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 if (mg->requeue_holder) {
950 clear_dirty(cache, mg->new_oblock, mg->cblock);
951 cell_defer(cache, mg->new_ocell, true);
954 * The block was promoted via an overwrite, so it's dirty.
956 set_dirty(cache, mg->new_oblock, mg->cblock);
957 bio_endio(mg->new_ocell->holder, 0);
958 cell_defer(cache, mg->new_ocell, false);
960 cleanup_migration(mg);
964 static void copy_complete(int read_err, unsigned long write_err, void *context)
967 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
968 struct cache *cache = mg->cache;
970 if (read_err || write_err)
973 spin_lock_irqsave(&cache->lock, flags);
974 list_add_tail(&mg->list, &cache->completed_migrations);
975 spin_unlock_irqrestore(&cache->lock, flags);
980 static void issue_copy_real(struct dm_cache_migration *mg)
983 struct dm_io_region o_region, c_region;
984 struct cache *cache = mg->cache;
985 sector_t cblock = from_cblock(mg->cblock);
987 o_region.bdev = cache->origin_dev->bdev;
988 o_region.count = cache->sectors_per_block;
990 c_region.bdev = cache->cache_dev->bdev;
991 c_region.sector = cblock * cache->sectors_per_block;
992 c_region.count = cache->sectors_per_block;
994 if (mg->writeback || mg->demote) {
996 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
997 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1000 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1001 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1005 DMERR_LIMIT("issuing migration failed");
1006 migration_failure(mg);
1010 static void overwrite_endio(struct bio *bio, int err)
1012 struct dm_cache_migration *mg = bio->bi_private;
1013 struct cache *cache = mg->cache;
1014 size_t pb_data_size = get_per_bio_data_size(cache);
1015 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1016 unsigned long flags;
1018 dm_unhook_bio(&pb->hook_info, bio);
1023 mg->requeue_holder = false;
1025 spin_lock_irqsave(&cache->lock, flags);
1026 list_add_tail(&mg->list, &cache->completed_migrations);
1027 spin_unlock_irqrestore(&cache->lock, flags);
1032 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1034 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1035 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1037 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1038 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1039 generic_make_request(bio);
1042 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1044 return (bio_data_dir(bio) == WRITE) &&
1045 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1048 static void avoid_copy(struct dm_cache_migration *mg)
1050 atomic_inc(&mg->cache->stats.copies_avoided);
1051 migration_success_pre_commit(mg);
1054 static void issue_copy(struct dm_cache_migration *mg)
1057 struct cache *cache = mg->cache;
1059 if (mg->writeback || mg->demote)
1060 avoid = !is_dirty(cache, mg->cblock) ||
1061 is_discarded_oblock(cache, mg->old_oblock);
1063 struct bio *bio = mg->new_ocell->holder;
1065 avoid = is_discarded_oblock(cache, mg->new_oblock);
1067 if (writeback_mode(&cache->features) &&
1068 !avoid && bio_writes_complete_block(cache, bio)) {
1069 issue_overwrite(mg, bio);
1074 avoid ? avoid_copy(mg) : issue_copy_real(mg);
1077 static void complete_migration(struct dm_cache_migration *mg)
1080 migration_failure(mg);
1082 migration_success_pre_commit(mg);
1085 static void process_migrations(struct cache *cache, struct list_head *head,
1086 void (*fn)(struct dm_cache_migration *))
1088 unsigned long flags;
1089 struct list_head list;
1090 struct dm_cache_migration *mg, *tmp;
1092 INIT_LIST_HEAD(&list);
1093 spin_lock_irqsave(&cache->lock, flags);
1094 list_splice_init(head, &list);
1095 spin_unlock_irqrestore(&cache->lock, flags);
1097 list_for_each_entry_safe(mg, tmp, &list, list)
1101 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1103 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1106 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1108 unsigned long flags;
1109 struct cache *cache = mg->cache;
1111 spin_lock_irqsave(&cache->lock, flags);
1112 __queue_quiesced_migration(mg);
1113 spin_unlock_irqrestore(&cache->lock, flags);
1118 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1120 unsigned long flags;
1121 struct dm_cache_migration *mg, *tmp;
1123 spin_lock_irqsave(&cache->lock, flags);
1124 list_for_each_entry_safe(mg, tmp, work, list)
1125 __queue_quiesced_migration(mg);
1126 spin_unlock_irqrestore(&cache->lock, flags);
1131 static void check_for_quiesced_migrations(struct cache *cache,
1132 struct per_bio_data *pb)
1134 struct list_head work;
1136 if (!pb->all_io_entry)
1139 INIT_LIST_HEAD(&work);
1140 if (pb->all_io_entry)
1141 dm_deferred_entry_dec(pb->all_io_entry, &work);
1143 if (!list_empty(&work))
1144 queue_quiesced_migrations(cache, &work);
1147 static void quiesce_migration(struct dm_cache_migration *mg)
1149 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1150 queue_quiesced_migration(mg);
1153 static void promote(struct cache *cache, struct prealloc *structs,
1154 dm_oblock_t oblock, dm_cblock_t cblock,
1155 struct dm_bio_prison_cell *cell)
1157 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1160 mg->writeback = false;
1163 mg->requeue_holder = true;
1164 mg->invalidate = false;
1166 mg->new_oblock = oblock;
1167 mg->cblock = cblock;
1168 mg->old_ocell = NULL;
1169 mg->new_ocell = cell;
1170 mg->start_jiffies = jiffies;
1172 inc_nr_migrations(cache);
1173 quiesce_migration(mg);
1176 static void writeback(struct cache *cache, struct prealloc *structs,
1177 dm_oblock_t oblock, dm_cblock_t cblock,
1178 struct dm_bio_prison_cell *cell)
1180 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1183 mg->writeback = true;
1185 mg->promote = false;
1186 mg->requeue_holder = true;
1187 mg->invalidate = false;
1189 mg->old_oblock = oblock;
1190 mg->cblock = cblock;
1191 mg->old_ocell = cell;
1192 mg->new_ocell = NULL;
1193 mg->start_jiffies = jiffies;
1195 inc_nr_migrations(cache);
1196 quiesce_migration(mg);
1199 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1200 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1202 struct dm_bio_prison_cell *old_ocell,
1203 struct dm_bio_prison_cell *new_ocell)
1205 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1208 mg->writeback = false;
1211 mg->requeue_holder = true;
1212 mg->invalidate = false;
1214 mg->old_oblock = old_oblock;
1215 mg->new_oblock = new_oblock;
1216 mg->cblock = cblock;
1217 mg->old_ocell = old_ocell;
1218 mg->new_ocell = new_ocell;
1219 mg->start_jiffies = jiffies;
1221 inc_nr_migrations(cache);
1222 quiesce_migration(mg);
1226 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1227 * block are thrown away.
1229 static void invalidate(struct cache *cache, struct prealloc *structs,
1230 dm_oblock_t oblock, dm_cblock_t cblock,
1231 struct dm_bio_prison_cell *cell)
1233 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1236 mg->writeback = false;
1238 mg->promote = false;
1239 mg->requeue_holder = true;
1240 mg->invalidate = true;
1242 mg->old_oblock = oblock;
1243 mg->cblock = cblock;
1244 mg->old_ocell = cell;
1245 mg->new_ocell = NULL;
1246 mg->start_jiffies = jiffies;
1248 inc_nr_migrations(cache);
1249 quiesce_migration(mg);
1252 /*----------------------------------------------------------------
1254 *--------------------------------------------------------------*/
1255 static void defer_bio(struct cache *cache, struct bio *bio)
1257 unsigned long flags;
1259 spin_lock_irqsave(&cache->lock, flags);
1260 bio_list_add(&cache->deferred_bios, bio);
1261 spin_unlock_irqrestore(&cache->lock, flags);
1266 static void process_flush_bio(struct cache *cache, struct bio *bio)
1268 size_t pb_data_size = get_per_bio_data_size(cache);
1269 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1271 BUG_ON(bio->bi_iter.bi_size);
1273 remap_to_origin(cache, bio);
1275 remap_to_cache(cache, bio, 0);
1281 * People generally discard large parts of a device, eg, the whole device
1282 * when formatting. Splitting these large discards up into cache block
1283 * sized ios and then quiescing (always neccessary for discard) takes too
1286 * We keep it simple, and allow any size of discard to come in, and just
1287 * mark off blocks on the discard bitset. No passdown occurs!
1289 * To implement passdown we need to change the bio_prison such that a cell
1290 * can have a key that spans many blocks.
1292 static void process_discard_bio(struct cache *cache, struct bio *bio)
1294 dm_block_t start_block = dm_sector_div_up(bio->bi_iter.bi_sector,
1295 cache->discard_block_size);
1296 dm_block_t end_block = bio_end_sector(bio);
1299 end_block = block_div(end_block, cache->discard_block_size);
1301 for (b = start_block; b < end_block; b++)
1302 set_discard(cache, to_dblock(b));
1307 static bool spare_migration_bandwidth(struct cache *cache)
1309 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1310 cache->sectors_per_block;
1311 return current_volume < cache->migration_threshold;
1314 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1316 atomic_inc(bio_data_dir(bio) == READ ?
1317 &cache->stats.read_hit : &cache->stats.write_hit);
1320 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1322 atomic_inc(bio_data_dir(bio) == READ ?
1323 &cache->stats.read_miss : &cache->stats.write_miss);
1326 static void issue_cache_bio(struct cache *cache, struct bio *bio,
1327 struct per_bio_data *pb,
1328 dm_oblock_t oblock, dm_cblock_t cblock)
1330 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1331 remap_to_cache_dirty(cache, bio, oblock, cblock);
1335 static void process_bio(struct cache *cache, struct prealloc *structs,
1339 bool release_cell = true;
1340 dm_oblock_t block = get_bio_block(cache, bio);
1341 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1342 struct policy_result lookup_result;
1343 size_t pb_data_size = get_per_bio_data_size(cache);
1344 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1345 bool discarded_block = is_discarded_oblock(cache, block);
1346 bool passthrough = passthrough_mode(&cache->features);
1347 bool can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
1350 * Check to see if that block is currently migrating.
1352 cell_prealloc = prealloc_get_cell(structs);
1353 r = bio_detain(cache, block, bio, cell_prealloc,
1354 (cell_free_fn) prealloc_put_cell,
1355 structs, &new_ocell);
1359 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1360 bio, &lookup_result);
1362 if (r == -EWOULDBLOCK)
1363 /* migration has been denied */
1364 lookup_result.op = POLICY_MISS;
1366 switch (lookup_result.op) {
1369 inc_miss_counter(cache, bio);
1372 * Passthrough always maps to the origin,
1373 * invalidating any cache blocks that are written
1377 if (bio_data_dir(bio) == WRITE) {
1378 atomic_inc(&cache->stats.demotion);
1379 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1380 release_cell = false;
1383 /* FIXME: factor out issue_origin() */
1384 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1385 remap_to_origin_clear_discard(cache, bio, block);
1389 inc_hit_counter(cache, bio);
1391 if (bio_data_dir(bio) == WRITE &&
1392 writethrough_mode(&cache->features) &&
1393 !is_dirty(cache, lookup_result.cblock)) {
1394 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1395 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1398 issue_cache_bio(cache, bio, pb, block, lookup_result.cblock);
1404 inc_miss_counter(cache, bio);
1405 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1406 remap_to_origin_clear_discard(cache, bio, block);
1411 atomic_inc(&cache->stats.promotion);
1412 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1413 release_cell = false;
1416 case POLICY_REPLACE:
1417 cell_prealloc = prealloc_get_cell(structs);
1418 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1419 (cell_free_fn) prealloc_put_cell,
1420 structs, &old_ocell);
1423 * We have to be careful to avoid lock inversion of
1424 * the cells. So we back off, and wait for the
1425 * old_ocell to become free.
1427 policy_force_mapping(cache->policy, block,
1428 lookup_result.old_oblock);
1429 atomic_inc(&cache->stats.cache_cell_clash);
1432 atomic_inc(&cache->stats.demotion);
1433 atomic_inc(&cache->stats.promotion);
1435 demote_then_promote(cache, structs, lookup_result.old_oblock,
1436 block, lookup_result.cblock,
1437 old_ocell, new_ocell);
1438 release_cell = false;
1442 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1443 (unsigned) lookup_result.op);
1448 cell_defer(cache, new_ocell, false);
1451 static int need_commit_due_to_time(struct cache *cache)
1453 return jiffies < cache->last_commit_jiffies ||
1454 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1457 static int commit_if_needed(struct cache *cache)
1461 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1462 dm_cache_changed_this_transaction(cache->cmd)) {
1463 atomic_inc(&cache->stats.commit_count);
1464 cache->commit_requested = false;
1465 r = dm_cache_commit(cache->cmd, false);
1466 cache->last_commit_jiffies = jiffies;
1472 static void process_deferred_bios(struct cache *cache)
1474 unsigned long flags;
1475 struct bio_list bios;
1477 struct prealloc structs;
1479 memset(&structs, 0, sizeof(structs));
1480 bio_list_init(&bios);
1482 spin_lock_irqsave(&cache->lock, flags);
1483 bio_list_merge(&bios, &cache->deferred_bios);
1484 bio_list_init(&cache->deferred_bios);
1485 spin_unlock_irqrestore(&cache->lock, flags);
1487 while (!bio_list_empty(&bios)) {
1489 * If we've got no free migration structs, and processing
1490 * this bio might require one, we pause until there are some
1491 * prepared mappings to process.
1493 if (prealloc_data_structs(cache, &structs)) {
1494 spin_lock_irqsave(&cache->lock, flags);
1495 bio_list_merge(&cache->deferred_bios, &bios);
1496 spin_unlock_irqrestore(&cache->lock, flags);
1500 bio = bio_list_pop(&bios);
1502 if (bio->bi_rw & REQ_FLUSH)
1503 process_flush_bio(cache, bio);
1504 else if (bio->bi_rw & REQ_DISCARD)
1505 process_discard_bio(cache, bio);
1507 process_bio(cache, &structs, bio);
1510 prealloc_free_structs(cache, &structs);
1513 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1515 unsigned long flags;
1516 struct bio_list bios;
1519 bio_list_init(&bios);
1521 spin_lock_irqsave(&cache->lock, flags);
1522 bio_list_merge(&bios, &cache->deferred_flush_bios);
1523 bio_list_init(&cache->deferred_flush_bios);
1524 spin_unlock_irqrestore(&cache->lock, flags);
1526 while ((bio = bio_list_pop(&bios)))
1527 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1530 static void process_deferred_writethrough_bios(struct cache *cache)
1532 unsigned long flags;
1533 struct bio_list bios;
1536 bio_list_init(&bios);
1538 spin_lock_irqsave(&cache->lock, flags);
1539 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1540 bio_list_init(&cache->deferred_writethrough_bios);
1541 spin_unlock_irqrestore(&cache->lock, flags);
1543 while ((bio = bio_list_pop(&bios)))
1544 generic_make_request(bio);
1547 static void writeback_some_dirty_blocks(struct cache *cache)
1552 struct prealloc structs;
1553 struct dm_bio_prison_cell *old_ocell;
1555 memset(&structs, 0, sizeof(structs));
1557 while (spare_migration_bandwidth(cache)) {
1558 if (prealloc_data_structs(cache, &structs))
1561 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1565 r = get_cell(cache, oblock, &structs, &old_ocell);
1567 policy_set_dirty(cache->policy, oblock);
1571 writeback(cache, &structs, oblock, cblock, old_ocell);
1574 prealloc_free_structs(cache, &structs);
1577 /*----------------------------------------------------------------
1579 * Dropping something from the cache *without* writing back.
1580 *--------------------------------------------------------------*/
1582 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
1585 uint64_t begin = from_cblock(req->cblocks->begin);
1586 uint64_t end = from_cblock(req->cblocks->end);
1588 while (begin != end) {
1589 r = policy_remove_cblock(cache->policy, to_cblock(begin));
1591 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
1595 } else if (r == -ENODATA) {
1596 /* harmless, already unmapped */
1600 DMERR("policy_remove_cblock failed");
1607 cache->commit_requested = true;
1610 atomic_set(&req->complete, 1);
1612 wake_up(&req->result_wait);
1615 static void process_invalidation_requests(struct cache *cache)
1617 struct list_head list;
1618 struct invalidation_request *req, *tmp;
1620 INIT_LIST_HEAD(&list);
1621 spin_lock(&cache->invalidation_lock);
1622 list_splice_init(&cache->invalidation_requests, &list);
1623 spin_unlock(&cache->invalidation_lock);
1625 list_for_each_entry_safe (req, tmp, &list, list)
1626 process_invalidation_request(cache, req);
1629 /*----------------------------------------------------------------
1631 *--------------------------------------------------------------*/
1632 static bool is_quiescing(struct cache *cache)
1634 return atomic_read(&cache->quiescing);
1637 static void ack_quiescing(struct cache *cache)
1639 if (is_quiescing(cache)) {
1640 atomic_inc(&cache->quiescing_ack);
1641 wake_up(&cache->quiescing_wait);
1645 static void wait_for_quiescing_ack(struct cache *cache)
1647 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
1650 static void start_quiescing(struct cache *cache)
1652 atomic_inc(&cache->quiescing);
1653 wait_for_quiescing_ack(cache);
1656 static void stop_quiescing(struct cache *cache)
1658 atomic_set(&cache->quiescing, 0);
1659 atomic_set(&cache->quiescing_ack, 0);
1662 static void wait_for_migrations(struct cache *cache)
1664 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1667 static void stop_worker(struct cache *cache)
1669 cancel_delayed_work(&cache->waker);
1670 flush_workqueue(cache->wq);
1673 static void requeue_deferred_io(struct cache *cache)
1676 struct bio_list bios;
1678 bio_list_init(&bios);
1679 bio_list_merge(&bios, &cache->deferred_bios);
1680 bio_list_init(&cache->deferred_bios);
1682 while ((bio = bio_list_pop(&bios)))
1683 bio_endio(bio, DM_ENDIO_REQUEUE);
1686 static int more_work(struct cache *cache)
1688 if (is_quiescing(cache))
1689 return !list_empty(&cache->quiesced_migrations) ||
1690 !list_empty(&cache->completed_migrations) ||
1691 !list_empty(&cache->need_commit_migrations);
1693 return !bio_list_empty(&cache->deferred_bios) ||
1694 !bio_list_empty(&cache->deferred_flush_bios) ||
1695 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1696 !list_empty(&cache->quiesced_migrations) ||
1697 !list_empty(&cache->completed_migrations) ||
1698 !list_empty(&cache->need_commit_migrations) ||
1702 static void do_worker(struct work_struct *ws)
1704 struct cache *cache = container_of(ws, struct cache, worker);
1707 if (!is_quiescing(cache)) {
1708 writeback_some_dirty_blocks(cache);
1709 process_deferred_writethrough_bios(cache);
1710 process_deferred_bios(cache);
1711 process_invalidation_requests(cache);
1714 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1715 process_migrations(cache, &cache->completed_migrations, complete_migration);
1717 if (commit_if_needed(cache)) {
1718 process_deferred_flush_bios(cache, false);
1721 * FIXME: rollback metadata or just go into a
1722 * failure mode and error everything
1725 process_deferred_flush_bios(cache, true);
1726 process_migrations(cache, &cache->need_commit_migrations,
1727 migration_success_post_commit);
1730 ack_quiescing(cache);
1732 } while (more_work(cache));
1736 * We want to commit periodically so that not too much
1737 * unwritten metadata builds up.
1739 static void do_waker(struct work_struct *ws)
1741 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1742 policy_tick(cache->policy);
1744 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1747 /*----------------------------------------------------------------*/
1749 static int is_congested(struct dm_dev *dev, int bdi_bits)
1751 struct request_queue *q = bdev_get_queue(dev->bdev);
1752 return bdi_congested(&q->backing_dev_info, bdi_bits);
1755 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1757 struct cache *cache = container_of(cb, struct cache, callbacks);
1759 return is_congested(cache->origin_dev, bdi_bits) ||
1760 is_congested(cache->cache_dev, bdi_bits);
1763 /*----------------------------------------------------------------
1765 *--------------------------------------------------------------*/
1768 * This function gets called on the error paths of the constructor, so we
1769 * have to cope with a partially initialised struct.
1771 static void destroy(struct cache *cache)
1775 if (cache->next_migration)
1776 mempool_free(cache->next_migration, cache->migration_pool);
1778 if (cache->migration_pool)
1779 mempool_destroy(cache->migration_pool);
1781 if (cache->all_io_ds)
1782 dm_deferred_set_destroy(cache->all_io_ds);
1785 dm_bio_prison_destroy(cache->prison);
1788 destroy_workqueue(cache->wq);
1790 if (cache->dirty_bitset)
1791 free_bitset(cache->dirty_bitset);
1793 if (cache->discard_bitset)
1794 free_bitset(cache->discard_bitset);
1797 dm_kcopyd_client_destroy(cache->copier);
1800 dm_cache_metadata_close(cache->cmd);
1802 if (cache->metadata_dev)
1803 dm_put_device(cache->ti, cache->metadata_dev);
1805 if (cache->origin_dev)
1806 dm_put_device(cache->ti, cache->origin_dev);
1808 if (cache->cache_dev)
1809 dm_put_device(cache->ti, cache->cache_dev);
1812 dm_cache_policy_destroy(cache->policy);
1814 for (i = 0; i < cache->nr_ctr_args ; i++)
1815 kfree(cache->ctr_args[i]);
1816 kfree(cache->ctr_args);
1821 static void cache_dtr(struct dm_target *ti)
1823 struct cache *cache = ti->private;
1828 static sector_t get_dev_size(struct dm_dev *dev)
1830 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1833 /*----------------------------------------------------------------*/
1836 * Construct a cache device mapping.
1838 * cache <metadata dev> <cache dev> <origin dev> <block size>
1839 * <#feature args> [<feature arg>]*
1840 * <policy> <#policy args> [<policy arg>]*
1842 * metadata dev : fast device holding the persistent metadata
1843 * cache dev : fast device holding cached data blocks
1844 * origin dev : slow device holding original data blocks
1845 * block size : cache unit size in sectors
1847 * #feature args : number of feature arguments passed
1848 * feature args : writethrough. (The default is writeback.)
1850 * policy : the replacement policy to use
1851 * #policy args : an even number of policy arguments corresponding
1852 * to key/value pairs passed to the policy
1853 * policy args : key/value pairs passed to the policy
1854 * E.g. 'sequential_threshold 1024'
1855 * See cache-policies.txt for details.
1857 * Optional feature arguments are:
1858 * writethrough : write through caching that prohibits cache block
1859 * content from being different from origin block content.
1860 * Without this argument, the default behaviour is to write
1861 * back cache block contents later for performance reasons,
1862 * so they may differ from the corresponding origin blocks.
1865 struct dm_target *ti;
1867 struct dm_dev *metadata_dev;
1869 struct dm_dev *cache_dev;
1870 sector_t cache_sectors;
1872 struct dm_dev *origin_dev;
1873 sector_t origin_sectors;
1875 uint32_t block_size;
1877 const char *policy_name;
1879 const char **policy_argv;
1881 struct cache_features features;
1884 static void destroy_cache_args(struct cache_args *ca)
1886 if (ca->metadata_dev)
1887 dm_put_device(ca->ti, ca->metadata_dev);
1890 dm_put_device(ca->ti, ca->cache_dev);
1893 dm_put_device(ca->ti, ca->origin_dev);
1898 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1901 *error = "Insufficient args";
1908 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1912 sector_t metadata_dev_size;
1913 char b[BDEVNAME_SIZE];
1915 if (!at_least_one_arg(as, error))
1918 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1921 *error = "Error opening metadata device";
1925 metadata_dev_size = get_dev_size(ca->metadata_dev);
1926 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1927 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1928 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1933 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1938 if (!at_least_one_arg(as, error))
1941 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1944 *error = "Error opening cache device";
1947 ca->cache_sectors = get_dev_size(ca->cache_dev);
1952 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1957 if (!at_least_one_arg(as, error))
1960 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1963 *error = "Error opening origin device";
1967 ca->origin_sectors = get_dev_size(ca->origin_dev);
1968 if (ca->ti->len > ca->origin_sectors) {
1969 *error = "Device size larger than cached device";
1976 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1979 unsigned long block_size;
1981 if (!at_least_one_arg(as, error))
1984 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
1985 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1986 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1987 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1988 *error = "Invalid data block size";
1992 if (block_size > ca->cache_sectors) {
1993 *error = "Data block size is larger than the cache device";
1997 ca->block_size = block_size;
2002 static void init_features(struct cache_features *cf)
2004 cf->mode = CM_WRITE;
2005 cf->io_mode = CM_IO_WRITEBACK;
2008 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2011 static struct dm_arg _args[] = {
2012 {0, 1, "Invalid number of cache feature arguments"},
2018 struct cache_features *cf = &ca->features;
2022 r = dm_read_arg_group(_args, as, &argc, error);
2027 arg = dm_shift_arg(as);
2029 if (!strcasecmp(arg, "writeback"))
2030 cf->io_mode = CM_IO_WRITEBACK;
2032 else if (!strcasecmp(arg, "writethrough"))
2033 cf->io_mode = CM_IO_WRITETHROUGH;
2035 else if (!strcasecmp(arg, "passthrough"))
2036 cf->io_mode = CM_IO_PASSTHROUGH;
2039 *error = "Unrecognised cache feature requested";
2047 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2050 static struct dm_arg _args[] = {
2051 {0, 1024, "Invalid number of policy arguments"},
2056 if (!at_least_one_arg(as, error))
2059 ca->policy_name = dm_shift_arg(as);
2061 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2065 ca->policy_argv = (const char **)as->argv;
2066 dm_consume_args(as, ca->policy_argc);
2071 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2075 struct dm_arg_set as;
2080 r = parse_metadata_dev(ca, &as, error);
2084 r = parse_cache_dev(ca, &as, error);
2088 r = parse_origin_dev(ca, &as, error);
2092 r = parse_block_size(ca, &as, error);
2096 r = parse_features(ca, &as, error);
2100 r = parse_policy(ca, &as, error);
2107 /*----------------------------------------------------------------*/
2109 static struct kmem_cache *migration_cache;
2111 #define NOT_CORE_OPTION 1
2113 static int process_config_option(struct cache *cache, const char *key, const char *value)
2117 if (!strcasecmp(key, "migration_threshold")) {
2118 if (kstrtoul(value, 10, &tmp))
2121 cache->migration_threshold = tmp;
2125 return NOT_CORE_OPTION;
2128 static int set_config_value(struct cache *cache, const char *key, const char *value)
2130 int r = process_config_option(cache, key, value);
2132 if (r == NOT_CORE_OPTION)
2133 r = policy_set_config_value(cache->policy, key, value);
2136 DMWARN("bad config value for %s: %s", key, value);
2141 static int set_config_values(struct cache *cache, int argc, const char **argv)
2146 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2151 r = set_config_value(cache, argv[0], argv[1]);
2162 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2165 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2167 cache->origin_sectors,
2168 cache->sectors_per_block);
2170 *error = "Error creating cache's policy";
2178 #define DEFAULT_MIGRATION_THRESHOLD 2048
2180 static int cache_create(struct cache_args *ca, struct cache **result)
2183 char **error = &ca->ti->error;
2184 struct cache *cache;
2185 struct dm_target *ti = ca->ti;
2186 dm_block_t origin_blocks;
2187 struct dm_cache_metadata *cmd;
2188 bool may_format = ca->features.mode == CM_WRITE;
2190 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2195 ti->private = cache;
2196 ti->num_flush_bios = 2;
2197 ti->flush_supported = true;
2199 ti->num_discard_bios = 1;
2200 ti->discards_supported = true;
2201 ti->discard_zeroes_data_unsupported = true;
2202 /* Discard bios must be split on a block boundary */
2203 ti->split_discard_bios = true;
2205 cache->features = ca->features;
2206 ti->per_bio_data_size = get_per_bio_data_size(cache);
2208 cache->callbacks.congested_fn = cache_is_congested;
2209 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2211 cache->metadata_dev = ca->metadata_dev;
2212 cache->origin_dev = ca->origin_dev;
2213 cache->cache_dev = ca->cache_dev;
2215 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2217 /* FIXME: factor out this whole section */
2218 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2219 origin_blocks = block_div(origin_blocks, ca->block_size);
2220 cache->origin_blocks = to_oblock(origin_blocks);
2222 cache->sectors_per_block = ca->block_size;
2223 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2228 if (ca->block_size & (ca->block_size - 1)) {
2229 dm_block_t cache_size = ca->cache_sectors;
2231 cache->sectors_per_block_shift = -1;
2232 cache_size = block_div(cache_size, ca->block_size);
2233 cache->cache_size = to_cblock(cache_size);
2235 cache->sectors_per_block_shift = __ffs(ca->block_size);
2236 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
2239 r = create_cache_policy(cache, ca, error);
2243 cache->policy_nr_args = ca->policy_argc;
2244 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2246 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2248 *error = "Error setting cache policy's config values";
2252 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2253 ca->block_size, may_format,
2254 dm_cache_policy_get_hint_size(cache->policy));
2256 *error = "Error creating metadata object";
2262 if (passthrough_mode(&cache->features)) {
2265 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2267 *error = "dm_cache_metadata_all_clean() failed";
2272 *error = "Cannot enter passthrough mode unless all blocks are clean";
2278 spin_lock_init(&cache->lock);
2279 bio_list_init(&cache->deferred_bios);
2280 bio_list_init(&cache->deferred_flush_bios);
2281 bio_list_init(&cache->deferred_writethrough_bios);
2282 INIT_LIST_HEAD(&cache->quiesced_migrations);
2283 INIT_LIST_HEAD(&cache->completed_migrations);
2284 INIT_LIST_HEAD(&cache->need_commit_migrations);
2285 atomic_set(&cache->nr_migrations, 0);
2286 init_waitqueue_head(&cache->migration_wait);
2288 init_waitqueue_head(&cache->quiescing_wait);
2289 atomic_set(&cache->quiescing, 0);
2290 atomic_set(&cache->quiescing_ack, 0);
2293 atomic_set(&cache->nr_dirty, 0);
2294 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2295 if (!cache->dirty_bitset) {
2296 *error = "could not allocate dirty bitset";
2299 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2301 cache->discard_block_size = cache->sectors_per_block;
2302 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2303 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2304 if (!cache->discard_bitset) {
2305 *error = "could not allocate discard bitset";
2308 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2310 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2311 if (IS_ERR(cache->copier)) {
2312 *error = "could not create kcopyd client";
2313 r = PTR_ERR(cache->copier);
2317 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2319 *error = "could not create workqueue for metadata object";
2322 INIT_WORK(&cache->worker, do_worker);
2323 INIT_DELAYED_WORK(&cache->waker, do_waker);
2324 cache->last_commit_jiffies = jiffies;
2326 cache->prison = dm_bio_prison_create(PRISON_CELLS);
2327 if (!cache->prison) {
2328 *error = "could not create bio prison";
2332 cache->all_io_ds = dm_deferred_set_create();
2333 if (!cache->all_io_ds) {
2334 *error = "could not create all_io deferred set";
2338 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2340 if (!cache->migration_pool) {
2341 *error = "Error creating cache's migration mempool";
2345 cache->next_migration = NULL;
2347 cache->need_tick_bio = true;
2348 cache->sized = false;
2349 cache->invalidate = false;
2350 cache->commit_requested = false;
2351 cache->loaded_mappings = false;
2352 cache->loaded_discards = false;
2356 atomic_set(&cache->stats.demotion, 0);
2357 atomic_set(&cache->stats.promotion, 0);
2358 atomic_set(&cache->stats.copies_avoided, 0);
2359 atomic_set(&cache->stats.cache_cell_clash, 0);
2360 atomic_set(&cache->stats.commit_count, 0);
2361 atomic_set(&cache->stats.discard_count, 0);
2363 spin_lock_init(&cache->invalidation_lock);
2364 INIT_LIST_HEAD(&cache->invalidation_requests);
2374 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2379 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2382 for (i = 0; i < argc; i++) {
2383 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2392 cache->nr_ctr_args = argc;
2393 cache->ctr_args = copy;
2398 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2401 struct cache_args *ca;
2402 struct cache *cache = NULL;
2404 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2406 ti->error = "Error allocating memory for cache";
2411 r = parse_cache_args(ca, argc, argv, &ti->error);
2415 r = cache_create(ca, &cache);
2419 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2425 ti->private = cache;
2428 destroy_cache_args(ca);
2432 static int cache_map(struct dm_target *ti, struct bio *bio)
2434 struct cache *cache = ti->private;
2437 dm_oblock_t block = get_bio_block(cache, bio);
2438 size_t pb_data_size = get_per_bio_data_size(cache);
2439 bool can_migrate = false;
2440 bool discarded_block;
2441 struct dm_bio_prison_cell *cell;
2442 struct policy_result lookup_result;
2443 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
2445 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2447 * This can only occur if the io goes to a partial block at
2448 * the end of the origin device. We don't cache these.
2449 * Just remap to the origin and carry on.
2451 remap_to_origin(cache, bio);
2452 return DM_MAPIO_REMAPPED;
2455 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2456 defer_bio(cache, bio);
2457 return DM_MAPIO_SUBMITTED;
2461 * Check to see if that block is currently migrating.
2463 cell = alloc_prison_cell(cache);
2465 defer_bio(cache, bio);
2466 return DM_MAPIO_SUBMITTED;
2469 r = bio_detain(cache, block, bio, cell,
2470 (cell_free_fn) free_prison_cell,
2474 defer_bio(cache, bio);
2476 return DM_MAPIO_SUBMITTED;
2479 discarded_block = is_discarded_oblock(cache, block);
2481 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2482 bio, &lookup_result);
2483 if (r == -EWOULDBLOCK) {
2484 cell_defer(cache, cell, true);
2485 return DM_MAPIO_SUBMITTED;
2488 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2490 return DM_MAPIO_SUBMITTED;
2493 r = DM_MAPIO_REMAPPED;
2494 switch (lookup_result.op) {
2496 if (passthrough_mode(&cache->features)) {
2497 if (bio_data_dir(bio) == WRITE) {
2499 * We need to invalidate this block, so
2500 * defer for the worker thread.
2502 cell_defer(cache, cell, true);
2503 r = DM_MAPIO_SUBMITTED;
2506 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2507 inc_miss_counter(cache, bio);
2508 remap_to_origin_clear_discard(cache, bio, block);
2510 cell_defer(cache, cell, false);
2514 inc_hit_counter(cache, bio);
2515 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2517 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
2518 !is_dirty(cache, lookup_result.cblock))
2519 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2521 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2523 cell_defer(cache, cell, false);
2528 inc_miss_counter(cache, bio);
2529 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2531 if (pb->req_nr != 0) {
2533 * This is a duplicate writethrough io that is no
2534 * longer needed because the block has been demoted.
2537 cell_defer(cache, cell, false);
2538 return DM_MAPIO_SUBMITTED;
2540 remap_to_origin_clear_discard(cache, bio, block);
2541 cell_defer(cache, cell, false);
2546 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2547 (unsigned) lookup_result.op);
2549 r = DM_MAPIO_SUBMITTED;
2555 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2557 struct cache *cache = ti->private;
2558 unsigned long flags;
2559 size_t pb_data_size = get_per_bio_data_size(cache);
2560 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2563 policy_tick(cache->policy);
2565 spin_lock_irqsave(&cache->lock, flags);
2566 cache->need_tick_bio = true;
2567 spin_unlock_irqrestore(&cache->lock, flags);
2570 check_for_quiesced_migrations(cache, pb);
2575 static int write_dirty_bitset(struct cache *cache)
2579 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2580 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2581 is_dirty(cache, to_cblock(i)));
2589 static int write_discard_bitset(struct cache *cache)
2593 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2594 cache->discard_nr_blocks);
2596 DMERR("could not resize on-disk discard bitset");
2600 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2601 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2602 is_discarded(cache, to_dblock(i)));
2611 * returns true on success
2613 static bool sync_metadata(struct cache *cache)
2617 r1 = write_dirty_bitset(cache);
2619 DMERR("could not write dirty bitset");
2621 r2 = write_discard_bitset(cache);
2623 DMERR("could not write discard bitset");
2627 r3 = dm_cache_write_hints(cache->cmd, cache->policy);
2629 DMERR("could not write hints");
2632 * If writing the above metadata failed, we still commit, but don't
2633 * set the clean shutdown flag. This will effectively force every
2634 * dirty bit to be set on reload.
2636 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2638 DMERR("could not write cache metadata. Data loss may occur.");
2640 return !r1 && !r2 && !r3 && !r4;
2643 static void cache_postsuspend(struct dm_target *ti)
2645 struct cache *cache = ti->private;
2647 start_quiescing(cache);
2648 wait_for_migrations(cache);
2650 requeue_deferred_io(cache);
2651 stop_quiescing(cache);
2653 (void) sync_metadata(cache);
2656 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2657 bool dirty, uint32_t hint, bool hint_valid)
2660 struct cache *cache = context;
2662 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2667 set_dirty(cache, oblock, cblock);
2669 clear_dirty(cache, oblock, cblock);
2674 static int load_discard(void *context, sector_t discard_block_size,
2675 dm_dblock_t dblock, bool discard)
2677 struct cache *cache = context;
2679 /* FIXME: handle mis-matched block size */
2682 set_discard(cache, dblock);
2684 clear_discard(cache, dblock);
2689 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2691 sector_t size = get_dev_size(cache->cache_dev);
2692 (void) sector_div(size, cache->sectors_per_block);
2693 return to_cblock(size);
2696 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2698 if (from_cblock(new_size) > from_cblock(cache->cache_size))
2702 * We can't drop a dirty block when shrinking the cache.
2704 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2705 new_size = to_cblock(from_cblock(new_size) + 1);
2706 if (is_dirty(cache, new_size)) {
2707 DMERR("unable to shrink cache; cache block %llu is dirty",
2708 (unsigned long long) from_cblock(new_size));
2716 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2720 r = dm_cache_resize(cache->cmd, new_size);
2722 DMERR("could not resize cache metadata");
2726 cache->cache_size = new_size;
2731 static int cache_preresume(struct dm_target *ti)
2734 struct cache *cache = ti->private;
2735 dm_cblock_t csize = get_cache_dev_size(cache);
2738 * Check to see if the cache has resized.
2740 if (!cache->sized) {
2741 r = resize_cache_dev(cache, csize);
2745 cache->sized = true;
2747 } else if (csize != cache->cache_size) {
2748 if (!can_resize(cache, csize))
2751 r = resize_cache_dev(cache, csize);
2756 if (!cache->loaded_mappings) {
2757 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2758 load_mapping, cache);
2760 DMERR("could not load cache mappings");
2764 cache->loaded_mappings = true;
2767 if (!cache->loaded_discards) {
2768 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2770 DMERR("could not load origin discards");
2774 cache->loaded_discards = true;
2780 static void cache_resume(struct dm_target *ti)
2782 struct cache *cache = ti->private;
2784 cache->need_tick_bio = true;
2785 do_waker(&cache->waker.work);
2791 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
2792 * <cache block size> <#used cache blocks>/<#total cache blocks>
2793 * <#read hits> <#read misses> <#write hits> <#write misses>
2794 * <#demotions> <#promotions> <#dirty>
2795 * <#features> <features>*
2796 * <#core args> <core args>
2797 * <policy name> <#policy args> <policy args>*
2799 static void cache_status(struct dm_target *ti, status_type_t type,
2800 unsigned status_flags, char *result, unsigned maxlen)
2805 dm_block_t nr_free_blocks_metadata = 0;
2806 dm_block_t nr_blocks_metadata = 0;
2807 char buf[BDEVNAME_SIZE];
2808 struct cache *cache = ti->private;
2809 dm_cblock_t residency;
2812 case STATUSTYPE_INFO:
2813 /* Commit to ensure statistics aren't out-of-date */
2814 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2815 r = dm_cache_commit(cache->cmd, false);
2817 DMERR("could not commit metadata for accurate status");
2820 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2821 &nr_free_blocks_metadata);
2823 DMERR("could not get metadata free block count");
2827 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2829 DMERR("could not get metadata device size");
2833 residency = policy_residency(cache->policy);
2835 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
2836 (unsigned)(DM_CACHE_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
2837 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2838 (unsigned long long)nr_blocks_metadata,
2839 cache->sectors_per_block,
2840 (unsigned long long) from_cblock(residency),
2841 (unsigned long long) from_cblock(cache->cache_size),
2842 (unsigned) atomic_read(&cache->stats.read_hit),
2843 (unsigned) atomic_read(&cache->stats.read_miss),
2844 (unsigned) atomic_read(&cache->stats.write_hit),
2845 (unsigned) atomic_read(&cache->stats.write_miss),
2846 (unsigned) atomic_read(&cache->stats.demotion),
2847 (unsigned) atomic_read(&cache->stats.promotion),
2848 (unsigned long) atomic_read(&cache->nr_dirty));
2850 if (writethrough_mode(&cache->features))
2851 DMEMIT("1 writethrough ");
2853 else if (passthrough_mode(&cache->features))
2854 DMEMIT("1 passthrough ");
2856 else if (writeback_mode(&cache->features))
2857 DMEMIT("1 writeback ");
2860 DMERR("internal error: unknown io mode: %d", (int) cache->features.io_mode);
2864 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2866 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
2868 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2870 DMERR("policy_emit_config_values returned %d", r);
2875 case STATUSTYPE_TABLE:
2876 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2878 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2880 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2883 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2884 DMEMIT(" %s", cache->ctr_args[i]);
2885 if (cache->nr_ctr_args)
2886 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2896 * A cache block range can take two forms:
2898 * i) A single cblock, eg. '3456'
2899 * ii) A begin and end cblock with dots between, eg. 123-234
2901 static int parse_cblock_range(struct cache *cache, const char *str,
2902 struct cblock_range *result)
2909 * Try and parse form (ii) first.
2911 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
2916 result->begin = to_cblock(b);
2917 result->end = to_cblock(e);
2922 * That didn't work, try form (i).
2924 r = sscanf(str, "%llu%c", &b, &dummy);
2929 result->begin = to_cblock(b);
2930 result->end = to_cblock(from_cblock(result->begin) + 1u);
2934 DMERR("invalid cblock range '%s'", str);
2938 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
2940 uint64_t b = from_cblock(range->begin);
2941 uint64_t e = from_cblock(range->end);
2942 uint64_t n = from_cblock(cache->cache_size);
2945 DMERR("begin cblock out of range: %llu >= %llu", b, n);
2950 DMERR("end cblock out of range: %llu > %llu", e, n);
2955 DMERR("invalid cblock range: %llu >= %llu", b, e);
2962 static int request_invalidation(struct cache *cache, struct cblock_range *range)
2964 struct invalidation_request req;
2966 INIT_LIST_HEAD(&req.list);
2967 req.cblocks = range;
2968 atomic_set(&req.complete, 0);
2970 init_waitqueue_head(&req.result_wait);
2972 spin_lock(&cache->invalidation_lock);
2973 list_add(&req.list, &cache->invalidation_requests);
2974 spin_unlock(&cache->invalidation_lock);
2977 wait_event(req.result_wait, atomic_read(&req.complete));
2981 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
2982 const char **cblock_ranges)
2986 struct cblock_range range;
2988 if (!passthrough_mode(&cache->features)) {
2989 DMERR("cache has to be in passthrough mode for invalidation");
2993 for (i = 0; i < count; i++) {
2994 r = parse_cblock_range(cache, cblock_ranges[i], &range);
2998 r = validate_cblock_range(cache, &range);
3003 * Pass begin and end origin blocks to the worker and wake it.
3005 r = request_invalidation(cache, &range);
3017 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3019 * The key migration_threshold is supported by the cache target core.
3021 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3023 struct cache *cache = ti->private;
3028 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3029 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3034 return set_config_value(cache, argv[0], argv[1]);
3037 static int cache_iterate_devices(struct dm_target *ti,
3038 iterate_devices_callout_fn fn, void *data)
3041 struct cache *cache = ti->private;
3043 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3045 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3051 * We assume I/O is going to the origin (which is the volume
3052 * more likely to have restrictions e.g. by being striped).
3053 * (Looking up the exact location of the data would be expensive
3054 * and could always be out of date by the time the bio is submitted.)
3056 static int cache_bvec_merge(struct dm_target *ti,
3057 struct bvec_merge_data *bvm,
3058 struct bio_vec *biovec, int max_size)
3060 struct cache *cache = ti->private;
3061 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3063 if (!q->merge_bvec_fn)
3066 bvm->bi_bdev = cache->origin_dev->bdev;
3067 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3070 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3073 * FIXME: these limits may be incompatible with the cache device
3075 limits->max_discard_sectors = cache->discard_block_size;
3076 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3079 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3081 struct cache *cache = ti->private;
3082 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3085 * If the system-determined stacked limits are compatible with the
3086 * cache's blocksize (io_opt is a factor) do not override them.
3088 if (io_opt_sectors < cache->sectors_per_block ||
3089 do_div(io_opt_sectors, cache->sectors_per_block)) {
3090 blk_limits_io_min(limits, 0);
3091 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3093 set_discard_limits(cache, limits);
3096 /*----------------------------------------------------------------*/
3098 static struct target_type cache_target = {
3100 .version = {1, 4, 0},
3101 .module = THIS_MODULE,
3105 .end_io = cache_end_io,
3106 .postsuspend = cache_postsuspend,
3107 .preresume = cache_preresume,
3108 .resume = cache_resume,
3109 .status = cache_status,
3110 .message = cache_message,
3111 .iterate_devices = cache_iterate_devices,
3112 .merge = cache_bvec_merge,
3113 .io_hints = cache_io_hints,
3116 static int __init dm_cache_init(void)
3120 r = dm_register_target(&cache_target);
3122 DMERR("cache target registration failed: %d", r);
3126 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3127 if (!migration_cache) {
3128 dm_unregister_target(&cache_target);
3135 static void __exit dm_cache_exit(void)
3137 dm_unregister_target(&cache_target);
3138 kmem_cache_destroy(migration_cache);
3141 module_init(dm_cache_init);
3142 module_exit(dm_cache_exit);
3144 MODULE_DESCRIPTION(DM_NAME " cache target");
3145 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3146 MODULE_LICENSE("GPL");