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 /*----------------------------------------------------------------*/
92 #define PRISON_CELLS 1024
93 #define MIGRATION_POOL_SIZE 128
94 #define COMMIT_PERIOD HZ
95 #define MIGRATION_COUNT_WINDOW 10
98 * The block size of the device holding cache data must be
99 * between 32KB and 1GB.
101 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
102 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
105 * FIXME: the cache is read/write for the time being.
107 enum cache_metadata_mode {
108 CM_WRITE, /* metadata may be changed */
109 CM_READ_ONLY, /* metadata may not be changed */
114 * Data is written to cached blocks only. These blocks are marked
115 * dirty. If you lose the cache device you will lose data.
116 * Potential performance increase for both reads and writes.
121 * Data is written to both cache and origin. Blocks are never
122 * dirty. Potential performance benfit for reads only.
127 * A degraded mode useful for various cache coherency situations
128 * (eg, rolling back snapshots). Reads and writes always go to the
129 * origin. If a write goes to a cached oblock, then the cache
130 * block is invalidated.
135 struct cache_features {
136 enum cache_metadata_mode mode;
137 enum cache_io_mode io_mode;
147 atomic_t copies_avoided;
148 atomic_t cache_cell_clash;
149 atomic_t commit_count;
150 atomic_t discard_count;
154 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
155 * the one-past-the-end value.
157 struct cblock_range {
162 struct invalidation_request {
163 struct list_head list;
164 struct cblock_range *cblocks;
169 wait_queue_head_t result_wait;
173 struct dm_target *ti;
174 struct dm_target_callbacks callbacks;
176 struct dm_cache_metadata *cmd;
179 * Metadata is written to this device.
181 struct dm_dev *metadata_dev;
184 * The slower of the two data devices. Typically a spindle.
186 struct dm_dev *origin_dev;
189 * The faster of the two data devices. Typically an SSD.
191 struct dm_dev *cache_dev;
194 * Size of the origin device in _complete_ blocks and native sectors.
196 dm_oblock_t origin_blocks;
197 sector_t origin_sectors;
200 * Size of the cache device in blocks.
202 dm_cblock_t cache_size;
205 * Fields for converting from sectors to blocks.
207 uint32_t sectors_per_block;
208 int sectors_per_block_shift;
211 struct bio_list deferred_bios;
212 struct bio_list deferred_flush_bios;
213 struct bio_list deferred_writethrough_bios;
214 struct list_head quiesced_migrations;
215 struct list_head completed_migrations;
216 struct list_head need_commit_migrations;
217 sector_t migration_threshold;
218 wait_queue_head_t migration_wait;
219 atomic_t nr_migrations;
221 wait_queue_head_t quiescing_wait;
223 atomic_t quiescing_ack;
226 * cache_size entries, dirty if set
228 dm_cblock_t nr_dirty;
229 unsigned long *dirty_bitset;
232 * origin_blocks entries, discarded if set.
234 dm_dblock_t discard_nr_blocks;
235 unsigned long *discard_bitset;
236 uint32_t discard_block_size; /* a power of 2 times sectors per block */
239 * Rather than reconstructing the table line for the status we just
240 * save it and regurgitate.
242 unsigned nr_ctr_args;
243 const char **ctr_args;
245 struct dm_kcopyd_client *copier;
246 struct workqueue_struct *wq;
247 struct work_struct worker;
249 struct delayed_work waker;
250 unsigned long last_commit_jiffies;
252 struct dm_bio_prison *prison;
253 struct dm_deferred_set *all_io_ds;
255 mempool_t *migration_pool;
256 struct dm_cache_migration *next_migration;
258 struct dm_cache_policy *policy;
259 unsigned policy_nr_args;
261 bool need_tick_bio:1;
264 bool commit_requested:1;
265 bool loaded_mappings:1;
266 bool loaded_discards:1;
269 * Cache features such as write-through.
271 struct cache_features features;
273 struct cache_stats stats;
276 * Invalidation fields.
278 spinlock_t invalidation_lock;
279 struct list_head invalidation_requests;
282 struct per_bio_data {
285 struct dm_deferred_entry *all_io_entry;
288 * writethrough fields. These MUST remain at the end of this
289 * structure and the 'cache' member must be the first as it
290 * is used to determine the offset of the writethrough fields.
294 struct dm_hook_info hook_info;
295 struct dm_bio_details bio_details;
298 struct dm_cache_migration {
299 struct list_head list;
302 unsigned long start_jiffies;
303 dm_oblock_t old_oblock;
304 dm_oblock_t new_oblock;
311 bool requeue_holder:1;
314 struct dm_bio_prison_cell *old_ocell;
315 struct dm_bio_prison_cell *new_ocell;
319 * Processing a bio in the worker thread may require these memory
320 * allocations. We prealloc to avoid deadlocks (the same worker thread
321 * frees them back to the mempool).
324 struct dm_cache_migration *mg;
325 struct dm_bio_prison_cell *cell1;
326 struct dm_bio_prison_cell *cell2;
329 static void wake_worker(struct cache *cache)
331 queue_work(cache->wq, &cache->worker);
334 /*----------------------------------------------------------------*/
336 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
338 /* FIXME: change to use a local slab. */
339 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
342 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
344 dm_bio_prison_free_cell(cache->prison, cell);
347 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
350 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
356 p->cell1 = alloc_prison_cell(cache);
362 p->cell2 = alloc_prison_cell(cache);
370 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
373 free_prison_cell(cache, p->cell2);
376 free_prison_cell(cache, p->cell1);
379 mempool_free(p->mg, cache->migration_pool);
382 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
384 struct dm_cache_migration *mg = p->mg;
393 * You must have a cell within the prealloc struct to return. If not this
394 * function will BUG() rather than returning NULL.
396 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
398 struct dm_bio_prison_cell *r = NULL;
404 } else if (p->cell2) {
414 * You can't have more than two cells in a prealloc struct. BUG() will be
415 * called if you try and overfill.
417 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
429 /*----------------------------------------------------------------*/
431 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
435 key->block = from_oblock(oblock);
439 * The caller hands in a preallocated cell, and a free function for it.
440 * The cell will be freed if there's an error, or if it wasn't used because
441 * a cell with that key already exists.
443 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
445 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
446 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
447 cell_free_fn free_fn, void *free_context,
448 struct dm_bio_prison_cell **cell_result)
451 struct dm_cell_key key;
453 build_key(oblock, &key);
454 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
456 free_fn(free_context, cell_prealloc);
461 static int get_cell(struct cache *cache,
463 struct prealloc *structs,
464 struct dm_bio_prison_cell **cell_result)
467 struct dm_cell_key key;
468 struct dm_bio_prison_cell *cell_prealloc;
470 cell_prealloc = prealloc_get_cell(structs);
472 build_key(oblock, &key);
473 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
475 prealloc_put_cell(structs, cell_prealloc);
480 /*----------------------------------------------------------------*/
482 static bool is_dirty(struct cache *cache, dm_cblock_t b)
484 return test_bit(from_cblock(b), cache->dirty_bitset);
487 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
489 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
490 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
491 policy_set_dirty(cache->policy, oblock);
495 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
497 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
498 policy_clear_dirty(cache->policy, oblock);
499 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
500 if (!from_cblock(cache->nr_dirty))
501 dm_table_event(cache->ti->table);
505 /*----------------------------------------------------------------*/
507 static bool block_size_is_power_of_two(struct cache *cache)
509 return cache->sectors_per_block_shift >= 0;
512 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
513 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
516 static dm_block_t block_div(dm_block_t b, uint32_t n)
523 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
525 uint32_t discard_blocks = cache->discard_block_size;
526 dm_block_t b = from_oblock(oblock);
528 if (!block_size_is_power_of_two(cache))
529 discard_blocks = discard_blocks / cache->sectors_per_block;
531 discard_blocks >>= cache->sectors_per_block_shift;
533 b = block_div(b, discard_blocks);
538 static void set_discard(struct cache *cache, dm_dblock_t b)
542 atomic_inc(&cache->stats.discard_count);
544 spin_lock_irqsave(&cache->lock, flags);
545 set_bit(from_dblock(b), cache->discard_bitset);
546 spin_unlock_irqrestore(&cache->lock, flags);
549 static void clear_discard(struct cache *cache, dm_dblock_t b)
553 spin_lock_irqsave(&cache->lock, flags);
554 clear_bit(from_dblock(b), cache->discard_bitset);
555 spin_unlock_irqrestore(&cache->lock, flags);
558 static bool is_discarded(struct cache *cache, dm_dblock_t b)
563 spin_lock_irqsave(&cache->lock, flags);
564 r = test_bit(from_dblock(b), cache->discard_bitset);
565 spin_unlock_irqrestore(&cache->lock, flags);
570 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
575 spin_lock_irqsave(&cache->lock, flags);
576 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
577 cache->discard_bitset);
578 spin_unlock_irqrestore(&cache->lock, flags);
583 /*----------------------------------------------------------------*/
585 static void load_stats(struct cache *cache)
587 struct dm_cache_statistics stats;
589 dm_cache_metadata_get_stats(cache->cmd, &stats);
590 atomic_set(&cache->stats.read_hit, stats.read_hits);
591 atomic_set(&cache->stats.read_miss, stats.read_misses);
592 atomic_set(&cache->stats.write_hit, stats.write_hits);
593 atomic_set(&cache->stats.write_miss, stats.write_misses);
596 static void save_stats(struct cache *cache)
598 struct dm_cache_statistics stats;
600 stats.read_hits = atomic_read(&cache->stats.read_hit);
601 stats.read_misses = atomic_read(&cache->stats.read_miss);
602 stats.write_hits = atomic_read(&cache->stats.write_hit);
603 stats.write_misses = atomic_read(&cache->stats.write_miss);
605 dm_cache_metadata_set_stats(cache->cmd, &stats);
608 /*----------------------------------------------------------------
610 *--------------------------------------------------------------*/
613 * If using writeback, leave out struct per_bio_data's writethrough fields.
615 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
616 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
618 static bool writethrough_mode(struct cache_features *f)
620 return f->io_mode == CM_IO_WRITETHROUGH;
623 static bool writeback_mode(struct cache_features *f)
625 return f->io_mode == CM_IO_WRITEBACK;
628 static bool passthrough_mode(struct cache_features *f)
630 return f->io_mode == CM_IO_PASSTHROUGH;
633 static size_t get_per_bio_data_size(struct cache *cache)
635 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
638 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
640 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
645 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
647 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
650 pb->req_nr = dm_bio_get_target_bio_nr(bio);
651 pb->all_io_entry = NULL;
656 /*----------------------------------------------------------------
658 *--------------------------------------------------------------*/
659 static void remap_to_origin(struct cache *cache, struct bio *bio)
661 bio->bi_bdev = cache->origin_dev->bdev;
664 static void remap_to_cache(struct cache *cache, struct bio *bio,
667 sector_t bi_sector = bio->bi_iter.bi_sector;
669 bio->bi_bdev = cache->cache_dev->bdev;
670 if (!block_size_is_power_of_two(cache))
671 bio->bi_iter.bi_sector =
672 (from_cblock(cblock) * cache->sectors_per_block) +
673 sector_div(bi_sector, cache->sectors_per_block);
675 bio->bi_iter.bi_sector =
676 (from_cblock(cblock) << cache->sectors_per_block_shift) |
677 (bi_sector & (cache->sectors_per_block - 1));
680 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
683 size_t pb_data_size = get_per_bio_data_size(cache);
684 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
686 spin_lock_irqsave(&cache->lock, flags);
687 if (cache->need_tick_bio &&
688 !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
690 cache->need_tick_bio = false;
692 spin_unlock_irqrestore(&cache->lock, flags);
695 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
698 check_if_tick_bio_needed(cache, bio);
699 remap_to_origin(cache, bio);
700 if (bio_data_dir(bio) == WRITE)
701 clear_discard(cache, oblock_to_dblock(cache, oblock));
704 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
705 dm_oblock_t oblock, dm_cblock_t cblock)
707 check_if_tick_bio_needed(cache, bio);
708 remap_to_cache(cache, bio, cblock);
709 if (bio_data_dir(bio) == WRITE) {
710 set_dirty(cache, oblock, cblock);
711 clear_discard(cache, oblock_to_dblock(cache, oblock));
715 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
717 sector_t block_nr = bio->bi_iter.bi_sector;
719 if (!block_size_is_power_of_two(cache))
720 (void) sector_div(block_nr, cache->sectors_per_block);
722 block_nr >>= cache->sectors_per_block_shift;
724 return to_oblock(block_nr);
727 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
729 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
732 static void issue(struct cache *cache, struct bio *bio)
736 if (!bio_triggers_commit(cache, bio)) {
737 generic_make_request(bio);
742 * Batch together any bios that trigger commits and then issue a
743 * single commit for them in do_worker().
745 spin_lock_irqsave(&cache->lock, flags);
746 cache->commit_requested = true;
747 bio_list_add(&cache->deferred_flush_bios, bio);
748 spin_unlock_irqrestore(&cache->lock, flags);
751 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
755 spin_lock_irqsave(&cache->lock, flags);
756 bio_list_add(&cache->deferred_writethrough_bios, bio);
757 spin_unlock_irqrestore(&cache->lock, flags);
762 static void writethrough_endio(struct bio *bio, int err)
764 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
766 dm_unhook_bio(&pb->hook_info, bio);
769 * Must bump bi_remaining to allow bio to complete with
770 * restored bi_end_io.
772 atomic_inc(&bio->bi_remaining);
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 cell_defer(cache, mg->old_ocell, false);
894 clear_dirty(cache, mg->old_oblock, mg->cblock);
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 cell_defer(cache, mg->new_ocell, true);
952 bio_endio(mg->new_ocell->holder, 0);
953 cell_defer(cache, mg->new_ocell, false);
955 clear_dirty(cache, mg->new_oblock, mg->cblock);
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;
982 o_region.bdev = cache->origin_dev->bdev;
983 o_region.count = cache->sectors_per_block;
985 c_region.bdev = cache->cache_dev->bdev;
986 c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
987 c_region.count = cache->sectors_per_block;
989 if (mg->writeback || mg->demote) {
991 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
992 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
995 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
996 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1000 DMERR_LIMIT("issuing migration failed");
1001 migration_failure(mg);
1005 static void overwrite_endio(struct bio *bio, int err)
1007 struct dm_cache_migration *mg = bio->bi_private;
1008 struct cache *cache = mg->cache;
1009 size_t pb_data_size = get_per_bio_data_size(cache);
1010 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1011 unsigned long flags;
1016 spin_lock_irqsave(&cache->lock, flags);
1017 list_add_tail(&mg->list, &cache->completed_migrations);
1018 dm_unhook_bio(&pb->hook_info, bio);
1019 mg->requeue_holder = false;
1020 spin_unlock_irqrestore(&cache->lock, flags);
1025 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1027 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1028 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1030 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1031 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1032 generic_make_request(bio);
1035 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1037 return (bio_data_dir(bio) == WRITE) &&
1038 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1041 static void avoid_copy(struct dm_cache_migration *mg)
1043 atomic_inc(&mg->cache->stats.copies_avoided);
1044 migration_success_pre_commit(mg);
1047 static void issue_copy(struct dm_cache_migration *mg)
1050 struct cache *cache = mg->cache;
1052 if (mg->writeback || mg->demote)
1053 avoid = !is_dirty(cache, mg->cblock) ||
1054 is_discarded_oblock(cache, mg->old_oblock);
1056 struct bio *bio = mg->new_ocell->holder;
1058 avoid = is_discarded_oblock(cache, mg->new_oblock);
1060 if (!avoid && bio_writes_complete_block(cache, bio)) {
1061 issue_overwrite(mg, bio);
1066 avoid ? avoid_copy(mg) : issue_copy_real(mg);
1069 static void complete_migration(struct dm_cache_migration *mg)
1072 migration_failure(mg);
1074 migration_success_pre_commit(mg);
1077 static void process_migrations(struct cache *cache, struct list_head *head,
1078 void (*fn)(struct dm_cache_migration *))
1080 unsigned long flags;
1081 struct list_head list;
1082 struct dm_cache_migration *mg, *tmp;
1084 INIT_LIST_HEAD(&list);
1085 spin_lock_irqsave(&cache->lock, flags);
1086 list_splice_init(head, &list);
1087 spin_unlock_irqrestore(&cache->lock, flags);
1089 list_for_each_entry_safe(mg, tmp, &list, list)
1093 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1095 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1098 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1100 unsigned long flags;
1101 struct cache *cache = mg->cache;
1103 spin_lock_irqsave(&cache->lock, flags);
1104 __queue_quiesced_migration(mg);
1105 spin_unlock_irqrestore(&cache->lock, flags);
1110 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1112 unsigned long flags;
1113 struct dm_cache_migration *mg, *tmp;
1115 spin_lock_irqsave(&cache->lock, flags);
1116 list_for_each_entry_safe(mg, tmp, work, list)
1117 __queue_quiesced_migration(mg);
1118 spin_unlock_irqrestore(&cache->lock, flags);
1123 static void check_for_quiesced_migrations(struct cache *cache,
1124 struct per_bio_data *pb)
1126 struct list_head work;
1128 if (!pb->all_io_entry)
1131 INIT_LIST_HEAD(&work);
1132 if (pb->all_io_entry)
1133 dm_deferred_entry_dec(pb->all_io_entry, &work);
1135 if (!list_empty(&work))
1136 queue_quiesced_migrations(cache, &work);
1139 static void quiesce_migration(struct dm_cache_migration *mg)
1141 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1142 queue_quiesced_migration(mg);
1145 static void promote(struct cache *cache, struct prealloc *structs,
1146 dm_oblock_t oblock, dm_cblock_t cblock,
1147 struct dm_bio_prison_cell *cell)
1149 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1152 mg->writeback = false;
1155 mg->requeue_holder = true;
1156 mg->invalidate = false;
1158 mg->new_oblock = oblock;
1159 mg->cblock = cblock;
1160 mg->old_ocell = NULL;
1161 mg->new_ocell = cell;
1162 mg->start_jiffies = jiffies;
1164 inc_nr_migrations(cache);
1165 quiesce_migration(mg);
1168 static void writeback(struct cache *cache, struct prealloc *structs,
1169 dm_oblock_t oblock, dm_cblock_t cblock,
1170 struct dm_bio_prison_cell *cell)
1172 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1175 mg->writeback = true;
1177 mg->promote = false;
1178 mg->requeue_holder = true;
1179 mg->invalidate = false;
1181 mg->old_oblock = oblock;
1182 mg->cblock = cblock;
1183 mg->old_ocell = cell;
1184 mg->new_ocell = NULL;
1185 mg->start_jiffies = jiffies;
1187 inc_nr_migrations(cache);
1188 quiesce_migration(mg);
1191 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1192 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1194 struct dm_bio_prison_cell *old_ocell,
1195 struct dm_bio_prison_cell *new_ocell)
1197 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1200 mg->writeback = false;
1203 mg->requeue_holder = true;
1204 mg->invalidate = false;
1206 mg->old_oblock = old_oblock;
1207 mg->new_oblock = new_oblock;
1208 mg->cblock = cblock;
1209 mg->old_ocell = old_ocell;
1210 mg->new_ocell = new_ocell;
1211 mg->start_jiffies = jiffies;
1213 inc_nr_migrations(cache);
1214 quiesce_migration(mg);
1218 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1219 * block are thrown away.
1221 static void invalidate(struct cache *cache, struct prealloc *structs,
1222 dm_oblock_t oblock, dm_cblock_t cblock,
1223 struct dm_bio_prison_cell *cell)
1225 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1228 mg->writeback = false;
1230 mg->promote = false;
1231 mg->requeue_holder = true;
1232 mg->invalidate = true;
1234 mg->old_oblock = oblock;
1235 mg->cblock = cblock;
1236 mg->old_ocell = cell;
1237 mg->new_ocell = NULL;
1238 mg->start_jiffies = jiffies;
1240 inc_nr_migrations(cache);
1241 quiesce_migration(mg);
1244 /*----------------------------------------------------------------
1246 *--------------------------------------------------------------*/
1247 static void defer_bio(struct cache *cache, struct bio *bio)
1249 unsigned long flags;
1251 spin_lock_irqsave(&cache->lock, flags);
1252 bio_list_add(&cache->deferred_bios, bio);
1253 spin_unlock_irqrestore(&cache->lock, flags);
1258 static void process_flush_bio(struct cache *cache, struct bio *bio)
1260 size_t pb_data_size = get_per_bio_data_size(cache);
1261 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1263 BUG_ON(bio->bi_iter.bi_size);
1265 remap_to_origin(cache, bio);
1267 remap_to_cache(cache, bio, 0);
1273 * People generally discard large parts of a device, eg, the whole device
1274 * when formatting. Splitting these large discards up into cache block
1275 * sized ios and then quiescing (always neccessary for discard) takes too
1278 * We keep it simple, and allow any size of discard to come in, and just
1279 * mark off blocks on the discard bitset. No passdown occurs!
1281 * To implement passdown we need to change the bio_prison such that a cell
1282 * can have a key that spans many blocks.
1284 static void process_discard_bio(struct cache *cache, struct bio *bio)
1286 dm_block_t start_block = dm_sector_div_up(bio->bi_iter.bi_sector,
1287 cache->discard_block_size);
1288 dm_block_t end_block = bio_end_sector(bio);
1291 end_block = block_div(end_block, cache->discard_block_size);
1293 for (b = start_block; b < end_block; b++)
1294 set_discard(cache, to_dblock(b));
1299 static bool spare_migration_bandwidth(struct cache *cache)
1301 sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1302 cache->sectors_per_block;
1303 return current_volume < cache->migration_threshold;
1306 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1308 atomic_inc(bio_data_dir(bio) == READ ?
1309 &cache->stats.read_hit : &cache->stats.write_hit);
1312 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1314 atomic_inc(bio_data_dir(bio) == READ ?
1315 &cache->stats.read_miss : &cache->stats.write_miss);
1318 static void issue_cache_bio(struct cache *cache, struct bio *bio,
1319 struct per_bio_data *pb,
1320 dm_oblock_t oblock, dm_cblock_t cblock)
1322 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1323 remap_to_cache_dirty(cache, bio, oblock, cblock);
1327 static void process_bio(struct cache *cache, struct prealloc *structs,
1331 bool release_cell = true;
1332 dm_oblock_t block = get_bio_block(cache, bio);
1333 struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1334 struct policy_result lookup_result;
1335 size_t pb_data_size = get_per_bio_data_size(cache);
1336 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1337 bool discarded_block = is_discarded_oblock(cache, block);
1338 bool passthrough = passthrough_mode(&cache->features);
1339 bool can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
1342 * Check to see if that block is currently migrating.
1344 cell_prealloc = prealloc_get_cell(structs);
1345 r = bio_detain(cache, block, bio, cell_prealloc,
1346 (cell_free_fn) prealloc_put_cell,
1347 structs, &new_ocell);
1351 r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1352 bio, &lookup_result);
1354 if (r == -EWOULDBLOCK)
1355 /* migration has been denied */
1356 lookup_result.op = POLICY_MISS;
1358 switch (lookup_result.op) {
1361 inc_miss_counter(cache, bio);
1364 * Passthrough always maps to the origin,
1365 * invalidating any cache blocks that are written
1369 if (bio_data_dir(bio) == WRITE) {
1370 atomic_inc(&cache->stats.demotion);
1371 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1372 release_cell = false;
1375 /* FIXME: factor out issue_origin() */
1376 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1377 remap_to_origin_clear_discard(cache, bio, block);
1381 inc_hit_counter(cache, bio);
1383 if (bio_data_dir(bio) == WRITE &&
1384 writethrough_mode(&cache->features) &&
1385 !is_dirty(cache, lookup_result.cblock)) {
1386 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1387 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1390 issue_cache_bio(cache, bio, pb, block, lookup_result.cblock);
1396 inc_miss_counter(cache, bio);
1397 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1398 remap_to_origin_clear_discard(cache, bio, block);
1403 atomic_inc(&cache->stats.promotion);
1404 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1405 release_cell = false;
1408 case POLICY_REPLACE:
1409 cell_prealloc = prealloc_get_cell(structs);
1410 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1411 (cell_free_fn) prealloc_put_cell,
1412 structs, &old_ocell);
1415 * We have to be careful to avoid lock inversion of
1416 * the cells. So we back off, and wait for the
1417 * old_ocell to become free.
1419 policy_force_mapping(cache->policy, block,
1420 lookup_result.old_oblock);
1421 atomic_inc(&cache->stats.cache_cell_clash);
1424 atomic_inc(&cache->stats.demotion);
1425 atomic_inc(&cache->stats.promotion);
1427 demote_then_promote(cache, structs, lookup_result.old_oblock,
1428 block, lookup_result.cblock,
1429 old_ocell, new_ocell);
1430 release_cell = false;
1434 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1435 (unsigned) lookup_result.op);
1440 cell_defer(cache, new_ocell, false);
1443 static int need_commit_due_to_time(struct cache *cache)
1445 return jiffies < cache->last_commit_jiffies ||
1446 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1449 static int commit_if_needed(struct cache *cache)
1453 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1454 dm_cache_changed_this_transaction(cache->cmd)) {
1455 atomic_inc(&cache->stats.commit_count);
1456 cache->commit_requested = false;
1457 r = dm_cache_commit(cache->cmd, false);
1458 cache->last_commit_jiffies = jiffies;
1464 static void process_deferred_bios(struct cache *cache)
1466 unsigned long flags;
1467 struct bio_list bios;
1469 struct prealloc structs;
1471 memset(&structs, 0, sizeof(structs));
1472 bio_list_init(&bios);
1474 spin_lock_irqsave(&cache->lock, flags);
1475 bio_list_merge(&bios, &cache->deferred_bios);
1476 bio_list_init(&cache->deferred_bios);
1477 spin_unlock_irqrestore(&cache->lock, flags);
1479 while (!bio_list_empty(&bios)) {
1481 * If we've got no free migration structs, and processing
1482 * this bio might require one, we pause until there are some
1483 * prepared mappings to process.
1485 if (prealloc_data_structs(cache, &structs)) {
1486 spin_lock_irqsave(&cache->lock, flags);
1487 bio_list_merge(&cache->deferred_bios, &bios);
1488 spin_unlock_irqrestore(&cache->lock, flags);
1492 bio = bio_list_pop(&bios);
1494 if (bio->bi_rw & REQ_FLUSH)
1495 process_flush_bio(cache, bio);
1496 else if (bio->bi_rw & REQ_DISCARD)
1497 process_discard_bio(cache, bio);
1499 process_bio(cache, &structs, bio);
1502 prealloc_free_structs(cache, &structs);
1505 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1507 unsigned long flags;
1508 struct bio_list bios;
1511 bio_list_init(&bios);
1513 spin_lock_irqsave(&cache->lock, flags);
1514 bio_list_merge(&bios, &cache->deferred_flush_bios);
1515 bio_list_init(&cache->deferred_flush_bios);
1516 spin_unlock_irqrestore(&cache->lock, flags);
1518 while ((bio = bio_list_pop(&bios)))
1519 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1522 static void process_deferred_writethrough_bios(struct cache *cache)
1524 unsigned long flags;
1525 struct bio_list bios;
1528 bio_list_init(&bios);
1530 spin_lock_irqsave(&cache->lock, flags);
1531 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1532 bio_list_init(&cache->deferred_writethrough_bios);
1533 spin_unlock_irqrestore(&cache->lock, flags);
1535 while ((bio = bio_list_pop(&bios)))
1536 generic_make_request(bio);
1539 static void writeback_some_dirty_blocks(struct cache *cache)
1544 struct prealloc structs;
1545 struct dm_bio_prison_cell *old_ocell;
1547 memset(&structs, 0, sizeof(structs));
1549 while (spare_migration_bandwidth(cache)) {
1550 if (prealloc_data_structs(cache, &structs))
1553 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1557 r = get_cell(cache, oblock, &structs, &old_ocell);
1559 policy_set_dirty(cache->policy, oblock);
1563 writeback(cache, &structs, oblock, cblock, old_ocell);
1566 prealloc_free_structs(cache, &structs);
1569 /*----------------------------------------------------------------
1571 * Dropping something from the cache *without* writing back.
1572 *--------------------------------------------------------------*/
1574 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
1577 uint64_t begin = from_cblock(req->cblocks->begin);
1578 uint64_t end = from_cblock(req->cblocks->end);
1580 while (begin != end) {
1581 r = policy_remove_cblock(cache->policy, to_cblock(begin));
1583 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
1587 } else if (r == -ENODATA) {
1588 /* harmless, already unmapped */
1592 DMERR("policy_remove_cblock failed");
1599 cache->commit_requested = true;
1602 atomic_set(&req->complete, 1);
1604 wake_up(&req->result_wait);
1607 static void process_invalidation_requests(struct cache *cache)
1609 struct list_head list;
1610 struct invalidation_request *req, *tmp;
1612 INIT_LIST_HEAD(&list);
1613 spin_lock(&cache->invalidation_lock);
1614 list_splice_init(&cache->invalidation_requests, &list);
1615 spin_unlock(&cache->invalidation_lock);
1617 list_for_each_entry_safe (req, tmp, &list, list)
1618 process_invalidation_request(cache, req);
1621 /*----------------------------------------------------------------
1623 *--------------------------------------------------------------*/
1624 static bool is_quiescing(struct cache *cache)
1626 return atomic_read(&cache->quiescing);
1629 static void ack_quiescing(struct cache *cache)
1631 if (is_quiescing(cache)) {
1632 atomic_inc(&cache->quiescing_ack);
1633 wake_up(&cache->quiescing_wait);
1637 static void wait_for_quiescing_ack(struct cache *cache)
1639 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
1642 static void start_quiescing(struct cache *cache)
1644 atomic_inc(&cache->quiescing);
1645 wait_for_quiescing_ack(cache);
1648 static void stop_quiescing(struct cache *cache)
1650 atomic_set(&cache->quiescing, 0);
1651 atomic_set(&cache->quiescing_ack, 0);
1654 static void wait_for_migrations(struct cache *cache)
1656 wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1659 static void stop_worker(struct cache *cache)
1661 cancel_delayed_work(&cache->waker);
1662 flush_workqueue(cache->wq);
1665 static void requeue_deferred_io(struct cache *cache)
1668 struct bio_list bios;
1670 bio_list_init(&bios);
1671 bio_list_merge(&bios, &cache->deferred_bios);
1672 bio_list_init(&cache->deferred_bios);
1674 while ((bio = bio_list_pop(&bios)))
1675 bio_endio(bio, DM_ENDIO_REQUEUE);
1678 static int more_work(struct cache *cache)
1680 if (is_quiescing(cache))
1681 return !list_empty(&cache->quiesced_migrations) ||
1682 !list_empty(&cache->completed_migrations) ||
1683 !list_empty(&cache->need_commit_migrations);
1685 return !bio_list_empty(&cache->deferred_bios) ||
1686 !bio_list_empty(&cache->deferred_flush_bios) ||
1687 !bio_list_empty(&cache->deferred_writethrough_bios) ||
1688 !list_empty(&cache->quiesced_migrations) ||
1689 !list_empty(&cache->completed_migrations) ||
1690 !list_empty(&cache->need_commit_migrations) ||
1694 static void do_worker(struct work_struct *ws)
1696 struct cache *cache = container_of(ws, struct cache, worker);
1699 if (!is_quiescing(cache)) {
1700 writeback_some_dirty_blocks(cache);
1701 process_deferred_writethrough_bios(cache);
1702 process_deferred_bios(cache);
1703 process_invalidation_requests(cache);
1706 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1707 process_migrations(cache, &cache->completed_migrations, complete_migration);
1709 if (commit_if_needed(cache)) {
1710 process_deferred_flush_bios(cache, false);
1713 * FIXME: rollback metadata or just go into a
1714 * failure mode and error everything
1717 process_deferred_flush_bios(cache, true);
1718 process_migrations(cache, &cache->need_commit_migrations,
1719 migration_success_post_commit);
1722 ack_quiescing(cache);
1724 } while (more_work(cache));
1728 * We want to commit periodically so that not too much
1729 * unwritten metadata builds up.
1731 static void do_waker(struct work_struct *ws)
1733 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1734 policy_tick(cache->policy);
1736 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1739 /*----------------------------------------------------------------*/
1741 static int is_congested(struct dm_dev *dev, int bdi_bits)
1743 struct request_queue *q = bdev_get_queue(dev->bdev);
1744 return bdi_congested(&q->backing_dev_info, bdi_bits);
1747 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1749 struct cache *cache = container_of(cb, struct cache, callbacks);
1751 return is_congested(cache->origin_dev, bdi_bits) ||
1752 is_congested(cache->cache_dev, bdi_bits);
1755 /*----------------------------------------------------------------
1757 *--------------------------------------------------------------*/
1760 * This function gets called on the error paths of the constructor, so we
1761 * have to cope with a partially initialised struct.
1763 static void destroy(struct cache *cache)
1767 if (cache->next_migration)
1768 mempool_free(cache->next_migration, cache->migration_pool);
1770 if (cache->migration_pool)
1771 mempool_destroy(cache->migration_pool);
1773 if (cache->all_io_ds)
1774 dm_deferred_set_destroy(cache->all_io_ds);
1777 dm_bio_prison_destroy(cache->prison);
1780 destroy_workqueue(cache->wq);
1782 if (cache->dirty_bitset)
1783 free_bitset(cache->dirty_bitset);
1785 if (cache->discard_bitset)
1786 free_bitset(cache->discard_bitset);
1789 dm_kcopyd_client_destroy(cache->copier);
1792 dm_cache_metadata_close(cache->cmd);
1794 if (cache->metadata_dev)
1795 dm_put_device(cache->ti, cache->metadata_dev);
1797 if (cache->origin_dev)
1798 dm_put_device(cache->ti, cache->origin_dev);
1800 if (cache->cache_dev)
1801 dm_put_device(cache->ti, cache->cache_dev);
1804 dm_cache_policy_destroy(cache->policy);
1806 for (i = 0; i < cache->nr_ctr_args ; i++)
1807 kfree(cache->ctr_args[i]);
1808 kfree(cache->ctr_args);
1813 static void cache_dtr(struct dm_target *ti)
1815 struct cache *cache = ti->private;
1820 static sector_t get_dev_size(struct dm_dev *dev)
1822 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1825 /*----------------------------------------------------------------*/
1828 * Construct a cache device mapping.
1830 * cache <metadata dev> <cache dev> <origin dev> <block size>
1831 * <#feature args> [<feature arg>]*
1832 * <policy> <#policy args> [<policy arg>]*
1834 * metadata dev : fast device holding the persistent metadata
1835 * cache dev : fast device holding cached data blocks
1836 * origin dev : slow device holding original data blocks
1837 * block size : cache unit size in sectors
1839 * #feature args : number of feature arguments passed
1840 * feature args : writethrough. (The default is writeback.)
1842 * policy : the replacement policy to use
1843 * #policy args : an even number of policy arguments corresponding
1844 * to key/value pairs passed to the policy
1845 * policy args : key/value pairs passed to the policy
1846 * E.g. 'sequential_threshold 1024'
1847 * See cache-policies.txt for details.
1849 * Optional feature arguments are:
1850 * writethrough : write through caching that prohibits cache block
1851 * content from being different from origin block content.
1852 * Without this argument, the default behaviour is to write
1853 * back cache block contents later for performance reasons,
1854 * so they may differ from the corresponding origin blocks.
1857 struct dm_target *ti;
1859 struct dm_dev *metadata_dev;
1861 struct dm_dev *cache_dev;
1862 sector_t cache_sectors;
1864 struct dm_dev *origin_dev;
1865 sector_t origin_sectors;
1867 uint32_t block_size;
1869 const char *policy_name;
1871 const char **policy_argv;
1873 struct cache_features features;
1876 static void destroy_cache_args(struct cache_args *ca)
1878 if (ca->metadata_dev)
1879 dm_put_device(ca->ti, ca->metadata_dev);
1882 dm_put_device(ca->ti, ca->cache_dev);
1885 dm_put_device(ca->ti, ca->origin_dev);
1890 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1893 *error = "Insufficient args";
1900 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1904 sector_t metadata_dev_size;
1905 char b[BDEVNAME_SIZE];
1907 if (!at_least_one_arg(as, error))
1910 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1913 *error = "Error opening metadata device";
1917 metadata_dev_size = get_dev_size(ca->metadata_dev);
1918 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1919 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1920 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1925 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1930 if (!at_least_one_arg(as, error))
1933 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1936 *error = "Error opening cache device";
1939 ca->cache_sectors = get_dev_size(ca->cache_dev);
1944 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1949 if (!at_least_one_arg(as, error))
1952 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1955 *error = "Error opening origin device";
1959 ca->origin_sectors = get_dev_size(ca->origin_dev);
1960 if (ca->ti->len > ca->origin_sectors) {
1961 *error = "Device size larger than cached device";
1968 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1971 unsigned long block_size;
1973 if (!at_least_one_arg(as, error))
1976 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
1977 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1978 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1979 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1980 *error = "Invalid data block size";
1984 if (block_size > ca->cache_sectors) {
1985 *error = "Data block size is larger than the cache device";
1989 ca->block_size = block_size;
1994 static void init_features(struct cache_features *cf)
1996 cf->mode = CM_WRITE;
1997 cf->io_mode = CM_IO_WRITEBACK;
2000 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2003 static struct dm_arg _args[] = {
2004 {0, 1, "Invalid number of cache feature arguments"},
2010 struct cache_features *cf = &ca->features;
2014 r = dm_read_arg_group(_args, as, &argc, error);
2019 arg = dm_shift_arg(as);
2021 if (!strcasecmp(arg, "writeback"))
2022 cf->io_mode = CM_IO_WRITEBACK;
2024 else if (!strcasecmp(arg, "writethrough"))
2025 cf->io_mode = CM_IO_WRITETHROUGH;
2027 else if (!strcasecmp(arg, "passthrough"))
2028 cf->io_mode = CM_IO_PASSTHROUGH;
2031 *error = "Unrecognised cache feature requested";
2039 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2042 static struct dm_arg _args[] = {
2043 {0, 1024, "Invalid number of policy arguments"},
2048 if (!at_least_one_arg(as, error))
2051 ca->policy_name = dm_shift_arg(as);
2053 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2057 ca->policy_argv = (const char **)as->argv;
2058 dm_consume_args(as, ca->policy_argc);
2063 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2067 struct dm_arg_set as;
2072 r = parse_metadata_dev(ca, &as, error);
2076 r = parse_cache_dev(ca, &as, error);
2080 r = parse_origin_dev(ca, &as, error);
2084 r = parse_block_size(ca, &as, error);
2088 r = parse_features(ca, &as, error);
2092 r = parse_policy(ca, &as, error);
2099 /*----------------------------------------------------------------*/
2101 static struct kmem_cache *migration_cache;
2103 #define NOT_CORE_OPTION 1
2105 static int process_config_option(struct cache *cache, const char *key, const char *value)
2109 if (!strcasecmp(key, "migration_threshold")) {
2110 if (kstrtoul(value, 10, &tmp))
2113 cache->migration_threshold = tmp;
2117 return NOT_CORE_OPTION;
2120 static int set_config_value(struct cache *cache, const char *key, const char *value)
2122 int r = process_config_option(cache, key, value);
2124 if (r == NOT_CORE_OPTION)
2125 r = policy_set_config_value(cache->policy, key, value);
2128 DMWARN("bad config value for %s: %s", key, value);
2133 static int set_config_values(struct cache *cache, int argc, const char **argv)
2138 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2143 r = set_config_value(cache, argv[0], argv[1]);
2154 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2157 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2159 cache->origin_sectors,
2160 cache->sectors_per_block);
2162 *error = "Error creating cache's policy";
2171 * We want the discard block size to be a power of two, at least the size
2172 * of the cache block size, and have no more than 2^14 discard blocks
2173 * across the origin.
2175 #define MAX_DISCARD_BLOCKS (1 << 14)
2177 static bool too_many_discard_blocks(sector_t discard_block_size,
2178 sector_t origin_size)
2180 (void) sector_div(origin_size, discard_block_size);
2182 return origin_size > MAX_DISCARD_BLOCKS;
2185 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2186 sector_t origin_size)
2188 sector_t discard_block_size;
2190 discard_block_size = roundup_pow_of_two(cache_block_size);
2193 while (too_many_discard_blocks(discard_block_size, origin_size))
2194 discard_block_size *= 2;
2196 return discard_block_size;
2199 #define DEFAULT_MIGRATION_THRESHOLD 2048
2201 static int cache_create(struct cache_args *ca, struct cache **result)
2204 char **error = &ca->ti->error;
2205 struct cache *cache;
2206 struct dm_target *ti = ca->ti;
2207 dm_block_t origin_blocks;
2208 struct dm_cache_metadata *cmd;
2209 bool may_format = ca->features.mode == CM_WRITE;
2211 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2216 ti->private = cache;
2217 ti->num_flush_bios = 2;
2218 ti->flush_supported = true;
2220 ti->num_discard_bios = 1;
2221 ti->discards_supported = true;
2222 ti->discard_zeroes_data_unsupported = true;
2224 cache->features = ca->features;
2225 ti->per_bio_data_size = get_per_bio_data_size(cache);
2227 cache->callbacks.congested_fn = cache_is_congested;
2228 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2230 cache->metadata_dev = ca->metadata_dev;
2231 cache->origin_dev = ca->origin_dev;
2232 cache->cache_dev = ca->cache_dev;
2234 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2236 /* FIXME: factor out this whole section */
2237 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2238 origin_blocks = block_div(origin_blocks, ca->block_size);
2239 cache->origin_blocks = to_oblock(origin_blocks);
2241 cache->sectors_per_block = ca->block_size;
2242 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2247 if (ca->block_size & (ca->block_size - 1)) {
2248 dm_block_t cache_size = ca->cache_sectors;
2250 cache->sectors_per_block_shift = -1;
2251 cache_size = block_div(cache_size, ca->block_size);
2252 cache->cache_size = to_cblock(cache_size);
2254 cache->sectors_per_block_shift = __ffs(ca->block_size);
2255 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
2258 r = create_cache_policy(cache, ca, error);
2262 cache->policy_nr_args = ca->policy_argc;
2263 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2265 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2267 *error = "Error setting cache policy's config values";
2271 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2272 ca->block_size, may_format,
2273 dm_cache_policy_get_hint_size(cache->policy));
2275 *error = "Error creating metadata object";
2281 if (passthrough_mode(&cache->features)) {
2284 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2286 *error = "dm_cache_metadata_all_clean() failed";
2291 *error = "Cannot enter passthrough mode unless all blocks are clean";
2297 spin_lock_init(&cache->lock);
2298 bio_list_init(&cache->deferred_bios);
2299 bio_list_init(&cache->deferred_flush_bios);
2300 bio_list_init(&cache->deferred_writethrough_bios);
2301 INIT_LIST_HEAD(&cache->quiesced_migrations);
2302 INIT_LIST_HEAD(&cache->completed_migrations);
2303 INIT_LIST_HEAD(&cache->need_commit_migrations);
2304 atomic_set(&cache->nr_migrations, 0);
2305 init_waitqueue_head(&cache->migration_wait);
2307 init_waitqueue_head(&cache->quiescing_wait);
2308 atomic_set(&cache->quiescing, 0);
2309 atomic_set(&cache->quiescing_ack, 0);
2312 cache->nr_dirty = 0;
2313 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2314 if (!cache->dirty_bitset) {
2315 *error = "could not allocate dirty bitset";
2318 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2320 cache->discard_block_size =
2321 calculate_discard_block_size(cache->sectors_per_block,
2322 cache->origin_sectors);
2323 cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2324 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2325 if (!cache->discard_bitset) {
2326 *error = "could not allocate discard bitset";
2329 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2331 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2332 if (IS_ERR(cache->copier)) {
2333 *error = "could not create kcopyd client";
2334 r = PTR_ERR(cache->copier);
2338 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2340 *error = "could not create workqueue for metadata object";
2343 INIT_WORK(&cache->worker, do_worker);
2344 INIT_DELAYED_WORK(&cache->waker, do_waker);
2345 cache->last_commit_jiffies = jiffies;
2347 cache->prison = dm_bio_prison_create(PRISON_CELLS);
2348 if (!cache->prison) {
2349 *error = "could not create bio prison";
2353 cache->all_io_ds = dm_deferred_set_create();
2354 if (!cache->all_io_ds) {
2355 *error = "could not create all_io deferred set";
2359 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2361 if (!cache->migration_pool) {
2362 *error = "Error creating cache's migration mempool";
2366 cache->next_migration = NULL;
2368 cache->need_tick_bio = true;
2369 cache->sized = false;
2370 cache->invalidate = false;
2371 cache->commit_requested = false;
2372 cache->loaded_mappings = false;
2373 cache->loaded_discards = false;
2377 atomic_set(&cache->stats.demotion, 0);
2378 atomic_set(&cache->stats.promotion, 0);
2379 atomic_set(&cache->stats.copies_avoided, 0);
2380 atomic_set(&cache->stats.cache_cell_clash, 0);
2381 atomic_set(&cache->stats.commit_count, 0);
2382 atomic_set(&cache->stats.discard_count, 0);
2384 spin_lock_init(&cache->invalidation_lock);
2385 INIT_LIST_HEAD(&cache->invalidation_requests);
2395 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2400 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2403 for (i = 0; i < argc; i++) {
2404 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2413 cache->nr_ctr_args = argc;
2414 cache->ctr_args = copy;
2419 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2422 struct cache_args *ca;
2423 struct cache *cache = NULL;
2425 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2427 ti->error = "Error allocating memory for cache";
2432 r = parse_cache_args(ca, argc, argv, &ti->error);
2436 r = cache_create(ca, &cache);
2440 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2446 ti->private = cache;
2449 destroy_cache_args(ca);
2453 static int cache_map(struct dm_target *ti, struct bio *bio)
2455 struct cache *cache = ti->private;
2458 dm_oblock_t block = get_bio_block(cache, bio);
2459 size_t pb_data_size = get_per_bio_data_size(cache);
2460 bool can_migrate = false;
2461 bool discarded_block;
2462 struct dm_bio_prison_cell *cell;
2463 struct policy_result lookup_result;
2464 struct per_bio_data *pb;
2466 if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2468 * This can only occur if the io goes to a partial block at
2469 * the end of the origin device. We don't cache these.
2470 * Just remap to the origin and carry on.
2472 remap_to_origin_clear_discard(cache, bio, block);
2473 return DM_MAPIO_REMAPPED;
2476 pb = init_per_bio_data(bio, pb_data_size);
2478 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2479 defer_bio(cache, bio);
2480 return DM_MAPIO_SUBMITTED;
2484 * Check to see if that block is currently migrating.
2486 cell = alloc_prison_cell(cache);
2488 defer_bio(cache, bio);
2489 return DM_MAPIO_SUBMITTED;
2492 r = bio_detain(cache, block, bio, cell,
2493 (cell_free_fn) free_prison_cell,
2497 defer_bio(cache, bio);
2499 return DM_MAPIO_SUBMITTED;
2502 discarded_block = is_discarded_oblock(cache, block);
2504 r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2505 bio, &lookup_result);
2506 if (r == -EWOULDBLOCK) {
2507 cell_defer(cache, cell, true);
2508 return DM_MAPIO_SUBMITTED;
2511 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2513 return DM_MAPIO_SUBMITTED;
2516 r = DM_MAPIO_REMAPPED;
2517 switch (lookup_result.op) {
2519 if (passthrough_mode(&cache->features)) {
2520 if (bio_data_dir(bio) == WRITE) {
2522 * We need to invalidate this block, so
2523 * defer for the worker thread.
2525 cell_defer(cache, cell, true);
2526 r = DM_MAPIO_SUBMITTED;
2529 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2530 inc_miss_counter(cache, bio);
2531 remap_to_origin_clear_discard(cache, bio, block);
2533 cell_defer(cache, cell, false);
2537 inc_hit_counter(cache, bio);
2539 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
2540 !is_dirty(cache, lookup_result.cblock))
2541 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2543 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2545 cell_defer(cache, cell, false);
2550 inc_miss_counter(cache, bio);
2551 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2553 if (pb->req_nr != 0) {
2555 * This is a duplicate writethrough io that is no
2556 * longer needed because the block has been demoted.
2559 cell_defer(cache, cell, false);
2560 return DM_MAPIO_SUBMITTED;
2562 remap_to_origin_clear_discard(cache, bio, block);
2563 cell_defer(cache, cell, false);
2568 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2569 (unsigned) lookup_result.op);
2571 r = DM_MAPIO_SUBMITTED;
2577 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2579 struct cache *cache = ti->private;
2580 unsigned long flags;
2581 size_t pb_data_size = get_per_bio_data_size(cache);
2582 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2585 policy_tick(cache->policy);
2587 spin_lock_irqsave(&cache->lock, flags);
2588 cache->need_tick_bio = true;
2589 spin_unlock_irqrestore(&cache->lock, flags);
2592 check_for_quiesced_migrations(cache, pb);
2597 static int write_dirty_bitset(struct cache *cache)
2601 for (i = 0; i < from_cblock(cache->cache_size); i++) {
2602 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2603 is_dirty(cache, to_cblock(i)));
2611 static int write_discard_bitset(struct cache *cache)
2615 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2616 cache->discard_nr_blocks);
2618 DMERR("could not resize on-disk discard bitset");
2622 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2623 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2624 is_discarded(cache, to_dblock(i)));
2632 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2635 struct cache *cache = context;
2636 return dm_cache_save_hint(cache->cmd, cblock, hint);
2639 static int write_hints(struct cache *cache)
2643 r = dm_cache_begin_hints(cache->cmd, cache->policy);
2645 DMERR("dm_cache_begin_hints failed");
2649 r = policy_walk_mappings(cache->policy, save_hint, cache);
2651 DMERR("policy_walk_mappings failed");
2657 * returns true on success
2659 static bool sync_metadata(struct cache *cache)
2663 r1 = write_dirty_bitset(cache);
2665 DMERR("could not write dirty bitset");
2667 r2 = write_discard_bitset(cache);
2669 DMERR("could not write discard bitset");
2673 r3 = write_hints(cache);
2675 DMERR("could not write hints");
2678 * If writing the above metadata failed, we still commit, but don't
2679 * set the clean shutdown flag. This will effectively force every
2680 * dirty bit to be set on reload.
2682 r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2684 DMERR("could not write cache metadata. Data loss may occur.");
2686 return !r1 && !r2 && !r3 && !r4;
2689 static void cache_postsuspend(struct dm_target *ti)
2691 struct cache *cache = ti->private;
2693 start_quiescing(cache);
2694 wait_for_migrations(cache);
2696 requeue_deferred_io(cache);
2697 stop_quiescing(cache);
2699 (void) sync_metadata(cache);
2702 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2703 bool dirty, uint32_t hint, bool hint_valid)
2706 struct cache *cache = context;
2708 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2713 set_dirty(cache, oblock, cblock);
2715 clear_dirty(cache, oblock, cblock);
2720 static int load_discard(void *context, sector_t discard_block_size,
2721 dm_dblock_t dblock, bool discard)
2723 struct cache *cache = context;
2725 /* FIXME: handle mis-matched block size */
2728 set_discard(cache, dblock);
2730 clear_discard(cache, dblock);
2735 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2737 sector_t size = get_dev_size(cache->cache_dev);
2738 (void) sector_div(size, cache->sectors_per_block);
2739 return to_cblock(size);
2742 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2744 if (from_cblock(new_size) > from_cblock(cache->cache_size))
2748 * We can't drop a dirty block when shrinking the cache.
2750 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2751 new_size = to_cblock(from_cblock(new_size) + 1);
2752 if (is_dirty(cache, new_size)) {
2753 DMERR("unable to shrink cache; cache block %llu is dirty",
2754 (unsigned long long) from_cblock(new_size));
2762 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2766 r = dm_cache_resize(cache->cmd, cache->cache_size);
2768 DMERR("could not resize cache metadata");
2772 cache->cache_size = new_size;
2777 static int cache_preresume(struct dm_target *ti)
2780 struct cache *cache = ti->private;
2781 dm_cblock_t csize = get_cache_dev_size(cache);
2784 * Check to see if the cache has resized.
2786 if (!cache->sized) {
2787 r = resize_cache_dev(cache, csize);
2791 cache->sized = true;
2793 } else if (csize != cache->cache_size) {
2794 if (!can_resize(cache, csize))
2797 r = resize_cache_dev(cache, csize);
2802 if (!cache->loaded_mappings) {
2803 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2804 load_mapping, cache);
2806 DMERR("could not load cache mappings");
2810 cache->loaded_mappings = true;
2813 if (!cache->loaded_discards) {
2814 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2816 DMERR("could not load origin discards");
2820 cache->loaded_discards = true;
2826 static void cache_resume(struct dm_target *ti)
2828 struct cache *cache = ti->private;
2830 cache->need_tick_bio = true;
2831 do_waker(&cache->waker.work);
2837 * <#used metadata blocks>/<#total metadata blocks>
2838 * <#read hits> <#read misses> <#write hits> <#write misses>
2839 * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2840 * <#features> <features>*
2841 * <#core args> <core args>
2842 * <#policy args> <policy args>*
2844 static void cache_status(struct dm_target *ti, status_type_t type,
2845 unsigned status_flags, char *result, unsigned maxlen)
2850 dm_block_t nr_free_blocks_metadata = 0;
2851 dm_block_t nr_blocks_metadata = 0;
2852 char buf[BDEVNAME_SIZE];
2853 struct cache *cache = ti->private;
2854 dm_cblock_t residency;
2857 case STATUSTYPE_INFO:
2858 /* Commit to ensure statistics aren't out-of-date */
2859 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2860 r = dm_cache_commit(cache->cmd, false);
2862 DMERR("could not commit metadata for accurate status");
2865 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2866 &nr_free_blocks_metadata);
2868 DMERR("could not get metadata free block count");
2872 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2874 DMERR("could not get metadata device size");
2878 residency = policy_residency(cache->policy);
2880 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2881 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2882 (unsigned long long)nr_blocks_metadata,
2883 (unsigned) atomic_read(&cache->stats.read_hit),
2884 (unsigned) atomic_read(&cache->stats.read_miss),
2885 (unsigned) atomic_read(&cache->stats.write_hit),
2886 (unsigned) atomic_read(&cache->stats.write_miss),
2887 (unsigned) atomic_read(&cache->stats.demotion),
2888 (unsigned) atomic_read(&cache->stats.promotion),
2889 (unsigned long long) from_cblock(residency),
2892 if (writethrough_mode(&cache->features))
2893 DMEMIT("1 writethrough ");
2895 else if (passthrough_mode(&cache->features))
2896 DMEMIT("1 passthrough ");
2898 else if (writeback_mode(&cache->features))
2899 DMEMIT("1 writeback ");
2902 DMERR("internal error: unknown io mode: %d", (int) cache->features.io_mode);
2906 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2908 r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2910 DMERR("policy_emit_config_values returned %d", r);
2915 case STATUSTYPE_TABLE:
2916 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2918 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2920 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2923 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2924 DMEMIT(" %s", cache->ctr_args[i]);
2925 if (cache->nr_ctr_args)
2926 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2936 * A cache block range can take two forms:
2938 * i) A single cblock, eg. '3456'
2939 * ii) A begin and end cblock with dots between, eg. 123-234
2941 static int parse_cblock_range(struct cache *cache, const char *str,
2942 struct cblock_range *result)
2949 * Try and parse form (ii) first.
2951 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
2956 result->begin = to_cblock(b);
2957 result->end = to_cblock(e);
2962 * That didn't work, try form (i).
2964 r = sscanf(str, "%llu%c", &b, &dummy);
2969 result->begin = to_cblock(b);
2970 result->end = to_cblock(from_cblock(result->begin) + 1u);
2974 DMERR("invalid cblock range '%s'", str);
2978 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
2980 uint64_t b = from_cblock(range->begin);
2981 uint64_t e = from_cblock(range->end);
2982 uint64_t n = from_cblock(cache->cache_size);
2985 DMERR("begin cblock out of range: %llu >= %llu", b, n);
2990 DMERR("end cblock out of range: %llu > %llu", e, n);
2995 DMERR("invalid cblock range: %llu >= %llu", b, e);
3002 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3004 struct invalidation_request req;
3006 INIT_LIST_HEAD(&req.list);
3007 req.cblocks = range;
3008 atomic_set(&req.complete, 0);
3010 init_waitqueue_head(&req.result_wait);
3012 spin_lock(&cache->invalidation_lock);
3013 list_add(&req.list, &cache->invalidation_requests);
3014 spin_unlock(&cache->invalidation_lock);
3017 wait_event(req.result_wait, atomic_read(&req.complete));
3021 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3022 const char **cblock_ranges)
3026 struct cblock_range range;
3028 if (!passthrough_mode(&cache->features)) {
3029 DMERR("cache has to be in passthrough mode for invalidation");
3033 for (i = 0; i < count; i++) {
3034 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3038 r = validate_cblock_range(cache, &range);
3043 * Pass begin and end origin blocks to the worker and wake it.
3045 r = request_invalidation(cache, &range);
3057 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3059 * The key migration_threshold is supported by the cache target core.
3061 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3063 struct cache *cache = ti->private;
3068 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3069 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3074 return set_config_value(cache, argv[0], argv[1]);
3077 static int cache_iterate_devices(struct dm_target *ti,
3078 iterate_devices_callout_fn fn, void *data)
3081 struct cache *cache = ti->private;
3083 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3085 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3091 * We assume I/O is going to the origin (which is the volume
3092 * more likely to have restrictions e.g. by being striped).
3093 * (Looking up the exact location of the data would be expensive
3094 * and could always be out of date by the time the bio is submitted.)
3096 static int cache_bvec_merge(struct dm_target *ti,
3097 struct bvec_merge_data *bvm,
3098 struct bio_vec *biovec, int max_size)
3100 struct cache *cache = ti->private;
3101 struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3103 if (!q->merge_bvec_fn)
3106 bvm->bi_bdev = cache->origin_dev->bdev;
3107 return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3110 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3113 * FIXME: these limits may be incompatible with the cache device
3115 limits->max_discard_sectors = cache->discard_block_size * 1024;
3116 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3119 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3121 struct cache *cache = ti->private;
3122 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3125 * If the system-determined stacked limits are compatible with the
3126 * cache's blocksize (io_opt is a factor) do not override them.
3128 if (io_opt_sectors < cache->sectors_per_block ||
3129 do_div(io_opt_sectors, cache->sectors_per_block)) {
3130 blk_limits_io_min(limits, 0);
3131 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3133 set_discard_limits(cache, limits);
3136 /*----------------------------------------------------------------*/
3138 static struct target_type cache_target = {
3140 .version = {1, 2, 0},
3141 .module = THIS_MODULE,
3145 .end_io = cache_end_io,
3146 .postsuspend = cache_postsuspend,
3147 .preresume = cache_preresume,
3148 .resume = cache_resume,
3149 .status = cache_status,
3150 .message = cache_message,
3151 .iterate_devices = cache_iterate_devices,
3152 .merge = cache_bvec_merge,
3153 .io_hints = cache_io_hints,
3156 static int __init dm_cache_init(void)
3160 r = dm_register_target(&cache_target);
3162 DMERR("cache target registration failed: %d", r);
3166 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3167 if (!migration_cache) {
3168 dm_unregister_target(&cache_target);
3175 static void __exit dm_cache_exit(void)
3177 dm_unregister_target(&cache_target);
3178 kmem_cache_destroy(migration_cache);
3181 module_init(dm_cache_init);
3182 module_exit(dm_cache_exit);
3184 MODULE_DESCRIPTION(DM_NAME " cache target");
3185 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3186 MODULE_LICENSE("GPL");