * a cell with that key already exists.
448 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
450 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
451 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
452 cell_free_fn free_fn, void *free_context,
453 struct dm_bio_prison_cell **cell_result)
456 struct dm_cell_key key;
458 build_key(oblock, &key);
459 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
461 free_fn(free_context, cell_prealloc);
466 static int get_cell(struct cache *cache,
468 struct prealloc *structs,
469 struct dm_bio_prison_cell **cell_result)
472 struct dm_cell_key key;
473 struct dm_bio_prison_cell *cell_prealloc;
475 cell_prealloc = prealloc_get_cell(structs);
477 build_key(oblock, &key);
478 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
480 prealloc_put_cell(structs, cell_prealloc);
485 /*----------------------------------------------------------------*/
487 static bool is_dirty(struct cache *cache, dm_cblock_t b)
489 return test_bit(from_cblock(b), cache->dirty_bitset);
492 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
494 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
495 atomic_inc(&cache->nr_dirty);
496 policy_set_dirty(cache->policy, oblock);
500 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
502 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
503 policy_clear_dirty(cache->policy, oblock);
504 if (atomic_dec_return(&cache->nr_dirty) == 0)
505 dm_table_event(cache->ti->table);
509 /*----------------------------------------------------------------*/
511 static bool block_size_is_power_of_two(struct cache *cache)
513 return cache->sectors_per_block_shift >= 0;
516 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
517 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
520 static dm_block_t block_div(dm_block_t b, uint32_t n)
527 static void set_discard(struct cache *cache, dm_oblock_t b)
531 atomic_inc(&cache->stats.discard_count);
533 spin_lock_irqsave(&cache->lock, flags);
534 set_bit(from_oblock(b), cache->discard_bitset);
535 spin_unlock_irqrestore(&cache->lock, flags);
538 static void clear_discard(struct cache *cache, dm_oblock_t b)
542 spin_lock_irqsave(&cache->lock, flags);
543 clear_bit(from_oblock(b), cache->discard_bitset);
544 spin_unlock_irqrestore(&cache->lock, flags);
547 static bool is_discarded(struct cache *cache, dm_oblock_t b)
552 spin_lock_irqsave(&cache->lock, flags);
553 r = test_bit(from_oblock(b), cache->discard_bitset);
554 spin_unlock_irqrestore(&cache->lock, flags);
559 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
564 spin_lock_irqsave(&cache->lock, flags);
565 r = test_bit(from_oblock(b), cache->discard_bitset);
566 spin_unlock_irqrestore(&cache->lock, flags);
571 /*----------------------------------------------------------------*/
573 static void load_stats(struct cache *cache)
575 struct dm_cache_statistics stats;
577 dm_cache_metadata_get_stats(cache->cmd, &stats);
578 atomic_set(&cache->stats.read_hit, stats.read_hits);
579 atomic_set(&cache->stats.read_miss, stats.read_misses);
580 atomic_set(&cache->stats.write_hit, stats.write_hits);
581 atomic_set(&cache->stats.write_miss, stats.write_misses);
584 static void save_stats(struct cache *cache)
586 struct dm_cache_statistics stats;
588 stats.read_hits = atomic_read(&cache->stats.read_hit);
589 stats.read_misses = atomic_read(&cache->stats.read_miss);
590 stats.write_hits = atomic_read(&cache->stats.write_hit);
591 stats.write_misses = atomic_read(&cache->stats.write_miss);
593 dm_cache_metadata_set_stats(cache->cmd, &stats);
596 /*----------------------------------------------------------------
598 *--------------------------------------------------------------*/
601 * If using writeback, leave out struct per_bio_data's writethrough fields.
603 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
604 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
606 static bool writethrough_mode(struct cache_features *f)
608 return f->io_mode == CM_IO_WRITETHROUGH;
611 static bool writeback_mode(struct cache_features *f)
613 return f->io_mode == CM_IO_WRITEBACK;
616 static bool passthrough_mode(struct cache_features *f)
618 return f->io_mode == CM_IO_PASSTHROUGH;
621 static size_t get_per_bio_data_size(struct cache *cache)
623 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
626 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
628 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
633 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
635 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
638 pb->req_nr = dm_bio_get_target_bio_nr(bio);
639 pb->all_io_entry = NULL;
644 /*----------------------------------------------------------------
646 *--------------------------------------------------------------*/
647 static void remap_to_origin(struct cache *cache, struct bio *bio)
649 bio->bi_bdev = cache->origin_dev->bdev;
652 static void remap_to_cache(struct cache *cache, struct bio *bio,
655 sector_t bi_sector = bio->bi_iter.bi_sector;
656 sector_t block = from_cblock(cblock);
658 bio->bi_bdev = cache->cache_dev->bdev;
659 if (!block_size_is_power_of_two(cache))
660 bio->bi_iter.bi_sector =
661 (block * cache->sectors_per_block) +
662 sector_div(bi_sector, cache->sectors_per_block);
664 bio->bi_iter.bi_sector =
665 (block << cache->sectors_per_block_shift) |
666 (bi_sector & (cache->sectors_per_block - 1));
669 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
672 size_t pb_data_size = get_per_bio_data_size(cache);
673 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
675 spin_lock_irqsave(&cache->lock, flags);
676 if (cache->need_tick_bio &&
677 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
679 cache->need_tick_bio = false;
681 spin_unlock_irqrestore(&cache->lock, flags);
684 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
687 check_if_tick_bio_needed(cache, bio);
688 remap_to_origin(cache, bio);
689 if (bio_data_dir(bio) == WRITE)
690 clear_discard(cache, oblock);
693 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
694 dm_oblock_t oblock, dm_cblock_t cblock)
696 check_if_tick_bio_needed(cache, bio);
697 remap_to_cache(cache, bio, cblock);
698 if (bio_data_dir(bio) == WRITE) {
699 set_dirty(cache, oblock, cblock);
700 clear_discard(cache, oblock);
704 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
706 sector_t block_nr = bio->bi_iter.bi_sector;
708 if (!block_size_is_power_of_two(cache))
709 (void) sector_div(block_nr, cache->sectors_per_block);
711 block_nr >>= cache->sectors_per_block_shift;
713 return to_oblock(block_nr);
716 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
718 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
721 static void issue(struct cache *cache, struct bio *bio)
725 if (!bio_triggers_commit(cache, bio)) {
726 generic_make_request(bio);
731 * Batch together any bios that trigger commits and then issue a
732 * single commit for them in do_worker().
734 spin_lock_irqsave(&cache->lock, flags);
735 cache->commit_requested = true;
736 bio_list_add(&cache->deferred_flush_bios, bio);
737 spin_unlock_irqrestore(&cache->lock, flags);
740 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
744 spin_lock_irqsave(&cache->lock, flags);
745 bio_list_add(&cache->deferred_writethrough_bios, bio);
746 spin_unlock_irqrestore(&cache->lock, flags);
751 static void writethrough_endio(struct bio *bio, int err)
753 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
755 dm_unhook_bio(&pb->hook_info, bio);
762 dm_bio_restore(&pb->bio_details, bio);
763 remap_to_cache(pb->cache, bio, pb->cblock);
766 * We can't issue this bio directly, since we're in interrupt
767 * context. So it gets put on a bio list for processing by the
770 defer_writethrough_bio(pb->cache, bio);
774 * When running in writethrough mode we need to send writes to clean blocks
775 * to both the cache and origin devices. In future we'd like to clone the
776 * bio and send them in parallel, but for now we're doing them in
777 * series as this is easier.
779 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
780 dm_oblock_t oblock, dm_cblock_t cblock)
782 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
786 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
787 dm_bio_record(&pb->bio_details, bio);
789 remap_to_origin_clear_discard(pb->cache, bio, oblock);
792 /*----------------------------------------------------------------
793 * Migration processing
795 * Migration covers moving data from the origin device to the cache, or
797 *--------------------------------------------------------------*/
798 static void free_migration(struct dm_cache_migration *mg)
800 mempool_free(mg, mg->cache->migration_pool);
803 static void inc_nr_migrations(struct cache *cache)
805 atomic_inc(&cache->nr_migrations);
808 static void dec_nr_migrations(struct cache *cache)
810 atomic_dec(&cache->nr_migrations);
813 * Wake the worker in case we're suspending the target.
815 wake_up(&cache->migration_wait);
818 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
821 (holder ? dm_cell_release : dm_cell_release_no_holder)
822 (cache->prison, cell, &cache->deferred_bios);
823 free_prison_cell(cache, cell);
826 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
831 spin_lock_irqsave(&cache->lock, flags);
832 __cell_defer(cache, cell, holder);
833 spin_unlock_irqrestore(&cache->lock, flags);
838 static void cleanup_migration(struct dm_cache_migration *mg)
840 struct cache *cache = mg->cache;
842 dec_nr_migrations(cache);
845 static void migration_failure(struct dm_cache_migration *mg)
847 struct cache *cache = mg->cache;
850 DMWARN_LIMIT("writeback failed; couldn't copy block");
851 set_dirty(cache, mg->old_oblock, mg->cblock);
852 cell_defer(cache, mg->old_ocell, false);
854 } else if (mg->demote) {
855 DMWARN_LIMIT("demotion failed; couldn't copy block");
856 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
858 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
860 cell_defer(cache, mg->new_ocell, true);
862 DMWARN_LIMIT("promotion failed; couldn't copy block");
863 policy_remove_mapping(cache->policy, mg->new_oblock);
864 cell_defer(cache, mg->new_ocell, true);
867 cleanup_migration(mg);
870 static void migration_success_pre_commit(struct dm_cache_migration *mg)
873 struct cache *cache = mg->cache;
876 cell_defer(cache, mg->old_ocell, false);
877 clear_dirty(cache, mg->old_oblock, mg->cblock);
878 cleanup_migration(mg);
881 } else if (mg->demote) {
882 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
883 DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
884 policy_force_mapping(cache->policy, mg->new_oblock,
887 cell_defer(cache, mg->new_ocell, true);
888 cleanup_migration(mg);
892 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
893 DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
894 policy_remove_mapping(cache->policy, mg->new_oblock);
895 cleanup_migration(mg);
900 spin_lock_irqsave(&cache->lock, flags);
901 list_add_tail(&mg->list, &cache->need_commit_migrations);
902 cache->commit_requested = true;
903 spin_unlock_irqrestore(&cache->lock, flags);
906 static void migration_success_post_commit(struct dm_cache_migration *mg)
909 struct cache *cache = mg->cache;
912 DMWARN("writeback unexpectedly triggered commit");
915 } else if (mg->demote) {
916 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
921 spin_lock_irqsave(&cache->lock, flags);
922 list_add_tail(&mg->list, &cache->quiesced_migrations);
923 spin_unlock_irqrestore(&cache->lock, flags);
927 policy_remove_mapping(cache->policy, mg->old_oblock);
928 cleanup_migration(mg);
932 if (mg->requeue_holder)
933 cell_defer(cache, mg->new_ocell, true);
935 bio_endio(mg->new_ocell->holder, 0);
936 cell_defer(cache, mg->new_ocell, false);
938 clear_dirty(cache, mg->new_oblock, mg->cblock);
939 cleanup_migration(mg);
943 static void copy_complete(int read_err, unsigned long write_err, void *context)
946 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
947 struct cache *cache = mg->cache;
949 if (read_err || write_err)
952 spin_lock_irqsave(&cache->lock, flags);
953 list_add_tail(&mg->list, &cache->completed_migrations);
954 spin_unlock_irqrestore(&cache->lock, flags);
959 static void issue_copy_real(struct dm_cache_migration *mg)
962 struct dm_io_region o_region, c_region;
963 struct cache *cache = mg->cache;
964 sector_t cblock = from_cblock(mg->cblock);
966 o_region.bdev = cache->origin_dev->bdev;
967 o_region.count = cache->sectors_per_block;
969 c_region.bdev = cache->cache_dev->bdev;
970 c_region.sector = cblock * cache->sectors_per_block;
971 c_region.count = cache->sectors_per_block;
973 if (mg->writeback || mg->demote) {
975 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
976 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
979 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
980 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
984 DMERR_LIMIT("issuing migration failed");
985 migration_failure(mg);
989 static void overwrite_endio(struct bio *bio, int err)
991 struct dm_cache_migration *mg = bio->bi_private;
992 struct cache *cache = mg->cache;
993 size_t pb_data_size = get_per_bio_data_size(cache);
994 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
997 dm_unhook_bio(&pb->hook_info, bio);
1002 mg->requeue_holder = false;
1004 spin_lock_irqsave(&cache->lock, flags);
1005 list_add_tail(&mg->list, &cache->completed_migrations);
1006 spin_unlock_irqrestore(&cache->lock, flags);
1011 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1013 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1014 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1016 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1017 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1018 generic_make_request(bio);
1021 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1023 return (bio_data_dir(bio) == WRITE) &&
1024 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1027 static void avoid_copy(struct dm_cache_migration *mg)
1029 atomic_inc(&mg->cache->stats.copies_avoided);
1030 migration_success_pre_commit(mg);
1033 static void issue_copy(struct dm_cache_migration *mg)
1036 struct cache *cache = mg->cache;
1038 if (mg->writeback || mg->demote)
1039 avoid = !is_dirty(cache, mg->cblock) ||
1040 is_discarded_oblock(cache, mg->old_oblock);
1042 struct bio *bio = mg->new_ocell->holder;
1044 avoid = is_discarded_oblock(cache, mg->new_oblock);
1046 if (!avoid && bio_writes_complete_block(cache, bio)) {
1047 issue_overwrite(mg, bio);
1052 avoid ? avoid_copy(mg) : issue_copy_real(mg);
1055 static void complete_migration(struct dm_cache_migration *mg)
1058 migration_failure(mg);
1060 migration_success_pre_commit(mg);
1063 static void process_migrations(struct cache *cache, struct list_head *head,
1064 void (*fn)(struct dm_cache_migration *))
1066 unsigned long flags;
1067 struct list_head list;
1068 struct dm_cache_migration *mg, *tmp;
1070 INIT_LIST_HEAD(&list);
1071 spin_lock_irqsave(&cache->lock, flags);
1072 list_splice_init(head, &list);
1073 spin_unlock_irqrestore(&cache->lock, flags);
1075 list_for_each_entry_safe(mg, tmp, &list, list)
1079 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1081 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1084 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1086 unsigned long flags;
1087 struct cache *cache = mg->cache;
1089 spin_lock_irqsave(&cache->lock, flags);
1090 __queue_quiesced_migration(mg);
1091 spin_unlock_irqrestore(&cache->lock, flags);
1096 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1098 unsigned long flags;
1099 struct dm_cache_migration *mg, *tmp;
1101 spin_lock_irqsave(&cache->lock, flags);
1102 list_for_each_entry_safe(mg, tmp, work, list)
1103 __queue_quiesced_migration(mg);
1104 spin_unlock_irqrestore(&cache->lock, flags);
1109 static void check_for_quiesced_migrations(struct cache *cache,
1110 struct per_bio_data *pb)
1112 struct list_head work;
1114 if (!pb->all_io_entry)
1117 INIT_LIST_HEAD(&work);
1118 if (pb->all_io_entry)
1119 dm_deferred_entry_dec(pb->all_io_entry, &work);
1121 if (!list_empty(&work))
1122 queue_quiesced_migrations(cache, &work);
1125 static void quiesce_migration(struct dm_cache_migration *mg)
1127 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1128 queue_quiesced_migration(mg);
1131 static void promote(struct cache *cache, struct prealloc *structs,
1132 dm_oblock_t oblock, dm_cblock_t cblock,
1133 struct dm_bio_prison_cell *cell)
1135 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1138 mg->writeback = false;
1141 mg->requeue_holder = true;
1142 mg->invalidate = false;
1144 mg->new_oblock = oblock;
1145 mg->cblock = cblock;
1146 mg->old_ocell = NULL;
1147 mg->new_ocell = cell;
1148 mg->start_jiffies = jiffies;
1150 inc_nr_migrations(cache);
1151 quiesce_migration(mg);
1154 static void writeback(struct cache *cache, struct prealloc *structs,
1155 dm_oblock_t oblock, dm_cblock_t cblock,
1156 struct dm_bio_prison_cell *cell)
1158 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1161 mg->writeback = true;
1163 mg->promote = false;
1164 mg->requeue_holder = true;
1165 mg->invalidate = false;
1167 mg->old_oblock = oblock;
1168 mg->cblock = cblock;
1169 mg->old_ocell = cell;
1170 mg->new_ocell = NULL;
1171 mg->start_jiffies = jiffies;
1173 inc_nr_migrations(cache);
1174 quiesce_migration(mg);
1177 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1178 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1180 struct dm_bio_prison_cell *old_ocell,
1181 struct dm_bio_prison_cell *new_ocell)
1183 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1186 mg->writeback = false;
1189 mg->requeue_holder = true;
1190 mg->invalidate = false;
1192 mg->old_oblock = old_oblock;
1193 mg->new_oblock = new_oblock;
1194 mg->cblock = cblock;
1195 mg->old_ocell = old_ocell;
1196 mg->new_ocell = new_ocell;
1197 mg->start_jiffies = jiffies;
1199 inc_nr_migrations(cache);
1200 quiesce_migration(mg);
1204 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1205 * block are thrown away.
1207 static void invalidate(struct cache *cache, struct prealloc *structs,
1208 dm_oblock_t oblock, dm_cblock_t cblock,
1209 struct dm_bio_prison_cell *cell)
1211 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1214 mg->writeback = false;
1216 mg->promote = false;
1217 mg->requeue_holder = true;
1218 mg->invalidate = true;
1220 mg->old_oblock = oblock;
1221 mg->cblock = cblock;
1222 mg->old_ocell = cell;
1223 mg->new_ocell = NULL;
1224 mg->start_jiffies = jiffies;
1226 inc_nr_migrations(cache);
1227 quiesce_migration(mg);
1230 /*----------------------------------------------------------------
1232 *--------------------------------------------------------------*/
1233 static void defer_bio(struct cache *cache, struct bio *bio)
1235 unsigned long flags;
1237 spin_lock_irqsave(&cache->lock, flags);
1238 bio_list_add(&cache->deferred_bios, bio);
1239 spin_unlock_irqrestore(&cache->lock, flags);
1244 static void process_flush_bio(struct cache *cache, struct bio *bio)
1246 size_t pb_data_size = get_per_bio_data_size(cache);
1247 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1249 BUG_ON(bio->bi_iter.bi_size);
1251 remap_to_origin(cache, bio);
1253 remap_to_cache(cache, bio, 0);
1259 * People generally discard large parts of a device, eg, the whole device
1260 * when formatting. Splitting these large discards up into cache block
1261 * sized ios and then quiescing (always neccessary for discard) takes too
1264 * We keep it simple, and allow any size of discard to come in, and just
1265 * mark off blocks on the discard bitset. No passdown occurs!
1267 * To implement passdown we need to change the bio_prison such that a cell
1268 * can have a key that spans many blocks.
1270 static void process_discard_bio(struct cache *cache, struct bio *bio)
1272 dm_block_t start_block = dm_sector_div_up(bio->bi_iter.bi_sector,
1273 cache->sectors_per_block);
1274 dm_block_t end_block = bio_end_sector(bio);
1277 end_block = block_div(end_block, cache->sectors_per_block);
1279 for (b = start_block; b < end_block; b++)
1280 set_discard(cache, to_oblock(b));
1285 static bool spare_migration_bandwidth(struct cache *cache)
1287 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1288 cache->sectors_per_block;
1289 return current_volume < cache->migration_threshold;
1292 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1294 atomic_inc(bio_data_dir(bio) == READ ?
1295 &cache->stats.read_hit : &cache->stats.write_hit);
1298 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1300 atomic_inc(bio_data_dir(bio) == READ ?
1301 &cache->stats.read_miss : &cache->stats.write_miss);
1304 static void issue_cache_bio(struct cache *cache, struct bio *bio,
1305 struct per_bio_data *pb,
1306 dm_oblock_t oblock, dm_cblock_t cblock)
1308 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1309 remap_to_cache_dirty(cache, bio, oblock, cblock);
1313 static void process_bio(struct cache *cache, struct prealloc *structs,
1317 bool release_cell = true;
1318 dm_oblock_t block = get_bio_block(cache, bio);
1319 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1320 struct policy_result lookup_result;
1321 size_t pb_data_size = get_per_bio_data_size(cache);
1322 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1323 bool discarded_block = is_discarded_oblock(cache, block);
1324 bool passthrough = passthrough_mode(&cache->features);
1325 bool can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
1328 * Check to see if that block is currently migrating.
1330 cell_prealloc = prealloc_get_cell(structs);
1331 r = bio_detain(cache, block, bio, cell_prealloc,
1332 (cell_free_fn) prealloc_put_cell,
1333 structs, &new_ocell);
1337 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1338 bio, &lookup_result);
1340 if (r == -EWOULDBLOCK)
1341 /* migration has been denied */
1342 lookup_result.op = POLICY_MISS;
1344 switch (lookup_result.op) {
1347 inc_miss_counter(cache, bio);
1350 * Passthrough always maps to the origin,
1351 * invalidating any cache blocks that are written
1355 if (bio_data_dir(bio) == WRITE) {
1356 atomic_inc(&cache->stats.demotion);
1357 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1358 release_cell = false;
1361 /* FIXME: factor out issue_origin() */
1362 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1363 remap_to_origin_clear_discard(cache, bio, block);
1367 inc_hit_counter(cache, bio);
1369 if (bio_data_dir(bio) == WRITE &&
1370 writethrough_mode(&cache->features) &&
1371 !is_dirty(cache, lookup_result.cblock)) {
1372 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1373 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1376 issue_cache_bio(cache, bio, pb, block, lookup_result.cblock);
1382 inc_miss_counter(cache, bio);
1383 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1384 remap_to_origin_clear_discard(cache, bio, block);
1389 atomic_inc(&cache->stats.promotion);
1390 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1391 release_cell = false;
1394 case POLICY_REPLACE:
1395 cell_prealloc = prealloc_get_cell(structs);
1396 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1397 (cell_free_fn) prealloc_put_cell,
1398 structs, &old_ocell);
1401 * We have to be careful to avoid lock inversion of
1402 * the cells. So we back off, and wait for the
1403 * old_ocell to become free.
1405 policy_force_mapping(cache->policy, block,
1406 lookup_result.old_oblock);
1407 atomic_inc(&cache->stats.cache_cell_clash);
1410 atomic_inc(&cache->stats.demotion);
1411 atomic_inc(&cache->stats.promotion);
1413 demote_then_promote(cache, structs, lookup_result.old_oblock,
1414 block, lookup_result.cblock,
1415 old_ocell, new_ocell);
1416 release_cell = false;
1420 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1421 (unsigned) lookup_result.op);
1426 cell_defer(cache, new_ocell, false);
1429 static int need_commit_due_to_time(struct cache *cache)
1431 return jiffies < cache->last_commit_jiffies ||
1432 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1435 static int commit_if_needed(struct cache *cache)
1439 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1440 dm_cache_changed_this_transaction(cache->cmd)) {
1441 atomic_inc(&cache->stats.commit_count);
1442 cache->commit_requested = false;
1443 r = dm_cache_commit(cache->cmd, false);
1444 cache->last_commit_jiffies = jiffies;
1450 static void process_deferred_bios(struct cache *cache)
1452 unsigned long flags;
1453 struct bio_list bios;
1455 struct prealloc structs;
1457 memset(&structs, 0, sizeof(structs));
1458 bio_list_init(&bios);
1460 spin_lock_irqsave(&cache->lock, flags);
1461 bio_list_merge(&bios, &cache->deferred_bios);
1462 bio_list_init(&cache->deferred_bios);
1463 spin_unlock_irqrestore(&cache->lock, flags);
1465 while (!bio_list_empty(&bios)) {
1467 * If we've got no free migration structs, and processing
1468 * this bio might require one, we pause until there are some
1469 * prepared mappings to process.
1471 if (prealloc_data_structs(cache, &structs)) {
1472 spin_lock_irqsave(&cache->lock, flags);
1473 bio_list_merge(&cache->deferred_bios, &bios);
1474 spin_unlock_irqrestore(&cache->lock, flags);
1478 bio = bio_list_pop(&bios);
1480 if (bio->bi_rw & REQ_FLUSH)
1481 process_flush_bio(cache, bio);
1482 else if (bio->bi_rw & REQ_DISCARD)
1483 process_discard_bio(cache, bio);
1485 process_bio(cache, &structs, bio);
1488 prealloc_free_structs(cache, &structs);
1491 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1493 unsigned long flags;
1494 struct bio_list bios;
1497 bio_list_init(&bios);
1499 spin_lock_irqsave(&cache->lock, flags);
1500 bio_list_merge(&bios, &cache->deferred_flush_bios);
1501 bio_list_init(&cache->deferred_flush_bios);
1502 spin_unlock_irqrestore(&cache->lock, flags);
1504 while ((bio = bio_list_pop(&bios)))
1505 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1508 static void process_deferred_writethrough_bios(struct cache *cache)
1510 unsigned long flags;
1511 struct bio_list bios;
1514 bio_list_init(&bios);
1516 spin_lock_irqsave(&cache->lock, flags);
1517 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1518 bio_list_init(&cache->deferred_writethrough_bios);
1519 spin_unlock_irqrestore(&cache->lock, flags);
1521 while ((bio = bio_list_pop(&bios)))
1522 generic_make_request(bio);
1525 static void writeback_some_dirty_blocks(struct cache *cache)
1530 struct prealloc structs;
1531 struct dm_bio_prison_cell *old_ocell;
1533 memset(&structs, 0, sizeof(structs));
1535 while (spare_migration_bandwidth(cache)) {
1536 if (prealloc_data_structs(cache, &structs))
1539 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1543 r = get_cell(cache, oblock, &structs, &old_ocell);
1545 policy_set_dirty(cache->policy, oblock);
1549 writeback(cache, &structs, oblock, cblock, old_ocell);
1552 prealloc_free_structs(cache, &structs);
1555 /*----------------------------------------------------------------
1557 * Dropping something from the cache *without* writing back.
1558 *--------------------------------------------------------------*/
1560 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
1563 uint64_t begin = from_cblock(req->cblocks->begin);
1564 uint64_t end = from_cblock(req->cblocks->end);
1566 while (begin != end) {
1567 r = policy_remove_cblock(cache->policy, to_cblock(begin));
1569 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
1573 } else if (r == -ENODATA) {
1574 /* harmless, already unmapped */
1578 DMERR("policy_remove_cblock failed");
1585 cache->commit_requested = true;
1588 atomic_set(&req->complete, 1);
1590 wake_up(&req->result_wait);
1593 static void process_invalidation_requests(struct cache *cache)
1595 struct list_head list;
1596 struct invalidation_request *req, *tmp;
1598 INIT_LIST_HEAD(&list);
1599 spin_lock(&cache->invalidation_lock);
1600 list_splice_init(&cache->invalidation_requests, &list);
1601 spin_unlock(&cache->invalidation_lock);
1603 list_for_each_entry_safe (req, tmp, &list, list)
1604 process_invalidation_request(cache, req);
1607 /*----------------------------------------------------------------
1609 *--------------------------------------------------------------*/
1610 static bool is_quiescing(struct cache *cache)
1612 return atomic_read(&cache->quiescing);
1615 static void ack_quiescing(struct cache *cache)
1617 if (is_quiescing(cache)) {
1618 atomic_inc(&cache->quiescing_ack);
1619 wake_up(&cache->quiescing_wait);
1623 static void wait_for_quiescing_ack(struct cache *cache)
1625 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
1628 static void start_quiescing(struct cache *cache)
1630 atomic_inc(&cache->quiescing);
1631 wait_for_quiescing_ack(cache);
1634 static void stop_quiescing(struct cache *cache)
1636 atomic_set(&cache->quiescing, 0);
1637 atomic_set(&cache->quiescing_ack, 0);
1640 static void wait_for_migrations(struct cache *cache)
1642 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1645 static void stop_worker(struct cache *cache)
1647 cancel_delayed_work(&cache->waker);
1648 flush_workqueue(cache->wq);
1651 static void requeue_deferred_io(struct cache *cache)
1654 struct bio_list bios;
1656 bio_list_init(&bios);
1657 bio_list_merge(&bios, &cache->deferred_bios);
1658 bio_list_init(&cache->deferred_bios);
1660 while ((bio = bio_list_pop(&bios)))
1661 bio_endio(bio, DM_ENDIO_REQUEUE);
1664 static int more_work(struct cache *cache)
1666 if (is_quiescing(cache))
1667 return !list_empty(&cache->quiesced_migrations) ||
1668 !list_empty(&cache->completed_migrations) ||
1669 !list_empty(&cache->need_commit_migrations);
1671 return !bio_list_empty(&cache->deferred_bios) ||
1672 !bio_list_empty(&cache->deferred_flush_bios) ||
1673 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1674 !list_empty(&cache->quiesced_migrations) ||
1675 !list_empty(&cache->completed_migrations) ||
1676 !list_empty(&cache->need_commit_migrations) ||
1680 static void do_worker(struct work_struct *ws)
1682 struct cache *cache = container_of(ws, struct cache, worker);
1685 if (!is_quiescing(cache)) {
1686 writeback_some_dirty_blocks(cache);
1687 process_deferred_writethrough_bios(cache);
1688 process_deferred_bios(cache);
1689 process_invalidation_requests(cache);
1692 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1693 process_migrations(cache, &cache->completed_migrations, complete_migration);
1695 if (commit_if_needed(cache)) {
1696 process_deferred_flush_bios(cache, false);
1699 * FIXME: rollback metadata or just go into a
1700 * failure mode and error everything
1703 process_deferred_flush_bios(cache, true);
1704 process_migrations(cache, &cache->need_commit_migrations,
1705 migration_success_post_commit);
1708 ack_quiescing(cache);
1710 } while (more_work(cache));
1714 * We want to commit periodically so that not too much
1715 * unwritten metadata builds up.
1717 static void do_waker(struct work_struct *ws)
1719 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1720 policy_tick(cache->policy);
1722 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1725 /*----------------------------------------------------------------*/
1727 static int is_congested(struct dm_dev *dev, int bdi_bits)
1729 struct request_queue *q = bdev_get_queue(dev->bdev);
1730 return bdi_congested(&q->backing_dev_info, bdi_bits);
1733 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1735 struct cache *cache = container_of(cb, struct cache, callbacks);
1737 return is_congested(cache->origin_dev, bdi_bits) ||
1738 is_congested(cache->cache_dev, bdi_bits);
1741 /*----------------------------------------------------------------
1743 *--------------------------------------------------------------*/
1746 * This function gets called on the error paths of the constructor, so we
1747 * have to cope with a partially initialised struct.
1749 static void destroy(struct cache *cache)
1753 if (cache->next_migration)
1754 mempool_free(cache->next_migration, cache->migration_pool);
1756 if (cache->migration_pool)
1757 mempool_destroy(cache->migration_pool);
1759 if (cache->all_io_ds)
1760 dm_deferred_set_destroy(cache->all_io_ds);
1763 dm_bio_prison_destroy(cache->prison);
1766 destroy_workqueue(cache->wq);
1768 if (cache->dirty_bitset)
1769 free_bitset(cache->dirty_bitset);
1771 if (cache->discard_bitset)
1772 free_bitset(cache->discard_bitset);
1775 dm_kcopyd_client_destroy(cache->copier);
1778 dm_cache_metadata_close(cache->cmd);
1780 if (cache->metadata_dev)
1781 dm_put_device(cache->ti, cache->metadata_dev);
1783 if (cache->origin_dev)
1784 dm_put_device(cache->ti, cache->origin_dev);
1786 if (cache->cache_dev)
1787 dm_put_device(cache->ti, cache->cache_dev);
1790 dm_cache_policy_destroy(cache->policy);
1792 for (i = 0; i < cache->nr_ctr_args ; i++)
1793 kfree(cache->ctr_args[i]);
1794 kfree(cache->ctr_args);
1799 static void cache_dtr(struct dm_target *ti)
1801 struct cache *cache = ti->private;
1806 static sector_t get_dev_size(struct dm_dev *dev)
1808 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1811 /*----------------------------------------------------------------*/
1814 * Construct a cache device mapping.
1816 * cache <metadata dev> <cache dev> <origin dev> <block size>
1817 * <#feature args> [<feature arg>]*
1818 * <policy> <#policy args> [<policy arg>]*
1820 * metadata dev : fast device holding the persistent metadata
1821 * cache dev : fast device holding cached data blocks
1822 * origin dev : slow device holding original data blocks
1823 * block size : cache unit size in sectors
1825 * #feature args : number of feature arguments passed
1826 * feature args : writethrough. (The default is writeback.)
1828 * policy : the replacement policy to use
1829 * #policy args : an even number of policy arguments corresponding
1830 * to key/value pairs passed to the policy
1831 * policy args : key/value pairs passed to the policy
1832 * E.g. 'sequential_threshold 1024'
1833 * See cache-policies.txt for details.
1835 * Optional feature arguments are:
1836 * writethrough : write through caching that prohibits cache block
1837 * content from being different from origin block content.
1838 * Without this argument, the default behaviour is to write
1839 * back cache block contents later for performance reasons,
1840 * so they may differ from the corresponding origin blocks.
1843 struct dm_target *ti;
1845 struct dm_dev *metadata_dev;
1847 struct dm_dev *cache_dev;
1848 sector_t cache_sectors;
1850 struct dm_dev *origin_dev;
1851 sector_t origin_sectors;
1853 uint32_t block_size;
1855 const char *policy_name;
1857 const char **policy_argv;
1859 struct cache_features features;
1862 static void destroy_cache_args(struct cache_args *ca)
1864 if (ca->metadata_dev)
1865 dm_put_device(ca->ti, ca->metadata_dev);
1868 dm_put_device(ca->ti, ca->cache_dev);
1871 dm_put_device(ca->ti, ca->origin_dev);
1876 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1879 *error = "Insufficient args";
1886 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1890 sector_t metadata_dev_size;
1891 char b[BDEVNAME_SIZE];
1893 if (!at_least_one_arg(as, error))
1896 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1899 *error = "Error opening metadata device";
1903 metadata_dev_size = get_dev_size(ca->metadata_dev);
1904 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1905 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1906 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1911 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1916 if (!at_least_one_arg(as, error))
1919 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1922 *error = "Error opening cache device";
1925 ca->cache_sectors = get_dev_size(ca->cache_dev);
1930 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1935 if (!at_least_one_arg(as, error))
1938 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1941 *error = "Error opening origin device";
1945 ca->origin_sectors = get_dev_size(ca->origin_dev);
1946 if (ca->ti->len > ca->origin_sectors) {
1947 *error = "Device size larger than cached device";
1954 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1957 unsigned long block_size;
1959 if (!at_least_one_arg(as, error))
1962 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
1963 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1964 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1965 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1966 *error = "Invalid data block size";
1970 if (block_size > ca->cache_sectors) {
1971 *error = "Data block size is larger than the cache device";
1975 ca->block_size = block_size;
1980 static void init_features(struct cache_features *cf)
1982 cf->mode = CM_WRITE;
1983 cf->io_mode = CM_IO_WRITEBACK;
1986 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1989 static struct dm_arg _args[] = {
1990 {0, 1, "Invalid number of cache feature arguments"},
1996 struct cache_features *cf = &ca->features;
2000 r = dm_read_arg_group(_args, as, &argc, error);
2005 arg = dm_shift_arg(as);
2007 if (!strcasecmp(arg, "writeback"))
2008 cf->io_mode = CM_IO_WRITEBACK;
2010 else if (!strcasecmp(arg, "writethrough"))
2011 cf->io_mode = CM_IO_WRITETHROUGH;
2013 else if (!strcasecmp(arg, "passthrough"))
2014 cf->io_mode = CM_IO_PASSTHROUGH;
2017 *error = "Unrecognised cache feature requested";
2025 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2028 static struct dm_arg _args[] = {
2029 {0, 1024, "Invalid number of policy arguments"},
2034 if (!at_least_one_arg(as, error))
2037 ca->policy_name = dm_shift_arg(as);
2039 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2043 ca->policy_argv = (const char **)as->argv;
2044 dm_consume_args(as, ca->policy_argc);
2049 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2053 struct dm_arg_set as;
2058 r = parse_metadata_dev(ca, &as, error);
2062 r = parse_cache_dev(ca, &as, error);
2066 r = parse_origin_dev(ca, &as, error);
2070 r = parse_block_size(ca, &as, error);
2074 r = parse_features(ca, &as, error);
2078 r = parse_policy(ca, &as, error);
2085 /*----------------------------------------------------------------*/
2087 static struct kmem_cache *migration_cache;
2089 #define NOT_CORE_OPTION 1
2091 static int process_config_option(struct cache *cache, const char *key, const char *value)
2095 if (!strcasecmp(key, "migration_threshold")) {
2096 if (kstrtoul(value, 10, &tmp))
2099 cache->migration_threshold = tmp;
2103 return NOT_CORE_OPTION;
2106 static int set_config_value(struct cache *cache, const char *key, const char *value)
2108 int r = process_config_option(cache, key, value);
2110 if (r == NOT_CORE_OPTION)
2111 r = policy_set_config_value(cache->policy, key, value);
2114 DMWARN("bad config value for %s: %s", key, value);
2119 static int set_config_values(struct cache *cache, int argc, const char **argv)
2124 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2129 r = set_config_value(cache, argv[0], argv[1]);
2140 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2143 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2145 cache->origin_sectors,
2146 cache->sectors_per_block);
2148 *error = "Error creating cache's policy";
2156 #define DEFAULT_MIGRATION_THRESHOLD 2048
2158 static int cache_create(struct cache_args *ca, struct cache **result)
2161 char **error = &ca->ti->error;
2162 struct cache *cache;
2163 struct dm_target *ti = ca->ti;
2164 dm_block_t origin_blocks;
2165 struct dm_cache_metadata *cmd;
2166 bool may_format = ca->features.mode == CM_WRITE;
2168 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2173 ti->private = cache;
2174 ti->num_flush_bios = 2;
2175 ti->flush_supported = true;
2177 ti->num_discard_bios = 1;
2178 ti->discards_supported = true;
2179 ti->discard_zeroes_data_unsupported = true;
2180 /* Discard bios must be split on a block boundary */
2181 ti->split_discard_bios = true;
2183 cache->features = ca->features;
2184 ti->per_bio_data_size = get_per_bio_data_size(cache);
2186 cache->callbacks.congested_fn = cache_is_congested;
2187 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2189 cache->metadata_dev = ca->metadata_dev;
2190 cache->origin_dev = ca->origin_dev;
2191 cache->cache_dev = ca->cache_dev;
2193 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2195 /* FIXME: factor out this whole section */
2196 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2197 origin_blocks = block_div(origin_blocks, ca->block_size);
2198 cache->origin_blocks = to_oblock(origin_blocks);
2200 cache->sectors_per_block = ca->block_size;
2201 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2206 if (ca->block_size & (ca->block_size - 1)) {
2207 dm_block_t cache_size = ca->cache_sectors;
2209 cache->sectors_per_block_shift = -1;
2210 cache_size = block_div(cache_size, ca->block_size);
2211 cache->cache_size = to_cblock(cache_size);
2213 cache->sectors_per_block_shift = __ffs(ca->block_size);
2214 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
2217 r = create_cache_policy(cache, ca, error);
2221 cache->policy_nr_args = ca->policy_argc;
2222 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2224 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2226 *error = "Error setting cache policy's config values";
2230 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2231 ca->block_size, may_format,
2232 dm_cache_policy_get_hint_size(cache->policy));
2234 *error = "Error creating metadata object";
2240 if (passthrough_mode(&cache->features)) {
2243 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2245 *error = "dm_cache_metadata_all_clean() failed";
2250 *error = "Cannot enter passthrough mode unless all blocks are clean";
2256 spin_lock_init(&cache->lock);
2257 bio_list_init(&cache->deferred_bios);
2258 bio_list_init(&cache->deferred_flush_bios);
2259 bio_list_init(&cache->deferred_writethrough_bios);
2260 INIT_LIST_HEAD(&cache->quiesced_migrations);
2261 INIT_LIST_HEAD(&cache->completed_migrations);
2262 INIT_LIST_HEAD(&cache->need_commit_migrations);
2263 atomic_set(&cache->nr_migrations, 0);
2264 init_waitqueue_head(&cache->migration_wait);
2266 init_waitqueue_head(&cache->quiescing_wait);
2267 atomic_set(&cache->quiescing, 0);
2268 atomic_set(&cache->quiescing_ack, 0);
2271 atomic_set(&cache->nr_dirty, 0);
2272 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2273 if (!cache->dirty_bitset) {
2274 *error = "could not allocate dirty bitset";
2277 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2279 cache->discard_nr_blocks = cache->origin_blocks;
2280 cache->discard_bitset = alloc_bitset(from_oblock(cache->discard_nr_blocks));
2281 if (!cache->discard_bitset) {
2282 *error = "could not allocate discard bitset";
2285 clear_bitset(cache->discard_bitset, from_oblock(cache->discard_nr_blocks));
2287 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2288 if (IS_ERR(cache->copier)) {
2289 *error = "could not create kcopyd client";
2290 r = PTR_ERR(cache->copier);
2294 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2296 *error = "could not create workqueue for metadata object";
2299 INIT_WORK(&cache->worker, do_worker);
2300 INIT_DELAYED_WORK(&cache->waker, do_waker);
2301 cache->last_commit_jiffies = jiffies;
2303 cache->prison = dm_bio_prison_create(PRISON_CELLS);
2304 if (!cache->prison) {
2305 *error = "could not create bio prison";
2309 cache->all_io_ds = dm_deferred_set_create();
2310 if (!cache->all_io_ds) {
2311 *error = "could not create all_io deferred set";
2315 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2317 if (!cache->migration_pool) {
2318 *error = "Error creating cache's migration mempool";
2322 cache->next_migration = NULL;
2324 cache->need_tick_bio = true;
2325 cache->sized = false;
2326 cache->invalidate = false;
2327 cache->commit_requested = false;
2328 cache->loaded_mappings = false;
2329 cache->loaded_discards = false;
2333 atomic_set(&cache->stats.demotion, 0);
2334 atomic_set(&cache->stats.promotion, 0);
2335 atomic_set(&cache->stats.copies_avoided, 0);
2336 atomic_set(&cache->stats.cache_cell_clash, 0);
2337 atomic_set(&cache->stats.commit_count, 0);
2338 atomic_set(&cache->stats.discard_count, 0);
2340 spin_lock_init(&cache->invalidation_lock);
2341 INIT_LIST_HEAD(&cache->invalidation_requests);
2351 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2356 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2359 for (i = 0; i < argc; i++) {
2360 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2369 cache->nr_ctr_args = argc;
2370 cache->ctr_args = copy;
2375 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2378 struct cache_args *ca;
2379 struct cache *cache = NULL;
2381 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2383 ti->error = "Error allocating memory for cache";
2388 r = parse_cache_args(ca, argc, argv, &ti->error);
2392 r = cache_create(ca, &cache);
2396 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2402 ti->private = cache;
2405 destroy_cache_args(ca);
2409 static int cache_map(struct dm_target *ti, struct bio *bio)
2411 struct cache *cache = ti->private;
2414 dm_oblock_t block = get_bio_block(cache, bio);
2415 size_t pb_data_size = get_per_bio_data_size(cache);
2416 bool can_migrate = false;
2417 bool discarded_block;
2418 struct dm_bio_prison_cell *cell;
2419 struct policy_result lookup_result;
2420 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
2422 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2424 * This can only occur if the io goes to a partial block at
2425 * the end of the origin device. We don't cache these.
2426 * Just remap to the origin and carry on.
2428 remap_to_origin(cache, bio);
2429 return DM_MAPIO_REMAPPED;
2432 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2433 defer_bio(cache, bio);
2434 return DM_MAPIO_SUBMITTED;
2438 * Check to see if that block is currently migrating.
2440 cell = alloc_prison_cell(cache);
2442 defer_bio(cache, bio);
2443 return DM_MAPIO_SUBMITTED;
2446 r = bio_detain(cache, block, bio, cell,
2447 (cell_free_fn) free_prison_cell,
2451 defer_bio(cache, bio);
2453 return DM_MAPIO_SUBMITTED;
2456 discarded_block = is_discarded_oblock(cache, block);
2458 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2459 bio, &lookup_result);
2460 if (r == -EWOULDBLOCK) {
2461 cell_defer(cache, cell, true);
2462 return DM_MAPIO_SUBMITTED;
2465 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2467 return DM_MAPIO_SUBMITTED;
2470 r = DM_MAPIO_REMAPPED;
2471 switch (lookup_result.op) {
2473 if (passthrough_mode(&cache->features)) {
2474 if (bio_data_dir(bio) == WRITE) {
2476 * We need to invalidate this block, so
2477 * defer for the worker thread.
2479 cell_defer(cache, cell, true);
2480 r = DM_MAPIO_SUBMITTED;
2483 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2484 inc_miss_counter(cache, bio);
2485 remap_to_origin_clear_discard(cache, bio, block);
2487 cell_defer(cache, cell, false);
2491 inc_hit_counter(cache, bio);
2492 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2494 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
2495 !is_dirty(cache, lookup_result.cblock))
2496 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2498 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2500 cell_defer(cache, cell, false);
2505 inc_miss_counter(cache, bio);
2506 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2508 if (pb->req_nr != 0) {
2510 * This is a duplicate writethrough io that is no
2511 * longer needed because the block has been demoted.
2514 cell_defer(cache, cell, false);
2515 return DM_MAPIO_SUBMITTED;
2517 remap_to_origin_clear_discard(cache, bio, block);
2518 cell_defer(cache, cell, false);
2523 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2524 (unsigned) lookup_result.op);
2526 r = DM_MAPIO_SUBMITTED;
2532 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2534 struct cache *cache = ti->private;
2535 unsigned long flags;
2536 size_t pb_data_size = get_per_bio_data_size(cache);
2537 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2540 policy_tick(cache->policy);
2542 spin_lock_irqsave(&cache->lock, flags);
2543 cache->need_tick_bio = true;
2544 spin_unlock_irqrestore(&cache->lock, flags);
2547 check_for_quiesced_migrations(cache, pb);
2552 static int write_dirty_bitset(struct cache *cache)
2556 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2557 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2558 is_dirty(cache, to_cblock(i)));
2566 static int write_discard_bitset(struct cache *cache)
2570 r = dm_cache_discard_bitset_resize(cache->cmd, cache->sectors_per_block,
2571 cache->origin_blocks);
2573 DMERR("could not resize on-disk discard bitset");
2577 for (i = 0; i < from_oblock(cache->discard_nr_blocks); i++) {
2578 r = dm_cache_set_discard(cache->cmd, to_oblock(i),
2579 is_discarded(cache, to_oblock(i)));
2588 * returns true on success
2590 static bool sync_metadata(struct cache *cache)
2594 r1 = write_dirty_bitset(cache);
2596 DMERR("could not write dirty bitset");
2598 r2 = write_discard_bitset(cache);
2600 DMERR("could not write discard bitset");
2604 r3 = dm_cache_write_hints(cache->cmd, cache->policy);
2606 DMERR("could not write hints");
2609 * If writing the above metadata failed, we still commit, but don't
2610 * set the clean shutdown flag. This will effectively force every
2611 * dirty bit to be set on reload.
2613 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2615 DMERR("could not write cache metadata. Data loss may occur.");
2617 return !r1 && !r2 && !r3 && !r4;
2620 static void cache_postsuspend(struct dm_target *ti)
2622 struct cache *cache = ti->private;
2624 start_quiescing(cache);
2625 wait_for_migrations(cache);
2627 requeue_deferred_io(cache);
2628 stop_quiescing(cache);
2630 (void) sync_metadata(cache);
2633 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2634 bool dirty, uint32_t hint, bool hint_valid)
2637 struct cache *cache = context;
2639 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2644 set_dirty(cache, oblock, cblock);
2646 clear_dirty(cache, oblock, cblock);
2651 static int load_discard(void *context, sector_t discard_block_size,
2652 dm_oblock_t oblock, bool discard)
2654 struct cache *cache = context;
2657 set_discard(cache, oblock);
2659 clear_discard(cache, oblock);
2664 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2666 sector_t size = get_dev_size(cache->cache_dev);
2667 (void) sector_div(size, cache->sectors_per_block);
2668 return to_cblock(size);
2671 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2673 if (from_cblock(new_size) > from_cblock(cache->cache_size))
2677 * We can't drop a dirty block when shrinking the cache.
2679 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2680 new_size = to_cblock(from_cblock(new_size) + 1);
2681 if (is_dirty(cache, new_size)) {
2682 DMERR("unable to shrink cache; cache block %llu is dirty",
2683 (unsigned long long) from_cblock(new_size));
2691 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2695 r = dm_cache_resize(cache->cmd, new_size);
2697 DMERR("could not resize cache metadata");
2701 cache->cache_size = new_size;
2706 static int cache_preresume(struct dm_target *ti)
2709 struct cache *cache = ti->private;
2710 dm_cblock_t csize = get_cache_dev_size(cache);
2713 * Check to see if the cache has resized.
2715 if (!cache->sized) {
2716 r = resize_cache_dev(cache, csize);
2720 cache->sized = true;
2722 } else if (csize != cache->cache_size) {
2723 if (!can_resize(cache, csize))
2726 r = resize_cache_dev(cache, csize);
2731 if (!cache->loaded_mappings) {
2732 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2733 load_mapping, cache);
2735 DMERR("could not load cache mappings");
2739 cache->loaded_mappings = true;
2742 if (!cache->loaded_discards) {
2743 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2745 DMERR("could not load origin discards");
2749 cache->loaded_discards = true;
2755 static void cache_resume(struct dm_target *ti)
2757 struct cache *cache = ti->private;
2759 cache->need_tick_bio = true;
2760 do_waker(&cache->waker.work);
2766 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
2767 * <cache block size> <#used cache blocks>/<#total cache blocks>
2768 * <#read hits> <#read misses> <#write hits> <#write misses>
2769 * <#demotions> <#promotions> <#dirty>
2770 * <#features> <features>*
2771 * <#core args> <core args>
2772 * <policy name> <#policy args> <policy args>*
2774 static void cache_status(struct dm_target *ti, status_type_t type,
2775 unsigned status_flags, char *result, unsigned maxlen)
2780 dm_block_t nr_free_blocks_metadata = 0;
2781 dm_block_t nr_blocks_metadata = 0;
2782 char buf[BDEVNAME_SIZE];
2783 struct cache *cache = ti->private;
2784 dm_cblock_t residency;
2787 case STATUSTYPE_INFO:
2788 /* Commit to ensure statistics aren't out-of-date */
2789 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2790 r = dm_cache_commit(cache->cmd, false);
2792 DMERR("could not commit metadata for accurate status");
2795 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2796 &nr_free_blocks_metadata);
2798 DMERR("could not get metadata free block count");
2802 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2804 DMERR("could not get metadata device size");
2808 residency = policy_residency(cache->policy);
2810 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
2811 (unsigned)(DM_CACHE_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
2812 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2813 (unsigned long long)nr_blocks_metadata,
2814 cache->sectors_per_block,
2815 (unsigned long long) from_cblock(residency),
2816 (unsigned long long) from_cblock(cache->cache_size),
2817 (unsigned) atomic_read(&cache->stats.read_hit),
2818 (unsigned) atomic_read(&cache->stats.read_miss),
2819 (unsigned) atomic_read(&cache->stats.write_hit),
2820 (unsigned) atomic_read(&cache->stats.write_miss),
2821 (unsigned) atomic_read(&cache->stats.demotion),
2822 (unsigned) atomic_read(&cache->stats.promotion),
2823 (unsigned long) atomic_read(&cache->nr_dirty));
2825 if (writethrough_mode(&cache->features))
2826 DMEMIT("1 writethrough ");
2828 else if (passthrough_mode(&cache->features))
2829 DMEMIT("1 passthrough ");
2831 else if (writeback_mode(&cache->features))
2832 DMEMIT("1 writeback ");
2835 DMERR("internal error: unknown io mode: %d", (int) cache->features.io_mode);
2839 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2841 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
2843 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2845 DMERR("policy_emit_config_values returned %d", r);
2850 case STATUSTYPE_TABLE:
2851 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2853 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2855 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2858 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2859 DMEMIT(" %s", cache->ctr_args[i]);
2860 if (cache->nr_ctr_args)
2861 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2871 * A cache block range can take two forms:
2873 * i) A single cblock, eg. '3456'
2874 * ii) A begin and end cblock with dots between, eg. 123-234
2876 static int parse_cblock_range(struct cache *cache, const char *str,
2877 struct cblock_range *result)
2884 * Try and parse form (ii) first.
2886 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
2891 result->begin = to_cblock(b);
2892 result->end = to_cblock(e);
2897 * That didn't work, try form (i).
2899 r = sscanf(str, "%llu%c", &b, &dummy);
2904 result->begin = to_cblock(b);
2905 result->end = to_cblock(from_cblock(result->begin) + 1u);
2909 DMERR("invalid cblock range '%s'", str);
2913 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
2915 uint64_t b = from_cblock(range->begin);
2916 uint64_t e = from_cblock(range->end);
2917 uint64_t n = from_cblock(cache->cache_size);
2920 DMERR("begin cblock out of range: %llu >= %llu", b, n);
2925 DMERR("end cblock out of range: %llu > %llu", e, n);
2930 DMERR("invalid cblock range: %llu >= %llu", b, e);
2937 static int request_invalidation(struct cache *cache, struct cblock_range *range)
2939 struct invalidation_request req;
2941 INIT_LIST_HEAD(&req.list);
2942 req.cblocks = range;
2943 atomic_set(&req.complete, 0);
2945 init_waitqueue_head(&req.result_wait);
2947 spin_lock(&cache->invalidation_lock);
2948 list_add(&req.list, &cache->invalidation_requests);
2949 spin_unlock(&cache->invalidation_lock);
2952 wait_event(req.result_wait, atomic_read(&req.complete));
2956 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
2957 const char **cblock_ranges)
2961 struct cblock_range range;
2963 if (!passthrough_mode(&cache->features)) {
2964 DMERR("cache has to be in passthrough mode for invalidation");
2968 for (i = 0; i < count; i++) {
2969 r = parse_cblock_range(cache, cblock_ranges[i], &range);
2973 r = validate_cblock_range(cache, &range);
2978 * Pass begin and end origin blocks to the worker and wake it.
2980 r = request_invalidation(cache, &range);
2992 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
2994 * The key migration_threshold is supported by the cache target core.
2996 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2998 struct cache *cache = ti->private;
3003 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3004 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3009 return set_config_value(cache, argv[0], argv[1]);
3012 static int cache_iterate_devices(struct dm_target *ti,
3013 iterate_devices_callout_fn fn, void *data)
3016 struct cache *cache = ti->private;
3018 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3020 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3026 * We assume I/O is going to the origin (which is the volume
3027 * more likely to have restrictions e.g. by being striped).
3028 * (Looking up the exact location of the data would be expensive
3029 * and could always be out of date by the time the bio is submitted.)
3031 static int cache_bvec_merge(struct dm_target *ti,
3032 struct bvec_merge_data *bvm,
3033 struct bio_vec *biovec, int max_size)
3035 struct cache *cache = ti->private;
3036 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3038 if (!q->merge_bvec_fn)
3041 bvm->bi_bdev = cache->origin_dev->bdev;
3042 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3045 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3048 * FIXME: these limits may be incompatible with the cache device
3050 limits->max_discard_sectors = cache->sectors_per_block;
3051 limits->discard_granularity = cache->sectors_per_block << SECTOR_SHIFT;
3054 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3056 struct cache *cache = ti->private;
3057 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3060 * If the system-determined stacked limits are compatible with the
3061 * cache's blocksize (io_opt is a factor) do not override them.
3063 if (io_opt_sectors < cache->sectors_per_block ||
3064 do_div(io_opt_sectors, cache->sectors_per_block)) {
3065 blk_limits_io_min(limits, 0);
3066 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3068 set_discard_limits(cache, limits);
3071 /*----------------------------------------------------------------*/
3073 static struct target_type cache_target = {
3075 .version = {1, 4, 0},
3076 .module = THIS_MODULE,
3080 .end_io = cache_end_io,
3081 .postsuspend = cache_postsuspend,
3082 .preresume = cache_preresume,
3083 .resume = cache_resume,
3084 .status = cache_status,
3085 .message = cache_message,
3086 .iterate_devices = cache_iterate_devices,
3087 .merge = cache_bvec_merge,
3088 .io_hints = cache_io_hints,
3091 static int __init dm_cache_init(void)
3095 r = dm_register_target(&cache_target);
3097 DMERR("cache target registration failed: %d", r);
3101 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3102 if (!migration_cache) {
3103 dm_unregister_target(&cache_target);
3110 static void __exit dm_cache_exit(void)
3112 dm_unregister_target(&cache_target);
3113 kmem_cache_destroy(migration_cache);
3116 module_init(dm_cache_init);
3117 module_exit(dm_cache_exit);
3119 MODULE_DESCRIPTION(DM_NAME " cache target");
3120 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3121 MODULE_LICENSE("GPL");