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_dblock_t discard_nr_blocks;
241 unsigned long *discard_bitset;
242 uint32_t discard_block_size;
245 * Rather than reconstructing the table line for the status we just
246 * save it and regurgitate.
248 unsigned nr_ctr_args;
249 const char **ctr_args;
251 struct dm_kcopyd_client *copier;
252 struct workqueue_struct *wq;
253 struct work_struct worker;
255 struct delayed_work waker;
256 unsigned long last_commit_jiffies;
258 struct dm_bio_prison *prison;
259 struct dm_deferred_set *all_io_ds;
261 mempool_t *migration_pool;
262 struct dm_cache_migration *next_migration;
264 struct dm_cache_policy *policy;
265 unsigned policy_nr_args;
267 bool need_tick_bio:1;
270 bool commit_requested:1;
271 bool loaded_mappings:1;
272 bool loaded_discards:1;
275 * Cache features such as write-through.
277 struct cache_features features;
279 struct cache_stats stats;
282 * Invalidation fields.
284 spinlock_t invalidation_lock;
285 struct list_head invalidation_requests;
288 struct per_bio_data {
291 struct dm_deferred_entry *all_io_entry;
292 struct dm_hook_info hook_info;
295 * writethrough fields. These MUST remain at the end of this
296 * structure and the 'cache' member must be the first as it
297 * is used to determine the offset of the writethrough fields.
301 struct dm_bio_details bio_details;
304 struct dm_cache_migration {
305 struct list_head list;
308 unsigned long start_jiffies;
309 dm_oblock_t old_oblock;
310 dm_oblock_t new_oblock;
317 bool requeue_holder:1;
320 struct dm_bio_prison_cell *old_ocell;
321 struct dm_bio_prison_cell *new_ocell;
325 * Processing a bio in the worker thread may require these memory
326 * allocations. We prealloc to avoid deadlocks (the same worker thread
327 * frees them back to the mempool).
330 struct dm_cache_migration *mg;
331 struct dm_bio_prison_cell *cell1;
332 struct dm_bio_prison_cell *cell2;
335 static void wake_worker(struct cache *cache)
337 queue_work(cache->wq, &cache->worker);
340 /*----------------------------------------------------------------*/
342 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
344 /* FIXME: change to use a local slab. */
345 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
348 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
350 dm_bio_prison_free_cell(cache->prison, cell);
353 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
356 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
362 p->cell1 = alloc_prison_cell(cache);
368 p->cell2 = alloc_prison_cell(cache);
376 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
379 free_prison_cell(cache, p->cell2);
382 free_prison_cell(cache, p->cell1);
385 mempool_free(p->mg, cache->migration_pool);
388 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
390 struct dm_cache_migration *mg = p->mg;
399 * You must have a cell within the prealloc struct to return. If not this
400 * function will BUG() rather than returning NULL.
402 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
404 struct dm_bio_prison_cell *r = NULL;
410 } else if (p->cell2) {
420 * You can't have more than two cells in a prealloc struct. BUG() will be
421 * called if you try and overfill.
423 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
435 /*----------------------------------------------------------------*/
437 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
441 key->block = from_oblock(oblock);
445 * The caller hands in a preallocated cell, and a free function for it.
446 * The cell will be freed if there's an error, or if it wasn't used because
447 * a cell with that key already exists.
449 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
451 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
452 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
453 cell_free_fn free_fn, void *free_context,
454 struct dm_bio_prison_cell **cell_result)
457 struct dm_cell_key key;
459 build_key(oblock, &key);
460 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
462 free_fn(free_context, cell_prealloc);
467 static int get_cell(struct cache *cache,
469 struct prealloc *structs,
470 struct dm_bio_prison_cell **cell_result)
473 struct dm_cell_key key;
474 struct dm_bio_prison_cell *cell_prealloc;
476 cell_prealloc = prealloc_get_cell(structs);
478 build_key(oblock, &key);
479 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
481 prealloc_put_cell(structs, cell_prealloc);
486 /*----------------------------------------------------------------*/
488 static bool is_dirty(struct cache *cache, dm_cblock_t b)
490 return test_bit(from_cblock(b), cache->dirty_bitset);
493 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
495 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
496 atomic_inc(&cache->nr_dirty);
497 policy_set_dirty(cache->policy, oblock);
501 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
503 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
504 policy_clear_dirty(cache->policy, oblock);
505 if (atomic_dec_return(&cache->nr_dirty) == 0)
506 dm_table_event(cache->ti->table);
510 /*----------------------------------------------------------------*/
512 static bool block_size_is_power_of_two(struct cache *cache)
514 return cache->sectors_per_block_shift >= 0;
517 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
518 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
521 static dm_block_t block_div(dm_block_t b, uint32_t n)
528 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
530 uint32_t discard_blocks = cache->discard_block_size;
531 dm_block_t b = from_oblock(oblock);
533 if (!block_size_is_power_of_two(cache))
534 discard_blocks = discard_blocks / cache->sectors_per_block;
536 discard_blocks >>= cache->sectors_per_block_shift;
538 b = block_div(b, discard_blocks);
543 static void set_discard(struct cache *cache, dm_dblock_t b)
547 atomic_inc(&cache->stats.discard_count);
549 spin_lock_irqsave(&cache->lock, flags);
550 set_bit(from_dblock(b), cache->discard_bitset);
551 spin_unlock_irqrestore(&cache->lock, flags);
554 static void clear_discard(struct cache *cache, dm_dblock_t b)
558 spin_lock_irqsave(&cache->lock, flags);
559 clear_bit(from_dblock(b), cache->discard_bitset);
560 spin_unlock_irqrestore(&cache->lock, flags);
563 static bool is_discarded(struct cache *cache, dm_dblock_t b)
568 spin_lock_irqsave(&cache->lock, flags);
569 r = test_bit(from_dblock(b), cache->discard_bitset);
570 spin_unlock_irqrestore(&cache->lock, flags);
575 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
580 spin_lock_irqsave(&cache->lock, flags);
581 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
582 cache->discard_bitset);
583 spin_unlock_irqrestore(&cache->lock, flags);
588 /*----------------------------------------------------------------*/
590 static void load_stats(struct cache *cache)
592 struct dm_cache_statistics stats;
594 dm_cache_metadata_get_stats(cache->cmd, &stats);
595 atomic_set(&cache->stats.read_hit, stats.read_hits);
596 atomic_set(&cache->stats.read_miss, stats.read_misses);
597 atomic_set(&cache->stats.write_hit, stats.write_hits);
598 atomic_set(&cache->stats.write_miss, stats.write_misses);
601 static void save_stats(struct cache *cache)
603 struct dm_cache_statistics stats;
605 stats.read_hits = atomic_read(&cache->stats.read_hit);
606 stats.read_misses = atomic_read(&cache->stats.read_miss);
607 stats.write_hits = atomic_read(&cache->stats.write_hit);
608 stats.write_misses = atomic_read(&cache->stats.write_miss);
610 dm_cache_metadata_set_stats(cache->cmd, &stats);
613 /*----------------------------------------------------------------
615 *--------------------------------------------------------------*/
618 * If using writeback, leave out struct per_bio_data's writethrough fields.
620 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
621 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
623 static bool writethrough_mode(struct cache_features *f)
625 return f->io_mode == CM_IO_WRITETHROUGH;
628 static bool writeback_mode(struct cache_features *f)
630 return f->io_mode == CM_IO_WRITEBACK;
633 static bool passthrough_mode(struct cache_features *f)
635 return f->io_mode == CM_IO_PASSTHROUGH;
638 static size_t get_per_bio_data_size(struct cache *cache)
640 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
643 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
645 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
650 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
652 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
655 pb->req_nr = dm_bio_get_target_bio_nr(bio);
656 pb->all_io_entry = NULL;
661 /*----------------------------------------------------------------
663 *--------------------------------------------------------------*/
664 static void remap_to_origin(struct cache *cache, struct bio *bio)
666 bio->bi_bdev = cache->origin_dev->bdev;
669 static void remap_to_cache(struct cache *cache, struct bio *bio,
672 sector_t bi_sector = bio->bi_iter.bi_sector;
673 sector_t block = from_cblock(cblock);
675 bio->bi_bdev = cache->cache_dev->bdev;
676 if (!block_size_is_power_of_two(cache))
677 bio->bi_iter.bi_sector =
678 (block * cache->sectors_per_block) +
679 sector_div(bi_sector, cache->sectors_per_block);
681 bio->bi_iter.bi_sector =
682 (block << cache->sectors_per_block_shift) |
683 (bi_sector & (cache->sectors_per_block - 1));
686 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
689 size_t pb_data_size = get_per_bio_data_size(cache);
690 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
692 spin_lock_irqsave(&cache->lock, flags);
693 if (cache->need_tick_bio &&
694 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
696 cache->need_tick_bio = false;
698 spin_unlock_irqrestore(&cache->lock, flags);
701 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
704 check_if_tick_bio_needed(cache, bio);
705 remap_to_origin(cache, bio);
706 if (bio_data_dir(bio) == WRITE)
707 clear_discard(cache, oblock_to_dblock(cache, oblock));
710 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
711 dm_oblock_t oblock, dm_cblock_t cblock)
713 check_if_tick_bio_needed(cache, bio);
714 remap_to_cache(cache, bio, cblock);
715 if (bio_data_dir(bio) == WRITE) {
716 set_dirty(cache, oblock, cblock);
717 clear_discard(cache, oblock_to_dblock(cache, oblock));
721 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
723 sector_t block_nr = bio->bi_iter.bi_sector;
725 if (!block_size_is_power_of_two(cache))
726 (void) sector_div(block_nr, cache->sectors_per_block);
728 block_nr >>= cache->sectors_per_block_shift;
730 return to_oblock(block_nr);
733 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
735 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
738 static void issue(struct cache *cache, struct bio *bio)
742 if (!bio_triggers_commit(cache, bio)) {
743 generic_make_request(bio);
748 * Batch together any bios that trigger commits and then issue a
749 * single commit for them in do_worker().
751 spin_lock_irqsave(&cache->lock, flags);
752 cache->commit_requested = true;
753 bio_list_add(&cache->deferred_flush_bios, bio);
754 spin_unlock_irqrestore(&cache->lock, flags);
757 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
761 spin_lock_irqsave(&cache->lock, flags);
762 bio_list_add(&cache->deferred_writethrough_bios, bio);
763 spin_unlock_irqrestore(&cache->lock, flags);
768 static void writethrough_endio(struct bio *bio, int err)
770 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
772 dm_unhook_bio(&pb->hook_info, bio);
779 dm_bio_restore(&pb->bio_details, bio);
780 remap_to_cache(pb->cache, bio, pb->cblock);
783 * We can't issue this bio directly, since we're in interrupt
784 * context. So it gets put on a bio list for processing by the
787 defer_writethrough_bio(pb->cache, bio);
791 * When running in writethrough mode we need to send writes to clean blocks
792 * to both the cache and origin devices. In future we'd like to clone the
793 * bio and send them in parallel, but for now we're doing them in
794 * series as this is easier.
796 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
797 dm_oblock_t oblock, dm_cblock_t cblock)
799 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
803 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
804 dm_bio_record(&pb->bio_details, bio);
806 remap_to_origin_clear_discard(pb->cache, bio, oblock);
809 /*----------------------------------------------------------------
810 * Migration processing
812 * Migration covers moving data from the origin device to the cache, or
814 *--------------------------------------------------------------*/
815 static void free_migration(struct dm_cache_migration *mg)
817 mempool_free(mg, mg->cache->migration_pool);
820 static void inc_nr_migrations(struct cache *cache)
822 atomic_inc(&cache->nr_migrations);
825 static void dec_nr_migrations(struct cache *cache)
827 atomic_dec(&cache->nr_migrations);
830 * Wake the worker in case we're suspending the target.
832 wake_up(&cache->migration_wait);
835 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
838 (holder ? dm_cell_release : dm_cell_release_no_holder)
839 (cache->prison, cell, &cache->deferred_bios);
840 free_prison_cell(cache, cell);
843 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
848 spin_lock_irqsave(&cache->lock, flags);
849 __cell_defer(cache, cell, holder);
850 spin_unlock_irqrestore(&cache->lock, flags);
855 static void cleanup_migration(struct dm_cache_migration *mg)
857 struct cache *cache = mg->cache;
859 dec_nr_migrations(cache);
862 static void migration_failure(struct dm_cache_migration *mg)
864 struct cache *cache = mg->cache;
867 DMWARN_LIMIT("writeback failed; couldn't copy block");
868 set_dirty(cache, mg->old_oblock, mg->cblock);
869 cell_defer(cache, mg->old_ocell, false);
871 } else if (mg->demote) {
872 DMWARN_LIMIT("demotion failed; couldn't copy block");
873 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
875 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
877 cell_defer(cache, mg->new_ocell, true);
879 DMWARN_LIMIT("promotion failed; couldn't copy block");
880 policy_remove_mapping(cache->policy, mg->new_oblock);
881 cell_defer(cache, mg->new_ocell, true);
884 cleanup_migration(mg);
887 static void migration_success_pre_commit(struct dm_cache_migration *mg)
890 struct cache *cache = mg->cache;
893 clear_dirty(cache, mg->old_oblock, mg->cblock);
894 cell_defer(cache, mg->old_ocell, false);
895 cleanup_migration(mg);
898 } else if (mg->demote) {
899 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
900 DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
901 policy_force_mapping(cache->policy, mg->new_oblock,
904 cell_defer(cache, mg->new_ocell, true);
905 cleanup_migration(mg);
909 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
910 DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
911 policy_remove_mapping(cache->policy, mg->new_oblock);
912 cleanup_migration(mg);
917 spin_lock_irqsave(&cache->lock, flags);
918 list_add_tail(&mg->list, &cache->need_commit_migrations);
919 cache->commit_requested = true;
920 spin_unlock_irqrestore(&cache->lock, flags);
923 static void migration_success_post_commit(struct dm_cache_migration *mg)
926 struct cache *cache = mg->cache;
929 DMWARN("writeback unexpectedly triggered commit");
932 } else if (mg->demote) {
933 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
938 spin_lock_irqsave(&cache->lock, flags);
939 list_add_tail(&mg->list, &cache->quiesced_migrations);
940 spin_unlock_irqrestore(&cache->lock, flags);
944 policy_remove_mapping(cache->policy, mg->old_oblock);
945 cleanup_migration(mg);
949 clear_dirty(cache, mg->new_oblock, mg->cblock);
950 if (mg->requeue_holder)
951 cell_defer(cache, mg->new_ocell, true);
953 bio_endio(mg->new_ocell->holder, 0);
954 cell_defer(cache, mg->new_ocell, false);
956 cleanup_migration(mg);
960 static void copy_complete(int read_err, unsigned long write_err, void *context)
963 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
964 struct cache *cache = mg->cache;
966 if (read_err || write_err)
969 spin_lock_irqsave(&cache->lock, flags);
970 list_add_tail(&mg->list, &cache->completed_migrations);
971 spin_unlock_irqrestore(&cache->lock, flags);
976 static void issue_copy_real(struct dm_cache_migration *mg)
979 struct dm_io_region o_region, c_region;
980 struct cache *cache = mg->cache;
981 sector_t cblock = from_cblock(mg->cblock);
983 o_region.bdev = cache->origin_dev->bdev;
984 o_region.count = cache->sectors_per_block;
986 c_region.bdev = cache->cache_dev->bdev;
987 c_region.sector = cblock * cache->sectors_per_block;
988 c_region.count = cache->sectors_per_block;
990 if (mg->writeback || mg->demote) {
992 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
993 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
996 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
997 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1001 DMERR_LIMIT("issuing migration failed");
1002 migration_failure(mg);
1006 static void overwrite_endio(struct bio *bio, int err)
1008 struct dm_cache_migration *mg = bio->bi_private;
1009 struct cache *cache = mg->cache;
1010 size_t pb_data_size = get_per_bio_data_size(cache);
1011 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1012 unsigned long flags;
1014 dm_unhook_bio(&pb->hook_info, bio);
1019 mg->requeue_holder = false;
1021 spin_lock_irqsave(&cache->lock, flags);
1022 list_add_tail(&mg->list, &cache->completed_migrations);
1023 spin_unlock_irqrestore(&cache->lock, flags);
1028 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1030 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1031 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1033 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1034 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1035 generic_make_request(bio);
1038 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1040 return (bio_data_dir(bio) == WRITE) &&
1041 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1044 static void avoid_copy(struct dm_cache_migration *mg)
1046 atomic_inc(&mg->cache->stats.copies_avoided);
1047 migration_success_pre_commit(mg);
1050 static void issue_copy(struct dm_cache_migration *mg)
1053 struct cache *cache = mg->cache;
1055 if (mg->writeback || mg->demote)
1056 avoid = !is_dirty(cache, mg->cblock) ||
1057 is_discarded_oblock(cache, mg->old_oblock);
1059 struct bio *bio = mg->new_ocell->holder;
1061 avoid = is_discarded_oblock(cache, mg->new_oblock);
1063 if (!avoid && bio_writes_complete_block(cache, bio)) {
1064 issue_overwrite(mg, bio);
1069 avoid ? avoid_copy(mg) : issue_copy_real(mg);
1072 static void complete_migration(struct dm_cache_migration *mg)
1075 migration_failure(mg);
1077 migration_success_pre_commit(mg);
1080 static void process_migrations(struct cache *cache, struct list_head *head,
1081 void (*fn)(struct dm_cache_migration *))
1083 unsigned long flags;
1084 struct list_head list;
1085 struct dm_cache_migration *mg, *tmp;
1087 INIT_LIST_HEAD(&list);
1088 spin_lock_irqsave(&cache->lock, flags);
1089 list_splice_init(head, &list);
1090 spin_unlock_irqrestore(&cache->lock, flags);
1092 list_for_each_entry_safe(mg, tmp, &list, list)
1096 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1098 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1101 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1103 unsigned long flags;
1104 struct cache *cache = mg->cache;
1106 spin_lock_irqsave(&cache->lock, flags);
1107 __queue_quiesced_migration(mg);
1108 spin_unlock_irqrestore(&cache->lock, flags);
1113 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1115 unsigned long flags;
1116 struct dm_cache_migration *mg, *tmp;
1118 spin_lock_irqsave(&cache->lock, flags);
1119 list_for_each_entry_safe(mg, tmp, work, list)
1120 __queue_quiesced_migration(mg);
1121 spin_unlock_irqrestore(&cache->lock, flags);
1126 static void check_for_quiesced_migrations(struct cache *cache,
1127 struct per_bio_data *pb)
1129 struct list_head work;
1131 if (!pb->all_io_entry)
1134 INIT_LIST_HEAD(&work);
1135 if (pb->all_io_entry)
1136 dm_deferred_entry_dec(pb->all_io_entry, &work);
1138 if (!list_empty(&work))
1139 queue_quiesced_migrations(cache, &work);
1142 static void quiesce_migration(struct dm_cache_migration *mg)
1144 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1145 queue_quiesced_migration(mg);
1148 static void promote(struct cache *cache, struct prealloc *structs,
1149 dm_oblock_t oblock, dm_cblock_t cblock,
1150 struct dm_bio_prison_cell *cell)
1152 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1155 mg->writeback = false;
1158 mg->requeue_holder = true;
1159 mg->invalidate = false;
1161 mg->new_oblock = oblock;
1162 mg->cblock = cblock;
1163 mg->old_ocell = NULL;
1164 mg->new_ocell = cell;
1165 mg->start_jiffies = jiffies;
1167 inc_nr_migrations(cache);
1168 quiesce_migration(mg);
1171 static void writeback(struct cache *cache, struct prealloc *structs,
1172 dm_oblock_t oblock, dm_cblock_t cblock,
1173 struct dm_bio_prison_cell *cell)
1175 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1178 mg->writeback = true;
1180 mg->promote = false;
1181 mg->requeue_holder = true;
1182 mg->invalidate = false;
1184 mg->old_oblock = oblock;
1185 mg->cblock = cblock;
1186 mg->old_ocell = cell;
1187 mg->new_ocell = NULL;
1188 mg->start_jiffies = jiffies;
1190 inc_nr_migrations(cache);
1191 quiesce_migration(mg);
1194 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1195 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1197 struct dm_bio_prison_cell *old_ocell,
1198 struct dm_bio_prison_cell *new_ocell)
1200 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1203 mg->writeback = false;
1206 mg->requeue_holder = true;
1207 mg->invalidate = false;
1209 mg->old_oblock = old_oblock;
1210 mg->new_oblock = new_oblock;
1211 mg->cblock = cblock;
1212 mg->old_ocell = old_ocell;
1213 mg->new_ocell = new_ocell;
1214 mg->start_jiffies = jiffies;
1216 inc_nr_migrations(cache);
1217 quiesce_migration(mg);
1221 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1222 * block are thrown away.
1224 static void invalidate(struct cache *cache, struct prealloc *structs,
1225 dm_oblock_t oblock, dm_cblock_t cblock,
1226 struct dm_bio_prison_cell *cell)
1228 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1231 mg->writeback = false;
1233 mg->promote = false;
1234 mg->requeue_holder = true;
1235 mg->invalidate = true;
1237 mg->old_oblock = oblock;
1238 mg->cblock = cblock;
1239 mg->old_ocell = cell;
1240 mg->new_ocell = NULL;
1241 mg->start_jiffies = jiffies;
1243 inc_nr_migrations(cache);
1244 quiesce_migration(mg);
1247 /*----------------------------------------------------------------
1249 *--------------------------------------------------------------*/
1250 static void defer_bio(struct cache *cache, struct bio *bio)
1252 unsigned long flags;
1254 spin_lock_irqsave(&cache->lock, flags);
1255 bio_list_add(&cache->deferred_bios, bio);
1256 spin_unlock_irqrestore(&cache->lock, flags);
1261 static void process_flush_bio(struct cache *cache, struct bio *bio)
1263 size_t pb_data_size = get_per_bio_data_size(cache);
1264 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1266 BUG_ON(bio->bi_iter.bi_size);
1268 remap_to_origin(cache, bio);
1270 remap_to_cache(cache, bio, 0);
1276 * People generally discard large parts of a device, eg, the whole device
1277 * when formatting. Splitting these large discards up into cache block
1278 * sized ios and then quiescing (always neccessary for discard) takes too
1281 * We keep it simple, and allow any size of discard to come in, and just
1282 * mark off blocks on the discard bitset. No passdown occurs!
1284 * To implement passdown we need to change the bio_prison such that a cell
1285 * can have a key that spans many blocks.
1287 static void process_discard_bio(struct cache *cache, struct bio *bio)
1289 dm_block_t start_block = dm_sector_div_up(bio->bi_iter.bi_sector,
1290 cache->discard_block_size);
1291 dm_block_t end_block = bio_end_sector(bio);
1294 end_block = block_div(end_block, cache->discard_block_size);
1296 for (b = start_block; b < end_block; b++)
1297 set_discard(cache, to_dblock(b));
1302 static bool spare_migration_bandwidth(struct cache *cache)
1304 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1305 cache->sectors_per_block;
1306 return current_volume < cache->migration_threshold;
1309 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1311 atomic_inc(bio_data_dir(bio) == READ ?
1312 &cache->stats.read_hit : &cache->stats.write_hit);
1315 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1317 atomic_inc(bio_data_dir(bio) == READ ?
1318 &cache->stats.read_miss : &cache->stats.write_miss);
1321 static void issue_cache_bio(struct cache *cache, struct bio *bio,
1322 struct per_bio_data *pb,
1323 dm_oblock_t oblock, dm_cblock_t cblock)
1325 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1326 remap_to_cache_dirty(cache, bio, oblock, cblock);
1330 static void process_bio(struct cache *cache, struct prealloc *structs,
1334 bool release_cell = true;
1335 dm_oblock_t block = get_bio_block(cache, bio);
1336 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1337 struct policy_result lookup_result;
1338 size_t pb_data_size = get_per_bio_data_size(cache);
1339 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1340 bool discarded_block = is_discarded_oblock(cache, block);
1341 bool passthrough = passthrough_mode(&cache->features);
1342 bool can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
1345 * Check to see if that block is currently migrating.
1347 cell_prealloc = prealloc_get_cell(structs);
1348 r = bio_detain(cache, block, bio, cell_prealloc,
1349 (cell_free_fn) prealloc_put_cell,
1350 structs, &new_ocell);
1354 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1355 bio, &lookup_result);
1357 if (r == -EWOULDBLOCK)
1358 /* migration has been denied */
1359 lookup_result.op = POLICY_MISS;
1361 switch (lookup_result.op) {
1364 inc_miss_counter(cache, bio);
1367 * Passthrough always maps to the origin,
1368 * invalidating any cache blocks that are written
1372 if (bio_data_dir(bio) == WRITE) {
1373 atomic_inc(&cache->stats.demotion);
1374 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1375 release_cell = false;
1378 /* FIXME: factor out issue_origin() */
1379 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1380 remap_to_origin_clear_discard(cache, bio, block);
1384 inc_hit_counter(cache, bio);
1386 if (bio_data_dir(bio) == WRITE &&
1387 writethrough_mode(&cache->features) &&
1388 !is_dirty(cache, lookup_result.cblock)) {
1389 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1390 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1393 issue_cache_bio(cache, bio, pb, block, lookup_result.cblock);
1399 inc_miss_counter(cache, bio);
1400 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1401 remap_to_origin_clear_discard(cache, bio, block);
1406 atomic_inc(&cache->stats.promotion);
1407 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1408 release_cell = false;
1411 case POLICY_REPLACE:
1412 cell_prealloc = prealloc_get_cell(structs);
1413 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1414 (cell_free_fn) prealloc_put_cell,
1415 structs, &old_ocell);
1418 * We have to be careful to avoid lock inversion of
1419 * the cells. So we back off, and wait for the
1420 * old_ocell to become free.
1422 policy_force_mapping(cache->policy, block,
1423 lookup_result.old_oblock);
1424 atomic_inc(&cache->stats.cache_cell_clash);
1427 atomic_inc(&cache->stats.demotion);
1428 atomic_inc(&cache->stats.promotion);
1430 demote_then_promote(cache, structs, lookup_result.old_oblock,
1431 block, lookup_result.cblock,
1432 old_ocell, new_ocell);
1433 release_cell = false;
1437 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1438 (unsigned) lookup_result.op);
1443 cell_defer(cache, new_ocell, false);
1446 static int need_commit_due_to_time(struct cache *cache)
1448 return jiffies < cache->last_commit_jiffies ||
1449 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1452 static int commit_if_needed(struct cache *cache)
1456 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1457 dm_cache_changed_this_transaction(cache->cmd)) {
1458 atomic_inc(&cache->stats.commit_count);
1459 cache->commit_requested = false;
1460 r = dm_cache_commit(cache->cmd, false);
1461 cache->last_commit_jiffies = jiffies;
1467 static void process_deferred_bios(struct cache *cache)
1469 unsigned long flags;
1470 struct bio_list bios;
1472 struct prealloc structs;
1474 memset(&structs, 0, sizeof(structs));
1475 bio_list_init(&bios);
1477 spin_lock_irqsave(&cache->lock, flags);
1478 bio_list_merge(&bios, &cache->deferred_bios);
1479 bio_list_init(&cache->deferred_bios);
1480 spin_unlock_irqrestore(&cache->lock, flags);
1482 while (!bio_list_empty(&bios)) {
1484 * If we've got no free migration structs, and processing
1485 * this bio might require one, we pause until there are some
1486 * prepared mappings to process.
1488 if (prealloc_data_structs(cache, &structs)) {
1489 spin_lock_irqsave(&cache->lock, flags);
1490 bio_list_merge(&cache->deferred_bios, &bios);
1491 spin_unlock_irqrestore(&cache->lock, flags);
1495 bio = bio_list_pop(&bios);
1497 if (bio->bi_rw & REQ_FLUSH)
1498 process_flush_bio(cache, bio);
1499 else if (bio->bi_rw & REQ_DISCARD)
1500 process_discard_bio(cache, bio);
1502 process_bio(cache, &structs, bio);
1505 prealloc_free_structs(cache, &structs);
1508 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
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_flush_bios);
1518 bio_list_init(&cache->deferred_flush_bios);
1519 spin_unlock_irqrestore(&cache->lock, flags);
1521 while ((bio = bio_list_pop(&bios)))
1522 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1525 static void process_deferred_writethrough_bios(struct cache *cache)
1527 unsigned long flags;
1528 struct bio_list bios;
1531 bio_list_init(&bios);
1533 spin_lock_irqsave(&cache->lock, flags);
1534 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1535 bio_list_init(&cache->deferred_writethrough_bios);
1536 spin_unlock_irqrestore(&cache->lock, flags);
1538 while ((bio = bio_list_pop(&bios)))
1539 generic_make_request(bio);
1542 static void writeback_some_dirty_blocks(struct cache *cache)
1547 struct prealloc structs;
1548 struct dm_bio_prison_cell *old_ocell;
1550 memset(&structs, 0, sizeof(structs));
1552 while (spare_migration_bandwidth(cache)) {
1553 if (prealloc_data_structs(cache, &structs))
1556 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1560 r = get_cell(cache, oblock, &structs, &old_ocell);
1562 policy_set_dirty(cache->policy, oblock);
1566 writeback(cache, &structs, oblock, cblock, old_ocell);
1569 prealloc_free_structs(cache, &structs);
1572 /*----------------------------------------------------------------
1574 * Dropping something from the cache *without* writing back.
1575 *--------------------------------------------------------------*/
1577 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
1580 uint64_t begin = from_cblock(req->cblocks->begin);
1581 uint64_t end = from_cblock(req->cblocks->end);
1583 while (begin != end) {
1584 r = policy_remove_cblock(cache->policy, to_cblock(begin));
1586 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
1590 } else if (r == -ENODATA) {
1591 /* harmless, already unmapped */
1595 DMERR("policy_remove_cblock failed");
1602 cache->commit_requested = true;
1605 atomic_set(&req->complete, 1);
1607 wake_up(&req->result_wait);
1610 static void process_invalidation_requests(struct cache *cache)
1612 struct list_head list;
1613 struct invalidation_request *req, *tmp;
1615 INIT_LIST_HEAD(&list);
1616 spin_lock(&cache->invalidation_lock);
1617 list_splice_init(&cache->invalidation_requests, &list);
1618 spin_unlock(&cache->invalidation_lock);
1620 list_for_each_entry_safe (req, tmp, &list, list)
1621 process_invalidation_request(cache, req);
1624 /*----------------------------------------------------------------
1626 *--------------------------------------------------------------*/
1627 static bool is_quiescing(struct cache *cache)
1629 return atomic_read(&cache->quiescing);
1632 static void ack_quiescing(struct cache *cache)
1634 if (is_quiescing(cache)) {
1635 atomic_inc(&cache->quiescing_ack);
1636 wake_up(&cache->quiescing_wait);
1640 static void wait_for_quiescing_ack(struct cache *cache)
1642 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
1645 static void start_quiescing(struct cache *cache)
1647 atomic_inc(&cache->quiescing);
1648 wait_for_quiescing_ack(cache);
1651 static void stop_quiescing(struct cache *cache)
1653 atomic_set(&cache->quiescing, 0);
1654 atomic_set(&cache->quiescing_ack, 0);
1657 static void wait_for_migrations(struct cache *cache)
1659 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1662 static void stop_worker(struct cache *cache)
1664 cancel_delayed_work(&cache->waker);
1665 flush_workqueue(cache->wq);
1668 static void requeue_deferred_io(struct cache *cache)
1671 struct bio_list bios;
1673 bio_list_init(&bios);
1674 bio_list_merge(&bios, &cache->deferred_bios);
1675 bio_list_init(&cache->deferred_bios);
1677 while ((bio = bio_list_pop(&bios)))
1678 bio_endio(bio, DM_ENDIO_REQUEUE);
1681 static int more_work(struct cache *cache)
1683 if (is_quiescing(cache))
1684 return !list_empty(&cache->quiesced_migrations) ||
1685 !list_empty(&cache->completed_migrations) ||
1686 !list_empty(&cache->need_commit_migrations);
1688 return !bio_list_empty(&cache->deferred_bios) ||
1689 !bio_list_empty(&cache->deferred_flush_bios) ||
1690 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1691 !list_empty(&cache->quiesced_migrations) ||
1692 !list_empty(&cache->completed_migrations) ||
1693 !list_empty(&cache->need_commit_migrations) ||
1697 static void do_worker(struct work_struct *ws)
1699 struct cache *cache = container_of(ws, struct cache, worker);
1702 if (!is_quiescing(cache)) {
1703 writeback_some_dirty_blocks(cache);
1704 process_deferred_writethrough_bios(cache);
1705 process_deferred_bios(cache);
1706 process_invalidation_requests(cache);
1709 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1710 process_migrations(cache, &cache->completed_migrations, complete_migration);
1712 if (commit_if_needed(cache)) {
1713 process_deferred_flush_bios(cache, false);
1716 * FIXME: rollback metadata or just go into a
1717 * failure mode and error everything
1720 process_deferred_flush_bios(cache, true);
1721 process_migrations(cache, &cache->need_commit_migrations,
1722 migration_success_post_commit);
1725 ack_quiescing(cache);
1727 } while (more_work(cache));
1731 * We want to commit periodically so that not too much
1732 * unwritten metadata builds up.
1734 static void do_waker(struct work_struct *ws)
1736 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1737 policy_tick(cache->policy);
1739 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1742 /*----------------------------------------------------------------*/
1744 static int is_congested(struct dm_dev *dev, int bdi_bits)
1746 struct request_queue *q = bdev_get_queue(dev->bdev);
1747 return bdi_congested(&q->backing_dev_info, bdi_bits);
1750 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1752 struct cache *cache = container_of(cb, struct cache, callbacks);
1754 return is_congested(cache->origin_dev, bdi_bits) ||
1755 is_congested(cache->cache_dev, bdi_bits);
1758 /*----------------------------------------------------------------
1760 *--------------------------------------------------------------*/
1763 * This function gets called on the error paths of the constructor, so we
1764 * have to cope with a partially initialised struct.
1766 static void destroy(struct cache *cache)
1770 if (cache->next_migration)
1771 mempool_free(cache->next_migration, cache->migration_pool);
1773 if (cache->migration_pool)
1774 mempool_destroy(cache->migration_pool);
1776 if (cache->all_io_ds)
1777 dm_deferred_set_destroy(cache->all_io_ds);
1780 dm_bio_prison_destroy(cache->prison);
1783 destroy_workqueue(cache->wq);
1785 if (cache->dirty_bitset)
1786 free_bitset(cache->dirty_bitset);
1788 if (cache->discard_bitset)
1789 free_bitset(cache->discard_bitset);
1792 dm_kcopyd_client_destroy(cache->copier);
1795 dm_cache_metadata_close(cache->cmd);
1797 if (cache->metadata_dev)
1798 dm_put_device(cache->ti, cache->metadata_dev);
1800 if (cache->origin_dev)
1801 dm_put_device(cache->ti, cache->origin_dev);
1803 if (cache->cache_dev)
1804 dm_put_device(cache->ti, cache->cache_dev);
1807 dm_cache_policy_destroy(cache->policy);
1809 for (i = 0; i < cache->nr_ctr_args ; i++)
1810 kfree(cache->ctr_args[i]);
1811 kfree(cache->ctr_args);
1816 static void cache_dtr(struct dm_target *ti)
1818 struct cache *cache = ti->private;
1823 static sector_t get_dev_size(struct dm_dev *dev)
1825 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1828 /*----------------------------------------------------------------*/
1831 * Construct a cache device mapping.
1833 * cache <metadata dev> <cache dev> <origin dev> <block size>
1834 * <#feature args> [<feature arg>]*
1835 * <policy> <#policy args> [<policy arg>]*
1837 * metadata dev : fast device holding the persistent metadata
1838 * cache dev : fast device holding cached data blocks
1839 * origin dev : slow device holding original data blocks
1840 * block size : cache unit size in sectors
1842 * #feature args : number of feature arguments passed
1843 * feature args : writethrough. (The default is writeback.)
1845 * policy : the replacement policy to use
1846 * #policy args : an even number of policy arguments corresponding
1847 * to key/value pairs passed to the policy
1848 * policy args : key/value pairs passed to the policy
1849 * E.g. 'sequential_threshold 1024'
1850 * See cache-policies.txt for details.
1852 * Optional feature arguments are:
1853 * writethrough : write through caching that prohibits cache block
1854 * content from being different from origin block content.
1855 * Without this argument, the default behaviour is to write
1856 * back cache block contents later for performance reasons,
1857 * so they may differ from the corresponding origin blocks.
1860 struct dm_target *ti;
1862 struct dm_dev *metadata_dev;
1864 struct dm_dev *cache_dev;
1865 sector_t cache_sectors;
1867 struct dm_dev *origin_dev;
1868 sector_t origin_sectors;
1870 uint32_t block_size;
1872 const char *policy_name;
1874 const char **policy_argv;
1876 struct cache_features features;
1879 static void destroy_cache_args(struct cache_args *ca)
1881 if (ca->metadata_dev)
1882 dm_put_device(ca->ti, ca->metadata_dev);
1885 dm_put_device(ca->ti, ca->cache_dev);
1888 dm_put_device(ca->ti, ca->origin_dev);
1893 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1896 *error = "Insufficient args";
1903 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1907 sector_t metadata_dev_size;
1908 char b[BDEVNAME_SIZE];
1910 if (!at_least_one_arg(as, error))
1913 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1916 *error = "Error opening metadata device";
1920 metadata_dev_size = get_dev_size(ca->metadata_dev);
1921 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1922 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1923 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1928 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1933 if (!at_least_one_arg(as, error))
1936 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1939 *error = "Error opening cache device";
1942 ca->cache_sectors = get_dev_size(ca->cache_dev);
1947 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1952 if (!at_least_one_arg(as, error))
1955 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1958 *error = "Error opening origin device";
1962 ca->origin_sectors = get_dev_size(ca->origin_dev);
1963 if (ca->ti->len > ca->origin_sectors) {
1964 *error = "Device size larger than cached device";
1971 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1974 unsigned long block_size;
1976 if (!at_least_one_arg(as, error))
1979 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
1980 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1981 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1982 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1983 *error = "Invalid data block size";
1987 if (block_size > ca->cache_sectors) {
1988 *error = "Data block size is larger than the cache device";
1992 ca->block_size = block_size;
1997 static void init_features(struct cache_features *cf)
1999 cf->mode = CM_WRITE;
2000 cf->io_mode = CM_IO_WRITEBACK;
2003 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2006 static struct dm_arg _args[] = {
2007 {0, 1, "Invalid number of cache feature arguments"},
2013 struct cache_features *cf = &ca->features;
2017 r = dm_read_arg_group(_args, as, &argc, error);
2022 arg = dm_shift_arg(as);
2024 if (!strcasecmp(arg, "writeback"))
2025 cf->io_mode = CM_IO_WRITEBACK;
2027 else if (!strcasecmp(arg, "writethrough"))
2028 cf->io_mode = CM_IO_WRITETHROUGH;
2030 else if (!strcasecmp(arg, "passthrough"))
2031 cf->io_mode = CM_IO_PASSTHROUGH;
2034 *error = "Unrecognised cache feature requested";
2042 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2045 static struct dm_arg _args[] = {
2046 {0, 1024, "Invalid number of policy arguments"},
2051 if (!at_least_one_arg(as, error))
2054 ca->policy_name = dm_shift_arg(as);
2056 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2060 ca->policy_argv = (const char **)as->argv;
2061 dm_consume_args(as, ca->policy_argc);
2066 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2070 struct dm_arg_set as;
2075 r = parse_metadata_dev(ca, &as, error);
2079 r = parse_cache_dev(ca, &as, error);
2083 r = parse_origin_dev(ca, &as, error);
2087 r = parse_block_size(ca, &as, error);
2091 r = parse_features(ca, &as, error);
2095 r = parse_policy(ca, &as, error);
2102 /*----------------------------------------------------------------*/
2104 static struct kmem_cache *migration_cache;
2106 #define NOT_CORE_OPTION 1
2108 static int process_config_option(struct cache *cache, const char *key, const char *value)
2112 if (!strcasecmp(key, "migration_threshold")) {
2113 if (kstrtoul(value, 10, &tmp))
2116 cache->migration_threshold = tmp;
2120 return NOT_CORE_OPTION;
2123 static int set_config_value(struct cache *cache, const char *key, const char *value)
2125 int r = process_config_option(cache, key, value);
2127 if (r == NOT_CORE_OPTION)
2128 r = policy_set_config_value(cache->policy, key, value);
2131 DMWARN("bad config value for %s: %s", key, value);
2136 static int set_config_values(struct cache *cache, int argc, const char **argv)
2141 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2146 r = set_config_value(cache, argv[0], argv[1]);
2157 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2160 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2162 cache->origin_sectors,
2163 cache->sectors_per_block);
2165 *error = "Error creating cache's policy";
2173 #define DEFAULT_MIGRATION_THRESHOLD 2048
2175 static int cache_create(struct cache_args *ca, struct cache **result)
2178 char **error = &ca->ti->error;
2179 struct cache *cache;
2180 struct dm_target *ti = ca->ti;
2181 dm_block_t origin_blocks;
2182 struct dm_cache_metadata *cmd;
2183 bool may_format = ca->features.mode == CM_WRITE;
2185 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2190 ti->private = cache;
2191 ti->num_flush_bios = 2;
2192 ti->flush_supported = true;
2194 ti->num_discard_bios = 1;
2195 ti->discards_supported = true;
2196 ti->discard_zeroes_data_unsupported = true;
2197 /* Discard bios must be split on a block boundary */
2198 ti->split_discard_bios = true;
2200 cache->features = ca->features;
2201 ti->per_bio_data_size = get_per_bio_data_size(cache);
2203 cache->callbacks.congested_fn = cache_is_congested;
2204 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2206 cache->metadata_dev = ca->metadata_dev;
2207 cache->origin_dev = ca->origin_dev;
2208 cache->cache_dev = ca->cache_dev;
2210 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2212 /* FIXME: factor out this whole section */
2213 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2214 origin_blocks = block_div(origin_blocks, ca->block_size);
2215 cache->origin_blocks = to_oblock(origin_blocks);
2217 cache->sectors_per_block = ca->block_size;
2218 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2223 if (ca->block_size & (ca->block_size - 1)) {
2224 dm_block_t cache_size = ca->cache_sectors;
2226 cache->sectors_per_block_shift = -1;
2227 cache_size = block_div(cache_size, ca->block_size);
2228 cache->cache_size = to_cblock(cache_size);
2230 cache->sectors_per_block_shift = __ffs(ca->block_size);
2231 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
2234 r = create_cache_policy(cache, ca, error);
2238 cache->policy_nr_args = ca->policy_argc;
2239 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2241 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2243 *error = "Error setting cache policy's config values";
2247 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2248 ca->block_size, may_format,
2249 dm_cache_policy_get_hint_size(cache->policy));
2251 *error = "Error creating metadata object";
2257 if (passthrough_mode(&cache->features)) {
2260 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2262 *error = "dm_cache_metadata_all_clean() failed";
2267 *error = "Cannot enter passthrough mode unless all blocks are clean";
2273 spin_lock_init(&cache->lock);
2274 bio_list_init(&cache->deferred_bios);
2275 bio_list_init(&cache->deferred_flush_bios);
2276 bio_list_init(&cache->deferred_writethrough_bios);
2277 INIT_LIST_HEAD(&cache->quiesced_migrations);
2278 INIT_LIST_HEAD(&cache->completed_migrations);
2279 INIT_LIST_HEAD(&cache->need_commit_migrations);
2280 atomic_set(&cache->nr_migrations, 0);
2281 init_waitqueue_head(&cache->migration_wait);
2283 init_waitqueue_head(&cache->quiescing_wait);
2284 atomic_set(&cache->quiescing, 0);
2285 atomic_set(&cache->quiescing_ack, 0);
2288 atomic_set(&cache->nr_dirty, 0);
2289 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2290 if (!cache->dirty_bitset) {
2291 *error = "could not allocate dirty bitset";
2294 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2296 cache->discard_block_size = cache->sectors_per_block;
2297 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2298 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2299 if (!cache->discard_bitset) {
2300 *error = "could not allocate discard bitset";
2303 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2305 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2306 if (IS_ERR(cache->copier)) {
2307 *error = "could not create kcopyd client";
2308 r = PTR_ERR(cache->copier);
2312 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2314 *error = "could not create workqueue for metadata object";
2317 INIT_WORK(&cache->worker, do_worker);
2318 INIT_DELAYED_WORK(&cache->waker, do_waker);
2319 cache->last_commit_jiffies = jiffies;
2321 cache->prison = dm_bio_prison_create(PRISON_CELLS);
2322 if (!cache->prison) {
2323 *error = "could not create bio prison";
2327 cache->all_io_ds = dm_deferred_set_create();
2328 if (!cache->all_io_ds) {
2329 *error = "could not create all_io deferred set";
2333 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2335 if (!cache->migration_pool) {
2336 *error = "Error creating cache's migration mempool";
2340 cache->next_migration = NULL;
2342 cache->need_tick_bio = true;
2343 cache->sized = false;
2344 cache->invalidate = false;
2345 cache->commit_requested = false;
2346 cache->loaded_mappings = false;
2347 cache->loaded_discards = false;
2351 atomic_set(&cache->stats.demotion, 0);
2352 atomic_set(&cache->stats.promotion, 0);
2353 atomic_set(&cache->stats.copies_avoided, 0);
2354 atomic_set(&cache->stats.cache_cell_clash, 0);
2355 atomic_set(&cache->stats.commit_count, 0);
2356 atomic_set(&cache->stats.discard_count, 0);
2358 spin_lock_init(&cache->invalidation_lock);
2359 INIT_LIST_HEAD(&cache->invalidation_requests);
2369 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2374 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2377 for (i = 0; i < argc; i++) {
2378 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2387 cache->nr_ctr_args = argc;
2388 cache->ctr_args = copy;
2393 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2396 struct cache_args *ca;
2397 struct cache *cache = NULL;
2399 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2401 ti->error = "Error allocating memory for cache";
2406 r = parse_cache_args(ca, argc, argv, &ti->error);
2410 r = cache_create(ca, &cache);
2414 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2420 ti->private = cache;
2423 destroy_cache_args(ca);
2427 static int cache_map(struct dm_target *ti, struct bio *bio)
2429 struct cache *cache = ti->private;
2432 dm_oblock_t block = get_bio_block(cache, bio);
2433 size_t pb_data_size = get_per_bio_data_size(cache);
2434 bool can_migrate = false;
2435 bool discarded_block;
2436 struct dm_bio_prison_cell *cell;
2437 struct policy_result lookup_result;
2438 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
2440 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2442 * This can only occur if the io goes to a partial block at
2443 * the end of the origin device. We don't cache these.
2444 * Just remap to the origin and carry on.
2446 remap_to_origin(cache, bio);
2447 return DM_MAPIO_REMAPPED;
2450 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2451 defer_bio(cache, bio);
2452 return DM_MAPIO_SUBMITTED;
2456 * Check to see if that block is currently migrating.
2458 cell = alloc_prison_cell(cache);
2460 defer_bio(cache, bio);
2461 return DM_MAPIO_SUBMITTED;
2464 r = bio_detain(cache, block, bio, cell,
2465 (cell_free_fn) free_prison_cell,
2469 defer_bio(cache, bio);
2471 return DM_MAPIO_SUBMITTED;
2474 discarded_block = is_discarded_oblock(cache, block);
2476 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2477 bio, &lookup_result);
2478 if (r == -EWOULDBLOCK) {
2479 cell_defer(cache, cell, true);
2480 return DM_MAPIO_SUBMITTED;
2483 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2485 return DM_MAPIO_SUBMITTED;
2488 r = DM_MAPIO_REMAPPED;
2489 switch (lookup_result.op) {
2491 if (passthrough_mode(&cache->features)) {
2492 if (bio_data_dir(bio) == WRITE) {
2494 * We need to invalidate this block, so
2495 * defer for the worker thread.
2497 cell_defer(cache, cell, true);
2498 r = DM_MAPIO_SUBMITTED;
2501 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2502 inc_miss_counter(cache, bio);
2503 remap_to_origin_clear_discard(cache, bio, block);
2505 cell_defer(cache, cell, false);
2509 inc_hit_counter(cache, bio);
2510 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2512 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
2513 !is_dirty(cache, lookup_result.cblock))
2514 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2516 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2518 cell_defer(cache, cell, false);
2523 inc_miss_counter(cache, bio);
2524 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2526 if (pb->req_nr != 0) {
2528 * This is a duplicate writethrough io that is no
2529 * longer needed because the block has been demoted.
2532 cell_defer(cache, cell, false);
2533 return DM_MAPIO_SUBMITTED;
2535 remap_to_origin_clear_discard(cache, bio, block);
2536 cell_defer(cache, cell, false);
2541 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2542 (unsigned) lookup_result.op);
2544 r = DM_MAPIO_SUBMITTED;
2550 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2552 struct cache *cache = ti->private;
2553 unsigned long flags;
2554 size_t pb_data_size = get_per_bio_data_size(cache);
2555 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2558 policy_tick(cache->policy);
2560 spin_lock_irqsave(&cache->lock, flags);
2561 cache->need_tick_bio = true;
2562 spin_unlock_irqrestore(&cache->lock, flags);
2565 check_for_quiesced_migrations(cache, pb);
2570 static int write_dirty_bitset(struct cache *cache)
2574 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2575 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2576 is_dirty(cache, to_cblock(i)));
2584 static int write_discard_bitset(struct cache *cache)
2588 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2589 cache->discard_nr_blocks);
2591 DMERR("could not resize on-disk discard bitset");
2595 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2596 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2597 is_discarded(cache, to_dblock(i)));
2606 * returns true on success
2608 static bool sync_metadata(struct cache *cache)
2612 r1 = write_dirty_bitset(cache);
2614 DMERR("could not write dirty bitset");
2616 r2 = write_discard_bitset(cache);
2618 DMERR("could not write discard bitset");
2622 r3 = dm_cache_write_hints(cache->cmd, cache->policy);
2624 DMERR("could not write hints");
2627 * If writing the above metadata failed, we still commit, but don't
2628 * set the clean shutdown flag. This will effectively force every
2629 * dirty bit to be set on reload.
2631 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2633 DMERR("could not write cache metadata. Data loss may occur.");
2635 return !r1 && !r2 && !r3 && !r4;
2638 static void cache_postsuspend(struct dm_target *ti)
2640 struct cache *cache = ti->private;
2642 start_quiescing(cache);
2643 wait_for_migrations(cache);
2645 requeue_deferred_io(cache);
2646 stop_quiescing(cache);
2648 (void) sync_metadata(cache);
2651 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2652 bool dirty, uint32_t hint, bool hint_valid)
2655 struct cache *cache = context;
2657 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2662 set_dirty(cache, oblock, cblock);
2664 clear_dirty(cache, oblock, cblock);
2669 static int load_discard(void *context, sector_t discard_block_size,
2670 dm_dblock_t dblock, bool discard)
2672 struct cache *cache = context;
2674 /* FIXME: handle mis-matched block size */
2677 set_discard(cache, dblock);
2679 clear_discard(cache, dblock);
2684 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2686 sector_t size = get_dev_size(cache->cache_dev);
2687 (void) sector_div(size, cache->sectors_per_block);
2688 return to_cblock(size);
2691 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2693 if (from_cblock(new_size) > from_cblock(cache->cache_size))
2697 * We can't drop a dirty block when shrinking the cache.
2699 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2700 new_size = to_cblock(from_cblock(new_size) + 1);
2701 if (is_dirty(cache, new_size)) {
2702 DMERR("unable to shrink cache; cache block %llu is dirty",
2703 (unsigned long long) from_cblock(new_size));
2711 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2715 r = dm_cache_resize(cache->cmd, new_size);
2717 DMERR("could not resize cache metadata");
2721 cache->cache_size = new_size;
2726 static int cache_preresume(struct dm_target *ti)
2729 struct cache *cache = ti->private;
2730 dm_cblock_t csize = get_cache_dev_size(cache);
2733 * Check to see if the cache has resized.
2735 if (!cache->sized) {
2736 r = resize_cache_dev(cache, csize);
2740 cache->sized = true;
2742 } else if (csize != cache->cache_size) {
2743 if (!can_resize(cache, csize))
2746 r = resize_cache_dev(cache, csize);
2751 if (!cache->loaded_mappings) {
2752 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2753 load_mapping, cache);
2755 DMERR("could not load cache mappings");
2759 cache->loaded_mappings = true;
2762 if (!cache->loaded_discards) {
2763 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2765 DMERR("could not load origin discards");
2769 cache->loaded_discards = true;
2775 static void cache_resume(struct dm_target *ti)
2777 struct cache *cache = ti->private;
2779 cache->need_tick_bio = true;
2780 do_waker(&cache->waker.work);
2786 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
2787 * <cache block size> <#used cache blocks>/<#total cache blocks>
2788 * <#read hits> <#read misses> <#write hits> <#write misses>
2789 * <#demotions> <#promotions> <#dirty>
2790 * <#features> <features>*
2791 * <#core args> <core args>
2792 * <policy name> <#policy args> <policy args>*
2794 static void cache_status(struct dm_target *ti, status_type_t type,
2795 unsigned status_flags, char *result, unsigned maxlen)
2800 dm_block_t nr_free_blocks_metadata = 0;
2801 dm_block_t nr_blocks_metadata = 0;
2802 char buf[BDEVNAME_SIZE];
2803 struct cache *cache = ti->private;
2804 dm_cblock_t residency;
2807 case STATUSTYPE_INFO:
2808 /* Commit to ensure statistics aren't out-of-date */
2809 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2810 r = dm_cache_commit(cache->cmd, false);
2812 DMERR("could not commit metadata for accurate status");
2815 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2816 &nr_free_blocks_metadata);
2818 DMERR("could not get metadata free block count");
2822 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2824 DMERR("could not get metadata device size");
2828 residency = policy_residency(cache->policy);
2830 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
2831 (unsigned)(DM_CACHE_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
2832 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2833 (unsigned long long)nr_blocks_metadata,
2834 cache->sectors_per_block,
2835 (unsigned long long) from_cblock(residency),
2836 (unsigned long long) from_cblock(cache->cache_size),
2837 (unsigned) atomic_read(&cache->stats.read_hit),
2838 (unsigned) atomic_read(&cache->stats.read_miss),
2839 (unsigned) atomic_read(&cache->stats.write_hit),
2840 (unsigned) atomic_read(&cache->stats.write_miss),
2841 (unsigned) atomic_read(&cache->stats.demotion),
2842 (unsigned) atomic_read(&cache->stats.promotion),
2843 (unsigned long) atomic_read(&cache->nr_dirty));
2845 if (writethrough_mode(&cache->features))
2846 DMEMIT("1 writethrough ");
2848 else if (passthrough_mode(&cache->features))
2849 DMEMIT("1 passthrough ");
2851 else if (writeback_mode(&cache->features))
2852 DMEMIT("1 writeback ");
2855 DMERR("internal error: unknown io mode: %d", (int) cache->features.io_mode);
2859 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2861 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
2863 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2865 DMERR("policy_emit_config_values returned %d", r);
2870 case STATUSTYPE_TABLE:
2871 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2873 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2875 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2878 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2879 DMEMIT(" %s", cache->ctr_args[i]);
2880 if (cache->nr_ctr_args)
2881 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2891 * A cache block range can take two forms:
2893 * i) A single cblock, eg. '3456'
2894 * ii) A begin and end cblock with dots between, eg. 123-234
2896 static int parse_cblock_range(struct cache *cache, const char *str,
2897 struct cblock_range *result)
2904 * Try and parse form (ii) first.
2906 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
2911 result->begin = to_cblock(b);
2912 result->end = to_cblock(e);
2917 * That didn't work, try form (i).
2919 r = sscanf(str, "%llu%c", &b, &dummy);
2924 result->begin = to_cblock(b);
2925 result->end = to_cblock(from_cblock(result->begin) + 1u);
2929 DMERR("invalid cblock range '%s'", str);
2933 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
2935 uint64_t b = from_cblock(range->begin);
2936 uint64_t e = from_cblock(range->end);
2937 uint64_t n = from_cblock(cache->cache_size);
2940 DMERR("begin cblock out of range: %llu >= %llu", b, n);
2945 DMERR("end cblock out of range: %llu > %llu", e, n);
2950 DMERR("invalid cblock range: %llu >= %llu", b, e);
2957 static int request_invalidation(struct cache *cache, struct cblock_range *range)
2959 struct invalidation_request req;
2961 INIT_LIST_HEAD(&req.list);
2962 req.cblocks = range;
2963 atomic_set(&req.complete, 0);
2965 init_waitqueue_head(&req.result_wait);
2967 spin_lock(&cache->invalidation_lock);
2968 list_add(&req.list, &cache->invalidation_requests);
2969 spin_unlock(&cache->invalidation_lock);
2972 wait_event(req.result_wait, atomic_read(&req.complete));
2976 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
2977 const char **cblock_ranges)
2981 struct cblock_range range;
2983 if (!passthrough_mode(&cache->features)) {
2984 DMERR("cache has to be in passthrough mode for invalidation");
2988 for (i = 0; i < count; i++) {
2989 r = parse_cblock_range(cache, cblock_ranges[i], &range);
2993 r = validate_cblock_range(cache, &range);
2998 * Pass begin and end origin blocks to the worker and wake it.
3000 r = request_invalidation(cache, &range);
3012 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3014 * The key migration_threshold is supported by the cache target core.
3016 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3018 struct cache *cache = ti->private;
3023 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3024 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3029 return set_config_value(cache, argv[0], argv[1]);
3032 static int cache_iterate_devices(struct dm_target *ti,
3033 iterate_devices_callout_fn fn, void *data)
3036 struct cache *cache = ti->private;
3038 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3040 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3046 * We assume I/O is going to the origin (which is the volume
3047 * more likely to have restrictions e.g. by being striped).
3048 * (Looking up the exact location of the data would be expensive
3049 * and could always be out of date by the time the bio is submitted.)
3051 static int cache_bvec_merge(struct dm_target *ti,
3052 struct bvec_merge_data *bvm,
3053 struct bio_vec *biovec, int max_size)
3055 struct cache *cache = ti->private;
3056 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3058 if (!q->merge_bvec_fn)
3061 bvm->bi_bdev = cache->origin_dev->bdev;
3062 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3065 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3068 * FIXME: these limits may be incompatible with the cache device
3070 limits->max_discard_sectors = cache->discard_block_size;
3071 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3074 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3076 struct cache *cache = ti->private;
3077 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3080 * If the system-determined stacked limits are compatible with the
3081 * cache's blocksize (io_opt is a factor) do not override them.
3083 if (io_opt_sectors < cache->sectors_per_block ||
3084 do_div(io_opt_sectors, cache->sectors_per_block)) {
3085 blk_limits_io_min(limits, 0);
3086 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3088 set_discard_limits(cache, limits);
3091 /*----------------------------------------------------------------*/
3093 static struct target_type cache_target = {
3095 .version = {1, 4, 0},
3096 .module = THIS_MODULE,
3100 .end_io = cache_end_io,
3101 .postsuspend = cache_postsuspend,
3102 .preresume = cache_preresume,
3103 .resume = cache_resume,
3104 .status = cache_status,
3105 .message = cache_message,
3106 .iterate_devices = cache_iterate_devices,
3107 .merge = cache_bvec_merge,
3108 .io_hints = cache_io_hints,
3111 static int __init dm_cache_init(void)
3115 r = dm_register_target(&cache_target);
3117 DMERR("cache target registration failed: %d", r);
3121 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3122 if (!migration_cache) {
3123 dm_unregister_target(&cache_target);
3130 static void __exit dm_cache_exit(void)
3132 dm_unregister_target(&cache_target);
3133 kmem_cache_destroy(migration_cache);
3136 module_init(dm_cache_init);
3137 module_exit(dm_cache_exit);
3139 MODULE_DESCRIPTION(DM_NAME " cache target");
3140 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3141 MODULE_LICENSE("GPL");